UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Three essays in health economics : determinants of individual health, medical care use, and treatment Timmins, Lori L. 2015

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2015_february_timmins_lori.pdf [ 3.59MB ]
Metadata
JSON: 24-1.0167115.json
JSON-LD: 24-1.0167115-ld.json
RDF/XML (Pretty): 24-1.0167115-rdf.xml
RDF/JSON: 24-1.0167115-rdf.json
Turtle: 24-1.0167115-turtle.txt
N-Triples: 24-1.0167115-rdf-ntriples.txt
Original Record: 24-1.0167115-source.json
Full Text
24-1.0167115-fulltext.txt
Citation
24-1.0167115.ris

Full Text

Three Essays in Health Economics:Determinants of Individual Health,Medical Care Use, and TreatmentbyLori L. TimminsB.A. (Hons), The University of Winnipeg, 2003M.A., Queen’s University, 2004A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Economics)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)January 2015c© Lori L. Timmins 2015AbstractThis dissertation studies and identifies determinants of individual health. The first chapteranalyzes how the supply of medical care affects patient treatment and health outcomes, fo-cusing on how hospitals respond to the loss of a profitable service line. This chapter providesstrong evidence that hospital spillovers across service lines are empirically important andthat hospitals differentiate treatment by patient payer type. Hospitals practice both revenueaugmenting and cost-cutting behavior in other lines of care, targeting specific proceduresand payers according to their profitability. Specifically, they increase the number of surgicalprocedures and perform more marginal surgeries. The effects are concentrated in medical spe-cialties where there are more discretionary surgeries and higher profit margins. Furthermore,hospitals cut back on unprofitable treatment by reducing non-elective admissions and unin-sured elective care. Hospitals also increase the intensity of treatment among private payers.The second chapter of this dissertation investigates the demand side of health care, analyzingthe role that health insurance plays on primary medical care usage by young American adults.I find office-based physician visits and prescription drugs are not affected by insurance, butdental visits are. There is a small increase in out-of-pocket expenditures caused by insuranceloss, concentrated heavily at the top of the distribution. No change in health status or abilityto afford care is found. The findings shed light on the expected welfare benefits of recent UShealth care policies targeting young adults. The final chapter of this dissertation analyzesthe extent to which the early childhood environment shapes child health and developmentoutcomes and, specifically, whether universal childcare levels the playing field across children.I analyze the introduction of a universal childcare program in Quebec in 1997, testing itsimpact on the distribution of child health and development outcomes. I find that there islittle heterogeneity in the response to the policy across the distributions of child motor skillsand cognitive outcomes. I do, however, find evidence that it led to a reduction in child bodyweight at the upper end of the distribution.iiPrefaceThis dissertation is original, unpublished, independent work by the author, Lori L. Timmins.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 How Do Hospitals Respond to Financial Pain? Evidence from HospitalMarkets in Texas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Hospital Payments and Profitability . . . . . . . . . . . . . . . . . . . . . . . . 72.2.1 Payer Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2.2 Hospital Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.3 Specialty Hospitals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102.2.4 Possible Hospital Responses To Specialty Hospital Competition . . . . 132.3 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.4 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.4.2 Identifying the Marginal Impact of Increased Specialty Hospital Com-petition on Uncontested Care . . . . . . . . . . . . . . . . . . . . . . . 212.4.3 Estimating the Specialty Hospital Market Share . . . . . . . . . . . . . 252.4.4 Main Estimating Equation . . . . . . . . . . . . . . . . . . . . . . . . . 272.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.5.1 Volume of Admissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29ivTable of Contents2.5.2 Intensity of Treatment and Differences by Payer Type . . . . . . . . . . 342.5.3 Robustness to Alternative Specifications . . . . . . . . . . . . . . . . . 422.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433 How Much Does Health Insurance Matter for Young Adults?: Its Role onPrimary Medical Care Consumption . . . . . . . . . . . . . . . . . . . . . . . 453.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.2 Previous Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.3 Legislative Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.5 Empirical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.5.1 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.5.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.6.1 Change in Insurance Coverage Rates at age 19 . . . . . . . . . . . . . . 643.6.2 Change in Medical Care Consumption at age 19 . . . . . . . . . . . . . 713.6.3 Change in Routine Medical Expenditures at age 19 . . . . . . . . . . . 733.6.4 Change in Ability to Afford Medical Care at age 19 . . . . . . . . . . . 793.6.5 Change in Health Status at age 19 . . . . . . . . . . . . . . . . . . . . . 813.7 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 823.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904 Beyond the Mean: An Examination of Heterogenous Child Responses to aUniversal Childcare Policy in Quebec . . . . . . . . . . . . . . . . . . . . . . . 914.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 914.2 Previous Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.3 The Quebec Policy Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.4 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 994.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.5.1 The Standard DID Estimator . . . . . . . . . . . . . . . . . . . . . . . 1034.5.2 Quantile Difference-in-Differences (QDID) . . . . . . . . . . . . . . . . 1054.5.3 Estimating the QDID Model . . . . . . . . . . . . . . . . . . . . . . . . 1084.5.4 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164.6.1 The ITT Scaling Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . 1174.6.2 The Full Sample Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 1204.6.3 The Subsample Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 132vTable of Contents4.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1434.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1475 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150AppendicesA Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163A.1 Specialty Hospital Designation . . . . . . . . . . . . . . . . . . . . . . . . . . . 163A.2 Obtaining Patient to Hospital Distances . . . . . . . . . . . . . . . . . . . . . . 166A.2.1 Hospital Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166A.2.2 Patient Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167A.2.3 Patient to Hospital Distance . . . . . . . . . . . . . . . . . . . . . . . . 167A.3 Supplemental Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . 168B Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173B.1 Supplemental Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . 173viList of Tables2.1 Hospital Profitability by Medical Specialty . . . . . . . . . . . . . . . . . . . . . 102.2 Admissions by Hospital Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.3 Descriptive Statistics: Uncontested Medical Treatment by Insurance Type . . . 182.4 Top 15 Procedures in General Surgery . . . . . . . . . . . . . . . . . . . . . . . 192.5 Total Contested and Uncontested Hospital Admissions . . . . . . . . . . . . . . 312.6 Total Hospital Admissions by Surgery Type . . . . . . . . . . . . . . . . . . . . 322.7 Impact of Increased Specialty Competition on Share of Surgical Patients . . . . 352.8 Impact of Increased Specialty Competition on Types of Uncontested Surgeries . 372.9 Impact of Increased Specialty Competition on Length of Stay (LOS) . . . . . . 392.10 Impact of Increased Specialty Competition on Mortality Rate . . . . . . . . . . 413.1 Insurance Coverage by Age and Income Group . . . . . . . . . . . . . . . . . . 553.2 Medical Visits and Expenditures by Age and Income Group . . . . . . . . . . 573.3 Characteristics of Office-Based Physician Visits for Young Adults . . . . . . . 593.4 Health Status and Demographic Characteristics by Age and Income Group . . 603.5 Change in Insurance Coverage Rates at 19 . . . . . . . . . . . . . . . . . . . . 653.6 Change in Insurance Coverage Rates at 19 by Demographic Group . . . . . . . 683.7 Change in Insurance Coverage Rates at 19 by Income Group . . . . . . . . . . 703.8 Change in Medical Care Use at 19 . . . . . . . . . . . . . . . . . . . . . . . . . 723.9 Change in Expenditures on Office and Dental Visits at 19 . . . . . . . . . . . . 753.10 Quantile Treatment Effects for Office Expenditures at 19 . . . . . . . . . . . . 773.11 Change in Ability to Afford Care and Health Status at 19 . . . . . . . . . . . . 803.12 Robustness Checks for Other Variables Changing at 19 . . . . . . . . . . . . . 843.13 DID Estimates Examining Anticipatory Consumption Prior to Coverage Loss . 863.14 Robustness Checks for Misreported Insurance Coverage at 19 . . . . . . . . . . 894.1 MSD Percentiles by Time Period and Region . . . . . . . . . . . . . . . . . . . 1124.2 PPVT-R Percentiles by Time Period and Region . . . . . . . . . . . . . . . . . 1134.3 Body Weight Percentiles by Time Period and Region . . . . . . . . . . . . . . . 1144.4 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.5 Child Care Use Results by MSD, PPVT-R, and Body Weight Percentiles . . . 119viiList of Tables4.6 Standard DID Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1214.7 Quantile Treatment Effects for MSD . . . . . . . . . . . . . . . . . . . . . . . . 1254.8 Quantile Treatment Effects for PPVT-R . . . . . . . . . . . . . . . . . . . . . . 1264.9 Quantile Treatment Effects for Body Weight . . . . . . . . . . . . . . . . . . . . 1274.10 Quantile Treatment Effects for MSD - Subgroup Analysis . . . . . . . . . . . . 1344.11 Quantile Treatment Effects for PPVT-R - Subgroup Analysis . . . . . . . . . . 1354.12 Quantile Treatment Effects for Body Weight - Subgroup Analysis . . . . . . . . 136A.1 The Distribution of Predicted and Actual Specialty Market Shares . . . . . . . 169A.2 Impact of Increased Specialty Competition on Share of Surgical Patients . . . . 170A.3 Impact of Increased Specialty Competition on Length of Stay (LOS) . . . . . . 171A.4 Impact of Increased Specialty Competition on Mortality Rate . . . . . . . . . . 172B.1 Insurance Coverage by FPL Income Grouping- Descriptive Statistics . . . . . . 175B.2 Change in Medical Care Use at 19 by Gender . . . . . . . . . . . . . . . . . . . 176B.3 Change in Medical Care Use at 19, With and Without Covariates . . . . . . . 177B.4 Change in Expenditures on Office and Dental Visits at 19, With and WithoutCovariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178B.5 Change in Ability to Afford Care and Health Status at 19, With and WithoutCovariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179viiiList of Figures2.1 Number of Specialty Hospitals per county in Texas 1999-2007 . . . . . . . . . . 122.2 Distribution of Actual and Predicted Specialty Hospital Market Shares . . . . . 283.1 Insurance Coverage by Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663.2 Medical Care Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733.3 Expenditures on Office Visits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.4 QTE for OOP Expenditures on Office Visits . . . . . . . . . . . . . . . . . . . . 783.5 QTE for Total Expenditures on Office Visits . . . . . . . . . . . . . . . . . . . . 793.6 Indicators of Health Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 824.1 FFL Estimates for MSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1234.2 FFL Estimates for PPVT-R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1234.3 FFL Estimates for Body Weight . . . . . . . . . . . . . . . . . . . . . . . . . . 1244.4 Firpo Estimates for MSD, PPVT-R, Body Weight . . . . . . . . . . . . . . . . 1284.5 FFL Estimates with Collapsed Time Periods . . . . . . . . . . . . . . . . . . . 1304.6 FFL DDD Estimates for MSD and Body Weight . . . . . . . . . . . . . . . . . 1324.7 FFL Estimates by Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1374.8 FFL Estimates by Mother’s Education . . . . . . . . . . . . . . . . . . . . . . . 1384.9 FFL Estimates by Father’s Education . . . . . . . . . . . . . . . . . . . . . . . 1384.10 FFL Estimates by Father’s Wage . . . . . . . . . . . . . . . . . . . . . . . . . . 1394.11 FFL Estimates by Parenting Style- Hostile . . . . . . . . . . . . . . . . . . . . . 1404.12 FFL Estimates by Parenting Style- Aversive . . . . . . . . . . . . . . . . . . . . 1414.13 FFL Estimates by Parenting Style- Consistent . . . . . . . . . . . . . . . . . . . 1414.14 FFL Estimates by Family Functioning . . . . . . . . . . . . . . . . . . . . . . . 1424.15 FFL Estimates by Maternal Depression . . . . . . . . . . . . . . . . . . . . . . 1424.16 Trends in MSD Quantiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1454.17 Trends in PPVT-R Quantiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1464.18 Trends in Body Weight Quantiles . . . . . . . . . . . . . . . . . . . . . . . . . . 146A.1 Hospital Service Areas in Texas . . . . . . . . . . . . . . . . . . . . . . . . . . . 168B.1 Insurance Coverage by Demographic Group . . . . . . . . . . . . . . . . . . . . 173ixList of FiguresB.2 Type of Office Visit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174B.3 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174xAcknowledgementsThe past six years have been both rewarding and challenging. I would like to express mysincere gratitude to all of those who supported me in different ways throughout the Ph.D.program. Without them, I would not have been able to complete this dissertation.Most importantly, I would like to thank my supervisors Marit Rehavi and Kevin Milliganfor their extraordinary support and guidance. Both of them were always encouraging, un-wavering in their support, and gave me confidence at times when it was most needed. Theyplayed pivotal roles in helping me succeed and making me a stronger researcher. Kevin wasalways diligent in his supervision and most generous in sharing his knowledge, time, andresources. Marit has not only been a great supervisor, but a wonderful mentor and a truerole model in the field. I immensely benefitted from her expertise, her attention to detail,as well as her patience and kindness. Second, I would like to thank Nicole Fortin for alwaysbeing available and having my best interests in mind, and for her supervision throughout theprogram. I am extremely grateful to Josh Gottlieb, whose invaluable feedback and sharp eco-nomic intuition strengthened my work immensely. I am also grateful to Craig Riddell, DavidGreen, Thomas Lemieux, Nancy Gallini, and other faculty members for their valuable com-ments and feedbacks. I would like to express my gratitude to the Canadian Labour Marketand Skills Researcher Network (CLSRN) for their continued financial support throughout myPhD. I am also appreciative to Maureen Chin for all the administrative support and guidancethroughout the course of program. I thank the staff at the Center for Health Statistics at theTexas Department of State health Services for providing the data used in the dissertation andfor answering questions. I also thank researchers at the Research Data Centre for providingdata and for their assistance.I want to thank my outstanding colleagues and friends at UBC, with whom I shared manylong days, laughs, and coffees. Haimin, Terri, Antoine, Andrew, Dan, Alix, Donna, Jeanne,and last but not least, Jocelyne, all made this journey more fulfilling. I thank my colleaguesin the public finance reading group for their generous feedback and support in the last yearof the program.I would like to express my deepest gratitude to my parents for their unconditional support,encouragement, and love. I would not be where I am without either of them. I would alsolike to thank my sister, Melanie, for always supporting me and my family whenever needed.Finally, I thank my very best friend and my partner in crime, Andy, for his patience,xiAcknowledgementsdedication, and love throughout the course of the program. It wasn’t easy, but with himbehind me each step of the way, it was possible.xiiDedicationThis dissertation is dedicated to my mom, my number one fanFor her endless love, belief, and supportHer spirit, strength, and determination are with mexiiiChapter 1IntroductionThis dissertation studies and identifies determinants of health. It focuses on distinct dimen-sions that may contribute to health differences across individuals and over time. It analyzeshow the supply of medical care affects patient treatment and health outcomes, focusing onthe hospital setting and the impact of hospital competition. It also investigates the demandside of health care, by examining the role that health insurance plays on medical care usageby young American adults. Finally, it analyzes the extent to which the early childhood en-vironment, specifically the role of childcare, shapes child health and development outcomesand whether universal childcare levels the playing field across children. The unifying themeof each chapter is the focus on analyzing factors that may shape individual health.The first essay of this dissertation studies how hospitals respond to profit shocks andthe loss of profitable service lines. I use the penetration of specialty hospitals, which offera subset of procedures with high profit margins, as a supply shock for these services in ahospital market. I analyze incumbent hospital behavior in other service lines, focusing onwhether hospitals adjust their procedures and, if so, whether it varies by the patient’s payertype. I find that incumbent hospitals have a more sophisticated, targeted response thanfound in previous research. Greater specialty hospital market shares cause incumbents toincrease the number of surgical procedures and perform more marginal surgeries. This variesby service line and payer type. I find that the effects are concentrated in medical specialtieswhere there are more discretionary surgeries and higher profit margins. The increase is onlyamong private payers whose insurance reimburses hospitals more generously. Hospitals alsoincrease the intensity of treatment among private payers, by increasing their length of stayconditional on the procedure. Furthermore, I find that hospitals cut back on unprofitabletreatment by reducing emergency department admissions and uninsured elective care. Myfindings provide empirical evidence that hospitals cross-subsidize both across procedures andpatients. This suggests that hospital spillovers are empirically important and that just lookingat substitution within a service line when evaluating health care policies ignores importanthospital responses and subsequent welfare implications, particularly among different payertypes.In the next chapter, I shift from the supply of health care to the demand for medical care.In particular, I examine the role of health insurance in shaping young American adults’ pri-mary medical care use. I obtain casual estimates using a regression discontinuity framework1Chapter 1. Introductionthat exploits insurance policy rules where individuals cease being covered on their 19th birth-day. I find office-based physician visits and prescription drugs are not affected by insurance,but dental visits are. There is a small increase in out-of-pocket expenditures, concentratedheavily at the top of the distribution. No change in health status or ability to afford careis found. The findings shed light on the expected welfare benefits of recent US health carepolicies targeting young adults.The final chapter of this dissertation analyzes the extent to which the early childhoodenvironment affects child health and development. Its focus is on the role that childcare typeplays in shaping motor skills, cognitive outcomes, and physical health of young children andwhether universal childcare levels the playing field across children. In particular, I analyzethe impact of a universal childcare policy in Quebec on the distributions of child health anddevelopment outcomes. In 1997, the Quebec government began offering reduced rate spacesfor $5 a day which was accessible to families from all economic and educational backgrounds.I estimate the impact of the reform on the marginal distribution of outcomes in the short runusing a quantile difference-in-differences model for two-parent families. I find that there islittle heterogeneity in the response to the universal childcare policy across the distributionsof motor skills and cognitive outcomes. In fact, this study finds that the policy had littlesignificant effect on these outcomes at any point along the distributions, neither for the fullsample of children nor when the sample is split by child demographic characteristics. I do,however, find evidence that the universal childcare policy led to a reduction in child bodyweight at the upper end of the distribution. These results are robust to different specificationsand estimation techniques.2Chapter 2How Do Hospitals Respond toFinancial Pain? Evidence fromHospital Markets in Texas2.1 IntroductionIt is widely believed that cross-subsidization is the primary means by which hospitals provideunprofitable care. Revenue from profitable procedures and patients are used to subsidize un-profitable hospital visits.1 Implicit in this belief is that hospital departments do not operateindependently from one another. Yet the bulk of empirical studies that analyze changes inpolicies, profits, or prices targeting a particular hospital service line largely ignore how suchchanges may impact the provision of services in other hospital departments, and hospitalspillovers across service lines may be quantitatively large. Additionally, they may be concen-trated in particular procedures, types of visits, and patients. Failing to properly take intoaccount such spillovers could result in incomplete welfare predictions.To date, very little empirical evidence of hospital cross-subsidization exists. There is anextensive literature that shows cross-subsidization is prevalent in other industries, such as inairline, railway, and telecommunications.2 However, the health care industry is unique. First,it is heavily regulated, where prices are administratively set. Second, patients often do notpay the full cost of their treatment because of insurance. Third, hospitals price discriminate.Fourth, there is asymmetric information whereby physicians make clinical decisions on behalfof their patients about treatment. Therefore, the study of hospitals can add substantively tothe existing literature and to understanding hospital behavior.One of the few studies to examine hospital cross-subsidization is David et al. (2014),who find that hospitals reduce the volume of admissions in unprofitable departments, par-ticularly psychiatric, substance-abuse, and trauma care when they lose a profitable serviceline (cardiac care). The focus of their study, however, is largely on the extensive margin ofunprofitable department admissions. Hospitals may adjust in other ways, such as increasing1See Gruber (1994); David et al. (2014); Norton & Staiger (1994); Horwitz (2005); Banks et al. (1999).2See Chevalier (2004); Kaserman & Mayo (1994); Banks et al. (1999).32.1. Introductionthe volume of profitable procedures, particularly elective procedures and procedures for whichthere is more clinical discretion in the course of treatment.3 In addition to operating on theextensive margin, hospitals can alter the intensity of treatment for a given condition, such asperforming more marginal surgeries or adjusting the length of stay. Importantly, hospitalsmay differentially target patients by insurance type, providing treatment on a case-by-casebasis. In particular, they may try to augment revenue by increasing the quantity and inten-sity of profitable care to patients whose insurance reimburses most generously (e.g. privatepayers), and they may cut back on care to unprofitable patients (e.g. uninsured).A related open question is whether hospitals can differentially target medical treatmentby payer type. It has long been recognized that hospitals can differentially set fees acrosspayer types and price discriminate (Kessel (1958)). There is also a lengthy literature onthe incidence of “cost-shifting”, whereby hospitals increase fees among patients with moregenerous reimbursement to make up for losses from less generous insurers.4 Many studiesanalyze the impact of reimbursement changes on own payer outcomes or on substitutionacross payers.5 However, as McGuire & Pauly (1991) contend in their multi-product andmulti-payer model of physician behavior, physicians can both substitute towards patientswith more generous reimbursement and change the mix of services they provide to specificpayers.6 Few studies explicitly test for differential medical treatment by payer type. Dor& Farley (1996) is a noticeable exception, finding some evidence that service intensity andquality might differ among different types of publicly insured patients.7 Studies of office-based physician practices find the role of insurance is limited in individual treatment decisions(e.g. Glied & Zivin (2002)).8 However, hospitals have more administrative staff and greaterresources to target treatment at the individual level. Furthermore, they tend to performhigher cost procedures so the marginal benefit of tailoring treatment to an individual patientmay be higher.This paper contributes to the existing literature by providing a more complete picture ofthe nature of hospital spillovers, treatment differences across payer types, and more broadly, ofhow hospitals respond to financial shocks. In particular, this study makes three key contribu-tions. First, it tests for cross-subsidization and determines where spillovers are concentrated,3In this study, I define scheduled visits as “elective” and non-scheduled visits as “non-elective”.4See Dranove (1988); Zuckerman (1987); Wu (2010) and Frakt (2011).5For example, Cutler (1995b)’s study on Medicare’s shift to prospective payment uses a sample of Medicarepatient to examine its impact on patient outcomes. Others have used payers without a reimbursement changeas a control group (e.g. Langa & Sussman (1993)). However, estimates will be biased if they are also affected.6There is some evidence that lower Medicaid reimbursement rates reduce service levels across all payertypes in hospitals, with greatest effects among Medicaid patients (Dranove & White (1998)).7While recognized as a pioneering study, its approach and estimates have been scrutinized as being drivenby omitted variable bias (Danger & Frech (1997)). Dor & Farley (1996) acknowledge they face severe datalimitations which led to fairly homogenous private payer type groupings and imprecise estimates. My paperdoes not suffer from these issues.8See the findings of (Glied & Zivin (2002), Tai-Seale et al. (2007)), which are in contrast to Newhouse &Marquis (1978).42.1. Introductionanalyzing both the extensive and intensive margins. Like David et al. (2014), I examinechanges in the volume of admissions across service lines. In addition, however, I study theintensity of treatment within particular services lines and test for heterogeneity across patienttypes. This paper analyzes just how sophisticated is the hospital response to profit shocks andprovides a deeper understanding of hospital spillovers and hospital behavior more generally.Second, this paper contributes to the literature on whether hospitals differentiate treat-ment by payer type, treating patients on a case-by-case basis. Surprisingly, this has receivedlittle attention to date in the hospital setting. The existence of payer-specific differences inthe intensity of treatment would suggest that there is a range of acceptable treatment options.However, some may not only impose greater health care costs but also more patient risk.Finally, this study contributes to the broader literature on how hospitals respond to fi-nancial shocks. Existing influential studies largely analyze the hospital response to pricechanges that explicitly alter incentives between different medical procedures and/or differentpayer types (e.g. Dafny (2005), Cutler (1995b)). Many studies also analyze how hospitalsrespond to the receipt of lump sum government payments or to global budget shocks (e.g.Duggan (2000), Dranove et al. (2013)). Instead, this paper examines a novel type of financialshock, namely the loss of a profitable service line. I provide a more complete picture of hos-pital behavior by analyzing spillovers across service lines, cross subsidization, and treatmentdifferences across payer types.In this study, I take advantage of a unique natural institutional feature in Texas to an-alyze how hospitals respond to a decline in revenue from their most profitable service lines.Specifically, I use the penetration of specialty hospitals, which concentrate on a subset ofprocedures with high profit margins as a supply shock for these services in Texas health caremarkets.9 I measure the response of general hospitals to the loss of profitable service lines,with particular attention to their behavior in other service lines. I test whether some typesof visits are more affected than others, based on their overall profitability (e.g. surgical vs.non-surgical), the nature of the visit (e.g. elective vs. non-elective), and the intensity oftreatment. Importantly, I test whether the response varies by payer type. There has beena surge in specialty hospital penetration in Texas over the last decade, leading it to be thestate with the greatest number and proportion of specialty hospitals. This has been met withstrong opposition. Many policy debates center on their impact on general hospitals’ abilityto provide less profitable care. This study sheds light on how general hospitals are affected.Specialty hospital entry into a market is not random, nor is the market share it capturesonce entered. To address these issues, I estimate the predicted demand for specialty hospi-9Texas is the state with the greatest number and proportion of specialty hospitals. David et al. (2014)follow a similar approach, using the entry of specialty hospitals in Arizona to test if hospitals cut back onunprofitable admissions. They use hospitals in Colorado as their control group. While my study tests otherdimensions beyond the extensive margin of care, I also use a different empirical methodology. I test the generalhospital response by using within market variation in the specialty hospital market shares over time.52.1. Introductiontals in a market using a patient-level hospital choice model in combination with instrumentalvariables and exploit within market variation over time. I build on the two-step approachof Kessler & McClellan (2000), by first modelling patient demand for specialty services toestimate the predicted market share of specialty hospitals, and then estimating the impact ofpredicted specialty hospital market share on non-specialty services at general hospitals. Fol-lowing previous studies, I use patients’ geographic location of residence relative to specialtyhospitals as an instrument for hospital choice.10 Unlike previous studies, however, I look atmedical treatment for a different set of individuals than those used to predict the specialtyhospital market share. As such, the identifying assumption only requires that distances tospecialty hospitals for patients obtaining specialty care do not directly affect medical treat-ment outcomes for patients seeking non-specialty care at general hospitals (except throughspecialty hospital market share). I also use a rich set of patient, hospital, and market char-acteristics, with market fixed effects and time trends to account for any unobserved factorsthat may be correlated with both specialty hospital market entry and non-specialty medicaltreatment.This paper provides strong evidence that hospital spillovers are empirically important.I find that specialty hospitals steal patients from general hospitals, causing a reduction inspecialty admissions at general hospitals. In turn, general hospitals employ a sophisticated,targeted response in their non-specialty service lines. They practice both revenue augmentingand cost-cutting behavior and adjust treatment by payer type. In particular, I first findthat hospitals make up for the lost volume of specialty surgeries by increasing the number ofsurgeries performed in other service lines. They do this by performing more marginal surgeries.The effects are concentrated in general surgery, a relatively high profit medical specialty wherehospitals have more discretion in treatment due to clinical grey areas. Aligned with this, I findan increase in the number of elective (i.e. scheduled) general surgery admissions. In addition,I show that hospitals do not only augment revenue by increasing profitable procedures, butthey also cut back on unprofitable procedures, particularly the number of non-elective (i.e.emergency) admissions.Secondly, my results provide strong evidence that hospitals vary treatment by payer type,suggesting that they treat patients on a case-by-case basis. In particular, hospitals targetprivate payers whose insurance reimburses hospitals more generously. Increased specialtyhospital penetration leads to a greater proportion of private payers with non-specialty surgicaladmissions, both across and within hospital departments. Effects are concentrated in electivesurgeries, with large increases in the share of private payers with a general surgery admission.I also find strong evidence that hospitals treat private payers more intensively, by increasingtheir length of stay. This effect is not entirely driven by an increase in surgical procedures10The work of Kessler & McClellan (2000), Chernew et al. (2002), Li & Dor (2013), and Swanson (2012)all use distance between hospitals and patients as an instrumental variable as part of their empirical strategy.62.2. Hospital Payments and Profitabilityamongst private payers, as the length of stay increases even when factoring in patients’ medicalprocedures. A notable finding is that no change in the length of stay is found for publicpayers. While private payers reimburse hospitals for each additional hospital day (i.e. per-diem), public payers reimburse a lump sum amount per admission. Additionally, hospitals cutback on care to the uninsured, with a smaller proportion of uninsured having an elective visit.Unlike the literature on office-based physician practices, my findings suggest that hospitalsdo target treatment by payer type.Finally, I find suggestive evidence that increased specialty hospital competition may putpatients obtaining non-specialty care at greater medical risk, with an increase in the mortalityrate. This holds even when I control for observable measures of health severity. I cannot,however, rule out that this is being driven by unobservable changes in the composition ofpatient health within a given diagnosis group.The rest of the paper proceeds as follows. In the next section, I provide backgroundinformation on how hospitals are reimbursed for different procedures and payer types. I alsodiscuss the origins and growth of specialty hospitals in Texas. In Section 3, I describe thedata. In Sections 4 and 5, I present the empirical model and results. I conclude in Section 6.2.2 Hospital Payments and ProfitabilityIn this section, I provide background information on hospital reimbursement to motivate thedifferent margins which hospitals may adjust to the loss of a profitable service line. I outlinehow reimbursement differs across payer types, with private payers typically being the mostprofitable and the uninsured the least. I also discuss how profitability varies across medicalspecialties, with surgical care being highly profitable. Then, I outline the origins and growthof specialty hospitals in Texas, and I discuss the possible response to increased specialtyhospital competition by general hospitals in non-specialty care.2.2.1 Payer TypesThere is substantial variation in the prices paid by insurers to hospitals for care. WhileMedicare payment rates are publicly available, the prices paid by other insurers are difficultto observe. Although insurers typically do not pay the full hospital list charges, it is thoughtthat private payers reimburse at the highest rates, followed by Medicare, and then Medicaid(Morrisey (1994); Dor & Farley (1996)).11Public payers (i.e. Medicare and Medicaid) set payments to the providers. Medicare isthe largest health insurance program in the world, and all Americans over 65 years old areeligible for coverage. Medicare pays hospitals a lump sum per admission, with the amount11Ellis (2001) provides an excellent overview of hospital reimbursement in the U.S.72.2. Hospital Payments and Profitabilitydepending, in part, on the patient’s principal disease. The reimbursement scheme reflectsexpected resource use and is based on average costs, not marginal costs. Medicaid is a federaland state funded program that targets very low income families, specifically children andpregnant women near the federal poverty line. The Medicaid eligibility rules for Texas areamong the most stringent in the country.12 Texas is one of the states that has decided notto expand Medicaid under the Affordable Care Act. Medicaid is well known for providinglow reimbursement rates, often below hospitals’ costs (Chernew et al. (2002)). In Texas,hospitals are reimbursed by Medicaid in a similar fashion as Medicare, with a fixed amountper inpatient episode of treatment.In contrast to public insurers, private payers negotiate payments with providers througha bargaining process (Ho (2009); Clemens & Gottlieb (2013)). There is an array of privateinsurance plans. In my study, I separate private payers into Health Maintenance Organizations(HMOs) and non-HMOs.13 Those in the latter group include indemnity plans and PreferredProvider Organizations (PPO). These patients are considered to be the most lucrative tohospitals. Although there is some variation in payment, this group of patients are generallyconsidered to pay fee-for-service (FFS). That is, hospitals are paid for each service theyprovide and/or on a per-diem basis. Some private insurers also reimburse hospitals with alump sum payment. HMOs differ from other private insurers in how they are organized. Theycontract selectively with only some hospitals in a given area and exert stricter gatekeeping,requiring non-urgent hospital visits to be referred through a general practitioner and to bepre-authorized. There is variation in how HMOs reimburse hospitals. In general, however,HMOs pay hospitals similar to FFS, providing payment for each service and/or on a per-diembasis, although at more discounted rates. Some HMOs also reimburse hospitals with a lumpsum payment that is fixed for each inpatient visit, similar to Medicare.14Individuals without insurance either reimburse hospitals for some or all of the charges(self-pay) or are charity care (i.e. uncompensated care). Texas has the highest percentageof residents without health insurance in the country. The Census Bureau estimated 6.4million Texans had no health coverage in 2012 (25% of its population). Self-pay patientsare profitable only if hospitals are able to recoup their costs since, unlike private insurers,there is no bargaining process in prices. In general, however, uninsured patients are thoughtto be unprofitable for hospitals to care for. It is argued that hospitals provide unprofitable care12The current eligibility rules are: 133% of the federal poverty line for children aged 1-5; 100% of the federalpoverty line for children 6-18 years old; and 185% of the federal poverty line for children under 1 year andpregnant women. Adults with children are eligible only if family income is at or below 26% of the federalpoverty line.13Note that Medicare Advantage plans reimburse hospitals the same as traditional Medicare so these patientsare categorized as Medicare.14Kaiser Permanente, which is a well-known vertically integrated HMO system that has its own hospitalsand physician practices, did not operate in Texas throughout my sample period. It stopped operating in Texasin 1998.82.2. Hospital Payments and Profitabilityfor various reasons which may vary by the hospital type. Non-profit hospitals are believedto be socially motivated (Frank & Salkever (1991) and Gruber (1994)). In addition, theymust provide a certain level of uncompensated care (i.e. charitable care) in order to beexempt from local, state, and federal taxes. Meanwhile, Gray (1991) argues that for-profithospitals provide unprofitable care as a business decision. They do so to strengthen their localreputation and increase business in more profitable types of care; to reduce the likelihood ofcivil liability and Medicare sanctions; and to avoid tangible community penalties (Bankset al. (1997)). Medicare also provides funding to hospitals with a disproportionate number ofuninsured and Medicaid patients under the Disproportionate Share Hospital (DSH) program.In addition, it should be noted that under the Emergency Medical Treatment and Active LaborAct (EMTALA), hospitals are required to treat all patients with life-threatening medicalepisodes, regardless of their ability to pay. Patients cannot be discharged until they havebeen stabilized.15 Hospitals are not required to treat patients whose life is not in immediatejeopardy.2.2.2 Hospital ServicesIn addition to treating multiple types of payers, general hospitals provide a variety of medicalservices ranging from neurology to obstetrics to cardiology. There is significant variation inthe profitability of departments and procedures, with some generating huge rents and othersa loss for hospitals. One factor that contributes to this variation is the prevalence of admin-istered pricing in the medical care industry (Newhouse (2002); Horwitz (2005)). Medicarein particular creates differential rents across specialties. Medicare provides higher reimburse-ments to specialists whose work is predominately hospital based (as opposed to outpatientbased), such as cardiovascular surgeons or neurosurgeons. Administered prices are also noto-riously sticky. When procedures are first introduced, productivity tends to be low, but overtime productivity improves with learning-by-doing and the cost of technology falls, which alsocreates rents in some specialties (Newhouse (2002)). As described, the Medicare reimburse-ment scheme reflects expected resource use and is based on average costs, not marginal costs.This can create distortions by giving hospitals an incentive to expand services which have thelargest difference between average and marginal costs (Kim (2011)).A list of the most and least profitable hospital specialties is provided in Table 2.1. Thisinformation is based on the findings of Lindrooth et al. (2013), Horwitz (2005), and Resnicket al. (2005). Departments performing surgical-intensive procedures, such as thoracic surgery,cardiovascular surgery, and neurosurgery are the most profitable. General surgery is also ahighly profitable department, performing a range of procedures from gallbladder surgeries to15EMTALA was passed in 1986 by the U.S. Congress as part of the Consolidated Omnibus Budget Recon-ciliation Act (COBRA). All hospitals that accept Medicare payments must abide by this act or else they forgoMedicare payment. This means that in practice, the act applies to virtually all hospitals in the country.92.2. Hospital Payments and Profitabilitymastectomies, as is Urology, carrying out a large number of urethral and prostatic surgeries.Less profitable departments perform few surgeries, such as Otolaryngology (ears, nose andthroat) and Nephrology (kidneys). Emergency department and psychiatric admissions areunprofitable service lines. As discussed, hospitals are thought to use the charges from theirmost profitable procedures and patients to cross-subsidize unprofitable care (Gruber (1994),David et al. (2014), and Banks et al. (1999)). To date, however, there are limited empiricalstudies that test if this is the case.Table 2.1: Hospital Profitability by Medical SpecialtyMost ProfitableThoracic SurgeryCardiovascular SurgeryNeurosurgeryGeneral SurgeryProfitableSurgical OrthopedicsUrologyOncologyGynecologyGeneral MedicineLess ProfitablePulmonologyGastroenterologyNephrologyOtolaryngologyCardiologyNeurologyMedical OrthopedicsUnprofitableEmergency DepartmentHospice CarePsychiatryNotes: Profitability status was assigned by compiling information from Lin-drooth et al. (2013), Horwitz (2005), and Resnick et al. (2005). Lindroothet al. (2013) calculate Medicare markups to assign specialty profitability.Horwitz (2005) determines profitability using information from peer-reviewedmedical and social science literature, government reports, and interviews withhospital administrators and doctors. Resnick et al. (2005) use hospital fi-nance department data to determine the profitability of surgical specialties.2.2.3 Specialty HospitalsUnlike general hospitals which provide a range of services, specialty hospitals concentrate onprocedures performed in the most profitable specialties. They largely provide three types ofcare: cardiac, orthopedic, or surgical (cardiac and/or orthopedic is the most common type102.2. Hospital Payments and Profitabilityof surgery performed at surgical hospitals). A surge in specialty hospitals occurred followingthe passage of the Stark law in the Omnibus Budget Reconciliation Act (OBRA) of 1993,which declared that physician owners were allowed to refer patients to their own hospitalsprovided they had investment interest in the whole hospital.16 This led to significant growthin a new type of specialty hospital, namely physician-owned hospitals providing profitablesurgical procedures.17Despite their growth over the last 15 years, specialty hospitals are highly controversial.Proponents argue they are focused factories (Herzlinger (1997); Skinner (1974)). By offering alimited range of services, specialty hospitals allow physicians to produce care more efficientlyand with higher quality.18 Proponents also argue that specialty hospitals spur system-wideinnovation through increased competition (Barro et al. (2006)). Critics of specialty hospitals,meanwhile, contest that they cream skim the most profitable patients and undermine com-munity hospitals’ ability to subsidize the less profitable patients and services (US Congress(2006)).19 Physician investors argue that the primary reason they form a hospital is for greatercontrol in determining the course of medical treatment. Profits are said to be secondary (USCongress (2006)).Although this controversy has led many states to ban specialty hospitals, they have flour-ished in the state of Texas. Figure 2.1 shows the growth of specialty hospitals in Texas overtime and space.20 In 1999, 58% of patients lived within 50 miles of a specialty hospital inTexas. This figure rose to 84% by 2007. Between 1999 and 2007, the number of specialtyhospitals more than tripled from 14 to 50.21 Specialty hospitals are not only concentratedin larger urban areas, such as Dallas and Houston, they are also prevalent in small citiessuch as Amarillo, Edinburg, and Odessa. Additionally, while cardiac care is amongst themost profitable, orthopedic and surgical specialty hospitals are more widespread. Among the50 specialty hospitals that existed in 2007, 8 were cardiac, 27 were orthopedic, and 15 weresurgical.16Specifically, this provision was known as the “Stark II”, following “Stark I” of OBRA 1989 which bannedself-referrals for clinical laboratory services. The exemption described above is known as the “whole hospitalexception”.17In Texas, 91% of specialty hospitals are for profit, and among these, 93% are physician owned.18They are likely to be substitutes rather than complements to general hospitals, performing similar typesof routine surgeries (e.g. catherization, angioplasty, hip replacements).19Swanson (2012) finds evidence of patient sorting across hospital types by medical complexity, rather thancherry picking on unobserved severity.20See Appendix A for details on how specialty hospitals were defined and identified in the data.21The rate of growth of specialty hospitals slowed in the later years with a moratorium on new physician-owned specialty hospitals that were not already under development. In particular, Congress enacted theMedicare Prescription Drug, Improvement and Modernization Act (MMA) of 2003, which legislated a tem-porary 18 month moratorium on new specialty hospitals, beginning in November 2003. The purpose of themoratorium was to allow the secretary of Health and Human Services (HHS) and MedPAC time to study theimpacts of specialty hospitals and to make recommendations to Congress. The moratorium was extended bythe CMS until August 2006 when it began to accept new applications for specialty hospitals.112.2.HospitalPaymentsandProfitabilityFigure 2.1: Number of Specialty Hospitals per county in Texas 1999-2007!!!!!!!!DallasHoustonAustinAmarilloSan AntonioCorpus ChristiWichita FallsEdinburgSpecialty Hospitals0123456789Total Specialty Hospitals in Texas: 14(a) Specialty Hospitals in 1999!!!!! !!!!!!!LaredoOdessa AustinDallasLubbockEl PasoHoustonEdinburgAmarilloSan AntonioWichita FallsCorpus ChristiSpecialty Hospitals0123456789Total Specialty Hospitals in Texas: 50(b) Specialty Hospitals in 2007122.2. Hospital Payments and ProfitabilityTexas is the state with the greatest number and proportion of specialty hospitals in theU.S., making it a rich setting to analyze how general hospitals respond to increased competi-tion in their most profitable service lines and to test for hospital spillovers across services andpayer types.22 Although the growth in specialty hospitals across Texas has been phenome-nal, it is likely to be short lived. The Patient Protection and Affordable Care Act (ACA) of2010 has banned physician investment in hospitals due to their controversy, although existingspecialty hospitals can be grandfathered in.2.2.4 Possible Hospital Responses To Specialty Hospital CompetitionProfit maximization is one of the key objectives of hospitals, even among non profits (Frank& Salkever (2000)).23 In response to increased competition from specialty hospitals, generalhospitals may try to make up lost revenue by expanding profitable care and cutting backunprofitable care. There are multiple ways they can do this.First, general hospitals may respond to the increased competition in specialty services bychanging the mix of their non-specialty care. In particular, hospitals may divert resourcestowards higher profit procedures from unprofitable services. Frank & Salkever (2000) develop atheoretical model that shows that in the face of financial pressure, hospitals shift the supply ofservices in the direction of more profitable services. Although physicians clearly must ensurepatients receive adequate medical care, there is somewhat of a clinical grey area for someprocedures. For example, there are certain illnesses that have multiple treatment possibilities(e.g. gallstones) and there are some conditions that are discretionary (e.g. obesity procedures)which do not always necessitate surgical care. These types of procedures are more prevalentin general surgery, as opposed to say, neurosurgery, which has fewer clinical grey areas.24One response then to increased specialty hospital competition would be for general hospitalsto perform more marginal surgeries, particularly general surgeries.25 However, as noted byHorwitz et al. (2013), this type of behavior may cause overuse of some procedures withoutclear medical guidance and may contribute to rising health care costs.2622This rich setting enables me to use a different empirical methodology than David et al. (2014). Whereasthese authors use hospitals in another state as controls to test for cross-subsidization, I exploit within marketvariation over time.23Frank & Salkever (2000) carry out focus groups with hospital administrators and find that profit maxi-mization is one of the key objectives of hospitals, even among non-profits. This is consistent with the findingsof other authors, such as Duggan (2000) and Sloan (1998), who focus on non-profit hospital behavior.24Interestingly, David et al. (2014) finds that cardiac specialty hospital entry increases the number of neu-rosurgeries in a hospital market. This is somewhat surprising given the clinical guidelines for these procedures,such as craniotomies, are more stringent.25Since physicians are paid separately from hospitals for each procedure performed, the financial incentivesto expand surgeries are likely to be aligned between hospitals and physicians. Furthermore, hospitals can exertsignificant influence on physicians. In particular, physicians must forge and maintain good relationships withthe hospital so that they may obtain admitting privileges, a requirement to treat their patients there.26Horwitz et al. (2013) analyze cardiac products and find hospitals in more competitive cardiac marketstend to expand their capacity of higher profit cardiac services to more marginal populations. These authors132.3. Data DescriptionIn addition to increasing the volume of higher profit procedures, hospitals may cut back onunprofitable services. As Sloan (1998) suggests, the opportunity for cross-subsidizing unprof-itable care becomes more difficult as hospital markets become more competitive. Consistentwith this, David et al. (2014) find that hospitals reduce trauma care in the face of increasedcompetition in cardiac services.Third, general hospitals may respond by targeting the more profitable patients to increaserevenue, such as those with more generous insurance schemes or those who are healthier. Intheir theoretical model, McGuire & Pauly (1991) propose that income shocks may lead toa divergence in treatment intensity across patient payer types, with more lucrative patientsexperiencing greater intensity. Their model focuses on supply induced demand by physicians,and it would also likely be applicable in the hospital setting. Dor & Farley (1996) note thathospitals have considerable discretion over many aspects of care provided to individual pa-tients. At the time of admission, internal discharge and utilization review panels have accessto the payment source information, which is also included on patients’ medical records. Thedifferences in payment schemes across payer types may consequently result in hospitals treat-ing less generous payers at a lower marginal cost than more generous payers who have thesame medical condition. This suggests that in the face of increased specialty hospital competi-tion, hospitals may treat private payers more intensely through increased surgeries or throughextending their length of stay. As noted, hospitals are not reimbursed for additional days ofstay by public payers. Hospitals may also cut back on unprofitable patients, particularly inregards to emergency admissions and care to the uninsured.My paper focuses on these possible responses, examining if increased specialty hospitalcompetition causes general hospitals to change their service mix in non-specialty care andwhether their response varies by payer type. As noted by Altman et al. (2006), increasedcompetition could also cause hospitals to downsize, reduce services, and cut staffing-patientratios. This may lower the quality of care all round. Because of data limitations, I cannotexamine changes in hospital resources in detail. This study focuses primarily on testingwhether hospitals alter their service mix and differentiate treatment by payer type in the faceof the loss of a profitable service line.2.3 Data DescriptionThe primary source of data for this analysis is the Texas Inpatient Public Use Data Files(PUDF), which contain patient-level information on all inpatient hospital stays in Texas from1999 to 2007 (24,806,916 inpatient visits). These data are collected by the Texas Health CareInformation Council (THCIC), a branch of the Texas Department of State Health Services(DSHS) Center for Health Statistics. Detailed medical information surrounding the visit isfocus on new invasive cardiac products in the situation when neighboring hospitals already offer these services.142.3. Data Descriptionrecorded, including the principal diagnosis (ICD-9-CM codes), the diagnosis-related groups(CMS-DRGs), and the major diagnostic category (CMS-MDC) codes. The data include thelength of stay (LOS) and the discharge status (e.g. discharged home, died, transferred toanother facility). The type of admission is also recorded, and I follow the THCIC by referringto scheduled visits as “elective” and emergency/urgent admissions as “non-elective” in thisstudy.27 The data also contain information about the primary and secondary payer (e.g.Medicare, Medicaid, uninsured) as well as hospital charges (total and by type of service).Patient demographics (e.g. gender, five year age group, race) and approximate location ofresidence (e.g. five-digit zip codes and county) are also provided. To reduce computationalburden, a 25% random sample is used for the analysis.From the 25% sample, I exclude individuals residing outside of Texas as well as those miss-ing full five-digit zip codes in order to get precise measures of hospital-patient distances.28 Ialso drop patients with limited demographic information due to confidentiality reasons stip-ulated by the THCIC.29 Additionally, I exclude visits relating to pregnancy and newbornssince this group is quite different than the rest of the population in terms of medical careneeds. I also exclude visits to other types of specialty hospitals, such as rehabilitation andpsychiatric institutes, since they are not directly applicable to test spillovers across servicelines in a hospital.30For the main analysis, I only examine non-specialty services provided in general hospitals.I refer to these as “uncontested” care because they are the services in which specialty hospitalstypically do not compete with general hospitals for patients. I refer to specialty admissions as“contested” services, and I analyze “contested” admissions to both general and specialty hos-pitals. These exclusions result in a total of 3,611,497 admissions, with 2,426,684 observationsfor uncontested care at general hospitals and 1,184,813 observations for contested services atall hospitals.Each patient in the sample is grouped into one of approximately 570 Diagnoses RelatedGroups (DRGs). The mapping between diagnoses and DRGs is not unique. Patients withthe same diagnosis may be coded into different DRGs, depending on the treatment theyreceive (e.g. whether or not they have surgery) and whether they have complications and/orcomorbidities. DRGs were introduced in 1982 as part of Medicare’s move to prospective27It should be emphasized that “elective” does not necessarily imply the procedure is discretionary.28The last two digits of the patient’s zip code are suppressed if there are fewer than thirty patients includedin the zip code, while the entire zip code is suppressed if a hospital has fewer than fifty discharges in a quarteror if the main diagnosis indicates alcohol or drug use or an HIV diagnosis. Additionally, zip codes are missingfor patients from states other than Texas.29Demographic information is suppressed for those patients obtaining care for HIV and alcohol and druguse. While age is represented by 22 age groups for the general patient population (typically five year agegroups), there are only 5 groups for patients with alcohol and drug use or an HIV diagnosis.30It is possible that there is an effect on these types of hospitals from increased specialty hospital penetration,but the focus of this study is on cross-subsidization and spillovers across departments and these hospitals offera narrow scope of services.152.3. Data Descriptionpayment and are used to determine the amount hospitals should be reimbursed based onexpected resource usage. The hospital is paid a fixed amount that varies by DRG. Each DRGis assigned a payment weight which functions as a price and is based on the average resourcesused to treat patients in that DRG, relative to the average level of resources for all Medicarepatients. The weights are intended to account for cost variations between different types ofprocedures. More costly conditions are assigned higher DRG weights. For example, coronarybypass is assigned a DRG weight of 6.74, obesity procedures a weight of 1.91, and urinarytract infections a weight of 0.45.31 The DRGs are further grouped into 25 mutually exclusiveMajor Diagnostic Categories (MDCs), which generally correspond to a single organ system.The Texas PUDF includes DRGs and MDCs for all payer types. DRGs can also be groupedinto clinical specialties, which tend to correspond to hospital departments.32Hospital characteristics used in this study come from the American Hospital Association(AHA) Annual Survey Database.33 The AHA Annual Survey collects detailed information onhospitals’ organizational structure (e.g. non-profit, public, for-profit), services provided, thenumber of beds (total and by service line), personnel (e.g. number of physicians and nurses),and financial performance. A hospital was designated as a specialty hospital if at least 45percent of its discharges were cardiac, orthopedic or surgical in nature, or at least 66 percentof the hospital’s discharges fell into two major diagnosis-related categories (MDC), with theprimary one being either cardiac or orthopedic. This definition comes from the MedicarePayment Advisory Commission (MedPAC), with further details provided in Appendix A.For my analysis, I define admissions with MDCs of 5 (Cardiac) or 8 (Orthopedic) as“contested” services.34 All other admissions are labelled as “uncontested” admissions (i.e.non-specialty care). Table 2.2 shows that the bulk (67.39%) of specialty hospital admissionsare for contested services, whereas a much smaller proportion of general hospital admissions(21.89%) are contested.35 This table also shows the distribution of hospital admissions acrossmedical specialty. As expected, most specialty hospital admissions are in cardiology (27.80%)and orthopedics (29.96%). For general hospitals, the distribution of admissions across special-ties is more evenly distributed. Although, obstetrics and neonatology form the largest sharesof admissions, cardiology and orthopedics also play substantial roles, accounting for 12.30%and 6.87% of admissions, respectively. Other medical specialties such as general surgery,pulmonary, and general medicine form considerable shares of general hospitals’ admissions.I briefly discuss how medical treatment in uncontested services varies by payer type in31To obtain the DRG weights in this analysis, I use the 2007 mapping provided by Centers for Medicare &Medicaid Services (CMS).32I used data from the Massachusetts Health Data Consortium to map DRGs into clinical specialties.33These data were generously provided by the Texas Health Care Information Council (THCIC).34I follow David et al. (2014) and the Medicare Payment Advisory Commission (2005) in how I define“contested” admissions.35These statistics were derived using all patients in the 25% sample, except for those without five-digitTexan zip codes.162.3. Data DescriptionTable 2.2: Admissions by Hospital TypeGeneral Hospital Specialty Hospital% in Contested ServicesMDC=5 (Cardiac) or 8 (Orthopedic) 21.89 67.39% by Medical SpecialtyCardiology 12.30 27.80Dentistry 0.10 0.04Dermatology 0.22 0.13Endocrine 2.59 1.13Gastroenterology 6.09 2.84General medicine 4.23 1.84General surgery 7.94 7.65Gynecology 3.01 5.25Hematology 0.97 0.37Neonatology 15.19 0.40Nephrology 2.58 1.07Neurology 3.64 1.59Neurosurgery 1.14 2.42Obstetrics 16.49 0.5Oncology 1.46 0.51Ophthalmology 0.13 0.07Orthopedics 6.87 29.96Otolaryngology 0.75 0.62Psychiatry 1.53 0.05Pulmonary 7.82 3.31Rheumatology 0.27 0.87Thoracic surgery 1.69 5.41Transplants 0.06 0Urology 1.59 1.91Vascular surgery 1.32 4.29Total 100 100Observations 5,180,523 64,498Notes: Data come from the Texas Inpatient Public Use Data Files, years 1999-2007.Contested services are defined as a hospital admission with principle diagnosis/procedure(DRG) falling into Major Diagnostic Categories (MDC) of 5 (Cardiac) or 8 (Orthopedic).Data from the Massachusetts Health Data Consortium were used to map DRGs intospecific medical specialties.172.3. Data Descriptionthe raw sample. Table 2.3 shows that Medicare patients form the largest proportion of thesample at 44.64%, followed by FFS patients (23.40%), HMO patients (8.40%), Medicaid(11.23%), and the uninsured (8.98%). Although Medicare patients have a lower proportion ofuncontested surgeries (18.7%), they have greater illness severity as seen by a higher averageDRG weight, longer lengths of stay (6.413 days on average), and a higher death rate (4.6%).These results likely reflect, in part, that Medicare patients are older than the rest of thepopulation. An important observation is that HMO and FFS patients look strikingly similaracross all dimensions of care. They have similar rates of uncontested surgeries (approximately40%), lengths of stay (roughly 4.25 days), DRG weights, and rates of death. Among all payers,they have the greatest rates of surgery and elective visits. Another important observation fromTable 2.3 is that uninsured patients have the lowest rates of elective care. This is unsurprisinggiven they must pay out-of-pocket for treatment if hospitals don’t absorb the costs.Table 2.3: Descriptive Statistics: Uncontested Medical Treatment by Insurance TypeInsurance Type Obs % of sample Surgery Length of Stay Elective DRG Weight DiedMedicare 1,081,295 44.64 0.187 6.413 0.265 1.167 0.046(0.390) (7.085) (0.441) (0.862) (0.209)Medicaid 272,008 11.23 0.173 5.065 0.253 0.969 0.014(0.378) (7.421) (0.435) (0.896) (0.119)Private: HMO 203,360 8.40 0.401 4.263 0.391 1.101 0.017(0.490) (5.853) (0.488) (0.845) (0.129)Private: FFS 566,737 23.40 0.395 4.243 0.371 1.102 0.019(0.489) (14.438) (0.483) (0.895) (0.135)Uninsured 217,410 8.98 0.289 4.765 0.150 1.108 0.023(0.453) (8.055) (0.357) (0.968) (0.150)Other 81,418 3.36 0.332 4.989 0.293 1.147 0.022(0.471) (6.843) (0.455) (1.048) (0.148)Total 2,422,228 100 0.266 5.378 0.290 1.117 0.031(0.442) (9.411) (0.454) (0.892) (0.173)Notes: Data come from the Texas Inpatient Public Use Data Files, years 1999-2007. This table shows descriptivestatistics of the main outcome variables, by payer type. Means are shown with standard deviations in parentheses.Surgery, elective, and died are proportions of the relevant payer type. Length of stay is measured in days. DRG weightis as described in the text. The sample consists only of patients with uncontested admissions.As discussed, one possible response to increased specialty hospital competition is for gen-eral hospitals to increase the remaining profitable discretionary procedures. Table 2.4 sheds182.4. Empirical Approachlight on the type of care provided in the general surgery department, showing the top 15types of surgeries performed. As outlined earlier, this is a department that is both relativelyprofitable and that has more clinical grey areas relative to other surgical specialties. Cellulitis(a type of skin infection), laparoscopic cholecystectomy (gallbladder surgery), and bowel pro-cedures are among the most common types of procedures in the general surgery department.It should be noted that these procedures are used to treat illnesses that often have multipletreatment options, including non-surgical care. In the next section, I present the empiricalmethodology taken to analyze hospital spillovers across service lines, testing if general hospi-tals changed their service mix in uncontested care in response to increased specialty hospitalcompetition and testing whether the effects differ by payer type.Table 2.4: Top 15 Procedures in General SurgeryTop DRG DRG Title % of General Surgery1 277 Cellulitis Age >17 W Cc 7.42 494 Laparoscopic Cholecystectomy W/O Cc 7.23 148 Major Small & Large Bowel Procedures W Cc 6.94 493 Laparoscopic Cholecystectomy W Cc 6.15 278 Cellulitis Age >17 W/O Cc 6.16 167 Appendectomy W/O Complicated Principal Diag W/O Cc 5.87 288 O.R. Procedures For Obesity 3.18 415 O.R. Procedure For Infectious & Parasitic Diseases 2.99 149 Major Small & Large Bowel Procedures W/O Cc 2.310 279 Cellulitis Age 0-17 1.811 263 Skin Graft For Skin Ulcer Or Cellulitis W Cc 1.812 290 Thyroid Procedures 1.713 154 Stomach, Esophageal & Duodenal Procedures Age >17 W Cc 1.614 150 Peritoneal Adhesiolysis W Cc 1.515 165 Appendectomy W Complicated Principal Diagnosis W/O Cc 1.5Observations 408,333Notes: Data come from the Texas Inpatient Public Use Data Files, years 1999-2007. This table shows descriptivestatistics of the top 15 types of procedures done in the General Surgery department. The last column shows theproportion that a given procedure makes up out of total admissions in general surgery. Note that “W Cc” signifieswith comorbidities and complications while “W/O Cc” signifies without comorbidities and complications. Thesample consists only of patients with general surgeries.2.4 Empirical ApproachNext, I outline the main relationship of interest in this study and the empirical approachtaken. I discuss the potential challenges to obtaining unbiased estimates since the penetrationof specialty hospitals in a market may be endogenous. In particular, the location and timingof specialty hospital entry may not be random. My estimation strategy takes into accountdifferences across locations with and without specialty hospitals over time. Furthermore,conditional on entry, the share of patients obtaining specialty care at a specialty hospital192.4. Empirical Approachis unlikely to be exogenous. To address this concern, I build on the two-step approach ofKessler & McClellan (2000) by first modelling patient demand for specialty services to obtainthe predicted market share of specialty hospitals. Specifically, I employ a patient-level hospitalchoice multinomial model, instrumenting for patient demand. Then, I estimate the impactof predicted specialty hospital market shares on uncontested medical treatment at generalhospitals. Details of the empirical approach are provided below.2.4.1 OverviewThe primary relationship of interest is the extent to which hospital profits in contested servicesaffect uncontested medical treatment at general hospitals:Yipkjt = φppijt + ωipkjt (2.1)where Yipjkt is the medical treatment of individual i who is payer type p, residing in marketk and seeking uncontested care at general hospital j at time t. Medical treatment includesthe type of procedure, the length of stay, and mortality. Payer type categories are: Medicare,Medicaid, HMO, FFS, and uninsured. The market area used for analysis is defined as theHospital Service Area (HSA), with 208 HSAs in Texas.36 It should be noted that hospital j isnot constrained to be in market k, as patients may visit hospitals outside their own market.pijt are profits of hospital j at time t in contested services. The residual is given by ωipkjt.The extent of hospital spillovers across service lines is captured by φp, the coefficient onhospitals’ profits in contested services. As noted by the subscript, the impact is allowed tovary across payer types. However, hospital profits are not directly observed in the data. Ad-ditionally, there may be unobserved factors correlated with both pijt and Yipkjt, and ordinaryleast squares estimation of Equation (2.1) may lead to biased estimates. As such, I use themarket share of specialty hospitals as a shock to general hospitals’ most profitable services totest whether general hospitals adjust the medical treatment in uncontested care. Specifically,it is believed that pijt = f(SMKSkt) ∀ k, where SMKSkt is the specialty hospital marketshare. In this study, the specialty hospital market share is defined as the proportion of pa-tients residing in market k at year t obtaining their contested care at a specialty hospital(as opposed to a general hospital). This relationship implies that general hospitals’ profits incontested services are a function of specialty hospital market shares.In its most basic form, the relationship of interest is:Yipkjt = γpSMKSkt + uipkjt (2.2)36HSAs are local health care markets for hospital care. An HSA is a collection of zip codes whose residentsreceive most of their hospitalizations from the hospitals in that area. It is produced by the Dartmouth Atlasof Health Care. A map of HSAs in Texas is provided in Appendix A.202.4. Empirical Approachwhere Yipjkt and SMKSkt are defined before, and the error term is given by uipkjt. Throughoutthe empirical analysis, the parameters of interest are the coefficients on the specialty hospitalmarket share, γp. The coefficients can be interpreted as the marginal impact of increasedspecialty hospital competition on uncontested care at general hospitals.2.4.2 Identifying the Marginal Impact of Increased Specialty HospitalCompetition on Uncontested CareOne challenge of directly estimating γp in Equation (2.2) is that there may be unobserved fac-tors in the error term uipkjt that are correlated with both SMKSkt and uncontested medicaltreatment Yipkjt, making the specialty hospital market share endogenous. The entry of spe-cialty hospitals into a market and how much of the market they capture once they have enteredmay not be random. In particular, there may be unobserved hospital market characteristics(both fixed and time-varying), unobserved patient characteristics (health and preferences),and unobserved general hospital characteristics that impact both specialty hospital marketshares and uncontested outcomes.Specialty Hospital Entry: Exploiting within Market VariationIn terms of the location of specialty hospital entry, specialty hospitals likely only consider thepotential demand and revenue in the market for contested services, not the uncontested. Itis nonetheless possible that the demand for contested and uncontested services in a marketis correlated. For example, if specialty hospitals locate in areas where patients are generallyhealthier and health is correlated across the dimensions of contested and uncontested illnesses,then this would lead to biased estimates. This would also be the case if specialty hospitalslocate in higher income areas and income affects demand for both types of care. In addition, ifspecialty hospitals locate in markets where the overarching administration at general hospitalsis poor, this could also be problematic.I include a range of patient demographic, patient zip code, and hospital controls to accountfor factors that may affect both specialty hospital entry and uncontested medical outcomes. Inparticular, patient demographic characteristics included are: age, gender, race, and ethnicity.I also include a dummy variable for the payer type of the patient. The annual patient zip codecontrols are: income per capita and the proportion of residents who are: over 65 years, White,Black, Hispanic, urban, live below the federal poverty, and are native born.37 Furthermore, inall my analyses, I include hospital characteristics (total beds, for profit, and teaching hospital)to capture any factors that might be correlated with both the specialty hospital market share37The zip code data come from the U.S. Census, years 2000 and 2010. Zip code data are not released everyyear, so I interpolate between years to obtain annual measures.212.4. Empirical Approachand uncontested medical treatment at general hospitals.38To address the concern that specialty hospital entry into a given market may be driven byunobserved differences across markets, I include market fixed effects and market specific lineartime trends. Additionally, I include year fixed effects to capture shocks that are common to allpatients in a given year. Adding these controls accounts for any fixed unobserved differencesacross markets as well as any (linear) changes in unobserved differences. This is a similardesign to Li & Dor (2013) who use this approach to estimate the impact of the repeal ofCertificate of Need regulations on coronary procedures, while Finkelstein (2007) uses a similarmodel to analyze the introduction of Medicare.Once I add this rich set of controls, the main equation of interest becomes:Yipkjt = γpSMKSkt+αpI(p)+αkI(k)+αtI(Y eart)+θk[I(k) · t]+Xitβ+Zjtη+ ipkjt (2.3)where I(p), I(k), I(Y eart) are payer, market, and year fixed effects respectively and θkis the linear time trend of market k; Xit are observed characteristics of individual i at timet (including both patient demographic and patient zip code characteristics); and Zjt arecharacteristics of hospital j at time t.Equation (2.3) amounts to a differences-in-trend design, where the impact of specialtyhospital penetration is identified off deviations from trends within a market region. So longas specialty hospital entry is not due to any unobserved deviations from the trend which arecorrelated with uncontested medical treatment and are not being captured in the extensiveset of patient demographic, zip code, and hospital controls, then this approach addresses thepossible endogenous entry decision.39The Specialty Hospital Market Share: Modelling UnobservedHeterogeneityEven once the possible endogenous entry decision has been taken into account, there may stillbe unobserved heterogeneity that affects the volume of patients admitted to specialty hospitalsfor contested care (and consequently the specialty hospital market share). For example,unobserved changes in individual preferences and health (off the deviations from trend) mayaffect where patients choose to obtain contested care. If these changes are also correlated withuncontested medical treatment, then estimates will be biased since how much of the marketspecialty hospitals capture would not be random. Formally, if ipkjt = vipkjt + ωipkjt, where38In the event that general hospitals anticipate the specialty hospital entry, then it seems there would be adownward bias of the estimates since then general hospitals will begin to prior to entry.39It should be noted that if specialty hospitals locate next to general hospitals with a poor administration,and this is not being captured by the controls, then it seems likely there would be a downward bias in termsof the sophistication of hospital response. Similarly if specialty hospitals locate in areas where patients arehealthier in unobservable ways that deviate from the market specific trend, this would lead to a downward biasin the intensity which general hospitals treat these healthier patients.222.4. Empirical Approachipkjt is from Equation (2.3), vipkjt are unobserved preferences of individual i who is payertype p residing in market k at time t, and ωipkjt is the true error term, then the concern isthat cov(vipkjt, SMKSkt) 6= 0, leading to biased estimates.To address this possibility, I extend the two step estimator developed by Kessler & McClel-lan (2000). I first construct predicted specialty hospital market shares using a multinomialchoice model for contested services. The probability that a patient attends a given hospitalfor contested care is a function of observed hospital and patient characteristics, as well as thedistance between the patient’s residence and the hospital location. In the next step, I estimatethe marginal impact of increased specialty competition on uncontested medical treatment us-ing the estimated specialty hospital market shares derived in the first step. That is, ratherthan using the actual specialty hospital market share for the analysis, I use the predictedspecialty hospital market share.The approach is in the same spirit as previous studies that use distances to hospitals ina patient’s geographic region as instrumental variables.40 Essentially, the distances betweenhospitals and patients seeking care for contested services are being used to forecast the pre-dicted specialty hospital market share for uncontested patients in an area.41 Previous studieshave found distance to be a primary determinant of hospital choice (Burns & Wholey (1992);Luft et al. (1990)).The identifying assumption is that unobserved deviations from markettrends affecting uncontested medical treatment are uncorrelated with the distance betweenhospitals and patients seeking care for contested services. The exclusion restriction is arguablyless demanding in this study than previous work since I focus on the medical treatment ofuncontested patients, a different set of individuals than those used to obtain predicted marketshares.42I argue that conditional on the rich set of patient, zip code, hospital, and market controls,the distance between patients obtaining specialty care and hospitals that provide specialtyservices are orthogonal to uncontested care. That is, I argue conditional independence betweenuncontested medical treatment and these distances. Since hospital markets are relatively small(see map of HSAs in Appendix A), it seems likely that specialty hospitals choose their locationwith the intention of serving the demand of patients from all across that market. The primarythreat to conditional independence is the specialty hospital location choice within a market.As discussed, specialty hospitals likely choose their location based on the expected demand40See for example Kessler & McClellan (2000); Chernew et al. (2002); Li & Dor (2013); and Swanson(2012).41The multinomial choice model relies on the assumption that the choice decision between any two hospitalsis independent of irrelevant alternatives (IIA). However, unlike previous studies, my empirical approach doesnot solely achieve identification from the IIA assumption (i.e. by renormalizing probabilities once specialtyhospitals enter a market). Rather, patient to hospital distances for contested services are excluded regressorsin my analysis.42Formally, the exclusion restriction is that cov(D, ωipkjt)=0, where D is the distance between hospitalsand patients seeking contested care.232.4. Empirical Approachfor specialty services, not uncontested services, which helps alleviate this concern. They maynonetheless target areas with higher incomes or where people are healthier within a market.I control for patient and zip code characteristics which are the factors most likely to affectwhere specialty hospitals locate (e.g. income per capita, the % of people over 65 years).Furthermore, the market specific linear trend is included to capture any unobserved timevarying factors that may be correlated with specialty hospital penetration and uncontestedmedical treatment.The remaining concern is that specialty hospital location choice is correlated with un-observed deviations from the market trend that also impact uncontested medical treatment.I discuss two important features of hospitals that allay this concern. First, there is a lagbetween the decision for a specialty hospital to enter a market and when the hospital is op-erational. Physician owners must not only choose a site for the hospital, but even before itsconstruction, they must go through the process of obtaining a licence, undergoing an archi-tectural review, and securing construction approval, all which take considerable time. Theconstruction of the hospital itself takes many months, and upon its completion, it must passa comprehensive inspection to obtain final approval by the state health department and toobtain Medicare certification. The entire process can take years.43 Given the lag, it seemsunlikely that unobserved deviations from the trend would affect both the location decision andcontemporaneous uncontested medical treatment. It would have to be persistent deviationsfrom trend that would create endogeneity, though they would have to be small enough tonot affect the trend itself. One scenario where this might occur would be if physician ownerschoose to locate where they project future construction of residential units targeted at peoplewith high demand for not only their services but also uncontested care, such as a retirementcommunity for example. While this behavior cannot be ruled out, it would require significantand detailed foresight by the physician owners on future long term residential developmentsacross the market, beyond the information captured by the rich set of controls. Such com-prehensive insights on real estate market projections by physician owners seems doubtful. Inaddition, I find no evidence that specialty hospitals differentially locate in areas with a higherconcentration of elderly people, making this possibility unlikely.44 The second important fea-ture of specialty hospitals in Texas is that they often locate where there is plenty of space tobuild and expand. Rather then always locating in high density wealthy residential neighbour-hoods, they are often built on the periphery of cities or on highway corridors. Together, thesetwo insights on specialty hospitals support conditional independence of uncontested medical43In fact, it is due to this lag that the largest growth of specialty hospitals in Texas during the sampleperiod actually occurred near the end of the federal moratorium on new specialty hospitals (existing ones andthose that had already obtained licences could be grandfathered in) which was mandated between November2003 and August 2006.44On average, zip codes with a specialty hospital have 10.1% of individuals being 65 years or older, whilethose that do not have a specialty hospital, on average, have 13.6% of people being over 65 years.242.4. Empirical Approachtreatment and patient-hospitals distances.The estimates derived in this paper are also robust to endogenous hospital choice becauseI assign specialty hospital shares to where a patient lives, not to the hospital to which she isadmitted for uncontested care. This is important because hospital choice may be endogenousif changes in specialty hospital market shares cause patients to alter where they seek care foruncontested services, or if market shares are correlated with unobserved hospital quality orhospital characteristics.45 The details of the estimation strategy are described below.2.4.3 Estimating the Specialty Hospital Market ShareI first estimate the market share of specialty hospitals in contested services. As discussed, themarket area used for analysis is the HSA, and SMKSkt is the proportion of patients residingin HSA k at time t who obtain contested care (i.e. cardiac or orthopedic care) from specialtyhospitals. I specify a patient-level hospital choice model for patients seeking contested care.I model the hospital choice decision for contested care as a function of hospital and patientcharacteristics which are arguably orthogonal to uncontested patient outcomes.In particular, individual i’s indirect utility from choosing hospital j is given by:Uij = V (Dij ;Zj) +W (Xi;Zj) + ξij (2.4)where Dij is non-parametric function of the distance from individual i to hospital j; Zjare characteristics of hospital j; Xi are characteristics of individual i. The choice set for eachindividual is comprised of all hospitals within a 50 mile radius of her residence, or 100 mileradius for teaching hospitals, with patient location being approximated by the centroid of herzip code.46 Euclidean distances between patients’ residences and hospitals were calculatedusing GIS.47 Further details are provided in Appendix A.45For example, this would occur if patients observe a decline in contested services at general hospitals andbelieve that this provides information about its quality so obtain care elsewhere. Similarly, if patients seekinghigh quality, cutting edge care are likely to travel further to urban areas which have more specialty hospitals,then this would lead to biased estimates.46Nearly 95% of patients chose a hospital within 50 miles. Within the 50 mile radius, the median patienthad 23 hospitals to choose from and chose a hospital that was 7.80 miles from the centroid of her zip code.47Other measures of distance could also have been used, such as road distance or travel time. Euclideandistances will be correlated with these other measures. However, the computation of these alternatives issignificantly more demanding and requires the formulation of additional assumptions (e.g. travel routes, drivingspeeds, etc.). Using the Euclidean distance rests on the assumption that two points that are closer togetherinvolves lower travel costs (e.g. time). A potential downside is that Euclidean distance ignores the topographyof the land. Two points very close together may not be easily accessible if, say, mountains, large rivers, orcanyons separate them. In the state of Texas, these topographic factors are less of a concern. In areas wherethe majority of the state’s population is concentrated, around the Gulf of Mexico and just west, the terrainis flat, with gentle hills, and the rivers are narrow. With numerous bridges and paved roads, such naturalfeatures go unnoticed. The Great Plains extend into the north of the state and into the Texas Panhandle and,consequently, these parts are also relatively flat, with few natural barriers. The west of Texas is primarilydessert and does have some rugged, mountainous terrain and wide rivers. However, this region is largely not252.4. Empirical ApproachFor every i− j pair, V(.) is a nonparametric function of distance and hospital character-istics h = 1, ...,H.Vij =H∑h=1αhDijZhjSpecifically, Dij is a vector of four dummy variables indicating the quartile of distancewhich i− j pair falls into from the distribution of distances of all pairs. Zhj contains infor-mation on hospital characteristic h, such as indicators for whether hospital j is for profit, ateaching hospital, a specialty hospital, and the tercile of total hospital beds.From equation 2.4, W(.) is a nonparametric function of the interaction between individuali and hospital j’s characteristics:Wij =H∑h=1XiZhj γhwhere Zhj are as defined above, and the vector Xi includes age categories (grouped by fiveyears), gender, race (white, black, other), ethnicity and illness severity (minor, moderate,severe, or extreme). Note that individual characteristics Xi are fully interacted with thebinary hospital characteristics Zhj .I estimate the patient-level multinomial logit hospital choice model in Equation (2.4) usingmaximum likelihood, deriving estimates of parameters γh and αh for h = 1, ...,H. McFadden(1973) shows that the probability of individual i choosing hospital j, is given by:pij = Pr(Yij = 1) =exp(Vij +Wij)∑j∈Jiexp(Vij +Wij)(2.5)where Yij =1 if individual i is treated at hospital j and =0 otherwise, and Ji is the set ofhospitals within a 50 mile radius from patient i. To allow for differences in preferences overtime and across medical conditions, I estimate hospital choice separately for different years, fordifferent specialties (cardiac or orthopedic), and for those who do and do not obtain surgicalprocedures.48Following McFadden (1973), the expected market demand for hospital j in region k isgiven by:inhabited, making this less of a concern since few patients in my dataset reside here. Furthermore, since myanalyses focus on within-market variation, fixed topographic differences across hospital markets that may becorrelated with patient health outcomes will be controlled for. Swanson (2012) examines patient sorting inphysician owned specialty hospitals and shows that her results are robust to using both Euclidean and moresophisticated measures of distance.48In total, the model is estimated separately four times for each year.262.4. Empirical Approachdˆjk =∑i∈kpˆijAs such, the estimated market share of specialty hospitals in region k is:ˆSMKSk =∑j∈SPCkdˆjk∑j∈Jkdˆjk(2.6)where SPCk is the set of all specialty hospitals within 50 mile radius from each patientresiding in market k; Jk is the set of all hospitals (specialty and general) within 50 mile radiusfrom each patient in market k. The actual market share of specialty hospitals in region k is:SMKSk = ˆSMKSk + uˆk, where uˆk is the estimated residual of the specialty hospital marketshare in k.The distribution of predicted versus actual specialty hospital market share is shown inFigure 2.2. As can be seen, the distribution of specialty hospital market shares is highlyskewed to the right, with the average share being heavily driven by those markets with veryhigh specialty hospital penetration. The model somewhat overpredicts specialty hospitalmarket share at the lower end of the distribution. However, overall, it does a very good jobof predicting the specialty hospital market share throughout the distribution. In the sampleperiod, the average specialty hospital market share has increased by just under 3 percentagepoints, from 0.90% of the market to 3.75% between 1999 and 2007 (see Appendix A fordetails). Among markets with a positive (i.e. non-zero) specialty hospital market share atthe end of the sample period, the average specialty hospital market share has increased bynearly 5 percentage points, from 1.39% of the market in 1999 to 6.26% in 2007.2.4.4 Main Estimating EquationAfter having obtained the predicted market share and the estimated residual, I can then de-termine how general hospitals respond to changes in specialty hospital competition. I usethe method of two-stage residual inclusion, which involves using the actual specialty hospitalmarket share and the predicted residual rather than the predicted hospital market share. Asargued by Terza et al. (2008) and Wooldridge (2014), it is a simple approach that can beapplied to obtain consistent estimates for both linear and non-linear relationships.49 Unob-49This method was used since many dependent variables in the analysis are binary. Two-stage residualinclusion (2SRI) is a case of the conventional generic two-stage optimization estimator and was first proposed byHausman (1978). Terza et al. (2008) discuss why estimating nonlinear relationships with two-stage predictorsubstitution will not result in consistent estimates and advocate for 2SRI. As noted by Wooldridge (2014),2SRI involves parameterizing the underlying endogeneity and provides consistent estimates so long as thepredicted residuals act as a sufficient statistic for capturing this endogeneity. Although 2SRI allows for nonlinear272.4. Empirical ApproachFigure 2.2: Distribution of Actual and Predicted Specialty Hospital Market SharesNotes: This figure shows the actual and the predicted specialtyhospital market shares. The specialty hospital market share is theproportion of patients residing in a given HSA in a year who obtaincontested care (i.e. cardiac or orthopedic care) from a specialtyhospital (as opposed to a general hospital). The predicted specialtyhospital share is derived from a patient-level multinomial model. Datato construct the shares come from the Texas Inpatient Public Use DataFiles, years 1999-2007.282.5. Empirical Resultsserved factors affecting specialty hospital market shares are controlled for in the estimatedmarket share residual.The main estimating equation for the analysis is as follows:Yipkjt = γpSMKSkt + αpI(p) + αkI(HSAk) + αtI(Y eart) + θk[I(HSAk) · t]+Xiptβ +Zjtη + σuˆkt + ωipkjt (2.7)where uˆkt is the estimated residual of the specialty hospital market share in k at time t.The rest of the notation is as defined previously.50 In some specifications, hospital departmentfixed effects are included in the analysis as are controls for the comorbidity of patients andtheir DRG. Again, the parameters of interest are the γp’s, the marginal impact of increasedspecialty hospital competition for payer type p and whose estimates are shown in subsequenttables. Since specialty hospital shares are assigned to the market where the patient lives, asopposed to the hospital visited, the estimates are also robust to endogenous hospital choice.Standard errors are bootstrapped to account for the two-step estimation resulting in a gener-ated regressor. In particular, I use the block bootstrap where zip codes are the blocks.512.5 Empirical ResultsIn this section, I discuss and present the estimates derived from Equation (2.7) using themicro patient-level data. I examine whether changes in the specialty hospital market shareaffect the types of procedures, the length of stay, and the mortality rate of patients obtaininguncontested care. I also examine whether there is heterogeneity in the effects across payertypes to test whether hospitals differentiate treatment by payer type. Prior to carrying outthe patient-level analysis, I provide findings on changes in the extensive margin of hospitaladmissions, testing if specialty hospitals steal patients and analyzing if general hospitals cutback or ramp up other types of care in response. This analysis motivates and can shed lighton the results from the main analysis.2.5.1 Volume of AdmissionsI first analyze the extensive margin of hospital admissions to get a preliminary sense of generalhospitals’ response to specialty hospital competition. I aggregate the number of patients ina hospital who have particular types of admissions in a given year and then test the impactestimation (i.e. probit/logit), in practice, estimation in the main analysis was done with OLS to greatly reducecomputation time.50It should be noted that fully interacting patient and hospital characteristics as in the estimation ofthe predicted specialty hospital market shares leaves the estimated parameters from Equation 2.7 virtuallyunchanged.51The results presented in the remaining section are also robust to clustering at the HSA level.292.5. Empirical Resultsof specialty hospital market shares on the log of admissions.52 I assign the specialty hospitalmarket share to hospitals based on the HSA which they are located and then test if there isan impact on hospital admissions.53 I include the predicted residual, and I also control forhospital characteristics and time varying HSA demographic characteristics.54 In some cases,there are very few hospitals in an HSA which makes identifying HSA fixed effects and HSAtime trends very demanding for estimation at the hospital-level. Instead, I add fixed effects forthe Hospital Referral Region (HRR) in which the hospital is located as well as HRR specificlinear time trends.55The results of this analysis are shown in Tables 2.5 and 2.6. I first analyze the impact ofgreater specialty hospital penetration on contested hospital admissions at general hospitals.The first column of Table 2.5 shows evidence of a sizeable decline in admissions, with a 1percentage point change in specialty hospital market shares leading to a 1.063% decline incontested admissions at general hospitals. While the results are somewhat imprecise, theyare sizeable and are statistically significant at the 10% threshold. This finding suggests thatspecialty hospitals steal patients from general hospitals, leading to a shift in the volume ofcontested services away from general hospitals and towards specialty hospitals.5652I use the log of admissions as the dependent variable since the distribution of hospital admissions is heavilyskewed to the right due to the presence of very large hospitals.53Specifically, I use the predicted specialty hospital market shares derived for patients in a given HSA andassign these shares to all hospitals located in the same HSA.54In particular, I control for income per capita as well as the proportion of residents who are: over 65 years,White, Black, Hispanic, urban, live below the federal poverty, and are native born.55Each HRR is formed from various HSAs. HRRs are regional health care markets for tertiary medical carethat needs a major referral center. Dartmouth Atlas define the boundaries of HRRs by determining wherepatients were referred for major cardiovascular surgical procedures and for neurosurgery. There are 24 HRRsin Texas.56See Mankiw & Whinston (1986) and Li & Dor (2013) on the possible effects of specialty hospitals onincumbents’ contested service line.302.5. Empirical ResultsTable 2.5: Total Contested and Uncontested Hospital AdmissionsLog Contested Log Uncontested Log Uncontested Log UncontestedAdmissions Admissions Elective Non-ElectiveSMKS -1.063* 0.137 2.670** -1.058*(0.567) (0.346) (1.204) (0.529)Observations 2,353 2,353 2,241 2,340Notes: This table shows the change in contested and uncontested admissions at general hospitalsdue to specialty hospital market share (SMKS). The coefficient on SMKS is estimated with a linearprobability model using the method of two-stage residual inclusion. Hospitals are weighted by thetotal number of beds in their first year. Hospital controls include indicators for the tercile of bedsin first year; for profit; and teaching hospital. Annual HSA controls are also included (per capitaincome as well as the proportion of the population: 65+, White, Black, Hispanic, urban, high schoolgraduate, native born, below federal poverty line). Year fixed effects, HRR fixed effects, and HRRtime trends are included. Standard errors are clustered by HRR. * p<0.10, ** p<0.05, *** p<0.01.312.5.EmpiricalResultsTable 2.6: Total Hospital Admissions by Surgery TypeContested + Uncontested Uncontested General SurgeryLog Surgical Log General Log Other Log Non-Surgical Log Non-Elective Log ElectiveAdmissions Surgeries Surgeries AdmissionsSMKS 0.0888 0.973** 0.0005 0.162 -0.266 2.885**(0.675) (0.470) (0.660) (0.324) (0.563) (1.110)Observations 2,247 2,225 2,121 2,353 2,192 2,021Notes: This table shows the change in surgical admissions at general hospitals due to specialty hospital market share (SMKS). Thecoefficient on SMKS is estimated with a linear probability model using the method of two-stage residual inclusion. Hospitals are weightedby the total number of beds in their first year. Hospital controls include indicators for the tercile of beds in first year; for profit; andteaching hospital. Annual HSA controls are also included (per capita income as well as the proportion of the population: 65+, White,Black, Hispanic, urban, high school graduate, native born, below federal poverty line). Year fixed effects, HRR fixed effects, and HRRtime trends are included. Standard errors are clustered by HRR. * p<0.10, ** p<0.05, *** p<0.01.322.5. Empirical ResultsNext, I test if specialty hospitals caused general hospitals to shrink uncontested servicesacross the board due to the negative budget shock (column 2 of Table 2.5). I find no evidencethat the total volume of uncontested admissions was affected by increased specialty hospitalconcentration. There is, however, heterogeneity across the types of uncontested admissions.Column 3 shows a large increase in uncontested elective (i.e. scheduled) admissions, with a1 percentage point increase in specialty hospital market share causing a 2.670% increase inadmissions. This offsets a decline in non-elective (i.e. urgent) admissions, which is in theorder of 1.058%, as shown in Column 4. This is consistent with the findings of David et al.(2014), who also find a decline in trauma care due to specialty hospital entry. These resultssuggest that hospitals are admitting fewer patients from the emergency department and arereplacing these visits with more elective care.57As discussed, one way in which hospitals can try to make up lost revenue from the decline incontested admissions is to increase admissions in the remaining profitable procedures, whichare primarily surgeries. First, I test if there is a change in the total volume of surgeriesperformed at general hospitals. Column 1 of Table 2.6 shows no significant change in totalsurgeries. Next, I examine heterogeneity in the types of uncontested surgeries affected. I finda significant increase in the number of general surgery admissions, with a 1 percentage pointincrease in specialty hospital share causing a 0.973% increase in general surgery admissions(column 2). However, I find no changes in other types of surgeries (column 3). These resultsare aligned with general surgeries having more clinical grey areas than other surgeries. Tounderstand the nature of the increase in general surgical admissions, I analyze which types ofsurgical procedures are driving these results. I find some evidence of a decline in non-electivegeneral surgeries. However, this result is not statistically significant. I do, however, finda sizeable increase in elective general surgical procedures, in the order of 2.885% with a 1percentage point increase in specialty hospital shares (column 6).These findings provide evidence that specialty hospitals steal patients from general hos-pitals and the increased competition leads general hospitals to shrink back on unprofitablecare and ramp up discretionary, profitable procedures. It should be noted that although Ifind no evidence of a change in the total volume of uncontested visits, it is possible thatgreater specialty hospital penetration may cause some patients to forgo uncontested care atgeneral hospitals altogether. While I only observe individuals with an inpatient visit in thedata set and cannot test this directly, it seems most plausible that if this occurred, individualswould forgo elective discretionary care, in which case my findings would be a lower bound of57It is possible that some of this is a temporal shift in care (i.e. those who would have been admitted fromthe emergency department are being admitted later and are coded as an elective visit). My data do not includereadmissions, so I cannot test for this directly. However, subsequent findings suggest that this is unlikely thewhole story and that hospitals are strategically cutting back on unprofitable care and replacing it with higherprofit, discretionary procedures.332.5. Empirical Resultsthe true estimate.58 As discussed, the hospital-level analysis does not take into account thatthe location where patients obtain uncontested care may be endogenous to the penetrationof specialty hospitals. To address the potential endogeneity of hospital choice and to test ifthere is heterogeneity across patient payer types, I turn to the patient-level analysis usingindividual micro data.2.5.2 Intensity of Treatment and Differences by Payer TypeI now present the findings from the main analysis, which come from estimation of Equation(2.7) using individual patient-level data on uncontested care. This section analyzes howgeneral hospitals respond to the loss of high profit services. As discussed, specialty hospitalmarket shares are assigned to the location of the patient in this analysis. In addition, a richset of individual demographic characteristics are included as controls which is particularlyimportant to take into account when examining differences across payer types.Types of Procedures: SurgeriesTable 2.7 shows the impact of increased specialty hospital penetration on the share of patientswith an uncontested surgery. I first analyze surgeries as a whole (columns 1 and 2), and addin department fixed effects to test within individual hospital departments (columns 3 and 4).The first column shows that there is a slight increase in the share of patients with a surgicalprocedure. However, this effect is not statistically significant. I add in specialty hospitalmarket shares and payer interactions to test for differences across payer types (column 2). Ifind no effect for Medicare patients (the base category), Medicaid, and uninsured patients.However, I find that a 1 percentage point increase in specialty hospital market share leads toa 0.219 percentage point increase in HMO patients receiving a surgery and a 0.139 percentagepoint increase for FFS patients.These results show an overall increase in the share of private payers with an uncontestedsurgical procedure. I next test whether, on average, there is a differential increase in surgerieswithin an individual hospital department. This is to better understand if the overall resultsare being driven by a subset of departments or if hospitals are increasing the intensity oftreatment within individual departments. I find no overall effect in the share of patients witha surgery in a department (column 3). However, I find there is heterogeneity across differentpayers. In particular, the intensity of treatment increases for private payers, with a greaterproportion having a surgery. Specifically, a 1 percentage point increase in specialty hospitalmarket share causes a 0.090 percentage point increase in the share of HMO patients with asurgery in a department and a 0.081 percentage point increase in the share of FFS patients.58It also seems unlikely patients would forgo emergency department visits requiring hospital admission,which are the relatively more severe types of emergency visit.342.5. Empirical ResultsTable 2.7: Impact of Increased Specialty Competition on Share ofSurgical PatientsOverall Within DepartmentSurgery Surgery Surgery SurgerySMKS 0.0027 -0.0411 0.0065 -0.0174(.0249) (0.0285) (0.0133) (0.0138)SMKS x Medicaid 0.0147 0.0161(0.0377) (0.0122)SMKS x Private: HMO 0.219*** 0.0896***(0.0440) (0.0192)SMKS x Private: FFS 0.139*** 0.0808***(0.0265) (0.0125)SMKS x Uninsured 0.0013 0.0056(0.0342) (0.0192)Observations 2,295,064 2,295,064 2,275,489 2,275,489Payer Dummies Yes Yes Yes YesPayer Interactions No Yes No YesDept FE No No Yes YesMean 0.266St. Dev 0.442Notes: This table shows the change in the proportion of patients with a surgicaladmission in a HSA due to specialty hospital market share (SMKS). The coeffi-cient on SMKS is estimated with a linear probability model using the method oftwo-stage residual inclusion. The base category for payer type is Medicare. Pa-tient demographic characteristics (gender dummy, five year age group dummies,race dummies, Hispanic dummy) and hospital characteristics (total beds indica-tors, for profit, teaching dummy) are included. Annual zip code characteristicsare included (per capita income as well as the proportion of the population: 65+,White, Black, Hispanic, urban, high school graduate, native born, below federalpoverty line). Year fixed effects, HSA fixed effects, and HSA linear time trendsare included. Standard errors are block bootstrapped where the blocks are zipcodes. * p<0.10, ** p<0.05, *** p<0.01.352.5. Empirical ResultsTo understand the nature of surgeries being affected among the private payers, I analyzedifferent types of surgeries in Table 2.8. First I estimate changes in the DRG weight for thesample of individuals with a surgery (column 1). Department fixed effects are included sothese represent average changes within a department. I find evidence there is a differentialdecline in the average DRG weight for private payers (relative to Medicare patients), in theorder of approximately 0.002 of a standard deviation with a 1 percentage point increase inspecialty hospital market shares. Since patients whose condition is the most severe are likelyto still obtain complex surgeries, the relative reduction in the average DRG for private patientssuggests that hospital departments are expanding surgical care to more marginal patients.59I next analyze the effects on the proportion of individuals with an elective surgery. Ifind that there is a significant increase in the share of private payers with an elective surgery(column 2), where a 1 percentage point increase in specialty hospital market share increaseselective surgeries in departments by 0.132 percentage points for HMO patients and 0.129percentage points for FFS. These findings also support the evidence that hospital departmentsare performing relatively more discretionary surgeries amongst private payers. In addition,there is evidence of a decline in the share of uninsured patients with an elective procedure, inthe order of 0.067 percentage points, suggesting that hospitals are cutting back on unprofitablecare.Columns 3-5 of Table 2.8 provide evidence on which surgical departments are drivingthese changes. Most noticeable is the large increase in the share of private payers with generalsurgeries (column 3). Relative to Medicare patients, a 1 percentage point increase in specialtyhospital market share increases the share of private payers with a surgery in the general surgerydepartment by 0.142 percentage points for HMO patients and 0.119 percentage points for FFSpatients. The evidence for other types of surgeries, such as in gynecology, neurosurgery, andurology being impacted is more limited. As discussed previously, general surgeries tend tohave more clinical grey areas than other surgeries.59As discussed, I find no evidence that the volume of uncontested admissions is affected by greater specialtyhospital penetration. Furthermore, I find no differential effect for private payers in terms of changes in theoverall average DRG. These findings both suggest that differences in the severity mix are not completely drivingthis result.362.5. Empirical ResultsTable 2.8: Impact of Increased Specialty Competition on Types of Uncontested SurgeriesDRG Elective General Gynecology Neuro- UrologyWeight Surgery Surgery Surgery surgery SurgerySMKS 0.895*** 0.0414 -0.0506*** -0.0032 -0.0030 -0.0031(0.183) (0.0274) (0.0188) (0.0144) (0.0101) (0.0095)SMKS x Medicaid -0.309 -0.0012 0.0481** -0.0381*** 0.0037 0.0002(0.247) (0.0179) (0.0244) (0.0124) (0.0072) (0.0075)SMKS x Private: HMO -1.188*** 0.132*** 0.142*** 0.0357 0.0149* -0.0076(0.171) (0.0309) (0.0269) (0.0247) (0.0090) (0.0123)SMKS x Private: FFS -1.192*** 0.129*** 0.119*** -0.0299* 0.0222*** 0.0155**(0.148) (0.0188) (0.0171) (0.0167) (0.0065) (0.0069)SMKS x Uninsured -0.945*** -0.0666*** 0.0759*** -0.0569*** 0.0076 0.0069(0.241) (0.0182) (0.0283) (0.0150) (0.0078) (0.0078)Observations 533,032 2,275,489 2,295,064 2,295,064 2,295,064 2,295,064Sample Surgical All All All All AllSample Mean 1.695 0.139 0.135 0.060 0.021 0.025St. Dev 1.518 0.346 0.342 0.237 0.142 0.155Notes: This table shows the change in the DRG weight for surgical admissions and the change in the proportionof patients with particular types of surgical admissions in a HSA due to specialty hospital market share (SMKS).The coefficient on SMKS is estimated with a linear probability model using the method of two-stage residualinclusion. The base category for payer type is Medicare. Patient demographic characteristics (gender dummy, fiveyear age group dummies, race dummies, Hispanic dummy) and hospital characteristics (total beds indicators, forprofit, teaching dummy) are included. Department fixed effects are included when the DRG weight and electivesurgery are the dependent variables. Annual zip code characteristics are included (per capita income as well asthe proportion of the population who are: 65+, White, Black, Hispanic, urban, high school graduate, nativeborn, below federal poverty line). Year fixed effects, HSA fixed effects, and HSA linear time trends are included.Standard errors are block bootstrapped where the blocks are zip codes. * p<0.10, ** p<0.05, *** p<0.01.372.5. Empirical ResultsIntensity of Treatment: Length of StayThe evidence thus far suggests that hospitals respond to the loss of admissions in their prof-itable service lines by increasing profitable procedures among the most profitable patients. Inext test for differential effects in the intensity of treatment across payer types. To measureintensity of treatment, I use the length of stay (in days). Table 2.9 shows a small increase inthe average length of stay in a hospital (column 1). Specifically, a 1 percentage point increasein specialty hospital market share increases length of stay by 0.0123 days, or 0.0013 of a stan-dard deviation. This effect is driven by private payers (column 2), where a 1 percentage pointincrease specialty hospital market share increases the length of stay by 0.0270 days (or 0.0029of a standard deviation) for HMO patients and 0.0220 days (0.0023 of a standard deviation)for FFS patients.These findings may be driven by an increase in the share of private payers with surgeries.Columns 3-6 test whether this is the case. I first add department fixed effects to see if theincrease in length of stay persists, on average, within a department. I find the overall increasein length of stay remains (column 3), and that the differential increase among private payersstill holds (column 4). Next, I condition on a patient’s DRG.60 If surgeries are completelydriving the increased length of stay among private payers, then we should see the effectdisappear when we condition on the DRG. However, columns 6 shows that this is not thecase. There is still a significant differential increase in private payers’ length of stay. Inparticular, conditional on patients’ DRG, HMO patients experience an increase of 0.0176days (0.0019 standard deviations) and FFS patients an increase of 0.0125 (0.0013 standarddeviations) from a 1 percentage point increase in specialty hospital market share. As notedpreviously, only private payers reimburse for extra hospital days, while public payers do not.While the magnitude of these effects are not large, these findings clearly suggest hospitals dodifferentiate treatment by payer type and try to make up for lost revenue.60It should be noted that DRG itself may be endogenous. However, its inclusion as a control helps shedlight on better understanding the hospital response.382.5. Empirical ResultsTable 2.9: Impact of Increased Specialty Competition on Length of Stay (LOS)Overall Within Department Within DRGLOS LOS LOS LOS LOS LOSSMKS 1.232** 0.435 1.110** 0.534 0.820** 0.383(0.483) (0.465) (0.450) (0.430) (0.408) (0.400)SMKS x Medicaid 0.519 0.246 0.0691(0.615) (0.561) (0.497)SMKS x Private: HMO 2.698*** 2.103*** 1.758***(0.468) (0.416) (0.372)SMKS x Private: FFS 2.202*** 1.629*** 1.248***(0.388) (0.333) (0.303)SMKS x Uninsured 1.383*** 1.030** 0.786*(0.452) (0.402) (0.425)Observations 2,295,062 2,295,062 2,275,487 2,275,487 2,295,062 2,295,062Payer Dummies Yes Yes Yes Yes Yes YesPayer Interactions No Yes No Yes No YesDept FE No No Yes Yes No NoDRG FE No No No No Yes YesMean 5.378St. Dev 9.411Notes: This table shows the change in the length of stay in a HSA due to specialty hospital market share(SMKS). The coefficient on SMKS is estimated with a linear probability model using the method of two-stage residual inclusion. The base category for payer type is Medicare. Patient demographic characteristics(gender dummy, five year age group dummies, race dummies, Hispanic dummy) and hospital characteristics(total beds indicators, for profit, teaching dummy) are included. Annual zip code characteristics areincluded (per capita income as well as the proportion of the population who are: 65+, White, Black,Hispanic, rural, high school graduate, native born, below federal poverty line). Year fixed effects, HSAfixed effects, and HSA linear time trends are included. Standard errors are block bootstrapped where theblocks are zip codes. * p<0.10, ** p<0.05, *** p<0.01.392.5. Empirical ResultsImpact on Mortality RateFinally, I examine if greater specialty hospital penetration has an impact on patients’ deathrate. Column 1 of Table 2.10 shows there is an increase in the average death rate of patients.Testing for heterogeneity across payers, I find the effect is primarily concentrated amonguninsured and private payers (column 2). Among uninsured patients, a 1 percentage pointincrease in specialty hospital market share increases the proportion of uninsured who die by0.03 percentage points. However, this could be due to composition effects in terms of the typesof patients being admitted. As shown previously, there is a decline in the share of uninsuredpatients with an elective visit, suggesting that this result may be driven, at least in part, byuninsured patients who are admitted having higher severity. In columns 3 and 4, I controlfor whether the patient was noted to have a comorbidity.61 The increase in mortality ratepersists even when factoring in patients’ underlying health.The increase in mortality rate may be due to hospitals performing riskier types of proce-dures, specifically more surgeries. In the last two columns of Table 2.10, I add DRG fixedeffects, controlling for the type of treatment. The increase in mortality rate remains evenwhen conditioning on the DRG. Specifically, a 1 percentage point increase in specialty hos-pital market share results in an increase in mortality in the order of 0.04 percentage pointsfor uninsured patients and 0.03 for privately insured patients. If DRGs entirely capture theseverity or underlying risk of patient health, then the findings could suggest that hospitalquality may decline in the face of increased specialty competition. That is, the increase inmortality is not entirely being driven by an increase in the intensity of types of proceduresbeing performed or in the sickness of patients admitted. If, however, there is unobservedheterogeneity in patient risk within a given DRG, then the increase in the mortality rate maybe from increased mismatch between patients and the types of procedures performed.6261I do this using the DRG code and indicate whether the patient had a DRG with comorbidities and/orcomplications. Again, it should be noted that DRGs themselves may be endogenous, but their inclusion helpsshed light on hospital behavior.62I focus on the mortality rate for patients obtaining uncontested services. It is possible that the mortalityrate of patients obtaining contested care may also be affected, particularly for cardiac patients. I do notexplicitly examine this in my study. However, existing evidence suggests that markets with specialty hospitalsexperience a decline in average mortality rate for contested care. For example, Swanson (2012) finds thatoverall cardiac patient mortality would increase if specialty hospitals were eliminated from their markets.402.5. Empirical ResultsTable 2.10: Impact of Increased Specialty Competition on Mortality RateOverall Comorbidity FE Within DRGDied Died Died Died Died DiedSMKS 0.0178* 0.0076 0.0184** 0.0081 0.0060 -0.0089(0.0093) (0.0108) (0.0093) (0.0107) (0.0082) (0.0094)SMKS x Medicaid 0.0154* 0.0157* 0.0175**(0.0092) (0.0092) (0.0081)SMKS x Private: HMO 0.0210* 0.0207* 0.0328***(0.0109) (0.0109) (0.0104)SMKS x Private: FFS 0.0177** 0.0177** 0.0310***(0.0078) (0.0078) (0.0075)SMKS x Uninsured 0.0325*** 0.0318*** 0.0392***(0.0122) (0.0122) (0.0112)Observations 2,290,179 2,290,179 2,290,179 2,290,179 2,290,179 2,290,179Payer Dummies Yes Yes Yes Yes Yes YesPayer Interactions No Yes No Yes No YesComorbidity FE No No Yes Yes No NoDRG FE No No No No Yes YesMean 0.031St. Dev 0.173Notes: This table shows the change in mortality rates in a HSA due to specialty hospital market share(SMKS). The coefficient on SMKS is estimated with a linear probability model using the method of two-stage residual inclusion. The base category for payer type is Medicare. Patient demographic characteristics(gender dummy, five year age group dummies, race dummies, Hispanic dummy) and hospital characteristics(total beds indicators, for profit, teaching dummy) are included. Annual zip code characteristics are included(per capita income as well as the proportion of the population who are: 65+, White, Black, Hispanic, rural,high school graduate, native born, below federal poverty line). Year fixed effects, HSA fixed effects, and HSAlinear time trends are included. Standard errors are block bootstrapped where the blocks are zip codes. *p<0.10, ** p<0.05, *** p<0.01.412.6. Discussion2.5.3 Robustness to Alternative SpecificationsI test whether the same pattern of results hold with alternative specifications. First, I estimateEquation (2.7) using zip code fixed effects rather than HSA fixed effects. The parameterestimates are virtually identical to when HSA fixed effects are used across all outcomes.In another specification, I estimate Equation (2.7) without the HSA specific linear trend.Qualitatively, the results are quite similar to the main specification. Although the baseestimates (i.e. Medicare) slightly change compared to the main estimates, the differentialeffects between payer types remain virtually unchanged. Finally, I test alternative patient zipcode controls using the raw 2000 and 2010 measures (rather than their annual interpolationsas in the main specification). Again, the estimates are robust to this specification and arevirtually unchanged compared to the main estimation. The estimates from these alternativespecifications are provided in Tables A.2 to A.4 of the Appendix.632.6 DiscussionMy findings suggest that hospitals respond to the loss of a profitable service line by alteringtheir service mix. Specifically, I find they expand surgical procedures, performing procedureswith lower marginal benefit. Between 1999 and 2007, the magnitude of this expansion wasan increase in the share of private payers with a surgical procedure in the order of 0.70-1.07percentage points in those markets with non-zero specialty hospital market shares.64While hospitals might have previously been constrained in the care they provide and thefreed resources now allow them to expand other services lines, it is unlikely that pent updemand is the whole story. The increase in treatment intensity is concentrated primarilyamong the most profitable procedures and privately insured patients. Furthermore, there is adecline in unprofitable care, notably non-elective care, particularly to the uninsured. Together,these findings suggest that hospitals are not simply shifting resources but are strategicallytargeting specific procedures and patients to augment revenue. This is further supported bythe increase in the length of stay only among private payers who reimburse per diem, evenwhen taking into account increased surgeries.The analysis also clearly shows that hospitals do differentiate treatment by payer type,providing care on a case by case basis to patients. These findings suggest that there aremultiple treatment options that are consistent with acceptable medical practices. At the63The effects are present for both non-profit and for-profit hospitals. Although, the intensity varies by typeof hospital. In future analyses, I will use AHA data to investigate whether the hospital response varies withtheir financial health, focusing on operating margins, equity to revenue ratios, and financial liquidity. Theseanalyses can shed light on the extent to which bankruptcy risk and the hospital budget constraint affects theprovision of care and patient health.64As mentioned, the average increase in the specialty hospital market share was 5 percentage points between1999 and 2007 in those markets with non-zero specialty hospital shares at the end of the sample period.422.7. Conclusionsame time, some types of care are more costly and pose greater medical risks. In particular,surgical procedures are major interventions. Certainly, the increase in mortality rate foundis a concern there may be a mismatch among patients and procedures. Between 1999 and2007, I find that greater specialty hospital penetration increased the mortality rate by 0.15percentage points for private payers in markets with non-zero specialty hospital market sharesand 0.2 percentage points for uninsured patients. This is a sizeable effect relative to the meanmortality rate for these patients. The reduction in non-elective care provided by generalhospitals is also disconcerting as individuals may no longer be getting necessary care. However,it is also possible that some of care in the previous service mix was wasteful, so shrinkingsome services may actually increase efficiency.Two fundamental considerations should be kept in mind when assessing the welfare impli-cations of greater specialty hospital penetration. First, private patients are not paying the fullcost of additional services since they are insured, so the social marginal benefits of additionalcare may be less than the social marginal costs (standard moral hazard). Secondly, thereis asymmetric information between physicians and patients (the principle-agent problem),meaning that patients may not fully understand the expected benefits of different types ofcare. The expansion of surgical procedures may very well reflect supply induced demand andmay be socially wasteful.65 As discussed, more intense treatment often also carries greaterrisks. The increase in mortality rate is certainly a concern and may reflect this risk. Fur-thermore, more intense treatment poses additional health care costs. One argument put forthto ban specialty hospitals is that they are responsible for increasing health care costs amongspecialty services in a market. My findings suggest that they may also be driving up costsin non-specialty services, by increasing the volume of elective and general surgeries amongstprivate payers, procedures which are quite costly.2.7 ConclusionMuch of the existing literature ignores hospital spillovers when examining policy changes andshocks to profits in specific service lines. For example, many studies analyze how hospitalsrespond to changes in Medicare reimbursement for a given procedure (e.g. angioplasty),focusing on the impact on that particular procedure as well as the volume of procedures that65It seems unlikely that changes in prices would drive all the results. I do not observe prices paid by privateinsurance companies. However, the existing literature suggests that private prices tend to follow Medicareprices (Clemens & Gottlieb (2013)), which are notoriously sticky and are administratively set (partly underthe influence of congress). Nevertheless, it is certainly possible that specialty hospital entry could affectprivate prices in a market. Since I am examining the immediate, short run response of general hospitals toincreased specialty hospital competition, the negotiation process between private insurers and each hospitalwould have to be take place and be completed very quickly for price changes to have such an immediate impact.Furthermore, it seems that the prices most likely to be affected would be those for specialty services. My studyinstead focuses on the hospital response in non-specialty services, finding that hospitals have a sophisticatedand targeted response in the short run.432.7. Conclusionare substitutes (e.g. coronary artery bypass surgery). Little research analyzes the extent towhich the hospital adjusts other services in response to such changes (e.g. non-cardiac care).This paper finds strong evidence that hospitals cross-subsidize and differentiate medicaltreatment across payer types. It contributes to the existing literature by providing a morecomplete picture of hospital spillovers and treatment differences across payer types. It alsosheds light on the broader issue of how hospitals respond to financial shocks. I find that thehospital response to the loss of a profitable service line is very sophisticated. Hospitals practiceboth revenue augmenting and cost-cutting behavior in other lines of care, targeting specificprocedures and payers according to their profitability. Specifically, they increase the numberof surgical procedures and perform more marginal surgeries. This varies with the serviceline and the payer type. The effects are concentrated in medical specialties where there aremore discretionary surgeries and higher profit margins. Hospitals also increase the intensity oftreatment among private payers, by increasing their length of stay. Furthermore, hospitals cutback on unprofitable treatment by reducing non-elective admissions and uninsured electivecare.The findings of this paper suggest that hospital responses to financial shocks are muchmore sophisticated and targeted than previously thought. Hospitals are able to adjust theirmix of services and are able to differentiate treatment by payer type. My findings suggest thatfocusing only on substitution effects within a service line, as much of the existing literaturehas done, ignores important hospital responses and leads to incomplete welfare implications,particularly among different payer groups.44Chapter 3How Much Does Health InsuranceMatter for Young Adults?: Its Roleon Primary Medical CareConsumption3.1 IntroductionIn 2010, almost one third of individuals aged 19 to 29 years were without health insurance inthe United States, making it the age group with the highest proportion of uninsured. In fact,young adults are grossly over-represented amongst the uninsured, comprising 13 million ofthe 47 million Americans without insurance (National Conference of State Legislatures 2011).Numerous factors likely contribute to the low take-up of insurance among young adults,including entry-level wages, jobs without employer sponsored insurances, and high healthpremiums that are unaffordable for a group just at the start of their careers. Importantly,young adults form a relatively healthy group that is less dependent on receiving medicalservices so the cost of insurance may outweigh the perceived benefits.Recent federal and state policy has sought to target the low insurance rates among youngadults. The Patient Protection and Affordable Care Act (PPACA), the principle legislation ofhealth care reform, extended dependent coverage so that individuals can now remain on theirparents’ insurance plans until the age of 26, a sizeable increase over the traditional age of 19years. This comes on the heels of numerous state mandates also extending dependent coverage.The new federal law came into effect September 2010, and emerging research is finding thatyoung adults are now 30 percent more likely to be on their parents’ plans compared to beforethe law (Antwi, Moriya, Simon 2012). The explicit goal of health reform is to increase theaccess to and affordability of medical care among individuals, particularly those most in needsuch as individuals with pre-existing conditions. It is still too early for a thorough evaluationof expanded coverage among young adults; however, there are two clear potential welfareimplications. First, there is likely to be a societal impact in having these lower medical riskindividuals in the insurance pool through the cross-subsidization of higher risk individuals,453.1. Introductionsuch as the elderly and those with chronic conditions, and keeping premiums low. The reformmay also reduce costs paid by others for uncompensated care by uninsured young adults.66Secondly, there may be private welfare gains to young adults themselves in terms of increasedaccess to medical care and, in turn, improvements in their health.This study aims to shed light on the private welfare effects by examining the causal impactof young adults’ insurance status on their medical care utilization, out-of-pocket medicalexpenditures, and health outcomes. If expanded insurance coverage only increases moneyspent on premiums and leads to no differences in access to healthcare, then this questionsthe private welfare benefits of recent policies. Prior to health reform, young adults couldweigh the cost of health insurance with its expected benefits. Unlike other types of insurance,there is a large part of medical spending that is predictable, such as annual preventativescreenings, routine dental visits, and prescription drugs. In the case of emergency care,hospitals must treat individuals regardless of their insurance status or ability to pay fortreatment. A distinguishing feature of young adults is that they tend to be healthy, consumelittle medical care, and have low out-of-pocket medical expenditures. In any given month,only 14 percent of 19 to 25 year olds have a medical visit, which is a third of the rate forthose over 60.67 In fact, 40 percent of young adults consume no medical care in a given year.Less than 4 percent of uninsured young adults report having difficulty affording necessarymedical care, with their average annual out-of-pocket medical expenditures being small, atapproximately $260 (in 2011 dollars). This weighs against the average annual premium of$3,380 for PPACA extended dependent coverage. The distribution of medical expenditures isheavily right skewed, suggesting that the cost of health insurance is more than what the vastmajority of uninsured individuals would otherwise expect to pay out-of-pocket for care.68If young adults are risk neutral, it may be rational to forgo insurance and not pay the highpremiums if the costs exceed the expected benefits. At the same time, one of the main reasonsindividuals purchase health insurance is to minimize the risk associated with the possibilityof a negative medical shock and to smooth income/consumption over time. Even youngadults with extreme events, however, have relatively low out-of-pocket expenditures, with thetop one percent of uninsured young adults spending an average of $7,200 out-of-pocket in66Note that these external impacts may not necessarily increase societal welfare and may just be transfersbetween individuals.67A medical visit includes doctor, dental, hospital, and emergency department visits. Approximately 32percent of adults over 30 years have a medical visit in a given month, while 45 percent of adults over 60 yearshave a visit. The statistics in this paragraph are derived using data from the Medical Expenditure PanelSurvey (MEPS), years 1997 to 2006. This dataset is described in more detail later on in the text.68The median total out-of-pocket expenditures for uninsured young adults is nearly $0, the 75th percentile$175, the 90th percentile $630, and the 99th percentile $3,800. Previous studies suggest that adverse selectionin the health insurance markets leads to actuarially unfair premiums for young adults (Levine, McKnight, andHeep 2011). This is likely to continue under the PPACA policy, where caps have been introduced on the ratioof insurance premiums between older individuals to those of young adults, which is expected to effectivelyresult in the young insured population simply subsidizing the older population in the risk pool (see Damlerand Houchens 2012).463.1. Introductiona year, which is considerably lower than the $17,100 spent by the top one percent of thoseover 60. Risk averse individuals consequently might prefer to self-insure by setting asideprecautionary savings rather than buying health insurance. Alternatively, they could jointhe growing number of individuals using catastrophic insurance plans with high deductibles.However, young adults have low levels of wealth-at-risk for bankruptcy. Thus, bankruptcy actsas an implicit form of high deductible health insurance for low asset households by crowdingout conventional health insurance coverage (Mahoney 2012). Given there may exist less costlyalternatives for young adults to smooth their income over time, it may be rational for someyoung adults to forgo health insurance.69This study examines how having health insurance affects young adults’ medical care usage,focusing on primary care, such as physician office visits, dental care, and prescription use.Such care is not only important for treating non-urgent illnesses and obtaining screenings,but also plays a key role in referring patients to specialists and helping them navigate themedical system. Also, many psychiatric disorders are not revealed until young adulthood,underlying the importance of access to care. There may be additional spillovers for youngadults in obtaining routine care, such as forming ties with primary care physicians, takingresponsibility for their own care, and habit formation for preventative screenings. These maynot necessarily result in immediate health benefits for young adults, but are likely to have apositive impact on their health in the future as they age.70Individuals with insurance may differ from those without in many unobserved ways suchas medical risks, discount rates, and risk aversion. To obtain causal estimates of the impact ofhealth insurance on primary medical use, I exploit insurance policy rules governing the eligibil-ity of young adults. Prior to recent laws, private health insurers would typically cease coveragefor dependents at 19 years, reflecting regulations in the federal tax code. Additionally, the twomain public insurance programs for children, Medicaid and SCHIP, both reclassify children asadults the day they turn 19. These policies create quasi-experimental variation in insurancestatus which I exploit in a regression discontinuity framework by comparing individuals justyounger than 19 years to those just over.This paper complements a recent study by Anderson, Dobkin, and Gross (2012) who usethe same regression discontinuity design to examine the impact of losing coverage at age 19on emergency department and hospital visits. They find that an absence of insurance leads tolarge declines in both emergency department visits and inpatient hospital admissions. Theirfindings suggest that uninsured individuals do not substitute emergency department care forprimary care or, if they do, the substitution is swamped by a reduction in regular emergencyvisits. If individuals are not receiving primary care in a hospital setting, then a key question69It should be noted that bankruptcy acting as a form of health insurance may be individually optimal, butit clearly is not socially optimal. The recent federal mandates will curb this practice.70This is particularly true in the case of preventative care for women as females not only have more recom-mended preventative screenings, but these screenings start earlier in life.473.1. Introductionbecomes whether they are consuming this type of care in a cheaper setting, such as physicianoffices, or are they simply forgoing these types of care altogether? If it is the former, thenthere are efficiency gains to be had in making individuals more price sensitive; however, if itis the latter, then there is the concern that young adults are not getting the necessary carethey require.This study provides a complete picture of how young adults’ overall medical care consump-tion is affected by insurance coverage by accounting for care outside of the hospital setting.Emergency visits and hospitalization are extreme events and rare for young adults; therefore,we cannot expect them to consume primary care in the same manner.71 Also, it is importantto examine how more general measures of day-to-day health are affected as hospital visitscapture more severe illnesses. Previous research on health insurance has largely focused onthe effects of public insurance expansions; however, these policies target individuals with verydifferent medical risks (young children, pregnant women, and the elderly). It is unlikely thatyoung adults will be affected by insurance coverage in the same way. Furthermore, thesestudies cannot isolate the causal impact of insurance status from crowd-out effects associatedwith individuals moving between different insurance schemes. In the context of recent federaland state policy, it is of particular interest to understand the impact of having insurance andto isolate the effects for young adults. This study addresses these issues.This paper finds that the 19th birthday played a significant role in insurance coveragerates of young American adults over the last decade. The estimated decline is on the orderof 3 to 5 percentage points and is driven primarily by the loss of private rather than publicinsurance. Consistent with federal tax regulations, the effect is strongest for non-studentsand those living at home. Those with severe chronic conditions are amongst the least affectedby private insurance loss, which is noteworthy given this group has been targeted in recenthealth policies. Interestingly, even young adults from families with high incomes experience asignificant and sizeable decline in private insurance (4 to 5 percentage points), suggesting thatit may be a conscious decision for some to forgo health insurance. This study finds no effect ofinsurance loss on office-based physician visits and new prescription drug use; although, thereis an increase of approximately $28 per month in out-of-pocket office expenditures amongstthose who lose insurance, on average, which is heavily driven by those at the very top end ofthe expenditure distribution. There is a significant 21 percent point decline in dental visitsand a decline of $46 per month in total dental care spending amongst those who lose insurance,which may reflect the perceived discretionary nature of dental care spending. Importantly,no change is found in the self-reported ability of young adults to afford necessary medicalcare. Interestingly, I find no evidence of anticipation effects whereby individuals stock up on71For example, in any given month, 1.40 percent of young adults aged 17 to 21 visit the emergency depart-ment, while 0.27 percent have a hospital inpatient visit. These figures compare to the 10.51 percent of youngadults who have an office visit. These statistics are derived using the MEPS, years 1997-2006.483.2. Previous Literaturemedical care prior losing their insurance, suggesting that young adults are myopic in theirprimary medical care consumption. Although this study only considers short run effects, noimmediate effect of insurance loss is detected for health status. It is possible that the healthbenefits of insurance occur in the future. Given that those losing insurance at age 19 areamongst the healthiest and most able to afford care, the findings in this paper are perhapsnot surprising. Overall, these results question the expected private welfare benefits of recentfederal health policies for most young adults, particularly when there may exist less costlyalternatives to smooth income over time.The remainder of the paper proceeds as follows. First, an overview of previous work inthis area and the legislative background are provided, followed by a description of the dataand the empirical methodology. The estimation results are then presented and are shown tobe robust to a variety of potential confounding factors. A discussion of the policy implicationsconcludes.3.2 Previous LiteratureMany studies examining the impact of insurance coverage on medical care consumption andhealth outcomes compare insured individuals to uninsured, generally finding that the insuredare less likely to have adverse health outcomes, preventable health problems, progressed dis-ease states when diagnosed, and lower mortality rates (Hoffman and Paradise 2008; Hadley2003). Similarly, the insured are more likely to have a regular physician, receive timely care,and get preventative screenings (Institute of Medicine 2001; Buchmueller et al. 2005). Whilethese studies provide insight on the relationship between insurance and medical outcomes,they confound the types of people who choose to purchase insurance and the effect of insur-ance itself.One exception is the RAND Health Insurance Experiment (Newhouse 1982), which ran-domly assigned individuals to insurance schemes with different cost-sharing rules. As theorypredicts, cost-sharing resulted in less total spending on care, with one third fewer physicianvisits and hospitalizations compared to those with free care (Brook et al. 1983; Keeler 1992).Those with cost-sharing plans had more minor health problems, but no differences in serioushealth conditions were observed (Keeler 1992). Given that the focus was on cost-sharing,rather than insurance coverage itself, the RAND study may be limited in understanding theeffects of recent policies which aim to expand insurance. Also, it’s been over 30 years sinceit took place. In a more recent randomized insurance experiment, Finkelstein et al. (2011)use a unique lottery in Oregon that allowed low-income adults to apply for Medicaid, findinginsurance led to improved self-reported health as well as more primary care, preventativescreenings, and hospital visits.A smaller group of studies have sought to address the endogeneity of insurance take-up493.2. Previous Literaturein non-experimental settings.72 The more credible have used quasi-experimental variation in-duced by changes in eligibility rules for Medicaid and Medicare, finding that public insuranceexpansions lead to increased medical care use and better health. For example, relaxing Med-icaid restrictions for low-income children increased hospital admissions and physician visits,yet lowered the rates of mortality and avoidable hospitalizations (Curie and Gruber 1996;Dafny and Gruber 2005). Disruptions in Medicaid coverage for low-income individuals led tofewer physician visits, more unmet medical needs, and increased medical debt (Carlson et al.2006). Another branch of studies exploit the jump in Medicare coverage at 65 years, the agemost individuals become eligible, finding eligibility raises the number of hospital procedures,total list charges, and routine doctor visits, increasing more for those previously uninsuredand those with supplementary coverage (Card et al. 2008, 2009; McWilliams et al. 2003).These studies have been quite informative in understanding how individuals respond to publicinsurance expansions. At the same time, there are some limitations in the context of recentUS policy developments where private insurance is playing an increasingly important role.First, public insurance programs target very different populations (young children, the verylow income, and the elderly) than young adults who are more likely to be uninsured and havedifferent medical risks; therefore, previous studies may not easily generalize to this group.73Additionally, many individuals who gain insurance through public schemes are often insuredbeforehand, making it difficult to isolate the causal effect of having insurance, versus nothaving insurance, as noted by Anderson, Dobkin, and Gross (2012).74This study contributes to the existing literature by examining the impact of young adults’insurance status on medical care consumption and health outcomes in a causal framework.Exploiting quasi-experimental variation arising from rules used by both private and publicinsurers where individuals cease being covered on their 19th birthday, I compare those justunder 19 years to those just over in a regression discontinuity framework. These policy ruleswere first exploited by Anderson, Dobkin, and Gross (2012) (ADG herein) who examine theeffect of insurance on emergency department and hospital inpatient visits. Using a uniquedataset of hospital records from seven states, ADG find that having insurance leads to a40 percent increase in emergency department visits and a 61 percent increase in inpatienthospital admissions. The authors conclude that the newly uninsured likely do not substitute72See Freeman et al. 2008 for a comprehensive overview. Many have used potentially problematic identifica-tion strategies. For example, longitudinal data with individual fixed effects cannot control for unobserved timevarying individual characteristics which may be correlated with insurance status and health outcomes. Instru-mental variables such as self-employment status, job characteristics, or immigration status are of debatablevalidity because they may have their own direct effects on health outcomes.73The population group in the Oregon lottery experiment, who were primarily very low income adults, forexample, was found to be an exceptionally high disease group relative to the general population (Finkelsteinet al. 2011).74For example, the number of individuals who move from private coverage to Medicare at age 65 is six timesas large as those gaining insurance (Card et al. 2008), while a nontrivial 25 percent of individuals with privateinsurance schemes took-up Medicaid when they became eligible (Busch and Duchovny 2005).503.3. Legislative Backgroundemergency department care for primary care. What cannot be addressed in the ADG study iswhether young adults make up this lost care in other settings or whether they simply forgo italtogether. As discussed above, primary care plays a key role in treating non-urgent illnesses,diagnosing psychiatric disorders, and obtaining preventative screenings. Additionally, it actsas a gateway to the medical care system, particularly specialist care.This study provides a deeper understanding of how young adults’ overall medical careuse is affected by health insurance and builds upon the work of ADG by accounting forconsumption outside of the hospital setting. Using a nationally representative survey, I focuson office-based physician visits, prescription drug use, and dental care, looking at both theincidence of consumption as well as expenditures. Day-to-day health and the ability to affordcare are also examined. This study is among a handful that can shed light on how youngadults are affected by health insurance, which is particularly relevant given recent federal andstate policies aim to reduce the number of uninsured young adults. Moreover, this studyfocuses on a group of individuals most likely to be uninsured, isolates the impact of losinginsurance coverage on health outcomes, and considers the role of private insurance.3.3 Legislative BackgroundPrior to recent federal and state regulations, the 19th birthday has been the critical milestoneat which young adults are often dropped from their parents’ employer sponsored policiesor from public insurance programs. Employer-sponsored health insurance is the mainstayof most family and dependent coverage. The federal tax code allowed tax-free coverage ofchildren up to age 19 (or 24 if a full time student) so long as they lived at home for more thanhalf the year and were financially dependent on their parents (Department of the TreasuryInternal Revenue Service 2009).75 Therefore, until recent legislation, most plans only covereddependents under 19 years of age unless they were full time students. Even if employers didoffer coverage to children over 19 years, there is a strong disincentive for parents to keep themon their plans because it would count as a taxable benefit given their children no longer qualifyas dependents under the federal tax law (Levine et al. 2011, Barber and Nguyen 2009).In the last five years, many states began to mandate extended dependent coverage beyond19 years. Prior to 2006, only four states mandated extended coverage; however, 31 states didby 2010 (Cantor et al. 2012).76 Dependent coverage requirements now vary across states, butthe majority has a limiting age of dependency less than 26 years (typically 21 or 25 years)with different ages for students and imposes the dependent to be unmarried. Some states also75Most states also imposed the additional requirement that the dependent be unmarried.76Utah was the first state to extend dependent coverage up to age 26 in 1995, North Dakota followed suitin 1995 with the upper age limit of 22 for nonstudents. In 2003, New Mexico raised the age to 25 years, whileTexas did the same in 2004. Two states increased the eligibility age for dependent coverage in 2006 (Coloradoraising it to 25 years and New Jersey to 30 years), while fourteen states raised it in 2007.513.3. Legislative Backgroundrequire the dependent to reside with the parent or at least in the same state. The PPACApolicy of extended dependent coverage was implemented in September 2010, superseding thelaws of many states whose dependent eligibility laws were more stringent. In particular, thefederal law has a limiting age of 26 years regardless of student status, allows married youngadults to stay on their parents’ plans, and does not impose any residency requirements. Whilestate laws generally do not apply to self-funded insurance plans, the federal law does. Thenational tax code has been changed, allowing tax free coverage under 26 years. Emergingresearch is finding a large and significant increase in the probability of young adults beingcovered under their parents’ plans since PPACA extended dependent coverage (30 percentmore likely than before) with the effects being concentrated among non-students and Whites(Antwi, Moriya, and Simon 2012).Rules governing public insurance also play a role in young adults losing health insurancewhen they turn 19. Under both Medicaid and SCHIP, children are defined as those less than19 years of age and are reclassified as adults on their 19th birthday. Medicaid is means-tested, targeting the poor and the near poor, and has different income eligibility criteriaacross demographic groups, with more stringent requirements for adults.77 There is a lotof variation in the eligibility rules across states. However, many use the federally mandatedincome eligibility limits of 133% and 100% of Federal Poverty Line (FPL) for young andold children, respectively, in defining their Medicaid eligible population (Baicker and Rehavi2010).SCHIP, on the other hand, provides states with federal funds to expand health insuranceexclusively among children. Enacted in 1997 by the Balanced Budget Act, it targets childrenjust above the poverty threshold (the near poor), with incomes that exceed Medicaid eligibilityrequirements. Rollout of SCHIP varied across the country, but all states had begun to enrollchildren into these programs by the end of 1999, with a modest take-up (Rosenbach et al.2003; Gruber and Simon 2008).78 States adopted SCHIP income eligibility limits between133% and 350% of the FPL, with most at 200% or less.79 Once children hit their 19thbirthday, they become subject to the more stringent Medicaid eligibility criteria for adults.77Mandatory Medicaid populations include children under 6 years old who are below 133% of the FPL,children 6 to 19 years below 100% of the FPL, pregnant women below 133% of the FPL, and parents withincomes below state welfare eligibility levels (typically <50% of the FPL). States have discretion to cover op-tional populations, such as children and parents above mandatory coverage limits and receive federal matchingpayments if they do.78Gruber and Simon (2008) estimate the initial take-up of SCHIP to be modest at 5 to 15 percent, likelya result of most newly eligible children (80 percent) already having private insurance. They find very largecrowd-out effects, with the number of privately insured children falling by about 60 percent as much as thepublicly insured increased.79Specifically, 27 states adopted at exactly 200%, ten between 133% and 200%, and 13 between 200% and350% (Kronebusch and Elbel 2005). States also had the option of implementing SCHIP as either a Medicaidexpansion (i.e. use the same benefit packages, managed care organizations, and providers as in their existingMedicaid program), as a separate program, or as a combination of the two. Among the states, 15 chose tointegrate it as part of Medicaid, 16 as a separate program, and 19 states as a combination program.523.4. DataMost states do not provide public coverage to childless adults since they do not receive anyfederal funds to do so.80Young adults have traditionally been at risk of becoming uninsured when they turn 19years old, aging out of both their parent’s insurance plans and public insurance programs.They typically have low-wage, entry-level, and temporary jobs that do not offer employer-sponsored insurance (Schwartz and Schwartz 2008). They are also relatively healthy with lowmedical risks, so the cost of premiums may exceed the expected benefits. Additionally, withthe exception of those with chronic illnesses, not buying insurance in the young adulthoodperiod does not preclude them from purchasing it in the future when they may have moremedical needs. These factors likely all contribute to young adults having among the lowestrates of insurance coverage.3.4 DataThe data used in this study come from the Medical Expenditure Panel Survey (MEPS), acomprehensive dataset on health care utilization, insurance coverage, and medical expendi-tures by individuals of all ages. It is produced by the Agency for Healthcare Research andQuality. MEPS draws from a nationally representative sample of US families, with a rollingpanel design. A family is interviewed five times (rounds) over two full calendar years. Indi-viduals who leave the original family unit are followed and remain in the survey. Every year,a new panel of approximately 15,000 individuals is added to the survey. Thus, two panelsare overlapping at any given point in time, resulting in roughly 30,000 individuals being in-terviewed each year. Initiated in 1996, the MEPS has interviewed 16 panels of individualsto date. The sample used in my analysis includes all individuals 17 to 21 years old. I onlylook at years 1997 to 2006 due to state and federal legislation. In particular, SCHIP was onlyimplemented in 1997 and is the main public insurer for older children. Additionally, giventhat many states laws extended dependent coverage beyond 19 years in 2007 or later, I onlyinclude up to 2006.The MEPS is well suited for the present study as it collects detailed information onindividuals’ insurance coverage and medical care consumption over the entire two years theyare in the survey. Respondents are asked about their insurance coverage and the type (e.g.employer sponsored, Medicaid/SCHIP, etc.) in each calendar month. I examine the impactof turning 19 on whether the individual has any type of medical insurance plan (private orpublic); whether the individual has private insurance; and whether the individual is coveredunder public insurance. I create three dummy variables representing coverage in a given80Childless adults do not form part of Medicaid’s “optional” groups. The few states that do provide publicinsurance to this group generally provide limited coverage far below Medicaid benefits (Schwartz and Damico2010).533.4. Datamonth. Given the two year panel, there are up to 24 observations per individual for insurancecoverage.Young adults’ insurance coverage rates vary across age and income. Given the legislationdescribed above, we would expect lower insurance rates for those over 19 years compared toyounger individuals. Family income relative to the FPL is the main measure for whetherindividuals qualify for public insurance, so we would also expect higher income groups tohave greater rates of private insurance and lower rates of public insurance compared to poorerindividuals. Table 3.1 confirms these expected patterns. The insurance coverage rate is shownfor all young adults, those 19 years or younger, and those older than 19. Two samples areexamined: the full sample of 17 to 21 year olds and those with family incomes 125% or more ofthe FPL.81 As will be shown in the next section, most of the change in insurance coverage atage 19 comes from a decline in private insurance, with very little change in public insurance.As such, for most of the subsequent analysis I focus on individuals at or above 125% of theFPL, who are less likely to qualify for public insurance. Approximately 68 percent of all17-21 year olds have health insurance, with older individuals being less likely (59 percent)compared to the younger age group (75 percent). Almost 50 percent of all individuals haveprivate insurance, whereas just over 20 percent have public. For the sample of individualsabove 125% FPL, 72 percent are covered, and a greater proportion has private insuranceversus public, with the size of the decline in private insurance between age groups beinglarger compared to the full sample.In addition to insurance coverage, the MEPS asks individuals about the medical servicesthey used over the round, such as physician visits, outpatient services, and prescription refills,as well as the frequency with which they used them. Information on expenditures and sourceof payment for care is also collected. To supplement and verify the accuracy of informationreceived from individuals, MEPS obtains data from medical providers which individuals re-ported to have visited, such as physicians, hospitals, and pharmacies. Collected informationincludes date of visit, reason for visit, diagnosis, and payment information.To measure non-urgent medical care consumption, I focus on office-based physician vis-its, prescription use, and dental visits. Office visits cover non-emergency medical care thatoccurs in a variety of settings such as doctor’s/group practice offices (general practitionersand specialists), medical and surgical centers, community health clinics, and laboratory orx-ray facilities. New prescription drugs are examined rather than refills because informationon the exact date a prescription is filled is only collected for new prescriptions.82 In terms of81The MEPS categorizes individuals into one of five groups: i) the poor (100% or less of the FPL); ii) thenear poor (100-124% of FPL); iii) low income (125-199% of FPL); iv) middle income (200-399% of FPL); andv) high income (400% or more of FPL). Table B.1 breaks down insurance coverage by each grouping, showingthose under 125% of FPL predominately have public insurance, while those over 125% of FPL primarily haveprivate insurance.82In regards to prescriptions refills, information is obtained only on the round they are filled, not the543.4. DataTable 3.1: Insurance Coverage by Age and Income GroupInsurance Type 17-21 Years ≤ 19 Years >19 YearsFull SampleAny Insurance 0.6775 0.7548 0.5876(0.4675) (0.4302) (0.4923)Private Insurance 0.4897 0.5256 0.4479(0.4999) (0.4993) (0.4973)Public Insurance 0.2064 0.2486 0.1573(0.4047) (0.4322) (0.3641)Observations 232,539 124,273 108,266Family Income>125% FPLAny Insurance 0.7216 0.7924 0.6387(0.4482) (0.4056) (0.4804)Private Insurance 0.6144 0.6616 0.5591(0.4867) (0.4732) (0.4965)Public Insurance 0.1266 0.1507 0.0984(0.3325) (0.3577) (0.2978)Observations 164,865 88,392 76,473Notes: All insurance variables are coded as 0/1 dummy variables at theindividual level, so these statistics reflect the proportion of individuals meet-ing the specific criteria in a month. Standard errors in parentheses. Those19 years and under comprise of 17 to 19 year olds, while those over 19 yearsare between 19 and 21. Since insurance coverage information is collectedeach month, an individual forms up to 24 observations in the analysis. Datacome from the Medical Expenditure Panel, years 1997-2006.553.4. Datadental care, I group together visits to general dentists, dental hygienists, and orthodontists,and care may be for routine check-ups, treatment, or accidents. Visits and prescriptions re-lating to pregnancies are excluded from the analysis due to the special nature of maternitycare in insurance coverage.83 Medical care use at the month level is the focus of the analysis,specifically whether an individual has a particular type of consumption (0/1 dummy variable)and if so, the amount of monthly expenditures.84 I focus on out-of-pocket (OOP) and totalexpenditures (i.e. the sum of out-of-pocket, private and/or public insurance amounts paid,and third party payers).85 Expenditures for all individuals are examined, as well as for onlythe subsample who report having a visit in a given month in order to narrow the analysis toheavier users of medical care. All expenditures are adjusted for inflation and are expressed inyear 2000 dollars.Table 3.2 shows that approximately 11 percent of young adults have an office visit in agiven calendar month, 4 percent fill a new prescription, and 5 percent have a dental visit.Younger and higher income individuals are slightly more likely to consume medical care.Monthly medical care expenditures are quite low for young adults, with the mean OOPexpenditures being only $3.58 for office visits ($34.02 for those with a visit) and $5.01 fordental visits ($95.54 for those with a visit). Average total monthly expenditures are $15.86for office visits ($150.91 for those with a visit) and $11.44 for dental care ($218.05 for thosewith a visit). Average unconditional expenditures are lower for older individuals, likely drivenby their having fewer visits on average. The figures in this table overall seem to suggest thatyoung adults are low consumers of primary medical care.To better understand the nature of young adults’ office-based physician visits, Table 3.3provides more detailed information, showing that about half are to seek diagnosis or treatment,while 15 percent are for general checkups, and roughly 9 percent are for mental counseling(Panel a). Among visits recorded to be with specialists, nearly 50 percent were by general,family or pediatric practitioners, 9 percent by gynecologists, and 7 percent by psychiatrists(Panel b). The top medical conditions diagnosed or treated in office settings, shown in thelast panel, are: injuries such as sprains and strains (17 percent), respiratory conditions suchas allergies or asthma (17 percent), and mental disorders such as depression and anxiety (16exact date, making it impossible to accurately construct total prescription drug use and expenditures for eachcalendar month.83In particular, most private plans do not cover pregnancies for dependents, while pregnant women mayqualify for coverage under Medicaid.84Again, there are up to 24 observations per individual on medical care use and expenditures. For individualsnot interviewed in a given round and for whom no medical care consumption is found, I code the visit andexpenditure information as missing. Alternatively, for those who are interviewed each round and for whom nomedical use is found, I code the visit and expenditure information both as zeros, indicating they have had noconsumption in each month covered by the round.85For non-fee for service managed care plans (i.e. capitation arrangements), the MEPs imputes totalexpenditures as if the provider were reimbursed on a discounted fee-for-service basis using similarly completedevents paid for on a fee-for-service. As discussed above, there is no monthly expenditure information onprescription drugs.563.4. DataTable 3.2: Medical Visits and Expenditures by Age and Income GroupFull Sample Income>125% of FPLVariable 17-21 Years ≤ 19 Years >19 Years 17-21 Years ≤ 19 Years >19 YearsMedical Care ConsumptionAny Office Visit 0.1051 0.1150 0.0936 0.1140 0.1265 0.0993(0.3067) (0.3190) (0.2913) (0.3178) (0.3324) (0.2991)Any New Prescription 0.0418 0.0431 0.0404 0.0433 0.0450 0.0413(0.2002) (0.2030) (0.1970) (0.2035) (0.2074) (0.1989)Any Dentist Visit 0.0525 0.0629 0.0403 0.0613 0.0738 0.0466(0.2229) (0.2428) (0.1966) (0.2398) (0.2614) (0.2107)Medical Care ExpenditureOOP $ on Office Visit 3.58 3.72 3.41 4.26 4.65 3.80(48.77) (47.15) (50.60) (53.82) (54.59) (52.91)Total $ on Office Visit 15.86 17.04 14.49 17.49 19.42 15.21(142.50) (137.29) (148.32) (142.37) (144.59) (139.69)OOP $ on Office (If Visit) 34.02 32.35 36.40 37.34 36.72 38.26(146.98) (135.69) (161.71) (155.49) (149.60) (163.92)Total $ on Office (If Visit) 150.91 148.21 154.76 153.38 153.52 153.17(415.75) (380.17) (461.82) (396.19) (380.35) (418.78)OOP $ on Dentist Visit 5.01 5.95 3.92 6.15 7.45 4.63(85.08) (96.14) (70.04) (93.27) (106.97) (74.01)Total $ on Dentist Visit 11.44 13.92 8.56 13.73 16.76 10.17(129.91) (148.00) (104.95) (142.53) (162.94) (113.92)OOP $ on Dentist (If Visit) 95.54 94.56 97.32 100.41 101.01 99.29(359.65) (372.21) (335.69) (364.05) (381.62) (328.99)Total $ on Dentist (If Visit) 218.05 221.17 212.38 224.01 227.05 218.37(526.02) (549.85) (479.74) (533.37) (558.64) (482.94)Observations 232,539 124,273 108,266 164,865 88,392 76,473Notes: The medical care consumption variables were coded as 0/1 dummy variables at the individual level, so these statistics reflect theproportion of individuals meeting the specific criteria in a month. Medical care expenditures are over a month and are real values expressedin 2000 dollars. Standard errors in parentheses. Those 19 years and under comprise of 17 to 19 year olds, while those over 19 yearsare between 19 and 21 years of age. Since medical care consumption information is collected each month, an individual forms up to 24observations in the analysis. Data come from the Medical Expenditure Panel, years 1997-2006.573.4. Datapercent).The other outcomes examined in this study are individuals’ ability to afford necessarymedical care and their health status. Individuals are asked each round whether they forwent ordelayed necessary care because they could not afford it. This provides a deeper understandingof how insurance affects medical care use, specifically necessary care, which medical visits alonecannot capture as not all visits are for essential care (i.e. the case of sensitive care). A verysmall proportion (1.5 percent) report not being able to afford necessary care, as shown inTable 3.4, with older individuals having slightly more of problem (1.7 percent) compared toyounger individuals (1.2 percent). The survey respondent (i.e. the reference person) is alsoasked each round about the perceived health status of family members as well as the days theymissed school or work due to illness. Physical health is rated on a 5 point scale (excellent,very good, good, fair, poor). To ease interpretation, I normalize the scale to be between 0 and1, where those in excellent health receive a value of 1, those in poor health a value of 0, andthose in between some increment of 0.25.86 I also create two dummies indicating whether anindividual is in excellent health (1 if excellent; 0 otherwise) or at least very good health (1 ifexcellent or very good; 0 otherwise). This information is collected each round, so individualsform up to five observations in the analysis. Table 3.4 shows that young adults are generallyconsidered to be quite healthy, with 71 percent in at least very good health, 40 percent inexcellent health, and a mean perceived health index measure of 0.76. Younger individuals haveslightly higher perceived health than those over 19 years, as does the higher income sample.Information on whether an individual misses school or work due to illness is perhaps a moreobjective measure of health status, so I create two dummies indicating whether the individualmissed any school or work in a round due to illness.87 Almost 20 percent of students missschool and 20 percent of working young adults miss work because of sickness.MEPS also collects basic demographic information each round, such as employment andstudent status, living arrangements, marital status, and having a chronic illness.88 The re-mainder of Table 3.4 shows the demographic characteristics of the sample. As expected,those over 19 years of age are less likely to be a student and are more likely to be workingand have moved out of their parents’ homes. I use a stringent classification of chronic illness,defined as a functional, sensory, or cognitive limitation that has persisted for more than 3months, with a small proportion (roughly 4 percent) meeting this criteria. Family income iscollected annually, and I deflate it to year 2000 real dollars. Those above 125% of the FPLhave considerably higher average family incomes compared to the full sample ($63,200 versus$47,900), a greater proportion work and are students, while relatively fewer have moved out.The following section outlines the methodology used in the analysis.86Those in very good, good, and fair health, receive a value of 0.75, 0.50 and 0.25 respectively.87In constructing these dummies, I only include individuals who reported being in school or work (e.g. thosewho aren’t students get a missing value for whether an individual misses school).88In the public use data, which I use in this study, there is no information on which state individuals reside.583.4. DataTable 3.3: Characteristics of Office-Based Physician Visits for Young AdultsPanel (a): Top Reasons for CareMain Type of Care Received % of Visits1. Diagnosis or Treatment 51.392. General Checkup 15.243. Psychotherapy/Mental Health Counseling 9.284. Follow-Up 7.875. Immunizations or Shots 4.826. Other 11.40Panel (b): Top Specialty VisitsDoctor Specialty % of Visits1. General Practice 20.172. Family Practice 17.783. Pediatrician 10.454. Gynecology 8.675. Psychiatry 7.456. Dermatology 5.077. Other 50.58Panel (c): Top Medical Conditions Diagnosed/TreatedCondition % of Visits1. Injury (e.g. strains, sprains) 17.242. Disease of Respiratory System (i.e. allergies, asthma, sinus infections) 17.093. Mental Disorder (i.e. depressive disorder, anxiety) 16.334. Disease of Musculoskeletal System (i.e. back disorders) 11.155. Disease of the Skin and Subcutaneous Tissue (i.e. acne) 6.476. Infectious and Parasitic Diseases (i.e. strep throat, chlamydia) 5.777. Symptoms, Signs, and Ill-Defined Conditions (i.e. malaise, fatigue) 5.378. Diseases of the Genitourinary System (i.e. urinary tract, bladder infections) 5.349. Other 65.90Notes: The sample includes all individuals with an office visit between the ages of 17 and 21. Percentagesare calculated from visits with reported information. Data come from the Medical Expenditures Panel, years1997-2006.593.4. DataTable 3.4: Health Status and Demographic Characteristics by Age and Income GroupFull Sample Income>125% of FPLVariable 17-21 Years ≤ 19 Years >19 Years 17-21 Years ≤ 19 Years >19 YearsIndicators of Health StatusCan’t Afford Care 0.0146 0.0123 0.0171 0.0116 0.0104 0.0130(0.1198) (0.1102) (0.1298) (0.1072) (0.1015) (0.1135)Health Index 0.7634 0.7746 0.7504 0.7846 0.7942 0.7735(0.2324) (0.2306) (0.2337) (0.2226) (0.2218) (0.2229)Very Good Health 0.7148 0.7282 0.6994 0.7552 0.7665 0.7420(0.4515) (0.4449) (0.4585) (0.4300) (0.4230) (0.4376)Excellent Health 0.3956 0.4216 0.3657 0.4265 0.4510 0.3979(0.4890) (0.4938) (0.4816) (0.4946) (0.4976) (0.4895)Miss School 0.1966 0.2284 0.1322 0.1966 0.2333 0.1269(0.3974) (0.4198) (0.3387) (0.3975) (0.4230) (0.3329)Miss Work 0.1989 0.1892 0.2055 0.1914 0.1805 0.1992(0.3992) (0.3917) (0.4041) (0.3934) (0.3847) (0.3994)Demographic CharacteristicsMale 0.4917 0.4992 0.4831 0.5129 0.5137 0.5120(0.4999) (0.5000) (0.4997) (0.4998) (0.4998) (0.4999)White 0.7564 0.7530 0.7605 0.7797 0.7793 0.7802(0.4292) (0.4313) (0.4268) (0.4144) (0.4147) (0.4141)Student 0.5302 0.6583 0.3818 0.5828 0.7069 0.4379(0.4991) (0.4743) (0.4858) (0.4931) (0.4552) (0.4961)Moved Out 0.3184 0.2069 0.4464 0.2608 0.1547 0.3835(0.4658) (0.4050) (0.4971) (0.4391) (0.3616) (0.4862)Work 0.4716 0.3820 0.5760 0.5300 0.4352 0.6413(0.4992) (0.4859) (0.4942) (0.4991) (0.4958) (0.4796)Chronic Illness 0.0440 0.0355 0.0491 0.0339 0.0286 0.0370(0.2051) (0.1850) (0.2160) (0.1810) (0.1668) (0.1888)Family Income 47,934 49,252 46,422 63,198 64,624 61,551(42,893) (42,604) (43,174) (42,022) (41,277) (42,809)Observations 49,114 26,293 22,821 34,998 18,825 16,173Notes: The health index is as defined in the text. The other health status variables and demographic characteristics are coded as 0/1dummy variables at the individual level, so these statistics reflect the proportion of individuals meeting the specific criteria. Family incomeis in real values, expressed in 2000 dollars, and is reported annually for each individual. Those 19 years and under comprise of 17 to 19year olds, while those over 19 years are between 19 and 21 years of age. Given that individuals were sampled multiple times and the abovevariables are collected each round, an individuals forms up to 5 observations in the analysis. Data come from the Medical ExpenditurePanel, years 1997-2006.603.5. Empirical Methodology3.5 Empirical Methodology3.5.1 IdentificationThe primary relationship of interest in this study focuses on the impact of health insurancecoverage on routine medical care consumption and health status, which can be represented inthe following reduced form model:Yi = α0 + α1Di + iHere, Yi is the outcome of interest (i.e. medical care consumption, medical expenditures,or health status) for individual i; Di is a 0/1 dummy variable for whether the individual hashealth insurance. The error term i represents all other factors affecting the outcome. Thecoefficient of interest in this study is α1, which measures the impact of insurance coverage onthe medical outcome of interest. As discussed, it is difficult in practice to get a consistentestimate of α1 as insurance take-up is likely endogenous. In particular, there are likely unob-served factors in i, such as discount rates or medical risks, which are correlated with bothDi and Yi .The identification strategy employed in this study to obtain an unbiased estimate of α1is a regression discontinuity (RD) design where individuals just under 19 years old, who aremore likely to be covered by health insurance, are compared to individuals just over 19 yearsold, who are at risk of having lost their insurance. Given that individuals have no control overtheir age, the insurance policy rules described above create an exogenous source of variationin coverage around 19 years of age. Clearly, turning 19 years old is not the sole determinantof insurance status; therefore, it is a fuzzy regression discontinuity design. As outlined in Leeand Lemieux (2010), the fuzzy RD can be described by the two equation system:Yij = α0 + α1Dij + f(Aij) + vij (3.1)Dij = γ0 + γ1Tij + g(Aij) + uij (3.2)Here, Aij represents the age of individual i, in months, relative to 19 years (Aij = 0 ifexactly 19 years). The indicator j reflects the grouping structure inherent in the sample,where all individuals of the same age (i.e. born in same year and month) belong to thesame group. The function f(·) represents the relationship between age and outcome Y ;Tij = 1[Aij > 0] is an indicator for whether an individual is older than 19 years;89 and g(·)describes the relationship between age and health insurance coverage. vij and uij are error89Note that in practice, insurers typically allow individuals to remain on their insurance plans until theend of the month they turn 19 years old. It is for this reason that there is a strict inequality in the indicatorfunction of Tij .613.5. Empirical Methodologyterms. Substituting (3.2) into (3.1) gives the following reduced form relationship:Yij = β0 + β1Tij + h(Aij) + zij (3.3)where β1 = α1γ1 can be interpreted as an intent-to-treat estimate. Thus, the impact ofhaving health insurance, α1, can be derived by estimating (3.3) to obtain β1, and then scalingup this estimate by the proportion of individuals who actually lose insurance upon turning19, γ1 which can be estimated in (3.2).The interpretation of the fuzzy RD estimate requires some attention. First, as in anyRD design, the estimated short run impact of health insurance on outcomes can only beidentified at the cutoff of 19 years old. Secondly, just as in the case of Two Stage LeastSquares (2SLS), the estimate of α1 can be interpreted as a Local Average Treatment Effect(LATE) under certain conditions, as will be discussed. The LATE measures the average effectfor those individuals who had insurance prior to turning 19 years old but who age out of theirinsurance plans on their 19th birthday (i.e. the “compliers” in language of Imbens and Angrist1994). Therefore, the fuzzy RD estimate only measures the average effect for a subgroup ofthe population.903.5.2 EstimationThis study uses nonparametric methods to estimate the impact of health insurance coverageon routine medical care use. However, global polynomial estimates are also presented forcomparison.91 Estimation is done using local linear regression, which has better propertiesat boundary points compared to the standard nonparametric kernel regression and is conse-quently better suited for RD analysis (see Porter 2003 and Fan and Gijbels 1992). Specifically,local linear regression is used to estimate the limits on each side of the discontinuity, with theestimated treatment effect being the difference in the two. This was done in one step usingweighted least squares, with weights given by the triangle (edge) kernel such that ages closer90It isn’t clear to what extent the results generalize to the whole young adult population. The RD estimatesmeasure the average effect for the group of individuals who lose insurance at age 19. The average effect ofhealth insurance at age 19 may differ among individuals whose insurance status does not actually change ontheir 19th birthday (i.e. the “never-takers” or the “always-takers”). Similarly, the impact of insurance maybe different for the compliers had they instead lost insurance at age 18, age 25, or even at 19.5 years. Assuch, the extent to which the results are generalizable to the whole young adult population depends on bothhow individuals change over time as they age as well as the degree to which the compliers differ from the“always-takers” and the “never-takers”, in both observable and unobservable ways. As will be shown, there issignificant heterogeneity across population groups in the extent of insurance loss at age 19.91Estimation of the fuzzy RD is usually done with either global polynomial regressions or nonparametricmethods. It is particularly important to use the correct functional form of the forcing variable (i.e. age) asmisspecification creates a bias (see Lee and Lemieux 2010 for a comprehensive overview). Many researchersfavor nonparametric estimation over polynomial regression because it is better at accommodating regressionfunctions of various forms, is less sensitive to outliers, and puts more weight on data in the neighborhood ofthe cutoff, which is aligned with the intuition inherent in RD design (see Imbens and Lemieux 2008; Hahn,Todd, and van der Klaauw 2001).623.5. Empirical Methodologyto the cutoff of 19 years are given more weight than those further away.92Critical to local linear estimation is the choice of the bandwidth, h, the key tuning param-eter in nonparametric settings. The optimal choice amounts to balancing bias and precisionof the estimates. There are many competing approaches to selecting the optimal bandwidth,including cross-validation methods (Ludwig and Miller 2007; Imbens and Lemieux 2008); andplug-in approaches (Fans and Gijbels 1996). An overview of these methods is provided inLee and Lemieux (2010). This study selects the bandwidth using a procedure by Imbensand Kalyanaraman (IK) (2009), who develop a data dependent bandwidth choice rule thatis optimal for local estimation at the threshold, taking into account the kernel employed (thetriangle kernel here).93 Results are reported using local linear estimation with the IK band-width and half of the IK bandwidth.94 The IK optimal bandwidth varies with the dependentvariable, ranging from 0.43 of a year for insurance coverage; 0.81 for medical visits; and 1.53for health status. All RD figures shown use the IK optimal bandwidth. Additionally, resultsusing a bandwidth of 1 year are reported because: 1) It allows easier comparison of effectsacross different outcomes and demographic groups as well as to other empirical studies, par-ticularly ADG, who focus on the one year window before and after young adults lose insuranceat age 19,95 and 2) Many health providers recommend annual screenings for their patientsand provide prescription refills for up to 12 months (e.g. Pap tests, contraception). Giventhat the focus of this paper is on non-emergent and routine medical care visits, including awhole year on each side of the cutoff of 19 years may be important for deriving estimates. Inaddition to presenting nonparametric estimates, I also show results of parametric estimationusing third-order polynomials with splines, where the full sample of 17 to 21 year olds isincluded.96Analysis will focus on the reduced form estimates of Equations (3.1) and (3.2); however,the Wald estimates of α1 (i.e. where health insurance coverage is instrumented with an indi-cator for being over 19 years), are also given in the text for interpretation when necessary. Assuggested by Imbens and Kalyanaraman (2009), given that the discontinuity in the treatmentregression is typically more precisely estimated than the discontinuity in outcomes, I simplyuse the optimal bandwidth for the outcome variable in calculating the Wald estimates. Stan-dard errors were clustered by the forcing variable, Aij , to reflect the within group correlation92In particular, the weighting function used is given by wtij = max(0, 1−abs(mij)), where mij = (0−Aij)/h,Aij is age relative to the cutoff of 19 years and h is the bandwidth. The edge kernel is known to be boundaryoptimal (see Cheng, Fan, and Marron 1997).93That is, the IK optimal bandwidth is chosen in conjunction with using the triangle kernel.94In practice, the IK bandwidth was determined using the rd (Nichols 2011) and rdob (Fuji, Imbens, andKalyanaraman 2009) ado packages in Stata.95ADG estimate a linear polynomial in age, with splines, using a rectangular kernel and a one year band-width.96A third order polynomial was used since odd order polynomials have better boundary properties thaneven ordered, which is critical for RD estimation. Given the curvature of the outcome variables across age, alinear specification was not appropriate in this application.633.6. Resultsamong all individuals in group j, the importance of which outlined by Lee and Card (2008).Additionally, due to the panel nature of the data set, individuals are observed at multiple agemonths. As such, it is also important to account for serial correlation amongst individualswhen calculating the standard errors of the estimates. Thus, standard errors were two-wayclustered by age (in months) and individual, using an estimator proposed by Cameron, Gel-bach, and Miller (2011) to account for multi-way clustering that is non-nested.97It should be noted that in the RD design, covariates need not be included in estimation;however, they may help with variance reduction. In this paper, I present estimates both withand without covariates, showing they are very similar. The controls included are: dummiesfor gender, white race, live in a MSA, full-time student, married, still live with parents, surveyyear, as well as log of family income (in 2000 dollars).3.6 Results3.6.1 Change in Insurance Coverage Rates at age 19The first set of results examines the extent to which the 19th birthday played a role ininsurance coverage amongst young adults. The regression discontinuity coefficients at age19 using the full sample of individuals are reported in Table 3.5 for various specificationsdiscussed above. There is a significant decline in insurance coverage at 19 years of age, on theorder of 3 to 5 percentage points. The IK bandwidth produces an estimate of 2.71 percentagepoints while the bandwidth of one year estimates a decline of 5.11 percentage points (withoutcontrols). These results are similar to the findings of ADG who use an alternate data setand estimate a 4 to 8 percentage point decline in coverage.98 Given the mean of insurancecoverage is approximately 68 percent, these results translate into a 4 to 7.5 percent declineamongst the insured, which is not trivial. Most of the decline in insurance is driven from adrop in private (2.75 to 3.75 percentage points) rather than public insurance (0.93 to 1.43percentage points). This holds across all specifications, both nonparametric and parametric.The RD estimates are quite similar when controls are and are not included and this holdsfor virtually all results; therefore, in the interest of space, I show only estimates withoutcovariates in subsequent tables. Those that include covariates are shown in the Appendix.97This was done using the cgmreg (Cameron, Gelbach, and Miller 2011) and ivreg2 (Baum, Schaffer, andStillman 2002) ado packages in Stata.98To make the estimation further comparable to the approach taken by ADG, I also estimated using arectangular kernel for weighting rather than a triangular with a one year bandwidth. This produces anestimated decline of 6.05 percentage points in insurance and falls within their estimated range.643.6.ResultsTable 3.5: Change in Insurance Coverage Rates at 19Specification Any Insurance Private Insurance Public InsuranceMean 0.6775 0.4897 0.2064Nonparametric: IK -0.0271*** -0.0268*** -0.0275*** -0.0248*** -0.0093*** -0.0108***(0.0033) (0.0034) (0.0037) (0.0034) (0.0028) (0.0025)Observations 50,539 48,984 68,845 66,732 68,845 66,732Bandwidth 0.427 0.427 0.665 0.665 0.632 0.632Nonparametric: IK / 2 -0.0192*** -0.0214*** -0.0160*** -0.0151*** -0.0078*** -0.0089***(0.0003) (0.0010) (0.0003) (0.0008) (0.0001) (0.0005)Observations 22,978 22,266 32,198 31,193 32,198 31,193Bandwidth 0.213 0.213 0.333 0.333 0.316 0.316Nonparametric: 1 yr -0.0511*** -0.0499*** -0.0375*** -0.0341*** -0.0143*** -0.0168***(0.0069) (0.0066) (0.0040) (0.0038) (0.0019) (0.0021)Observations 105,720 102,480 105,720 102,480 105,720 102,480Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Parametric: Cubic Spline -0.0390*** -0.0374*** -0.0335*** -0.0286*** -0.0044 -0.0078*(0.0046) (0.0044) (0.0064) (0.0042) (0.0043) (0.0037)Observations 225,027 217,549 225,027 217,549 225,028 217,549Covariates No Yes No Yes No YesNotes: The RD coefficients at age 19 are reported. Standard errors are in parentheses and were two-way clustered byindividual and age (in months). Nonparametric estimation was done using local linear regression. All variables are forinsurance coverage in a given month. Covariates include dummies for male, white, msa, full-time student, married, neverleave home, survey year, as well as log of family income. Data come from the Medical Expenditure Panel, years 1997-2006.* significant at 10%; ** significant at 5%; *** significant at 1%.653.6. ResultsFigure 3.1 shows how well nonparametric estimation with the IK bandwidth fit the data.The circles are the unconditional averages of insurance coverage for each age (measured inmonths), while the solid line gives the predicted values in Equation (3.2). Panels (a) and(b) reveal sharp discontinuities for any insurance coverage and private insurance at age 19,whereas panel (c) shows only a modest decline for public insurance. Local linear estimationwith the IK bandwidth generally seems to fit the raw data fairly well near the threshold of19 years.Figure 3.1: Insurance Coverage by Type.55.65.75.85Proportion Covered 17 18 19 20 21Age(a): Any Insurance Coverage .35.45.55.65Proportion Covered 17 18 19 20 21Age(b): Private Insurance.05.15.25.35Proportion Covered 17 18 19 20 21Age(c): Public InsuranceNotes: The dependent variable is having medical insurance coverage at anytime during the calendar month. Panel (a) is any insurance coverage, (b)private insurance coverage, (c) public insurance coverage. The circles plotlocal averages at each age month, while the solid lines are fitted values fromlocal linear regressions of the dependent variable using a triangular kerneland the IK bandwidth.As outlined in the previous section, the insurance legislation affects demographic groupsdifferently. In particular, non-students and those who live at home are expected to be most atrisk of private insurance loss at age 19 given the dependent coverage requirements. Addition-ally, those at higher incomes are less likely to qualify for public insurance so should experiencelittle change in public insurance. Given states provide coverage for pregnant women and par-ents of low income households, females may be less at risk of losing public insurance relativeto males. Indeed, the RD estimates in Tables 3.6 and 3.7 confirm these patterns, where Equa-tion 3.3 was estimated separately for different demographic groups, using the IK optimal663.6. Resultsbandwidth for the subsample and a bandwidth of one year.99Table 3.6 shows that non-students are more likely to lose insurance coverage, private in-surance in particular (4.61 percentage points), compared to students (1.43 percentage points).Additionally, the decline in private insurance is greater for those who remain at home com-pared to those who have moved out (3.37 versus 2.06 percentage points). These findingsare also aligned with recent research by Antwi, et. al (2012) who find that the PPACAextended dependent coverage most affected insurance take-up for non-students compared tostudents.100 Table 3.6 also shows that males experience slightly more of a decline in coveragethan females, and there is no change in public insurance for females. There is also a greaterdecline in public insurance for blacks compared to whites, which may be explained by blackshaving a higher incidence of eligibility for public insurance prior to age 19. A notable result isthat individuals with severe chronic illnesses are primarily losing public insurance, with only asmall change in private insurance. Although the standard errors on these estimates are noisierdue to the small number of young adults with severe chronic conditions, these results suggestextended dependent coverage will have a small effect on their insurance status compared toother demographic groups.Table 3.7 examines insurance coverage changes by income group and reinforces the resultsof Table 3.5, where the loss of insurance is driven primarily by a reduction in private ratherthan public insurance for most young adults, with the exception of the poorest individualswho are targeted by public insurance programs. For example, the top two income quartilesexperience a decline of 4.49 and 5.14 percentage points in private insurance, with little changein public insurance (approximately 0.8 percentage points). Conversely, the decline in coveragefor the first income quartile is primarily driven by public insurance loss (3.67 percentagepoints). It is also noteworthy that among all income groups, those in the highest incomequartile, who are most able to afford health insurance, experience the greatest overall declinein coverage (4.83 percentage points), suggesting that it may be a conscious decision on theirpart to not buy insurance. To zero in on the impact of losing insurance on the 19th birthday,the remaining analysis will focus on individuals who are most likely to lose private insurancecoverage, rather than public insurance, given the results above. Thus, I exclude individualsbelow 125% of the FPL and instead focus only on those whose private insurance coverage ismost affected once they turn 19.101 This amounts to just over 70 percent of the full sample of99It should be noted that one general observation in Tables 3.6 and 3.7 is that the IK optimal bandwidthsin the split samples are larger than in the full sample, a consequence of there being fewer observations. It is inpart for this reason that a common bandwidth of one year is provided to facilitate comparison across groups.100Figure B.1 in the Appendix provides a graphical representation of these results.101As discussed above, most public insurance rules target young adults who are near the poverty threshold.There is significant variation across states in the Medicaid income eligibility requirements for older children.Young teenage parents themselves can also qualify for Medicaid under different eligibility rules. As such, thereis likely to not only be large differences between individuals with public vs. private insurance around thecutoff, but differences within those who have public insurance. This further motivates focusing on the group of673.6. ResultsTable 3.6: Change in Insurance Coverage Rates at 19 by Demographic GroupAny Insurance Private Insurance Public InsuranceIK BW = 1 year IK BW = 1 year IK BW = 1 yearSubsampleFemales -0.0335*** -0.0416*** -0.0381*** -0.0383*** -0.0041 -0.0041(0.0055) (0.0050) (0.0046) (0.0045) (0.0029) (0.0029)Observations 40,330 54,488 54,488 54,488 54,488 54,488Bandwidth 0.676 1.000 0.983 1.000 0.990 1.000Males -0.0342*** -0.0617*** -0.0308*** -0.0366*** -0.0141*** -0.0254***(0.0050) (0.0076) (0.0046) (0.0038) (0.0037) (0.0021)Observations 28,770 51,232 42,176 51,232 28,770 51,232Bandwidth 0.507 1.000 0.810 1.000 0.580 1.000White -0.0385*** -0.0549*** -0.0383*** -0.0441*** -0.0100*** -0.0115***(0.0050) (0.0064) (0.0056) (0.0044) (0.0023) (0.0021)Observations 45,330 80,039 66,103 80,039 66,103 80,039Bandwidth 0.549 1.000 0.769 1.000 0.798 1.000Black -0.0213** -0.0395*** -0.0207*** -0.0193** -0.0153* -0.0201***(0.0080) (0.0063) (0.0060) (0.0062) (0.0064) (0.0055)Observations 13,455 18,346 19,996 18,346 15,087 18,346Bandwidth 0.679 1.000 1.041 1.000 0.823 1.000Student -0.0231*** -0.0318*** -0.0143*** -0.0154*** -0.0139*** -0.0169***(0.0047) -0.0032 (0.0036) (0.0034) (0.0031) (0.0023)Observations 32,085 50,485 45,754 50,485 41,099 50,485Bandwidth 0.610 1.000 0.915 1.000 0.751 1.000Non-Student -0.0428*** -0.0696*** -0.0461*** -0.0582*** -0.0104*** -0.0129***(0.0056) (0.0102) (0.0057) (0.0089) (0.0031) (0.0027)Observations 31,690 54,662 41,062 54,662 45,673 54,662Bandwidth 0.551 1.000 0.736 1.000 0.816 1.000Live with Parent -0.0346*** -0.0580*** -0.0337*** -0.0430*** -0.0125*** -0.0167***(0.0042) (0.0070) (0.0057) (0.0060) (0.0034) (0.0022)Observations 34,304 72,613 53,217 72,613 46,893 72,613Bandwidth 0.458 1.000 0.708 1.000 0.606 1.000Moved Out -0.0336*** -0.0353*** -0.0206*** -0.0238*** -0.0099** -0.0099**(0.0046) (0.0059) (0.0041) (0.0039) (0.0034) (0.0034)Observations 33,107 33,107 30,382 33,107 33,107 33,107Bandwidth 0.937 1.000 0.837 1.000 0.996 1.000Chronic Illness -0.0722*** -0.0721*** -0.0285* -0.0237 -0.0458*** -0.0688***(0.0121) (0.0113) (0.0134) (0.0122) (0.0116) (0.0105)Observations 4,832 4,123 3,797 4,123 3,128 4,123Bandwidth 1.232 1.000 0.856 1.000 0.732 1.000Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clusteredby individual and age (in months). All coefficients are derived from nonparametric local linear regression with the IK optimalbandwidth for the subsample and then a bandwidth of one year. All variables are for insurance coverage in a given month. Datacome from the Medical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.683.6. Results17-21 year olds.102 The first row of Table 3.7 provides the RD estimates of insurance loss forthis group, showing that most of the change in insurance coverage (3.01 to 5.10 percentagepoints) is driven by a decline in private insurance (3.44 to 4.19 percentage points).103 AnyWald statistics provided in the later sections of the paper will consequently be based oninsurance changes for the subsample of individuals above 125% of the FPL.individuals likely to have private insurance since they are a more homogenous group and the cutoff is cleaner.Table B.1 shows that the predominate source of coverage for those under 125% of the FPL (i.e. the poor andnear poor) is public insurance, whereas it is private insurance for those above 125% of the FPL.102It should be noted that the subsequent analysis was also estimated by including those at 100% to 125% ofFPL (i.e. dropping only the poor), and the resulting estimates were nearly identical to the ones shown belowthough in some cases were less precise.103It should be noted that since the federal poverty line takes into account family composition, those under125% of the FPL are not solely in the first income quartile. While they are predominately in the first andsecond income quartiles, some individuals are also in the third and fourth.693.6.ResultsTable 3.7: Change in Insurance Coverage Rates at 19 by Income GroupAny Insurance Private Insurance Public InsuranceIK BW = 1 year IK BW = 1 year IK BW = 1 yearSubsampleIncome>125% FPL -0.0301*** -0.0510*** -0.0344*** -0.0419*** -0.0077** -0.0115***(0.0041) (0.0070) (0.0048) (0.0044) (0.0024) (0.0020)Observations 42,041 74,619 54,961 74,619 54,961 74,619Bandwidth 0.505 1.000 0.724 1.000 0.693 1.0001st Quartile -0.0425*** -0.0521*** -0.0127** -0.0130*** -0.0367*** -0.0371***(0.0054) (0.0064) (0.0040) (0.0039) (0.0040) (0.0040)Observations 24,180 29,253 29,253 29,253 29,253 29,253Bandwidth 0.774 1.000 0.966 1.000 0.981 1.0002nd Quartile -0.0254** -0.0327*** -0.0280*** -0.0266*** -0.0054 -0.0066(0.0078) (0.0064) (0.0052) (0.0054) (0.0055) (0.0049)Observations 16,612 22,556 26,585 22,556 18,574 22,556Bandwidth 0.739 1.000 1.086 1.000 0.799 1.0003rd Quartile -0.0431*** -0.0536*** -0.0514*** -0.0453*** -0.0082* -0.0091**(0.0068) (0.0080) (0.0105) (0.0063) (0.0036) (0.0035)Observations 20,008 24,209 32,541 24,209 24,209 24,209Bandwidth 0.797 1.000 1.255 1.000 0.951 1.0004th Quartile -0.0483*** -0.0548*** -0.0449*** -0.0497*** -0.0084*** -0.0086***(0.0056) (0.0056) (0.0058) (0.0050) (0.0022) (0.0021)Observations 19,291 29,702 21,862 29,702 29,702 29,702Bandwidth 0.656 1.000 0.748 1.000 0.958 1.000Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clusteredby individual and age (in months). All coefficients are derived from nonparametric local linear regression with the IK optimal bandwidthfor the subsample and a bandwidth of one year. All variables are for insurance coverage in a given month. Data come from the MedicalExpenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.703.6. Results3.6.2 Change in Medical Care Consumption at age 19Next, I examine the impact of insurance loss on medical care consumption, namely office vis-its, new prescription drugs, and dental visits. Specific types of office visits are also examined,such as those for diagnosis or treatment, checkups, and mental counseling. The reduced formRD estimates are shown in Table 3.8. Across all specifications, there is no significant effect ofturning 19 on total office visits and new prescription drugs. These estimates are precise, beingclose to zero with small standard errors relative to the mean.104 There is, however, a declinein dental visits on the order of 1 percentage point, as shown in all specifications. The corre-sponding Wald estimate, derived by scaling up the reduced form estimate by the estimatedchange in insurance coverage amongst those with reported medical visits is approximately21 percentage points. I also examined whether these results differ by gender, as females andmales require different medical screenings and may have different health behavior. I find thereis no change in office visits and prescription drugs for both genders, but the overall decline indental visits is driven by females who have more visits (see Table B.2 in the Appendix). Thesignificant decline of dental visits but not other forms of routine care may be explained bythe perceived discretionary nature of dental care spending. Figure 3.2 plots the raw data formedical care consumption, along with the predicted values from the nonparametric analysisusing the IK optimal bandwidth.Although there is no change in overall office visits, it is possible that this result maskschanges in specific types of care. For example, upon losing insurance, individuals may forgomedical care that is less urgent and whose benefits are less salient, such as annual checkups.105Table 3.8 shows this is not the case, where there is no change in office visits for checkups andmental counseling. Additionally, there is no change in visits for diagnosis/treatment as shownacross all specifications, with the exception of half the IK bandwidth. These results largelysuggest that young adults still obtain the same mix of office care once they lose insurance.106104I can rule out declines in use greater than 1.77 percentage points for office visits and 0.83 percentage pointsfor prescription drugs at the 95 percent confidence level. These are the lower bounds of the 95-percent confidenceinterval of the estimates, representing the smallest magnitudes whose value I can distinguish statistically fromthe point estimates.105It should be acknowledged that the coding of the type of office visit may be affected by insurance coverage.106A graphical representation of these results is shown in Figure B.2 of the Appendix.713.6.ResultsTable 3.8: Change in Medical Care Use at 19Office VisitsSpecification All Diagnosis or Treatment Checkup Mental Counseling New Prescriptions Dental VisitsMean 0.1140 0.0622 0.0245 0.0109 0.0433 0.0613Nonparametric: IK -0.0065 -0.0076 -0.0026 -0.0004 -0.0027 -0.0103**(0.0057) (0.0045) (0.0031) (0.0006) (0.0029) (0.0039)Observations 61,501 74,619 61,501 61,501 68,073 74,619Bandwidth 0.812 0.939 0.820 0.778 0.904 0.944Nonparametric: IK / 2 -0.0046 -0.0102** -0.0012 -0.001 -0.0032 -0.0074*(0.0054) (0.0036) (0.0040) (0.0010) (0.0025) (0.0038)Observations 29,131 35,588 29,131 29,131 35,588 35,588Bandwidth 0.406 0.470 0.410 0.389 0.452 0.472Nonparametric: 1 yr -0.0064 -0.0072 -0.0032 -0.0004 -0.0027 -0.0103**(0.0058) (0.0045) (0.0029) (0.0009) (0.0029) (0.0038)Observations 74,619 74,619 74,619 74,619 74,619 74,619Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Parametric: Cubic Spline -0.0045 -0.0079 -0.0023 0.0005 -0.0019 -0.0133**(0.0066) (0.0050) (0.0036) (0.0017) (0.0034) (0.0047)Observations 160,243 160,243 160,243 160,243 160,243 160,243Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clustered by individual and age (inmonths). All nonparametric coefficients are derived using local linear regression. All variables respond to medical care use in a given month. The sample consistsof individuals above 125% of the Federal Poverty Line. Data come from the Medical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at5%; *** significant at 1%.723.6. ResultsFigure 3.2: Medical Care Consumption.07.09 .11.13.15Had Office Visit 17 18 19 20 21Age(a): Office Visits 0.02.04.06.08Filled New Prescription 17 18 19 20 21Age(b): New Prescription Drugs.02.04.06.08 .1Had Dental Visit 17 18 19 20 21Age(c): Dental VisitsNotes: The dependent variable is having a particular medical visit in a givencalendar month. Panel (a) is office-based physician visits, (b) newprescription drugs, and (c) dental visit. The circles plot local averages ateach age month, while the solid lines are fitted values from local linearregressions of the dependent variable using a triangular kernel and the IKbandwidth.3.6.3 Change in Routine Medical Expenditures at age 19The preceding section concludes that there is no significant change in office-based physicianvisits, but there is a significant and sizeable decline in dental visits upon losing insurance atage 19. This section examines how out-of-pocket and total expenditures in a given month areaffected. Table 3.9 shows the RD reduced form estimates for office and dental expenditures.I examine the unconditional average effect using the full sample (“All”) as well as the averageeffect among those who report having a visit in a given month (“If Visit”). Given there islikely to be heterogeneity in medical care use across individuals, the conditional average effectmay be more informative in how heavier users of medical care are affected by insurance loss.Young adults’ OOP expenditures on office visits are modestly affected by insurance lossat age 19, with an estimated increase of $1.36 per month (IK bandwidth) and this estimate is(marginally) significant. The associated Wald statistic corresponds to an increase of approx-imately $28 per month. Among those who report having an office visit, there is an averageincrease of $14.59.107 There is also no change in total expenditures on office visits in all spec-107I do not cite the Wald estimates here since it is not clear which scaling factor to use given that only those733.6. Resultsifications but half the IK bandwidth, where there is a very small increase. Figure 3.3 presentsthese results in graphical form, with panel (a) showing a very slight increase in average OOPmonthly expenditures, panel (b) providing a very clear increase in OOP among those with avisit, and panels (c) and (d) showing no change in total expenditures. The results point to avery modest mean effect on OOP expenditures for office visits upon turning 19 years.Figure 3.3: Expenditures on Office Visits1357Dollars per Month (in 2000 $) 17 18 19 20 21Age(a): OOP All 2030405060Dollars per Month (in 2000 $) 17 18 19 20 21Age(b): OOP If Visit51015 20 2530Dollars per Month (in 2000 $) 17 18 19 20 21Age(c): Total All 120140160180200Dollars per Month (in 2000 $) 17 18 19 20 21Age(d): Total If VisitNotes: This figure examines real out-of-pocket (OOP) and totalexpenditures (in 2000 dollars) on office visits in a calendar month. Panels(a) and (c) include all individuals, while panels (b) and (d) include onlythose who have a visit. The circles plot local averages at each age month,while the solid lines are fitted values from local linear regressions of thedependent variable using a triangular kernel and the IK bandwidth.with a visit are being considered.743.6.ResultsTable 3.9: Change in Expenditures on Office and Dental Visits at 19Office Expenditures Dental ExpendituresOut-of-Pocket Total Out-of-Pocket TotalSpecification All If Visit All If Visit All If Visit All If VisitMean $4.26 $37.34 $17.49 $153.38 $6.15 $100.41 $13.73 $224.01Nonparametric: IK 1.36* 14.59*** 0.56 5.11 -0.66 4.19 -3.11* -25.79(0.55) (3.05) (0.93) (11.25) (1.02) (18.72) (1.40) (20.84)Observations 68,073 8,185 54,961 9,686 147,079 6,262 140,484 7,107Bandwidth 0.882 0.932 0.732 1.146 1.901 1.391 1.82 1.559Nonparametric: IK / 2 1.56*** 16.48*** 1.90*** 10.76 -0.90 -6.56 -3.74 -28.49(0.40) (0.66) (0.51) (11.67) (1.45) (23.71) (2.03) (26.93)Observations 35,588 3,810 29,131 4,521 74,619 3,087 68,073 3,441Bandwidth 0.441 0.466 0.366 0.573 0.951 0.695 0.91 0.780Nonparametric: 1 yr 1.22* 13.73*** 0.47 13.23 -0.86 4.22 -3.79* -28.05(0.57) (3.29) (1.27) (10.58) (1.42) (21.54) (1.93) (25.39)Observations 74,619 8,185 74,619 8,185 74,619 4,178 74,619 4,178Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000Parametric: Cubic Spline 2.00* 19.94*** 1.99 23.32 -1.48 -0.17 -4.65* -29.72(0.81) (5.41) (1.96) (14.99) (1.69) (26.53) (2.33) (30.75)Observations 160,243 18,264 160,243 18,264 160,243 9,817 160,243 9,817Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clustered byindividual and age (in months). All nonparametric coefficients are derived using local linear regression. All variables respond to medicalexpenditures in a given month. The sample consists of individuals above 125% of the Federal Poverty Line. “All” includes all individuals inthe sample, while “If Visit” includes only those with a particular visit. Data come from the Medical Expenditure Panel, years 1997-2006. *significant at 10%; ** significant at 5%; *** significant at 1%.753.6. ResultsThe second half of Table 3.9 shows the average impact of losing insurance on dentalexpenditures in a month. The results show no change in OOP dental expenditures. Thereis, however, a marginally significant decline in average total dental expenditures of $3.11per month (IK bandwidth), which corresponds to a Wald statistic of $46 per month, likelycapturing fewer visits amongst the uninsured. Additionally, there is a sizeable decrease in totalexpenditures amongst those who report a visit, suggesting that individuals without insuranceobtain less care when they do see the dentist; however, these estimates are imprecise.While informative, the average effects may be a misleading characterization of how mostindividuals are affected by insurance loss. In particular, losing insurance at age 19 may havemore severe consequences for those at the top end of the spending distribution, comparedto those with little expenditures. Table 3.10 shows considerable heterogeneity in OOP andtotal expenditures across individuals, reporting the 10th, 25th, the median, the 75th, and the90th percentiles amongst those who report an office visit.108 For both types of expenditures,it is apparent that the data are heavily right skewed: the mean is greater than the 75thpercentile, signifying that those at the upper end of the distribution are driving up the meanconsiderably. Additionally, the maximum expenditure values are quite high. These generalpatterns suggest it may be important to go beyond the mean estimates shown in Table 3.9.To obtain a richer description of how young adults’ medical spending risk is impacted byinsurance coverage, I explore heterogeneity in office expenditures using quantile estimation. Iestimate the following for each threshold , where τ is some real number such that τ∈(0,1):Qτj = θτ0 + θτ1Tj +mτ (Aj) + wj (3.4)Here, Qτj is the medical expenditure quantile at threshold τ for age group j, mτ (Aj) isa threshold-specific smooth function of age, and wj is the error term. The RD coefficientsof interest are θτ1 , which is the conditional quantile treatment effects (QTE) at threshold τ .Equation (3.4) was estimated using the standard Koenker and Bassett estimator.109 Thestandard errors on the RD coefficients are computed using the empirical standard deviationof 199 bootstrap repetitions of the quantile treatment estimates. In the analysis, I includeonly those individuals with an office visit who are within the optimal IK bandwidth calculatedfor the average effect, shown in Table 3.9, in order to facilitate comparison between the meaneffect and the QTEs. An important caveat in any quantile analysis is the treatment effectfor a particular individual cannot be identified without invoking additional assumptions suchas rank preservation (Heckman et al. 1997). Given that the rank preservation assumption isquite strong, this study does not attempt to identify treatment effects at the individual level,and instead, focuses on the distributional effects.108Given that OOP expenditures were affected only for office visits and that the total dental expenditureswere more imprecise, I focus only on office based provider visits for the RD quantile analysis.109This was done using the IVQTE command in Stata, developed by Froelich and Melly (2010).763.6. ResultsTable 3.10: Quantile Treatment Effects for Office Expenditures at 19OOP Expenditures Total ExpendituresSpecification Expenditures QTE Expenditures QTE10th Percentile 0.00 0.00 15.72 0.755(0.00) (3.47)25th Percentile 0.00 0.00 40.23 0.560(0.00) (2.12)50th Percentile 10.52 2.68** 70.04 -2.25(1.25) (3.27)75th Percentile 30.78 7.74* 141.10 -11.98(4.18) (8.32)90th Percentile 81.11 -3.93 292.95 13.74(8.60) (27.57)Mean 37.34 14.59*** 153.38 5.11(155.49) (3.05) (396.19) (11.25)Min 0 - 0 -Max 8,459.11 - 15,238.24 -Observations 18,268 8,185 18,268 9,686Mean (≤95th Percentile) 19.92 1.70 143.81 -5.09*(27.16) (1.61) (358.08) (2.80)Notes: The Quantile RD coefficients at age 19 are reported, without controls. Standard errors arein parentheses and were two-way clustered by individual and age (in months). All values are in 2000US Dollars. Quantile treatment effects were estimated using the IK optimal bandwidth calculatedfor the average effect, as shown in Table 3.9. All variables respond to medical expenditures ina given month. The sample consists of individuals above 125% of the Federal Poverty Line whoreport having an office visit. Data come from the Medical Expenditure Panel, years 1997-2006. *significant at 10%; ** significant at 5%; *** significant at 1%.773.6. ResultsFigure 3.4: QTE for OOP Expenditures on Office Visits-50050100 150Dollars 1 10 20 30 40 50 60 70 80 90 99OOP Expenditure Percentile, Prior to 19 yearsNotes: The dependent variable is real out-of-pocket expenditures (in 2000dollars) on office-based medical visits in a given calendar month, amongthose who report to have a visit. The graph figure the estimated quantiletreatment effect (QTE) at various thresholds. The dashed lines are 95%confidence intervals. The estimating sample was derived using the IKbandwidth for the RD of average out-of-pocket expenditures.Figures 3.4 and 3.5 show the estimates (solid line) at each percentile along with the con-fidence intervals (dashed lines), while Table 3.10 provides the effects at selected thresholds.The results show that the average effect on OOP expenditures for office visits seems to overes-timate the impact for the vast majority of individuals. In particular, an estimated increase ofjust under $2.68 is estimated for the median OOP office expenditure, while $7.74 is estimatedfor the 75th percentile. These estimates are both significantly different than zero.110 Giventhat the average effect ($14.59) is considerably higher than these estimates, this suggests thatthose at the very top of the distribution are contributing heavily to the size of the averageeffects. Figure 3.4 shows that large, positive effects are found above the 95th percentile; al-though, they have a high variance. To obtain more representative estimates for the averageeffect, I drop those individuals above the 95th percentile and re-estimate the mean effect. Thelast row in Table 3.10 shows no average change in expenditures for those individuals belowthe 95th percentile, supporting that those at the very top end of the distribution were indeeddriving up the average effect for the full sample. In terms of total expenditures on office visits,110For reasons discussed above, I do not report the Wald statistics for this analysis since it only considersthose with an office visit.783.6. ResultsFigure 3.5: QTE for Total Expenditures on Office Visits-400-2000 200400Dollars 1 10 20 30 40 50 60 70 80 90 99Total Expenditure Percentile, Prior to 19 yearsNotes: The dependent variable is total expenditures (in 2000 dollars) onoffice-based medical visits in a given calendar month, among those whoreport to have a visit. The figure shows the estimated quantile treatmenteffect (QTE) at various thresholds. The dashed lines are 95% confidenceintervals. The estimating sample was derived using the IK bandwidth forthe RD of average total expenditures.a tight zero effect is found across most of the distribution as shown in Figure 3.5, which isconsistent with the average effect found in Table 3.9.3.6.4 Change in Ability to Afford Medical Care at age 19I also examine the impact of insurance status on the difficulty of young adults being ableto afford medical care. These results are shown in the first column of Table 3.11. There isno change in individuals reporting to have problems obtaining necessary care once they turn19 years, and this holds true across all specifications. Panel (a) of Figure 3.6 shows a lot ofvariation across cell means in terms of the raw data, but there is no clear change in the abilityto afford care at the threshold of 19 years. These results suggest that even in the absence ofinsurance coverage, the average young adult is still able to afford necessary care.793.6.ResultsTable 3.11: Change in Ability to Afford Care and Health Status at 19Specification Can’t Afford Care Health Index Very Good Health Excellent Health Missed School Missed WorkMean 0.0116 0.7846 0.7552 0.4265 0.1966 0.1914Nonparametric: IK -0.0076 -0.0016 0.0081 -0.0098 -0.0147 0.0227(0.0053) (0.0057) (0.0083) (0.0146) (0.0139) (0.0157)Observations 13,070 15,263 19,465 19,465 8,628 7,661Bandwidth 1.077 1.232 1.526 1.532 1.214 1.305Nonparametric: IK / 2 -0.0032 -0.0019 -0.0003 -0.0155 -0.0298* 0.0002(0.0064) (0.0076) (0.0105) (0.0155) (0.0151) (0.0196)Observations 6,736 7,824 9,935 9,935 4,197 3,684Bandwidth 0.538 0.616 0.763 0.766 0.607 0.652Nonparametric: 1 yr -0.0081 -0.0023 0.0035 -0.0132 -0.0182 0.0170(0.0055) (0.0059) (0.0088) (0.0151) (0.0144) (0.0181)Observations 11,983 12,070 12,070 12,070 6,661 5,615Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Parametric: Cubic Spline -0.0078 -0.0070 -0.0100 -0.0169 -0.0193 0.0103(0.0068) (0.0087) (0.0170) (0.0182) (0.0181) (0.0248)Observations 25,639 25,810 25,810 25,810 15,341 11,848Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clustered by individual and age (inmonths). All nonparametric coefficients are derived using local linear regression. The sample consists of individuals above 125% of the Federal Poverty Line. Datacome from the Medical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.803.6. Results3.6.5 Change in Health Status at age 19The final set of outcomes I examine are measures of general health status, particularly per-ceived physical health and days missed work or school. The reduced form RD estimates areshown in Table 3.11. The estimates for the health index, being in very good health, and beingin excellent health are close to zero across all specifications.111 The case of being in at leastvery good health is shown in Panel (b) of Figure 3.6. Table 3.11 also shows that insuranceloss at age 19 does not change the probability of missing any school or work school due toillness. The estimates are small and negative in the case of missing school. However, theyare only marginally significant in the case of half the IK bandwidth. Panel (c) in Figure 3.6shows no noticeable change in missing school at age 19, while panel (d) does the same formissing work due to illness. One caveat to this analysis is that only the immediate impactof insurance loss is being examined, given the nature of the RD design employed. Althoughno effects on health are found, there may still be long term effects of not having insurance,particularly since health is a stock and not a flow.112111I can rule out declines of more than 0.82 percentage points for being in at least very good health at the95 percent confidence level, which is quite small relative to the mean.112As discussed previously, there may be additional spillovers for young adults in obtaining routine care, suchas forming ties with primary care physicians, taking responsibility for their own care, and habit formation forpreventative screenings. These will not necessarily result in immediate health benefits for young adults, butare likely to occur in the long run as they age.813.7. Robustness ChecksFigure 3.6: Indicators of Health Status0 .01.02 .03.04Proportion 17 18 19 20 21Age(a): Problem Affording Care .7.72.74.76.78.8Proportion 17 18 19 20 21Age(b): Very Good Health.1.15.2 .25.3Proportion 17 18 19 20 21Age(c): Miss School .1.15.2.25.3Proportion 17 18 19 20 21Age(d): Miss WorkNotes: This figure examines indicators of health, with panel (a) showingproblem affording care, (b) being in at least very good health, (c) missingschool due to illness, and (d) missing work due to illness. The circles plotlocal averages at each age month, while the solid lines are fitted values fromlocal linear regressions of the dependent variable using a triangular kerneland the IK bandwidth.3.7 Robustness ChecksThere are two conditions under which the fuzzy RD gives an unbiased estimate of the LATE:monotonicity and excludability. Monotonicity rules out that some uninsured individuals takeup insurance on their 19th birthday. Excludability implies that turning 19 years old canonly impact medical outcomes through the loss of health insurance. The latter conditionamounts to assuming that E[vij |Aij = a] is continuous at a = 19 (see Equation (3.1)) andrules out other factors correlated with health outcomes to change discontinuously on the 19thbirthday. This assumption could be violated if say, living arrangements, school attendance,or health lifestyle behavior changed discontinuously at 19 years old. However, given that ageis measured in months, as opposed to years, it is unlikely that these factors would changediscontinuously within one or two months of turning 19 years old. As noted by ADG, the mostobvious confounder might be high school graduation and the ensuing transition to college oremployment. Given that graduation typically occurs at the end of June in a year and birthdaysare distributed throughout the year, these factors should not bias the estimates.As a robustness check, I examine whether certain covariates change discontinuously at 19years. I estimate Equation (3.3) with possible confounders as dependent variables. First, I823.7. Robustness Checksexamine the incidence of being a student, leaving home, and being married. As discussedabove, these factors affect individuals’ eligibility for insurance coverage and may affect healthoutcomes. No discrete change at age 19 was found for these variables as shown in Table3.12.113Additionally, I test whether any discrete changes in the MEPS survey structure occuronce individuals turn 19 years which may be confounding the estimates. In particular, I firstanalyze whether the composition of people in the sample changes at the age 19 cutoff.114One concern might be that survey attrition jumps at the threshold, whether due to familycomposition changes or young adults leaving home, and the estimates are picking up changesin the sample composition. As such, I test if there is a jump in the number of young adultsin the sample at age 19. I find little evidence this is the case with no change found across allspecifications, with the exception of half the IK bandwidth which has a very small, marginaleffect. Next, I test if the incidence of being the reference person at the interview roundchanges discontinuously at 19 years as changes in the family structure could be confoundingthe estimates.115 The IK optimal bandwidth and the bandwidth of 1 year estimate a precisezero effect for being the reference person. This suggests that changes in the family structureare not confounding the true estimates. Overall, my findings provide strong support for thevalidity of the RD design. Given that observed individual and survey characteristics are notchanging discontinuously at age 19, this reduces the concern that unobservable confoundersmight be changing at the cutoff and confounding the estimates.An additional robustness check I perform examines whether there are any anticipationeffects, such as individuals “stocking up” on medical care services prior to turning 19. If youngadults expect to lose insurance in the near future, it would be rational for them to get theirmedical needs attended prior to the loss of coverage. In this scenario, the estimates presentedabove would be upward biased. This is particularly relevant in the case of dental care, whichis a more discretionary form of spending and where a significant decline in consumption wasfound upon turning 19 years. If, however, young adults are myopic in their demand for healthcare, then there would be no anticipatory consumption and the estimates above remain valid.I exploit the panel dimension of the data set to examine whether individuals differentiallyconsume medical care in months closer to the time they lose their coverage. For each indi-vidual, I construct three dummy variables denoting whether the individual loses insuranceone month (L1), two months (L2), and three months (L3) in the future. For individuals whonever lose coverage while in the survey, the value of all three variables is each zero. I follow113The graphical representation of these results is shown in Figure B.3 in the Appendix.114MEPs collects information on all those in a Reporting Unit, which is defined as a person or a group ofindividuals who are related by family association. If individuals leave the original Reporting Unit, MEPs triesto follow them and keep them in the survey.115The reference person is defined as the household member in the Reporting Unit who owns or rents thehome. In student households, the student is the reference person. The reference person could change if thefamily dissolves or if there is significant changes in the family composition.833.7. Robustness ChecksTable 3.12: Robustness Checks for Other Variables Changing at 19Specification Student Married Moved Out No. of Individuals Reference PersonMean 0.5828 0.0368 0.2608 718.9683 0.1294Nonparametric: IK 0.0120 -0.0043 -0.0117 28.4106 0.0004(0.0119) (0.0055) (0.0086) (20.6328) (0.0064)Observations 21,997 14,815 16,294 9,167 17,726Bandwidth 1.270 0.909 0.922 0.559 1.061Nonparametric: IK / 2 0.0088 -0.0101 -0.0041 19.3246* 0.0028(0.0158) (0.0058) (0.0149) (9.7899) (0.0081)Observations 10,543 7,747 7,750 5,012 9,167Bandwidth 0.635 0.455 0.461 0.279 0.531Nonparametric: 1 yr 0.0131 -0.0045 -0.0116 28.3441 -0.0002(0.0127) (0.0054) (0.0081) (21.5801) (0.0067)Observations 16,252 16,287 16,294 16,294 16,294Bandwidth 1.000 1.000 1.000 1.000 1.000Parametric: Cubic Spline 0.0058 -0.0085 -0.0258 33.9253 -0.0046(0.0168) (0.0054) (0.0139) (28.5712) (0.0090)Observations 34,527 34,985 34,998 34,998 34,998Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-wayclustered by individual and age (in months). All nonparametric coefficients are derived using local linear regression. Thesample consists of individuals above 125% of the Federal Poverty Line. Data come from the Medical Expenditure Panel, years1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.an approach similar to Gross (2010) who uses a difference-in-difference estimator, comparingmedical care consumption in a given month across individuals who lose insurance coverage atdifferent periods (or not at all). The estimating equation is given by:Yimt = δ0 + δ1L1imt + γ2L2imt + γ3L3imt + φXimt + Li + µi +m+ t+mt+ ωimt (3.5)The outcomes examined Yimt are office and dental visits within the month, as well as fillingout new prescription drugs in the month. Fixed effects for month (m) and year (t) are included,as are the interaction of month and year (mt), in order to control for any differences affectingmedical care consumption across time common to all individuals i. Additionally, a dummyfor whether the individual loses insurance (Li) and 11 dummy variables indicating the monthin of insurance loss (µi) and are included to take into account any fixed differences betweenindividuals who lose coverage at different times of the year and who do not lose coverage at all.Individual characteristics are captured in the vector Ximt, including age, gender, student andmarital status, and log of real family income. In this model, the identifying assumption is thatthe timing of insurance loss is exogenous such that there are no unobserved factors correlated843.7. Robustness Checkswith both the timing of insurance loss and medical consumption. While there is no way toverify this, the goal is that, having controlled for other factors affecting health outcomes,the estimates will shed light on whether there is an indication of anticipatory consumption.Positive and significant signs on coefficients, γ1, γ2, and γ3 would suggest that young adultsare forward looking and are stocking up on their medical consumption.853.7.RobustnessChecksTable 3.13: DID Estimates Examining Anticipatory Consumption Prior to Coverage LossOffice VisitsExplanatory Variables All Diagnosis or Treatment Checkup Mental Counseling New Prescriptions Dental VisitsLose in One Month -0.0098 -0.0110 0.0023 -0.0023 -0.0008 -0.0007(0.0108) (0.0076) (0.0052) (0.0040) (0.0060) (0.0074)Lose in Two Months 0.0147 0.0024 0.0036 0.0031 0.0102* 0.0079(0.0097) (0.0085) (0.0046) (0.0037) (0.0051) (0.0081)Lose in Three Months -0.0154 -0.0055 -0.0009 -0.0024 0.0011 -0.0093(0.0125) (0.0085) (0.0044) (0.0042) (0.0070) (0.0074)Observations 104,107 104,107 104,107 104,107 104,107 104,105Notes: DID estimates are presented above on the impact of medical care consumption prior to insurance loss. Standard errors are in parentheses and weretwo-way clustered by individual and age (in months). The sample includes all those who are observed as having insurance at one point throughout the paneland are at 125% of the FPL. Covariates include dummies for male, white, msa, full-time student, married, never leave home, as well as log of family income.Additionally, calendar month and year fixed effects are included, as are the interaction of month and year. Finally, all estimates include dummies for month ofinsurance loss and whether an individual ever loses insurance are included. Data come from the Medical Expenditure Panel, years 1997-2006. * significant at10%; ** significant at 5%; *** significant at 1%.863.7. Robustness ChecksTable 3.13 shows the resulting estimates from this analysis. This table shows that nopositive, significant effects can be found for office and dental visits in regards to losing insur-ance one month, two months, and three months prior to insurance loss. The evidence alsoshows that individuals do not stock up on specific types of office care, such as diagnosis/treat-ment, checkups, or mental counseling. In the case of new prescription drug use, there is nosignificant effect for one and three months prior to losing coverage, where the estimates aretight zeros. While there is a positive effect found for two months prior to insurance loss, theeffect is only marginally significant at the 10% level. Additionally, given that no significanteffect was found in terms of new prescription drug use at age 19, anticipatory consumption isrelatively less of a concern in this case. The finding that young adults generally do not stockup on medical care and are myopic in their consumption confirms the results of Gross (2010)who reaches this conclusion using another dataset on teenage medical care consumption.116While these results cannot directly rule out anticipatory consumption prior to age 19, theyare nevertheless reassuring to support the validity of the RD research design.Another concern might be a misreporting of insurance coverage, particularly around theage 19 threshold. This could potentially bias the RD estimates of insurance changes above.Furthermore, a possible scenario might be that individuals who are not aware they have justlost coverage at age 19 consume care they otherwise wouldn’t. This would result in downwardbiases in the second order estimates above. To explore this possibility, I compare individuals’reported insurance coverage with the source of payment information in the office and dentalprovider records.117 As outlined above, the provider payment information notes the amountpaid by the individual, by private insurance providers, and by public insurance plans. Inthe analysis, I only include those who report an office or dental visit. First, I construct adummy variable indicating whether the individual reports to have insurance, yet there is nopayment made by an insurance provider for the visit. Insurance plans vary, and not all medicalprocedures are covered by plans; therefore this indicator does not necessarily represent theproportion of individuals who are misreporting insurance coverage. In fact, roughly 45 percentof all visits by individuals reporting to be covered have no payment made by an insuranceprovider.118 The goal here, however, is to examine whether there is a discontinuous change atage 19 in the proportion of individuals who have a value of 1 for this indicator, which could besuggestive evidence of an increase in misreporting around the threshold. I use the same RDestimation methods as earlier, using local linear regression with the IK optimal bandwidth.I also examine whether this indicator varies by those who report having private insurance116Gross (2010) performs multiple robustness checks that support his main set of results, suggesting that thelack of anticipatory consumption may be partly explained by young adults being uncertain as to exactly whenthey lose their coverage.117As a reminder, individuals are asked whether they had insurance on any day in the calendar month, notwhether they had it the entire calendar month.118Most of this is driven by dental visits, where over 80 percent of those reporting to be covered have noinsurance payment made to the provider, compared to 20 percent in the case of office-based physician visits.873.7. Robustness Checks(47 percent have no insurance payment for the visit) and public insurance (35 percent haveno insurance payment for the visit). I also construct a dummy variable for those who reportnot having insurance at any time in the month, yet where there is a payment made by aninsurance provider on the visit. This amounts to just under 30 percent of all those with avisit who report not having any insurance coverage.The RD estimates are presented in Table 3.14, showing that there is no significant changein the proportion of individuals who report having any insurance with no payment madeto the medical provider by an insurance plan, and this holds true when private and publicinsurance are considered separately. The last column of Table 3.14 shows that no significantchange can be detected in the proportion of those who report not having insurance but whohave a payment made to the provider by an insurance plan. These results suggest that thereis no discontinuous change in the misreporting of insurance coverage status upon turning 19years, providing further validity to the RD design used in this study.119119One caveat to this analysis is that the MEPS collects insurance coverage retrospectively for each monthat the round interviews, so there is a gap between the time of reporting and the month in which the visit tookplace. As such, it may still be the case that individuals do not know they have lost coverage, yet consume carethey wouldn’t otherwise. Given the nature of the MEPS, however, this issue cannot be further explored.883.7.RobustnessChecksTable 3.14: Robustness Checks for Misreported Insurance Coverage at 19Report Any Coverage, Report Private Coverage, Report Public Coverage, Report No Coverage,No Insurance Payment No Insurance Payment No Insurance Payment Have Insurance PaymentSpecificationMean 0.4542 0.4676 0.3481 0.2906Nonparametric: IK -0.0103 -0.0049 -0.0159 0.0227(0.0178) (0.0182) (0.0357) (0.0353)Observations 9,682 10,201 2,316 3,655Bandwidth 0.986 1.124 1.468 2.072Nonparametric: IK / 2 -0.0144 -0.0087 -0.0053 0.0448(0.0196) (0.0192) (0.0211) (0.0371)Observations 4,492 4,687 1,057 1,926Bandwidth 0.493 0.562 0.734 1.036Nonparametric: 1 yr -0.01 -0.0101 -0.0041 0.0439(0.0178) (0.0185) (0.0274) (0.0371)Observations 9,682 8,546 1,466 1,793Bandwidth 1.000 1.000 1.000 1.000Parametric: Cubic Spline -0.0329 -0.0419 0.0211 0.0749(0.0280) (0.0286) (0.0451) (0.0487)Observations 22,334 19,639 3,421 3,655Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-way clustered by individual and age (inmonths). All nonparametric coefficients are derived using local linear regression. Variables are as described in text. The sample consists of all individuals above125% of the Federal Poverty Line who report having an office or dental visit. Data come from the Medical Expenditure Panel, years 1997-2006. * significant at10%; ** significant at 5%; *** significant at 1%.893.8. Concluding Remarks3.8 Concluding RemarksThis paper finds that the 19th birthday played a significant role in health insurance coveragerates in the US over the last decade. The estimated reduction in insurance coverage at age19 is 3 to 5 percentage points, driven by a loss of dependent coverage (3 to 4 percentagepoints). The change in public insurance is moderate at less than 1 percentage point. The lossof insurance is strongest for non-students and those living at home. Additionally, those withsevere chronic conditions are amongst the least affected by private insurance loss at age 19,which is a noteworthy result given they are one of groups targeted in recent federal healthpolicies. Interestingly, even young adults from families who are most able to afford insuranceexperience a significant and sizeable decline in private insurance, suggesting that it may be aconscious decision on their part to forgo insurance.This study finds no effect of insurance loss on office-based physician visits and new pre-scription drug use. Given that those most affected by insurance loss at age 19 (i.e. thecompliers) are higher income and amongst the healthiest, these results are not surprising.There is an average increase of $28 per month in out-of-pocket office expenditures amongstthose who lose insurance. However, most of the increase in spending occurs at the very top endof the distribution. I do find a significant decline in dental visits on the order of 21 percentagepoint amongst those who lose insurance, which results in lower total dental expenditures (adecline of $46 per month on average amongst the uninsured). The drop in dental visits butnot other forms of non-emergency care may be due to the perceived discretionary nature ofdental spending. Importantly, there is no change in young adults reporting to have a problemaffording necessary medical care or in their health status. I also find no evidence that youngadults stock up on medical care prior to insurance loss, suggesting they are myopic in theirmedical care consumption.These findings show that, interestingly, young adults are not making up the lost hospitaland emergency department care observed by Anderson, Dobkin, and Gross (2012) in othermedical settings. At the same time, the type of care that is forgone in hospitals very well couldbe sensitive care that was not needed in the first place. Overall, the findings in this papershow that the vast majority of young adults are not heavily impacted by health insurance interms of their routine medical care consumption, expenditures, and short-run health. One ofthe primary benefits of health insurance comes from minimizing the risk associated with anegative health shock, allowing individuals to smooth income over time. As discussed, thereare alternative means that may also provide this benefit and which may be less costly forsome young adults. My findings suggest that the gains of increased insurance coverage mayprimarily be external, rather than to most young adults themselves.90Chapter 4Beyond the Mean: An Examinationof Heterogenous Child Responses toa Universal Childcare Policy inQuebec4.1 IntroductionWith rising rates of maternal employment, there has been a parallel growth in the demandfor accessible and affordable childcare across developed nations. Under mounting pressure tomeet these demands, many governments are adopting the explicit goal of expanding childcarecoverage, particularly to families with young children. A central debate in Canada that hascontinued unabated over the past two decades is a plan for a national childcare program. Inrecent federal elections, in particular, there has been much discussion on expanding the numberof government regulated childcare spaces and providing childcare subsidies to a broader rangeof the population. Although childcare subsidies are not a recent phenomenon, largely havingbeen targeted at low income families in the past, policy makers, the public, and researchersalike are all increasingly directing their attention to the role of publicly funded universalchildcare subsidies in improving childcare coverage.While the costs and merits of universal childcare subsidies have been the source of manyheated debates in the political arena, unfortunately very little research to date has actuallybeen carried out on their impact on child health and developmental outcomes. With a grow-ing body of evidence finding that the early childhood environment plays a key influential rolein long run health and skill formation and that inequalities are set early in life (e.g. Greggand Machin, 2000; Cuhna and Heckman, 2007), proponents of universal childcare argue thatthe policy may assist in equalizing skills across children. They argue it would benefit dis-advantaged children in particular through the provision of an enriched environment outsideof the home. It is consequently important that evaluations of universal childcare programsare able to take into account differential responses to the policy. The small, but quicklyemerging, literature examining the effect of universal childcare policies on child health and914.1. Introductiondevelopment outcomes finds mixed results on the average outcomes across children (Baker,Gruber, and Milligan 2008; Havnes and Mogstad 2009; Datta Gupta and Simonsen 2010).However, the bulk of the existing research examines the mean impacts of the policy, largelyignoring heterogenous responses, and consequently does not directly assess whether one ofthe main justifications given for universal childcare programs, namely that they help equalizeskills across children, has any empirical support.This study aims to fill a gap in the existing literature by using a quantile difference-in-differences (QDID) model to identify heterogeneous responses to a universal childcare pro-gram in Quebec, Canada in terms of child health and developmental outcomes. In 1997, thegovernment of Quebec introduced universal subsidies for childcare, where families from alleducational and economic backgrounds became eligible for the heavily subsidized spaces of$5 a day. Along with vast reductions in parental fees, the policy also included an expansionin the number of regulated childcare spaces and stiffer requirements for childcare providersto obtain government subsidies. It is, in fact, this childcare model that many politicians atthe federal level have discussed adopting for the rest of Canada. Given this policy cost theQuebec government millions of tax dollars to implement and would likewise cost the Canadiangovernment billions more to adopt at the national level, it is crucial that the merits of theprogram are properly evaluated, including its impact on child health and development.This research extends the work of Baker, Gruber, and Milligan (2008) who use a standarddifference-in-differences (DID) estimator to find that the Quebec policy led to worse outcomes,on average, for young children in terms of problematic behavior, health, and motor skills.These results are interesting in light that household resources were effectively raised by thepolicy, with increased maternal labour supply and cheaper childcare, and that the policycreated a large shift from informal care to care in registered centres, which are found to beof higher quality than other forms of care in Quebec (Japel et al. 2005). While the Baker,Gruber, and Milligan (BGM herein) study provides one of the first evaluations of a universalchildcare program, there is limited understanding on exactly what mechanisms generatedthe negative mean estimates of the Quebec policy. Given universal childcare programs arequite expensive to implement and are receiving increased public attention in many developedcountries besides just Canada, the findings of BGM beckon more research to be done in thearea so as to develop a clearer picture on precisely how universal childcare programs affectchildren’s health and development.In this paper I examine the existence of heterogenous responses by children to the universalchildcare program and gain a deeper understanding of the effects of the Quebec Policy onmotor skills, cognitive development, and child health. The existence of differential responsesto the childcare program may shed light on important mechanisms at play. Additionally,given that one of the most common goals and justifications for universal childcare policies isthat they level the playing field across children, it is important to go beyond the mean. As924.1. IntroductionHeckman, Smith, and Clements (1997) discuss, knowledge on the distributional impacts isoften critical for evaluation as mean impact estimates cannot always provide the necessaryinformation to compute the true gains of a program because they can mask large variations inindividual responses. A case in point is the widely cited work of Bitler, Gelbach, and Hoynes(2006) who analyze the distributional characteristics of a welfare reform in the U.S. to findthat there was substantial heterogeneity in the response to the policy change and that theirkey empirical findings could not otherwise be revealed by simply performing mean impactanalysis. Thus, in some circumstances, it is necessary to go beyond the mean as averageimpacts can miss a great deal.This paper makes three broad contributions to the literature. First, it analyzes the extentto which universal childcare differentially affects the distribution of child health and devel-opment. While existing studies largely focus on the average effects of childcare and earlyeducation policies, I analyze the extent to which these policies level the playing field for chil-dren by examining the entire distribution of outcomes. Second, this study tests the impact ofthe program on children’s body weight, an outcome that has received little attention in thechildcare literature to date. If universal care altered the mix of activities in which childrenwere engaged or impacted their diet, then child weight could be affected. With growing ratesof child obesity, a deeper understanding of the factors that shape this trend is of particularpolicy interest and could help identify potential solutions. Third, my study sheds light on theextent to which the roll out of large scale social programs may influence their impact on thedistribution of outcomes by contrasting the short run impacts to the longer run effects.This paper goes beyond the mean impact estimates of the BGM paper by using a quantiledifference-in-differences (QDID) estimator to evaluate the effect of the Quebec universal child-care reform on the entire marginal distribution of motor skills, cognitive, and health outcomesfor children in Quebec. To identify the effect of policy, the QDID approach uses the entirepre and post-policy distributions of child health and developmental outcomes in the otherCanadian provinces to estimate a “counterfactual” distribution of development outcomes inQuebec that would have existed in the post-reform period in the absence of the childcare pol-icy. The method used in this paper allows me to test whether the impact of reform is constantacross the distribution of child outcomes, or whether the reform led to larger changes at someparts of the distribution. The data used in this study come from the National LongitudinalSurvey of Children and Youth (NLSCY), which is a large, nationally representative Canadiansurvey that collects detailed information on children’s development and environment frombirth. It is the same dataset used by BGM.The findings in this study suggest that overall, there is little heterogeneity in the responseto Quebec’s universal childcare policy, at least in terms of motor skills and cognitive outcomesin the short run. In fact, this study finds that the policy had little significant effect on theseoutcomes at any point along the distributions, neither when the full sample of children is used934.2. Previous Researchnor when the sample is split by child demographic characteristics. I do find evidence thatthe childcare policy led to a reduction in the body weight of children in the upper part of thedistribution. These results are robust to different specifications and estimation techniques.This study is complementary to recent work by Kottelenberg and Lehrer (2014) whoindependently focus on the long run distributional impacts of the Quebec childcare program.These authors analyze the impact of the policy on children from both single and two-parenthouseholds on a subset of outcomes I also test. They examine a longer time period after thepolicy was introduced and find evidence of improvements in the lower end of the distribution ofcognitive skills among single-parent children but a decline in the lower end for children in two-parent households. They suggest that changes in the home environment play a role in shapingthese outcomes. The differences between our findings suggest that the roll out of the universalchildcare mattered. In particular, it may take considerable time to set up high quality childcare centres, to find and train appropriate staff, and to develop a curriculum. My paperdeepens our understanding of the program and, more broadly, may inform our understandingof how social programs affect participant outcomes differentially over the program’s life.The remainder of the paper is as follows: Section 2 presents previous research on theimpact of childcare and early education programs on child development, as well as a briefoverview of the literature on non-maternal care more generally. Section 3 then describes theQuebec universal childcare policy in more detail. The following section describes the NLSCYand the primary variables of interest, while Section 5 presents the main empirical methodologyused in the analysis (the QDID estimator), as well as descriptive statistics. Section 6 presentsthe findings of the study, with Section 7 providing an interpretation of the results. Concludingremarks and directions for future research are given in Section 8.4.2 Previous ResearchThere is a small but quickly growing literature on the effects of universal childcare policies onchild outcomes, most of which focus on mean impacts. From this small collection of research,there is no real consensus on the merits of universal care, with the studies finding mixedresults. Using a difference-in-differences estimator, Havnes and Mogstad (2009) find that theintroduction of universal care in Norway led to strong, positive long run outcomes in termsof higher educational attainment, greater labour market participation, and a reduction inwelfare dependency. Conversely, Datta Gupta and Simonsen (2010) find that the expansionof universal pre-school and family day care in Denmark had no mean effect on child non-cognitive outcomes for children in pre-school, but family day care worsened outcomes forboys of low educated mothers. Again, BGM found that the universal childcare program inQuebec resulted in poorer average child outcomes in terms of aggression, illness, and motordevelopment.944.2. Previous ResearchThe only known studies to date which has explicitly examined the distributional impactsof a universal childcare program is work by Havnes and Mogstad (2010) and the recent studyby Kottelenberg and Lehrer (2014).120 Havnes and Mogstad (2010) examine the effect of alarge scale, heavily subsidized childcare program in Norway on subsequent adult earnings.Using a threshold difference-in-differences model, they find that although the mean impactof the program was insignificant, there were significant, positive effects over most of theearnings distribution. Kottelenberg and Lehrer (2014) analyze the long run effects of theQuebec childcare program on motor skills and cognitive development. They find that universalchildcare boosted test scores for the most disadvantaged children. These studies demonstratehow simply examining mean impacts in the context of childcare may mask much of the policy’simpacts.The effect of targeted (as opposed to universal) childcare subsidies on maternal laboursupply and childcare use has also received some attention, with most studies finding they leadto increased maternal employment and formal childcare use (Meyers et al. 2002; Tekin, 2005;Blau and Tekin, 2008). Although less attention has been given to the effect of these subsidieson child outcomes, a recent study by Herbst and Tekin (2010) finds that childcare subsidiesto low income U.S. families have negative effects on children’s math and reading scores andlead to greater behavior problems, which the authors argue likely arise from parents choosinglow quality childcare.Several studies investigate the mean impact of expansions of universal early educationprograms, with most finding positive effects. For example, Cascio (2009) finds that the ex-pansion of universal kindergarten in the U.S. in the 1960s and 1970s led to lower high schooldrop-out and institutionalization rates for Whites, but had little effects for African Amer-icans. Gormley and Gayer (2005) and Fitzpatrick (2008) examine the impact of universalpre-kindergarten in U.S. states and both find these programs lead to higher child test scores,with the greatest gains accruing to disadvantaged children. The findings of Berlinksi et al.(2009) echo these positive results in the context of a universal pre-primary education programin Argentina.A large body of research examines the effect of non-maternal care more generally onchild outcomes, with the evidence pointing to differential impacts across specific domainsof child development. Many studies find non-maternal care is associated with poorer childsocio-emotional adjustment, in terms of increased rates of at-risk levels of assertiveness, ex-ternalizing behavior problems, and aggression (Bates et al. 1994; Belsky 2001; NICHD 2002,2004, 2006, 2007). Additionally, longer hours of early non-maternal care is linked with lessharmonious parent-child relations and more conflict with adults, as marked by greater lev-els of disobedience and non-compliance (Belsky 2001; NICHD 2003). On the other hand,120Upon completion of the first draft of this paper, the mimeo by Kottelenberg and Lehrer (2014) wasdiscovered.954.2. Previous Researchnon-maternal care has also been associated with more positive, cooperative, and skilled peerplay (Scarr and Eisenberg 1993; NICHD 1998, 2001). In terms of cognitive and motor skills,centre-based care is associated with stronger pre-academic math, reading, memory, and lan-guage skills, while informal care is linked with poorer cognitive outcomes (NICHD 2000, 2002;Bernal and Keane 2010; Hickman 2006). The timing of care also seems to matter for cogni-tive development, with more hours in centre care throughout infancy being associated withlower pre-academic test scores at 4.5 years of age, although more time in centre care duringtoddlerhood is associated with stronger language skills (NICHD 2004).While many of the studies outlined above examine the impact of childcare and early yearsprograms on average child outcomes, only the studies by Havnes and Mogstad (2010) andKottelenberg and Lehrer (2014) explicitly examine distributional impacts and they focus onlong run outcomes. The most common approach taken to investigate heterogenous responsesto early childcare policies is subsample analysis where average effects are allowed to varyacross child demographic and family characteristics. However, as shown in the study byBitler, Gelbach, and Hoynes (2006) for a welfare reform in the U.S., simply performing meanimpact analysis on defined subgroups of the population may not fully reveal the impact of thepolicy; in their study, the intra-group variation in quantile treatment effects greatly exceededthe inter-group variation in mean impacts, and the authors note that simple mean impactanalysis would not have revealed their key findings.My study takes an approach similar to Havnes and Mogstad (2010) by analyzing theprogram’s effect on the entire distribution of child outcomes using a QDID estimator. Kot-telenberg and Lehrer (2014) analyze the long run impact Quebec’s universal childcare policyon the distribution of child development outcomes using Athey and Imben (2006)’s changes-in-change estimator. My study focuses on impacts in the short run and uses an alternativemethodological approach (the QDID). The other key difference between the two studies isthat I focus solely on children from two-parent households. I do this because there were othersocial programs in Quebec and other provinces that were introduced around the same timeas Quebec’s universal childcare program. These policies largely targeted children from singleparent families. As such, I focus on children from two-parent households to ensure that othersocial programs are not confounding the true effects of the universal childcare policy. I alsoperform a much richer analysis to test if distributional impacts vary across specific groups ofchildren, such as by parental education, income, and family functioning. Furthermore, I testif the program affected child body weight. It is unlikely childcare affected children’s height,so my study can help identify factors that may or may not contribute to growing child obesityrates and is certainly of particular policy interest.964.3. The Quebec Policy Change4.3 The Quebec Policy ChangeIn 1997, the province of Quebec experienced a major transformation of its early childhoodcare and education system, known as the new Family Policy initiative. At the heart of thereform was an overhaul of the early childcare setting, an expansion of school-age childcareprograms, and the introduction of full-day kindergarten. This study is concerned with thefirst aspect of the Family Policy, namely the restructuring of the childcare system for youngchildren not yet of school age. Prior to the policy change, the demand for regulated childcarespaces surpassed the number available, leaving the majority of young children in the provincewithout access to monitored care of a known quality. Given that the government providedfinancial exemptions primarily to the poor, middle income families in particular had limitedaccess to care as they often did not have sufficient resources to pay for it (Tougas, 2002).With the goal of fostering child well-being and development through improved educationalchildcare, the government of Quebec undertook a significant restructuring of the childcaresystem in the fall of 1997. Improvements in both the quality of and the access to regulatedcare were central to this initiative, with sweeping reductions in parental fees, an expansion inthe number of regulated childcare spaces, and stiffer requirements for childcare providers toobtain government subsidies.The introduction of reduced rate spaces to families of all economic backgrounds was akey aspect of the new Family Policy. All children aged 0 to 4 became eligible for subsidiesin regulated childcare spaces. Under the new scheme, parents only had to contribute $5 perday per child for a regulated childcare space in the first few years of the program, which wasmodestly increased to $7 a day in early 2004. Under this reduced rate pay scheme, parentswere allowed to leave their children in care a maximum of 10 hours per day and 261 daysper year. For very low income families, fees were waived for up to 23 hours of care a weekand additional compensation of $3 a day was given to those accessing a $5 a day space. Theintroduction of the reduced fees occurred in stages, with reduced rate spaces initially beingmade available to 4-year olds exclusively in September, 1997. These spaces then becameaccessible to 3-year olds in September 1998, 2-year olds in September 1999, and by September2000, all children under 5 years (0-59 months) became eligible for reduced rate care. Giventhat all families were eligible for the reduced rate spaces and access was not tied to parents’employment, educational, or income levels, the reform essentially amounted to a universalregulated childcare system. However, the effective price of childcare varied by family income.Since low income families were already receiving large, targeted subsidies prior to the reform,they experienced little gains. Rather, the largest reductions in childcare prices accrued tomiddle and high income earning families.121121BGM note that prior to 1997, refundable childcare tax credits were provided at a rate that dependedon family income, varying from %75 for those with the lowest incomes to %26 for those with family incomes974.3. The Quebec Policy ChangeAlthough the introduction of the new subsidy scheme was staggered by age, the excessdemand for regulated childcare spaces became exacerbated in the post reform period. Sincethe bulk of regulated care spaces became available at the newly subsidized rate, queues beganto form. The government sought to address this shortfall by expanding government subsidiesto nonprofit, community-based organizations called centres de la petite enfance (CPE). CPE’swere responsible for overseeing regulated care throughout the community in both centre andfamily home settings. In general, the centres served as the organizational nodes of the CPEs,while home based providers throughout the community formed as a network affiliated withthe neighbourhood centre. Typically, children over 2 years of age were placed in centrecare, while home care providers attracted the younger children. These agencies were initiallycreated out of the existing non-profit centres and family home care agencies, but over time,new centres and family home providers were created. The expansion of care in family homes,in particular, became integral to increasing the number of regulated spaces in the province.While the government was successful in more than doubling the number of subsidized spacesfrom approximately 74,000 in early 1997 to over 189,000 in early 2005, growth was relativelyslow in the initial year of the program. As outlined in LeFebvre and Merrigan (2008), growthin subsidized spaces was less than 4% from 1997 to 1998 and most of the available spaces wentto accommodate families who were already using the existing regulated facilities. It wasn’tuntil the second year of the program that the increase in subsidized spaces really took off,growing at over 25%, before tapering off to about 7% growth in 2005.In addition to increasing the number of regulated childcare spaces and reducing parentalfees, the policy also led to significant changes in the centre and home care environments.To obtain government funding, all childcare agencies affiliated with CPE’s became subjectto a range of newly established regulations, including stricter requirements for the physicalenvironment and layout, the number of caregivers per child of a given age, and the educationaland training requirements of the childcare providers. The subsidies given to CPE’s by thegovernment were quite substantial, making up roughly 80% of these agencies’ operating costs.The CPEs were also required to implement the government’s educational program, whichwas based on a version of American High/Scope Educational Approach, whose aim is toensure the well-rounded development of children across all aspects of their personality andeffective motor, language, and socio-emotional skills (Tougas, 2002). Although most childcareproviders already had an educational curriculum in place, many were required to modify theirprograms to meet the stricter requirements.greater than $48,000.984.4. Data Description4.4 Data DescriptionThe data used in this study come from the National Longitudinal Survey of Children andYouth (NLSCY), a nationally representative Canadian survey which collects detailed informa-tion on children’s development and environment from birth through adulthood. It is producedby Statistics Canada. The study is designed to collect information about factors influencing achild’s social, health, and behavioural development and to monitor the impact of these factorson the child’s development over time. An extensive range of data are consequently collected inthe NLSCY, including measures of cognitive and motor development, socio-emotional skills,family economic and educational background, health, and childcare characteristics. Most ofthe information is obtained from parents on behalf of their children through a household inter-view. Direct measures of cognitive and motor development are collected by the interviewerswho directly administer tests and assess the children.The NLSCY includes both a longitudinal and cross-sectional component and sampleschildren of all ages every two years, with eight data collections (called cycles) having takenplace to date. All samples of the NLSCY were drawn from the Labour Force Survey’s (LFS)sample of respondent households. In addition to following the original longitudinal cohort ofchildren who were first sampled in 1994, the NLSCY places a particular focus on monitoringthe early childhood period by adding and following a new sample of infants and young childrenat each cycle, who are primarily aged 0-5 years old.The sample which will be used for the analysis consists of children less than 5 years of age(i.e. 0 to 59 months) in two parent households from all provinces across Canada. It is this agegroup who would be most affected by Quebec’s universal childcare policy, while the exclusionof those five years and older helps avoid confounding the effects of universal childcare withthose due to concurrent changes in Quebec’s kindergarten system and school-age childcareprograms under the new Family Policy. There are multiple reasons for which only two parentfamilies are included in the analysis. First, many single parent families in Quebec were alreadyreceiving heavily subsidized childcare prior to the new Family Policy. Additionally, as BGMdescribe, there were changes in the Quebec welfare system that targeted single mothers whichwere being introduced at the same time as the new Family Policy. Similarly, some otherprovinces in Canada were making changes to their welfare systems in this period, and giventhat a greater proportion of single parent families access these systems, they are more likelyto be affected by the changes. Such contemporaneous policy changes both in Quebec and therest of Canada consequently make it difficult to isolate the effect of the universal childcaresubsidy in Quebec on child health and developmental outcomes for children in single familyhomes. It is for this reason that only children from two parent families are considered in thisstudy. Given that the work and childcare decisions of single parents are likely to be quitedifferent from those in two-parent households, this is an additional reason to focus exclusively994.4. Data Descriptionon only one group.To isolate the impact of Quebec’s policy and net out the effects of the increase in thereduced rate childcare fee in Quebec in 2004 (from $5 a day to $7 a day), only data obtainedprior to 2004 are analyzed in this study. Additionally, given that the expansion in newsubsidized childcare spaces was quite slow in the first couple of years of the reform, withfamilies already using the existing spaces being prioritized and regulated spaces being createdin already existing centres and family homes, I treat the post-reform period as commencingin the fourth data collection cycle (from September 2000 onwards). Observations collected inthe third wave (October 1998 - June 1999) are consequently not included in the post-reformsample because the concern that a large proportion of these children did not have actuallyhave access to a regulated reduced rate childcare space in Quebec. This procedure was alsotaken in the BGM analysis, which will facilitate comparisons between the results of this studyand theirs. Thus, the pre-reform sample consists of children in Cycles 1 and 2 (data fromDecember 1994-April 1997), while the post-reform sample includes children in Cycles 4 and5 (data from September 2000-June 2003). In total, 35,950 children meet the criteria listedabove in terms of age, family type, and NLSCY cycle and form the main sample in this study.It should be noted that Kottelenberg and Lehrer (2014) also use data from Cycles 6 and 7 intheir study, extending the analysis up to 2007.The key child health and developmental measures which are the focus of this study are:i) the Motor and Social Development (MSD) Scale, ii) the Peabody Picture Vocabulary Test-Revised (PPVT-R), and iii) child body weight. The MSD Scale is designed to measuremotor, social, and cognitive development of children aged 0-47 months. It consists of a set of15 questions which vary by the age of the child, asking the person most knowledgeable aboutthe child, usually the mother, whether or not the child is able to perform a specific task. Thescales are standardized by Statistics Canada in one month age groups, with the mean MSDscore being 100 and a standard deviation of 15 across all age groupings.122 The PPVT-Rwas designed to measure receptive or hearing vocabulary for children aged 4-5 years and isa widely used scale for measures of verbal intelligence and school readiness. The PPVT-R was administered by the interviewer through a computer assisted interview. StatisticsCanada standardizes the scores by two month age groups so as to allow comparisons acrossage groups, with a mean of 100 and a standard deviation of 15 for all age groupings. The thirdoutcome of interest is child body weight, which is parent-reported at the time of the surveyand is measured in kilograms. This outcome is not examined in BGM. However, as discussed,differences in activities, physical resources, and diets across child care types may influencechild’s weight. While BGM also investigate other health and non-cognitive outcomes in their122Note that since the full sample of children are not being used in the main analysis (i.e. I use only twoparent households), the sample means and standard deviations are not exactly 100 and 15, respectively, forboth the MSD and the PPVT. This can be seen in Tables 4.1 and 4.2.1004.5. Empirical Strategymean DID estimates, the estimation of quantile treatment effects for these outcomes becomesmore difficult as these variables are discrete and exhibit significant heaping. As such, thisstudy focuses exclusively on the continuous measures of motor and cognitive skills describedabove and on child body weight.A range of information on childcare use and care arrangements was also collected in theNLSCY. Details on the type of care, the number of hours per week in care, as well as basiccharacteristics of caregivers and the care environment are included in the survey. This in-formation is reported by the parents of the child. There is no information on the childcarecurriculum used by the care givers, if any, and there is limited information available to con-struct detailed childcare quality measures. Furthermore, data on the price which families paidfor care were not included in the NLSCY until the seventh cycle. Additionally, no informationwas collected on whether the child had a reduced rate childcare space. Consequently, thereis no knowledge of whether a child living in Quebec in the post-reform period was actuallydirectly impacted by the Family Policy in terms of a change in their childcare arrangements.As will be discussed below, this paper circumvents this lack of information by estimating anintention-to-treat (ITT) effect, which is a common approach taken in the empirical literaturewhen only random assignment to treatment is observed but the actual take-up of treatment isnot. Further details on the ITT, the empirical methodology used in this paper, and descriptivestatistics are discussed in the section below.4.5 Empirical StrategyThis paper uses a difference-in-differences (DID) model to estimate the impact of the universalchildcare program in Quebec on the entire distribution of child health and developmentaloutcomes, as measured by body weight and the MSD and PPVT-R scores. Children inQuebec are observed before and after the 1997 policy change and form the treatment groupin this study. The control group will be made up of children of the same age from all otherCanadian provinces, where there were no major childcare policy changes throughout theperiod of analysis that targeted children from two parent families. It is the effect of the newFamily Policy on the treatment group (i.e. Quebec) which will be the focus of this study.As it will be discussed in detail below, identification is achieved by using the comparisongroup’s pre- and post distributions to construct a “counterfactual” distribution of outcomesthat would have prevailed in the treated group in the absence of the policy. In this study,the quantile treatment effect will be defined as the horizontal distance between the observedmarginal distribution of outcomes in Quebec in the post-reform period and the counterfactualdistribution. This approach then essentially permits the estimation of the policy change onany feature of the Quebec distribution.It should be noted clearly that although heterogeneity is allowed to exist across children1014.5. Empirical Strategywith respect to differential treatment effects and time trends in this study, the treatment effectfor a particular individual child cannot be identified without invoking additional, strongerassumptions. In particular, one assumption sometimes made in the literature to identify theeffect for an individual is that an observed child would maintain her rank in the distributionregardless of her actual treatment status. This is referred to as the “rank preservation”assumption in the literature (Heckman et al. 1997). When rank preservation holds, thenthe horizontal difference between the two marginal distributions will identify the individualtreatment effect for those at a given threshold. However, given that the rank preservationassumption is quite strong, this study does not attempt to identify treatment effects at theindividual level, and instead, the focus is on the distributional effects of the childcare reform.As such, all quantile treatment effects in this study should simply be thought of as identifyingthe difference in quantiles, at a given threshold level, between the observed and counterfactualin Quebec in the post-reform period.As was touched upon above, due to a lack of information in the NLSCY on whether a childactually receives subsidized care, this study estimates an intention-to-treat (ITT) effect ratherthan the treatment on the treated (TT) effect. The ITT gives the full impact of the universalchildcare policy on the developmental outcomes of all children in Quebec eligible for subsidizedcare, regardless of whether or not their childcare arrangements were actually affected by thenew Family Policy. Usually, however, the TT effect is of most interest to policy makers asit measures the change in outcomes for those whose childcare arrangements were affected(i.e. the treated). As will be discussed below, there are various ways in which treatment canbe defined with the new Family Policy, and estimates will be obtained to measure exactlywhat proportion of children are “treated” under various definitions of treatment. Althoughthe TT effect is often of most interest, there are a couple of advantages of examining theITT rather than the TT. First, estimation of the ITT circumvents potential endogeneityissues, as clearly the take-up of regulated, reduced fee childcare spaces is not exogenous.Additionally, the ITT captures any peer effects resulting from the new Family Policy, wherebythe childcare arrangement of one child is allowed to affect the outcomes of another child. Underthe assumption that the developmental outcomes of untreated children were unaffected by thenew Family Policy (i.e. no peer effects), then the ITT and the TT differ only by some scalingfactor. This scaling factor is given by the inverse of the proportion of eligible children who areactually treated in Quebec, and multiplying the ITT by the scaling factor gives the TT. Whenthe proportion of eligible children that are treated approaches unity, then the ITT approachesthe TT.In the following subsections, I will briefly provide an overview of the standard DID model,which is used to identify the average ITT effect of the universal childcare program on Quebecchildren. This is the estimation strategy used by BGM. Then, I extend its main ideas tothe quantile difference-in-differences model (QDID) which will be used to identify heteroge-1024.5. Empirical Strategyneous treatment effects across the distribution of outcomes of children in Quebec. Descriptivestatistics for the child developmental scores (MSD and PPVT-R) and body weight will thenbe presented, along with statistics on covariates, childcare arrangements, and the householdenvironment. The following section of the paper will then reveal the estimates of the impactof the new Family Policy on the children of Quebec.4.5.1 The Standard DID EstimatorThe typical notation used for the standard DID is as follows: Child i belongs to groupGi ∈ {0, 1} where G = 1 if the child lives in the treatment province (i.e Quebec) and G = 0otherwise. In a simple model where there are only two time periods (i.e. pre and post-reform),child i is observed in Ti ∈ {0, 1}, where T = 0 denotes the pre-reform period and T = 1 thepost-reform period. Also, let Yi denote child i’s observed MSD, PPVT-R, or body weightoutcome. Thus, for a given child i, the triplet (Gi, Ti, Yi) is observed. Following the potentialoutcomes literature motivated by Rubin (1978), let Y 0i denote the outcome of individual iwhen she is not treated and Y 1i denote the outcome of this individual when she does receivetreatment. Clearly only Y 0i or Y1i is observed at a given point in time, but not both. Let Iidenote an indicator for whether child i is treated, with Ii = 1 if she is and Ii = 0 otherwise. Tosimplify matters, assume for the moment that all children eligible for universal childcare areactually treated, and later on, further notation will be introduced to relax this assumption.Then, the observed outcome for child i is given by:Yi = Y0i + (Y1i − Y0i )Ii (4.1)In the standard DID model, if outcomes are linear in covariates, X, then the outcome forchild i in the absence of treatment can be written as:Y 0i = α+ γTi + δGi +Xiβ + εi (4.2)where γ represents the time effect and δ represents the group fixed effect; Xi is a 1xk vectorof covariates for child i; β is a kx1 vector of coefficients on these covariates; and εi is anunobserved component that affects outcomes.Note that by definition, the average treatment on the treated (TT) effect, ∆DID, is givenby:∆DID ≡ E[Y 1i |G = 1, T = 1]− E[Y0i |G = 1, T = 1] (4.3)The problem in estimating the above is that the last term on the right hand side, namelyE[Y 0i |G = 1, T = 1], is not observed. The focus of the standard DID model is consequentlyhow to construct a proper counterfactual to estimate this unobserved term. In the standardDID model, the unobserved component εi is assumed to be independent of group assignment1034.5. Empirical Strategyand time, εi⊥(Gi, Ti, ), meaning that the underlying distributions of unobservables is identicalacross all groups and time periods so that universal childcare eligibility status isn’t related tounobservables (i.e. the unconfoundedness assumption). Under this assumption, the averagetreatment on the treated effect (conditional on X) in the standard DID model is:∆DID|X = [E[Yi|G = 1, T = 1, X]− E[Yi|G = 1, T = 0, X]]−[E[Yi|G = 0, T = 1, X]− E[Yi|G = 0, T = 0, X]](4.4)and the unconditional average treatment on the treated effect is given by:∆DID = E[∆DID|X]= [E[Yi|G = 1, T = 1]− E[Yi|G = 1, T = 0]]− [E[Yi|G = 0, T = 1]− E[Yi|G = 0, T = 0]](4.5)Thus, the identifying assumption used to generate the counterfactual for the average outcomeof the treated group in the absence of treatment is that there is a common time trend acrossQuebec and the other provinces which is unrelated to the policy change. Equation (4.5) aboveshows how subtracting the average difference in outcomes over time in the control group fromthe treatment group removes this common time trend and identifies the treatment effect.That is, the identifying assumption amounts to assuming:E[Y 0i |G = 1, T = 1]− E[Y0i |G = 1, T = 0]= E[Y 0i |G = 0, T = 1]− E[Y0i |G = 0, T = 0](4.6)In practice, the standard DID estimator is often obtained by assuming the treatment effect isconstant across individuals, such that ∆DID = Y 1i − Y0i for all i and then running a simpleOLS on the following model to estimate ∆DID:Yi = α+ γTi + δGi + ∆DIDIi +Xiβ + i (4.7)To estimate the standard DID effect in this study for the ITT, I simply extend the two-period,two-group model above to the case where there are multiple time periods (four in total forNLSCY cycles 1, 2, 4, and 5) and multiple groups (10 provinces in total). Additionally, Irelax the assumption that all Quebec children eligible for universal care were actually treatedin the post-reform period. To do this, I simply replace the indicator Ii in Equation (4.7)with an indicator for whether the child is eligible for universal, subsidized care, denoted byELIGi. ELIG = 1 if the child is eligible for subsidized care (i.e. is observed in Quebec inthe post-reform period and is of eligible age) and ELIG = 0 otherwise. Then, I estimate the1044.5. Empirical Strategyfollowing model using OLS:Yi = α+4∑k=1γkTki +10∑j=1δjGji + θELIGi +Xiβ + εi (4.8)where Tki for k ∈ {1, 2, 4, 5} denotes the NLSCY cycle in which child i is observed; γk isthe coefficient associated with time period k; Gji denotes the province of residence of childi where j ∈ {1, 2,..., 10}; and δj is the province fixed effect for province j. Here, θ is theprimary coefficient of interest and is an estimate of the average ITT effect. Again, underthe assumption of no externalities or peer effects, θ will approach ∆DID as the proportionof eligible children who are actually treated approaches unity. This basic model is extendedto the case where heterogeneous responses to the universal childcare reform are of primaryinterest.4.5.2 Quantile Difference-in-Differences (QDID)Consider again the simple two-group, two-period model described above. To estimate thequantile treatment effects, further notation must be introduced.123 To ease notational burden,I drop the subscript i and treat (Y,G, T ) as a vector of random variables. Further, it is assumedthat:Y 0gt →d Y0| G = g, T = t Y 1gt →d Y1| G = g, T = tand Ygt →d Y | G = g, T = twhere →d is shorthand for “distributed as”. The (unconditional) cumulative distributionfunctions corresponding to the above are denoted by FY 0, gt, FY 1, gt, and FY,gt respectively.Additionally, let the inverses of the distribution functions (i.e. the quantile functions) bedenoted byq0gt(τ) = F−1Y 0, gt(τ) q1gt(τ) = F−1Y 1, gt(τ)and qgt(τ) = F−1Y,gt(τ)where τ is some real number such that τ ∈ (0, 1) and is the threshold level of interest. Thedistributions of outcomes which are observed are: FY 0, 10, FY 1, 11, FY 0, 00, and FY 0, 01 as are theirrespective quantile functions. The distribution of outcomes which is not observed is FY 0, 11that is, the distribution of outcomes for children in Quebec in the post-reform period thatwould exist in the absence of the new Family Policy. This study is concerned with estimatingthis counterfactual distribution, which will be denoted by FCY 0, 11, and its inverse, denoted byqC11(τ).123The notation used in this section is based on a model by Athey and Imbens (2006).1054.5. Empirical StrategyThe approach taken in this study to estimate the counterfactual distribution FCY 0, 11 usesthe quantile difference-in-differences model (QDID), where quantile changes in the comparisongroup over time at a given threshold level, τ , are used to identify the counterfactual quantilefor the treated group. As mentioned previously, in this study the quantile treatment effectfor a given τ is defined as the horizontal distance between the distribution functions of thepost-reform treatment group and its counterfactual. That is,∆QDID(τ) = q111(τ)− qC11(τ) (4.9)where ∆QDID(τ) is the quantile treatment effect in the QDID model for a given thresholdlevel τ . In the QDID model, the counterfactual quantile at the τ -th percentile is constructedas:qC11(τ) = q010(τ) + [q001(τ)− q000(τ)] (4.10)Just as in the standard DID model, the identifying assumption for the QDID estimatorto give an unbiased estimate of the impact of the childcare reform for a given threshold level,τ , is a common time trend assumption as outlined in Equation (4.10). Here, however, theassumption is more stringent than in the standard DID model in that a common trend isassumed to hold at each threshold level, τ , whereby it is assumed that the change in thequantile value at the τ -th threshold would be same between the treatment group and thecontrol group in the absence of the reform. It should be noted that the QDID does not putany limitations on differences in the shape of the distribution functions between the treatmentand control groups at a given point of time. Rather, the QDID achieves identification of thetreatment effect by putting restrictions on the changes in these distributions within eachgroup over time. Also, just as in the standard DID model, unconfoundedness is assumed tohold, so that i⊥(Gi, Ti).Given that fixed effects for each province and each NLSCY cycle are controlled for in theestimation, as will be seen explicitly below, the effect of the new Family Policy in Quebecfor a given threshold value, τ is identified by the change in quantiles in Quebec, relative toother provinces, in the post-reform period (cycles 4 and 5) compared to the pre-reform period(cycles 1 and 2). Thus, it should be noted that a disadvantage of both the standard DID andthe QDID models is that any Quebec-specific shocks that coincide with the 1997 childcarepolicy will bias the estimates as neither of the models are able to separately identify thisshock from the introduction of the policy. Similarly, if any other policies were implementedeither in Quebec or the rest of Canada during this time which affected child outcomes, therewould be a bundling problem as the DID estimator cannot disentangle the new Family Policyfrom any other policy, resulting in biased estimates. Related to this, if Quebec had differentlabour market trends compared to the rest of Canada and these trending labour marketcharacteristics affected child outcomes, then again the DID will give biased estimates of the1064.5. Empirical Strategyuniversal childcare policy. Recent research suggests that family income has a significant,positive causal effect on children’s development.124 If this is the case then, given that theearly 1990s in Canada were characterized by a deep recession and the early 2000’s by higheconomic growth, any differential trends between Quebec and the rest of Canada in terms ofthe improving labour market could plausibly result in different child development trends whichwould again violate the assumptions of the DID framework. Similarly, if family income has acausal impact on child outcomes and if parental work preferences were changing differentiallyacross the regions over time, again the standard DID and the QDID estimators may be biased.It is unlikely that social programs in Quebec or other provinces are confounding the ef-fects, particularly since I focus on two-parent families. First, no other provinces implementeduniversal childcare programs in this period. Second, although there were changes in theQuebec welfare system introduced at the same time as the new Family Policy, they largelytargeted single mothers (see BGM for more details). Third, while the National Child Ben-efit (NCB) program was introduced in 1998, it provided reduced amounts for families withincomes above a threshold (initially $15,921). Some provinces also introduced small earnedincome or family supplements as part of the NCB, but all were income-tested. In some cases,provinces also partially integrated social assistance with the NCB.125 Since these programsare all income-tested, they are accessed at a much higher rate by single parent families. Forexample, Milligan and Stabile (2007) find that 43.5% of single Canadian women are on socialassistance which compares to only 4.4% of married women. They also find that husbands’income tends to push families over the income threshold for the NCB program ($25,921 in1998), above which families no longer received supplements. In particular, they estimate that74.7% of single mothers receive some NCB supplement while only 17.8% of married mothersdo. Given that I focus on children from two-parent families, these findings allay the concernthat the Quebec Universal childcare policy is being confounded with other social programs.An additional requirement for both the standard DID and the QDID estimator to giveunbiased estimates is that the introduction of the Quebec policy must really have been ex-ogenous to child development outcomes. That is, it cannot be that the introduction of thepolicy was in response to contemporaneous labour and child development conditions; other-wise, there would be issues of reversal causality, leading to biased estimators. BGM note thatthey find little evidence to suggest that the Quebec policy arose from any contemporaneousdevelopments in Canada or the rest of Canada and was instead the result of a lengthy publicdiscourse, suggesting that such political endogeneity is unlikely. In addition, the DID esti-mator rules out the existence of any pre-treatment effects, whereby Quebec parents reactedin anticipation of the policy prior to its actual introduction. Again, if there were any pre-124See Dahl and Lochner (2008) who use changes in tax credits in the US to find that family income has apositive, significant effect on children’s children math and reading scores.125A detailed overview of the NCB is provided by Milligan and Stabile (2011).1074.5. Empirical Strategytreatment effects, both the standard DID and the QDID would result in biased estimates ofthe new Family policy.While there are similarities between the standard DID and the QDID models, under theidentifying assumptions outlined above, the QDID approach allows the estimation of thetreatment effect across the treatment group’s entire distribution of outcomes, whereas thestandard DID only examines the mean treatment effect. The standard DID will only renderthe same estimates as the QDID estimator in the case where there’s no heterogeneity.It should be emphasized that the QDID model estimates treatment effects at variousquantiles of the marginal distribution rather than of the conditional distribution as is madeclear in Equation (4.9) where ∆QDID(τ) is not a function of any covariates. However, asFrolich and Melly (2010) note, including covariates in the analysis can help increase theefficiency of the estimators and can also control for any systematic differences in the set ofobservable covariates between Quebec and the other provinces which may have motivatedthe introduction of the new Family Policy in the first place. Consequently, covariates will beincluded in the estimation of the Quebec childcare policy. The steps taken to estimate theQDID are described next.4.5.3 Estimating the QDID ModelIn order to derive the QDID estimates across the distribution of the outcome variable, twomethods were used to evaluate the robustness of the results to the estimation technique. Thefirst involves running a series of regressions of a transformation of the outcome variable on theset of covariates and treatment status indicator, using a recent estimation technique proposedby Firpo, Fortin and Lemieux (2009). This approach will be referred to as the FFL approach.The second estimation method involves an estimation procedure proposed by Firpo (2007),where covariates are used to construct observational weights and estimation does not requireany computation of densities, unlike the FFL approach. This method will be referred to asthe Firpo (2007) approach. Each is described in detail below.The FFL approach is a relatively new regression method that can be used to evaluate theimpact of changes in explanatory variables on the quantiles of the unconditional distributionof an outcome. As is well known, the standard conditional quantile regression model (e.g.Koenker and Bassett 1978) is not particularly helpful for estimating unconditional quantiletreatment effects because, unlike the standard OLS regression, the average of conditionalquantiles estimates is not equal to the unconditional quantile and the difference between thetwo can often be very large. The FFL methodology addresses this issue by estimating theeffect of a change in covariates on the unconditional quantile using the recentered influencefunction (RIF) as the dependent variable in a linear regression framework. In particular, theinfluence function (IF) provides the influence or contribution of each data point to the τ -th1084.5. Empirical Strategyquantile of Y , qτ , and is given by IF (Y, qτ , FY ) = (τ − 1{Y ≤ qτ})/fY (qτ ), where 1 is theindicator function and fy(qτ ) is the density of y evaluated at qτ . Adding back the value ofthe τ -th quantile to the influence function then gives the recentered influence function (RIF).Thus, the RIF can be written as follows:RIFi(Yi, qτ , FY ) = qτ + (τ − 1{Y ≤ qτ})/fY (qτ )Note that the expected value of the RIF will be qτ itself. Importantly, qτ can be expressedin terms of the conditional expectation of RIF given a set of covariates X using the law ofiterated expectations.FFL show that by using the RIF as the dependent variable in an OLS regression on aset of covariates, the estimated coefficients on the covariates give the unconditional quantile(partial) effects. It should be noted that running a regression of RIF on a set of covariates Xamounts to running a linear probability model for whether the observed outcome of individuali, Yi, is above the quantile of interest (i.e. Pr[Y ≥ qτ ]), but here in the case of RIFs, thecoefficients must be divided by the density evaluated at that quantile.Prior to running regressions, an estimate of the RIF must first be derived. This involvesestimating both qτ and fY (qτ ), which can be done with the usual τ -th sample quantile (e.g. asoutlined by Koenker and Bassett 1978) and a kernel density estimator, respectively. Then, fora given value of τ , the following regression is run with OLS to estimate the quantile treatmenteffect of the reform at τ -th quantile:126R̂IFi(Yi, qτ , FY ) = ατ +4∑k=1γτkTki +10∑j=1δτjGji + θτELIGi +Xiβτ + τi (4.11)All variables are defined as in the standard DID model, except here each coefficient rep-resents the effect of a change in a given covariate on the unconditional quantile, where thethreshold level is given by τ . In the empirical analysis, X consists of the following covari-ates: child age and gender; parental education (grouped into high school dropout, high schoolgraduate, some post-high school, and university degree); parental age (grouped into 5-yearcategories, starting with 16-20 and ending with 46+); parental immigration status; the size ofthe urban area (grouped into five categories of population size: rural, under 30,000, 30,000-99,999, 100,000-499,999, and 500,000+); and number of older and younger siblings (eachgrouped into three categories: zero, one, and two or more). Note that because income isendogenous to the labour supply response, which the Quebec policy likely affected, it is notincluded in the analysis; although, the inclusion of parental education will partly control forfamily socioeconomic background.126Note this is done automatically with the rifreg package in Stata developed by FFL.1094.5. Empirical StrategyThe parameter of interest in Equation (4.11) is θτ , which gives the ITT estimate of thequantile treatment effect of being eligible for universal childcare at the τ -th threshold level.In the empirical analysis, I estimate the impact of the reform at all 1-99 percentiles. Notethat the estimator derived from (11) will be consistent so long as Pr[Y ≥ qτ ] is linear incovariates X. In their 2009 paper, FFL discuss how to implement more flexible estimatorswith the RIF.To estimate the τ -th quantile, I use the Gaussian kernel in all specifications along withthe “optimal” bandwidth (i.e. the bandwidth that minimizes the mean integrated squarederror in a given specification). I take into account that estimation of Equation 4.11 is donewith a generated variable (i.e. the density and quantiles are being estimated) by obtainingstandard errors of the estimates using 100 draws from the original data and bootstrappingover the entire process.The second method to estimating the quantile effects follows an approach developed byFirpo (2007), which he outlines is only appropriate when selection into treatment is randomor may be based on observable characteristics. As discussed above, there is little evidenceto suggest that the introduction of the new Family Policy was related to contemporaneouslabour and child development outcomes in Quebec, and consequently, the Firpo method seemsappropriate to use in the present case. There are two steps for the Firpo approach to beimplemented. First, a propensity score is estimated for being in the treatment group (i.e.Quebec), which is denoted by P (X). This was done by using the predicted value of a logitregression of being in the treatment group on the covariates X. It should be noted that theFirpo method does require the assumption of common support, meaning 0 < P (X) < 1,which implies that for all values of X, both treatment and comparison assignment have apositive probability of occurrence. The second step of the Firpo method involves computingthe sample quantiles for each group in each time period in the usual fashion (e.g. Koenkerand Bassett 1978) by minimizing a sum of check functions, except here, the check functionsare weighted by a factor relating to the probability of being in the treatment group (i.e. aninverse probability weighting scheme). As Firpo (2007) outlines, the weighting function foran individual is given by:Wi =QiN · Pˆ (X)+1−QiN · (1− Pˆ (X))(4.12)where Qi is an indicator for whether the child lives in Quebec (with Q = 1 if she doesand Q = 0 if she does not), N is the total number of children in the sample, and Pˆ (X) is theestimate of the propensity score obtained in the first step.Consider again the simple two group, two period case. Then, for a given group g ∈ {0, 1}1104.5. Empirical Strategyat time t ∈ {0, 1}, the estimate of the τ -th quantile is given byqˆgt(τ) = argminqN∑i=1Wi · ρτ (Yi − q) (4.13)where the check function ρτ (·) evaluated at a real number a is ρτ (a) = a · (τ − 1{a ≤ 0}).As Firpo (2007) points out, the weights used in the check functions reflect the fact that thedistribution of the covariates differs between the comparison and treatment groups.The counterfactual quantile for the treated group using the Firpo method in the two-group,two time-period simplification is then given by:qC11(τ) = qˆ010(τ) + [qˆ001(τ)− qˆ000(τ)] (4.14)The quantile treatment effect is then defined as in Equation (4.9) by plugging in qC11(τ)above and qˆ111(τ), as derived in (13). In the estimation, multiple groups and time periods wereused to derive the quantile treatment effects. In particular, the quantile treatment effectswere calculated for 36 different combinations obtained by varying the comparison province (9possible provinces) and by varying the pre/post-reform time period (4 possibilities: Cycles 1and 4; Cycles 1 and 5; Cycles 2 and 4; Cycles 2 and 5).127 As Athey and Imbens (2006) pointout, each of these combinations should provide consistent estimates of the actual treatmenteffect. The overall quantile treatment effect then was derived as a weighted average from the36 different combinations, where the weights are based on the number of children observed ineach province in a given time period. Again, the impact of the reform is estimated at all 1-99percentiles using the Firpo method.Although the two approaches taken to estimate the quantile treatment effects are different,each relies on the same identification assumption outlined in Equation (4.10). The primarydifference between the two is that the Firpo approach is more flexible than the linear FFLestimating equation in (4.11). The estimates of the quantile treatment effects using each esti-mation technique, along with estimates from the standard DID, are provided in the followingsection. First, however, the descriptive statistics for the dependent variables (MSD, PPVT-R,and body weight) and the control variables are presented below.127Note: The Firpo estimation was carried out using the ivqte command developed by Frolich and Melly(2010). Given this is a multistage estimator, the standard errors were bootstrapped based on 199 draws fromthe original sample (with replacement) whereby observations were independently drawn within each provinceand each NLSCY cycle so as to ensure that each bootstrap sample has the same proportion of observationsfrom each province and each cycle as in the original dataset.1114.5. Empirical Strategy4.5.4 Descriptive StatisticsTables 4.1 to 4.3 show the values of the MSD scores , PPVT-R scores, and body weight atdifferent percentiles in Quebec before and after the introduction of the universal childcareprogram in 1997. The differences in percentiles between Quebec and the rest of Canada inthe pre and post-reform periods are also provided in these tables. Table 4.1 shows that thevalues of MSD scores in Quebec at various percentiles are lower after the reform comparedto before, dropping by 1-2 points. For the mean, the average MSD score slightly increased.Additionally, the pre-reform MSD values are lower in Quebec than the rest of Canada at almostall percentile levels, in addition to at the mean, with the exception of the 10th percentile,where the scores were the same. Interestingly, the last two columns of Table 4.1 show thatthe gap between percentiles across regions grew larger in the post-reform period, particularlyat the lower percentiles, where the difference increased by three points at the 25th percentileand two points at the 10th percentile, leaving Quebec faring even worse in the post-reformperiod.Table 4.1: MSD Percentiles by Time Period and RegionLevel DifferenceQuebec Quebec-Rest of CanadaPre-Reform Post-Reform Pre-Reform Post-Reform10th Percentile 81 80 0 -225th Percentile 91 89 -1 -450th Percentile 101 100 -1 -275th Percentile 109 108 -2 -390th Percentile 116 115 -2 -2Mean 99.13 99.26 -1.61 -2.07Standard Deviation 14.70 14.52 -0.46 0.33No. of Children:Quebec 2,661 2,505Rest of Canada 10,834 11,496Notes: The pre-reform period corresponds to children observed in NLSCY Cycles 1 (1994-95) and2 (1996-97), while the post-reform period corresponds to children in NLSCY Cycles 4 (2000-01) and5 (2002-03). The outcome variable, MSD, is defined in the text. The percentiles were separatelycalculated for Quebec and the rest of Canada in a given time period.The same general patterns for the PPVT-R scores can be seen in Table 4.2, where thepre-reform scores in Quebec are lower in the post-reform period at the lower threshold levels.Again, the mean is slightly higher in Quebec in the post-reform period. Additionally, the restof Canada had higher scores at the 25th, 50th, and 90th percentiles in the pre-reform period.1124.5. Empirical StrategyAgain, examining the last two columns of Table 4.2, the same general pattern holds as inthe case of the MSD scores, where Quebec fares relatively worse in the post-reform periodcompared to the other provinces at almost all threshold levels (the only exception is the 90th).The relative decline in Quebec scores is again most stark at the lowest percentiles, where thegap between the two regions increased by five and three points respectively for the 10th and25th percentiles.Table 4.2: PPVT-R Percentiles by Time Period and RegionLevel DifferenceQuebec Quebec-Rest of CanadaPre-Reform Post-Reform Pre-Reform Post-Reform10th Percentile 81 78 0 -525th Percentile 90 89 -1 -450th Percentile 100 99 -1 -275th Percentile 111 111 1 -190th Percentile 119 121 -1 1Mean 100.53 100.93 0.38 -1.43Standard Deviation 15.14 14.61 0.75 -0.09No. of Children:Quebec 533 524Rest of Canada 2,132 2,226Notes: The pre-reform period corresponds to children observed in NLSCY Cycles 1 (1994-95) and 2(1996-97), while the post-reform period corresponds to children in NLSCY Cycles 4 (2000-01) and 5(2002-03). The outcome variable, PPVT-R, is defined in the text. The percentiles were separatelycalculated for Quebec and the rest of Canada in a given time period.Table 4.3 shows descriptive statistics for children’s body weight. The pre-reform bodyweight in Quebec is slightly lower than the post-reform period at the lower threshold levels(the 10th and the 25th), but is fairly similar at higher thresholds. The last two columnsof this table show that at the upper thresholds, Quebec children have relatively lower bodyweight than the rest of Canada in the post-reform period compared to the pre-reform period.However, at the lower thresholds, there is no difference between provinces in both periods.Descriptive statistics of covariates, childcare characteristics, and the home environment inQuebec and the rest of Canada before and after the reform are presented in Table 4.4. Thetop of the table shows the means and standard deviations of the covariates, X, included inthe analysis. With the exception of age (child and parents’) and number of siblings, whichare continuous variables, all the covariates have been expressed as 0/1 dummy variables forthe construction of this table. Table 4.4 shows that the values of the covariates are quite1134.5. Empirical StrategyTable 4.3: Body Weight Percentiles by Time Period and RegionLevel DifferenceQuebec Quebec-Rest of CanadaPre-Reform Post Reform Pre-Reform Post-Reform10th Percentile 8.17 9.07 0 025th Percentile 10.89 11.34 0 050th Percentile 13.61 13.61 0 075th Percentile 16.00 15.88 -0.33 -0.990th Percentile 18.14 18.14 0 -0.91Mean 13.6 13.8 0.02 -0.41Standard Deviation 4.06 3.70 -0.02 -0.24Observations:Quebec 3,251 3,045Rest of Canada 13,382 14,239Notes: The pre-reform period corresponds to children observed in NLSCY Cycles 1 (1994-95) and2 (1996-97), while the post-reform period corresponds to children in NLSCY Cycles 4 (2000-01) and5 (2002-03). The outcome variable, body weight, is defined in the text and measured in kilograms.The percentiles were separately calculated for Quebec and the rest of Canada in a given time period.similar in Quebec and the rest of Canada in both periods, with the exception of parentimmigration status where the proportion of immigrants is higher in the rest of Canada. Mostimportantly, however, is that there are no noticeable differential trends in these covariatesacross the treatment and control groups between the pre and post-reform periods. This isencouraging in that any substantial changes over time in the demographics of children acrossQuebec and the rest of Canada may suggest there are also unobserved compositional changesin a region, which would violate the assumptions outlined in the empirical strategy.Table 4.4 also shows descriptive statistics of childcare characteristics in Quebec and therest of Canada across time. Again, all variables with the exception of hours of care, which iscontinuous, have been expressed as 0/1 dummy variables for the construction of this table.As expected, there is a large increase in the proportion of Quebec children in care between thepre-reform (42%) and post-reform periods (62%), a trend which is not observed in the rest ofCanada to the same extent. The types of care that experience the largest proportion of growthin Quebec are i) institutional care (increase from 11% to 30%), which consists primarily ofcentre care, but also includes nursery and pre-school, and ii) licensed care in others’ homes(increase from 5% to 11%). Note that this is aligned with the new Family Policy in thatthe newly established CPE’s, which were injected with large amounts of government fundingfollowing the reform to increase the number of spaces, consisted of both regulated centre and1144.5. Empirical StrategyTable 4.4: Descriptive StatisticsQuebec Rest of CanadaPre-Reform Post-Reform Pre-Reform Post-ReformCovariates:Age 2.03 2.01 2.00 2.02[1.42] [1.41] [1.42] [1.41]Male 0.51 0.52 0.51 0.51[0.50] [0.50] [0.50] [0.50]Mother Age 30.92 31.21 31.69 32.32[4.87] [5.39] [5.09] [5.46]Father Age 33.52 33.98 34.07 34.86[5.39] [5.85] [5.65] [6.00]Mother High School Dropout 0.13 0.12 0.11 0.09[0.34] [0.33] [0.31] [0.29]Mother University Degree 0.20 0.27 0.20 0.28[0.40] [0.44] [0.40] [0.45]Father High School Dropout 0.17 0.16 0.14 0.11[0.37] [0.36] [0.34] [0.31]Father University Degree 0.19 0.24 0.22 0.26[0.40] [0.43] [0.41] [0.44]Mother Immigrant 0.09 0.12 0.22 0.24[0.28] [0.33] [0.41] [0.43]Father Immigrant 0.10 0.13 0.21 0.24[0.30] [0.33] [0.41] [0.43]No. of Older Siblings 0.71 0.71 0.80 0.76[0.74] [0.72] [0.76] [0.73]No. of Younger Siblings 0.27 0.23 0.26 0.25[0.49] [0.45] [0.48] [0.47]Rural Area 0.15 0.15 0.15 0.11[0.36] [0.36] [0.36] [0.31]Child Care Characteristics:In Child Care 0.42 0.62 0.41 0.46[0.49] [0.49] [0.49] [0.50]Care in Own Home 0.07 0.08 0.11 0.12[0.26] [0.27] [0.31] [0.33]Care in Others’ Home 0.23 0.25 0.24 0.25[0.42] [0.43] [0.42] [0.43]In Institutional Care 0.11 0.30 0.06 0.09[0.31] [0.46] [0.23] [0.29]Care in Other Home, Licensed 0.05 0.11 0.04 0.05[0.21] [0.31] [0.19] [0.22]Hours of Care/Week 13.79 21.08 12.13 13.17[19.76] [21.38] [18.62] [18.18]Household Environment:Parenting Scale- Hostile 8.29 8.61 9.19 8.78[3.85] [3.28] [3.74] [3.40]Parenting Scale- Aversity 4.34 3.94 5.19 4.56[2.01] [1.97] [2.30] [2.11]Parenting Scale- Consistency 14.04 14.13 14.70 15.36[3.27] [3.13] [3.40] [3.10]Family Functioning Scale 7.19 8.35 7.81 8.70[4.99] [5.01] [5.15] [4.87]Mother’s Depression Score 4.18 3.92 4.53 3.83[4.54] [4.78] [4.94] [4.43]No. of Obs. 3,407 3,305 14,005 15,233Notes: The pre-reform period corresponds to children observed in NLSCY Cycles 1 (1994-95) and 2 (1996-97), while the post-reform period corresponds to children in NLSCY Cycles 4 (2000-01) and 5 (2002-03). Thedescriptions of variables are defined in the text. Means and standard deviations were separately calculated forQuebec and the rest of Canada in a given time period. Standard deviations are in parentheses.1154.6. Resultsfamily home care. Additionally, Table 4.4 shows that there was a large increase in the numberof hours per week Quebec children spent in care, which rises from just under 14 hours perweek in the pre-reform period to over 21 hours in the post-reform period. Again, this trendis not observed in the rest of Canada over time.The last part of Table 4.4 shows measures of the household environment. The NLSCYcollects information on the quality of parent-child interactions and on the well-being of theparents by asking a series of questions to the parents. Although these measures are notof primary interest in this study, it seems plausible to expect that these factors might beaffected by the increased use of childcare in Quebec. In particular, BGM find that the policyresulted in the deterioration of the household environment, which they interpret likely aroseas a response to the elevated stress associated with increased rates of two parent workingfamilies and childcare use created by the policy.Three measures of parenting style are used to evaluate whether there were changes inthe household environment: i) Hostile and Ineffective Parenting, ii) Aversive and PunitiveParenting, and iii) Consistent Parenting. These measures are obtained from a series of parent-reported questions in families with children 2-4 years of age, which are then aggregated toform the above indices. The range for hostile parenting is 0-25 points, while the range of bothaversive and consistent parenting is 0-20, with higher scores indicating a greater presence ofthe particular characteristic. Between the pre and post-reform periods, the average degree ofhostile parenting worsens in Quebec slightly, while improvements in aversive and consistentparenting are made. For the rest of Canada, there are improvements in the averages of all threeparenting behaviour measures over the time periods. Questions on family functioning werealso collected in the NLSCY to provide an indication of the quality of family relationships.The range of this index is from 0-36, with higher scores indicating greater family dysfunction.As Table 4.4 shows, there was a greater average level of dysfunction in both Quebec andthe rest of Canada in the post-reform period, with the size of the deterioration in Quebecbeing slightly larger. Finally, parents were asked about their own feelings in the NLSCYand a measure of maternal depression was collected. The range of this variable is from 0-36,with higher scores indicating greater maternal depression. As the last row of this table shows,both regions found a decrease in average maternal depression between the pre and post-reformperiods, although the reduction was greater for the rest of Canada.4.6 ResultsThis section presents the results from the estimation techniques described above. First, esti-mates of the proportion of Quebec children who were treated in the post-reform period arepresented, under various definitions of treatment. This then informs on the value of the ITTscaling factor which can be used to derive the treatment on the treated (TT) effects, under the1164.6. Resultsassumptions outlined above. Then the standard DID estimates are presented, where the aver-age impacts of the universal childcare policy on Quebec children are presented. The quantiletreatment effects are revealed for the full sample, with some robustness checks performed, andthe section then concludes with subsample analyses where estimates are derived for groups ofchildren separated on their demographic and household characteristics.4.6.1 The ITT Scaling FactorIn order to obtain an idea of what the scaling factor would be to convert the ITT to the TTeffect, a series of OLS regressions were carried out to determine what proportion of eligiblechildren in Quebec were actually affected by the program in terms of changes in childcarearrangements. As BGM explain, it isn’t clear how “treatment” should be defined becausevarious factors could be affected by the universal childcare program and played a role on childhealth and developmental outcomes. For example, it could be defined as being in any typeof childcare, reflecting the parental response to reduced childcare prices by substituting theirown care for the care of others. This could alter the net resources received by children, andthe impact on child outcomes would depend on the quality of parental versus non-parentalcare. Treatment could also be defined as being in institutional or licensed care. This is maindefinition of treatment I use in my study and focuses on differences in resources across formalcare versus informal care. Furthermore, some children may also have experienced an increasein the intensity of care since the program was available for longer periods of care per day, sotreatment could alternatively be defined as the intensity of care. Finally, since the new FamilyPolicy sought to improve the curriculum of childcare, treatment could instead be defined aschanges in the quality of care within a childcare setting.Conversely, treatment could also be viewed as factors besides childcare itself, such asincreased parental employment and household income. If lower childcare prices allowed in-creased employment by reducing the opportunity costs of working, then treatment could bedefined as exposure to maternal employment. In this case, treatment would capture bothchanges in mother’s labour supply as well as household income. Greater household incomemay lead to positive effects on child development, but increased maternal employment mayalso lead to more parental stress in the home and less time engaged in productive home ac-tivities. Treatment could also be defined as an increase in mothers remaining at home andnot working but using care outside the home. Finally, for those families who were alreadyusing childcare and did not alter their arrangements, then the policy represents a pure incomeeffect.128128One possible response to Quebec’s new Family Policy is an increased fertility rate due to a reduction inthe costs of having children. To my knowledge, no existing empirical study has examined if this occurred.Milligan (2005) did find strong, positive impacts on fertility from the introduction of a pro-natalist transferpolicy introduced in Quebec in the late 1980s that paid up to $8,000 to families having a child. This suggeststhat Quebec families may be responsive to reductions in the cost of having children. If the new Family Policy1174.6. ResultsGiven this study focuses exclusively on childcare, treatment will be considered primarilyin terms of changes in childcare arrangements. Specifically, changes in the following types ofchildcare arrangements will be investigated to determine the proportion of children treated:i) Any type of care, ii) Institutional care, and iii) Institutional or licensed care outside thehome. While these treatment indicators are the focus of the study, it should be kept in mindthat changes in family income, labour supply decisions, the childcare curriculum, and theintensity of care may each play a role in shaping changes in child development and healthoutcomes associated with the new Family Policy.To determine the proportion of children that is “treated,” Equation (4.8) was estimatedusing a dummy variable for whether the child is in a particular childcare arrangement as thedependent variable. Given that the proportion of children who is treated might vary acrossMSD/PPVT-R/body weight quantiles, separate regressions are estimated for children basedon their MSD/PPVT-R/body weight. In particular, children are ranked by comparing theirscores with those of other children in the same cycle/province cell. Then, the proportion ofchildren who is treated is examined separately for those at the 10th, 25th, 50th, 75th, and90th percentiles of the cycle/province cells. So as to ensure that the sample size is sufficientlylarge for this analysis, children with scores within 5 percentiles above and below the thresholdof interest are included in the analysis. Thus, to determine the proportion of children treatedat the 10th percentile, children who are between the 5th and 15th percentiles within theircycle/province cell are included in the regression, while to determine the proportion of childrentreated at the 50th percentile, children between the 45th and 55th percentiles within thecycle/province cell are included. This was done separately for MSD, PPVT-R, and bodyweight. It should be noted that given the childcare reform plausibly affected the compositionof Quebec children at a particular point along the distribution of outcomes between the preand post-reform periods, it makes it difficult to directly compare children over time betweenthe treatment and comparison groups based on their percentile rank. Ideally, longitudinaldata would be better suited for this type of analysis with separation done on pre-reformMSD/PPVT-R/body weight; however, the necessary data to do this aren’t available in theNLSCY. Thus, cross-sectional data is only used and the analysis is carried out as describedabove; although, such compositional changes should be kept in mind when examining thederived ITT scaling factor for a given threshold level.increased fertility, there could be a reduction in investment per child among families with children since familyincome and parenting time would have to be spread among more children. In turn, this could lead to worsechild developmental and health outcomes. Conversely, it may be that more children per family leads to greaterparenting experience, which in turn improves parenting skills and results in better child outcomes, particularlyamong younger children. This could offset the reductions in income and time spent per child. Furthermore, therelationship between the number of children in a household and child developmental outcomes may be highlynon-linear. For example, in large households, older siblings may begin to assist and mentor younger siblingsonce family size increases. While fertility is not examined in this study, it is not clear exactly what impactfertility changes due to the new Family Policy would have on child health and developmental outcomes.1184.6.ResultsTable 4.5: Child Care Use Results by MSD, PPVT-R, and Body Weight PercentilesEstimates by MSDQuantile Estimates by PPVT-R Quantile Estimates by Body Weight QuantileIn Care Institutional Institutional/ In Care Institutional Institutional/ In Care Institutional Institutional/Care Licensed Care Care Licensed Care Care Licensed CareA. 10th QuantileELIG Dummy 0.12*** 0.19*** 0.21*** 0.11 0.08 0.04 0.17** 0.13*** 0.16***[0.03] [0.01] [0.02] [0.07] [0.05] [0.06] [0.08] [0.03] [0.05]No.of Obs 2,895 2,895 2,899 574 574 575 3,684 3,684 3,693B. 25th QuantileELIG Dummy 0.19*** 0.15*** 0.21*** 0.08* 0.08 0.20*** 0.20*** 0.20*** 0.25***[0.03] [0.02] [0.04] [0.05] [0.05] [0.04] [0.05] [0.02] [0.04]No.of Obs 2,857 2,857 2,866 588 588 589 3,763 3,763 3,777C. 50th QuantileELIG Dummy 0.13*** 0.13*** 0.19*** 0.17 0.34*** 0.36*** 0.17*** 0.14*** 0.19***[0.03] [0.03] [0.02] [0.12] [0.05] [0.06] [0.03] [0.02] [0.03]No.of Obs 2,923 2,923 2,926 586 586 588 2,702 2,702 2,711D. 75th QuantileELIG Dummy 0.27*** 0.23*** 0.31*** 0.11* 0.19** 0.27** 0.03 0.14*** 0.19***[0.05] [0.03] [0.04] [0.06] [0.09] [0.11] [0.05] [0.05] [0.05]No.of Obs 2,843 2,843 2,850 543 543 543 2,913 2,913 2,919E. 90th QuantileELIG Dummy 0.19*** 0.16*** 0.29*** 0.08 0.16*** 0.25*** 0.11*** 0.23*** 0.29***[0.04] [0.02] [0.02] [0.08] [0.05] [0.07] [0.02] [0.04] [0.03]No.of Obs 2,561 2,561 2,565 587 587 589 2,662 2,662 2,672F. Mean (Standard DD)ELIG Dummy 0.15*** 0.15*** 0.20*** 0.15*** 0.15*** 0.20*** 0.15*** 0.15*** 0.20***[0.03] [0.03] [0.03] [0.03] [0.03] [0.03] [0.03] [0.03] [0.03]No.of Obs 33,702 33,702 33,878 33,702 33,702 33,878 33,702 33,702 33,878Notes: Each column represents different dependent variables on child care arrangement. Each panel represents separate samples included in the estimation. The children are groupedinto samples based on their MSD/PPVT-R/Body Weight rank within province/cycle cells. The sample for a given percentile is different when grouping is done for MSD (left mostcolumns), PPVT-R (centre columns), and Body Weight (right most columns). For each dependent variable, the coefficient on the ELIG dummy is reported for separate regressionswith different samples. Also included in the regressions are a set of control variables including dummies for the child’s age and gender, number of older and younger siblings, mother’sage and education, father’s age and education, mother and father’s immigration status, the size of the urban area, NLSCY cycle dummies, and province dummies. Standard errors arein brackets and were clustered by province and cycle. * significant at 10%; ** significant at 5%; *** significant at 1%.1194.6. ResultsTable 4.5 provides estimates of the proportion of children who are treated using the threedifferent interpretations of being “treated” as discussed above. This table shows that theproportion treated (i.e. the coefficient on the ELIG dummy) derived by separating the samplebased on MSD scores varies from 12%, with “In Care” as the dependent variable at the 10thpercentile, to 31% in the case where “Institutional/Licensed Care” is the dependent variableand the 75th percentile is considered. Additionally, the coefficients are all significant atthe 99% confidence level and of the expected positive signs when splitting is done by MSDpercentiles. When separate regressions are estimated for children split by PPVT-R scores,the proportion treated ranges from 4% in the case where “Institutional/Licensed Care” is thedefinition of treatment at the 10th percentile to 36% in the case of “Institutional/LicensedCare” care being estimated for the 50th percentile of PPVT-R scores. Note, however, thatthe estimates for PPVT-R are statistically significantly different than zero less often thanthe case when MSD scores are considered, which is likely the result of larger standard errorson these coefficients which can partly be explained by a smaller sample of 4-year olds. Theestimates for body weight are similar in size as the MSD estimates and are derived from thefuller sample of all children under 5 years. Almost all estimates are statistically different thanzero.The average proportion of children who are treated includes all children in the analysis.Here, the proportion of children who are treatment varies from 15-20% based on the definitionof treatment and the estimates are significant. It can thus be concluded that the Quebec 1997universal childcare policy raised the proportion of children in care, in institutional care, aswell as institutional/licensed care in Quebec. In their study, BGM also examine broadermeasures of treatment for their analysis. They find that if treatment is defined as exposureto maternal labour supply, then 7.7% of children are treated. If it is defined as the decreasein mothers caring for their children but not working, then 10% of children are treated.129Throughout this paper, treatment will be considered as the change in the proportion ofchildren in “Institutional/Licensed Care” (at a given threshold level), and the scaling factorfor the ITT that gives the TT effect at a given quantile is consequently the inverse of thisproportion.4.6.2 The Full Sample AnalysisThe results from the standard DID model (Equation 4.8 ) are provided in Table 4.6 and arebenchmarked against the BGM estimates. This table shows that the Family Policy had astatistically significant negative impact on the average MSD scores of Quebec children, but129BGM also find that the policy led to different effects by family socio-economic background. Since incomeis endogenous, they split the sample by parental education group and find that the reform led to strongereffects on childcare utilization and labour supply for the higher education group. However, they find mixedevidence of differences in average child outcomes across education groups.1204.6. Resultsno effect on PPVT-R scores. Specifically, the reform resulted in a reduction of MSD scores forchildren aged 0-3 years by 1.64 points, which is 11% of a standard deviation in MSD scores.When this result is scaled by the ITT factor of 5, the average treatment on the treated effectis estimated to be quite large at -8.21 points, which is reduction of nearly 55% of a standarddeviation. Although the change in the PPVT-R scores is insignificant, the relatively smallnumber of 4-year olds in the sample contributes to the imprecision. As can be seen fromthis table, the estimates obtained in this study are very similar to those in the BGM study.This table also shows that, on average, the universal childcare program led to a reduction inchildren’s body weight of nearly 0.3 kilograms. When scaled by the ITT factor, the averagetreatment on the treated effect is estimated to be quite sizeable at 1.44 kilograms, which is adecline of nearly 35% of a standard deviation.Table 4.6: Standard DID EstimatesTimmins BGMMSD PPVT-R Body Weight MSD PPVT-REligible -1.642*** 0.406 -0.288*** -1.647*** 0.36[0.473] [0.740] [0.065] [0.46] [0.75]Observations 26,036 5,198 32,141 26,176 5,210ITT 5 5 5 7-13 7-13Notes: The first two columns give the results of the present study (Timmins) while the last two showthe results of BGM. Within each set of results, the two columns represent different dependent variables ofdevelopmental outcomes (MSD, PPVT-R, or Body Weight). For each dependent variable, the coefficienton the ELIG dummy is reported for the standard DID estimation. Also included in the regressions area set of control variables including dummies for the child’s age and gender, number of older and youngersiblings, mother’s age and education, father’s age and education, mother and father’s immigration status,the size of the urban area, NLSCY cycle dummies, and province dummies. Standard errors are in bracketsand were clustered by province and cycle. The ITT scaling factor is the inverse of the proportion of childrentreated, where the present study defines treatment in terms of Institutional/Licensed Care, while the BGMstudy defines treatment in multiple ways, with the range of the proportion treated being bound between7.7% and 14.6% of children. * significant at 10%; ** significant at 5%; *** significant at 1%.The quantile treatment effect estimates using FFL are given in Figures 4.1 to 4.3 for MSDscores, PPVT-R scores, and body weight respectively, along with the 95% confidence intervalbands (the dashed lines). Figure 4.1 shows that the Family Policy had little effect on theQuebec distribution of MSD scores. The lower and upper ends of the distribution appearto have experienced a negative effect of the program, while the middle of the distributionexperienced a very small positive effect on the percentile values. However, as the confidenceintervals show, these estimated effects are not significantly different than zero at the 95%confidence level. Figure 4.2 shows that for most of the distribution, the Quebec reform had anegative effect on the PPVT-R quantiles, particularly around the 75th threshold level. Thesenegative effects are not significant across most of the distribution, with the smaller samplesize likely contributing to the imprecision of the estimates. However, the negative effect of1214.6. Resultsthe policy is significant at the 95% level around the 75th threshold level. Figure 4.3 showsa significant decline in body weight for those thresholds above the median, with a size thatgrows with the threshold.The first rows in Tables 4.7 to 4.9 show the estimated quantile treatment effects using FFLon Quebec’s distribution at selected threshold levels. As is aligned with the results from thefigures, the estimated effects are not significant at any of the threshold levels for MSD scores,and only the estimate on PPVT-R scores at the 75th threshold level is statistically significant.It should be noted that the standard errors on the PPVT-R estimated coefficients are quitelarge in comparison to the MSD coefficients. For example, for the MSD scores at the 25thpercentile, the standard errors on the estimated coefficient would imply that any (absolute)estimate of 1.33 points or greater (0.68*1.96) would be detected as being statistically differentthan zero at the 95% confidence level. This amounts to an effect size of 8.89% of a standarddeviation in MSD scores. However, at the 25th percentile of PPVT-R scores, the standarderrors on the estimated coefficient would require an estimate of 2.20 in absolute value tobe detected, which is nearly 15% of a standard deviation in PPVT-R scores. This generalpattern can also be seen across the other threshold levels besides the 25th, where the largerstandard errors put stricter requirements on the size of the estimated coefficients for statisticalsignificance compared to in the MSD analysis. As mentioned above, the larger imprecisionof the PPVT-R can be explained in part by differences in the sample sizes across the twoestimations (26,036 children aged 0-3 years for MSD and 5,198 children aged 4 years forPPVT-R).1224.6. ResultsFigure 4.1: FFL Estimates for MSD-10-505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-PolicyNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method. The dependent variableis the Motor and Social Development (MSD) Scale. The dashed lines are the95% confidence intervals.Figure 4.2: FFL Estimates for PPVT-R-10-505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-PolicyNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method. The dependent variableis the Peabody Picture Vocabulary Test-Revised (PPVT-R). The dashedlines are the 95% confidence intervals.1234.6. ResultsFigure 4.3: FFL Estimates for Body Weight-2-1012Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-PolicyNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method. The dependent variableis child body weight in kilograms. The dashed lines are the 95% confidenceintervals.1244.6.ResultsTable 4.7: Quantile Treatment Effects for MSDSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenFFL -1.25 -0.59 0.26 -0.24 0.04 26,036[0.92] [0.68] [0.54] [0.60] [0.50]Firpo -0.28 -0.22 -0.52 0.35 0.04 26,036[1.06] [0.85] [0.68] [0.68] [0.57]FFL- Collapsed Time Periods -1.24 -0.58 0.26 -0.25 0.04 26,036[0.93] [0.68] [0.54] [0.60] [0.50]FFL-DDD 4.07* 1.59 0.89 1.14 -1.23 35,397[2.12] [1.29] [1.20] [1.39] [1.05]ITT Scaling Factor: 4.76 4.76 5.26 3.23 3.45Notes: Each column represents different a threshold level for MSD scores. Each row represents a different estimation strategy. The coefficient on the ELIG dummy isreported for each threshold value and estimation strategy with robust standard errors in brackets. Standard errors were bootstrapped by resampling from the originalestimation sample 100 times. A set of control variables are included in all estimation techniques including dummies for the child’s age and gender, number of older andyounger siblings, mother’s age and education, father’s age and education, mother and father’s immigration status, the size of the urban area, NLSCY cycle dummies,and province dummies. The ITT scaling factor is the inverse of the proportion of children treated for the given threshold level, where treatment is defined in terms ofInstitutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at 1%.1254.6.ResultsTable 4.8: Quantile Treatment Effects for PPVT-RSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenFFL -0.24 -1.51 -1.01 -3.22** 0.48 5,198[1.89] [1.12] [1.06] [1.43] [1.60]Firpo -0.99 -1.52 -2.14 -2.89 0.12 5,198[3.73] [1.77] [1.61] [1.94] [1.87]FFL- Collapsed Time Periods -0.24 -1.51 -1.04 -3.15** 0.52 5,198[1.89] [1.22] [1.22] [1.43] [1.60]ITT Scaling Factor: 25 5 2.78 3.70 4Notes: Each column represents different a threshold level for PPVT-R scores. Each row represents a different estimation strategy. The coefficient on the ELIG dummyis reported for each threshold value and estimation strategy with robust standard errors in brackets. Standard errors were bootstrapped by resampling from the originalestimation sample 100 times. A set of control variables are included in all estimation techniques including dummies for the child’s age and gender, number of older andyounger siblings, mother’s age and education, father’s age and education, mother and father’s immigration status, the size of the urban area, NLSCY cycle dummies,and province dummies. The ITT scaling factor is the inverse of the proportion of children treated for the given threshold level, where treatment is defined in terms ofInstitutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at 1%.1264.6.ResultsTable 4.9: Quantile Treatment Effects for Body WeightSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenFFL -0.13 -0.12 -0.10 -0.24** -0.26** 32,141[0.12] [0.09] [0.08] [0.10] [0.12]Firpo 0.12 -0.18 -0.75*** -0.58*** -0.47*** 32,141[0.26] [0.2] [0.22] [0.16] [0.23]FFL- Collapsed Time Periods -0.13 -0.11 -0.09 -0.24** -0.26*** 32,141[0.12] [0.09] [0.08] [0.10] [0.12]FFL-DDD 0.01 -0.07 0.06 0.06 -0.04 42,779[0.19] [0.15] [0.12] [0.21] [0.29]ITT Scaling Factor: 6.25 4.00 5.26 5.26 3.45Notes: Each column represents different a threshold level for body weight. Each row represents a different estimation strategy. The coefficient on the ELIG dummy isreported for each threshold value and estimation strategy with robust standard errors in brackets. Standard errors were bootstrapped by resampling from the originalestimation sample 100 times. A set of control variables are included in all estimation techniques including dummies for the child’s age and gender, number of older andyounger siblings, mother’s age and education, father’s age and education, mother and father’s immigration status, the size of the urban area, NLSCY cycle dummies,and province dummies. The ITT scaling factor is the inverse of the proportion of children treated for the given threshold level, where treatment is defined in terms ofInstitutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at 1%.1274.6. ResultsTable 4.9 shows the FFL estimates for body weight. I find evidence there is a declineat all thresholds of the distribution, however they increase in size and become statisticallysignificant at the 95% confidence level with higher thresholds. I find a decline of 0.24 kg at the75th percentile and 0.26 kg at the 90th percentile. When scaled by the relevant ITT factor,this amounts to a decline of 1.26 kg (31% of a standard deviation) for the 75th percentileand 0.90 kg (22% of a standard deviation) for the 90th threshold. Relative to the mean,these effects are sizeable. They are also sizeable relative to the World Health Organizationstandards, with a mean weight-for-age of 18.17 for boys and 18.04 for girls for a 5 year old(World Health Organization, 2006).Figure 4.4: Firpo Estimates for MSD, PPVT-R, Body Weight-10-505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) MSD-10 -505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(b) PPVT-R-2-1012Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(c) Body WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the Firpo method. Panel (a) has MSD asthe dependent variable, panel (b) has PPVT-R as the dependent variable,and panel (c) has body weight as the dependent variable. The dashed linesare the 95% confidence intervals.The robustness of these results is verified against those obtained with the Firpo method.Given that the Firpo estimator is more flexible in how it conditions on covariates, we do notexpect the two approaches to yield the exact same results. Panel a) from Figure 4.4 shows theeffect of the program on the distribution of MSD scores using the Firpo estimation strategy.Just as in the FFL method, the program tends to have a small negative impact on MSDscores at the lower threshold level before giving a positive impact at around the 40th thresholdlevel. Here, the estimates tend to be much smaller in absolute size than the FFL method.Nonetheless, just as in the FFL approach, these estimates are not significantly different than1284.6. Resultszero, with the exception of just a few points which are only marginally significant. Panelb) shows the results for the PPVT-R distribution using the Firpo method. Similar to theFFL results, the effect of the program is negative for most of the distribution, becomingincreasingly negative at higher threshold levels before turning slightly positive at the upperend and then negative again thereafter. Once more, however, the estimates are too impreciseto give any significant effects, as shown by the large confidence intervals. Panel c) shows theresults for the body weight distribution using the Firpo method. Similar to the findings fromFFL approach, the effects appear larger at the highest threshold levels. The second row inTables 4.7 thru 4.9 provide the quantile treatment effect estimates at selected threshold levelsusing the Firpo method, which are all shown to be statistically insignificant for MSD andPPVT-R scores. For body weight, however, again the higher threshold levels experience thelargest reductions in body weight. The size of the estimates using this method are larger thanusing FFL and approximately double in size (0.58 vs 0.24 kg for the 75th percentile and 0.47vs 0.26 for the 90th percentile).One concern with DID estimation is the correct computation of standard errors. AsBertrand, Duflo, and Mullainathan (2004) point out, there can often be serial correlationproblems that can lead to gross overrejection rates if not accounted for. In particular, theynote that the problem is most severe when the time series are long, when the dependentvariables are of the type that are highly positively serially correlated, and when the treatmentstatus changes little over time within a province. While the first two circumstances are lessapplicable to the analysis here, with only four time periods and dependent variables relatingto child development, the last point may be of concern. They note that aggregating the databy collapsing it into only two time periods, namely before and after the policy, works wellwhen the number of groups is small, as in the case here. The results of collapsing the timeperiods and using FFL are shown in the third rows of Tables 4.7 to 4.9. Comparing theseestimates with the FFL results in the first rows, it can be seen that aggregating the data intocollapsed time periods makes little difference in terms of the standard errors or the estimates.Again, the bulk of the estimates remain insignificant, while the PPVT-R distribution at the75th percentile of the Quebec distribution is still negatively affected by the policy. For bodyweight, the estimates are identical and are of the same level of statistical significance. Figure4.5 shows the impact of the policy across the entire distribution of outcomes for MSD (panela), PPVT-R (panel b), and body weight (panel c) in Quebec using the collapsed time periods,with the results being nearly identical to those found in Figures 4.1 to 4.3.Another estimation strategy used to test the robustness of the estimates in this study isa triple difference model (DDD). Since the introduction of the Family Policy was staggeredacross age groups, this permits an added dimension of variation to be exploited by comparingacross age groups. Given that the NLSCY cycle 3 sampling occurred from October 1998 -June 1999 and that only three and four year olds were eligible for subsidized care at this time,1294.6. Resultsanother potential control group in the same province can be used to evaluate the programfor the third cycle (two year olds and younger in Quebec who were not eligible at this time).The advantage of the triple difference model is that, in addition to province and time fixedeffects encompassed in the QDID model, as well as age fixed effects, differential time trendsare allowed to exist across provinces and age groups, as are differential age fixed effects acrossprovinces. What this approach does put a restriction on, however, is that there cannot be anydifferential time trend for children of different ages who live in the same province besides thoseaccounted for above. That is, this estimator will be unbiased if the effect of age on outcomesdoes not shift differentially between the pre and post reform periods in Quebec versus the restof Canada. It is the omission of this three-way interaction that identifies the model. Giventhat MSD scores were only obtained for children up to four years of age and that PPVT-Rscores were only obtained for children aged 4-5 years, only MSD scores for children aged 3and under can be included in the DDD analysis. The identifying assumption of the DDDestimation may be less innocuous for body weight compared to MSD. In particular, given itsbiological nature, there may be unobserved factors that differentially affect the rate of growthfor different ages across provinces over time.Figure 4.5: FFL Estimates with Collapsed Time Periods-10-505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) MSD-10 -505 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(b) PPVT-R-2-1012Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(c) Body WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method with collapsed pre andpost time periods. Panel (a) has MSD as the dependent variable, panel (b)has PPVT-R as the dependent variable, and panel (c) has body weight asthe dependent variable. The dashed lines are the 95% confidence intervals.1304.6. ResultsEstimation was just as in Equation (4.11) but now age fixed effects are included as areall second order interactions between age, province, and time period so that the estimatingequation is:R̂IFi(Yi, qτ , FY ) = ατ +∑5k=1 γτkTki +∑10j=1 δτjGji +∑3s=0 ρτsasi +∑3s=0∑5k=1 µτskTkiasi+∑10j=1∑5k=1 piτjkTkiGji +∑10j=1∑3s=0 ϕτjsasiGji + θτELIGi +Xiβτ + τi(4.15)Here as represents age and s ∈ {0, 1, 2, 3}; ρτs is the coefficient associated with age as atthe τ -th threshold; and µτsk, piτjk, ϕτjs are the coefficients on the second order interactions.Still, interest lays in the coefficient on the ELIGi variable θτ .A word of caution using the DDD in this study is that it heavily depends on variationwithin NLSCY cycle 3 data. However, as mentioned previously, the roll out of the subsidizedchildcare spaces was quite slow in the initial period. In particular, there were severe capacityissues in that the demand for places greatly surpassed the available supply in the early years.Consequently, most of the initial subsidized spaces were created in centres and family homecare settings that already existed prior to the introduction of the policy and, as alreadypointed out, it was the children already receiving subsidized care who obtained priority inobtaining the limited number of available spaces. As such, it was likely that in the infancy ofthe universal childcare program, the same children had access to subsidized care in the samefacilities that existed prior to the policy, with little changes in the staff, location, and physicalenvironment. It was for these exact reasons that the third cycle of the NLSCY was droppedfrom the main empirical analysis. However, in order to use the DDD model, this cycle mustbe included to get the differential roll out of the policy across age groups. Thus, the results ofthis estimation must be examined with caution as it’s quite possible that a large proportionof 3-year olds observed in cycle 3 had limited access to subsidized care in practice and littlechanges in the arrangement for those who did relative to prior to the policy.The results for the DDD estimation strategy on the distribution of MSD scores and bodyweight are provided in Figure 4.6 and in the fourth row of Tables 4.7 and 4.9. As can beseen, the DDD strategy leads to different estimated coefficients compared to the QDID modeland larger standard errors. For MSD, this model produces positive effects for the bulk ofthe distribution before turning negative around the 80th percentile. However, the confidenceinterval bands of the DDD estimates are quite large, resulting in insignificant estimates acrossthe distribution, with the exception being at the 10th percentile where the estimate is onlymarginally significant and positive. For body weight, this model shows no change in bodyweight across the distribution. For the reasons discussed above, however, this is not thepreferred estimation model for the analysis.1314.6. ResultsFigure 4.6: FFL DDD Estimates for MSD and Body Weight-10-5 05 10Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) MSD-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(b) Body WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method in the DDD specification.Panel (a) has MSD as the dependent variable and panel (b) has body weightas the dependent variable. The dashed lines are the 95% confidenceintervals.4.6.3 The Subsample AnalysisTo investigate whether the same general pattern of insignificant effects of the reform on thedistribution of outcomes holds across children with different demographic and family charac-teristics, subsample analyses were carried out. In particular, Equation (4.11) was estimatedwith separate samples. The effect of the reform on boys versus girls was examined, as wereany differential effects across parents’ education. In this analysis, a parent is considered loweducated if he/she has a high school diploma or less and high educated otherwise. The exis-tence of differential impacts across father’s wage income is also examined, with a father beingconsidered as having a ”low wage income” if he is in the bottom 30% of his province/cyclecell and a ”high wage income” otherwise.130 Finally, for each measure of parenting skills andthe family functioning measure, the subsample is divided by separating those in the top 30%of poorer skills/interactions from the remaining 70% with stronger skills/interactions. Thissame approach was taken for maternal depression, with the sample separated by the top 30%of mothers with more depressive symptoms from the remaining 70% with lower depression.130Given that the reform likely had an effect on maternal labour supply and consequently household income,only father’s income, which is presumably more exogenous to the analysis, was examined for the subsampleanalyses.1324.6. ResultsAs was discussed previously, however, it is possible that the reform affected parenting skills,family functioning, and maternal depression, possibly through increased stress associated withgreater rates of two parent employment, as was suggested by BGM. As such, comparing chil-dren at a given point along the parenting skills/family functioning distribution before andafter the reform may be misleading if the composition of individuals at this point differen-tially change across treatment and comparison groups over time. Consequently, caution mustbe taken when interpreting the results. For the subsample analyses based on differences ingender, parental age, education, and father’s wage income, this is less of a concern as thesevariables are largely pre-determined.Table 4.10 shows the effects on particular MSD percentiles for the subsample analyses,while Tables 4.11 and 4.12 show them for PPVT-R and body weight percentiles. Figures 4.7to 4.15 provide the estimates and confidence bands across the entire distribution for bothoutcomes. While there are slight differences across the estimates for boys and girls for givenMSD percentiles, none of the estimates are statistically different from zero. The upper end ofthe PPVT-R distribution for girls is negatively impacted by the reform for a small amount ofthreshold levels, with the effect being significant at the 75th threshold level at -4.53, whichis 30% of a standard deviation in PPVT-R scores. For body weight, there are no sizeabledifferences across the distribution. However, the decline in body weight is almost double thesize for boys than girls at the 90th percentile (see Table 4.12).In general, the reform had no differential impact when the sample is split by maternaleducation. The bands forming the confidence region are particularly wide for the sample withlow educated mothers and PPVT-R scores are the outcome of interest. For body weight, themost notable difference between children from less and more educated mothers is at the higherthresholds (90th threshold and over), with children from higher educated mothers having alarger decline. Although the Quebec MSD distribution is not significantly affected by thereform for neither low or high educated fathers, PPVT-R scores are. In particular, the reformled to significant negative impacts on the middle portion of the PPVT-R distribution forchildren of low educated fathers (the 30th percentile to 60th percentile) as can be seen inpanel c) of Figure 4.9. At the 50th percentile, the estimated impact is -5.88 points, which isequivalent to almost 40% of a standard deviation of PPVT-R scores. For body weight, theeffects are concentrated at the upper part of the distribution for children from both low andhigh educated fathers. Figure 4.10 reveals that the reform had a significant negative impacton the lower end of MSD scores for children of low wage income fathers (panel a), with aninsignificant effect on the distribution of MSD scores for children of high wage fathers or onthe distributions of PPVT-R scores by father wage income. At the 10th percentile of MSDscores for children of low wage income fathers, the ITT estimate of the impact of the reformis -5.04 points, which is one third of a standard deviation. This is significant at the 99% level.At the 25th percentile of this distribution, the effect size is estimated to be -3.28 points or1334.6. ResultsTable 4.10: Quantile Treatment Effects for MSD - Subgroup AnalysisSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenBoys -1.32 -0.34 0.44 0.75 -0.18 13,254[1.27] [0.97] [0.78] [0.72] [0.71]Girls -2.07 -1.47 -0.51 -0.46 0.18 12,782[1.37] [0.98] [0.74] [0.68] [0.61]Mom Low Educated -2.94 0.07 1.06 0.16 -1.36 7,645[2.04] [1.33] [1.10] [1.00] [0.89]Mom High Educated -0.81 -0.77 -0.09 -0.46 0.49 18,391[1.12] [0.74] [0.62] [0.64] [0.52]Dad Low Educated -3.21* -0.37 1.11 0.27 0.38 8,948[1.73] [1.17] [0.93] [0.94] [0.83]Dad High Educated -0.45 -0.68 -0.27 -0.55 -0.18 17,088[1.17] [0.77] [0.67] [0.65] [0.53]Dad Low Income -5.04*** -3.28** -1.87 -0.29 1.09 5,595[1.90] [1.37] [1.28] [1.18] [1.06]Dad High Income -1.78 -1.56 -0.05 0.29 0.52 13,095[1.43] [0.99] [0.84] [0.8] [0.68]Hostile Parenting -4.33* -0.85 1.40 1.64 0.06 3,452[2.60] [1.87] [1.49] [1.35] [1.17]Non-Hostile Parenting -1.68 -1.46 0.59 0.54 0.81 9,940[1.72] [1.14] [0.74] [0.71] [0.57]Inconsistent Parenting -2.86 0.68 2.43* 1.30 0.76 4,622[2.38] [1.64] [1.30] [1.14] [0.98]Consistent Parenting -3.70** -2.66** 0.39 0.35 0.35 8,614[1.85] [1.22] [0.77] [0.78] [0.60]Aversive Parenting 1.86 2.95 1.42 2.34 2.44* 3,144[3.54] [2.34] [1.70] [1.50] [1.42]Non-Aversive Parenting -2.73* -1.41 0.67 0.92 0.77 10,318[1.59] [1.13] [0.70] [0.68] [0.54]Dysfunctional Family -3.08 -0.05 1.98 -0.34 -0.02 4,966[2.22] [1.65] [1.37] [1.29] [1.27]Non-Dysfunctional Family -1.16 -0.93 0.06 -0.37 0.07 20,495[1.11] [0.70] [0.60] [0.60] [0.49]High Maternal Depression -2.00 -2.55* 0.44 -1.21 -0.36 6,108[2.11] [1.45] [1.19] [1.05] [0.97]Low Maternal Depression -0.79 -0.68 0.37 -0.02 -0.02 16,380[1.22] [0.81] [0.66] [0.67] [0.56]ITT Scaling Factor: 4.76 4.76 5.26 3.23 3.45Notes: Each column represents different a threshold level for MSD scores. Each row represents a different subsample. The coefficient on the ELIG dummy usingFFL is reported for each threshold value and subsample with robust standard errors in brackets. A set of control variables are included in all estimation techniquesincluding dummies for the child’s age and gender, number of older and younger siblings, mother’s age and education, father’s age and education, mother and father’simmigration status, the size of the urban area, NLSCY cycle dummies, and province dummies. The ITT scaling factor is the inverse of the proportion of childrentreated for the given threshold level, where treatment is defined in terms of Institutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at1%.1344.6. ResultsTable 4.11: Quantile Treatment Effects for PPVT-R - Subgroup AnalysisSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenBoys 1.06 -1.21 -1.38 -1.48 0.42 2,561[2.36] [1.64] [1.70] [2.46] [2.38]Girls -0.55 0.47 -1.02 -4.53** 0.31 2,637[2.49] [1.77] [1.69] [2.09] [2.33]Mom Low Educated -1.68 -0.44 -2.72 -5.39* -2.99 1,577[3.75] [2.17] [2.07] [2.96] [2.99]Mom High Educated -0.07 0.71 -0.46 -2.27 0.63 3,621[1.83] [1.40] [1.43] [1.83] [1.92]Dad Low Educated -2.28 -2.23 -5.88*** -3.31 -2.18 1,839[3.12] [2.11] [1.89] [2.55] [2.80]Dad High Educated 0.71 1.66 1.36 -1.75 1.12 3,359[1.92] [1.44] [1.52] [1.96] [2.01]Dad Low Income 1.38 -3.13 -1.25 -3.81 -0.75 964[5.21] [3.67] [3.09] [4.92] [3.82]Dad High Income 0.04 0.18 -1.06 -3.52 -0.20 2,641[2.41] [1.78] [1.93] [2.29] [2.40]Hostile Parenting -1.98 -0.98 -1.34 -2.15 -1.67 1,206[3.43] [2.46] [2.53] [3.31] [3.41]Non-Hostile Parenting 0.09 -1.23 -0.53 -3.00 2.12 3,947[2.01] [1.40] [1.41] [1.83] [1.89]Inconsistent Parenting -1.63 -5.48** -2.88 -5.48* 0.18 1,598[2.93] [2.32] [2.28] [3.11] [3.17]Consistent Parenting 1.44 0.36 -0.46 -1.91 1.19 3,523[1.87] [1.40] [1.46] [1.82] [1.87]Aversive Parenting 1.41 -2.55 -0.07 1.08 3.02 1,037[3.84] [3.10] [2.74] [3.81] [4.25]Non-Aversive Parenting -0.69 -1.62 -0.91 -3.36* 0.59 4,135[1.96] [1.31] [1.37] [1.72] [1.82]Dysfunctional Family -3.16 -5.55 1.92 6.91 5.48 883[4.36] [3.50] [3.19] [4.90] [5.17]Non-Dysfunctional Family -0.74 -1.69 -1.75 -3.97** -0.05 4,254[1.92] [1.34] [1.36] [1.74] [1.79]High Maternal Depression -3.65 0.65 -0.21 -2.91 -0.07 1,126[3.84] [2.70] [2.66] [3.90] [3.95]Low Maternal Depression 1.54 -1.16 0.09 -2.65 -0.81 3,638[1.91] [1.41] [1.46] [1.84] [1.93]ITT Scaling Factor: 25.00 5.00 2.78 3.70 4.00Notes: Each column represents different a threshold level for PPVT-R scores. Each row represents a different subsample. The coefficient on the ELIG dummy usingFFL is reported for each threshold value and subsample with robust standard errors in brackets. A set of control variables are included in all estimation techniquesincluding dummies for the child’s age and gender, number of older and younger siblings, mother’s age and education, father’s age and education, mother and father’simmigration status, the size of the urban area, NLSCY cycle dummies, and province dummies. The ITT scaling factor is the inverse of the proportion of childrentreated for the given threshold level, where treatment is defined in terms of Institutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at1%.1354.6. ResultsTable 4.12: Quantile Treatment Effects for Body Weight - Subgroup AnalysisSpecification 10th Percentile 25th Percentile 50th Percentile 75th Percentile 90th Percentile No. of ChildrenBoys -0.25 -0.19 -0.21 -0.34* -0.63** 16,318[0.20] [0.13] [0.13] [0.19] [0.27]Girls -0.17 -0.08 -0.16 -0.20 -0.34** 15,823[0.20] [0.13] [0.12] [0.14] [0.17]Mom Low Educated -0.01 -0.17 -0.26* -0.67*** -0.09 9,404[0.25] [0.17] [0.15] [0.22] [0.25]Mom High Educated -0.12 -0.10 -0.05 -0.09 -0.25** 22,737[0.16] [0.10] [0.09] [0.12] [0.12]Dad Low Educated -0.01 -0.08 -0.07 -0.50*** -0.53** 11,052[0.23] [0.17] [0.14] [0.19] [0.22]Dad High Educated -0.11 -0.13 -0.12 -0.10 -0.15 21,089[0.17] [0.11] [0.10] [0.13] [0.13]Dad Low Income -0.16 -0.07 0.02 -0.20 -0.03 6,625[0.31] [0.21] [0.19] [0.24] [0.31]Dad High Income 0.15 -0.14 -0.21 -0.40** -0.50*** 16,090[0.22] [0.15] [0.14] [0.16] [0.19]Hostile Parenting 0.23 -0.15 0.20 -0.21 -0.32 4,787[0.27] [0.19] [0.18] [0.23] [0.49]Non-Hostile Parenting -0.16 -0.21* -0.14 -0.11 -0.79*** 14,236[0.16] [0.11] [0.10] [0.12] [0.22]Inconsistent Parenting 0.09 -0.24 -0.24 -0.32 -0.73* 6,350[0.23] [0.18] [0.17] [0.20] [0.38]Consistent Parenting -0.23 -0.20* 0.03 -0.01 -0.72** 12,484[0.18] [0.12] [0.11] [0.13] [0.29]Aversive Parenting -0.02 0.11 0.31 -0.04 -0.39 4,229[0.30] [0.18] [0.22] [0.26] [0.64]Non-Aversive Parenting -0.27 -0.28** -0.15 -0.12 -0.65*** 14,874[0.16] [0.11] [0.10] [0.12] [0.20]Dysfunctional Family -0.33 -0.20 -0.08 -0.34 -0.10 5,993[0.34] [0.22] [0.20] [0.23] [0.29]Non-Dysfunctional Family -0.08 -0.09 -0.12 -0.24** -0.39*** 25,489[0.14] [0.10] [0.09] [0.12] [0.15]High Maternal Depression -0.13 -0.27 -0.33* -0.26 0.14 7,415[0.30] [0.21] [0.18] [0.22] [0.26]Low Maternal Depression 0.04 -0.05 0.03 -0.12 -0.40** 20,709[0.17] [0.11] [0.10] [0.14] [0.17]ITT Scaling Factor: 4.76 4.76 5.26 3.23 3.45Notes: Each column represents different a threshold level for Body Weight. Each row represents a different subsample. The coefficient on the ELIG dummy usingFFL is reported for each threshold value and subsample with robust standard errors in brackets. A set of control variables are included in all estimation techniquesincluding dummies for the child’s age and gender, number of older and younger siblings, mother’s age and education, father’s age and education, mother and father’simmigration status, the size of the urban area, NLSCY cycle dummies, and province dummies. The ITT scaling factor is the inverse of the proportion of childrentreated for the given threshold level, where treatment is defined in terms of Institutional/Licensed Care. * significant at 10%; ** significant at 5%; *** significant at1%.1364.6. Results22% of a standard deviation and is also significant. For body weight, however, the declineis concentrated among the higher quantiles for children from high income fathers (-0.5 kg atthe 90th threshold). Taking into account the ITT scaling factors, these estimates are quitesizeable.Figure 4.7: FFL Estimates by Gender-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Boys: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Girls: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Boys: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Girls: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Boys: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Girls: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by gender. Panels (a) and(b) have MSD as the dependent variable, panels (c) and (d) have PPVT-Ras the dependent variable, and panels (e) and (f) have body weight as thedependent variable. The dashed lines are the 95% confidence intervals.1374.6. ResultsFigure 4.8: FFL Estimates by Mother’s Education-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Mom Low Educated: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Mom High Educated: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Mom Low Educated: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Mom High Educated: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Mom Low Educated: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Mom High Educated: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by mother’s education.Panels (a) and (b) have MSD as the dependent variable, panels (c) and (d)have PPVT-R as the dependent variable, and panels (e) and (f) have bodyweight as the dependent variable. The dashed lines are the 95% confidenceintervals.Figure 4.9: FFL Estimates by Father’s Education-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Dad Low Educated: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Dad High Educated: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Dad Low Educated: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Dad High Educated: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Dad Low Educated: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Dad High Educated: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by father’s education.Panels (a) and (b) have MSD as the dependent variable, panels (c) and (d)have PPVT-R as the dependent variable, and panels (e) and (f) have bodyweight as the dependent variable. The dashed lines are the 95% confidenceintervals.1384.6. ResultsFigure 4.10: FFL Estimates by Father’s Wage-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Dad Low Wage: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Dad High Wage: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Dad Low Wage: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Dad High Wage: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Dad Low Wage: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Dad High Wage: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by father’s income.Panels (a) and (b) have MSD as the dependent variable, panels (c) and (d)have PPVT-R as the dependent variable, and panels (e) and (f) have bodyweight as the dependent variable. The dashed lines are the 95% confidenceintervals.1394.6. ResultsIn terms of parenting skills, there are no differential effects for children of parents in thetop 30% of hostile parenting scores versus the remaining children for neither MSD or PPVT-R distributions. The same is largely true when the sample is split by the degree of aversiveparenting. With the exception of a few threshold levels, the reform had insignificant effects onthe distributions of MSD and PPVT-R of children from parents with more inconsistent par-enting versus consistent parenting. This pattern also holds when the sample is split by familyfunctioning scores and maternal depression, where the confidence intervals are exceptionallylarge for the PPVT-R distribution of children in dysfunctional families. For body weight,there is some evidence that the size and statistical significance varies across groups, with thelargest reduction in the upper quantile occurring for families with non-hostile parenting.Figure 4.11: FFL Estimates by Parenting Style- Hostile-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Hostile Parenting: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Non-Hostile Parenting: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Hostile Parenting: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Non-Hostile Parenting: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Hostile Parenting: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Non-Hostile Parenting: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by parenting style (hostileor not hostile). Panels (a) and (b) have MSD as the dependent variable,panels (c) and (d) have PPVT-R as the dependent variable, and panels (e)and (f) have body weight as the dependent variable. The dashed lines arethe 95% confidence intervals.1404.6. ResultsFigure 4.12: FFL Estimates by Parenting Style- Aversive-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Aversive Parenting: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Non-Aversive Parenting: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Aversive Parenting: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Non-Aversive  Parenting: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Aversive Parenting: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Non-Aversive  Parenting: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by parenting style(aversive or not aversive). Panels (a) and (b) have MSD as the dependentvariable, panels (c) and (d) have PPVT-R as the dependent variable, andpanels (e) and (f) have body weight as the dependent variable. The dashedlines are the 95% confidence intervals.Figure 4.13: FFL Estimates by Parenting Style- Consistent-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Inconsistent Parenting: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Consistent Parenting: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Inconsistent Parenting: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Consistent Parenting: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Inconsistent Parenting: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Consistent Parenting: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by parenting style(consistent or inconsistent). Panels (a) and (b) have MSD as the dependentvariable, panels (c) and (d) have PPVT-R as the dependent variable, andpanels (e) and (f) have body weight as the dependent variable. The dashedlines are the 95% confidence intervals.1414.6. ResultsFigure 4.14: FFL Estimates by Family Functioning-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) Dysfunctional: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Non-Dysfunctional: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) Dysfunctional: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Non-Dysfunctional: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) Dysfunctional: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Non-Dysfunctional: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by family dysfunction.Panels (a) and (b) have MSD as the dependent variable, panels (c) and (d)have PPVT-R as the dependent variable, and panels (e) and (f) have bodyweight as the dependent variable. The dashed lines are the 95% confidenceintervals.Figure 4.15: FFL Estimates by Maternal Depression-10-5 0510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(a) High Maternal Depression: MSD -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99MSD Percentile, Quebec Pre-Policy(b) Low Maternal Depression: MSD-10 -50 510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(c) High Maternal Depression: PPVT-R -10-50510Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99PPVT-R Percentile, Quebec Pre-Policy(d) Low Maternal Depression: PPVT-R-202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(e) High Maternal Depression: Weight -202Child Care Policy Effect 1 10 20 30 40 50 60 70 80 90 99Body Weight Percentile, Quebec Pre-Policy(f) Low Maternal Depression: WeightNotes: This figure shows the quantile difference-in-differences estimate (theITT) at various thresholds using the FFL method by maternal depression.Panels (a) and (b) have MSD as the dependent variable, panels (c) and (d)have PPVT-R as the dependent variable, and panels (e) and (f) have bodyweight as the dependent variable. The dashed lines are the 95% confidenceintervals.1424.7. Discussion4.7 DiscussionThe results above reveal that there was only little heterogeneity in the response to the universalchildcare policy in Quebec across the distribution of motor skills and cognitive outcomes. The75th percentile of PPVT-R scores in Quebec was significantly negatively affected, both for thewhole sample as well as for girls. Additionally, there was a differential response to the policyfor children of low educated fathers, where the middle portion of the PPVT-R distributionexperienced a sizeable decline in scores. The percentiles at lower thresholds of the MSDdistribution for children of low educated fathers were also negatively affected by the policy.These negative effects were not seen for children of high educated nor high wage incomefathers. Besides this handful of negative impacts, there was little heterogeneity across thedistribution of outcomes for the full sample or for the subsample analyses, with the majorityof the estimates of the program impact being statistically insignificant.Conversely, the findings did reveal sizeable declines in body weight, which were concen-trated at the upper end of the distribution and generally increased with the threshold. Thesefindings suggest that the universal childcare program may have altered the mix of componentsthat affect child weight, such as physical activities or food provided. Other factors affectedby the childcare reform, such as higher household income levels due to increased maternalsupply, could also play a role in reducing child body weight. It is unclear the extent to whichspecific factors may have contributed to the decline in child body weight. As discussed ear-lier, there is limited information in the NLSCY on the childcare curriculum or detailed qualityindicators, so I cannot examine the extent to which these were affected. A further challengeis that treatment is unlikely to be random, whether “treatment” is defined as increased ma-ternal labour supply or switching childcare arrangements. That is, the Quebec policy mayhave caused certain families to increase maternal labour supply for unobserved reasons thatare correlated with child body weight. Additionally, the choice of parents to choose particu-lar childcare arrangements and settings may be driven by unobserved factors that also affectchild health outcomes. This makes it challenging to separately identify which specific factorslowered child body weight at the top of the distribution. In my analyses, I circumvent suchendogeneity issues by focusing on the ITT estimates. To date, this the first known study totest if universal childcare has an impact on child body weight.There are multiple reasons for which little differential effects of the universal childcarepolicy are found across the motor skills and cognitive outcome distributions. First, it ispossible that there is quite simply no heterogeneity in the response to the universal childcarepolicy in Quebec in terms of MSD and PPVT-R outcomes. In this case, then the means doan accurate job of accounting for the policy impact. Although child behavioural outcomeswere not examined in this paper for reasons discussed above, BGM find they are significantlyaffected by the policy. It is possible that any heterogeneous response to Quebec’s childcare1434.7. Discussionpolicy is revealed in these outcomes, rather than MSD and PPVT-R scores. However, whilethe confidence interval bands were relatively tight for the MSD analysis which gives moreassurance in the insignificant results, they were not quite as tight for the PPVT-R estimation.This suggests that it may be that the Family Policy had an impact on the distribution ofPPVT-R scores in Quebec, but the power of the test statistics employed in this paper are toolow to reject the null of no effect. Thus, it may be that a larger sample size of 4-years oldsis required to detect any significant effects across the PPVT-R distribution of outcomes byreducing the imprecision of the estimates.Another possibility for which no heterogeneous impacts across the distribution of motorand cognitive outcomes were found is that only the short run effects of the program are beingexamined in this study. Kottelenberg and Lehrer (2014) find heterogeneity in these outcomeswhen they extend the analysis to later years. This suggests that effects occur in the longrun and that the manner in which large scale social programs are implemented could playa considerable role in their impact. Furthermore, they analyze a period after which therewas a price increase in universal childcare (from $5 to $7 per day). Although this increaseis small, if it affected the composition of children in childcare, this could be another sourceof the difference in findings.131 It is also possible that heterogeneity in treatment outcomesdo exist, but they only manifest when the children are older. As discussed above, Havnesand Mogstad (2010) find substantial heterogeneity in the response to a universal childcareprogram in Norway in terms of subsequent labour market and educational outcomes whenthe children were older. This study only considers children 4 years of age and younger, andif the effects of the program are revealed over time, it is unlikely they will be captured in theanalysis here.Changes in childcare quality could also be at play. Recent findings show that the qualityof care is an important determinant of the impact of non-maternal care on child outcomes(Burchinal 2000, Love et al. 2003). One reason perhaps that no heterogeneous effects werefound is that there was little change in the quality of childcare for children before and after thereform. As discussed above, much of the expansion of subsidized care spaces came from alreadyexisting non-profit centres and family homes. As noted, there could be improvements in thequality over time in the longer run. To obtain government funding, CPE’s were required meetrequirements on the educational curriculum, the physical environment, and the education ofthe caregivers. It could be that the reform did not affect the proportion of childcare giversabove that bar. Finally, given that many children moved from maternal care to non-maternalcare with the policy, the relative quality of maternal versus non-maternal care will vary greatlyacross families based on maternal parenting style and the characteristics of non-maternal care131Another source of the difference could be different methodological approaches. As discussed, I use theQDID whereas they use the changes-in-changes estimator. The two approaches impose different underlyingassumptions on the counterfactual distributions.1444.7. DiscussionFigure 4.16: Trends in MSD Quantiles7577 798183 85MSD Quantile 1 2 3 4 5WaveQuebecRest of Canada (a) 10th Quantile 87899193 95MSD Quantile 1 2 3 4 5Wave(b) 25th Quantile98100102 104MSD Quantile 1 2 3 4 5Wave(c) 50th Quantile 106108 110112114MSD Quantile 1 2 3 4 5Wave(d) 75th Quantile114116118120MSD Quantile 1 2 3 4 5Wave(e) 90th Quantile 9698100102104MSD Value 1 2 3 4 5Wave(f) MeanNotes: This figure shows the Motor and Social Development (MSD) Scale atdifferent thresholds across NLSCY cycles for Quebec and the Rest ofCanada (ROC).environment and would be difficult to measure.A final explanation for which heterogeneous impacts of the universal childcare policy inQuebec were not found in motor and cognitive outcomes is that the estimation strategyemployed in this paper to identify these effects is rested on assumptions which do not hold.As outlined above, the key identifying assumption is that a common time trend is assumed tohold at each threshold level, τ , between the treatment and comparison groups in the absenceof treatment. While this cannot be empirically verified in practice since we do not observe thetreated group in the absence of the universal childcare policy, the pre-existing trends betweenthe treatment and control groups can be compared in the years prior to 1997 childcare reformto provide some insight on the appropriateness of this assumption. Figures 4.16 to 4.18 showthe trends in the percentiles at selected threshold levels in Quebec and the rest of Canada forMSD scores, PPVT-R scores, and body weight respectively. As can be seen, the common timetrend appears to be more evident at certain threshold levels (e.g. the 50th percentile of MSDscores, and the 25th, 50th, and 75th percentiles of PPVT-R scores) while it doesn’t appearto hold at others. For body weight, however, the trends are very similar across Quebec andthe rest of Canada in the pre-policy period at all quantiles. Again, while there is no way toverify the accuracy of the common trend assumption, this evidence shows that the assumptionmight be more valid at certain threshold levels and outcome measures than others.1454.7. DiscussionFigure 4.17: Trends in PPVT-R Quantiles7577 798183 85PPVT-R Quantile 1 2 3 4 5WaveQuebecRest of Canada(a) 10th Quantile 86889092 94PPVT-R Quantile 1 2 3 4 5Wave(b) 25th Quantile9698100102104PPVT-R Quantile 1 2 3 4 5Wave(c) 50th Quantile 106108 110112114PPVT-R Quantile 1 2 3 4 5Wave(d) 75th Quantile116118120122PPVT-R Quantile 1 2 3 4 5Wave(e) 90th Quantile 9698100102104PPVT-R Value 1 2 3 4 5Wave(f) MeanNotes: This figure shows the Peabody Picture Vocabulary Test- Revised(PPVT-R) Scale at different thresholds across NLSCY cycles for Quebecand the Rest of Canada (ROC).Figure 4.18: Trends in Body Weight Quantiles4 68 1012 1416Body Weight Quantile 1 2 3 4 5WaveQuebecRest of Canada(a) 10th Quantile 681012 1416 18Body Weight Quantile 1 2 3 4 5Wave(b) 25th Quantile810 1214161820Body Weight Quantile 1 2 3 4 5Wave(c) 50th Quantile 10121416 182022Body Weight Quantile 1 2 3 4 5Wave(d) 75th Quantile12141618202224Body Weight Quantile 1 2 3 4 5Wave(e) 90th Quantile 1012141618 20Body Weight Value 1 2 3 4 5Wave(f) MeanNotes: This figure shows child body weight in kilograms at differentthresholds across NLSCY cycles for Quebec and the Rest of Canada (ROC).1464.8. Conclusion4.8 ConclusionThis study examines the impact of a universal childcare policy in Quebec on the distributionsof motor skills, cognitive development, and body weight of children in the province. Estimatingthe impact of the reform on the marginal distribution of outcomes using a quantile difference-in-differences model, I find a sizeable decline in child body weight at the upper end of thedistribution, with intention to treat estimates of just under 0.24 kg at the 75th threshold and0.26 kg at the 90th threshold. This amounts to treatment on the treated estimates of 1.26kg (31% of a standard deviation) and 0.90 kg (22% of a standard deviation) at the 75th and90th thresholds, respectively. These effects are sizeable relative to the WHO standards. Onaverage, the policy lead to an intention to treat estimate of just under -0.3 kg amongst Quebecchildren. Overall, these findings suggest that the universal childcare policy led to a changein the mix of activities which children were engaged and/or the changes in their diet. This isof particular interest with rising child obesity rates. For motor skills and cognitive outcomes,however, only a handful of estimates were significant in this study, where some percentilesin the upper portion of the Quebec PPVT-R distribution were negatively impacted by thepolicy, particularly for girls. Children of low income fathers also experienced a negativeimpact of the reform at the lower end of the MSD distribution, while the same was found forchildren of low educated fathers in the middle portion of the PPVT-R distribution. Besidesthis handful of significant estimates, there was little significant heterogeneity in the impact ofQuebec’s universal childcare policy. These results were robust to different specifications andestimation techniques. Some explanations for these results were discussed, including the timeframe examined in the study, the sample size used to obtain the PPVT-R estimates, and theidentifying assumption used to derive the estimates.The results presented in this paper are particularly relevant for ongoing policy debatein many developed countries today, where there are heated debates on the merits and costsof universally accessible subsidized care. Universal childcare programs are often justified inpart by the goal of leveling the playing field in child development. This paper is amongstthe first studies to examine whether there is evidence to support this argument and findslittle effect in the short run. Future work in the area should focus on making progress onunraveling what’s inside the “black box” that led to poorer average outcomes for Quebecchildren after the reform. The evidence in this paper suggests heterogeneous responses, atleast in terms of motor and cognitive outcomes, contribute little to this understanding. Inparticular, a structural model might be most promising in developing a better understanding ofthe mechanisms which generated the negative mean impacts of the Quebec universal childcarepolicy.147Chapter 5ConclusionThis dissertation provides an extensive analysis of determinants of health, focusing on specificdimensions that may contribute to health differences across individuals and over time.Chapter 2 analyzes the role of hospital competition in shaping patient treatment andhealth outcomes. In this chapter, I examine how hospitals respond to the loss of a profitableservice line, providing a more complete picture of the nature of hospital spillovers, treatmentdifferences across payer types, and more broadly, how hospitals respond to financial shocks.I find that the hospital response to the loss of a profitable service line is very sophisticated.Hospitals practice both revenue augmenting and cost-cutting behavior in other lines of care,targeting specific procedures and payers according to their profitability. Specifically, theyincrease the number of surgical procedures and perform more marginal surgeries. This varieswith the service line and the payer type. The effects are concentrated in medical specialtieswhere there are more discretionary surgeries and higher profit margins. Hospitals also increasethe intensity of treatment among private payers, by increasing their length of stay. Further-more, hospitals cut back on unprofitable treatment by reducing non-elective admissions anduninsured elective care.The findings of this chapter suggest that hospital responses to financial shocks are muchmore sophisticated and targeted than previously thought. Hospitals are able to adjust theirmix of services and are able to differentiate treatment by payer type. My findings suggestthat focusing only on substitution effects within a service line when evaluating health carepolicies, as much of the existing literature has done, ignores important hospital responses andleads to incomplete welfare implications, particularly among different payer groups.Chapter 3 analyzes the role of health insurance on shaping young American adults’ primarymedical care use. I obtain casual estimates using a regression discontinuity framework thatexploits insurance policy rules where individuals cease being covered on their 19th birthday.This chapter finds that the 19th birthday played a significant role in health insurance coveragerates in the US over the last decade. The estimated reduction in insurance coverage at age 19is 3 to 5 percentage points, driven by a loss of dependent coverage (3 to 4 percentage points).I find office-based physician visits and prescription drugs are not affected by insurance,but dental visits are. The drop in dental visits but not other forms of non-emergency care maybe due to the perceived discretionary nature of dental spending. There is a small increase inout-of-pocket expenditures, concentrated heavily at the top of the distribution. Importantly,148Chapter 5. Conclusionthere is no change in young adults reporting to have a problem affording necessary medicalcare or in their health status. I also find no evidence that young adults stock up on medicalcare prior to insurance loss, suggesting they are myopic in their medical care consumption.Overall, the findings in this chapter show that the vast majority of young adults are notheavily impacted by health insurance in terms of their routine medical care consumption,expenditures, and short-run health. Theses results shed light on the expected welfare benefitsof recent US health care policies targeting young adults.The final chapter of this dissertation analyzes the extent to which the early childhoodenvironment affects child health and development. In particular, I analyze the impact of auniversal childcare policy in Quebec on the distributions of child health and developmentoutcomes, testing if it leveled the playing field across children. I estimate the impact of thereform on the marginal distribution of outcomes in the short run using a quantile difference-in-differences model for two parent families.I find that there is little heterogeneity in the response to the universal childcare policyacross the distributions of motor skills and cognitive outcomes. In fact, this study finds thatthe policy had little significant effect on these outcomes at any point along the distributions,neither for the full sample of children nor when the sample is split by child demographiccharacteristics. I do, however, find evidence that the universal childcare policy led to areduction in child body weight at the upper end of the distribution. These results are robustto different specifications and estimation techniques.The results presented in this chapter are particularly relevant for ongoing policy debatein many developed countries today, where there are heated debates on the merits and costsof universally accessible subsidized care. Universal childcare programs are often justified inpart by the goal of leveling the playing field in child development. This study is amongst thefirst to test whether there is evidence to support this argument, particularly in the Canadiancontext, and finds little effect in the short run.149BibliographyAltman, Stuart H, Shactman, David, & Eilat, Efrat. 2006. Could US hospitals go the way ofUS airlines? Health Affairs, 25(1), 11–21.Anderson, Michael, Dobkin, Carlos, & Gross, Tal. 2012. The effect of health insurance cov-erage on the use of medical services. American Economic Journal: Economic Policy, 4(1),1–27.Antwi, Yaa Akosa, Moriya, Asako S, & Simon, Kosali. 2012. Effects of federal policy to insureyoung adults: Evidence from the 2010 Affordable Care Act dependent coverage mandate.NBER Working Papers Series.Athey, S, & Imbens, GW. 2006. Identification and inference in nonlinear difference-in-differences models. Econometrica, 74(2), 431–497.Baicker, Katherine, & Rehavi, Marit. 2010. The Effect of Medicaid Enrollment on Children:Is Conditional Coverage Enough?Baker, Michael, Gruber, Jonathan, & Milligan, Kevin. 2008. Universal child care, maternallabor supply, and family well-being. Journal of Political Economy, 116(4), 709–745.Banks, Dwayne A, Paterson, Mary, & Wendel, Jeanne. 1997. Uncompensated hospital care:charitable mission or profitable business decision? Health Economics, 6(2), 133–143.Banks, Dwayne A, Foreman, Stephen E, & Keeler, Theodore E. 1999. Cross-subsidization inhospital care: some lessons from the law and economics of regulation. Health Matrix, 9, 1.Barber, C, & Nguyen, Q.C. 2009. Expanding Coverage for Dependents. Community CatalystInc, February Issue.Barro, Jason R, Huckman, Robert S, & Kessler, Daniel P. 2006. The effects of cardiacspecialty hospitals on the cost and quality of medical care. Journal of Health Economics,25(4), 702–721.Bates, John E., Marvinney, Denny, Kelly, Timothy, Dodge, Kenneth, Bennett, D., & Pettit,G. 1994. Child care history and kindergarten adjustment. Developmental Psychology, 30(5),690–700.150BibliographyBaum, C, Schaffer, M.E, & Stillman, S. 2002. IVREG2: Stata module for extended instru-mental variables/2SLS and GMM estimation. Tech. rept. Statistical Software ComponentsS425401, Boston College Department of Economics.Belsky, J. 2001. Emanuel Miller lecture developmental risks (still) associated with early childcare. Journal of Child Psychology, 42(7), 845–59.Belsky, Jay, Vandell, Deborah Lowe, Burchinal, Margaret, Clarke-Stewart, K Alison, McCart-ney, Kathleen, & Owen, Margaret Tresch. 2007. Are there long-term effects of early childcare? Child Development, 78(2), 681–701.Bernal, Raquel, & Keane, Michael P. 2010. Child care choices and children’s cognitive achieve-ment: The case of single mothers. Universidad de los Andes and UTS, working paper.Bertrand, M, Duflo, E, & Mullainathan, S. 2004. How much should we trust differences-in-differences estimates? Quarterly Journal of Economics, 119(1), 249–275.Blau, David, & Tekin, Erdal. 2007. The determinants and consequences of child care subsidiesfor single mothers in the USA. Journal of Population Economics, 20(4), 719–741.Brook, Robert Henry, Ware Jr, John E, Rogers, William H, Keeler, Emmett B, Davies,Allyson R, Donald, Cathy A, Goldberg, George A, Lohr, Kathleen N, Masthay, Patricia C,& Newhouse, Joseph P. 1983. Does free care improve adults’ health? results from a ran-domized controlled trial. The New England Journal of Medicine, 309(23), 1426–1434.Buchmueller, Thomas C, Grumbach, Kevin, Kronick, Richard, & Kahn, James G. 2005.The effect of health insurance on medical care utilization and implications for insuranceexpansion: A review of the literature. Medical Care Research and Review, 62(1), 3–30.Burchinal, M R, Roberts, J E, Riggin, R, Zeisel, SA, Neebe, E, & Bryant, D. 2000. Relatingquality of center-based child care to early cognitive and language development longitudi-nally. Child Development, 71(2), 339–57.Burns, Lawton R, & Wholey, Douglas R. 1992. The impact of physician characteristics inconditional choice models for hospital care. Journal of Health Economics, 11(1), 43–62.Busch, Susan H, & Duchovny, Noelia. 2005. Family coverage expansions: impact on insurancecoverage and health care utilization of parents. Journal of health economics, 24(5), 876–890.Cameron, A Colin, Gelbach, Jonah B, & Miller, Douglas L. 2011. Robust inference withmultiway clustering. Journal of Business & Economic Statistics, 29(2).151BibliographyCantor, Joel C, Belloff, Dina, Monheit, Alan C, DeLia, Derek, & Koller, Margaret. 2012.Expanding Dependent Coverage for Young Adults: Lessons from State Initiatives. Journalof health politics, policy and law, 37(1), 99–128.Card, David, Dobkin, Carlos, & Maestas, Nicole. 2008. The Impact of Nearly Universal Insur-ance Coverage on Health Care Utilization: Evidence from Medicare. American EconomicReview, 98(5), 2242–2258.Card, David, Dobkin, Carlos, & Maestas, Nicole. 2009. Does Medicare Save Lives? TheQuarterly Journal of Economics, 124(2), 597–636.Carlson, Matthew J, DeVoe, Jennifer, & Wright, Bill J. 2006. Short-term impacts of coverageloss in a Medicaid population: early results from a prospective cohort study of the OregonHealth Plan. The Annals of Family Medicine, 4(5), 391–398.Cascio, E. 2009. Do investments in universal early education pay off? Long-term effects ofintroducing kindergartens into public schools. NBER Working Papers Series, 14951.Cheng, Ming-Yen, Fan, Jianqing, Marron, James S, et al. . 1997. On automatic boundarycorrections. The Annals of Statistics, 25(4), 1691–1708.Chernew, Michael, Gowrisankaran, Gautam, & Fendrick, A Mark. 2002. Payer type andthe returns to bypass surgery: evidence from hospital entry behavior. Journal of HealthEconomics, 21(3), 451–474.Chevalier, Judith. 2004. What Do We Know About Cross-subsidization? Evidence fromMerging Firms. Advances in Economic Analysis & Policy, 4(1).Chollet, Deborah, Liu, Su, Gimm, Gilbert, Fahlman, Cheryl, Felland, Laurie, Gerland, An-neliese, Banker, Michelle, & Liebhaber, Allison. 2006. Analysis of Niche Hospitals in Texasand the Financial Impact on General Hospitals. Mathematica Policy Research.Clemens, Jeffrey, & Gottlieb, Joshua D. 2013. Bargaining in the Shadow of a Giant: Medicare’sInfluence on Private Payment Systems. NBER Working Paper.Cunha, F, & Heckman, James. 2007. The technology of skill formation. The AmericanEconomic Review, 97(2), 31–47.Cunha, F, Heckman, JJ, Lochner, L, & Masterov, D. 2006. Interpreting the evidence on lifecycle skill formation. In: Hanushek, E, & Welch, F (eds), Handbook of the Economics ofEducation. Amsterdam: North Holland.Currie, Janet, & Gruber, Jonathan. 1996. Health Insurance Eligibility, Utilization of MedicalCare, and Child Health*. The Quarterly journal of economics, 111(2), 431–466.152BibliographyCutler, David M. 1995a. The cost and financing of health care. The American EconomicReview, 85(2), 32–37.Cutler, David M. 1995b. The Incidence of Adverse Medical Outcomes Under ProspectivePayment. Econometrica, 63(1), 29–50.Dafny, Leemore, & Gruber, Jonathan. 2005. Public insurance and child hospitalizations:access and efficiency effects. Journal of Public Economics, 89(1), 109–129.Dafny, Leemore S. 2005. How Do Hospitals Respond to Price Changes? The AmericanEconomic Review, 95(5), 1525–1547.Dahl, G, & Lochner, L. 2008. The impact of family income on child achievement: Evidencefrom the Earned Income Tax Credit. NBER Working Paper Series, 14599.Damler, R, & Houchens, P. 2012. ACA Premium Impact Variability of Individual MarketPremium Rate Changes. Tech. rept. Milliman Health Care Exchange Issue Brief: IndianaExchange Policy Committee.Danger, Kenneth L, & Frech, H.E. 1997. Payment source and the cost of hospital care:evidence from a multiproduct cost function with multiple payers: Comment. UCSB WorkingPaper.Datta Gupta, Nabanita, & Simonsen, Marianne. 2010. Non-cognitive child outcomes anduniversal high quality child care. Journal of Public Economics, 94(Feb.), 30–43.David, Guy, Lindrooth, Richard, Helmchen, Lorens A, & Burns, Lawton R. 2014. Do HospitalsCross Subsidize? Journal of Health Economics, 37, 198–218.Dor, Avi, & Farley, Dean E. 1996. Payment source and the cost of hospital care: evidence froma multiproduct cost function with multiple payers. Journal of Health Economics, 15(1),1–21.Dranove, David. 1988. Pricing by non-profit institutions: the case of hospital cost-shifting.Journal of Health Economics, 7(1), 47–57.Dranove, David, & White, William D. 1998. Medicaid-dependent hospitals and their patients:How have they fared? Health Services Research, 33(2 Pt 1), 163.Dranove, David, Garthwaite, Craig, & Ody, Christopher. 2013. How do hospitals respond tonegative financial shocks? The impact of the 2008 stock market crash. Tech. rept. NationalBureau of Economic Research.153BibliographyDuggan, Mark G. 2000. Hospital Ownership and Public Medical Spending. The QuarterlyJournal of Economics, 115(4), 1343–1373.Ellis, Randall P. 2001. Hospital payment in the United States: an overview and discussionof current policy issues. In: Paris: Paper prepared for international Conference on settingprices for disease: lessons from foreign experience.Fan, Jianqing, & Gijbels, Irene. 1992. Variable bandwidth and local linear regressionsmoothers. The Annals of Statistics, 2008–2036.Fan, Jianqing, & Gijbels, Irene. 1996. Local polynomial modelling and its applications: mono-graphs on statistics and applied probability 66. Vol. 66. CRC Press.Finkelstein, Amy. 2007. The aggregate effects of health insurance: Evidence from the intro-duction of Medicare. The Quarterly Journal of Economics, 122(1), 1–37.Finkelstein, Amy, Taubman, Sarah, Wright, Bill, Bernstein, Mira, Gruber, Jonathan, New-house, Joseph P, Allen, Heidi, Baicker, Katherine, et al. . 2011. The Oregon health insuranceexperiment: evidence from the first year. NBER Working Papers Series.Firpo, S. 2007. Efficient semiparametric estimation of quantile treatment effects. Economet-rica, 75(1), 259–276.Firpo, S, Fortin, NM, & Lemieux, T. 2009. Unconditional quantile regressions. Econometrica,77(3), 953–973.Frakt, Austin B. 2011. How much do hospitals cost shift? A review of the evidence. MilbankQuarterly, 89(1), 90–130.Frank, Richard, & Salkever, David S. 2000. Market Forces, Diversification of Activity, and theMission of No t-for-Profit Hospitals. Pages 195–226 of: The Changing Hospital Industry:Comparing For-Profit and Not-for-Profit Institutions. University of Chicago Press.Frank, Richard G, & Salkever, David S. 1991. The supply of charity services by nonprofithospitals: Motives and market structure. The Rand Journal of Economics, 430–445.Frank, Richard G, & Zeckhauser, Richard J. 2007. Custom-made versus ready-to-wear treat-ments: behavioral propensities in physicians choices. Journal of Health Economics, 26(6),1101–1127.Freeman, Joseph D, Kadiyala, Srikanth, Bell, Janice F, & Martin, Diane P. 2008. The causaleffect of health insurance on utilization and outcomes in adults: a systematic review of USstudies. Medical care, 46(10), 1023–1032.154BibliographyFro¨lich, M, & Melly, B. 2010. Estimation of quantile treatment effects with STATA. StataJournal.Fro¨lich, Markus, & Melly, Blaise. 2010. Estimation of quantile treatment effects with Stata.Stata Journal, 10(3), 423.Fuji, Daisuke, Imbens, Guido, & Kalyanaraman, Karthik. 2009. Notes for matlab and stataregression discontinuity software. Manuscript. http://www. economics. harvard. edu/facul-ty/imbens/files/rd software 09aug4. pdf.Glied, Sherry, & Zivin, Joshua Graff. 2002. How do doctors behave when some (but not all)of their patients are in managed care? Journal of Health Economics, 21(2), 337–353.Government of Quebec. 1997. An act respecting the Ministere de la Famille et de l’Enfanceand amending the Act respecting child day care. Statutes of Quebec.Gray, Bradford H. 1991. The profit motive and patient care: The changing accountability ofdoctors and hospitals. Harvard University Press.Gregg, P., & Machin, S. 2000. Childhood disadvantage and success or failure in the youthlabour market. In: Blanchflower, D, & Freeman, R (eds), Youth employment and joblessnessin advanced countries. Chicago: University of Chicago Press.Gross, Tal. 2010. Using Insurance Before you Lose It: Health Care Consumption at the Endof Coverage.Gruber, Jonathan. 1994. The effect of competitive pressure on charity: Hospital responses toprice shopping in California. Journal of Health Economics, 13(2), 183–211.Gruber, Jonathan, & Simon, Kosali. 2008. Crowd-out 10 years later: Have recent publicinsurance expansions crowded out private health insurance? Journal of health economics,27(2), 201–217.Hadley, Jack. 2003. Sicker and poorerThe consequences of being uninsured: A review of theresearch on the relationship between health insurance, medical care use, health, work, andincome. Medical Care Research and Review, 60(2 suppl), 3S–75S.Hahn, Jinyong, Todd, Petra, & Van der Klaauw, Wilbert. 2001. Identification and estimationof treatment effects with a regression-discontinuity design. Econometrica, 69(1), 201–209.Hausman, Jerry A. 1978. Specification tests in econometrics. Econometrica: Journal of theEconometric Society, 1251–1271.155BibliographyHavnes, T, & Mogstad, M. 2009. No child left behind. Universal child care and children’slong-run outcomes. Discussion Papers 582, Statistics Norway.Havnes, Tarjei, & Mogstad, Magne. 2010. Is universal child care leveling the playing field?Evidence from non-linear difference-in-differences. IZA Discussion Papers 4978.Heckman, James J, Smith, Jeffrey, & Clements, Nancy. 1997a. Making the most out ofprogramme evaluations and social experiments: Accounting for heterogeneity in programmeimpacts. The Review of Economic Studies, 64(4), 487–535.Heckman, JJ, Smith, Jeffrey, & Clements, Nancy. 1997b. Making the most out of programmeevaluations and social experiments: Accounting for heterogeneity in programme impacts.The Review of Economic, 64(4), 487–535.Herbst, Chris M., & Tekin, Erdal. 2010. Child care subsidies and child development. Eco-nomics of Education Review, 29(Aug.), 618–638.Herzlinger, Regina. 1997. Market Driven Health Care Who Wins Who Loses in the Transfor-mation of America’s Largest Service. Industry Perseus Books Reading.Hickman, L. N. 2006. Who should care for our children?: The effects of home versus centercare on child cognition and social adjustment. Journal of Family Issues, 27(5), 652–684.Ho, Katherine. 2009. Insurer-Provider Networks in the Medical Care Market. AmericanEconomic, 99(1), 393–430.Hoffman, Catherine, & Paradise, Julia. 2008. Health insurance and access to health care inthe United States. Annals of the New York Academy of Sciences, 1136(1), 149–160.Horwitz, Jill R. 2005. Making profits and providing care: Comparing nonprofit, for-profit,and government hospitals. Health affairs, 24(3), 790–801.Horwitz, Jill R, Nichols, Austin, Nallamothu, Brahmajee K, Sasson, Comilla, & Iwashyna,Theodore J. 2013. Expansion of invasive cardiac services in the United States. Circulation,128(8), 803–810.Imbens, Guido, & Kalyanaraman, Karthik. 2011. Optimal bandwidth choice for the regressiondiscontinuity estimator. The Review of Economic Studies, rdr043.Imbens, Guido W, & Angrist, Joshua D. 1994. Identification and estimation of local averagetreatment effects. Econometrica: Journal of the Econometric Society, 467–475.Imbens, Guido W, & Lemieux, Thomas. 2008. Regression discontinuity designs: A guide topractice. Journal of econometrics, 142(2), 615–635.156BibliographyInstitute of Medicine. 2001. Coverage Matters: Insurance and Health Care. Tech. rept.Washington, DC.Japel, Christa, Tremblay, RE, & Coˆte´, Sylvana. 2005. Quality counts! Assessing the qualityof daycare services based on the Quebec Longitudinal Study of Child Development. IRPPChoices, 11(5).Kaserman, David L, & Mayo, John W. 1994. Cross-Subsidies in Telecommunications: Road-blocks on the Road to More Intelligent Telephone Pricing. Yale J. on Reg., 11, 119.Keeler, Emmett B. 1992. Effects of cost sharing on use of medical services and health. Journalof Medical Practice Management, 8.Kessel, Reuben A. 1958. Price discrimination in medicine. Journal of Law and Economics,1, 20–53.Kessler, Daniel P, & McClellan, Mark B. 2000. Is hospital competition socially wasteful? TheQuarterly Journal of Economics, 115(2), 577–615.Kim, Daeho. 2011. Medicare Payment Reform and Hospital Costs: Evidence from the Prospec-tive Payment System and the Treatment of Cardiac Disease. Brown University WorkingPaper.Kisker, E, Hofferth, S, Phillips, D, & Farquhar, E. 1991. A profile of chilcare settings: Earlyeducation and care in 1990. United States Department of Education. Office of the UnderSecretary.Koenker, Roger, & Bassett, Gilbert. 1978. Regression Quantiles. Econometrica, 46(1), 33–50.Koenker, Roger, & Bassett Jr, Gilbert. 1978. Regression quantiles. Econometrica: journal ofthe Econometric Society, 33–50.Kottelenberg, Michael J, & Lehrer, Steven F. 2010. Reinvestigating who Benefits and WhoLoses from Universal Childcare in Canada. Queen’s University, mimeo.Kronebusch, Karl, & Elbel, Brian. 2004. Simplifying childrens Medicaid and SCHIP. HealthAffairs, 23(3), 233–246.Langa, Kenneth M, & Sussman, Elliot J. 1993. The effect of cost-containment policies on ratesof coronary revascularization in California. New England Journal of Medicine, 329(24),1784–1789.Lee, David S, & Card, David. 2008. Regression discontinuity inference with specification error.Journal of Econometrics, 142(2), 655–674.157BibliographyLee, David S, & Lemieux, Thomas. 2010. Regression Discontinuity Designs in Economics.Journal of Economic Literature, 48(2), 281–355.Lefebvre, Pierre, & Merrigan, Philip. 2008. Child-care policy and the labor supply of motherswith young children: A natural experiment from Canada. Journal of Labor Economics,26(3), 519–548.Levine, Phillip B, McKnight, Robin, & Heep, Samantha. 2011. How effective are publicpolicies to increase health insurance coverage among young adults? American EconomicJournal: Economic Policy, 3(1), 129–156.Li, Suhui, & Dor, Avi. 2013. How Do Hospitals Respond to Market Entry? Evidence from aDeregulated Market for Cardiac Revascularization. NBER Working Paper.Lindrooth, Richard C, Konetzka, R Tamara, Navathe, Amol S, Zhu, Jingsan, Chen, Wei,& Volpp, Kevin. 2013. The Impact of Profitability of Hospital Admissions on Mortality.Health Services Research.Loeb, S, Bridges, M, Bassok, D, Fuller, B, & Rumberger, R. 2007. How much is too much? Theinfluence of preschool centers on children’s social and cognitive development. Economics ofEducation Review, 26(1), 52–66.Love, John M, Harrison, Linda, Sagi-Schwartz, Abraham, van IJzendoorn, Marinus H, Ross,Christine, Ungerer, Judy, Raikes, Helen, Brady-Smith, Christy, Boller, Kimberly, Brooks-Gunn, Jeanne, Constantine, Jill, Kisker, Ellen Eliason, Paulsell, Diane, & Chazan-Cohen,Rachel. 2003. Child care quality matters: How conclusions may vary with context. ChildDevelopment, 74(4), 1021–33.Ludwig, Jens, & Miller, Douglas L. 2007. Does Head Start Improve Children’s Life Chances?Evidence from a Regression Discontinuity Design. The Quarterly Journal of Economics,159–208.Luft, Harold S, Garnick, Deborah W, Mark, David H, Peltzman, Deborah J, Phibbs, Ciaran S,Lichtenberg, Erik, & McPhee, Stephen J. 1990. Does quality influence choice of hospital?JAMA: the Journal of the American Medical Association, 263(21), 2899–2906.Mahoney, Neale. 2012. Bankruptcy as implicit health insurance. NBER Working PapersSeries.Mankiw, N Gregory, & Whinston, Michael D. 1986. Free entry and social inefficiency. TheRAND Journal of Economics, 48–58.McFadden, D. 1973. Conditional logit analysis of qualitative choice behavior. Pages 105–142of: Zarembka, P. (ed), Frontiers in Econometrics. New York: Academic press.158BibliographyMcGuire, Thomas G, & Pauly, Mark V. 1991. Physician response to fee changes with multiplepayers. Journal of Health Economics, 10(4), 385–410.McWilliams, J Michael, Zaslavsky, Alan M, Meara, Ellen, & Ayanian, John Z. 2003. Impactof Medicare coverage on basic clinical services for previously uninsured adults. JAMA,290(6), 757–764.Medicare Payment Advisory Commission. 2005. Report to the Congress: Physician-OwnedSpecialty Hospitals.Meyers, Marcia K, Heintze, Theresa, & Wolf, Douglas. 2002. Child care subsidies and theemployment of welfare recipients. Demography, 39(1), 165–79.Milligan, Kevin. 2005. Subsidizing the stork: New evidence on tax incentives and fertility.Review of Economics and Statistics, 87(3), 539–555.Milligan, Kevin, & Stabile, Mark. 2007. The integration of child tax credits and welfare:Evidence from the Canadian National Child Benefit program. Journal of Public Economics,91(1), 305–326.Milligan, Kevin, & Stabile, Mark. 2011. Do Child Tax Benefits Affect the Well-being ofChildren? Evidence from Canadian Child Benefit Expansions. American Economic Journal:Economic Policy, 175–205.Morrisey, Michael A. 1994. Cost shifting in health care: Separating evidence from rhetoric.American Enterprise Institute.National Conference of State Legislatures. 2011. Covering Young Adults Through TheirParents or Guardians Health Policy.Newhouse, Joseph P. 1982. A Summary of the Rand Health Insurance Study*. Annals of theNew York Academy of Sciences, 387(1), 111–114.Newhouse, Joseph P. 2002. Pricing the Priceless: A Health Care Conundrum. Cambridge,Massachusetts: The MIT Press.Newhouse, Joseph P, & Marquis, M Susan. 1978. The norms hypothesis and the demand formedical care. Journal of Human Resources, 159–182.NICHD Early Child Care Research Network. 1998. Early child care and self-control, com-pliance, and problem behavior at twenty-four and thirty-six months. Child Development,69(4), 1145–1170.159BibliographyNICHD Early Child Care Research Network. 2000. The relation of child care to cognitive andlanguage development. Child Development, 71(4), 960–980.NICHD Early Child Care Research Network. 2001. Child care and children’s peer interactionat 24 and 36 months: The NICHD study of early child care. Child Development, 72(5),1478–1500.NICHD Early Child Care Research Network. 2002. Early child care and children’s developmentprior to school entry: Results from the NICHD Study of Early Child Care. AmericanEducational Research Journal, 39(1), 133–164.NICHD Early Child Care Research Network. 2003. Does amount of time spent in child carepredict socioemotional adjustment during the transition to kindergarten? Child Develop-ment, 74(4), 976–1005.NICHD Early Child Care Research Network. 2004. Type of child care and children’s devel-opment at 54 months. Early Childhood Research Quarterly, 19(2), 203–230.NICHD Early Child Care Research Network. 2006. Child-care effect sizes for the NICHDstudy of early child care and youth development. American Psychologist, 61(2), 99–116.Nichols, Austin. 2012. rd: Stata module for regression discontinuity estimation. StatisticalSoftware Components.Norton, Edward C, & Staiger, Douglas O. 1994. How hospital ownership affects access to carefor the uninsured. The Rand Journal of Economics, 171–185.Porter, Jack. 2003. Estimation in the regression discontinuity model. Unpublished Manuscript,Department of Economics, University of Wisconsin at Madison, 5–19.Resnick, Andrew S, Corrigan, Diane, Mullen, James L, & Kaiser, Larry R. 2005. Surgeoncontribution to hospital bottom line: not all are created equal. Annals of surgery, 242(4),530.Rosenbach, M, Ellwood, M, Irvin, C, Young, C, Conroy, W, Quinn, B, , & Kell, M. 2003.Implementation of the State Childrens Health Insurance Program, Synthesis of State Eval-uations, Background for the Report to Congress. Tech. rept. Mathematica Policy ResearchInc, MPR Reference 8644-100.Rubin, DB. 1978. Bayesian inference for causal effects: The role of randomization. The Annalsof Statistics, 6, 34–58.Scarr, S, & Eisenberg, M. 1993. Child care research: Issues, perspectives, and results. Annualreview of psychology, 44(Jan.), 613–44.160BibliographySchwartz, K, & Damico, A. 2010. Aging Out of Medicaid: What Is the Risk of BecomingUninsured. Tech. rept. Kaiser Commission on Medicaid and the Uninsured.Schwartz, K, & Schwartz, T. 2008. Uninsured Young Adults: A Profile and Overview ofCoverage Options. Tech. rept. Kaiser Commission on Medicaid and the Uninsured.Shonkoff, J, & DA, Phillips. 2000. From neurons to neighborhoods: The science of early child-hood development. Committtee on Integrating the Science of Early Childhood Development,Board on Children, Youth, and Families.Skinner, Wickham. 1974. The focused factory. Harvard Business Review.Sloan, Frank A. 1998. Commercialism in nonprofit hospitals. Journal of Policy Analysis andManagement, 17(2), 234–252.Swanson, Ashley. 2012. Physician Ownership and Incentives: Evidence from Cardiac Care.University of Pennsylvania, mimeo.Tai-Seale, Ming, McGuire, Thomas G, & Zhang, Weimin. 2007. Time allocation in primarycare office visits. Health Services Research, 42(5), 1871–1894.Tekin, E. 2005. Child care subsidy receipt, employment, and child care choices of singlemothers. Economics Letters, 89(1), 1–6.Terza, Joseph V, Basu, Anirban, & Rathouz, Paul J. 2008. Two-stage residual inclusionestimation: addressing endogeneity in health econometric modeling. Journal of HealthEconomics, 27(3), 531–543.Tougas, Jocelyne. 2002. Reforming Que´bec’s early childhood care and education: The firstfive years. CRRU Occasional Paper, 17.US Congress. 2006. Physician-owned specialty hospitals: profits before patients?: Hearingbefore the Committee on Finance, United States Senate, One Hundred Ninth Congress,second session. May 17.Wooldridge, Jeffrey M. 2014. Quasi-maximum likelihood estimation and testing for nonlinearmodels with endogenous explanatory variables. Journal of Econometrics.Woolridge, Jeffrey M. 2002. Econometric analysis of cross section and panel data.World Health Organization. 2006. WHO child growth standards: length/height-for-age,weight-for-age, weight-for-length, weight-for-height and body mass index-for-age: methodsand development.161Wu, Vivian Y. 2010. Hospital cost shifting revisited: new evidence from the balanced budgetact of 1997. International Journal of Health Care Finance and Economics, 10(1), 61–83.Zuckerman, Stephen. 1987. Commercial insurers and all-payer regulation: Evidence on hos-pitals’ responses to financial need. Journal of Health Economics, 6(3), 165–187.162Appendix AAppendix to Chapter 2A.1 Specialty Hospital DesignationSpecialty hospitals were identified following the definition outlined in the Medicare PaymentAdvisory Commission (2005)’s Report to the Congress. A hospital was designated as a spe-cialty hospital if at least 45 percent of the hospitals discharges were cardiac, orthopedicor surgical in nature, or at least 66 percent of the hospitals discharges fell into two majordiagnosis-related categories (MDC), with the primary one being either cardiac or orthopedic.This definition is the most widespread and is used in numerous other governmental reports,including those by the Secretary of the Department of Health and Human Services (HHS) andthe Center for Medicaid and Medicare Services (CMS). This designation is also aligned withthe description of a specialty hospital provided in Section 507 of the Medicare PrescriptionDrug, Improvement, and Modernization Act of 2003 (MMA) as well as the one outlined inTexas Senate Bill 872. The MMAs definition only considers physician-owned hospitals tobe specialty hospitals, while the Texas Senate Bill excludes public hospitals as well as thosehospitals for which the majority of inpatient claims are for major diagnosis-related groupsrelating to rehabilitation, psychiatry, alcohol and drug treatment, or children or newborns.Additionally, it should be noted that the Texas Senate Bill only classifies specialty hospitalsusing the higher threshold of two thirds (roughly 0.66 as used above) for the top two MDCs orsurgical cases. Thus, while my approach will capture all hospitals designated as specialty us-ing the Texas Senate Bill, it will also include additional hospitals since the Medicare PaymentAdvisory Commission (MedPAC) definition is somewhat less stringent.The steps I take to identify specialty hospitals are as follows. I first derived the totaldischarges in a year for each hospital. Then, to isolate the concentration of services offeredin the hospital, I examined the distribution of medical diagnoses. Specifically, I constructedthree specialty indices for each hospital for each year based on the definition of specialtyhospital above:Specialty Index 1 is the proportion of total hospital discharges that fall in the most com-mon Major Diagnostic Category (MDC) in the year. This index only considers hospitals withtop MDCs being cardiac or orthopedic and is missing for all others. A hospital was classi-fied as a cardiac specialty hospital in the year if its most common MDC was Diseases and163A.1. Specialty Hospital DesignationDisorders of the Circulatory System (MDC 5) and if 45% of its cases fell into this category.Similarly, orthopedic hospitals must have its most common MDC being Diseases and Disor-ders of the Musculoskeletal System and Connective Tissue (MDC 8), with 45% of its cases inthis group. Hospitals with Specialty Index 1 of 0.45 or greater are consequently designatedas being specialty.Specialty Index 2 is the proportion of total hospital discharges with surgical DRGs in theyear (as identified using the CMSs annual list of DRGs). A hospital was classified as a surgicalspecialty hospital in the year if this index was 0.45 or greater (i.e. 45% or more of dischargesinvolved a surgical procedure) and if it was not identified as a particular type of specialtyusing Specialty Index 1.Specialty Index 3 is the proportion of total hospital discharges that that the top twoMDCs make up in the year. It only considers those with the most common MDC being eitherCardiac (MDC 5) or Orthopedic (MDC 8). A hospital was classified as a specialty hospitalin the year if the specialty index was 0.66 or greater. Again, it was identified as a particulartype (cardiac or orthopedic) based on the most common MDC. Although all three indices wereused to determine which hospitals were specialty hospitals, there were no hospitals identifiedas a specialty using Index 3 that were not already identified using Indices 1 and 2. Thus, themain criteria effectively used to determine specialty hospital’s was whether at least 45 percentof a hospitals discharges were cardiac, orthopedic, or surgical.All hospitals that were identified as a specialty were then examined thoroughly. If thehospital was publicly owned, it was removed from the list of specialty hospitals. I used questionB1 from the Annual Hospital Survey to establish the type of organization that is responsiblefor controlling the operation of the hospital. Any reporting to be government operated (codes12-16 and 41-48) were removed from the specialty list. Additionally, a hospital that had amajority of inpatient claims being for discharges relating to rehabilitation, psychiatry, alcoholand drug treatment, or children or newborns, was removed from the list of specialty hospitalsand excluded from analysis. Hospitals whose primary focus was on surgeries not covered byMedicare (such as bariatric surgery) were also removed. Specifically, I used the question onAHA Annual Survey of Hospitals that asks hospitals to indicate the type of services that bestwhat is provided to the majority of their patients. Hospitals were excluded from the analysisif they identified as being either psychiatric (code 22), an institute for the mentally retarded(code 62), tuberculosis or other respiratory diseases (code 33), cancer (41), rehabilitation(46), chronic diseases (48), acute long-term care (80 and 90), or alcoholism/other chemicaldependency (82). Additionally, I examined the share of DRGs in each hospital in a year thatfell into rehabilitation, psychiatry, alcohol/drug treatment, children/newborns, and bariatricsurgery. This was done primarily to validate the AHA information and also to examine164A.1. Specialty Hospital Designationhospitals that were not in the AHA Annual Survey. Those with very high shares in theexcluded categories were removed from the analysis.Additionally, there were a number of hospitals that were on the margin of being a specialtyhospital, meeting the threshold in some years but not others. For these hospitals, I followedthe approach taken by Chollet et al. (2006), using a case by case basis. A hospital wasdesignated as specialty if it was just under the threshold in earlier years but was well aboveit in later years. Conversely, if a hospital was above the threshold in the earlier years but thespecialty index gradually fell over time to below the threshold, it was classified as a specialtyhospital only in those earlier years where it met the specialty criteria.Another challenge was that the Texas IPUDFs do not include hospitals with fewer than50 inpatient discharges per quarter or those that report with other facilities. In such cases,I used discharge information for the quarters whenever available as well as in-depth websearches and AHA information to establish if a hospital was a specialty. If the hospital wasclearly above the threshold in the periods it was in the discharge data files, it was consideredto be a specialty throughout the sample. The AHA data had detailed information on thecharacteristics of most Texan hospitals, including those that were not in the PUDF, whichwas also quite useful in identifying specialty hospitals not appearing in the discharge data.I ran a probit model to obtain a propensity score for being a specialty hospital using AHAvariables, such as total beds, physician ownership, total births, as the explanatory variablesand an indicator for being a (non borderline) specialty hospital as the dependent variable.This helped identify some hospitals missing in the PUDF data as well as borderline hospitalsas being specialty. Additionally, I spent considerable time looking up individual hospitalsthrough web searches to see if it self-identified as a specialty or whether there was strongqualitative evidence to indicate it was a specialty hospital.Upon developing a preliminary list of specialty hospitals, I compared it to those pro-duced by other organizations. In particular, I examined the lists provided in the 2006 SenateHearings on Physician-Owned Specialty Hospitals to ensure I was not missing any specialtyhospitals (US Congress (2006)). All hospitals listed in the report as specialty (i.e. specialty)were on my list; although, my list also included hospitals that were not owned by physicians(i.e. other investor-owned and in one case non-profit). Additionally, I obtained a list of ex-isting hospitals that identified as specialty in 2012 from the Regulatory Licensing Unit of theTexas Department of State Health Services. Reassuringly, I had classified all hospitals on thatlist as specialty hospitals. Although the Chollet et al. (2006) study does not provide a listof specialty hospitals and uses only the somewhat more stringent Texas Senate Bill definitionof specialty hospital, I compared the number of specialty hospitals and their general location(i.e. county) for the time period of their study using only the Texas Senate Bill criteria.Again, my approach produced very similar results in terms of the quantity and location ofspecialty hospitals in Texas.165A.2. Obtaining Patient to Hospital DistancesA.2 Obtaining Patient to Hospital DistancesA.2.1 Hospital LocationThe location of hospitals was obtained using information from the American Hospital Asso-ciations (AHA) Annual Survey of Hospitals, the Texas Health Care Information Collection(THCIC) database, and researcher collected data. The AHA Annual Survey of Hospitalsand the THCIC information were kindly provided by the Texas Department of State HealthServices (DSHS). Both data sources contain annual information on all licensed hospitals inTexas, including the physical address of hospitals. It is mandatory for all licensed hospitalsto respond to the AHA Annual Survey. As such, the bulk of the hospital addresses were ob-tained from the AHA Annual Survey. In the case where a hospital is licensed as part of a mainhospital, only the main hospital reports to the AHA. As such, I used data from the THCICdatabase to fill in addresses for these hospitals whenever possible. A small subset of hospitalsdid not appear in either the AHA Annual Survey or the THCIC data files, so I performedthorough internet searches for these hospitals to obtain an address. The majority of hospitalsappeared in both the AHA Annual Survey and the THCIC data files, so I cross-checked theAHA information against the THCIC information. If the addresses differed across sources,I verified the correct address through rigorous internet searches. It should be noted that ina small number of cases, hospitals changed location over the sample period, either movinginto a brand new structure (largely in rural areas) or moving into an existing building thathad previously housed a hospital (more common in urban areas). In these cases, the yearthe hospital moved was noted, with the old and new location being used in the appropriatetime period. Once the hospital locations were verified and collected, GIS software was usedto convert the addresses into longitudinal coordinates. The software used was ArcGIS 10.1developed by ESRI. ArcGIS can be used to manage attribute data, in this case addresses, anddisplay them geographically by geocoding. Specifically, the hospital addresses were geocodedin ArcGIS 10.1 using the 10.0 North America Address Locator. This locator is based onNAVTEQ Q3 2011 reference data for North America and was last updated in June 2012. Inalmost all cases, the hospital addresses matched correctly to the points plotted by ArcGIS.In some cases, however, the address had to be slightly altered prior to geocoding for ArcGISto correctly identify the location (i.e. giving a street name adjacent to the actual street orslightly changing the street number). Many robustness checks were done to ensure that thelocation obtained correctly matched the hospitals address, such as comparing the address andcoordinates generated by ArcGIS to those in Google Maps.166A.2. Obtaining Patient to Hospital DistancesA.2.2 Patient LocationThe analysis is restricted to patients living in Texas. This was determined using informationcollected by hospitals on patients listed state of residence and zip codes. Individuals denotedas residing outside of Texas were excluded from the analysis (132,336 inpatient visits). Thefull five digit zip code was recorded for 94.19% of Texan patients. In order to preserve patientconfidentiality, the DSHS suppressed the last two digits of a zip code if there were fewer thanthirty patients in the zip code in a discharge quarter. The entire zip code was suppressed if ahospital had less than 50 discharges a quarter or if the ICD-9 code indicated sensitive medicalconditions (i.e. alcohol or drug abuse or an HIV diagnosis). Although some patients withmissing zip codes had county of residence, I only included patients with a full five digit zipcode in the analysis to ensure a high level of precision in patient residence.The location of patients residences were approximated with longitudinal coordinates thatwere derived in ArcGIS using the centroid of the zip code for those patients with full five digitzip codes. Zip Codes are not geographic features but are instead a collection of mail deliveryroutes for the US Postal Service. As such, to obtain a geographic representation of the zipcodes to match to the patient-level data, ZCTA area shapefiles for all of Texas were obtainedfrom the US Census Bureau for the years 2000 and 2010. ZCTA regions are geographicalareas produced by the US Census Bureau based on the most prevalent postal zip code withina fixed geographic area. As such, while the match between ZCTA areas and zip codes is notexact, there is significant overlap. In order to calculate the centroids of the ZTCA boundaries,I used the Feature to Point tool in ArcGIS which creates a feature class containing centroidpoints generated from the boundary polygon line of the ZCTA area.A.2.3 Patient to Hospital DistanceTo derive distances between patients’ residences and hospitals, the centroids of the ZCTAsand the hospital locations were projected using a UTM Projected Coordinates System (NAD1983 HARN UTM Zone 14N). The distances were calculated using the Point Distance tool inArcGIS, which provides Euclidean distances (i.e. as the crow flies). Non-teaching hospitalsthat were more than 50 miles from the patient residence were dropped from her choice set.Teaching hospitals that were more than 100 miles from the patient were also dropped.167A.3. Supplemental Figures and TablesA.3 Supplemental Figures and TablesFigure A.1: Hospital Service Areas in Texas!!!!! !!!!!!DallasHoustonAustinAmarilloEl PasoOdessaLubbockSan AntonioCorpus ChristiWichita FallsEdinburgNotes: This figure shows the boundaries of Hospital Service Areas (HSA) in Texas. HSAs are local healthcare markets for hospital care. An HSA is a collection of zip codes whose residents receive most of theirhospitalizations from the hospitals in that area. It is produced by the Dartmouth Atlas of Health Care.168A.3.SupplementalFiguresandTablesTable A.1: The Distribution of Predicted and Actual Specialty Market SharesYear Mean Std. Dev 25th Percentile Median 75th Percentile OLS Coefficient R squaredActual Specialty Market Share1999 All 0.009 0.031 0 0 0Cardiac 0.010 0.043 0 0 0Orthopedic 0.006 0.019 0 0 02003 All 0.027 0.049 0 0.005 0.032Cardiac 0.020 0.054 0 0 0.003Orthopedic 0.034 0.070 0 0 0.0452007 All 0.037 0.056 0 0.013 0.056Cardiac 0.024 0.060 0 0 0.009Orthopedic 0.058 0.092 0 0.019 0.095Overall All 0.029 0.057 0 0 0.033Cardiac 0.022 0.064 0 0 0.001Orthopedic 0.036 0.075 0 0 0.047Predicted Specialty Market Share1999 All 0.007 0.014 0 0 0.008 1.288 0.382Cardiac 0.006 0.015 0 0 0.005 1.703 0.345Orthopedic 0.009 0.016 0 0 0.012 0.548 0.1982003 All 0.022 0.037 0 0.010 0.029 0.795 0.357Cardiac 0.018 0.034 0 0.005 0.021 0.728 0.211Orthopedic 0.031 0.047 0 0.014 0.043 0.813 0.2952007 All 0.039 0.044 0 0.025 0.065 0.892 0.504Cardiac 0.032 0.037 0 0.023 0.050 0.645 0.159Orthopedic 0.050 0.059 0 0.029 0.086 0.996 0.417Overall All 0.028 0.045 0 0.006 0.040 0.824 0.438Cardiac 0.025 0.048 0 0.003 0.033 0.607 0.208Orthopedic 0.033 0.052 0 0.006 0.048 0.925 0.408Notes: The specialty market share is defined as the proportion of patients in the HSA that are admitted to specialty hospitals.The predicted specialty market shares are estimated using maximum likelihood and are derived from a patient-level multinomialhospital choice model for patients seeking care in specialty services (i.e. MDC=5 or MDC=8). The choice set includes all hospitalswithin a 50 mile radius from the patient (or 100 miles for teaching hospitals). A patient’s indirect utility function is specified as anon-parametric function of hospital-patient distance quartiles, fully interacted with patient and hospital characteristics. Estimationwas done separately across years and across type of care (cardiac surgical, cardiac non-surgical, orthopedic surgical, and orthopedicnon-surgical). The coefficient and the R-squared from an OLS regression of actual market share on predicted market share are shownin the last two columns.169A.3.SupplementalFiguresandTablesTable A.2: Impact of Increased Specialty Competition on Share of Surgical PatientsOverall Within DepartmentSpecification: Main Main Zip code No Other Main Main Zip code No Other(no bootstrap) FE Trend Zip code (no bootstrap) FE Trend Zip codeSMKS -0.0411 -0.0411 -0.0370 -0.129*** -0.0408 -0.0174 -0.0174 -0.0157 -0.0461*** -0.0171(0.0285) (0.0266) (0.0263) (0.0230) (0.0265) (0.0138) (0.0116) (0.0117) (0.0105) (0.0116)SMKS x Medicaid 0.0147 0.0147 0.0175 0.0305 0.0153 0.0161 0.0161 0.0180 0.0230* 0.0160(0.0377) (0.0383) (0.0379) (0.0365) (0.0382) (0.0122) (0.0119) (0.0119) (0.0118) (0.0118)SMKS x Private: HMO 0.219*** 0.219*** 0.214*** 0.208*** 0.220*** 0.0896*** 0.0896*** 0.0893*** 0.0863*** 0.0898***(0.0440) (0.0447) (0.0432) (0.0440) (0.0445) (0.0192) (0.0181) (0.0178) (0.0181) (0.0181)SMKS x Private: FFS 0.139*** 0.139*** 0.134*** 0.135*** 0.140*** 0.0808*** 0.0808*** 0.0793*** 0.0785*** 0.0812***(0.0265) (0.0258) (0.0254) (0.0257) (0.0257) (0.0125) (0.0118) (0.0116) (0.0118) (0.0117)SMKS x Uninsured 0.0013 0.0013 -0.0009 -0.0064 0.0012 0.0056 0.0056 0.0043 0.0052 0.0053(0.0342) (0.0374) (0.0366) (0.0373) (0.0372) (0.0192) (0.0194) (0.0190) (0.0194) (0.0193)Observations 2,295,064 2,295,064 2,295,064 2,295,064 2,295,062 2,275,489 2,275,489 2,275,489 2,275,489 2,275,487Mean 0.266St. Dev 0.442Payer Dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes YesPayer Interactions Yes Yes Yes Yes Yes Yes Yes Yes Yes YesDept Dummies No No No No No Yes Yes Yes Yes YesNotes: This table shows the change in the proportion of patients with a surgical admission in a HSA due to specialty hospital market share (SMKS) under various specifications.The coefficient on SMKS is estimated with a linear probability model using the method of two-stage residual inclusion. The base category for payer type is Medicare. Yearfixed effects are included. Patient demographic characteristics (gender dummy, five year age group dummies, race dummies, Hispanic dummy) and hospital characteristics (totalbeds indicators, for profit, teaching dummy) are included. “Main” reports the original estimates in the main specification. “Zip code FE” employs zip code fixed effects ratherthan HSA fixed effects. “No Trend” removes the HSA specific linear trend included in the main estimation. “Other Zip Code” uses the raw 2000 and 2010 census zip code, asopposed to their linear interpolation as in the main specification. Standard errors are clustered by zip code with the main specification bootstrapped. * p<0.10, ** p<0.05, ***p<0.01.170A.3.SupplementalFiguresandTablesTable A.3: Impact of Increased Specialty Competition on Length of Stay (LOS)Overall Within DRGSpecification: Main Main Zip code No Other Main Main Zip code No Other(no bootstrap) FE Trend Zip code (no bootstrap) FE Trend Zip codeSMKS 0.435 0.435 0.439 -0.269 0.431 0.383 0.383 0.392 0.627 0.381(0.465) (0.493) (0.493) (0.459) (0.493) (0.400) (0.383) (0.383) (0.381) (0.383)SMKS x Medicaid 0.519 0.519 0.581 0.623 0.509 0.0691 0.0691 0.113 0.0220 0.0660(0.615 (0.567) (0.571) (0.545) (0.566) (0.497) (0.455) (0.459) (0.438) (0.455)SMKS x Private: HMO 2.698*** 2.698*** 2.578*** 2.969*** 2.724*** 1.758*** 1.758*** 1.647*** 2.081*** 1.773***(0.468) (0.454) (0.450) (0.453) (0.452) (0.372) (0.395) (0.393) (0.396) (0.394)SMKS x Private: FFS 2.202*** 2.202*** 2.123*** 2.213*** 2.213*** 1.248*** 1.248*** 1.179*** 1.277*** 1.250***(0.388) (0.394) (0.390) (0.400) (0.393) (0.303) (0.311) (0.311) (0.312) (0.311)SMKS x Uninsured 1.383*** 1.383*** 1.435*** 1.408*** 1.388*** 0.786* 0.786* 0.869** 0.773* 0.791*(0.452) (0.472) (0.479) (0.474) (0.472) (0.425) (0.401) (0.405) (0.402) (0.402)Observations 2,295,062 2,295,062 2,295,062 2,295,062 2,295,060 2,295,062 2,295,062 2,295,062 2,295,062 2,295,060Mean 5.376St. Dev 9.414Payer Dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes YesPayer Interactions Yes Yes Yes Yes Yes Yes Yes Yes Yes YesDept FE No No No No No No No No No NoDRG FE No No No No No Yes Yes Yes Yes YesNotes: This table shows the change in the length of stay in a HSA due to specialty hospital market share (SMKS). The coefficient on SMKS is estimated with a linearprobability model using the method of two-stage residual inclusion. The base category for payer type is Medicare. Year fixed effects are included. Patient demographiccharacteristics (gender dummy, five year age group dummies, race dummies, Hispanic dummy) and hospital characteristics (total beds indicators, for profit, teaching dummy)are included. “Main” reports the original estimates in the main specification. “Zip code FE” employs zip code fixed effects rather than HSA fixed effects. “No Trend”removes the HSA specific linear trend included in the main estimation. “Other Zip Code” uses the raw 2000 and 2010 census zip code, as opposed to their linear interpolationas in the main specification. Standard errors are clustered by zip code with the main specification bootstrapped. * p<0.10, ** p<0.05, *** p<0.01.171A.3.SupplementalFiguresandTablesTable A.4: Impact of Increased Specialty Competition on Mortality RateOverall Within DRGSpecification: Main Main Zip code No Other Main Main Zip code No Other(no bootstrap) FE Trend Zip code (no bootstrap) FE Trend Zip codeSMKS 0.0076 0.0076 0.0087 -0.0104 0.0076 -0.0089 -0.0089 -0.0081 -0.0124* -0.0089(0.0108) (0.0108) (0.0109) (0.0082) (0.0108) (0.0094) (0.0094) (0.0094) (0.0075) (0.0094)SMKS x Medicaid 0.0154* 0.0154* 0.0155* 0.0189** 0.0153* 0.0175** 0.0175** 0.0180** 0.0193*** 0.0175**(0.0092) (0.0090) (0.0091) (0.0088) (0.0090) (0.0081) (0.0073) (0.0074) (0.0071) (0.0073)SMKS x Private: HMO 0.0210* 0.0210** 0.0202** 0.0200** 0.0212** 0.0328*** 0.0328*** 0.0316*** 0.0327*** 0.0328***(0.0109) (0.0096) (0.0096) (0.0097) (0.0096) (0.0104) (0.0092) (0.0092) (0.0093) (0.0092)SMKS x Private: FFS 0.0177** 0.0177** 0.0164** 0.0180** 0.0177** 0.0310*** 0.0310*** 0.0297*** 0.0313*** 0.0310***(0.0078) (0.0074) (0.0073) (0.0075) (0.0074) (0.0075) (0.0067) (0.0067) (0.0068) (0.0067)SMKS x Uninsured 0.0325*** 0.0325*** 0.0336*** 0.0324*** 0.0325*** 0.0392*** 0.0392*** 0.0403*** 0.0391*** 0.0393***(0.0122) (0.0115) (0.0115) (0.0115) (0.0115) (0.0112) (0.0106) (0.0106) (0.0106) (0.0106)N 2,290,179 2,290,179 2,290,179 2,290,179 2,290,177 2,290,179 2,290,179 2,290,179 2,290,179 2,290,177Mean 0.0307St. Dev 0.1724Payer Dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes YesPayer Interactions Yes Yes Yes Yes Yes Yes Yes Yes Yes YesDept FE No No No No No No No No No NoDRG FE No No No No No Yes Yes Yes Yes YesNotes: This table shows the change in the mortality rate in a HSA due to specialty hospital market share (SMKS). The coefficient on SMKS is estimated with a linear probabilitymodel using the method of two-stage residual inclusion. The base category for payer type is Medicare. Year fixed effects are included. Patient demographic characteristics (genderdummy, five year age group dummies, race dummies, Hispanic dummy) and hospital characteristics (total beds indicators, for profit, teaching dummy) are included.“Main”reports the original estimates in the main specification. “Zip code FE” employs zip code fixed effects rather than HSA fixed effects. “No Trend” removes the HSA specific lineartrend included in the main estimation. “Other Zip Code” uses the raw 2000 and 2010 census zip code, as opposed to their linear interpolation as in the main specification.Standard errors are clustered by zip code with the main specification bootstrapped. * p<0.10, ** p<0.05, *** p<0.01.172Appendix BAppendix to Chapter 3B.1 Supplemental Figures and TablesFigure B.1: Insurance Coverage by Demographic Group.45.55 .65.75.85Proportion Covered 17 18 19 20 21Age(a): Non Students .45.55.65.75.85Proportion Covered 17 18 19 20 21Age(b): Students.45.55.65.75 .85Proportion Covered 17 18 19 20 21Age(c): Moved Out .45.55.65.75.85Proportion Covered 17 18 19 20 21Age(d): Living at HomeNotes: The dependent variable is having medical insurance coverage at anytime during the calendar month. The analysis is done by differentdemographic groups: Non-students, Students, Moved Out, and Living atHome. The circles plot local averages at each age month, while the solidlines are fitted values from local linear regressions of the dependent variableusing a triangular kernel and the IK bandwidth for the subsample.173B.1. Supplemental Figures and TablesFigure B.2: Type of Office Visit.04.05.06.07 .08.09Had Office Visit 17 18 19 20 21Age(a): Diagnosis/Treatment 0.01 .02.03.04.05Had Office Visit 17 18 19 20 21Age(b): Checkup0.01.02.03.04 .05Had Office Visit 17 18 19 20 21Age(c): Mental CounselingNotes: The dependent variable here is a particular type of office based visitin a given calendar month. Panel (a) is a visit for a diagnosis or treatment,(b) is a checkup, and (c) is for psychotherapy or mental counseling. Thecircles plot local averages at each age month, while the solid lines are fittedvalues from local linear regressions of the dependent variable using atriangular kernel and the IK bandwidth.Figure B.3: Robustness Checks.4 .5.6.7 .8.9Proportion 17 18 19 20 21Age(a): Student 0.1.2.3.4.5Proportion 17 18 19 20 21Age(b): Moved Out600700800 900Individuals 17 18 19 20 21Age(c): No. of Individuals 0.1.2.3Proportion 17 18 19 20 21Age(d): Reference PersonNotes: This figure examines whether any other observable factors whichmay bias the estimates are changing at age 19, namely being a student(panel a), moving out of the home (panel b), the number of individuals inthe survey (panel c), and being the reference person (panel d). The circlesplot local averages at each age month, while the solid lines are fitted valuesfrom local linear regressions of the dependent variable using a triangularkernel and the IK bandwidth.174B.1. Supplemental Figures and TablesTable B.1: Insurance Coverage by FPL Income Grouping- DescriptiveStatisticsInsurance Type 17-21 Years ≤19 Years >19 YearsPoor: Family Income <100% FPLAny Insurance 0.572 0.661 0.468(0.495) (0.473) (0.499)Private Insurance 0.169 0.175 0.161(0.375) (0.38) (0.368)Public Insurance 0.419 0.504 0.320(0.493) (0.500) (0.467)Observations 52,926 28,373 24,553Near Poor: Family Income 100% - 124% FPLAny Insurance 0.556 0.659 0.448(0.497) (0.474) (0.497)Private Insurance 0.225 0.220 0.231(0.418) (0.414) (0.421)Public Insurance 0.348 0.457 0.234(0.476) (0.498) (0.423)Observations 14,748 7,508 7,240Low Income: Family Income 125% - 199% FPLAny Insurance 0.558 0.648 0.462(0.497) (0.477) (0.499)Private Insurance 0.334 0.362 0.304(0.472) (0.481) (0.46)Public Insurance 0.245 0.308 0.176(0.430) (0.462) (0.381)Observations 41,808 21,472 20,336Middle Income: Family Income 200% - 399% FPLAny Insurance 0.706 0.784 0.609(0.456) (0.411) (0.488)Private Insurance 0.608 0.666 0.536(0.488) (0.472) (0.499)Public Insurance 0.116 0.137 0.090(0.32) (0.344) (0.286)Observations 67,878 37,341 30,537High Income: Family Income >399% FPLAny Insurance 0.861 0.905 0.810(0.346) (0.293) (0.392)Private Insurance 0.829 0.869 0.783(0.376) (0.338) (0.412)Public Insurance 0.052 0.056 0.048(0.223) (0.23) (0.214)Observations 55,179 29,579 25,600Notes: These statistics reflect the proportion of individuals with a specific insurancetype in a given month. Standard errors in parentheses. Those 19 years and undercomprise of 17 to 19 year olds, while those over 19 years are between 19 and 21years of age. Since insurance coverage information is collected each month, anindividual forms up to 24 observations in the analysis. Data come from the MedicalExpenditure Panel, years 1997-2006.175B.1. Supplemental Figures and TablesTable B.2: Change in Medical Care Use at 19 by GenderOffice VisitsSpecification All Diagnosis or Checkup Mental New Prescriptions Dental VisitsTreatment CounselingPanel (a): AllMean 0.1140 0.0622 0.0245 0.0109 0.0433 0.0613Nonparametric: IK -0.0065 -0.0076 -0.0026 -0.0004 -0.0027 -0.0103**(0.0057) (0.0045) (0.0031) (0.0006) (0.0029) (0.0039)Observations 61,501 74,619 61,501 61,501 68,073 74,619Bandwidth 0.812 0.939 0.820 0.778 0.904 0.944Nonparametric: 1 yr -0.0064 -0.0072 -0.0032 -0.0004 -0.0027 -0.0103**(0.0058) (0.0045) (0.0029) (0.0009) (0.0029) (0.0038)Observations 74,619 74,619 74,619 74,619 74,619 74,619Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Panel (b): MalesMean 0.0890 0.0491 0.0166 0.0096 0.0328 0.0532Nonparametric: IK -0.0019 -0.0063 -0.0021 0.0013 0.0031 -0.0007(0.0053) (0.0046) (0.0029) (0.0009) (0.0028) (0.0050)Observations 41,216 37,854 37,854 31,125 27,769 44,577Bandwidth 1.041 0.947 0.980 0.814 0.686 1.116Nonparametric: 1 yr -0.0017 -0.0060 -0.0021 0.0014 0.0006 0.0004(0.0054) (0.0046) (0.0029) (0.0009) (0.0027) (0.0046)Observations 37,854 37,854 37,854 37,854 37,854 37,854Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Panel (c): FemalesMean 0.1403 0.0759 0.0327 0.0123 0.0543 0.0697Nonparametric: IK -0.0112 -0.0086 -0.0041 -0.0021 -0.0056 -0.0210***(0.0082) (0.0076) (0.0047) (0.0014) (0.0053) (0.0044)Observations 36,765 46,421 33,567 33,567 43,189 39,966Bandwidth 0.979 1.179 0.906 0.887 1.087 1.054Nonparametric: 1 yr -0.0111 -0.0085 -0.0043 -0.0021 -0.0062 -0.0213***(0.0083) (0.0076) (0.0045) (0.0016) (0.0055) (0.0045)Observations 36,765 36,765 36,765 36,765 36,765 36,765Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Notes: Notes: The RD coefficients at age 19 are reported, without covariates. Standard errors are in parentheses and were two-wayclustered by individual and age (in months). All nonparametric coefficients are derived using local linear regression. All variables respondto medical care use in a given month. The sample consists of individuals above 125% of the Federal Poverty Line. Data come from theMedical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.176B.1.SupplementalFiguresandTablesTable B.3: Change in Medical Care Use at 19, With and Without CovariatesOffice VisitsSpecification All Diagnosis or Checkup Mental New Prescriptions Dental VisitsTreatment CounselingMean 0.1140 0.0622 0.0245 0.0109 0.0433 0.0613Panel (a): Nonparametric, IK bandwidthWithout Covariates (Table 3.8)RD coefficient -0.0065 -0.0076 -0.0026 -0.0004 -0.0027 -0.0103**(0.0057) (0.0045) (0.0031) (0.0006) (0.0029) (0.0039)Observations 61,501 74,619 61,501 61,501 68,073 74,619Bandwidth 0.812 0.939 0.820 0.778 0.904 0.944With CovariatesRD coefficient -0.0073 -0.0079 -0.0028 -0.0003 -0.0030 -0.0108**(0.0058) (0.0046) (0.0029) (0.0008) (0.0029) (0.0039)Observations 61,018 74,043 61,018 61,018 67,546 74,043Bandwidth 0.812 0.939 0.820 0.778 0.904 0.944Panel (b): Nonparametric, 1 year bandwidthWithout Covariates (Table 3.8)RD coefficient -0.0064 -0.0072 -0.0032 -0.0004 -0.0027 -0.0103**(0.0058) (0.0045) (0.0029) (0.0009) (0.0029) (0.0038)Observations 74,619 74,619 74,619 74,619 74,619 74,619Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000With CovariatesRD coefficient -0.0072 -0.0075 -0.0035 -0.0003 -0.0031 -0.0108**(0.0059) (0.0046) (0.0028) (0.0009) (0.0029) (0.0038)Observations 74,043 74,043 74,043 74,043 74,043 74,043Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Notes: The RD coefficients at age 19 are reported, with and without covariates. The covariates included are: dummies for gender, white race, livein a MSA, full-time student, married, still live with parents, survey year, as well as log of family income (in 2000 dollars). Standard errors are inparentheses and were two-way clustered by individual and age (in months). All nonparametric coefficients are derived using local linear regression.All variables respond to medical care use in a given month. The sample consists of individuals above 125% of the Federal Poverty Line. Data comefrom the Medical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.177B.1.SupplementalFiguresandTablesTable B.4: Change in Expenditures on Office and Dental Visits at 19, With and Without CovariatesOffice Expenditures Dental ExpendituresOut-of-Pocket Total Out-of-Pocket TotalSpecification All If Visit All If Visit All If Visit All If VisitMean $4.26 $37.34 $17.49 $153.38 $6.15 $100.41 $13.73 $224.01Panel (a): Nonparametric, IK bandwidthWithout Covariates (Table 3.9)RD coefficient 1.36* 14.59*** 0.56 5.11 -0.66 4.19 -3.11* -25.79(0.55) (3.05) (0.93) (11.25) (1.02) (18.72) (1.40) (20.84)Observations 68,073 8,185 54,961 9,686 147,079 6,262 140,484 7,107Bandwidth 0.882 0.932 0.732 1.146 1.901 1.391 1.82 1.559With CovariatesRD coefficient 1.33* 14.61*** 0.43 5.61 -0.67 7.53 -3.19* -20.88(0.54) (2.83) (0.91) (12.08) (1.03) (18.93) (1.40) (19.90)Observations 67,546 8,151 54,528 9,646 145,793 6,244 139,279 7,084Bandwidth 0.880 0.930 0.730 1.150 1.900 1.390 1.820 1.560Panel (b): Nonparametric, 1 year bandwidthWithout Covariates (Table 3.9)RD coefficient 1.22* 13.73*** 0.47 13.23 -0.86 4.22 -3.79* -28.05(0.57) (3.29) (1.27) (10.58) (1.42) (21.54) (1.93) (25.39)Observations 74,619 8,185 74,619 8,185 74,619 4,178 74,619 4,178Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000With CovariatesRD coefficient 1.19* 13.75*** 0.35 13.93 -0.86 7.67 -3.87* -22.44(0.56) (3.04) (1.27) (11.32) (1.43) (21.98) (1.93) (24.35)Observations 74,043 8,151 74,043 8,151 74,043 4,166 74,043 4,166Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000Notes: The RD coefficients at age 19 are reported, with and without covariates. The covariates included are: dummies for gender, white race, livein a MSA, full-time student, married, still live with parents, survey year, as well as log of family income (in 2000 dollars). Standard errors are inparentheses and were two-way clustered by individual and age (in months). All nonparametric coefficients are derived using local linear regression.All variables respond to medical expenditures in a given month. The sample consists of individuals above 125% of the Federal Poverty Line. “All”includes all individuals in the sample, while “If Visit” includes only those with a particular visit. Data come from the Medical Expenditure Panel,years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.178B.1.SupplementalFiguresandTablesTable B.5: Change in Ability to Afford Care and Health Status at 19, With and Without CovariatesSpecification Can’t Afford Care Health Index Very Good Health Excellent Health Missed School Missed WorkMean 0.0116 0.7846 0.7552 0.4265 0.1966 0.1914Panel (a): Nonparametric, IK bandwidthWithout Covariates (Table 3.11)RD coefficient -0.0076 -0.0016 0.0081 -0.0098 -0.0147 0.0227(0.0053) (0.0057) (0.0083) (0.0146) (0.0139) (0.0157)Observations 13,070 15,263 19,465 19,465 8,628 7,661Bandwidth 1.077 1.232 1.526 1.532 1.214 1.305With CovariatesRD coefficient -0.0076 -0.003 0.0054 -0.0110 -0.0131 0.0211(0.0053) (0.0053) (0.0067) (0.0155) (0.0134) (0.0146)Observations 13,044 15,247 19,445 19,445 8,620 7,654Bandwidth 1.077 1.232 1.526 1.532 1.214 1.305Panel (b): Nonparametric, 1 year bandwidthWithout Covariates (Table 3.11)RD coefficient -0.0081 -0.0023 0.0035 -0.0132 -0.0182 0.0170(0.0055) (0.0059) (0.0088) (0.0151) (0.0144) (0.0181)Observations 11,983 12,070 12,070 12,070 6,661 5,615Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000With CovariatesRD coefficient -0.0081 -0.004 0.0000 -0.0153 -0.0158 0.0161(0.0055) (0.0053) (0.0059) (0.0161) (0.0138) (0.0169)Observations 11,958 12,056 12,056 12,056 6,654 5,610Bandwidth 1.000 1.000 1.000 1.000 1.000 1.000Notes: The RD coefficients at age 19 are reported, with and without covariates. The covariates included are: dummies for gender, white race, live in a MSA, full-time student,married, still live with parents, survey year, as well as log of family income (in 2000 dollars). Standard errors are in parentheses and were two-way clustered by individual andage (in months). All nonparametric coefficients are derived using local linear regression. The sample consists of individuals above 125% of the Federal Poverty Line. Datacome from the Medical Expenditure Panel, years 1997-2006. * significant at 10%; ** significant at 5%; *** significant at 1%.179

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0167115/manifest

Comment

Related Items