Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

The dynamic behaviour of income assistance recipients in British Columbia Barrett, Garry Fergus 1996

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

Item Metadata

Download

Media
831-ubc_1996-090418.pdf [ 6.41MB ]
Metadata
JSON: 831-1.0087822.json
JSON-LD: 831-1.0087822-ld.json
RDF/XML (Pretty): 831-1.0087822-rdf.xml
RDF/JSON: 831-1.0087822-rdf.json
Turtle: 831-1.0087822-turtle.txt
N-Triples: 831-1.0087822-rdf-ntriples.txt
Original Record: 831-1.0087822-source.json
Full Text
831-1.0087822-fulltext.txt
Citation
831-1.0087822.ris

Full Text

THE DYNAMIC BEHAVIOUR OF INCOME ASSISTANCE RECIPIENTS IN BRITISH COLUMBIA by GARRY FERGUS BARRETT B.Ec. (Hons), The University of Sydney, 1988 M.A., The University of British Columbia, 1990 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE STUDIES (Department of Economics) We accept this thesis as conforming to the required standard THE UNIVERSITY OF BRITISH COLUMBIA February, 1996 © Garry Fergus Barrett, 1996 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of €^COAOiWlCS> The University of British Columbia Vancouver, Canada Date DE-6 (2/88) 11 ABSTRACT This thesis investigates the labour market behaviour and program participation of social assistance recipients. The research is based on the analysis of a unique data set derived from the adrriinistration of the income assistance programs in British Columbia (B.C.) for the period 1980-1992 to analyse the length of time individuals and families spend on welfare. In chapter two the growth in social assistance caseloads and expenditures is documented and the institutional features of the program are laid out. The B.C. income assistance administrative data are then used to examine the changing demographic composition of the caseload. It is found that a substantial proportion of the B.C. caseload is comprised of employable single men and women without children. Further, an examination of program exit rates reveals that most welfare spells are relatively short; however, there is a very high incidence of recidivism. The deterrninants of the length of welfare spells are analysed in more detail in chapter three. More sophisticated econometric models are estimated which control for observed characteristics and "unobserved heterogeneity" (omitted variables). The estimated models are used to test for several forms of program dependence. It is found that there is strong evidence of negative duration dependence in program exit rates. This implies that the program, to some extent, acts as a "trap" whereby exit from the program becomes less likely the longer an individual remains on the program. Further, the evidence indicates the presence of negative occurrence dependence and negative lagged duration dependence, which imply that participating in welfare has a scarring effect on recipients' labour market careers. Policy implications of these finding are discussed. The research presented in chapter four uses data generated from the operation of the Unemployment Insurance (UI) and income assistance programs to test for broader forms of program dependence. Specifically, this research examines whether participating in welfare leads to greater reliance on Unemployment Insurance. After controlling for UI program parameters, demand conditions and individual characteristics, it is found that the UI exit rate for individuals with a recent welfare history was almost identical to that of other beneficiaries. iii TABLE OF CONTENTS Abstract ii Table of Contents iii List of Tables v List of Figures vi Acknowledgment vii Chapter One. Introduction and Overview 1 1.1. Outline of the contents and contribution of each chapter 4 a) Chapter 2-The dynamics of welfare participation in British Columbia 4 b) Chapter 3-The duration of income assistance spells in British Columbia: tests for state dependence 5 c) Chapter 4-Unemployment Insurance spells and welfare receipt in British Columbia 6 Chapter Two. The Dynamics of Welfare Participation in British Columbia 8 2.1. Introduction 8 2.2. Institutional Features 11 2.3. The Data 14 2.4. The Length of Completed Welfare Spells 17 a) welfare exit patterns 18 b) welfare dependence 20 c) business cycle effects 21 2.5. The Length of Off-Welfare Spells 23 2.6. Characteristics of Repeat Welfare Spells 25 2.7. Conclusions 28 Chapter Three. The Duration of Income Assistance Spells in British Columbia: Tests of State Dependence 47 3.1. Introduction 47 3.2. Theoretical Models 49 a) Static Model of Program Participation 49 b) A Dynamic Model of Program Participation 51 c) The Implications of Unobserved Heterogeneity 55 3.3. Methods 56 3.4. The Data 60 3.5. Empirical Results 66 a) Empirical Hazard Rate Estimates 66 b) Duration Model Estimates 67 i. Single men and women 67 ii. Couples with and without children 74 iii. Lone parent families 78 3.6. Conclusions 82 iv Chapter Four. Unemployment Insurance Spells and Welfare Receipt in British Columbia 111 4.1. Introduction 111 4.2. Economic Model 112 4.3. Econometric Methods 116 4.4. The Data 118 4.5. Empirical Results 122 a) Summary Statistics 122 b) Duration Model Estimates 124 c) Baseline Hazard Functions and Expected Duration 128 4.6. Conclusions 130 Chapter 5. Conclusions 145 Bibliography 150 LIST OF TABLES Table Page 2.1 General Assistance Recipients (including dependents) as a Percentage of the Provincial Population, 1970-1992. 31 2.2 General Assistance Expenditures Per Capita, 1970-1992. (1992 constant dollars) 32 2.3 Percent of Total Person-Months on Welfare in British Columbia by Year 33 2.4 Welfare Survivor Function Estimates 34 2.5 Welfare Survivor Function Estimates-Differentiating by Recession Start Date 3 5 2.6 Survivor Functions for Off-Welfare Spells 36 2.7 Survivor Functions for Repeat Welfare Spells 37 2.8 Survivor Functions for Repeat Welfare Experiences by Length of Previous Off-Welfare Spell 38 3.1 Potential Income Assistance Benefits in British Columbia, 1980-1992 (at 1992 constant dollars) 86 3.2 Descriptive Statistics for Single Men and Women Sample 87 3.3 Descriptive Statistics for Couples Sample 88 3.4 Descriptive Statistics for Lone Parent Family Sample 89 3.5 Duration Model Estimates, Single Men and Women Sample 90 3.6 Average (Per Month) Baseline Hazard Rate Estimates, Single Men and Women Sample 92 3.7 Average (Per Month) Baseline Hazard Rate Estimates, Single Men and Women Subsample 93 3.8 Differences in Baseline Hazard Rates (and Wald Test Statistics), Single Men and Women Sample 94 3.9 Duration Model Estimates, Couples With and Without Children Sample 95 3.10 Average (Per Month) Baseline Hazard Rate Estimates, Couples With and Without Children Sample 97 3.11 Average (Per Month) Baseline Hazard Rate Estimates, Couples With and Without Children Sample 98 3.12 Differences in Baseline Hazard Rates (and Wald Test Statistics), Couples With and Without Children Sample 99 3.13 Duration Model Estimates, Lone Parent Family Sample 100 3.14 Average (Per Month) Baseline Hazard Rate Estimates, Lone Parent Family Sample 102 3.15 Average (Per Month) Baseline Hazard Rate Estimates, Lone Parent Family Sample 103 3.16 Differences in Baseline Hazard Rates (and Wald Test Statistics), Lone Parent Family Sample 104 4.1 Unemployment Insurance Spell Sample Statitics 132 4.2 Duration Model Estimates, Non-Welfare History Sample 134 4.3 Duration Model Estimates, Welfare History Sample 136 4.4 Baseline Hazard Rate Estimates 138 4.5 The Expected Duration of UI Spells and the Marginal Impact of Covariates 139 vi LIST OF FIGURES Figure Page 2.1 BC Social Assistance Caseload, Jan. 1980-Dec. 1992 39 2.2 Welfare Exit Rates by Family Type, Employable Men 40 2.3 Welfare Exit Rates by Family Type, Employable Women 41 2.4 Welfare Exit Rates by Family Type, Employable Couples 42 2.5 Welfare Exit Rates by Family Type, Unemployable Men 43 2.6 Welfare Exit Rates by Family Type, Unemployable Women 44 2.7 Welfare Exit Rates by Family Type, Unemployable Couples 45 2.8 Off-Welfare Exit Rates 46 3.1 Budget Set with Welfare Program 105 3.2 Empirical Hazard Rate Function: BC Welfare Spells 106 3.3 Empirical Survival Function: BC Welfare Spells 107 3.4 Baseline Hazards: Single Men and Women Sample 108 3.5 Baseline Hazards: Couples With and Without Children Sample 109 3.6 Baseline Hazards: Lone Parent Family Sample 110 4.1 UI Benefit Receipt Empirical Hazard Function 140 4.2 UI Benefit Receipt Empirical Survival Function 141 4.3 Time Until Exhaustion Empirical Hazard Function 142 4.4 Baseline Hazard Functions: Weeks Until Exhaustion Constant 143 4.5 Baseline Hazard Functions: Time Until Exhaustion Varying with Duration 144 vii ACKNOWLEDGMENT During the course of researching and writing this thesis I have received assistance and wise counsel from many sources. First and foremost I thank the British Columbia Ministry of Social Services (MSS) for providing access to the data on which the thesis is based. I am very grateful to thank Bill Warburton and Nick Bailey, of MSS, for answering numerous questions concerning the data and the operation of the income assistance programs in British Columbia. I am especially grateful to my thesis supervisor, Craig Riddell, for his insightful comments, constructive suggestions and encouragement at every stage of this project. The members of my supervisory committee, Denise Doiron and David Green, were likewise very supportive, were always very quick in providing comments on written work and made innumerable good suggestions for improving the thesis. I also thank John Cragg, David Donaldson, Martin Dooley and Jon Kesselman for their comments on the material in individual chapters. I wish to thank my fellow grad students, and the participants in the Department of Economics Labour Lunch Workshop during 1994-1995, for making the department such a stimulating environment in which to study. Last, but most importantly, I thank Robyn Dowling for the perpetual encouragement, support and distraction she provided throughout the course of the thesis. 1 Chapter 1. Introduction and Overview The present system of social assistance programs in Canada is based on the Canada Assistance Plan (CAP) of 1966. CAP consolidated the existing set of disparate, categorical welfare programs with comprehensive provisions which sought to ensure the availability of financial assistance for all individuals and families in need. In recent times there has been a dramatic increase in the CAP caseload. For example, over the 1980's the total number of social assistance recipients in Canada increased by over 100 percent; in British Columbia the increase was over 125 percent. Over the same period there was a corresponding increase in total real expenditures under CAP by 90 percent for Canada as a whole, and by 100 in British Columbia. As a result, the social assistance programs have come under increasing public scrutiny, with the federal government currently considering major reforms to the income security system, of which social assistance is an important component.1 Despite the significance of the social assistance system (or welfare, as it is popularly known), both in terms of the people it directly benefits and the public resources devoted to it, relatively little is known about how people interact with the system. The major contribution of this dissertation is the analysis of a unique data set derived from the administration of the welfare program in British Columbia (B.C.) over the period 1980-1992. Specifically, the data is used to analyse the determinants of the duration of welfare spells in B.C., with a focus on testing for various forms of program dependency. 1 The Canada Assistance Plan was replaced with the federal Canada Health and Social Transfer (C.H.S.T) on March 31,1996. With the C.H.S.T., the federal government transfers a block of funds to each province for the combined areas of health, education and social assistance. The B.C. government revamped the provincial welfare programs with the B.C. Benefits Program, the introduction of which coincided with the expiration of CAP. Most other provinces in Canada have implemented, or are in the process of initiating, major changes to their welfare programs. 2 A major reason for the lack of knowledge regarding how people interact with the welfare system has been the lack of suitable, publicly available data. Only recently have researchers begun examining the incidence of social assistance receipt using annual, cross sectional survey data. For instance, Allen (1993) and Charette and Meng (1994) analysed the incidence of welfare receipt among female headed households using the Census for 1985 and the Labour Market Activity Survey (LMAS) for 1989 respectively. Dooley (1994) analysed changes in the annual incidence of welfare income among lone mothers using multiple cross-sections of the Survey of Consumer Finances (SCF) for the period 1973-1991. These studies provide new and important information on the characteristics of female heads and lone mothers who participate in the welfare program in a given year. However, the studies do not examine the other family types, such as single men and women, who represent a substantial portion of the welfare caseload at a point in time and account for much of the dramatic increase in the caseload in recent years. Furthermore, these studies do not provide any guide to the dynamics of participation, such as whether individuals remain reliant on the program for a long period of time. Lastly, it is well known that welfare income is subject to substantial underreporting in the above annual survey data, which raises concerns regarding the reliability of the research findings. The research presented in this dissertation is based on a large longitudinal data set which only recently became available to researchers outside of government. The data record monthly information on 10 percent of individuals who ever received welfare benefits in BC during the period 1980-1992. These data have a number of properties which make them highly appropriate for the analysis of the dynamics of welfare participation. Firstly, the data follow a very large cross-section of people (over 87,000) for an extended period of time (156 months). Therefore, the analysis is not beset with the problem of a small sample of welfare recipients, common to 3 analyses based on annual surveys, and permits examination of all household types rather than just lone mothers. Secondly, the data have a fine level of time aggregation, monthly, corresponding to the time period by which the program is administered. Therefore, it is possible to accurately date the beginning and end dates of welfare spells and hence determine the precise length of spells. This is an important property for the analysis of welfare dynamics and is rare for longitudinal survey data sets, such as the Panel Survey of Income Dynamics and the National Longitudinal Survey in the United States, which only record annual information on welfare participation. Thirdly, the administrative data were generated from the computerised case records of the BC welfare program and therefore provide accurate information on program participation. Although all data likely contain measurement error, an advantage of the administrative data is that they are not subject to the problems of underreporting or recall error as typically experienced in annual, retrospective survey data. However, the data used in the dissertation have several limitations. Like most administrative data, there is no information on individuals when they are not participating in the program. In particular, there is no information on the labour market status or income sources of individuals when they are off welfare, and consequently it is not possible to identify the event (for example, employment or marriage) leading to a recipient's exit from the program. Additionally, the data contain only limited demographic information on participants. The information is specifically that which is used in the administration of the program. As a result, information such as the recipient's level of education or the age of dependent children, information which would help improve our understanding of the people who receive welfare and would be useful in evaluating potential program reforms, is not recorded in the data set. In an attempt to address 4 this problem, proxies are used for some variables (such as the minimum wage as a measure of potential labour market earnings). Furthermore, the potential biases introduced by omitted variables are considered and appropriate techniques are implemented in the analysis. The following section briefly outlines the content of each of the substantive chapters, and highlights the major findings. The material presented in Chapter 2 is largely descriptive in nature and is included as background to the analyses presented in Chapters 3 and 4, which contain the main contributions of this thesis. 1.1. Outline of content and contribution of each chapter a) Chapter 2: The dynamics of welfare participation in British Columbia The research in Chapter 2 uses essentially descriptive techniques to document the various dimensions of the time pattern of welfare participation in B.C. This research provides new information on who uses welfare, distinguishing between the set of people on welfare at a point in time and the population on welfare over a period of time. This information is important since it challenges conventional perceptions of who uses welfare, perceptions which are based on anecdotal evidence at best. In addition this analysis provides a basis for considering alternative program reforms aimed at more effectively promoting the transition from welfare to work while reducing the cost of the program. More specifically, the research in chapter 2 documents the major trends in welfare caseloads and expenditures in Canada over recent decades and details the changes in the demographic composition of the B.C. welfare caseload during the period 1980-1992. Next, the length of completed welfare spells are summarised and program exit rates are compared across demographic groups. The research then describes the extent of welfare recidivism and examines 5 the length of time between spells (off-welfare spells), plus the length of repeat welfare spells, to highlight the episodic nature of many individuals' and familys' interaction with the program. b) Chapter 3. The duration of income assistance spells in British Columbia: tests for state dependence The major contribution of the thesis is in the research presented in Chapter 3. This research uses sophisticated econometric methods to parameterise the effects of demographic characteristics and labour market conditions on the program exit rate. The contribution of this work is that is uses high quality data and recent econometric techniques to measure the sensitivity of welfare exit rates to personal characteristics, developments in the labour market (such as the state of the business cycle and the level of wages) and program parameters (the benefit level). This analysis provides new information on the determinants of the length of time individuals and families spend on the program. Additionally, the research examines the evidence for various notions of "welfare dependency" where participation on the program subsequently leads to greater reliance on the program. This is an important contribution of the research for it provides new results on whether the welfare acts as a "trap" for those who enter the program, with the probability of escaping the system declining with the length of time on the program. It is also possible that the experience of being on welfare has a scarring effect on a person's labour market career (possibly due to changes in a person's preferences or perhaps due to employer screening) with the damaging effects increasing with greater interaction with the system over time. The presence of these different forms of welfare dependency may then suggest different policy responses from government. The research presented in chapter 3 also makes a contribution to empirical analysis of welfare spells. The most influential paper in this area is the work by Blank (1989) using data for the United States. Blank analysed the duration of the first observed welfare spell by single mothers, using a 72 month time window. In the analysis based on parametric proportional hazard models, Blank controlled for unobserved heterogeneity in the baseline hazard and tested for duration dependence and the impact of program parameters. The analysis in chapter 3 uses semiparametric methods (allowing a more flexible form for the baseline) and controls for the effects of unobserved heterogeneity on both the estimates of the baseline hazard and the covariate parameters. Furthermore, the longer time period of the B.C. administrative data, compared to Blank's sample, provide a large sample of observations on repeat spells and hence permit the testing for more general forms of state dependence. Likewise, the size of the B.C. sample (over 167,000 spell observations) means that it contains sufficient cross-sectional and time series variation to identify the effects of labour market conditions and program parameters on welfare exit rates, which was not empirically feasible with Blank's sample of 508 observations. c) Chapter 4. Unemployment Insurance spells and welfare receipt in British Columbia The research in chapter 4 extends the analysis of the dependency effects of welfare by examining whether the receipt of social assistance subsequently induces longer spells on Unemployment Insurance (UI). That is, this analysis empirically tests for a form of program dependence whereby welfare participation spills over into greater reliance on the UI program. This interaction between welfare and UI is examined because the programs represent the most important income security programs for Canada's working-age population. An important contribution of the research presented in chapter 4 is that it analyses the 7 welfare and UI programs together, as key components of the Canadian income security system. There is an established, and expanding, literature examining the use and behavioural effects of the UI program in Canada, and as noted above researchers have recently begun examining the incidence of welfare receipt in Canada. However, this work, with few exceptions,2 has examined each program separately without considering them as part of a broader income security system. Furthermore, the analysis contributes to the empirical literature on the dependency effects of welfare participation. In chapter 3, various avenues by which welfare participation may lead to greater reliance on welfare in the future were examined. In chapter 4, the potential for welfare participation to influence an individual's participation in the Unemployment Insurance program, corresponding to a broader form of public-assistance dependence, is examined. The results of the empirical analysis provide new information on how individuals interact with both programs and can contribute to the design of policies aimed at reducing dependence on UI. For example, if a significant negative effect of welfare participation on subsequent UI spells is found then it may be effective to target special policies at UI beneficiaries with a welfare history. Alternatively, if no effect of welfare on subsequent UI behaviour is found then special programs for claimants with a welfare history may be unwarranted. The main exceptions are Bruce et.al. (1993) and Fortin et.al. (1994). Chapter 2. The Dynamics of Welfare Participation in British Columbia1 8 2.1 Introduction The Canadian social assistance programs were established under the Canada Assistance Plan of 1966 to provide financial assistance to all individuals and families in need. Over the past two decades the number of social assistance recipients has increased by almost 120 percent and total government spending on welfare has more than tripled in real terms. Consequently the welfare programs have come under increasing public and governmental scrutiny, and the federal and provincial governments have either begun implementing, or are considering, major reforms to the social assistance programs. In order to evaluate alternative welfare reform proposals it is important to be informed about the time pattern of welfare use and how it varies with recipients' characteristics. For example, recipients may primarily use welfare as a form of transitional support, easing a financial crisis during a brief period while out of the labour force. Alternatively, recipients may use welfare as a substitute for labour market income and thereby remain on the program for a long period of time. If the welfare population is comprised of both types of users, then it is important for the design of effective policies to be able to identify and target the different groups. Despite the significance of the welfare system in Canada, both in terms of the people it directly benefits and the public resources devoted to it, there is very little knowledge of how people interact with the system. A major reason for the lack of published research on the use of 1 After writng an ealier version of this chapter, I discovered that another reseacher had independently undertaken a similair analysis. Subsequently, we collaberated in revising the research and hence the present version of Chapter 2 is coauthored with Michael Cragg. 9 social assistance in Canada has been the lack of suitable data sets. To study the dynamics of welfare participation it is necessary to have a panel data set that both follows individuals for a relatively long period of time and records program participation information. Until recently, there has been no publicly available Canadian longitudinal data set.2 Several researchers have recently examined the incidence of social assistance receipt using annual, cross-sectional survey data. Allen (1993) and Charette and Meng (1994) analysed the incidence of welfare participation among Canadian female headed households using the Census for 1985 and the Labour Market Activity Survey (LMAS) for 1989 respectively. Dooley (1994) analysed changes in the incidence of social assistance receipt among lone mothers with multiple cross-sections of the Survey of Consumer Finances (SCF) from the period 1973-1991.3 The findings of these studies are generally consistent with a static model of welfare participation and provide important information on the characteristics of lone mother families that participate in the welfare programs. However these studies do not examine other household types, especially single men and women who represent a substantial portion of the welfare caseload at a point in time and account for much of the dramatic increase in the caseload in recent years. Furthermore, these studies do not provide any guide to the dynamics of participation, such as whether welfare recipients remain on the caseload for a long period of time. To examine the intensity of welfare use, and whether individuals are reliant on the program over a period of time, 2 The first Canadian panel survey was Statistics Canada's Labour Market Activity Survey. The initial panel followed individuals for the years 1986 and 1987. The second, and last, panel followed individuals from 1988 to 1990. Although this data source has rich information on the labour market behaviour of the interviewees, the short time period of the panels severely limits their usefulness in analysing the dynamics of welfare participation. Furthermore, sources at Statistics Canada suggest that up to 20 - 30% of social assistance is unreported in the LMAS. There is substantial under reporting of social assistance income in the Census and SCF data as well as the LMAS, which raises the issue of the accuracy of the model estimates based on single cross-sections. Dooley's (1994) focus on changes in the incidence of welfare receipt using multiple cross-sections of the SCF should minimise the biases that may result from the under reporting of welfare receipt. 10 it is necessary to directly model the distribution of welfare spell durations. The vast majority of information about the experiences of low-income individuals with labour market and welfare programs has been generated using U.S. data (see Moffit (1992) for a review). The general conclusions of this literature are that most welfare recipients are single mothers, that most welfare spells are short (60 percent are shorter than 2 years) while the majority of use is through long spells (Bane and Ellwood (1983,1994), Gritz and MaCurdy (1992) ), and that the notion of welfare being a migration magnet is unwarranted (Walker (1993) ). Finally, entry into welfare is generally not due to divorce or childbearing but rather changes in labour market status. Exit is through marriage and work (Bane and Ellwood (1994), Gritz and MaCurdy (1992)). None of these conclusions need hold for Canada because institutionally the two systems are very different. For example, single men and couples without children are eligible to receive benefits in Canada but not in the U.S., and the Canadian welfare system is relatively much more generous (see Blank and Hanratty, 1993). Hence in an environment of policy reform, we are largely uninformed about the dynamics of welfare use in Canada. The research presented in this chapter utilises a unique data set, derived from the administration of social assistance programs in the Province of British Columbia, to analyse the time pattern of welfare receipt over the period 1980-1992. From the raw monthly caseload data continuous spells of welfare receipt are constructed. The dynamic pattern of welfare is addressed by analysing the length of completed welfare spells in B.C.. The duration of welfare spells is first summarised according to various demographic characteristics and then the information is refined to describe recidivism, the length of time between welfare spells and the incidence and length of repeat spells. 11 This chapter is organised as follows. In the next section the institutional features of the Canadian welfare system are outlined. In section 3 the data used in the analysis are described, and descriptive statistics are presented. In section 4, results from the estimation of simple, nonparametric duration models are used to explain the time pattern of welfare exits. In section 5 the length of time off-welfare (the welfare re-entry rate) is examined and the high incidence of welfare recidivism is documented. In section 6 the length of repeat welfare spells is analysed. The final section concludes the chapter and draws out several policy implications of the main findings. 2.2. Institutional Features The universal social assistance programs in operation in Canada until April, 1996, were established under the Canada Assistance Plan (CAP) of 1966.4 The stated objective of the Canadian welfare system is to provide financial assistance to individuals and families whose resources are inadequate to meet their needs, irrespective of the cause of the hardship. Under CAP, the federal government sets broad guidelines on the eligibility criteria and implementation of the "needs test" and undertakes to share equally with the provinces the costs of those programs.5 The provinces are responsible for administering the welfare programs and have much discretion in determining the rules and benefit structures of their separate programs. Consequently there are considerable differences in the welfare programs across the provinces and territories of Canada.6 4 The Canada Assistance Plan was superceded by the federal Canada Health and Social Transfer on April 1,1996. Since 1991 the Federal government has placed an upper limit on the total payments under CAP to the three "have" provinces of British Columbia, Alberta and Ontario. As a result, the Federal government presently funds significantly less than half the CAP program costs in these provinces. 6 In the provinces of Nova Scotia, Ontario and Manitoba there is a further division between the province and municipalities in the design and administration of welfare programs. 12 In assessing a household's needs, the individual's or family's employability is first established, which determines the asset exemption level applicable to the household.7 If the value of household assets is lower than the maximum allowable level, then the household's budgetary needs are calculated. The provisions of CAP prescribe that the assessment of need must take into account the basic living items of food, shelter, clothing, utilities, and personal and household expenses. The provinces set maximum allowable amounts for each of these items. Next, the resources available to the household to meet those requirements are calculated. The resources include earned income, alimony and maintenance payments and government transfers such as unemployment insurance and pension income. A deficit between assessed needs and available resources qualifies the household for assistance. The actual amount of assistance paid depends on the employability status, family status and size of the household.8 Since the inception of CAP there has been a dramatic rise in the number of Canadians in receipt of welfare at a point in time. Table 2.1 presents the number of general assistance recipients, as a proportion of the provincial or territorial population, for selected years between 1970 and 1992. Over this period, the fraction of the Canadian population in receipt of welfare increased by 3.8 percentage points from 5.71% to 9.54%.9 The onset of the recessions in 1974-75, 1981-82 and 1990-1992 resulted in large increases in the welfare caseload. Disturbingly, there were no significant declines in the caseloads in the years of economic and employment Households classified as employable are generally subject to lower asset exemption levels in the assessment of benefit eligibility. For example, the annual welfare benefits in British Columbia in 1992 were $6308 for single employable men and women, $11373 for an employable single parent with one child and $14389 for an employable couple with two dependent children. Recipients classified as unemployable received an extra $50 per month in benefits. See National Council of Welfare (1993a) for a comparison of the maximum annual welfare benefits payable to a number of household types for each province and territory. The absolute number of welfare recipients increased from approximately 1.2 million people in 1970 to over 2.7 million in 1992, a rise of almost 120 percent. 13 growth subsequent to the recessions. Corresponding to the trend in the welfare caseload has been a substantial increase in the amount of real resources spent on the welfare programs. Table 2.2 presents the total federal-provincial expenditures in per capita terms (at constant 1992 prices) for the general assistance programs in Canada. Over the 20 years from 1970 to 1990 the per capita expenditures on the general assistance welfare programs for all of Canada increased by 151 percent.10 The largest increase in per capita welfare costs occurred in New Brunswick (305 percent), Ontario (184 percent) and Quebec (177 percent).11 The large increases in welfare beneficiaries and program costs have provided a major impetus to reform of the system. The social assistance program administered by the B.C. government, collectively known as GAIN (Guaranteed Available Income for Need) is composed of six separate programs. Three are supplementary programs: GAIN for the handicapped, GAIN for Seniors, and Old Age Security. The other programs provide income assistance: Child in the home of a relative, Age 60-64 benefits and Basic Income Assistance (IA). Although Basic IA forms the residual category for individuals ineligible for the other 5 programs, it represents the predominant share of the total GAIN caseload over the 1980-1992 period. Given the general nature of the Basic IA program and its predominance in the total GAIN caseload, as shown in Figure 2.1, it is the primary focus of this paper.12 As noted above, the employability status of an individual is important in deterrnining their eligibility for assistance as well as the level of benefits. In B.C., during the period analysed here 1 0 Real expenditures on general assistance programs in Canada increased by over 260 percent in absolute dollars during this period. Per capita welfare expenditures in British Columbia increased by 150 percent from 1970-1990, which was very close to the Canadian average. The Basic IA program also accounts for most of the monthly variation in the total welfare caseload; the correlation coefficient between the monthly Basic IA and Total GAIN caseload over the 13 year period is 0.99 . (1980-1992), a person was classified as employable if he\she was not (i) 65 years of age or older (ii) temporarily or permanently unable to work due to medical reasons (iii) a single parent with one dependent child under six months of age or two or more dependent children under 12 years of age13 or (iv) a single parent required to stay at home to care for a disabled child. 2.3. The Data The data used in the analysis are derived from the monthly case records of the social assistance programs in B.C. The raw data are a ten percent random sample of all individuals with an IA history in B.C. during the period of January 1980 to December 1992. The sample consists of 87,288 individual records. Each record contains the individual's (or principal claimant's) birth date, sex and a variable indicating under which B.C. social assistance program, if any, the individual received benefits for each month of the thirteen year period. Additionally, the records include variables indicating the individual's family type, number of dependants14 and employability status for the months that the person was in receipt of social assistance. Table 2.3 provides cross-sectional summary statistics on how the caseload has changed over the decade by presenting the ratio of category specific person-months to the total person-months on the rolls in the various time periods. A number of attributes from this simple summary are worth noting. First, the use of IA reflects it being a safety net for all family types rather than a categorical program like Aid to Families with Dependent Children (AFDC) in the U.S. which is targeted towards children. Only one quarter of IA cases involve single mothers and only 30 percent are associated with families with children. Second, single individuals without dependants 13 The definition of unemployable in relation to the age of the youngest child is most stringent in B.C. and Alberta. In the other provinces and territories, generally a lone parent is classified as unemployable if the youngest child is under school age. However, in 1994 B.C. changed this to under 12 years of age. 1 4 It is noted that the data set does not contain any information on the age of dependent children. 15 represent over 60 percent of welfare recipients in any given month. Third, the use of welfare is evenly distributed over the age spectrum: at a point in time, a quarter are less than 26 years of age, 33 percent are between 26 and 36 and 41 percent are over 3 6.15 Finally, over the last decade more than half of the caseload in a given month has been represented by individuals classified as employable. These facts are very different to the public's perception that welfare is used principally by young single mothers. There appear to be two dominant trends in the dynamics of the aggregate caseload. The top panel of the table shows that over the last decade the proportion of beneficiaries who were employable steadily rose from 38 percent in 1980-82 to 64 percent in 1991-92. Furthermore, the proportion of the caseload accounted for by single males rose from 34 percent in 1980-82 to 44 percent in 1991-92. Every other type of household has fallen as a fraction of the total. These trends coincided with the proportion of total GAIN cases receiving Basic Income Assistance rising from 72 percent in 1980-82 to 81 percent in 1991-92. The age structure of the caseload has changed less dramatically over the decade. While most recipients are over age 25, this fraction has only risen by two percentage points to 74 percent. The source of this change has been an increase in the relative number of recipients between 26 and 36 years of age. To analyse the dynamics of welfare receipt, IA spells were constructed. A spell of IA is defined as a sequence of consecutive months of Basic IA receipt. Some care must be taken in selecting spell data that are appropriate for valid statistical analysis. Spells which began prior to the start of the data period were excluded from the analysis. To include such left-censored spells would require specification of the actual distribution that generated them. In order to avoid For families on IA, the age of the principal claimant only is recorded. As a result, the average age of all recipients of IA would be substantially less than that for principal applicants, since the former includes the dependent children in families that receive IA. 16 making strong distributional assumptions, left-censored spells were dropped from the sample and the results are therefore conditional on all spells beginning after January 1980. For these spells it is possible to determine the precise length of time on welfare unless the spell progressed beyond December 1992. Although such spells are "right-censored" the statistical methods allow for this and thereby provide unbiased estimates of the completed spell distribution. For each spell of IA there is also information on the recipient's sex, marital status, employability status and the number of dependent children at the commencement of the spell. A spell of "off-IA" is defined as the time between the end of an IA spell and the commencement of a new IA spell. From the original set of 87, 288 individuals, a sample of 164,894 Employable and 41,103 Unemployable IA spells was generated. A number of features of the spell sample merit discussion. First, only 20 percent of spells are by individuals or families classified as Unemployable although they account for nearly half of the population in receipt of welfare at a point in time. This is possible because spells by "Unemployables" are longer, which is clearly evident: 75 percent of the IA spells by Employables are from 1-6 months in length whereas only 52 percent of the spells by Unemployables are of this length. A second feature of the spell sample is that single men account for 53 percent of all spells, couples for 16 percent of the total, and single parent families for only 16 percent. Thus, it is clear from a comparison with the statistics on welfare receipt at a point in time (single parents comprise a quarter of the caseload) that single parents have longer spells on average than other family types. Third, over 73 percent of all spells over this time period were by single men (53 percent) or single women (20 percent) without children, further reinforcing the fact that single parents are not the predominant users of welfare. 17 2.4. The Length of Completed Welfare Spells To summarise the distribution of completed welfare spells hazard rate and survivor functions are estimated. The welfare hazard or exit rate at duration T, h(T), is the probability that a welfare spell will end at duration T conditional on it being at least T months in duration. The survivor function at duration J is the probability that a welfare spell is at least T months in duration. Formally, the hazard rate is defined as h(T) = f(T)/(l-F(T)) and the survivor probability is given by S(T) = 1-F(T) where f(.) is the density function and F(.) is the distribution function. The hazard rate directly measures the welfare exit rate as a function of time on the program while the survivor function provides a natural measure of program dependence (ie. the proportion of spells that are "long"). The hazard and survivor functions provide equivalent characterisations of the completed spell distribution and are intuitive and empirically tractable (see Kiefer, 1988). The empirical analysis is based on the empirical hazard rate and survivor probability function estimators.16 The empirical hazard and survival function estimators are nonparametric in that they do not require the specification of a functional form for the underlying spell distribution, and are straightforward to implement. For example, the estimate of the hazard rate at 6 months duration is given by the number of spells that actually terminate at 6 months divided by the Also known as the Kaplan-Meier estimators. 18 number of spells that are 6 months or longer in duration. Furthermore, the empirical hazard and survival function estimators allow for right-censored spells. (a) Welfare Exit Patterns The analysis proceeds by estimating the hazard rate and survivor functions for the sample of welfare spells stratified by the demographic characteristics of recipients. The welfare hazard rate functions are plotted in Figures 2.2-2.7 and the corresponding survivor functions are presented in Table 2.4. There are several striking features of the hazard function plots. Firstly, a significant feature of Figures 2.2-2.7 is the relatively high value of the exit rates, especially during the initial months of a spell. For example, employable single men with no children have an exit probability of 28% at 1 month, 27% at 2 months and 25% at 3 months. Unemployable single women with one dependent child have an exit probability of 16% at 1 month, 15% at 2 months and 13% at 3 months. Therefore the hazard functions reveal that the welfare population is very dynamic for there is a relatively high probability of program exit, for members of all demographic groups, especially in the initial months of a spell. Secondly, by comparing the hazard rate functions across the various demographic groups it is evident that single men, women and couples without children have very similar spell dynamics, especially after the first 12 months of a spell (couples have a higher initial exit rate). However, the presence of dependent children is associated with substantial differences in exit rates: couples have higher exit rates than single fathers who have higher exit rates than single The set of spells that are 6 months or longer in duration represent those spells at risk of terminating at 6 months and therefore are known as the "risk set" at 6 months duration. 19 mothers. While it is unclear why single fathers should leave welfare more quickly than single mothers, the finding for couples is explainable with a model of lower fixed costs of employment for couples: couples do not need to pay for outside child-care. This view is reinforced by Figure 2.2 where the exit rates are presented for groups differentiated by parental status. While the exit rates are lower for single fathers and mothers than for single men and women, the exit rates are virtually the same for couples irrespective of the presence of dependent children. Another important feature of the hazard functions is that the exit rates generally decline with spell length. That is, the longer individuals and families remain on the program the less likely they are to leave it. There are two broad classes of explanation for this finding of negative duration dependence. One is that there may be true duration dependence whereby the experience of being on welfare changes a recipients' preferences or constraints such that they become increasingly less likely to exit the program. Such a behavioural effect of welfare receipt is consistent with models in which welfare spell duration is associated with changes in welfare stigma or individual preferences more generally, human capital atrophy or employer screening. Alternatively, the observed duration dependence may be a statistical artefact, reflecting the effects of individual characteristics or "population heterogeneity". For example, consider a population composed of two types of individuals, the highly motivated and the less motivated. The highly motivated are more likely to exit welfare early, leaving behind a population of recipients composed of an increasing proportion of individuals with low motivation. In these circumstances the nature of the welfare population changes with spell duration, and if the characteristic (motivation) is not controlled for, negative duration dependence will be observed in the aggregate hazard rate even in the absence of true duration dependence.18 18 The distinction between true state dependence and population heterogeneity is laid out more formally in Chapter 3 below. 20 As a first approach to discriminating between these two competing explanations the hazard rate functions are estimated for the sample stratified by an array of observed characteristics. Figures 2.2-2.7 show that the observed negative duration dependence is common to all demographic groups and therefore is not attributable to gender, family type, number of dependent children nor employability status. In a further attempt to control for population heterogeneity, and to understand how welfare is being used, the sample is stratified along additional dimensions in order to control for business cycle effects, the number of welfare spells experienced by a recipient and the time spent off the welfare by multiple spell recipients. These results are discussed more fully below; however, to anticipate the results, it is found that the hazard and survival functions exhibit negative duration dependence after controlling for these observable characteristics. To further discriminate between the hypotheses of true duration dependence or population heterogeneity it is necessary to control for unobserved individual characteristics or "unobserved heterogeneity". This requires using maximum likelihood techniques to estimate duration models which control for observed covariates mixed with a distribution representing the unobserved heterogeneity component. This line of analysis is taken up in chapter 3, where semiparametric proportional hazard models with gamma distributed heterogeneity are estimated and it is found that allowing for unobserved population heterogeneity reduces, but does not eliminate, the extent of negative duration dependence. (b) Welfare dependence While the hazard rates illustrate the basic exit pattern over time, the survivor functions provide one measure of welfare dependency. Survivor functions for welfare usage are presented 21 in Table 2.4. For employable single men without children, only 40% of spells are ongoing after 3 months. That is, a full 60% of spells end within the first 3 months. For employable single women and couples, without children, over 62% of spells end within the first 3 months. As evidenced by the hazard functions, the presence of dependent children is associated with longer spells. Even so, over 50% of the welfare spells by employable single men and couples with children end within 3 months. For employable single women with dependent children, 39% of spells are ongoing after 3 months, 33% after 12 months, 12% after 4 years and 6.9% remain in progress after 6 years. The impact of dependent children on the welfare survival probabilities is much less pronounced for two parent households, which suggests that the long-term welfare receipt by single mothers may be due to problems of overcoming the fixed costs of employment. Furthermore, there are substantial differences in spell length by recipient's employability status. Approximately 40% of spells by unemployables last a full year compared to less than 20% of all spells by those classified as employable. As expected, this latter finding implies that the recipients' inability to work and the presence of very young children in the household are associated with longer welfare spells. Overall, the estimates of the survival functions confirm that most welfare recipients do not remain reliant on the program for a long, continuous period of time. The majority of welfare spells end within 3 months and only approximately 15% remain in progress after 1 year. Of those ongoing after this time, most are experienced by single parents. (c) Business Cycle Effects A further potential source of population heterogeneity that may account for the observed negative duration dependence in the hazard functions is the time aggregation inherent in treating 22 spells as homogeneous irrespective of the calendar date when they commenced. It is possible that individuals who entered the welfare program during a recession remain on the program longer than individuals who entered during an upswing of the business cycle due to weaker labour demand and fewer job opportunities. That is, it is reasonable to expect that welfare spells commenced during a recession will be longer on average, other things equal, than spells that were commenced during the upswing of the business cycle, and thereby partly account for the negative duration dependence observed in the hazard function. Although business cycle conditions may affect the probability of a given welfare spell ending, the occurrence of a recession may also affect the composition of the welfare population. For instance the onset of a recession may be associated with a group of normally financially self-sufficient individuals becoming unemployed and ultimately relying on welfare for a brief period of time, prior to regaining employment. Such individuals may have stronger links to the workforce than the "average" welfare recipient and consequently may experience an atypically short welfare spell. Therefore it is possible that by affecting the composition of the welfare population recessions may be associated with shorter, rather than longer, welfare spell durations. Table 2.5 presents the survivor function values, for the different family types, for the sample of welfare spells stratified according to whether the spell commenced during a recession. Two recessions were experienced during the data period, and are dated as July 1981 - November 1982 and April 1990 - November 1992.19 Spells experienced by employable single men without children which commenced outside of a recession had a higher survival probability than those commenced during a recession. For instance, the survival probabilities were 41.6 % and 33.0% at 3 months, 25.0% and 18.0% at 6 months, and 12.4% and 8.7% after 12 months, for spells 19 Although Statistics Canada does not officially date recessions, the convention is to date recessions according to turning points in aggregate employment. 23 commenced in non-recessionary and recessionary periods respectively. Therefore, for this group, spells commenced in periods of higher unemployment had a shorter expected duration. Additionally, the finding of a lower welfare survival probabilities for spells commenced during a recession is found for all demographic groups. Indeed, even spells by single mothers classified as unemployable had a lower survivor probability if the spell commenced during a recession. For example, spells commenced outside of the recessionary periods have a survivor probability of 64.1% at 3 months, 47.5% at 6 months and 31.6% at 12 months compared to the corresponding survival probabilities for spell commenced during a recession of 60.4% at 3 months, 45.8% at 6 months and 30.7% at 12 months. This finding of an association between recessions and shorter welfare spells is consistent with the results of the analysis by O'Neill et.al. (1984) of AFDC spell durations, where it was found that the unemployment rate had a positive impact on exit rates, and likely reflects compositional effects of recessions on the welfare population. 2.5. The Length of Off-Welfare Spells Features of the time off-welfare before another spell begins provide further clues about how welfare is used. Table 2.6 presents the survival function estimates for the off-welfare spell20durations for various demographic groups. The estimates in the table show that the probability of remaining off welfare, conditional on having previously been in receipt of welfare during the data period, declines substantially over the first 12 months off the program. For single men without children, 71.4% of the off-welfare spells are ongoing after 3 months and only 45.0% are still ongoing after 12 months. That is, a full 55% have returned to welfare within the first 20 Given the two state framework of the statistical model, the time off-welfare corresponds to the time between consecutive welfare spells and hence the off-welfare exit rate corresponds to the welfare re-entry rate. 24 year of leaving the program. Furthermore, 65.1% have returned within 2 years and a substantial 72.3% have returned within 4 years.21 Therefore although single men tend to have very short welfare spells, they also experience a very high rate of return to welfare. Single women, without children, have a slightly higher survival probability and hence a marginally lower probability of eventually returning to welfare compared to single men without children. The survival function estimates for couples without children are very similar to those for single women without children. Therefore the estimates for the time off-welfare survival functions clearly reveal that there is a very high incidence of repeat use of welfare which is experienced by all groups. The presence of dependent children is associated with lower off-welfare survival probabilities and hence a higher incidence of eventually returning to welfare. For example, 45.6% of welfare exits by single women without children have returned to welfare within the first 12 months while 50.0% of the exits by single mothers have returned within the same period. Overall, the presence of dependent children indicates that families are even less able to achieve financial independence and permanently leave welfare. Therefore, although most welfare spells appear to be relatively short, there is a very high incidence of recidivism. Continuous, long-term use and dependency on welfare is not a characteristic of most welfare recipients. However welfare does not appear to be effective in providing transitional support, for most welfare users periodically return to the program. Either many individuals are chronically poor and neither able to gain self-sufficiency nor end a repeating cycle of poverty, or, many individuals are using welfare in conjunction with temporary or Therefore, 55% of off-welfare spells end within the first year, 10.1% end in the second year and a further 7.2% end during the third and fourth years. 25 seasonal jobs. The off-welfare hazard function is plotted in Figure 2.8. The figure clearly shows the decline in off-welfare exit rates over time, especially during the first 12 months. However, the observed negative duration dependence in time off-welfare does have an important caveat. After the initial steep decline in the probability of returning to welfare at the end of the first year off welfare, the probability of returning to welfare rises temporarily. This might be a result of seasonal use of welfare by some individuals. To develop an understanding about which demographic groups show the presence of seasonal returns, the off-welfare hazard functions were estimated for the sample broken down by whether individuals are employable or unemployable, single or married, and with or without children. The striking feature of the results is that the seasonal pattern arises only for employable couples and single individuals without children. Hence, in an economy where seasonal jobs are common, the hypothesis that welfare is being utilised as a form of unemployment insurance22 merits further research. 2.6. Characteristics of Repeat Welfare Spells Given the high incidence of recidivism documented in the previous section, in this section the length of repeat welfare spells is analysed, and compared to the length of initial spells on the program. Additionally, by separating out repeat users of welfare the analysis provides a further exploration of population heterogeneity. Table 2.7 presents survivor function estimates for the sample stratified by the number of the spell in the sequence of spells observed for an individual. The general pattern of the survival 22 That is, such jobs may be not be covered by the regular Unemployment Insurance program, or at least the workers are only able to establish short duration claims. 26 function estimates show that the higher the spell number the greater the survival probabilities and expected durations. The more frequently individuals or couples without children move on and off welfare, the longer are the subsequent stays on welfare. This suggests that there may be "occurrence dependence" in welfare receipt, which parallels Corak's (1993a,b) findings of occurrence dependence in Unemployment Insurance spells in Canada over the period 1971-1 9 8 9 23 The results for the different family types with children have a similar pattern to those for families without children. Repeat spells tend to be progressively longer. An exception is that spells numbered 5 or higher have lower survival probabilities and expected duration than spells numbered 4. Moreover, for single mothers, spells numbered 5 or higher actually have the shortest expected duration. This surprising result likely reflects sample selection bias due to the fixed time period of the data.24 To experience 5 or more spells of welfare receipt during the 13 year data period, individuals and families must move on and off the program relatively quickly. As a result, families which experience 5 or more spells in the data period must have relatively short spells compared to the average spell length for their demographic group. In a further attempt to isolate individuals who are prone to using welfare, and control for potential sources of population heterogeneity, the sample of repeat spells was divided according to the length of time recipients spent off the program immediately prior to returning. The duration of time off-welfare was divided into 4 intervals: less than 3 months, 3-6 months, 7-23 However, this finding must be interpreted with caution due to the potential effects of unobserved heterogeneity. As discussed in relation to duration dependence, the observed occurrence dependence may not be a true behavioural impact of repeat welfare experiences but may simply reflect omitted variables (such as motivation) which are correlated with both longer spells and a greater incidence of recidivism. The analysis in chapter 3 formally tests for the presence of occurrence dependence after controlling for unobserved heterogeneity. 24 -The ideal data set would observe all individuals for their entire working life. This would provide a complete record of individuals' welfare history and no selection bias would arise from a truncated period of observation With the computerisation of program administration, administrative data sets should approach this ideal over time. 27 months and 24 or more months. The breakdown of welfare survival probabilities by the length of the previous off-welfare spell are presented in Table 2.8. Recall that in section 3 it was found that the longer individuals and families were off welfare, the less likely they were to eventually return to the program. This finding is consistent with individuals leaving welfare for employment, where they then accumulate work experience and human capital. Their employment then becomes more secure, and return to welfare less likely, the longer they remain employed. By the same reasoning, it is expected that the longer an individual remains off-welfare, the shorter will be a subsequent welfare spell if they do return to the program. That is, due to the accumulation of human capital and work experience, repeat spells are predicted to be shorter the longer the recipient remained off the program. The top panel of Table 2.8 presents the results for single men without children. For those who were off-welfare for only 1 or 2 months prior to returning, 49% remained on welfare for at least 3 months, 32% for at least 6 months and 17% for at least a year. However, for the group who remained off welfare for over 2 years, only 40% remained on welfare for 3 months, 24% for 6 months and 12% for a year. Therefore, the survival probabilities for repeat welfare spells are substantially lower the longer the recipient had been off welfare. These findings are consistent with the more successful integration into the labour market by individuals who remain off welfare longer. The magnitude and pattern of the survival probabilities by time off welfare for single women and couple without children are very similar to those outlined for single men without children. The presence of dependent children is associated with a uniform 5-10% increase in the survival probabilities for the corresponding family type. Therefore, the presence of dependent children is not associated with a difference in the pattern between time off welfare and the length 28 of repeat spells, but is associated with longer overall stays on welfare. An interesting feature of the survival probability estimates in Table 2.8 is evident by comparing the survival probabilities across the consecutive off-welfare duration categories. The survival probabilities for the 1-2 months off welfare interval are very similar to the 3-6 months interval and likewise the survival probabilities for the 7-24 and 24+ months intervals are very similar. The greatest difference between the survival probabilities is for the 3-6 month interval and the 7-24 months interval. Therefore the first 6 months off welfare appears to be a crucial period for individuals and families making the transition off welfare. The events during this time period have greatest impact on whether they successfully reintegrate into the labour market. Furthermore, this highlights the potential importance of policies designed to provide support services in this initial period after leaving welfare. The provision of services and assistance during this key transition period may have importance consequences for individual's and families' future reliance on the program and hence may have a significant impact on program expenditures over a period of time. 2.7. Conclusions The research presented in this chapter represents a first step in filling the large gap in knowledge concerning how individuals and families use social assistance in Canada. Only recently have researchers begun to examine the decision of lone parent families to participate in Canadian welfare programs using cross-sectional surveys. The research presented in this chapter utilised a unique longitudinal data set derived from the administration of the welfare programs in British Columbia to examine the dynamics of welfare participation. A number of patterns emerged from the data. It was found that most welfare spells are 29 shorter than six months while over 15 percent last longer than a year. Further, very few welfare cases last more than four years, and those involved families with children. Single mothers and fathers have longer spells than either couples (with and without children) or childless single men and women. Additionally, there have been large changes in the caseload composition: the proportion of the caseload which consists of those who are employable has steadily risen from 38 percent in 1980-82 to 64 percent in 1991-92, single males have risen by 10 percentage points from 34 percent of the caseload in 1980-82, while the proportions of all other types of household have fallen. The age structure of the caseload is virtually unchanged over the decade: over 70 percent are over age 25. Finally, a quarter of welfare recipients are back on the welfare rolls within three months of leaving, while a full 50 percent return within a year. We also find that for single individuals and couples without children there is a significant fraction of the population who display a seasonal pattern to their welfare use. Several important policy issues are raised by the findings in this chapter. First, for the large majority of recipients, "welfare dependence," defined in terms of remaining on welfare for a long period of time, does not accurately characterise their experience on welfare. Most spells are relatively short. However the very high incidence of repeat use, especially within the first 6 to 12 months after leaving welfare, suggests the need for governments to implement more active labour market policies targeted to these individuals to help them become independent and permanently self-sufficient. Additionally, there is a subset of single parent families who do remain on the welfare roll for several years continuously. This group accounts for an important fraction of the caseload at a point in time and for a substantial portion of the welfare budget over a period of time. It is likely these families face significant fixed costs of employment. As raised by Bane and Ellwood (1994) 30 in the US context, an important issue for public debate is whether it is desirable that welfare acts as a subsidy to these families or whether a more effective policy targeted specifically at this group, taking account of their special needs and characteristics, would be more effective. Future work needs to be undertaken to more fully understand population heterogeneity, and other labour supply characteristics, that may account for our findings of negative duration dependence. The research presented in Chapter 3 takes up this task and tests for the presence of duration dependence, plus other forms of state dependence, using more sophisticated econometric techniques which allow for the potential impact of unobserved individual characteristics. Another factor which needs to be better understood is the relationship between the UI and IA systems: how much IA use is generated by UI exhaustion, and what are the implications of welfare participation for subsequent reliance on UI? The research presented in Chapter 4 pursues this latter line of inquiry, and tests for the potential spill-over effects from welfare participation to future dependence on UI. Table 2.1 General Assistance Recipients (including dependents) as a Percentage of the Provincial Population , 1970-1992. Province 1970 1975 1980 1985 1990 1992 Newfoundland 15.96 11.32 8.45 8.45 8.27 10.24 P.E.I. 8.04 7.12 7.58 7.49 6.56 8.97 Nova Scotia 6.03 6.45 5.99 8.29 8.65 10.01 New Brunswick 7.95 8.19 9.36 9.52 9.04 10.38 Quebec 7.09 6.56 7.84 10.59 7.92 9.43 Ontario 4.32 4.03 4.05 5.20 6.53 11.13 Manitoba 4.97 5.51 4.40 5.79 6.04 7.23 Saskatchewan 5.52 4.93 4.27 6.22 5.35 5.99 Alberta 4.68 4.30 3.46 5.15 5.82 7.11 B.C. 4.82 6.46 4.46 8.95 6.55 8.03 Yukon n.a. n.a. 4.49 6.10 3.57 5.61 N.W.T.2 13.19 11.13 13.45 16.84 16.61 Canada 5.71 5.52 5.43 7.41 6.95 9.54 Source: Canada Assistance Plan, Annual Report 1970-1992 and Statistics Canada, The Labour Force, 1970-1992. Notes: (1) The figures correspond to the number of recipients of general social assistance for the month of March of each fiscal year divided by the relevant provincial or territorial population. (2) The Northwest Territories did not establish welfare programs under CAP until 1973-74. 32 Table 2.2 General Assistance Expenditures Per Capita, 1970-1992 (1992 constant dollars).1 Province 1970 1975 1980 1985 1990 1992 Newfoundland 251.62 237.45 196.10 197.80 220.76 303.92 P.E.I. 92.94 139.95 196.77 221.70 225.19 297.87 Nova Scotia 93.64 138.87 164.17 209.64 258.30 311.10 New Brunswick 84.00 236.67 288.52 354.50 340.78 344.35 Quebec 118.23 194.62 258.91 409.94 327.37 396.38 Ontario 88.62 142.69 142.09 201.05 251.39 2 Manitoba 116.23 130.88 115.27 184.40 199.57 251.85 Saskatchewan 89.67 132.34 164.17 224.44 200.73 203.67 Alberta 116.07 141.63 138.38 232.22 286.61 2 B.C. 113.24 234.97 224.53 386.72 283.36 2 Yukon 67.04 54.55 61.22 109.76 107.14 217.82 N.W.T. — 428.24 224.84 225.45 367.00 367.41 Canada 106.62 171.20 188.33 284.03 268.40 2 Source: Canada Assistance Plan, Annual Report 1970-1992 and Statistics Canada, The Labour Force, 1970-1992. Notes:(l) The amounts correspond to two times the Federal Governments payments to the provinces and territories for general assistance under CAP for the respective fiscal years divided by the provincial or territorial population. (2) The figures are not available due to the Federal Government's ceiling on total payments under the CAP. Table 2.3 - Percent of Total Person-Months on Welfare in British Columbia by Year 1980-92 1980-82 1983-84 1985-86 1987-88 1989-90 1991-92 Type of Benefit Employable 54.3 38.4 54.2 56.1 52.1 58.5 63.8 Unemployable 45.7 61.5 45.8 43.9 48.0 41.6 36.2 Unable to Work 42.6 57.4 43.3 41.3 44.9 38.0 33.3 Adult Care 2.9 3.6 2.3 2.4 2.9 3.4 2.8 Medical 0.1 0.1 0.1 0.1 0.1 0.1 0.1 Transient 0.1 0.5 0.2 0.0 0.0 0.0 0.0 Type of Household Couples 13.1 14.2 15.8 14.3 12.7 11.2 10.8 with children 8.1 7.9 10.2 9.2 8.0 6.7 6.7 no children 5.0 6.3 5.6 5.0 4.7 4.5 4.1 Single Women 46.3 50.9 43.2 43.5, 46.7 48.7 45.5 with children 24.1 27.4 21.7 22.3 24.9 25.7 23.4 no children 22.1 23.5 21.5 21.2 21.8 23.0 22.1 Single Men 40.5 34.0 41.1 42.3 40.6 40.1 43.7 with children 1.3 1.0 1.3 1.4 1.4 1.3 1.5 no children 39.2 33.1 39.8 40.9 39.2 38.8 42.2 Program Type Basic IA 78.5 72.1 79.3 81.0 79.7 77.3 80.5 Other than Basic IA 21.5 27.9 20.7 19.0 20.3 22.8 19.5 CIHR 1.6 2.2 1.4 1.1 1.4 1.8 1.9 Age 60-64 3.4 5.2 3.4 3.1 2.9 3.4 2.8 GFH 11.8 14.0 11.3 10.6 11.7 12.8 11.1 GFS 1.7 2.8 2.2 1.6 1.4 1.3 0.9 OAS 0.1 0.2 0.1 0.1 0.1 0.1 0.1 Age of Recipient Less than 21 13.6 14.4 15.9 13.7 12.2 12.3 12.9 Between 21 and 26 14.5 14.1 16.3 15.8 14.3 12.8 13.5 Between 26 and 36 30.8 27.4 29.7 30.8 31.8 31.9 32.6 Over 36 41.2 44.1 38.1 39.7 41.8 43.1 41.1 34 Table 2.4 - Welfare Survivor Function Estimates Months 1 3 6 12 24 48 72 Employable Single Men no children 0.720 0.396 0.234 0.116 0.046 0.014 0.006 children 0.767 0.479 0.311 0.182 0.105 0.041 0.021 Single Women no children 0.709 0.378 0.223 0.114 0.050 0.018 0.008 children 0.825 0.612 0.466 0.334 0.220 0.121 0.069 Couples no children 0.669 0.350 0.203 0.107 0.043 0.014 0.007 children 0.701 0.400 0.251 0.131 0.062 0.023 0.010 Unemployable Single Men no children 0.757 0.542 0.409 0.281 0.175 0.092 0.057 children 0.816 0.616 0.426 0.286 0.173 0.087 0.046 Single Women no children 0.768 0.554 0.427 0.303 0.196 0.113 0.080 children 0.840 0.633 0.471 0.314 0.181 0.080 0.042 Couples no children 0.751 0.508 0.391 0.267 0.189 0.113 0.077 children 0.735 0.472 0.330 0.198 0.110 0.049 0.030 35 Table 2.5 - Welfare Survivor Functions - Differentiating By Recession Start Date Months 1 3 6 12 24 48 72 Employable Single Men no recession no children 0.732 0.416 0.250 0.124 0.049 0.014 0.006 children 0.770 0.500 0.324 0.189 0.106 0.040 0.021 recession no children 0.678 0.330 0.180 0.087 0.037 0.020 0.012 children 0.754 0.405 0.265 0.161 0.105 0.050 0.020 Single Women no recession no children 0.719 0.395 0.237 0.120 0.051 0.018 0.008 children 0.829 0.623 0.479 0.343 0.225 0.121 0.066 recession no children 0.672 0.315 0.173 0.090 0.048 0.026 0.014 children 0.811 0.577 0.426 0.305 0.207 0.136 0.102 Couples no recession no children 0.678 0.370 0.221 0.118 0.048 0.015 0.007 children 0.712 0.417 0.263 0.136 0.062 0.023 0.009 recession no children 0.636 0.284 0.144 0.072 0.029 0.015 0.006 children 0.659 0.333 0.201 0.108 0.062 0.030 0.013 Unemployable Single Men no recession no children 0.772 0.566 0.431 0.295 0.181 0.095 0.058 children 0.814 0.606 0.422 0.270 0.163 0.083 0.042 recession no children 0.713 0.474 0.347 0.242 0.157 0.088 0.058 children 0.827 0.653 0.440 0.347 0.213 0.071 0.071 Single Women no recession no children 0.774 0.570 0.446 0.325 0.205 0.117 0.084 children 0.842 0.641 0.475 0.316 0.179 0.078 0.042 recession no children 0.752 0.508 0.369 0.242 0.171 0.107 0.067 children 0.835 0.604 0.458 0.307 0.187 0.092 0.042 Couple no recession no children 0.754 0.517 0.398 0.283 0.200 0.120 0.084 children 0.747 0.475 0.332 0.204 0.114 0.050 0.031 recession no children 0.742 0.485 0.374 0.227 0.163 0.077 0.039 children 0.698 0.463 0.322 0.181 0.101 0.046 0.029 Table 2.6 - Survivor Functions for Off-Welfare Spells Months 1 3 6 12 24 48 72 Single Men no children 0.855 0.714 0.598 0.450 0.349 0.277 0.248 children 0.841 0.699 0.579 0.441 0.343 0.268 0.242 Single Women no children 0.875 0.763 0.665 0.544 0.448 0.371 0.338 children 0.838 0.716 0.613 0.492 0.397 0.319 0.279 Couples no children 0.865 0.760 0.667 0.551 0.462 0.395 0.363 children 0.853 0.735 0.632 0.496 0.396 0.318 0.283 37 Table 2.7 - Survivor Functions for Repeat Welfare Spells Child Spell Status Number MONTHS 1 3 6 12 24 48 72 Single Men with first 0.75 0.44 0.28 0.16 0.10 0.04 0.02 children second 0.75 0.51 0.34 0.22 0.13 0.05 0.03 third 0.78 0.53 0.39 0.21 0.13 0.04 0.02 fourth 0.81 0.51 0.34 0.24 0.11 0.07 0.03 more than 5 0.79 0.52 0.33 0.20 0.12 0.06 0.02 without first 0.67 0.36 0.21 0.11 0.05 0.02 0.01 children second 0.73 0.42 0.26 0.14 0.06 0.02 0.01 third 0.75 0.43 0.27 0.14 0.06 0.02 0.01 fourth 0.76 0.45 0.28 0.15 0.07 0.03 0.01 more than 5 0.77 0.45 0.29 0.15 0.07 0.03 0.01 Single Worn en with first 0.83 0.62 0.47 0.33 0.21 0.11 0.06 children second 0.84 0.65 0.50 0.35 0.22 0.11 0.06 third 0.83 0.63 0.49 0.35 0.23 0.12 0.07 fourth 0.83 0.61 0.45 0.30 0.18 0.10 0.06 more than 5 0.81 0.58 0.42 0.28 0.17 0.09 0.05 without first 0.69 0.38 0.23 0.13 0.06 0.03 0.02 children second 0.73 0.42 0.27 0.16 0.09 0.04 0.03 third 0.74 0.44 0.29 0.16 0.09 0.04 0.03 fourth 0.74 0.44 0.29 0.18 0.09 0.05 0.03 more than 5 0.76 0.45 0.30 0.17 0.09 0.04 0.03 Couples with first 0.69 0.39 0.25 0.14 0.07 0.03 0.01 children second 0.71 0.41 0.26 0.14 0.07 0.02 0.01 third 0.70 0.42 0.27 0.15 0.07 0.03 0.01 fourth 0.73 0.42 0.27 0.14 0.07 0.03 0.01 more than 5 0.72 0.40 0.24 0.12 0.05 0.02 0.01 without first 0.63 0.32 0.19 0.10 0.04 0.02 0.01 children second 0.68 0.35 0.21 0.12 0.06 0.03 0.02 third 0.70 0.38 0.24 0.14 0.07 0.03 0.01 fourth 0.72 0.43 0.26 0.15 0.08 0.04 0.02 more than 5 0.74 0.46 0.28 0.16 0.07 0.03 0.02 38 Table 2.8 - Survivor Functions for Repeat Welfare Experiences By Length of Previous Off-Welfare Spell Without Children Length of Previous Off-Welfare Spell MONTHS 1 3 6 12 24 48 72 Single Men Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous > = 24 months 0.65 0.34 0.20 0.12 0.06 0.03 0.02 0.77 0.49 0.32 0.17 0.08 0.03 0.01 0.74 0.41 0.25 0.13 0.06 0.02 0.01 0.72 0.39 0.23 0.12 0.05 0.02 0.01 0.72 0.40 0.24 0.12 0.05 0.01 0.00 Single Women Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous >= 24 months 0.68 0.37 0.24 0.14 0.08 0.05 0.04 0.76 0.50 0.34 0.20 0.11 0.05 0.03 0.73 0.43 0.28 0.16 0.09 0.04 0.02 0.72 0.38 0.24 0.13 0.07 0.04 0.02 0.72 0.39 0.24 0.12 0.05 0.02 0.01 Couples Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous >= 24 months 0.63 0.32 0.19 0.11 0.06 0.03 0.02 0.73 0.46 0.29 0.17 0.09 0.04 0.02 0.70 0.38 0.23 0.13 0.06 0.03 0.02 0.66 0.34 0.20 0.11 0.05 0.02 0.01 0.67 0.34 0.21 0.10 0.04 0.01 0.01 With Children Single Men Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous >= 24 months 0.78 0.49 0.33 0.18 0.16 0.09 0.07 0.81 0.60 0.41 0.22 0.13 0.07 0.04 0.77 0.49 0.34 0.22 0.11 0.05 0.02 0.77 0.46 0.30 0.19 0.11 0.04 0.01 0.74 0.44 0.27 0.16 0.10 0.02 0.01 Single Women Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous >= 24 months 0.85 0.66 0.53 0.41 0.30 0.19 0.13 0.83 0.64 0.48 0.33 0.20 0.10 0.06 0.82 0.61 0.46 0.31 0.20 0.11 0.07 0.83 0.60 0.45 0.31 0.19 0.10 0.06 0.83 0.60 0.44 0.30 0.17 0.07 0.03 Couples Single Welfare Spell Multiple Welfare Spells in Sample previous < 3 months 3 <= previous < 7 months 7 <= previous < 24 months previous >= 24 months 0.71 0.42 0.29 0.18 0.10 0.05 0.03 0.75 0.47 0.30 0.16 0.07 0.03 0.01 0.70 0.41 0.24 0.11 0.05 0.02 0.01 0.67 0.36 0.22 0.12 0.06 0.02 0.01 0.69 0.39 0.24 0.12 0.05 0.01 0.00 (S ( QOOL U l ) U 9 d Q S 9 S D Q 40 o LO o LO o LO o LO o C N CNJ •<— o o 6 6 6 6 d 6 d d d 41 o LO o LO o 1 0 o LO o CM CM o o d d d ' d d d d d d 9 } D ^ pJDZDf-] 42 CD C L h -o C\J CD CO 0) Cl L L - Z5 CO i > CD X , =5 • O o X L U CD o s— CD C L L d c c <D (D i_ i_ XI • — x: x: O x: O O + o z T — CM CXI 1^ CD CD O CD OO CN CD o CN 00 CN H CD o o LO o LO o LO o LO o CN CN <— o o 6 6 6 6 d d d d d 9 } D ^ p j D Z D H 43 o LO o LO O LO O LO O ro ro CM CXI •<— ^ O O 6 6 CD CD CD CD CD CD O 9 } D ^ p J D Z D H 44 CD Q _ h— j Q C CD E o CD C D c\i _ CD W ~ ° Z CD O Z5 cn o o X L U CD o s— CD C L E CD C c c 0) -^X ! X I • — _ c SZ O O o + o z CN CN CD CD O CD OO CN CD r o O r o CN 00 CN H CD o o LO o LO o LO o LO o r o r o CN CN •<— —^ o o d d d d d d d d d 8 } D y p J D Z D H 8 } D ^ p J D Z D H 46 i 1 1 1 1 r i 1 1 1 1 r r CXI 1^ CD CD O CD LO ' 00 CXJ CD O O ro CXI 00 CXI H CD LO 6 J i i L_ O o _i i I i i i i i o LO O o o 6 o 9 } D y p j D Z D H 47 Chapter 3. The duration of income assistance spells in British Columbia: tests for state dependence 3.1. Introduction The public discussions concerning welfare reforms have indicated that there is a widespread perception of "welfare dependency". At the most basic level, there is a common belief that individuals and families who use welfare remain on the program for long periods of time. In addition, there is a perceived notion that receiving welfare in itself contributes to greater reliance on the program in the future. That is, there is a perception of state dependence in welfare participation. However, the empirical evidence brought to bear in the media discussions of welfare use has been anecdotal at best. The research presented in this chapter utilises the BC social assistance administrative data to analyse the duration of welfare spells and test for the presence of different forms of welfare dependence. The objective of the research is to answer two questions: (1) What is the impact of demographic variables and labour market conditions on the duration of IA spells, and (2) Is there evidence of state dependence in IA receipt? In addressing the latter question, the empirical analysis tests for the presence of three specific forms of state dependence: (i) duration dependence, whereby the probability of leaving welfare depends on the elapsed time on welfare in the current spell; (ii) lagged duration dependence, whereby the probability of leaving welfare depends on the length of previous welfare spells; and (iii) occurrence dependence, whereby the probability of exiting welfare depends on the number of past welfare spells. These different forms of state dependence imply very different behavioural impacts of the program. For 48 example, if negative duration dependence is present then the welfare system acts as a trap, with the likelihood of escaping declining with time on the program. The presence of lagged duration and occurrence dependence imply that participating on the program "scars" an individual's labour market career, the effects of which escalate with greater interaction with the system over time. Consequently, the different forms of state dependence suggest different policy responses. The answers to these questions provide new information on the dynamics of IA receipt in BC and can therefore contribute to the design and evaluation of program reforms. The issue of state dependence is closely linked to that of unobserved individual characteristics or "unobserved heterogeneity" (Heckman and Borjas, 1980). There are variables that potentially influence the length of welfare spells that cannot be measured or, at least, are not observed by the analyst. It is possible that evidence of state dependence may simply reflect these unobserved individual characteristics rather than true state dependence. Therefore, in testing for the presence of state dependence, it is important to control for the potential effects of unobserved heterogeneity. To date, there are no published studies of the dynamics of welfare participation in Canada due to the lack of suitable data sets. To study the dynamics of welfare participation it is necessary to have a panel data set that both follows individuals for a relatively long period of time and records program participation information. Until very recently there has been no publicly available Canadian longitudinal data set. There is an established literature examining the dynamics of welfare spells in the United States (see Moffit (1992) for a thorough review). O'Neill etal. (1987) use annual data from the National Longitudinal Survey of Young Women to examine AFDC dynamics over the 1986-92 period. Controlling for a large array of covariates, the authors find that women from a lone 49 parent family, with out of wedlock births, who are black, of ill health, with more children and less work experience have systematically lower exit rates from welfare. Blank (1989) examined the dynamics of AFDC participation by lone mothers over the 1971-1976 period using monthly data from the control group of the Denver and Seattle Income Maintenance Experiment. Restricting the sample to the first observed spell, Blank estimated several parametric proportional hazard models and tested for the presence of duration dependence. Blank finds that inferences regarding duration dependence are sensitive to the functional form assumption of the duration model used, and that the more flexible models indicate negative duration dependence for a subset of the population. Although the methods of the studies of welfare dynamics in the U.S. are instructive, it is not clear how applicable the findings are given the substantial differences between the Canadian and U.S. welfare programs and labour markets (see Blank and Hanratty, 1993). The chapter is organized as follows. In the next section a dynamic model of welfare participation is presented which generates predictions regarding the presence of the three different forms of state dependence and lays out the potential problems introduced by unobserved heterogeneity. Section 3 outlines the methods used in the analysis. The data set is described in more detail in section 4 and the results of the empirical analysis are presented in Section 5. Conclusions and policy implications of the findings are discussed in the final section. 3.2. Theoretical Model (a) Static Model of Program Participation Cross-sectional studies of the incidence of welfare participation are based on an extension of the static labour supply model. A household simultaneously decides hours of work and 50 whether to receive welfare. Let U(y,-h;z) represent the household's utility defined over income (y) and hours of work (h) conditional on characteristics z. The budget set over which U(.) is maximised, in the presence of the welfare program, is illustrated in Figure 3.1. The budget line is nonlinear and disjoint with convex segments. For non-welfare households, income equals employment earnings (wh) plus nonlabour income (A7). When on welfare and not working, the household receives income equal to the welfare benefit B plus (J-tfJN where r^ is the benefit tax-back rate on unearned income. When on welfare and earning less than the earnings 'disregard' (BC) the budget line is given by y=B+(1-tjJN+wh, and when earning more than the disregard, y=B+(l-tN)N+wh(l-tJ+whJe where te is the tax back rate on earnings beyond the disregard and h0 is that number of hours of work generating earnings equal to the disregard. For hours of employment beyond h2, the break-even level of hours, households are no longer eligible for benefits. The household is assumed to maximise utility subject to the budget constraint shown in Figure 3.1. A household participates in the welfare program if the optimal hours of work, h*, is less than h2. Moffit (1983) expanded this standard labour supply model by allowing for stigma effects, or "psychic costs", of welfare participation. Moffit's model was motivated to account for the fact that many households that appear eligible to receive welfare, and who would be financially better off if they received welfare, decide not to participate on the program. According to this model, the household's utility function if given by U(.) = U(y,-h;z)-C(P) where C(.) represents the cost in utility of participating on welfare (P=l). Moffit (1983) and Blank (1985) adopt this framework and estimate a joint model of the welfare participation and labour supply of single mothers in the United States. Allen (1993), Charette and Meng (1994) and Dooley (1994) use the model to motivate a univariate probit analysis of welfare participation by 51 Canadian lone mothers. The static models of program participation generate predictions on the effects of the budget constraint variables, including the welfare program parameters, on the likelihood of program participation and the household's labour supply. The econometric findings of these studies generally corifirm the predicted effects while accounting for the fact that many eligible households decide not to participate in the program. (b) A Dynamic Model of Program Participation The static model of program participation provides important information on the association between personal characteristics and program parameters on the probability of welfare participation at a point in time. However, these models do not provide any guide to how long individuals or households are likely to remain on welfare. For example, given that a household participates on welfare, what characteristics are associated with relatively long or short expected durations? Likewise, what are the effects of program parameters or labour market conditions on the expected duration of welfare spells? It is important to gauge these effects in order to better understand "welfare dependency" and for the effective targeting of programs aimed at reducing caseloads and program expenditures over time. In this section, the static model of welfare participation with stigma effects is extended to a dynamic setting. For simplicity, assume the household has to decide between two discrete alternatives; working fulltime or receiving welfare. As in the static participation model, the choice of states is based on the comparison of household utility under each alternative. The household's ex-post indirect utility from working fulltime is given by (1) V° = V°(zpw,d) where z, are household characteristics at time t, w, is the individual's wage rate at time t and d, 52 represent recurrent (per period) costs of fulltime employment (such as child care). Assume that the offered wage at time t, w„ has a random element, e0 and a systematic component w(.) with wt = M>(T)+e, where T is the length of current employment (off-welfare) spell. Furthermore, Wjo (the partial derivative of w(.) with respect to T)l <> 0 due to the accumulation of experience and human capital while employed and e, is a white noise error term with distribution function given by F(e). It follows from the properties of the indirect utility function that Vw°>0, Ve°>0, VjozO, Vf<0. An individual's ex-post indirect utility when participating in the welfare program is given by (2) V? = V(z,B,Q where Bt are potential welfare benefits available at time t, and C, are the nonmonetary stigma costs of welfare participation. Let stigma costs Ct=C(T"), where V is the length of the current welfare spell. It is assumed that stigma decreases the longer the individual has been in receipt of welfare; that is, Cjw<0. In can be shown that VBW>0, Vcw<0, VTw=Vcw.Crw>0. The individuals' decision process is assumed to take the following form. New values of the offered wage, w„ arrive at random intervals. When this information arrives the individual chooses the preferred alternative. Following Oisen et.al. (1986) and Fortin et.al. (1994), it is assumed that the arrival of new information follows a Poisson process with constant rate v. Since the error term (shocks to the offered wage) are temporally independent, the instantaneous rate of exit from welfare at time t, conditional on the welfare spell having lasted until T", is given 1 Throughout this section, a subscripted variable is used to denote the partial derivative of a function with respect to that variable. by (3) X(xvT) = v[Prob(V°(x,e,)2V(x,B,r))] where xt = {z„d}. Equation (3) represents the hazard function for the model. Define implicitly as the critical value of et such that V°(.)=VW(.), and hence ^ =^(xpBt,T") . Therefore (3) may be rewritten as (4) X(xp D = v.Prob[et^(xpBp T)] - v.F[e°(xpBpT)]. It is straightforward to show that (5) Xc>0, XB<0, Xd<0, X>0, Xr<0. The higher the offered wage the greater is the relative attractiveness of employment and hence the higher is the welfare exit probability. Higher potential benefit levels and recurrent costs of employment have a negative effect on the exit rate from welfare. Stigma costs have a negative effect on the relative utility of welfare participation and hence a positive effect on the welfare exit rate. However, as the amount of time on the program V increases, the stigma costs decrease leading to a decrease in the hazard rate. Therefore, the dynamic model of welfare participation with stigma costs predicts negative duration dependence.2 By positing that human capital accumulates while employed, the model also generates predictions regarding off-welfare (employment) spell durations. The offered wage systematically increases with the time off welfare which implies that the re-entry rate into welfare (or employment hazard rate) is decreasing in the duration of the off-welfare spell. That is, the off-welfare hazard rate also exhibits negative duration dependence. 54 The model can be readily extended in several directions. Firstly, stigma costs may be influenced by an individual's interaction with the welfare system over a period of time as well as the length of the current welfare spell. That is, stigma may decline with the time spent on welfare in previous spells, lagged duration, or with the total number of spells. Let 5 index the spells experienced by an individual, ordered by starting date, and / denote lagged duration. The stigma costs are then given by C t = C ( T , l , s ) which implies that (6) A,*0, As*0. These extensions to the model generate the additional predictions of negative lagged duration dependence and negative occurrence dependence. The predictions of the model correspond to the three forms of state dependence examined by Ffeckman and Borjas (1980).3 Additionally, labour market conditions are likely to impact on the arrival rate of wage offers. To allow for cyclical and seasonal fluctuations, the arrival rate of wage offers may be specified as v= v(u„vpq) where u, denotes the unemployment rate, vr denotes excess labour demand or job vacancies and qt denotes seasonal effects.4 It is expected that vu^0 and vvkO which implies that Au*0 and AvzO. It is noted that the actual cause of any state dependencies found cannot be identified. The model of dynamic welfare participation generates predictions of state dependence due to stigma costs. Alternatively, the dynamic effects may be due to human capital atrophy while on welfare or because the transactions costs of participating in welfare decline over time as the individual interacts with the program. Similarly, the rules of the welfare system are complex and an individual may learn about the potential benefits and the ehgibility criteria over time through participation in the program. These alternative hypotheses are observationally equivalent to the model based on stigma effects. Therefore the empirical tests for state dependence should be interpreted as including these effects. 4 Introducing seasonal correlation into the arrival rate of wage offers alters the individuals' decision process. When making a decision regarding a wage offer the individual must now weigh the relative utilities of present employment versus welfare participation and the expected waiting time for the arrival of the next wage offer conditional on v, and qt. As a result, the individuals' decision process is no longer temporally separable and leads to a complex dynamic programming problem, as discussed by Olsen etal. (1986). This added complexity is not addressed in the theoretical model nor the empirical implementation. 55 c). The implications of unobserved heterogeneity For simplicity, assume that the population is composed of two types of individuals who differ with respect to the (unobserved) arrival rate of job offers, which may be due to differences in job search intensity or motivation. The types are labelled high (H) and low (L), according to the arrival rate, with i/T> tA It is straightforward to show that AH(T0)>AL(T0), The aggregate hazard rate function is then given by (7) A(T) = p"(T). A"(T) + (l-p"(T)). AL(T) The observed exit rate at duration T is the weighted average of the exit rates for the two groups, with the weights equal to the proportion of the welfare population at duration T accounted for by each type. The proportion of the welfare population with spells of at least T months in duration is a function of the survival probabilities, source population proportions and welfare entry rates of the two types. For simplicity, assume that the source population consists of an equal proportion of H and L types and that both types have a common, exogenous entry rate. The survival functions for the two types are defined as (8) S(T) = ni'Jl-W) fori=H,L. Since l*AH(t)>AL(t)zO VI then SH(t)<SL(t) Vt. The proportion of the welfare population with spells of duration T who are of type H is then given by (9) fCT) = SH(T) / (SH(T)+SL(T)) Differentiating (9) with respect to T yields (10) dfciyffi = S^/(S"+SL) - S^ffi+S))2. Since Sj is negative and Sf is positive by definition, the first term on the right-hand side in 56 equation (10) is negative and the second term positive and hence d^cryar < o. Since individuals of type H have a higher job offer arrival rate, and hence welfare exit rate, the welfare population is composed of a smaller fraction of type ^ individuals at longer durations. The implications of unobserved characteristics for tests of duration dependence are revealed by totally differentiating (7) with respect to duration T: (11) dA(T)/cT = A". dp"/ai + f. dA"/cT - AL. dp"/dl + p". dAH/ffT. To draw out the potential problems introduced by unobserved heterogeneity assume that there is no underlying duration dependence in the exit rate functions for the 2 types; that is, dM/dT^O for j=H,L. Equation (11) then simplifies to (12) dA(T)/cT = (AH- AL).6p"/6T < 0. Hence, even in the absence of true duration dependence, unobserved individual characteristics lead to the appearance of negative duration dependence in the observed hazard rate. As equation (12) shows, negative duration dependence may be observed solely because individuals most likely to exit due to the unmeasured characteristic depart early, leaving behind a population less and less likely to exit. Similarly, it can be shown that unobserved characteristics may account for negative lagged duration and occurrence dependence in the aggregate hazard rate in the absence of true state dependencies. Therefore, in testing for the presence of the different forms of state dependence it is important to control in the estimation for the potential impact of unobserved heterogeneity. 3.3. Methods 57 The empirical analysis in this chapter is conducted in a hazard function framework. The analysis proceeds by first estimating the empirical hazard rate function. The empirical hazard rate function reveals information on the overall shape of the IA hazard without imposing a parametric function on the underlying distribution. A limitation of the empirical hazard rate estimator is that it treats the population as homogeneous. Spell lengths potentially differ according to the characteristics of the recipient. To control for such covariates, and to gauge the effects of the covariates on the hazard rate, a duration model is used. The most common specification is the proportional hazard model in which hfl) = hjl).ap{z/fl} where h/T) is the hazard for person i, hJT) is the baseline hazard common to all individuals, zt is vector of observable characteristics (which may vary with T) and B is a parameter vector to be estimated. For different values of z/B, the hazard function for individualis shifted proportionally up or down relative to the baseline. The estimation approach implemented in this paper is an extension of Prentice and Gloeckler (1978) and is detailed in Meyer (1986; 1990) and Lancaster (1990:172-208). The baseline hazard is estimated nonparametrically as a piece-wise constant function. The time axis is divided into a finite number of intervals and a separate baseline hazard parameter is estimated for each interval. This approach provides a very flexible method for estimating the baseline hazard function and avoids the imposition of a parametric functional form on the baseline. This is an important advantage in the present context for it has been shown that misspecifying the baseline hazard is a major source of error in drawing inferences concerning the presence of duration dependence (Manton, Stallard and Vaupel, 1986; Blank, 1989). In addition, Heckman and 58 Singer (1986) have shown that misspecification of the baseline leads to inconsistent estimates of the parameter vector fi which is the basis for the tests of lagged duration and occurrence dependence. Further advantages of the piece-wise constant proportional hazard model is that it allows for right-censored spells, tied spell durations and it can be readily extended to incorporate the effects of unobserved heterogeneity. As Meyer (1986) has shown, unobserved heterogeneity may be introduced into the model either parametrically or nonparametrically. An alternative approach for estimating the proportional hazard model is the Cox (1972, 1975) partial likelihood approach. This method provides estimates of the ft vector without specifying the shape of the baseline hazard function. The baseline can be estimated in a second step using a variant of the Kaplan-Meier estimator (Kalbfleisch and Prentice, 1980:84-87). For this model, the likelihood function is interpreted as the partial likelihood of the ranks of ordered durations. However, the major limitation of the partial likelihood approach is that it cannot be readily extended to allow for unobserved heterogeneity. In particular, controlling for unobserved heterogeneity requires the evaluation of multiple integrals, up to the order of the dimension of the sample, which is computationally infeasible (Meyer, 1986; Han and Hausman, 1990). To derive the likelihood function for the piece-wise constant proportional hazard model note that the probability that a spell lasts at least until time t+1, conditional on it having lasted time t, is given by (13) PfT^t+l/T^tJ = expf-exp(zi(t)'fi+ y(t))J where (u) r(t)=iogf/';%(u)duj. The log likelihood function for a sample of N welfare spells is then 59 (15) L(Y,P) = ZNi=1 { 6,log[l-exp(-exp(Y(k)+z,(k)'P))] -m?exp[Y(t)+z,<typ]} where kt is the observed length of the /'th welfare spell, d, equals one if the spell terminates before being censored and is zero if the spell is censored. In maximizing the log likelihood the y(t) are treated as parameters to be estimated. The average baseline hazard rate over the interval [tj,tJ+1] is defined as exp(y(t))/(t^rt). In implementing this model it is necessary to censor any ongoing observations at some duration T. For the empirical analysis, all observations lasting 73 months or more are censored at 73 and treated as right censored. Approximately five percent of all welfare spells were ongoing at that point. The 72 month time period is then divided into 18 intervals, with separate baseline parameters estimated for each interval.5 The proportional hazard model can be extended to allow for unobserved individual characteristics. Assuming that the unobserved heterogeneity takes a multiplicative form, the hazard rate is given by h,(T) = excryexpizip} where 6i is a non-negative random variable assumed to be independent of zt. Maximum likelihood estimates of the parameter vector and baseline hazard are then obtained by conditioning the likelihood function on the 6t and then integrating over the distribution of 6. This approach requires specifying a distribution function for dt. A popular and convenient distribution for 0is the gamma, which gives a closed form expression for the likelihood function. Assuming 6 is distributed as according to the unit gamma with variance o2, a parameter to be estimated, the log likelihood is then given by 5 Each month from 1 to 6 is treated as a separate interval. The additional intervals correspond to months 7-8, 9-10,11-12,13-16,17-20,21-24,25-30,31-36,37-42,43-51,52-60 8^61-72. 60 (16) L(y,p, a2) = 2 7 ? = , log {[1+ o2 Ek;=fexp(Y(t)^zi(typ)r"-i -611+ a2 Z^expfrW+z/tywr"-'}-In section 6 below estimates from the maximisation of the log likelihoods of equations (15) and (16) are presented and compared. The piece-wise constant proportional hazard model enables testing for the presence of the three forms of state dependence denned above. Duration dependence is present if the hazard rate off IA varies with the elapsed spell length. Duration dependence therefore relates to the shape of the baseline hazard rate function. If the baseline hazard is downward sloping then negative duration dependence is present, with the probability of exiting from IA declining as the IA spell length increases. Lagged duration dependence implies that the length of previous IA spells influences the exit rate for subsequent spells. This hypothesis is tested by including lagged duration in the vector of covariates and testing whether the estimated coefficient is significantly different from zero. Lastly, occurrence dependence is present if the hazard rate off welfare depends on the number, as distinct from the duration, of previous welfare spells. The hypothesis of occurrence dependence is tested by including a set of dummy variables indicating the number of the spell, in the sequence of spells observed for an individual, in the covariate vector and then testing the joint significance of the set of coefficients. 3.4. The D a t a The analysis in this chapter is based on the B.C. social assistance administrative data, which was examined in Chapter 2. The raw data are a ten percent random sample of all 61 individuals with a history of social assistance receipt in BC during the period of January 1980 to December 1992. The sample consists of 87,288 person-specific records. Each record contains the individual's (or principal claimant's) birth date, sex, and variables indicating under which BC social assistance program, if any, the individual received benefits for each month of the thirteen year time period. Additionally, there are variables indicating the individual's family type, number of dependent children and employability status for the corresponding month the person was in receipt of social assistance. A spell of IA is defined as a sequence of consecutive months of Basic IA receipt. An exit is defined as one month not in receipt of IA benefits.6 For spells beginning after January 1980 it is possible to determine the precise length of the spell unless the spell is still in progress in December 1992. Such spells in progress at the end of the data period are right censored and the analysis takes this censoring into account. The spells of IA form the basic unit of the empirical analysis. For each spell there is information on duration and whether it is right-censored, as well as the recipient's sex and age, family type, employability status and number of dependent children at the commencement of the spell.7 To control for the impact of welfare benefits on spell durations, the maximum potential welfare payment available to the household, given their characteristics, was constructed using the program rules.8 Table 3.1 presents the BC maximum monthly welfare benefits (in Dec. 1992 It is possible that administrative coding errors may lead to the artificial appearance of short, repeat spells in the data. To check for this form of measurement error, exits were redefined as two consecutive months not in receipt of IA. The principal findings discussed below are not sensitive to this alternative definition of exits. Unfortunately the data do not include information on the age of children in families that receive welfare. 8 The nominal benefits were deflated by the monthly consumer price index for BC, with December 1992 as the base period. 62 constant dollars) payable for several household types for selected months between 1980 and 1992. The administrative data do not contain a measure of the wage that an individual may command in the labour market. As a proxy for the wage, the monthly real earnings that an individual would receive if employed full-time at the prevailing minimum wage is calculated. The potential benefits payable to a household are a function of family size and as a consequence potential benefits and number of dependent children are highly collinear.9 In order to separately identify the effects of the number of children and potential benefits on welfare exit rates, potential benefits are divided by household size. Real monthly earnings at the minimum wage are also divided by family size to maintain the comparability between welfare and employment income for the household. To control for business cycle effects the spells are matched with the provincial unemployment rate and help wanted index (Hwi) prevailing at the date of spell commencement. The prime-age male unemployment rate is used so as to ensure the measure is exogenous to the welfare program. The help wanted index is compiled monthly by Statistics Canada and is a measure of the job vacancies that are advertised in the leading newspapers in the major cities of BC. In comparing the path of the unemployment rate and help wanted index over the data period, the major difference is that the unemployment rate lags the help wanted index as a measure of general labour market conditions. The help wanted index moved contemporaneously with aggregate output, with dramatic declines in 1981 and 1990 coincident with the onset of the two recessions that occurred during the sample period. The unemployment rate tended to lag the help wanted index by 9 to 12 months over the data period. Furthermore, to control for seasonal variation in labour market conditions, dummy variables indicating which quarter of the calender For the full sample of spells, the correlation coefficient is 0.96. 63 year that the spell began were included in the estimation. The test of duration dependence is straightforward in a hazard function framework since the exit rate is estimated conditional on spell duration. As discussed above, the test of duration dependence is based on the shape of the baseline hazard function. Wald tests of the hypothesis that the hazard rate does not change between specific months of a spell, against the alternative that the hazard rate decreases, are implemented. Tests of lagged duration dependence and occurrence dependence are identified by information from the history of an individual's interaction with the welfare system. Specifically, lagged duration is defined as the length of the individual's previous welfare spell.10 To test for occurrence dependence, dummy variables indicating whether a given spell corresponds to the first, second or up to the fifth observed spell for an individual were constructed. It is possible that the information on a person's welfare history may, in part, reflect general calender time trends. For example, structural changes in the labour market, such as the declining position of low skilled and young workers, may have led to welfare spells in general becoming longer over the course of the data period.11 Since the welfare history measures will be influenced by conditions prevailing earlier in the sample period, the estimated coefficients on the spell number dummy variables and lagged duration may then represent a convolution of such calender time trends and the true impact of welfare history on program exit rates. As a first 1 0 An alternative way to define lagged duration is to count total time on the program in a given time window, such as 2 or 3 years prior to the start of a spell. Such measures of lagged duration do not lead to any qualitative differences in the results regarding the presence of state dependence. 1 1 In the late 1980's a backlog developed in the processing of UI claims. As a consequence of the delays, a significant number of UI eligible individuals experienced financial hardships and received welfareuntil the claims were processed, which may have taken up to several months. These recipients were on welfare due to the administrative practices of the UI program. As reported by Bruce atal. (1993) and Barrett et.al. (1995), the welfare spells by "UI pending" recipients were relatively short compared to the average spell duration. Unfortunately the "UI pending" recipients are not identified in the data; however, the calender time dummy variable should, in part, control for the impact of the rise of the "UI pending" caseload in the latter part of the data period on the welfare exit rate. 64 approach to ensuring that the estimated impacts of an individuals' welfare history on the exit rate is not contaminated by a general calender time trend, a dummy variable indicating whether the spell commenced prior to July 1986, the midpoint of the sample period, was constructed and included in the estimation. The sensitivity of the results regarding the presence of state dependence to the inclusion of the 1980:1-1986:6 dummy variable is examined. Furthermore, it is noted that in the duration models upon which the analysis is based, all the covariates are specified as time invariant, and take their beginning of spell value. A major finding from the analysis in chapter 2 was that the vast majority of spells are relatively short: over 70 percent of all spells end within 6 months. Consequently, the within-spell variation in time-varying covariates is minor, especially in comparison to the across-spell variation in the covariates. For this reason, combined with the computational advantages of fixed covariate duration models,12 all the covariates were treated as fixed within a spell for the bulk of the estimation. However, Heckman and Singer (1986: 56) report that, in the context of a parametric duration model, the treatment of time-varying covariates as fixed can lead to biased estimates and hence erroneous inferences. To examine whether specification error is introduced into the analysis by treating all covariates as fixed, random subsamples were drawn for each family type and duration models were re-estimated, alternately treating the covariates as tuiie-varying and fixed. The estimates from both specifications are compared and discussed in section 5 below. To anticipate the results, it is found that, with the exception of the seasonal dummy variables, the treatment of covariates as fixed rather than time varying leads to only minor differences in the coefficient estimates and has no impact on the inferences regarding the presence of state 12 Incorporating time varying covariates effectively increases the sample size by a factor of 18: there is a separate vector of covariate values for each interval of the baseline hazard. Given the very large number of observations (over 167,000 in total) it was not computationally feasible to estimate the duration models with the full sample and allow for time-varying covariates. 65 dependence. From the original set of individual records a sample of 167,096 separate IA spells was generated. The sample is divided into three groups by family type; Single men and women (without children), Couples with and without children and Lone parent families. Each family-type subsample is analysed separately. The samples analysed consist of 113,345 observations for single men and women, 25,621 for couples and 28,130 for lone parent families. Table 3.2 presents the descriptive statistics for the sample of spells by single men and women, stratified into initial and repeat spells. The main features of the sample are that the average spell durations are approximately 6 and 7 months for first and repeat spells, respectively, the majority of the spells are experienced by men and that, on average, rninimum wage earnings are substantially greater than potential welfare benefits. Additionally, approximately 40 percent of the spells experienced by singles correspond to an individuals first observed spell. This indicates that there is a very high incidence of repeat use of the welfare program, with 60 percent of all spells corresponding to repeat spells. Table 3.3 presents descriptive statistics for the sample of spells experienced by couples. The average spell is 5.2 and 6.4 months for first and repeat spells, respectively, which is marginally shorter than that for singles. Approximately 19 percent of the principal claimants were female and the average number of dependent children is 1.4. For couples, the average potential welfare benefits exceeded the average earnings of one family member employed fulltime at the miriimum wage. Similar to the singles sample, approximately 60 percent of the spells by couples represent repeat spells. Descriptive statistics for the lone parent family sample are presented in Table 3.4. The average duration of welfare spells experienced by lone parents is 12.6 and 12.4 months for first 66 and repeat spells, respectively, which is substantially greater than that found for singles and couples. Over 92 percent of the spells were experienced by lone mothers and the average number of dependent children is 1.6. Over 41 percent of initial spells, and 31 percent of repeat spells, were accounted for by parents classified as unemployable, which is substantially higher than the proportions found for the other family types, which in part reflects the presence of young dependent children in the household. Further, approximately two-thirds of all spells were repeat spells which indicates a higher incidence of recidivism among lone parents than other family types. 3.5. Empirical Results (a) Empirical Hazard Rate Estimates The empirical hazard rate functions for the three samples, with first and repeat spells pooled, are plotted in Figure 3.2. The plot shows that the hazard rate functions are downward sloping for the three spell samples. The hazard function for singles and couples are very similar and lie uniformly above that for the sample of spells by lone parent families. For all three groups the exit rate from welfare is relatively high over the first 9 to 12 months of a spell. For the first month of a spell, the exit rates are 0.25, 0.29 and 0.17 for singles, couples and lone parents, respectively. The corresponding exit rates at 3 months are 0.20, 0.20 and 0.12 respectively, then declining to 0.12, 0.13 and 0.12 by the sixth months of a spell. At 12 months, the hazard rates are lower at 0.07, 0.08 and 0.05 respectively, and then generally decline more gradually over longer durations. The empirical survival functions are plotted in Figure 3.3. The graph shows that 50 percent of the welfare spells by singles, couples and lone parents end within 3, 2 and 5 months, 67 respectively. After 9 months approximately 19 percent of the spells by singles, 18 percent of the spells by couples and 38 percent of the spells by lone parents are ongoing. By 12 months, the proportions are 16, 14 and 32 percent, respectively. This evidence indicates that the majority of welfare recipients remain on the program for a relatively short period. Continuous, long term dependence on welfare does not accurately describe most people's experience on welfare. However, there is a subset of the lone parent population who do have relatively long welfare spells. Almost 20 percent of the spells by lone parents are at least 2 years in duration, 10 percent are at least 4 years in duration and 6 percent are ongoing after 6 years. Given that lone parent families generally receive greater benefits than other family types, these long-term recipients do account for an important fraction of the resources devoted to welfare over a period of time. The shape of the empirical hazard and survival functions indicates negative duration dependence; the longer that an individual is on welfare the lower is the probability that they will exit the program. However, as discussed above, the apparent duration dependence may be a reflection of differences in individual characteristics rather than true state dependence. The results from the duration model estimation, which controls for observable characteristics and unobserved heterogeneity, are presented next. (b) Duration Model estimates i. Single men and women The estimates of the mixed piece-wise constant proportional hazard models, which correct for unobserved heterogeneity, for the sample of spells by single men and women are presented in Table 3.5. The results in column (1), based on initial spells, show that women, older individuals and individuals classified as unemployable have systematically lower welfare exit rates 68 and hence longer expected durations. Potential welfare benefits have the predicted negative effect on the hazard rate, although the coefficient is not significant. The minimum wage has a significant positive impact on the exit rate, indicating that recipients' exit behaviour is responsive to potential labour market earnings. The unemployment rate has a substantial, negative impact on the hazard rate and the help wanted index has a positive impact. The latter estimates indicate that the welfare exit rate is very sensitive to business cycle effects. Specification (1) also includes a set of dummy variables indicating the quarter of the year that the spell began. The estimated coefficients on the dummy variables show that there is a strong seasonal pattern in the welfare exit rate for singles. Holding other covariates constant, the hazard rate is greatest in the first quarter of the calender year and then declines over subsequent quarters. This pattern in the hazard rate corresponds to the seasonal fluctuation in aggregate output and employment in the BC economy. Turning to the sample of repeat spells, the results for the most general specification are presented in column (3). Compared to the results in column (1), potential benefits have a much larger, and significant, negative impact on the exit rate for repeat spells which suggests that the labour supply disincentive effects maybe greater for repeat users of the program. The minimum wage has a smaller, though significant, positive impact on the repeat spell hazard. The unemployment rate has a much smaller impact on the exit rate for repeat spells while the help wanted index has a substantially larger impact, relative to their effects on the initial spell exit rate. The unemployment rate and help wanted index both capture business cycle effects in the labour market. The differential impact of the two measures across first and repeat spells may be due to the interaction between the Unemployment Insurance (UI) program and IA. Following the onset of a recession many people who become unemployed are eligible to receive UI, and after 69 exhausting their UI entitlement and failing to find employment, some move onto welfare. However, individuals who recently had a spell on welfare may not be eligible for UI, due to not having the minimum number of weeks of employment in the qualifying period as a consequence of having been on welfare, and therefore may move onto IA much sooner. Since the unemployment rate is a lagged indicator of fluctuations in labour demand whereas the help wanted index is a contemporaneous measure, this interaction between IA and UI may account for the first spell hazard rate being relatively more sensitive to the unemployment rate and the repeat spell hazard being more sensitive to the help wanted index. Further comparing the results in columns (1) and (3), the coefficient on the 1980-1986 dummy variable is positive and significant for the first spell sample and is negative and significant for the repeat spell sample. That is, initial spells on welfare were in general shorter, while repeat spells were in general longer, during the first half of the data period. Model (3) includes information on the history of recipients' welfare participation. The estimated coefficients on the spell number dummy variables indicate that the greater is the spell number, in the sequence of spells for a given individual, the lower is the exit rate, which is consistent with negative occurrence dependence. Additionally, the coefficient on lagged duration indicates that the longer the length of the previous spell, the lower is the exit rate; in particular, each month on welfare in the previous spell is associated with a 2 percent decline in the exit rate for the current spell. The next step in the estimation was to formally test for the presence of the different forms of state dependence. The test of occurrence dependence is based on the joint significance of the set of spell number dummy variables. Column (4) presents the coefficient estimates for specification (3) with the spell number dummy variables excluded. The likelihood ratio test leads 70 to the rejection of the restriction that the spell number dummy variables are jointly insignificant. The test statistic is 31.48, and is distributed as chi-squared with 3 degrees of freedom under the null hypothesis of no occurrence dependence.13 Therefore, the evidence indicates that there is negative occurrence dependence in the welfare exit rate for single men and women. The results in column (4) show that the estimated impact of lagged duration on the hazard rate is robust to the exclusion to the spell number dummy variables. The likelihood ratio test leads to the rejection of the hypothesis of no lagged duration dependence, with a test statistic of 1690.74.14 Thus, there is also very strong evidence negative lagged duration dependence in the welfare participation of single men and women. The evidence suggests that participation in the welfare program does have a scarring effect on single men and womens' subsequent labour market career. The baseline hazard rate function estimates from specifications (1) and (3) are reported in Table 3.6, and the baseline hazard functions from specifications (1), (2), (3) and (6) are graphed Figure 3.4. The graph indicates the presence of negative duration dependence for both initial and repeat spells, even after controlling for unobserved heterogeneity. For example, the baseline for the model based on repeat spells and allowing for heterogeneity reveals that the exit rate decreased by 10 percent, from 0.30 to 0.27, over the first 3 months of a spell. The exit rate then declined a further 50 percent by the 12th month, and decreased a total of 68 percent by the end of the third year of a spell. To determine the statistical significance of the decline in the hazard over the length of a spell, Wald test statistics for the null hypothesis that the hazard rates are The critical values at the 0.05 and 0.01 level are 7.82 and 11.34, respectively. 1 4 The critical values of the chi-squared distribution with one degree of freedom at the 0.05 and 0.01 level are 6.635 and 3.841, respectively. The estimates for specification (3) with the coefficient on lagged duration restricted to zero are not presented in Table 3.5. 71 equal in 2 specific months, against the alternate that the hazard rate declines over time, were calculated. The change in the hazard rate between the given months of a spell, and the associated Wald statistics are reported in Table 3.8. The test statistics show that the decline in the baseline hazard rate for initial and repeat spells are highly statistically significant, especially over the first 36 months of a spell. Therefore, the Wald tests confirm the presence of negative duration dependence in the welfare participation of single men and women, which implies that the welfare program acts as a "trap" in the sense that the longer a person remains on the program, the less likely it is they will exit the program. Models (2) and (6) replicate specifications (1) and (3), respectively, when no allowance is made for unobserved heterogeneity. In general, controlling for heterogeneity leads to an increase in the absolute value, and hence the statistical significance, of the coefficient estimates; which is consistent with the findings of Meyer (1990) and Han and Hausman (1990). Furthermore, likelihood ratio tests support the presence of unobserved heterogeneity in both the initial and repeat spell samples. The estimates in columns (7) and (8) provide the basis for deterrmning whether treating all covariates as time invariant introduced specification error into the analysis. A 5 percent random sample of the set of repeat spells was drawn and 8 covariates (the benefit level, minimum wage earnings, the unemployment rate, the help wanted index, the 1980-1986 dummy variable and the three seasonal dummy variables) were allowed to vary within a spell.15 The specification for model (3) was re-estimated with the subsample; firstly, allowing for time varying covariates (model (7)) and then treating all covariates as fixed (model (8)). In general, allowing for the time For the time varying covariates, a separate value of the regressor was observed for each time interval of the baseline hazard function. For time intervals which spanned more than one month, the value of the regressor at the start of the interval was used. 72 varying covariates led to an improvement in the likelihood value and a uniform decline in the standard errors of the estimates. Furthermore, the estimated variance of the heterogeneity distribution was lower for the model with time varying covariates. Several conclusions may be drawn from the comparison of the coefficient estimates in columns (7) and (8). Firstly, the coefficient estimates in model (7), apart from the coefficients on the seasonal dummies, are within one standard error of the estimates in model (8) which treats all covariates as time-invariant. In addition, the coefficient estimates for the fixed covariates are numerically very similar in models (7) and (8). That is, the estimated impact of demographic characteristics and welfare history, as measured by lagged duration and spell number, are invariant to the treatment of the time-varying covariates. Therefore, it can be concluded the inferences regarding occurrence and lagged duration dependence are not sensitive to the treatment of time varying regressors. However, several coefficient estimates do vary substantially across models (7) and (8). In particular, the estimates for the seasonal dummy variables are much larger, and more statistically significant, in model (7). An important source of variation between spell duration and the seasonal indicator variables is lost when the set of seasonal dummies are assumed fixed at their start of spell value, which leads to biased estimates of the seasonal dummy variables.16 Therefore, the estimates for the seasonal dummy variables should not be considered reliable in the duration models where they are held fixed within a spell. To examine the implications of allowing for time varying covariates on the estimates of the baseline hazard function, the baseline hazards for models (7) and (8) are reported in Table 1 6 Additionally, the coefficients on the unemployment rate and the 1980-1986 dummy variable are somewhat larger (though not significantly different) in model (7) relative to model (8). 73 3.7. It is apparent that allowing for time varying covariates led to a uniform decline in the baseline hazards by approximately 3 percentage points. However, the decline in the baseline hazards over the length of a spell is very similar for the two models. Table 3.8 reports the differences in the hazard rate between various months, and the associated Wald test statistic for the hypothesis of no duration dependence, for the baseline hazard functions of models (7) and (8). Both baseline hazards functions indicate very similar rates of decline over time, and confirm the presence of duration dependence, especially over the first 24 months of a spell. Therefore, the baseline hazard estimates from the duration model with fixed covariates provided a reliable basis for tests of duration dependence. Overall, it can be concluded that, for the sample of welfare spells by single men and women, the treatment of all covariates as fixed did not lead to any serious specification error for the main variables of interest. In particular, the estimates relating to the different forms of state dependence were unaffected by the treatment of the time varying covariates. Likewise, the estimated effects of the demographic variables, program benefits, minimum wage and help wanted index were largely invariant to the inclusion of the time varying regressors.17 However, the coefficients on the calender time variables, especially the seasonal dummy variables, were substantially different. Hence the treatment of all covariates as time invariant does not appear to have caused any serious misspecification error in the estimates of the welfare hazard rate, subject to the qualification that the coefficient estimates on the seasonal dummy variables should be interpreted with caution. The comparison of estimates from models (3) and (8), where the latter model was spell. Though, the unemployment rate had a somewhat more pronounced impact of when allowed to vary within a 74 estimated with a 5 percent subsample of the observations used in (3), provides a guide to the impact of sampling variability on the duration model estimates. It is apparent that random sampling variability contributed to larger differences in the duration model estimates than did the treatment of time varying covariates as fixed, supporting the estimation strategy followed in this chapter. Additional models were estimated which are not reported in Table 3.5. Models (1) and (3) were estimated with the 1980-1986 dummy variable excluded. The exclusion of the dummy variable leads to a small decrease in the magnitude in the coefficient on the minimum wage earnings, and an increase in the magnitude of the coefficients on the unemployment rate and the help wanted index. This implies that the hazard rate was more responsive to changes in the minimum wage, and less responsive to cyclical factors, within the two subperiods compared to the data period as a whole. The sensitivity of these coefficients to the calender time dummy variable is not surprising given that the impact of the variables are identified solely by their time-series variation. However, the coefficients on the spell number dummies and lagged duration are entirely robust to the exclusion of the 1980-1986 dummy variable, and the results regarding the presence of occurrence and lagged duration dependence were not influenced by general calender time trends in the length of welfare spells over the sample period. Further, it is noted that the estimated variance of the heterogeneity distribution is larger when the 1980-1986 dummy variable is excluded. ii. Couples with and without children The estimates for the mixed piece-wise constant proportional hazard models for the sample of spells by couples with and without children are presented in Table 3.9. Column (1) 75 contains the estimates for the most general specification based on first observed spells. The estimates have a similar pattern to that reported for single men and women. In cases where the principal applicant is a woman, is older or is classified as unemployable the welfare exit rate is significantly lower. Surprisingly, the presence of dependent children has a large, positive effect on the welfare exit rate for couples. Potential welfare benefits have the predicted sign; however, in all of the specifications the effect is not statistically significant. Minimum wage earnings have an insignificant effect on the exit rate for initial spells. The unemployment rate has a substantial negative impact on the hazard rate for first spells while the help wanted index is insignificant, similar to that found for singles. The estimates in column (3) are for the sample of repeat spells. As found for initial spells, dependent children are associated with a higher exit rate, which is counterintuitive. The unemployment rate does not significantly affect the exit rate for repeat spells although the help wanted index has a significant positive effect. The differential impact of the unemployment rate and help wanted index across first and repeat spells is similar to that found for the singles sample. Likewise, a potential explanation of the differential impacts of the unemployment rate and help wanted index across spells is the interaction between UI and IA, with new welfare entrants more likely to have first used UI prior to beginning their IA spell. The estimates in column (3) include the set of spell number dummy variables. The coefficient estimates on these dummy variables are uniformly negative and individually significant. Lagged duration has a large, negative impact on the repeat spell hazard rate; an additional month on welfare in the previous spell is associated with a 2.3 percent decline in the exit rate for the subsequent spell. The coefficient estimate for the 1980-1986 dummy variable is significant and positive for the repeat spell sample but significant and negative for the sample of 76 first spells. In models not reported in Table 3.9, the exclusion of the 1980-1986 dummy variable does not lead to any changes in the inferences based on models (1) and (3). The estimates presented in columns (2) and (6) replicate specifications (1) and (3), respectively, except no allowance is made for unobserved heterogeneity. The hypotheses of no heterogeneity is rejected with the likelihood ratio test statistic of 59.98 and 39.41 for the initial and repeat spell samples respectively. Again, controlling for heterogeneity generally leads to an increase in the absolute value, and statistical significance, of the estimated impact of the covariates. Next, the formal hypothesis tests for the presence of state dependencies were implemented. In column (4), the estimation results for the model with the spell number dummy variables excluded are presented. Based on models (3) and (4), the hypothesis of no occurrence is rejected, with a likelihood ratio test statistic of 25.36. Furthermore, the hypothesis of no lagged duration dependence is rejected with a likelihood ratio test statistic of405.42. Therefore there is strong evidence of both negative occurrence and lagged duration dependence in the welfare participation of couples and two-parent families. The baseline hazard function estimates from specifications (1), (2), (3) and (6) are graphed in Figure 3.5, and the estimates from (1) and (3) are reported in Table 3.10. As shown in Figure 3.5, controlling for unobserved heterogeneity reduces the extent of the decline in the baseline hazard rate over the early months of a spell. The Wald statistics, for the test of the decline in the hazard rate over spell length, are presented in Table 3.12. The baseline hazard function for the sample of initial spells is imprecisely estimated, and only the decline between months 3 and 6 is statistically significant. The subsequent rise in the estimates of the hazard over months 17 to 72 is not statistically significant. The hazard function for repeat spells indicate the 77 exit rate declines by almost 6 percentage points over the first three months, and a total of 13 percentage points over the first six months, of a spell. The decline in the hazard over the first 24 months of the repeat spells was statistically significant. Therefore, the Wald tests confirm the presence evidence of negative duration dependence in the hazard rate function, especially within the first two years of repeat spells. The implications of treating all covariates as fixed were explored with a subsample of the spells by couples. A 10 percent random subsample of the repeat spells was drawn, and 8 covariates were permitted to vary within a spell. Specification (3) was re-estimated with the subsample, with the allowance made for the time varying covariates and also treating them as time invariant. The results are reported in columns (7) and (8), respectively, of Table 3.9. The conclusions drawn from a comparison of models (7) and (8) are very similar to those drawn from the same comparison for the sample of spells by singles. The estimates for models (7) and (8) are very similar, with the exception of the seasonal dummy variables. In particular, the coefficient estimates for the fixed covariates, including the welfare history variables which are the basis of the tests of lagged duration and occurrence dependence, are unaffected by the treatment of the time varying covariates. Further, apart from the seasonal dummy variable, the estimates for the time varying covariates in model (7) are within one standard error of the estimates when they are treated as fixed. However, the seasonal dummy variables have a substantially larger, and statistically significant, impact when allowed to vary within a spell. In order to examine the implications of allowing for time varying covariates on the estimates of the baseline hazard function, the baseline hazards for models (7) and (8) are reported in Table 3.11. Allowing for time varying covariates leads to a downward displacement in the baseline function by approximately 2-3 percent points. However, the actual decrease in the 78 hazard rate with spell length is very similar in the two models, as confirmed by the Wald test statistics reported in Table 3.12. Hence the two models also lead to the same inference regarding the presence of duration dependence. iii. Lone parent families The estimates of the mixed piece-wise constant proportional hazard models for the sample of spells by lone parent families are presented in Table 3.13. Column (1) contains the estimates for the most general specification for the sample of first observed spells. The estimates show that lone fathers have a substantially higher exit rate from welfare than lone mothers. The coefficient on age is positive for lone parents, in contrast to the negative effect found for singles and couples, which may reflect the effects of age of dependent children rather than the age of the parent. That is, as dependent children grow older the lone parent would have greater flexibility in pursuing, and accepting, employment opportunities and integrating back into the workforce. As expected, the presence of additional dependent children is associated with a significantly lower exit rate from welfare. Potential welfare benefits have a significant negative effect on the exit rate for lone parents, and the magnitude is greater than that estimated for either singles or couples. This suggests that the labour supply disincentive effects of the program benefits may be more acute for lone parents than other family types. Potential earnings at the minimum wage do not significantly affect the first spell exit rate. Furthermore, the coefficient on the unemployment rate is significant and negative, indicating that the exit behaviour of lone parents is sensitive to cyclical factors. The estimates in column (3) are for the most general specification for the repeat spell 79 sample. Compared to the estimates in (1), additional children have a substantially larger effect on the exit rate for repeat spells. The coefficient on minimum wage earnings is not significant for either the initial or repeat spell hazard, which suggests that the minimum wage may be a poor proxy for either the offered wage or reservation wage of lone parents. Given the opportunity costs of entering fulltime employment, in terms of the welfare benefits forgone and recurrent (per period) child care costs, it is not surprising that exit rate from welfare by lone parents does not respond to changes in the minimum wage. Further, as found for the other family types, the unemployment rate has a significant effect on the initial spell hazard while the help wanted index has a significant effect on the repeat spell hazard. Overall, the estimates indicate that the exit behaviour of lone parents is sensitive to program parameters and general labour market conditions. Model (3) includes information on the history of individuals' interaction with the welfare program. The coefficient on the spell number 3 dummy variable is negative; however, the coefficients on the spell number 4 and 5 dummies are positive.18 The positive sign on the higher spell number dummy variables indicate that those lone parents who have the greatest number of spells cycle on and off the program relatively quickly. In order to experience 4 or more spells within the data period, the length of the welfare spells by the most frequent repeaters are shorter on average than spells by all lone parent spells in general. Specification (3) also controls for lagged duration. The coefficient on lagged duration is negative and highly significant, and indicates that for each month on welfare in the previous spell, the exit rate is approximately 1.5 percent lower in the current spell. The coefficient on the 1980-1986 dummy variable is positive and highly significant for Though, only the coefficient on the spell number 4 dummy variable is individually significant. 80 both the initial and repeat spell hazard. Therefore, over the course of the data period, all spells by lone parents tended to become longer, on average. Additional results, for models not reported in Table 3.13, show that the inferences based on models (1) and (3) are not sensitive to the exclusion of this calender-time variable. Comparison of models (1) with (2) and (3) with (9) provides the basis for testing for potential unobserved heterogeneity. The hypotheses of no heterogeneity is rejected, with likelihood ratio test statistics of 51.567 and 17.76 for the initial and repeat spell samples, respectively. As found for the other family types, controlling for unobserved heterogeneity generally leads to an increase in the absolute value of the coefficient estimates for the included covariates. Formal tests for the presence of the different forms of state dependence were then implemented. The null hypothesis of no occurrence dependence was rejected at the 5 percent level, with a test statistic of 8.52. However, the joint significance of the spell number dummy variables may be due to a form of sample selection bias, due to the limited length of the data period. The hypothesis of no lagged duration dependence is strongly rejected, with a likelihood ratio test statistic of 259.70 . Therefore, as found for the other family types, there is strong evidence of negative lagged duration dependence in the welfare participation of lone parent. Lastly, the presence of duration dependence is examined. The baseline hazard function estimates for models (1), (2), (3) and (6) are graphed in Figure 3.6, and the estimates for models (1) and (3) are reported in Table 3.14. Controlling for unobserved heterogeneity reduces, but does not eliminate, the decline in the exit rate over spell duration. For example, When allowing for gamma heterogeneity, the first spell hazard rate decreases by over 4 percentage points from months 3 to 6, and by a further 1.7 percentage points over the subsequent 6 months. The repeat 81 spell hazard rate declines by over 8 percentage points over the first 6 months and a further 5.9 percentage points over the following 18 months. The decreases in the hazard rates over the length of a spell are statistically significant, as shown by the Wald test statistics reported in Table 3.16. However, the increase in the first spell hazard rate over months 24 to 60 and 72 are not statistically significant. Overall, the evidence supports the presence of negative duration dependence in the program exit behaviour of lone parent families. The consequences of treating all covariates as fixed vrithin a spell in the duration model estimation was also examined with the sample of spells by lone parent families. A 10 percent random subsample of the repeat spells was drawn, and 8 covariates were treated as time varying. The estimates for specification (3), based on the subsample and allowing for time varying covariates are presented in column (7), and the estimates for the subsample assuming all covariates are fixed are presented in column (8). It may be expected that since spells by lone parents are longer, on average, than spells by either singles or couples, that the duration model estimates for this group would be most sensitive to the treatment of time varying covariates. However, as a comparison of models (7) and (8) reveals, there is only a marginal improvement in the value of likelihood function. The effect of dependent children is more pronounced in the model with only fixed covariates; however, the estimated effects of dependent children in model (7) are within one standard error of those for model (8), and in both models the coefficients are insignificant. Additionally, the magnitude of the estimated coefficients for the other covariates are very similar across the two models, with the exception of the seasonal dummy variables. Therefore, although the estimates for the seasonal dummy variables appear to be biased in models which treat them as time invariant, the estimates for the other covariates appear to be robust to the treatment of the time 82 varying covariates. To consider the implications of allowing for time varying covariates on estimates of the baseline hazard function, the estimates from models (7) and (8) are presented in Table 3.15. Again, the baseline for the model with time varying covariates is located below that for the model with only fixed covariates. However, both baselines exhibit a similar rate of decline over spell length as shown by the difference in the point estimates of the hazards, and the associated Wald test statistics, presented in Table 3.16. The two sets of baseline hazards generate the same inferences regarding the presence of duration dependence. Thus, the estimates of the main variables of interest and inferences based on them, for the sample of spells by lone parents, proved not to be sensitive to the treatment of time varying covariates. 3.6. Conclusion In this chapter the distribution of IA spell duration in BC for the 1980-1992 period was described and analysed. It was found that a large proportion of IA spells were of relatively short duration, with median spell lengths of 3, 2 and 5 months for singles, couples and lone parent families, respectively. Additionally, it was found that over 60 percent of the sample of spells corresponded to a repeat spell for an individual or family, indicating a very high incidence of repeat use of IA. The chapter set out to answer two questions: (1) what is the impact of demographic characteristics and labour market conditions on the duration of IA spells; and (2) is there evidence of state dependence in IA receipt? With respect to the first question, it was generally found that women had a lower hazard off IA than their male counterparts. Older individuals and principal claimants for couples had lower hazards, whilst older lone parents had higher hazard 83 rates. However, the positive relationship between lone parent's age and the hazard off IA may be picking up a positive age effect for dependent children. Additional dependent children were found to have a negative effect on the hazard rate for lone parent families but a positive effect on the hazard for couples. As expected, individuals or principal claimants classified as unemployable were found to have a substantially lower hazard rate off IA. The hazard rates for all family types were found to be very sensitive to labour market conditions as measured by the unemployment rate and the help wanted index. The differential effects of these two variables across initial and repeat spells of welfare relates to the timing of their response to the business cycle and suggests one dimension of the interaction between IA and UI. In answer to the second question, the empirical hazard functions and piece-wise constant proportional hazard models without controls for unobserved heterogeneity indicated the presence of negative duration dependence for all the samples of welfare spells examined. Controlling for unobserved heterogeneity reduced but did not eliminate negative duration dependence in the baseline hazard functions. Consequently, much of the decline in the exit rate over spell duration appears to be the result of the program itself leading to changes in individuals' preferences or constraints which make them become more reliant on the program. Therefore, the evidence suggests that welfare, to some extent, acts as a "trap" whereby escape and financial independence becomes less likely the longer one remains on the program. Furthermore, after controlling for unobserved heterogeneity, there is strong evidence of negative lagged duration dependence. There was also strong evidence of negative occurrence dependence for singles and couples but positive occurrence dependence for lone parents, the latter which may be a reflection of sample selection bias. The findings regarding the presence of both negative occurrence dependence and lagged duration dependence imply that greater 84 interaction with the welfare system over time reduces an individual's economic independence. That is, participating in welfare appears to have a "scarring" effect on recipients' labour market careers. The harmful effects accumulate over time, with the number and length of prior spells.19 This suggests that policy makers should consider targeting particular programs that aid integration back into the workforce to recipients based on the history of their welfare participation over a period of time. However, there is a difficulty with the interpretation of results based on models which allow for unobserved heterogeneity in that, by definition, the individual characteristics in question are not identified. Although unobserved characteristics accounted for part of the observed decline in the welfare exit rate over the duration of a spell, policy makers can neither target programs toward individuals according to this characteristic nor can they attempt to encourage acquisition of the characteristic (if it is manipulable). A greater understanding of these results therefore requires investigating the personal characteristics which are not observed in the data used in this study. The results point to the need to acquire better data with information on individual characteristics such as education and human capital, individuals' employment histories, and perhaps measures of ability and motivation. Further policy implications follow from the findings presented in this study. First, the large majority of welfare recipients do not remain on the program for a long period of time. However, the very high incidence of repeat use and the implications of occurrence and lagged 19 The theorectical model presented in section 2 suggests that one avenue through which scarring may occur is the decline in stigma costs with greater participation in welfare over time. However, alternative avenues through which welfare participation may scar a person's labour market career are human capital atrophy, transactions costs, program learning effects and employer screening. The data are unable to discriminate between these competing hypotheses, and therefore no single intrepretation can be favoured. However, the competing intrepretations could be tested if information on recipients' employment histories and level of education were available. The testing of the competing interpretations would also provide an important guide to the design of policies to promote the transition from welfare to the workforce. 85 duration dependence reinforces the need for governments to focus attention on keeping people off welfare once they have exited, rather than just encourage quicker initial exits. That is, governments need to consider more active labour market policies targeted to these individuals to help them become independent and permanently self-sufficient. Furthermore, the differences in the impact of cyclical factors on the hazard rate for initial and repeat spells may indicate one dimension of the interaction between IA and UI. This underscores the need for the government to consider the network of income security programs in designing and evaluating reforms to any one program. This is especially important when exarnining welfare use; since IA is the income security program of last resort, the tightening of the eligibility conditions or reducing the duration of benefits paid under other programs will inevitably impact on the IA caseload. The findings regarding the presence of particular forms of state dependence in welfare exit rates can contribute to a better understanding of the dynamics of the aggregate welfare caseload. The recessions of 1981-82 and 1990-92 had a ratcheting effect on the total IA caseload, whereby the counter-cyclical rise in the caseload was not mirrored by a decline during subsequent periods of employment growth. The findings suggest that the onset of the recession was responsible for individuals and families entering welfare for the first time and then, due to the effects of the three forms of state dependence at the micro level, these individuals and families were prone to longer, repeat spells on welfare. Table 3.1. Potential Social Assistance Benefits in British Columbia, 1980-1992 (at 1992 constant dollars). Month Single, Employable. Couple,2 deps, Employable. Lone Parent, 1 dep, Employable. Jan. 1980 490 1250 880 Jan. 1981 479 1209 853 Jan. 1982 507 1171 851 Jan. 1983 555 1249 913 Jan. 1984 530 1194 873 Jan. 1985 513 1154 843 Jan. 1986 497 1119 818 Jan. 1987 496 1118 818 Jan. 1988 528 1194 898 Jan. 1989 510 1154 927 Jan. 1990 526 1168 941 Jan. 1991 524 1147 930 Jan. 1992 515 1127 914 Table 3.2. Descriptive Statistics for Single Men and Women Sample. Variable 1st Spell Repeat Spells Duration 6.053 7.023 Female 0.326 0.261 Age 32.651 36.024 Unemployable 0.133 0.114 Benefits($100) 5.290 5.265 Min.wg.($100) 9.195 9.035 Unemp.Rate 0.089 0.092 Hwi 0.545 0.569 D1980-1986 0.552 0.440 DSpelll 1.000 DSpell2 0.420 DSpelB 0.271 DSpeU4 0.183 DSpell5 0.126 Lagged Duration 6.163 DQuarterl 0.259 0.240 DQuarter2 0.240 0.232 DQuarter3 0.241 0.235 DQuarter4 0.260 0.293 Right cens. 0.057 0.077 Observations 48020 65325 Table 3.3. Descriptive Statistics for Couples Sample. Variable 1st Spell Repeat Spells Duration 5.722 6.417 Female 0.198 0.184 Dep. Children 1.346 1.468 Age 38.019 38.866 Unemployable 0.085 0.071 Benefits($100) 10.928 10.953 Min.wg.($100) 9.055 9.032 Unemp.Rate 0.088 0.095 Hwi 0.515 0.537 D1980-1986 0.666 0.536 DSpelll 1.000 DSpell2 0.421 DSpell3 0.270 DSpell4 0.185 DSpell5 0.124 Lagged Duration 5.954 DQuarterl 0.274 0.252 DQuarter2 0.220 0.219 DQuarter3 0.236 0.230 DQuarter4 0.270 0.299 Right cens. 0.046 0.054 Observations 10486 15135 Table 3.4. Descriptive Statistics for Lone Parent Family Sample. Variable 1st Spell Repeat Spells Duration 12.641 12.385 Female 0.933 0.921 Dep. Children 1.669 1.634 Age 40.227 37.113 Unemployable 0.415 0.311 Benefits($100) 10.209 10.114 Min.wg.($100) 9 .137 9.043 Unemp.Rate 0.085 0.088 Hwi 0.552 0.569 D1980-1986 0.587 0.425 DSpelll 1.000 DSpell2 0.431 DSpell3 0.276 DSpelW 0.175 DSpell5 0.117 Lagged Duration 10.892 DQuarterl 0.229 0.225 DQuarter2 0.243 0.238 DQuarter3 0.289 0.273 DQuarter4 0.239 0.264 Right cens. 0.125 0.146 Observations 10380 17750 Table 3.5. Duration Model Estimates, Single Men and Women Sample. Variable 1st Spell 1st Spell Rep. Spell Rep. Spell (1) (2) (3) (4) Female -0.1464** -0.1286** -0.1200** -0.1152** (0.0135) (0.0102) (0.0110) (0.0110) Age (lOyrs) -0.0893** -0.0757** -0.0807** -0.0847** (0.0065) (0.0047) (0.0054) (0.0055) Unemp'ble -0.2838** -0.2057** -0.2314** -0.2283** (0.0295) (0.0218) (0.0239) (0.0240) Benefit ($100) -0.0297 -0.0269 -0.0716** -0.0758** (0.0319) (0.0235) (0.0263) (0.0269) Min.Wg. ($100) 0.0815** 0.0667** 0.0282** 0.0305** (0.0135) (0.0103) (0.0102) (0.0102) Unemp.Rt.(%) -0.0657** -0.0429** -0.0172** -0.0183** (0.0045) (0.0032) (0.0035) (0.0035) Hwi (%) 0.0020** 0.0019** 0.0046** 0.0045** (0.0006) (0.0004) (0.0004) (0.0004) D19866 0.0546** 0.0224 -0.0657** -0.0610** (0.0194) (0.0153) (0.0148) (0.0145) DSpell3 -0.0202* (0.0119) DSpell4 -0.0542** (0.0138) DSpell5 -0.0769** (0.0159) Lag.Duration -0.2128** -0.2121** (10 months) (0.0063) (0.0063) DQuart.1 0.1382** 0.0968** 0.1258** 0.1278** (0.0184) (0.0140) (0.0144) (0.0145) DQuart.2 0.1032** 0.0714** 0.1317** 0.1329** (0.0178) (0.0136) (0.0137) (0.0138) DQuart.3 0.0550** 0.0318** 0.0563** 0.0563** (0.0178) (0.0137) (0.0137) (0.0138) Sigma2 0.3617** 0.2054** 0.2101** (0.0367) (0.0272) (0.0274) LLF -101346 -101420 -144692 -144676 Note:(l) Asymptotic standard errors are in parentheses. (2) * * denotes significance at the 1 percent level and * denotes significance at the 5 percent level. Table 3.5. Duration Model Estimates, Single Men and Women Sample Continued. Variable Rep. Spell Rep. Spell Rep. Spell Rep.Spell (5) (6) (7) (8) Female -0.1214** -0.1226** -0.2020** -0.1887** (0.0107) (0.0093) (0.0417) (0.0481) Age (lOyrs) -0.1111** -0.0722** -0.0745** -0.0718** (0.0056) (0.0045) (0.0197) (0.0222) Unemp'ble -0.2642** -0.1778** -0.2394** -0.2115* (0.0233) (0.0200) (0.0973) (0.1093) Benefit ($100) -0.0562** -0.0843** -0.0751 -0.0870 (0.0251) (0.0219) (0.1066) (0.1219) Min.Wg. ($100) -0.0593** 0.0214** 0.0058 0.0015 (0.0098) (0.0087) (0.0373) (0.0467) Unemp.Rt.(%) -0.0187** -0.0104** -0.0644** -0.0391** (0.0034) (0.0029) (0.0144) (0.0185) Hwi (%) 0.0030** 0.0041** 0.0040** 0.0034** (0.0004) (0.0004) (0.0017) (0.0019) D19866 -0.0663** -0.0702** -0.1085* -0.0445 (0.0142) (0.0128) (0.0527) (0.0626) DSpell3 -0.0156 -0.0925 -0.1097* (0.0101) (0.0471) (0.0533) DSpelW -0.0464** , -0.0467 -0.0532 (0.0117) (0.0506) (0.0577) DSpell5 -0.0683** -0.0437 -0.0446 (0.0135) (0.0604) (0.0691) Lag.Duration -0.1823** -0.2064 -0.2427** (10 months) (0.0049) (0.0246) (0.0274) DQuart.l 0.1290** 0.0981** 0.2239** 0.1414** (0.0139) (0.0121) (0.0576) (0.0632) DQuart.2 0.1307** 0.1044** 0.2526** 0.0708 (0.0133) (0.0115) (0.0537) (0.0586) DQuart.3 0.0490** 0.0424** 0.2578** -0.0112 (0.0132) (0.0117) (0.0548) (0.0579) Sigma2 0.1435** — 0.1387** 0.2124** (0.0316) (0.0733) (0.0935) LLF -145512 -144692 -6992 -7023 Note:(l) Asymptotic standard errors are in parentheses. (2) ** denotes significance at the 1 percent level and * denotes significance at the 5 percent level. Table 3.6. Average (Per Month) Baseline Hazard Rate Estimates, Single Men and Women Sample, Interval (months) Specification (1): 1st Spell Specification (3): Repeat Spells Hazard Std.Err. Hazard StdErr. 1 0.3830 0.0073 0.3039 0.0046 2 0.4108 0.0113 0.3351 0.0066 3 0.3531 0.0134 0.2740 0.0072 4 0.2934 0.0136 0.2244 0.0073 5 0.2595 0.0139 0.1988 0.0074 6 0.2296 0.0137 0.1791 0.0075 7-8 0.2210 0.0142 0.1677 0.0075 9-10 0.1960 0.0144 0.1502 0.0077 11-12 0.1789 0.0146 0.1277 0.0073 13-16 0.1716 0.0150 0.1194 0.0073 17-20 0.1586 0.0157 0.1061 0.0074 21-24 0.1448 0.0159 0.1030 0.0079 25-30 0.1472 0.0173 0.0896 0.0075 31-36 0.1353 0.0178 0.0825 0.0076 37-42 0.1350 0.0192 0.0737 0.0075 43-51 0.1580 0.0240 0.0788 0.0083 52-60 0.1491 0.0255 0.0715 0.0084 61-72 0.1365 0.0260 0.0714 0.0092 Table3.7. Average (Per Month) Baseline Hazard Rate Estimates, Single Men and Women Subsample. Interval (months) Specification (7): Repeat Spells Specification (8): Repeat Spells Hazard StdErr. Hazard StdErr. 1 0.2236 0.0239 0.2614 0.0315 2 0.2250 0.0246 0.2758 0.0340 3 0.1749 0.0209 0.2231 0.0303 4 0.1343 0.0177 0.1771 0.0266 5 0.1161 0.0165 0.1577 0.0257 6 0.1022 0.0158 0.1427 0.0253 7-8 0.0822 0.0127 0.1175 0.0213 9-10 0.0719 0.0123 0.1063 0.0213 11-12 0.0622 0.0118 0.0945 0.0209 13-16 0.0548 0.0105 0.0844 0.0193 17-20 0.0432 0.0093 0.0685 0.0175 21-24 0.0309 0.0077 0.0520 0.0152 25-30 0.0381 0.0094 0.0665 0.0197 31-36 0.0265 0.0076 0.0491 0.0165 37-42 0.0229 0.0073 0.0453 0.0168 43-51 0.0294 0.0087 0.0621 0.0225 52-60 0.0245 0.0085 0.0546 0.0227 61-72 0.0252 0.0091 0.0654 0.0288 94 Table 3.8. Differences in Baseline Hazard Rates (and Wald Test Statistics); Single Men and Women Sample. Difference in Hazard Rates: Specification (1) Specification (3) Specification (7) Specification (8) 1st Spells Repeat Spells Repeat Spells Repeat Spells 1 - 3 0.0299*** 0.0299*** 0.0487*** 0.0383** (5.086) (5.032) (3.100) (1.853) 3- 6 0.1235*** 0.0949*** 0.0727*** 0.0804*** (18.569) (18.569) (4.699) (3.913) 6- 12 0.0507*** 0.0514*** 0.0400*** 0.0482*** (10.370) (10.370) (3.220) (2.696) 12-24 0.0341*** 0.0247*** 0.0313*** 0.0425*** (5.220) (5.220) (3.243) (2.787) 24-36 0.0095*** 0.0205*** 0.0044 0.0029 (4.218) (4.218) (0.593) (0.219) 36-60 -0.0138 0.0110** 0.0020 -0.0055 (-1.935) (1.935) (0.252) (-0.318") Notes: (1) The Wald Statistic is for the test of the null hypothesis that Hazard (Month Tl) = Hazard (Month T2) against the alternative hypothesis that Hazard (Month Tl) > Hazard (Month T2). (2) The * * * denotes that the difference in the hazard rates is significant at the 1 percent level, * * denotes significance at the 5 percent level and * denotes significance at the 10 percent level. Table 3.9. Duration Model Estimates, Couples With and Without Children Sample. Variable 1st Spell (!) 1st Spell (2) Rep.Spell (3) Rep.Spell W Female Age (lOyrs) Dep.l Dep.2 Unemp'ble ** -0.1221 (0.0387) -0.1448** (0.0175) 0.4233** (0.1592) 0.5326** -0.0141 (0.0250) -0.0978** (0.0105) 0.3069** (0.1051) 0.4414** -0.1129 (0.0272) -0.0662 (0.0123) 0.1709* (0.1058) 0.2544* ** (0.2252) -0.3166 (0.0643) ** (0.1478) -0.1944 (0.0435) ** (0.1494) -0.2545 (0.0462) -0.1058 (0.0273) -0.0659** (0.0124) 0.1792* (0.1062) 0.2630* (0.1500) -0.2508** (0.0464) Benefit ($100) Min.Wg. ($100) Unemp.Rt. (%) Hwi (%) -0.1137 (0.0589) 0.0340 (0.0449) -0.0616** (0.0091) 0.0006 (0.0015) -0.0819 (0.0377) 0.0202 (0.0288) -0.0400 (0.0057) -0.0003 (0.0010) -0.0785 (0.0427) 0.0950** (0.0331) -0.0050 (0.0067) 0.0047** (0.0010) -0.0858' (0.0430) 0.1027** (0.0333) -0.0072 (0.0067) 0.0045** (0.0010) D19866 0.1170" (0.0493) 0.0383 (0.0343) -0.1431 (0.0326) -0.1274 (0.0325) DSpell3 DSpell4 DSpell5 -0.0454 (0.0259) -0.1423** (0.0302) -0.0944** (0.0358) Lag.Duration (10 months) -0.2299 (0.0144) -0.2296 (0.0144) DQuart.l DQuart.2 DQuart.3 0.2364 (0.0441) 0.0147 (0.0444) -0.0608 (0.0433) 0.1404 (0.0279) 0.0038 (0.0297) -0.0667** (0.0293) 0.1197 (0.0305) 0.1050** (0.0305) -0.0112 (0.0298) 0.1260 (0.0306) 0.1067** (0.0307) -0.0122 (0.0299) Sigma2 0.6322 (0.1144) 0.2965 (0.0652) 0.3039 (0.0660) ** LLF -21797 -21827 -32793 -32806 Table 3.9. Duration Model Estimates, Couples With and Without Children Sample Continued. Variable Rep. Spell Rep.Spell Rep.Spell Rep.Spell (5) (6) (7) (8) Female -0.1492** -0.1190** -0.2812** -0.2886** (0.0255) (0.0216) "(0.0857) (0.0937) Age (lOyrs) -0.0720** -0.0548** -0.0598* -0.0658* (0.0116) (0.0096) (0.0366) (0.0401) Dep.l 0.2638** 0.1352* 0.2149 0.2217 (0.0961) (0.0835) (0.1844) (0.2019) Dep.2 0.3708** 0.1954* 0.2986 0.3138 (0.1357) (0.1182) (0.2815) (0.3062) Unemp'ble -0.2628** -0.2098** -0.2598* -0.2838* (0.0439) (0.0359) (0.1320) (0.1451) Benefit ($100) -0.0765* -0.0682* -0.0962 -0.0450 (0.0385) (0.0333) (0.2359) (0.2667) Min.Wg. ($100) 0.1159** 0.0802** 0.2163 0.1660 (0.0301) (0.0257) (0.1714) (0.1976) Unemp.Rt. (%) -0.0083 0.0025 -0.0290 -0.0182 (0.0060) (0.0052) (0.0195) (0.0220) Hwi (%) 0.0030** 0.0038** 0.0025 0.0024 (0.0009) (0.0008) (0.0030) (0.0033) D19866 -0.0916** -0.1360** -0.1540* -0.1206 (0.0294) (0.0262) (0.0947) (0.1058) DSpell3 -0.0411* -0.0308 -0.0316 (0.0206) (0.0792) (0.0849) DSpeU4 -0.1185** -0.0551 -0.0715 (0.0237) (0.0916) (0.0983) DSpellS -0.0706** -0.1493 -0.1799* (0.0287) (0.1054) (0.1135) Lag.Duration -0.1823** -0.2060** -0.2269** (10 months) (0.0104) (0.0455) (0.0496) DQuart.l 0.1228** 0.0754** 0.1699* 0.1081 (0.0280) (0.0236) (0.0855) (0.1004) DQuart.2 0.1026** 0.0793** 0.1828** 0.0750 (0:0280) (0.0242) (0.0829) (0.1018) DQuart.3 -0.0113 -0.0156 0.1748* 0.0598 (0.0270) (0.0240) (0.0827) (0.1004) Sigma2 0.1455** 0.1785 0.2862 (0.0774) (0.1771) (0.2187) LLF -33006 -32807 -3196 -3207 Table 3.10. Average (Per Month) Baseline Hazard Rate Estimates, Couples With and Without Children Sample. Interval (months) Specification (1): 1st Spell Specification (3): Repeat Spells Hazard Std.Err. Hazard StdErr. 1 0.2750 0.0388 0.3427 0.0350 2 0.3143 0.0462 0.3671 0.0395 3 0.2827 0.0457 0.2846 0.0331 4 0.2603 0.0460 0.2489 0.0310 5 0.2301 0.0439 0.1980 0.0263 6 0.2012 0.0412 0.2061 0.0287 7-8 0.2166 0.0468 0.1846 0.0269 9-10 0.2159 0.0514 0.1782 0.0280 11-12 0.2058 0.0532 0.1742 0.0295 13-16 0.2123 0.0594 0.1411 0.0255 17-20 0.2080 0.0646 0.1328 0.0263 21-24 0.2122 0.0717 0.1152 0.0248 25-30 0.2126 0.0778 0.1188 0.0269 31-36 0.1960 0.0795 0.1086 0.0268 37-42 0.3013 0.1290 0.0983 0.0262 43-51 0.2842 0.1343 0.1211 0.0337 52-60 0.3763 0.1940 0.0952 0.0294 61-72 0.4063 0.2316 0.1054 0.0344 98 Table 3.11. Average (Per Month) Baseline Hazard Rate Estimates, Couples With and Without Children Interval (months) Specification (7): Repeat Spells Specification (8): Repeat Spells Hazard StdErr. Hazard Std.Err. 1 0.2385 0.0569 0.2636 0.0662 2 0.2383 0.0592 0.2749 0.0725 3 0.2142 0.0581 0.2561 0.0751 4 0.1636 0.0490 0.2000 0.0661 5 0.1333 0.0433 0.1655 0.0594 6 0.1149 0.0396 0.1444 0.0557 7-8 0.0835 0.0296 0.1069 0.0428 9-10 0.0749 0.0289 0.0974 0.0426 11-12 0.0979 0.0391 0.1295 0.0601 13-16 0.0695 0.0304 0.0937 0.0481 17-20 0.0643 0.0305 0.0889 0.0496 21-24 0.0358 0.0196 0.0508 0.0322 25-30 0.0749 0.0394 0.1079 0.0683 31-36 0.0680 0.0408 0.1071 0.0763 37-42 0.0517 0.0353 0.0831 0.0658 43-51 0.0302 0.0226 0.0537 0.0466 52-60 0.0311 0.0251 0.0590 0.0543 61-72 0.0258 0.0212 0.0500 0.0472 Subsample. 99 Table 3.12. Differences in Baseline Hazard Rates (and Wald Test Statistics); Couples with and without Children Sample. Difference in Hazard Rates: Specification (1) Specification (3) Specification (7) Specification (8) 1st Spells Repeat Spells Repeat Spells Repeat Spells 1 - 3 -0.0077 0.0581*** 0.0243 0.0075 (-0.308) (3.630) (0.727) (0.167) 3- 6 0.0815*** 0.0785*** 0.0993*** 0.1117*** (4.111) (5.501) (2.980) (2.787) 6- 12 -0.0046 0.0319*** 0.0170 0.0149 (-0.194) (2.394) (0.712) (0.474) 12-24 -0.0064 0.0590*** 0.0621** 0.0787** (-0.203) (4.276) (2.318) (2.088) 24-36 0.0162 0.0066 -0.0322 -0.0563 (0.502) (0.495) (-1.133) (-1.074) 36-60 -0.1803 0.0134 0.0369* 0.0481 (-1.402) (0.908) (1.300) (1.022) Notes: (1) The Wald Statistic is for the test of the null hypothesis that Hazard (Month Tl) = Hazard (Month T2) against the alternative hypothesis that Hazard (Month Tl) > Hazard (Month T2). (2) The * * * denotes that the difference in the hazard rates is significant at the 1 percent level, * * denotes significance at the 5 percent level and * denotes significance at the 10 percent level. Table 3.13. Duration Model Estimates, Lone Parent Family Sample. Variable 1st Spell 1st Spell Rep.Spell Rep.Spell (1) (2) (3) (4) Male 0.5100** 0.3505** 0.2855** 0.2862** (0.0745) (0.0399) (0.0396) (0.0396) Age (lOyrs) 0.0560** 0.0065 0.0970** 0.0990** (0.0200) (0.0126) (0.0139) (0.0139) Dep.2 -0.0055 0.0384 -0.1664** -0.1672** (0.1094) (0.0709) (0.0644) (0.0644) Dep.3+ -0.1801 -0.0796 -0.3212** -0.3235** (0.1736) (0.1137) (0.1023) (0.1022) Unemp'ble -0.2880** -0.1861** -0.2298** -0.2307** (0.0459) (0.0278) (0.0285) (0.0285) Benefit ($100) -0.1893* -0.1131* -0.1472** -0.1465** (0.1036) (0.0647) (0:0654) (0.0654) Min.Wg. ($100) 0.0991 0.0790 0.0203 0.0193 (0.0852) (0.0530) (0.0536) (0.0535) Unemp.Rt. (%) -0.0439** -0.0209** -0.0083 -0.0071 (0.0101) (0.0061) (0.0069) (0.0069) Hwi (%) 0.0006 0.0001 0.0046** 0.0047** (0.0014) (0.0010) (0.0009) (0.0009) D19866 0.2955** 0.1290** 0.1067** 0.0995** (0.0573) (0.0370) (0.0336) (0.0332) DSpell3 -0.0133 (0.0246) DSpell4 0.0701** (0.0296) DSpell5 0.0338 (0.0343) Lag.Duration -0.1051** -0.1066** (10 months) (0.0075) (0.0075) DQuart.1 0.1249** 0.0500 0.0537* 0.0517** (0.0500) (0.0314) (0.0301) (0.0301) DQuart.2 0.0401 0.0012 0.0443 0.0450 (0.0484) (0.0302) (0.0289) (0.0289) DQuart.3 -0.0203 -0.0461 -0.0002 -0.0005 (0.0465) (0.0302) (0.0279) (0.0279) Sigma2 0.8344** 0.3173** 0.3166** (0.1627) (0.0779) (0.0784) LLF -25875 -25901 -43591 -43596 Table 3.13. Duration Model Estimates, Lone Parent Fami y Sample Continued. Variable Rep.Spell Rep.Spell Rep.Spell Rep.Spell (5) (6) (7) (8) Male 0.3635** 0.2350** 0.1478 0.1552 (0.0449) (0.0296) (0.1458) (0.1545) Age (10 yrs) 0.0858** 0.0708** 0.2530** 0.2681** (0.0151) (0.0100) (0.0579) (0.0619) Dep.2 -0.1532** -0.1431** -0.3080 -0.5543 (0.0702) (0.0521) (0.2109) (0.2593) Dep.3+ -0.2950** -0.2729** -0.5021 -0.9292 (0.1111) (0.0828) (0.3371) (0.4172) Unemp'ble -0.2471** -0.1875** -0.1751* -0.1657 (0.0317) (0.0219) (0.0985) (0.1083) Benefit ($100) -0.1705** -0.1069* -0.2113 -0.2959 (0.0715) (0.0527) (0.1941) (0.2532) Min.Wg. ($100) 0.0545 0.0023 0.0078 0.0159 (0.0584) (0.0433) (0.1586) (0.2091) Unemp.Rt. (%) -0.0116 -0.0034 -0.0520** -0.0624** (0.0075) (0.0056) (0.0023) (0.0289) Hwi (%) 0.0044** 0.0037** 0.0011 0.0012 (0.0010) (0.0007) (0.0028) (0.0038) D19866 0.1486** 0.0633** 0.0856** 0.1431 (0.0365) (0.0274) (0.1028) (0.1311) DSpell3 -0.0176 -0.0351 -0.0337 (0.0201) (0.0947) (0.0993) DSpelW 0.0520** 0.1008 0.1057 (0.0237) (0.1120) (0.1188) DSpell5 0.0172 -0.0521 -0.0440 (0.0279) (0.1316) (0.1393) Lag.Duration -0.0890** -0.1343 -0.1411 (10 months) (0.0056) (0.0310) (0.0321) DQuart.1 0.0590* 0.0401* 0.1199 -0.0048 (0.0329) (0.0243) (0.0858) (0.1211) DQuart.2 0.0506* 0.0369 0.1718 0.0178 (0.0315) (0.0235) (0.0783) (0.1188) DQuart.3 0.0097 -0.0088 0.1482 -0.0113 (0.0304) (0.0228) (0.0793) (0.1147) Sigma2 0.4878** 0.7170** 0.8439** (0.1087) (0.3333) (0.3431) LLF -43732 -43600 -4300 -4301 Table 3.14. Average (Per Month) Baseline Hazard Rate Estimates, Lone Parent Family Sample. Interval (months) Specification (1): 1st Spell Specification (3): Repeat Spells Hazard StdErr. Hazard StdErr. 1 0.1942 0.0141 0.2015 0.0090 2 0.2077 0.0175 0.1932 0.0097 3 0.1923 0.0197 0.1568 0.0093 4 0.1739 0.0208 0.1301 0.0088 5 0.1536 0.0207 0.1227 0.0091 6 0.1493 0.0221 0.1170 0.0095 7-8 0.1381 0.0224 0.1065 0.0091 9-10 0.1341 0.0245 0.0905 0.0087 11-12 0.1322 0.0265 0.0834 0.0088 13-16 0.1229 0.0270 0.0738 0.0084 17-20 0.1126 0.0278 0.0642 0.0082 21-24 0.1288 0.0348 0.0709 0.0098 25-30 0.1204 0.0355 0.0664 0.0100 31-36 0.1309 0.0424 0.0587 0.0098 37-42 0.1563 0.0552 0.0626 0.0113 43-51 0.1535 0.0588 0.0546 0.0106 52-60 0.2016 0.0856 0.0632 0.0133 61-72 0.2576 0.1216 0.0772 0.0177 Table 3.15. Average (Per Month) Baseline Hazard Rate Estimates, Lone Parent Family Subsample. Interval (months) Specification (7): Repeat Spells Specification (8): Repeat Spells Hazard StdErr. Hazard StdErr. 1 0.1891 0.0319 0.2383 0.0480 2 0.1693 0.0323 0.2186 0.0498 3 0.1732 0.0379 0.2272 0.0585 4 0.1323 0.0337 0.1763 0.0517 5 0.1419 0.0397 0.1918 0.0608 6 0.1473 0.0449 0.2012 0.0691 7-8 0.1111 0.0374 0.1534 0.0575 9-10 0.0941 0.0354 0.1308 0.0546 11-12 0.0979 0.0394 0.1372 0.0606 13-16 0.1073 0.0472 0.1536 0.0739 17-20 0.0875 0.0432 0.1263 0.0676 21-24 0.1122 0.0601 0.1645 0.0949 25-30 0.1520 0.0900 0.2293 0.1465 31-36 0.1041 0.0710 0.1607 0.1167 37-42 0.1059 0.0782 0.1675 0.1318 43-51 0.1211 0.0949 0.1964 0.1642 52-60 0.1387 0.1198 0.2347 0.2150 61-72 0.2336 0.2230 0.3462 0.3086 104 Table 3.16. Differences in Baseline Hazard Rates (and Wald Test Statistics); Lone Parent Family Sample. Difference in Hazard Rates: Specification (1) Specification (3) Specification (7) Specification (8) 1st Spells Repeat Spells Repeat Spells Repeat Spells 1 - 3 0.0019 0.0447*** 0.0159 0.0111 (0.135) (6.344) (0.582) (0.308) 3- 6 0.0430*** 0.0398*** 0.0259 0.0260 (3.742) (6.173) (1.004) (0.737) 6- 12 0.0171* 0.0336*** 0.0494** 0.0640** (1.405) (5.754) (2.099) (1.935) 12-24 0.0034 0.0125*** -0.0143 -0.0273 (0.241) (2.476) (-0.465) (-0.576) 24-36 -0.0021 0.0122*** 0.0081 0.0038 (-0.144) (2.563) (0.286) (0.084) 36-60 -0.0707 -0.0045 -0.0346 -0.0740 (-1.497) (-0.6m f-0.564) C-0.642) Notes: (1) The Wald Statistic is for the test of the null hypothesis that Hazard (Month Tl) = Hazard (Month T2) against the alternative hypothesis that Hazard (Month Tl) > Hazard (Month T2). (2) The * * * denotes that the difference in the hazard rates is significant at the 1 percent level, * * denotes significance at the 5 percent level and * denotes significance at the 10 percent level. Figure 3.1. Budget Set with Welfare Program Hours Worked ~~• 106 107 c n o o r ^ c D L n N - n c N v - o 6 0 6 0 6 0 0 0 6 0 A ^ j n q D q o J d | D A i A j n s 108 9 } D ^ p j D Z D H 109 CD C L E o if) c CD ~o _c C J - M Z5 o L O TJ c o CD I. J Z •+-> gui !M Ll_ CO CD C L Z5 o o ( ) O o CD CD c n x : O en o CD -4-* CD (U x: £ D en O i _ CD CD CTl O c CD CD C L C L CO CO CD CD C L C L CO CO -t—' -t-J 0 o CD CD C L C L CD CD 01 Od I d) 0 <b i i CM C D C D o C D LO 00 C N C D o -4-C N O O CN H co o d d d CN d o o d 110 CD CL c c o (/) E o Li_ c CD CO o Q_ CD CD L_ c Loi cn Loi O i O i _ CD CD -t-> O CD c n i_ _C o CD i_ CD c n o CD -4-< J C O E CD a CD -+-» E JZ E CD o o E _ c c n c a o _ u _ u CD c "CD "CD C L C L ^ — CO CO CD CD CL C L -f—* o -M D CO CO CD CD C L CL CO CO CD CD T — 1— a: DC I I i i <0 4) Q) 0 i i II II u CM C D C D O CD -3-00 CN C D ro O ro CM 00 CM H CD o LO 6 6 ro 6 CN 6 o o d 9 } D y pJDZDf- j I l l Chapter 4. Unemployment Insurance Spells and Welfare Receipt in British Columbia 4.1. Introduction Unemployment insurance (UI) and social assistance (S A) are the two most important income security programs for Canada's working age population. In 1992 the combined expenditures for the two programs accounted for more than $3 2b or approximately 5 percent of GDP. The average number of beneficiaries of the programs during 1992 was 1.4 million for UI and approximately 2.7 million for social assistance1 or welfare. Over recent years there has been a growing concern for the fiscal sustainability and the efficacy of the income security programs. Indeed, the federal government is currently considering major reforms to the income security system in an effort to reduce costs and to better promote the transition from assistance to work. In considering alternate reform proposals it is important to first understand how individuals use and interact with the current programs. There is an established, and expanding, literature examining the use and behavioral effects of UI in Canada. These studies include the work of Ham and Rea (1987), Green and Riddell (1993) and Corak (1993a,b). Recently a number of researchers, such as Allen (1993), Bruce et. al. (1993), Charette and Meng (1994) and Dooley (1994), have begun describing and analysing how individuals and families interact with Canada's welfare programs. However, to date, there is very little work which directly examines UI and welfare together, as part of an income security system. This is an important gap because there is a widespread perception that the two income security programs increasingly overlap. 1 The number of beneficiaries of UI corresponds to the number of individual claimants receiving benefits whereas for social assistance the number of beneficiaries includes all family members. 112 The objective of this chapter is to analyse one form of interaction between UI and welfare in detail. Specifically, the research presented in this chapter empirically tests the hypothesis that the receipt of welfare increases the expected duration of subsequent UI spells. As detailed in the following section, this form of interaction may result from human capital atrophy, employer screening or from learning about program rules and administration. The hypothesis is tested in a hazard function framework using individual level data derived from the administration of the two programs. In particular, this study estimates the hazard rate, the probability of leaving UI conditional on the length of the current UI spell, using micro data on a sample of males in British Columbia for the period 1988-1992. The chapter is organised as follows. In section 2 an economic model of UI durations is presented and the prediction of welfare receipt leading to longer expected UI durations is discussed. In section 3 the econometric approach for estimating the model and testing the hypothesis is outlined and in section 4 the data set used in the estimation, and the specification of the independent variables, is described. The empirical results are presented in section 5, and section 6 concludes the chapter. 4.2. Economic Model A common approach to analysing the dynamics of unemployment and UI spells is with job search models, as shown in the survey by Devine and Kiefer (1991). Following the basic search model of Mortensen (1977), individuals are assumed to maximise the present value of utility, with utility defined as a function of income and leisure. Individuals face a stationary, known wage offer distribution and the job offer arrival rate, for a given search intensity, is assumed constant. Mortensen derives an expression for the hazard (or exit rate) from 113 unemployment for individual /' as proportional to (1) A, = st[l-F(w)] where 5 is search intensity, w is the reservation wage and F(.) is the cumulative distribution of wage offers. An increase in search intensity increases the arrival rate of job offers and hence the hazard rate. A decrease in the reservation wage increases the probability of a job offer being acceptable which also leads to an increase in the hazard rate. As an individual approaches UI benefit exhaustion, s is shown to increase while w decreases, both causing the hazard rate to increase. The search model extended to a non-stationary environment leads to a specification of the unemployment hazard as: where tQ denotes the calender date at which the spell began and t is the current duration of the spell. Given the nature of the data, a reduced form version of equation (2) is estimated. It is assumed that the search intensity (or job offer arrival rate) and reservation wage are functions of demographic characteristics, demand conditions in the labour market, parameters of the UI system and the duration of the current Ul-unemployment spell. Participation in the welfare program may be expected to have a negative effect on the UI hazard for several reasons. Firstly, the experience of being on welfare may have a "scarring" effect on an individual's employment career. While on welfare an individual is not in employment and so is unable to accumulate work experience while their human capital atrophies. Additionally, employers may use individuals' employment and welfare histories (if known) in determining job offers. Together, these effects would lead to a systematically lower arrival rate of job offers, ceteris paribus, for individuals with a history of welfare receipt. In turn, the lower (2) Afi,tJ « s/t\tJ.[l-F(wfl,tJ)] 114 job offer arrival rate implies a lower hazard and hence longer expected duration of unemployment and UI. Furthermore, the experience of being on welfare may influence an individual's subsequent search behaviour when unemployed. By participating on welfare an individual learns about the operation of that program, including the eligibility criteria and benefit rates. Knowledge of the welfare program and having greater certainty regarding transfer income when UI is exhausted may cause individuals to search less intensively when subsequently on UI. Therefore, there are numerous avenues (human capital, screening, learning) through which recent welfare participation may contribute to longer expected unemployment spell durations. Although it is not possible to identify particular source(s) of such program spillover effects, given the properties of the available data, this paper tests for evidence of welfare recipiency leading to longer expected UI spells. In analysing the duration of unemployment spells it is important to control for features of the UI system. For the purposes at hand, the most important features of the UI system relate to the potential duration of benefits and the benefit rate. For the time period under study, 1988-1992, there was a major change to the program which became effective on November 18, 1990. From 1988 to November 1990, the potential duration of benefits was determined according to a three phase structure: 1.Initial benefit phase, which provided one week of benefits for each week of insured employment, up to a maximum of 25 weeks, 2. Labour force extended phase, which provided for an additional week of benefits for each two weeks of insured employment in excess of 26 weeks, up to a maximum of 13 weeks, 3. Regional extended benefit phase, which provided for two weeks of benefits for each 0.5 115 percentage point increment in the regional unemployment rate in excess of 4.0 percent, up to a maximum of 32 weeks. The maximum duration of benefits from all three phases was 50 weeks. After November 18, 1990, the three phase structure for the duration of benefits was replaced with a single phase. The potential duration of benefit remained a function of the regional unemployment rate; however, the duration of benefits for a given unemployment rate and weeks of insured employment were generally lower than under the three phase structure. Additional changes introduced in November 1990 included an increase in the number of UI regions from 48 to 62. Although the number of regions in B.C. remained at six, the boundaries of the regions were changed. Further, over the entire data period, the benefit rate remained at 60 percent of average weekly earnings from the 20 most recent weeks of insured employment, subject to a maximum level of insured earnings.2 The program rules and the appropriate regional unemployment rates were used to construct the potential duration of benefits3 at the commencement of each spell. In turn, the number of weeks remaining before benefits exhaust, at the end of the spell, defined as potential benefit duration minus actual duration and weeks of disqualification, is constructed and used as a control variable in the estimation. In this way, two sources of time dependence in the unemployment hazard are identified, as discussed by Ham and Rea (1987). Firstly, there is a direct duration effect, whereby the hazard rate changes with the length of the spell holding weeks until exhaustion constant. Secondly, there is an indirect duration effect with the hazard rate changing due to the decrease in the weeks until exhaustion as a spell progresses. This latter 2 The maximum level of insurable earnings was increased yearly to reflect changes in average earnings. 3 Ham and Rea (1987) term this "initial entitlement". 116 effect is identified through variation in individuals' potential duration of benefits, which arises due to variation in regional unemployment rates across regions and over time and in individuals' employment histories. The total duration dependence effect is then given by the addition of the direct and indirect effects. 4.3. Econometric Methods The empirical analysis is conducted in a hazard function framework. The analysis proceeds by first estimating the empirical hazard rate function. The empirical hazard rate function reveals information on the overall shape of the UI hazard without imposing a parametric function on the underlying distribution. A limitation of the empirical hazard rate estimator is that it treats the population as homogeneous. Spell lengths differ according to potential duration of benefit payments and the characteristics of the beneficiary. To control for covariates, and to gauge the effects of the covariates on the hazard rate, the piece-wise constant proportional hazard model is estimated, which was detailed in Chapter 3, section 3. This model provides a very flexible method for estimating the baseline hazard function and avoids the imposition of a parametric functional form. The log likelihood function for the piece-wise constant proportional hazard model is given by (3) L(Y,P) = 27^ -1 { d.Jogfl-expf-expfyfkJ+z^WJ -r^expfydJ+z/tJW where kt is the observed length of the ith UI spell, z, is a vector of observable characteristics, and <5, equals one if the spell terminates before being censored and is zero if the spell is censored. In maximizing the log likelihood the vectors/? and y are parameters to be estimated. 117 In implementing this model it is necessary to censor any ongoing observations at some duration T*. For the empirical analysis, all observations lasting 50 weeks are censored at 49 and treated as right censored. Approximately 6 percent of all UI spells were ongoing at that point. A separate baseline parameter is estimated for each week of the 49 week interval. The proportional hazard model can be extended to allow for unobserved individual characteristics. Assuming that the unobserved heterogeneity takes a multiplicative form, the hazard rate is given by h,(T) = eXCO-expfz/fi) where 6t is a non-negative random variable assumed to be independent of zt. Maximum likelihood estimates of the parameter vector and baseline hazard are then obtained by conditioning the likelihood function on the dt and then integrating over the distribution of 6. With the duration model parameter estimates it is possible to derive the predicted impact of covariates on the expected duration of UI spells. The key quantity in the derivation is the predicted survivor function. The predicted survivor function for week t, for an individual with characteristics r„ is defined as: i (4) STft) = expi-Z-J, aprfW+z/Tyf]}. ; where a * denotes an estimated value. The expected duration is then given by the integral of the survivor function over V. (5) E(T) = ST(r). The expected UI spell duration for the baseline category (where z,^0) and the marginal impact of the covariates on expected duration is calculated for the samples and presented in section 4.5. 118 4.4. Data The data analysed in this chapter are a random sample of case records derived from the administration of the UI and social assistance programs in the province of British Columbia (B.C.). The UI data are from the Status Vector File of Human Resources Development Canada's UI Longitudinal File. The records in the file are a 10 percent random sample of all claims lodged inB.C. The UI data used in the analysis was restricted to the subset of regular UI claims initiated by men in B.C. between January 1988 and December 1992. Claims that were for developmental purposes or where the claimant participated in government sponsored training programs were dropped from the sample. This step was taken because it is important to control for the potential duration of UI benefits in the empirical analysis, and this variable could not be accurately constructed for this special subset of claims. Moreover, the behaviour of claimants involved in training is very different from the behaviour of UI claimants in general and merits a separate analysis. The advantages of the UI administrative data include the fine level of time aggregation (weekly) and the detailed and accurate information on weeks of benefit receipt, level of weekly benefits, weeks of insured employment and other variables used in the administration of the program. The data also cover a relatively long period of time. However, an important limitation of these data is that they provide no information on the labour market status of individuals when they are not participating in the program. Furthermore, the data contain relatively limited information on claimants' demographic characteristics. The data on welfare participation is based on case records of a random sample of individuals with a history of welfare receipt in B.C. between September 1985 and December 119 1992. Analogous to the construction of the UI file, the random sample was generated by drawing the case records for all individuals with their SIN ending in 5. The UI claims were matched with the welfare spell file using the masked SIN. In particular, each UI claim in the sample was matched against the welfare spell file to see if within a 24 month window, beginning 27 months prior to the commencement of the UI claim, the individual had recently received welfare benefits. The time window used to examine an individuals' welfare history was constructed so as to allow adequate time for a person to have been on welfare and in employment, in order to be eligible for UI. Additionally the time window ended three months prior to UI claim commencement so to avoid sampling predominantly "UI pending" welfare spells.4 These welfare spells are a product of administrative practices of the UI program rather than of individual behaviour and, in terms of the special characteristics of such welfare spells, also warrant a separate analysis. The set of UI claims is stratified into two groups, claimants with and without a recent welfare history, and the duration models were estimated separately for each group. The primary variable examined in the analysis is the number of weeks that benefits were received on a UI claim. This definition of a UI spell is the same as in Corak (1993b), who used the same data source for UI claims, and by Moffit (1985) and Katz and Meyer (1990) who used administrative data for the US. It is noted that this definition of a UI spell may not correspond to consecutive weeks of UI receipt. While the UI claim is open, an individual may not receive benefits in a given week due to full-time employment or disqualification. This definition of a UI spell is very general, aggregating the weeks of compensation paid on the UI claim. As argued by Moffit (1985), weeks of UI benefit receipt may be more informative than unemployment spells 4 The UI pending welfare spells accounted for an important part of the increase in the B.C. welfare caseload in the late 1980's (Barrett et.al. 1995). 120 when unemployment is briefly interrupted but followed by further unemployment. That is, weeks of benefit receipt may be a more appropriate measure since it aggregates periods of similar behaviour.5 An important variable in the analysis is the potential duration of benefits payable on the UI claim. This variable was constructed using the number of weeks of insured employment in the qualifying period and the regional unemployment rate and the function defining benefit entitlement as specified in the program rules. With this information combined with the number of weeks of disqualification or disentitlement, the time until exhaustion (weeks of benefit entitlement not utilised) was constructed. UI claims that ended in exhaustion (12 percent), were externally terminated or remained unterminated (8 percent) were treated as right-censored. Another variable important in the analysis of UI duration is the real level of benefits. UI benefits are a constant proportion (the benefit ratio) of insurable earnings, where insurable earnings are equal to weekly earnings up to a maximum (the maximum insurable earnings). Benefits for claimants with earnings above the maximum is given by the benefit ratio times the maximum insurable level. For the time period of the data, the benefit ratio was fixed at 0.60. Ideally, the empirical analysis would control for both an individuals previous wage6 and the benefit level. The search model outlined in section 2 predicts that the UI hazard would be positively related to the previous wage and negatively related to the UI benefit level. Unfortunately, the data do not include a measure of the previous wage apart from the benefit 5 Ham and Rea (1987) use the same data source to analyse the effects of UI parameters on unemployment spell durations. The authors define an unemployment spell as consecutive weeks of UI receipt with no employment earnings. As a result, they potentially generate multiple unemployment spells from one UI claim, and these spells are substantially shorter, on average, than the UI spells durations examined in Corak (1993) and in the present study. 6 As argued in Ham and Rea (1987), an individual's previous wage is a proxy for the mean of the wage offer distribution. 121 rate.7 Therefore the real benefit level is used in the estimation and the coefficient will reflect the net effect of the previous wage and UI benefit generosity on UI spell length. Several variables are included in the estimation to control for demand conditions. Firstly, the regional unemployment rate is included as an independent variable. The omission of the unemployment rate may result in the weeks until exhaustion variable reflecting the true effect of the UI entitlement provisions as well as local demand conditions. A set of six occupation and five industry dummy variables are included to control for differences in demand conditions across professional labour markets. Furthermore, the basic specification is altered to included a dummy variable for the period November 18 1990 to December 31 1992 corresponding to the period when entitlement provisions (and UI region boundaries) were altered8. The change in program rules coincided with the recession of April 1990 to October 1992. It is possible that the recession altered the composition of the UI population, which may cause a structural break in the hazard function over time, and as a result the coefficient on the 1990-92 dummy cannot be interpreted unambiguously as the impact of the amended UI rules. Additionally, a set of dummy variables indicating UI region are included to control for regional fixed effects. Given that the UI regions were changed in 1990, indicator variables are defined separately for the regions before and after the boundary changes were implemented; the coefficients on these variables therefore reflect the convolution of regional fixed effects and the November 1990-December 1992 time period dummy variable. 7 Individuals with wages above the maximum insurable earnings provide the only source of variation in wages and UI benefits. Unfortunately, for all claimants, only insurable earnings and the benefit rate are reported. It is possible to instrument for the wage of these worker by estimating a Tobit model using the sample of all claimants, with the censoring level of wages given by the maximum insurable level. This was the strategy for identifying separate wage and UI benefit effects adopted by Ham and Rea (1987); this method was not empirically successful and therefore is not followed here. 8 Given that the program changes introduced In Nov. 1990 generally resulted in the reduction in the maximum potential duration of benefits, for given unemployment rate and weeks of insured employment, the dummy variable is predicted to have a positive sign. 122 The only demographic information available in the data is the age of the claimant, which is controlled for by set of dummy variables. Lastly, the specifications estimated controlled for seasonal effects with the inclusion of variables indicating the quarter of the calender year in which the UI claim commenced. 4.5. Empirical Results a) Summary Statistics Summary statistics for the sample of UI spells are presented in Table 4.1. From the full sample of 38,870 spells, 87 percent were without a matched welfare spell while 13 percent did have a recent welfare history. The average UI spell was longer for the welfare history group, by 1.4 weeks at 27.8, while their potential duration of benefits was marginally less at 45.4 weeks. The welfare history group also had more weeks of benefit disqualification on average and therefore tended exit the program with fewer weeks of entitlement remaining. The welfare history group received substantially lower UI benefits on average, reflecting a lower level of insured weekly earnings. The regional unemployment rate was also marginally higher for the welfare history sample. In terms of the age distribution of claimants, the welfare history group had a substantially larger fraction in the 19-34 years age bracket (65 percent relative to 50 percent for the non-welfare history sample), which may account for part of the differential in insured earnings between the two groups. Further, the non-welfare group had a larger fraction of claimants drawn from managerial, clerical and processing occupations and a smaller fraction from service occupations. The non-welfare group was also disproportionately drawn from primary and non-market service industries while members of the welfare group were more likely from distributive and other service industries. 123 To summarise the distribution of UI spell durations, the empirical hazard rate functions for the two subsamples are plotted in Figure 4.1. The major features of the empirical hazard functions is the relatively low exit rates over the first 25 weeks of a spell, and then the large increase in the exit rate over longer durations. The non-welfare group hazard function generally lies above that for the welfare history group for spell lengths 1-24 weeks, although the two hazard function are very similar over the entire range of spell lengths. An alternative way to present the information on the duration of UI spells is by plotting the empirical survival curves, which are illustrated in Figure 4.2. The survival curves show more clearly that the welfare history sample tend to have longer UI spells. For the non-welfare history sample, approximately 50 percent of spells end within 29 weeks while for the welfare history sample 50 percent end within 32 weeks. The next step in the analysis is to take into account time until benefit exhaustion. The differences in the duration of UI spells may be due in part to differences in the length of benefit entitlements. Figure 4.3 illustrates the time until exhaustion empirical hazard function for the two subsamples. Note that these are not the same as Figure 4.1 with the time axis reversed, since there is variation in the potential duration of benefits. For example, the risk set for the estimation of the time until exhaustion hazard at 30 weeks is given by the set of claims that have an entitlement of at least 30 weeks and were terminated or right censored with more that 30 weeks of entitlement remaining. The general shape of the hazard functions in Figure 4.3 are consistent with the search model and show a marked increase in the exit rate from UI as benefit exhaustion approaches, especially in the last 10 weeks of a claim. 124 b) Duration model estimates The duration model estimates for the non-welfare history sample are presented in Table 4.2. In specification (1) the only explanatory variables are a quadratic in weeks until exhaustion. The coefficients on the weeks until exhaustion terms are highly significant. The positive signs of the coefficients indicate that the proportional effect of weeks until exhaustion on the baseline UI hazard rate is greater at short spell durations. However, the baseline hazards are very small at short durations and the greatest effect of weeks until exhaustion on the level of the UI exit rate occurs at the longer spell lengths. This effect is illustrated and discussed further in relation to duration dependence effects in subsection (c) below. Specification (2) includes controls for the regional unemployment rate, UI benefit level, beneficiaries age, occupation, industry and season of spell commencement. The inclusion of these additional variables increases the magnitude of the coefficients on the weeks until exhaustion terms, indicating the estimates in model (1) were biased toward zero by the omission of the controls for demand conditions and demographic characteristics. As expected, the unemployment rate has a very large and significant negative effect on the UI exit rate. The UI benefit level has a large negative effect on the hazard, indicating that the net impact of higher earnings and UI benefits is to lower the exit rate from UI. The coefficients on the age dummy variables generally show a positive relationship between age and the UI hazard. The coefficient are negative and significant for the 25-29 and 30-34 years age groups and the coefficients are positive but insignificant for the 45-49 and older age groups. The coefficients on the occupational dummy variables are highly significant and reveal substantial variation across occupations in the UI exit rate. Relative to the processing and machining baseline category, individuals from managerial and primary occupations have a 125 substantially higher exit rate. In regard to industry of previous employment, beneficiaries from non-market services have a significantly higher exit rate and those from distributive services have a significantly lower exit rate. Lastly, although not reported, the coefficients on the seasonal dummy variables indicate a significant seasonal pattern in the UI exit rate, with spells commenced in the fourth quarter of the calender year (the baseline category) having the highest exit rate. The next step in the estimation was to allow for a structural break in the hazard function in 1990, coinciding with the change in the UI entitlement provisions and the definition of the UI regions. Specification (3) is the same as (2) with the inclusion of a variable indicating whether the spell commenced after November 18, 1990. The coefficient on the 1990-92 dummy variable has a positive sign, as predicted. The likelihood ratio test statistic for the inclusion of the dummy variable is 5138.9, compared to a critical value of 6.64 at the one percent level of significance. Therefore, the model indicates that the program rule changes implemented in November 1990 coincided with a substantial reduction in UI spell durations; this observed change in the UI hazard effect may be due to the amended program rules or perhaps changes in the composition of the UI population resulting from the recession which began in 1990. The inclusion of the 1990-92 dummy variable generally increased the absolute size of the estimated coefficients on the other explanatory variables in the model. In particular, the magnitude of the coefficient on the linear weeks until exhaustion term, the unemployment rate and benefit level increased substantially. Further, there was an appreciable change in the coefficients on the occupation9 and industry dummy variables. Thus the effect of controlling for changes to the program rules for the November 1990-92 period on estimates for the other explanatory variables was analogous to effect of controlling for unobserved heterogeneity in 9 Although the coefficient on sales occupations became negative but insignificant 126 chapter 3. The most general specification is presented in column (4), which controls for regional fixed effects. The model therefore allows for region specific factors, such as the nature of the regional economy and labour market, which may affect the duration of UI spells. As noted above, since UI regions were changed in 1990, the set of region controls also allow for a structural break in the hazard function in November 1990. Comparison of models (2) and (4) overwhelmingly supports the presence of regional fixed effects. The null hypothesis that the regional fixed effects are jointly insignificant leads to a likelihood ratio test statistic of 7513.5, strongly rejecting the null.10 In comparison to specification (3) with the 1990-92 dummy variable, controlling for regional fixed effects leads to a further increase in the absolute size and hence significance of the coefficients on the other explanatory variables. The exception is the regional unemployment rate, which is marginally smaller in magnitude in (4). In specification (3), and (2), the coefficient on the unemployment rate is identified through variation in unemployment across regions and over time. The regional fixed effects remove the regional mean unemployment rates as a source of variation and the coefficient is only identified through variation over time. The estimates from the different specifications reveal that regions with higher unemployment tend to have longer UI spells. Even so, increases in unemployment within a region do have a substantial negative impact on UI durations. The same specifications were estimated with the welfare history sample, and the results are presented in Table 4.3. In specification (1) coefficient estimates on the quadratic in weeks until exhaustion are positive and individually significant. Additionally, the estimates are larger for The chi-squared critical value with 11 degrees of freedom at the one percent level of significance is 24.725. 127 the welfare history sample than the non-welfare history sample, indicating that the former are more responsive to the UI program parameters which define entitlement duration. Specification (2) includes controls for demand conditions, benefit level and individual's age. The coefficient on the unemployment rate is highly significant and negative and is larger in magnitude than the corresponding estimate for the non-welfare sample. Therefore UI spell durations by individuals with a recent welfare history are also more sensitive to the general labour conditions, which result from lack of employment experience and human capital or location in the secondary labour market. The estimates of the other variable in specification (2) are similar to the corresponding estimates for the non-welfare sample. In particular, the coefficients on the age dummy variables show that the hazard rate generally increases with age, though the estimates are less significant due to the smaller sample size. The estimates for the occupation and industry dummies are also very similar, although the other-services industry coefficient is substantially more negative. Specification (3) includes the 1990-92 dummy variable. The estimate of the variable is positive and highly significant. The likelihood ratio test statistic for the significance of the 1990-92 dummy variable is 682.5, strongly supporting the hypothesis of a structural change in the UI hazard in 1990. As found for the non-welfare sample, the inclusion of the 1990-92 dummy variable generally increased the absolute size of the other coefficient estimates. Again, the greatest differences were for the estimates of the impact of weeks until exhaustion, the unemployment rate and the benefit level. Indeed, the coefficient on the benefit level is much larger in size and is highly significant in (3), and implies that the recent welfare recipients tend to be more responsive to the benefit level than non-welfare recipients. Estimates of the model which controls for regional fixed effects are presented in column 128 (4). The comparison of specifications (2) and (4) strongly support the presence of region fixed effects. The likelihood ratio test statistic for the joint significance of the region controls is 982.8, which is highly significant at conventional confidence levels.11 As found for the previous sample, controlling for regional fixed effects tends to increase the absolute magnitude and statistical significance of the coefficients on the other explanatory variables. Comparing the estimates for specification (4) across the two samples, the major differences are that the welfare history sample appears to be more responsive to the program parameters and demand conditions in the regional labour market. Lastly, the estimate for the managerial occupational dummy is substantially larger for the non-welfare history sample (which has a larger fraction of claimants in this category) while the estimate for the other-service industry dummy variable is large and negative for the welfare sample (which has a larger fraction of claimants in this category). c) baseline hazard functions and expected duration The baseline hazard rate estimates for specifications (1) and (4), for both samples, are reported in Table 4.4. The estimates for specification (1) indicate positive duration dependence, especially over the 38-49 week range. Further, the estimates show that the non-welfare history hazard function lies uniformly above that for the welfare history sample. However, when the additional explanatory variables are included in the estimation, to control for demographic characteristics, demand conditions, the UI benefit level and regional fixed effects, the baseline hazard functions are very similar, with the welfare history hazard located above the non-welfare history hazard over the 45-49 week range. Therefore, after controlling for the observable characteristics of UI claims, the welfare and non-welfare history claimants exhibit almost identical 1 1 The chi-squared critical value with 11 degrees of freedom at the one percent level of significance is 24.725. 129 behaviour while on UI. To illustrate the direct duration dependence effect (the change in the hazard rate holding weeks until exhaustion constant), the baseline hazard from specification (4) are plotted in Figure 4.4. The baseline hazard functions have a similar shape to the empirical hazards plotted in Figure 4.1 although they are somewhat lower (in part due to the choice of the baseline category) and the baseline hazards are smoother since weeks until exhaustion is held constant. The total duration dependence effect is the direct change in the baseline hazard rate with spell length combined with the effect of changes in weeks until exhaustion as the spell progress. Figure 4.5 illustrates the baseline hazard rate functions when weeks until exhaustion is allowed to vary with spell length. The figure is constructed for the case of an individual with the baseline characteristics and a benefit entitlement of 50 weeks. The hazard functions for both samples are much higher than those illustrated in Figure 4.4, and the effect of the quadratic in weeks until exhaustion is to magnify the hazard rates, especially over spell lengths of28-49 weeks, with the effect greatest for the welfare history sample. With the baseline hazard estimates from specification (4) the survival probabilities and expected duration were calculated. Table 4.5 reports the expected duration for the baseline category for the two samples. The expected UI spell duration for the non-welfare sample at 17.11 weeks which is almost equal to that for the welfare sample at 17.08 weeks, reflecting the closeness of the baseline hazard functions. The expected durations clearly show individuals with a recent welfare history do not have longer UI spells, on average, relative to other claimants. Thus, there is no ostensible state dependence effect of welfare recipiency on UI spell durations. The marginal impact of the covariates on expected duration is also reported in Table 4.5. The effect of moving one week closer to exhaustion or, equivalently, of reducing the potential 130 duration of benefits by one week, is to reduce expected duration by 0.278 and 0.348 weeks for the non-welfare and welfare history sample, respectively. This is very similar to the predicted effect of 0.26 to 0.33 weeks found by Ham and Rea (1987). Further, the marginal effects of the unemployment rate and benefit level for the two samples underscore the greater sensitivity of individuals with a recent welfare history to labour market conditions and UI program parameters. The empirical analysis of UI spells presented in this chapter did not control for unobserved individual characteristics. Although a broad array of observable characteristic associated with UI claims were controlled for, it is possible that unobserved characteristics, such as individual's level of education or previous job tenure, are a determinant of UI spell length. It has been proven by Heckman and Singer (1983), and illustrated in chapter 3, that the omission of explanatory variables necessarily contributes to negative duration dependence (or less positive duration dependence) in the baseline hazard functions. Therefore, the true hazard functions may be steeper than the baseline hazard plots in Figures 4.4 and 4.5. However, it is likely that omitted characteristics which are positively correlated with UI spell durations are most concentrated among the welfare history sample, as evidenced by their welfare participation. Controlling for unobserved heterogeneity therefore would only reinforce the finding that recent welfare recipients do not remain on UI longer, ceteris paribus, than other beneficiaries. 4.6. Conclusions The research presented in this chapter examined the effect of welfare receipt on the duration of subsequent UI spells. In the context of a search model of UI unemployment spells, recent welfare recipients may be expected to have longer UI spells because of human capital atrophy, employer screening or learning effects. The raw data indicated that individuals who had 131 received welfare in the two years prior to a UI claim received UI benefits for approximately two weeks longer than other UI claimants. However, after controlling for program parameters, the unemployment rate and other demand conditions, including regional fixed effects, it was found that individuals with a recent welfare history exited UI at an almost identical rate to non-welfare history claimants. Furthermore, the exit rate for individuals with a recent welfare history was more sensitive to the UI program parameters and the state of the particular labour market in which they were located. The major policy implication of the finding that recent welfare recipients, in general, do not remain on UI longer than other beneficiaries is that they do not need to be specifically targeted with programs to encourage the transition off UI. However, this policy conclusion is subject to two caveats. Firstly, there may be particular subgroups of the welfare history population who have identifiable characteristics (such as lack of particular skills) which, although not identified in the data, are associated with lower exit rates offUI. In this case, it may be efficient to target those subgroups. Secondly, Corak (1993a, 1993b) highlighted the extensive incidence of repeat use of UI. It is possible that although recent welfare participation does not lead to longer UI spells it may contribute to UI recidivism and broader forms of program dependence. This represents an interesting topic for future research. Table 4.1. Unemployment 1 [nsurance Spell Sam pie Statistics Variable All Claims Non-Welf. S ample Welfare Sample UI Duration 26.6174 26.4405 27.8185 Potential Dur. 45.5560 45.5721 45.4466 Weeks Disentitled 0.7333 0.7074 0.9090 Weeks Until Exh. 18.2053 18.4242 16.7218 Unemploy. Rate (%) 10.5100 10.4900 10.6500 LnBenefit($100/CPI) 0.7833 0.8030 0.6500 19-24 years 0.1698 0.1593 0.2404 25-29 years 0.1840 0.1718 0.2244 30-34 years 0.1689 0.1663 0.1864 35-39 years 0.1375 0.1383 0.1326 40-44 years* 0.1125 0.1167 0.0842 45-49 years 0.0809 0.0843 0.0578 50-54 years 0.0615 0.0652 0.0362 55+ years 0.0848 0.0918 0.0380 Managerial 0.0992 0.1034 0.0706 Clerical 0.2588 0.2617 0.2394 Sales 0.0625 0.0620 0.0654 Services 0.1992 0.1895 0.2642 Primary 0.0307 0.0314 0.0262 Processing & Mach.* 0.2922 0.2963 0.2642 Transportation 0.0574 0.0556 0.0698 Notes: (1) An * denotes the baseline category. Table 4.1. Unemployment 1 nsurance Spell Sam pie Statistics Continued. Variable All Claims Non-Welf. S ample Welfare Sample Primary 0.1168 0.1211 0.0880 Construction 0.0218 0.0214 0.0250 Manufacturing* 0.0930 0.0933 0.0922 Distrib. Services 0.1804 0.1780 0.1970 Non-market Services 0.2508 0.2583 0.2000 Other Services 0.3371 0.3280 0.3988 DQuarter 1 0.2137 0.2113 0.2294 DQuarter 2 0.2165 0.2142 0.2324 DQuarter 3 0.2993 0.3034 0.2714 DQuarter 4* 0.2705 0.2710 0.2668 D1990-92 0.4227 0.4240 0.4131 Right cens. 0.2782 0.2757 0.2962 Observations 38,870 33,870 5,000 (% claims) (100%) (87.1%) (12.9%) Notes: (1) An * denotes the baseline category. 134 Table 4.2. Duration Model Estimates, Non-Welfare History Sample. Variable (1) (2) (3) (4) Wks Until Exh.(10 wks) 1.0229" 1.1164" 1.4325" 1.5641" (0.0291) (0.0263) (0.0247) (0.0246) Wks Until Exh.2 0.3778" 0.3854" 0.3737" 0.3746" (0.0069) (0.0062) (0.0059) (0.0148) Unemploy. Rate (10%) -1.6146" -2.4092" -2.2377" (0.0337) (0.0340) (0.0731) UI Benefits ($100/cpi) -0.1870 -0.2519" -0.2776" (0.0188) (0.0180) (0.0181) Age 19-24 years -0.0276 -0.0015" -0.0067" (0.0280) (0.0266) (0.0266) Age 25-29 years -0.0570" -0.0525" -0.0475" (0.0268) (0.0257) (0.0254) Age 30-34 years -0.0336" -0.0316" -0.0330" (0.0269) (0.0260) (0.0258) Age 35-39 years -0.0238 -0.0175 -0.0178" (0.0281) (0.0271) (0.0269) Age 45-49 years -0.0099 0.0150 0.0253" (.0314 (0.0303 (0.0303) Age 50-54 years 0.0315 0.0208 0.0373 (.0329 (0.0319 (0.0317) Age 55+ years 0.0278 0.0303 0.0415" (.0299 (0.0289 (0.0286) Managerial 0.3114 0.4080 0.4095 (.0623 (0.0581 (0.0619) Clerical -0.0853" 0.0642" 0.0974" (0.0203) (0.0199) (0.0201) Sales -0.1333" 0.0380" 0.0672" (0.0359) (0.0350) (0.0361) Notes:(l) Asymptotic standard errors are in parentheses. (2) * * denotes significance at the 1 percent level and * denotes significance at the 5 percent level. (3) Specifications (2), (3) and (4) included 3 seasonal dummy variables. 135 Table 4.2. Duration Model Estimates, Non-Welfare History Sample Continued. Variable (1) (2) (3) (4) Services -0.0048" -0.0231" -0.0177" (0.0236) (0.0221) (0.0212) Primary 0.5131 0.5836 0.5579 (0.0370) (0.0329 (0.0304) Transportation 0.0656 0.1382 0.1646 (.0331) (0.0317 (0.0301) Primary 0.0215 0.1480 0.1739 (.0622) (0.0581 (0.0618) Construction 0.0313 0.0071 0.0085 (.0497) (0.0483 (0.0467) Distributive Services -0.0740 -0.0935 -0.1044 (.0295) (0.0285 (0.0272) Non-market Services 0.0566 0.1130 0.1340 (.0277) (0.0271 (0.0264) Other Services -0.0064 0.0262 0.0221 (.0268) (0.0258 (0.0247) D90-92 1.0397" (0.0164) Regional Fixed Effects No No No Yes LLF -88660 -86323 -83754 -82567 Notes:(l) Asymptotic standard errors are in parentheses. (2) * * denotes significance at the 1 percent level and * denotes significance at the 5 percent level. (3) Specifications (2), (3) and (4) included 3 seasonal dummy variables. 136 Table 4.3. Duration Model Estimates, Welfare History Sample. Variable (1) (2) (3) (4) Wks Until Exh.(10 wks) 1.4571" 1.5801" 1.8303" 1.9665" (0.0778) (0.0688) (0.0662) (0.0670) Wks Until Exh.2 0.2532" 0.2604" 0.2545" 0.2526" (0.0184) (0.0162) (0.0154) (0.0158) Unemploy. Rate (10%) -1.8859" -2.5963" -2.5663" (0.0881) (0.0913) (0.1877) UI Benefits ($100/cpi) -0.0741 -0.2927" -0.3151" (0.0530) (0.0506) (0.0503) Age 19-24 years -0.0482 0.0162 0.0052 (0.0741) (0.0709) (0.0687) Age 25-29 years -0.0855 -0.0436 -0.0543 (0.0739) (0.0714) (0.0691) Age 30-34 years -0.0618 -0.0487 -0.0532 (0.0755) (0.0726) (0.0708) Age 35-39 years -0.0516 -0.0340 -0.0330 (0.0788) (0.0766) (0.0761) Age 45-49 years -0.01301 -0.0154 -0.0133 (0.0964) (0.0938) (0.0908) Age 50-54 years 0.0463 0.0410 0.0221 (0.1086) (0.1081) (0.1060) Age 55+years 0.0574 -0.0837 -0.1230 (0.1018) (0.1017) (0.1001) Managerial -0.0759 -0.0003 0.0304 (0.1508) (0.1412) (0.1368) Clerical -0.1187 -0.0010 0.0746 (0.0517) (0.0508) (0.0518) Sales -0.1238 0.0065 0.0706 (0.0879) (0.0862) (0.0891) Notes:(l) Asymptotic standard errors are in parentheses. (2) * * denotes significance at the 1 percent level and * denotes significance at the 5 percent level.(3) Specifications (2), (3) and (4) included 3 seasonal dummy variables. 137 Table 4.3. Duration Model Estimates, Welfare History Sample Continued. Variable (1) (2) (3) (4) Services -0.0853 -0.0645 -0.0241 (0.0576) (0.0551) (0.0544) Primary 0.1853 0.2044 0.2266 (0.1137) (0.1055) (0.1032) Transportation -0.0118 0.0722 0.1177 (0.0767) (0.0726) (6.0698) Primary 0.0897 0.1515 0.1678 (0.1454) (0.1350) (0.1290) Construction 0.0075 -0.0861 -0.0714 (0.1205) (0.1142) (0.1100) Distributive Services -0.0512 -0.1383 -0.1469 (0.0739) (0.0697) (0.0678) Non-market Services 0.0315 0.0483 0.0727 (0.0732) (0.0693) (0.0676) Other Services -0.1148 -0.1195 -6.1214 (0.0704) (0.0658) (0.0634) D90-92 0.9832" (0.0428) Regional Fixed Effects No No No Yes LLF -13389 -12958 -12617 -12467 Notes:(l) Asymptotic standard errors are in parentheses. (2) * * denotes significance at the 1 percent level and * denotes significance at the 5 percent level.(3) Specifications (2), (3) and (4) included 3 seasonal dummy variables. Table 4.4. Baseline Hazard Rate Estimates. Non-welfare History Sample Welfare History Sample Wk1 Specification(l) Specification )^ Specification(l) Specification(4) Hazard StdErr Hazard StdErr Hazard StdErr Hazard StdErr 18 0.0002 0.0000 0.0000 0.0000 0.0002 0.0000 0.0000 0.0000 19 0.0004 0.0000 0.0001 0.0000 0.0003 0.0000 0.0000 0.0000 20 0.0005 0.0000 0.0001 0.0000 0.0004 0.0000 0.0001 0.0000 21 0.0007 0.0000 0.0001 0.0000 0.0005 0.0000 0.0001 0.0000 22 0.0008 0.0000 0.0001 0.0000 0.0007 0.0001 0.0001 0.0000 23 0.0013 0.0001 0.0002 0.0007 0.0009 0.0001 0.0002 0.0000 24 0.0018 0.0001 0.0003 0.0000 0.0012 0.0002 0.0003 0.0000 25 0.0041 0.0002 0.0008 0.0000 0.0033 0.0004 0.0008 0.0000 26 0.0030 0.0002 0.0007 0.0000 0.0021 0.0003 0.0006 0.0001 27 0.0052 0.0002 0.0012 0.0001 0.0039 0.0005 0.0011 0.0002 28 0.0057 0.0003 0.0014 0.0001 0.0037 0.0005 0.0011 0.0002 29 0.0079 0.0003 0.0020 0.0001 0.0050 0.0006 0.0016 0.0002 30 0.0108 0.0004 0.0029 0.0002 0.0058 0.0007 0.0020 0.0003 31 0.0136 0.0005 0.0038 0.0002 0.0083 0.0009 0.0029 0.0004 32 0.0133 0.0006 0.0038 0.0002 0.0100 0.0011 0.0036 0.0005 33 0.0189 0.0008 0.0057 0.0003 0.0107 0.0012 0.0040 0.0006 34 0.0170 0.0008 0.0054 0.0003 0.0136 0.0015 0.0052 0.0007 35 0.0256 0.0010 0.0085 0.0004 0.0166 0.0017 0.0068 0.0009 36 0.0304 0.0012 0.0106 0.0005 0.0205 0.0021 0.0090 0.0012 37 0.0295 0.0013 0.0110 0.0007 0.0272 0.0027 0.0216 0.0116 38 0.0313 0.0014 0.0121 0.0008 0.0306 0.0031 0.0147 0.0019 39 0.0363 0.0016 0.0145 0.0009 0.0359 0.0036 0.0178 0.0024 40 0.0383 0.0017 0.0158 0.0009 0.0332 0.0037 0.0169 0.0024 41 0.0545 0.0023 0.0233 0.0012 0.0557 0.0054 0.0290 0.0037 42 0.0485 0.0023 0.0214 0.0012 0.0465 0.0053 0.0255 0.0036 43 0.0602 0.0027 0.0274 0.0015 0.0599 0.0065 0.0342 0.0047 44 0.0818 0.0034 0.0389 0.0021 0.0638 0.0071 0.0381 0.0055 45 0.1427 0.0044 0.0776 0.0039 0.1243 0.0102 0.0840 0.0110 46 0.1258 0.0052 0.0776 0.0041 0.1240 0.0128 0.0933 0.0125 47 0.1467 0.0063 0.0937 0.0050 0.1405 0.0153 0.1083 0.0150 48 0.1843 0.0079 0.1218 0.0065 0.1672 0.0188 0.1300 0.0185 49 0.2581 0.0107 0.1872 0.0105 0.2385 0.0257 0.2040 0.0300 (1) The hazard rate estimates and asymptotic standard errors at spell lengths 1-17 weeks for the models presented, although statistically significant, are equal to 0 to four decimal places. Table 4.5. The Expected Duration of UI Spells and the Margina Impact of Covariates Variable Non-Welf. S ample Welfare Sample Baseline 17.1146 17.0761 Wks.UntilExh.(lwk) -0.2776 -0.3475 Reg.Unemp.Rate (10%) 3.9973 4.5732 Ln Benefits (S100/CPI) 4.9449 5.5958 19-24 years 0.1199 -0.0929 25-29 years 0.8519 0.9709 30-34 years 0.5917 0.9526 35-39 years 0.3191 0.5906 45-49 years -0.4521 0.2370 50-54 years -0.6674 -0.3952 55+years -0.7415 2.2015 Managerial -7.0195 -0.5432 Clerical -1.7360 -1.3297 Sales -1.1997 -1.2583 Services 0.3162 0.4309 Primary -9.2563 -3.9861 Transportation -2.9180 -2.0912 Primary -3.0801 -2.9690 Construction -0.1519 1.2772 Distrib. Services 1.8705 2.6284 Non-market Services -2.3816 -1.2963 Other Services -0.3954 2.1733 DQuarter 1 -0.2199 -2.0508 DQuarter 2 1.3437 -0.5176 DQuarter 3 1.7655 1.3561 140 9 } D ^ p J D Z D H A } i | i q D q o J d | D A i A j n s 142 CD c o "o c Z3 Li_ TD i _ O N O X O O CL LxJ Z5 CD C LH .2 - M CO Z5 o _c X LU CD O LO LO o LO r o O r o LO CXI o CXI LO o H LO o CD o 03 CD CN O •<— O O O O O 6 6 6 CD 6 6 CD 6 6 9 1 D y p J D Z D H 144 c o Z5 Q CP C o > o LO O LO ro O n LO CNJ O CNJ LO O LO O CL 'CD o CD CD c CD CD 5 M — o co _^ CD CD O O C N 6 C N 6 O C N 6 C D b C N b oo o b o b o o b 9 } D ^ p J D Z D H Chapter 5. Conclusions 145 The research presented in this dissertation represents a first step in filling the large gap in knowledge concerning how individuals and families use social assistance in Canada. The research in this study utilises a unique longitudinal data set derived from the administration of the welfare programs in British Columbia to examine the dynamics of welfare participation and the interaction between welfare and Unemployment Insurance (UI). From the analysis in chapter 2, a number of patterns emerge from the data. Firstly, most welfare spells are shorter than six months while over 15 percent last longer than a year. Further, very few welfare cases last more than four years, and those involve families with children. Third, single mothers and fathers have longer spells than either couples (with and without children) or childless single men and women. Fourth, there have been large changes in the caseload composition: the proportion of the caseload who are employable has steadily risen from 38 percent in 1980-82 to 64 percent in 1991-92, single males have risen by 10 percentage points from 34 percent of the caseload in 1980-82, while the proportions of all other types of household have fallen. Finally, a quarter of welfare recipients are back on the welfare rolls within three months of leaving, while a full 50 percent return within a year. It is also found that for single individuals and couples without children there is a significant fraction of the population who display a seasonal pattern to their welfare use. A number of policy issues are raised by the findings presented in chapter 2. First, the vast large majority of recipients do not remain dependent on the program for an extended period of time. However the very high incidence of repeat use, especially within the first year after leaving welfare, highlights the need for governments to consider more active labour market policies 146 targeted to these individuals to help them become independent and permanently self-sufficient. Secondly, there is a subset of single parent families who do remain on the welfare rolls for several continuous years. This group accounts for an important fraction of the caseload at a point in time and for a substantial portion of the welfare budget over a period of time. It is likely these families face significant fixed costs of employment and that there are substantial disincentives to entering the labour market in terms of forgone welfare services and income. An important issue for public debate is whether it is desirable that welfare acts as a subsidy to these families or whether a more effective policy targeted specifically at this group, taking account of their special needs and characteristics, may be more efficient. The research presented in chapter 3 set out to answer two questions: (1) what is the impact of demographic characteristics and labour market conditions on the duration of welfare spells; and (2) is there evidence of state dependence in welfare receipt? With respect to the first question, it was generally found that women had a lower hazard off IA than their male counterparts. Older individuals and older principal claimants had lower hazards, whilst the converse was found for lone parent families. However, the positive relationship between lone parent's age and the hazard off IA may be picking up a positive age effect for dependent children. Dependent children were found to have a negative effect on the welfare exit rate for lone parent families but, surprisingly, a significant positive effect on the exit rate for couples. As expected, individuals or principal claimants classified as unemployable were found to have a substantially lower hazard rate off welfare. The hazard rates for all family types were found to be very sensitive to labour market conditions as measured by the unemployment rate and the help wanted index. The differential effects of these two variables across initial and repeat spells of welfare relates to the timing of their response to the business cycle and suggested one possible dimension 147 of the interaction between IA and UI. In answer to the second question, it was found that even after controlling for unobserved heterogeneity there was strong evidence of negative duration dependence in the welfare participation of all the demographic groups examined. That is, the decline in the exit rate with more time on the program is a behavioural impact of the program, due to changes in recipient's preferences or constraints such that they are more reliant on the program, rather than a statistical artefact associated with the changing composition of the sample. This finding implies that welfare, to a degree, act as a "trap" for the longer an individual remains on the program in a given spell, the less likely he/she is to "escape" the system and gain financial independence. Furthermore, after controlling for unobserved heterogeneity, there is also strong evidence of negative lagged duration dependence and negative occurrence dependence. The findings regarding occurrence and lagged duration dependence imply that the more individuals interact with the system over time, in terms of both greater number of spells and accumulated time on the program, the greater is their reliance on the program and the less is their economic independence. That is, participating in welfare appears to have a "scarring" effect on recipients' labour market careers. This suggests that it may be useful for policy makers to consider targeting particular programs that aid integration back into the workforce to recipients based on their history of welfare participation over a period of time. A number of additional policy implications follow from the findings presented in chapter 3. The differences in the impact of cyclical effects on the hazard rate for initial and repeat spells indicate one dimension of the interaction between IA and UI. This underscores the need for the government to consider the network of income security programs in designing and evaluating reforms to any one program. This is especially important when examining welfare use; since IA is the income security program of last resort, the tightening of the eligibility conditions or reducing the duration of benefits paid under other programs will inevitably impact on the IA caseload. The findings regarding the presence of particular forms of state dependence in welfare exit rates can contribute to a better understanding of the dynamics of the aggregate welfare caseload. The findings suggest that the onset of the recession was responsible for individuals and families entering welfare for the first time and then, due to the effects of state dependence at the micro level, these individuals and families were prone to longer, repeat spells on welfare. The research presented in chapter 4 examined one dimension of the UI-welfare interaction by analysing the effect of welfare receipt on the duration of subsequent UI spells. This research examined an additional form of program dependence, and tested whether welfare participation contributed to greater reliance upon the UI program. The raw data indicated that individuals who had received welfare in the two years prior to a UI claim received UI benefits for approximately two weeks longer than other UI claimants. However, after controlling for program parameters, the unemployment rate and other demand conditions, including regional fixed effects, it was found that individuals with a recent welfare history exited UI at an almost identical rate to non-welfare history claimants. Furthermore, the exit rate for individuals with a recent welfare history was more sensitive to the UI program parameters and the state of the particular labour market in which they were located. The major policy implication of the results of the analyses in chapter 4 is that welfare recipients, in general, do not need to be specifically targeted with programs to encourage the transition off UI. However, this policy conclusion is subject to two caveats. Firstly, there may be particular subgroups of the welfare history population who have identifiable characteristics (such as lack of particular skills) which, although not identified in the data, are associated with lower exit rates offUI. In this case, it may be efficective to target those subgroups. Secondly, Corak (1993a, 1993b) highlighted the extensive incidence of repeat use of UI. It is possible that although recent welfare participation does not lead to longer UI spells it may contribute to UI recidivism and broader forms of program dependence. 150 Bibliography Allen, D.W. (1993). "Welfare and the Family: The Canadian Experience", Journal of Labor Economics, v.ll(2:l), pp.s201-s223. Bane, Mary Jo and David T. Ellwood (1983). "The Dynamics of Dependence: The Routes to Self Sufficiency," Report prepared for the U.S. Department of Health and Human Services. Bane, Mary Jo and David T. Ellwood (1994). Welfare Realities: From Rhetoric to Reform, Cambridge, Harvard University Press. Barrett, G., D. Doiron, D. Green and W.C. Riddell (1995). "The Interaction of Unemployment Insurance and Social Assistance", Mimeo, Department of Economics, University of British Columbia. Blank, R.M. (1985). "The Impact of State Economic Differentials on Household Welfare and Labor Force Behavior", Journal of Public Economics, 28, 25-58. Blank, KM. (1989) "Analyzing the length of welfare spells", Journal of Public Economics, 39, 245-273. Blank, Rebecca M. and Maria J. Hanratty (1993). "Responding to Need: A Comparison of Social Safety Nets in Canada and the United States," in Small Differences that Matter: Labor Markets and Income Maintenance in Canada and the United States, eds. David Card and Richard B. Freeman, Chicago, University of Chicago Press. Bruce, R., N. Bailey, J. Cragg, E. Diewert, A. Nakamura and B. Warburton (1993)."Those Returning to Income Assistance, the UI Connection, and Training," Paper presented at the Canadian Employment Research Forum Workshop on Income Support, Ottawa. Charette, M.F. and R. Meng (1993). "The determinants of welfare participation of female heads of household in Canada", Canadian Journal of Economics, v. 27(2), pp.290-306. Corak, M. (1993a). "Unemployment Insurance Once Again: The Incidence of Repeat Participation in the Canadian U.I. Program", Canadian Public Policy, vl9(2), 162-176. Corak, M. (1993b). "Is Unemployment Insurance Addictive? Evidence from the Benefit Durations of Repeat Users", Industrial and Labor Relations Review, v.47, 63-72. Cox, D.R. (1972). "Regression models and life tables", Journal of the Royal Statistical Society B, 34, 187-202. Cox, D.R. (1975). "Partial Likelihood", Biometrika, v62(2), 269-276. 151 Cox, D.R. and D.Oakes (1984). Analysis of Survival Data, New York: Chapman and Hall. Devine, TJ. andN.M. Kiefer (1991). Empirical Labor Economics: The Search Approach, New York, Oxford University Press. Dooley, Martin D. (1994)."The Use of Social Assistance by Canadian Lone Mothers", Mimeo, Department of Economics, McMaster University. Ellwood, David T. (1986). "Targeting Would-be Long Term Recipients of AFDC," Report prepared for the U.S. Department of Health and Human Services. Fitzgerald, John (1992). "Welfare Durations and the Marriage Market: Evidence from the Survey of Income and Program Participation," Journal of Human Resources, v.27, pp. 545-561. Fortin, B., P. Lanoie and C. Laporte (1994). "Is Workers' Compensation Disguised Unemployment Insurance ?", mimeo, Laval University. Green, D.A. and W.C. Riddell (1993). "Qualifying for Unemployment Insurance: An Empirical Analysis", Department of Economics, University of British Columbia, Discussion Paper 93-33. Gritz, Mark and Thomas MaCurdy (1992). "Transitions from Welfare to Work," Report prepared for the U.S. Department of Health and Human Services. Ham, J.C. and S.A.Rea,Jr. (1987). "Unemployment Insurance and Male Unemployment Duration in Canada", Journal of Labor Economics, v.5(3), 325-353. Han, A. and JAHausman (1990). "Flexible Parametric Estimation of Duration and Competing Risk Models"', Journal of AppliedEconometrics, v.5(l), 1-28. Heckman, J.J. and G.J.Borgas (1980). "Does unemployment cause future unemployment? Definitions, questions and answers from a continuous time model of heterogeneity and state dependence", Economica, 47, 257-283. Heckman, J.J. and B.Singer (1984). "Social science duration analysis", in J.J.Heckman and B.Singer (eds.), 1985, Longitudinal Analysis of Labor Market Data, Cambridge: Cambridge University Press, 39-110. Hoffman, Saul D. (1987). "Patterns of Welfare Dependency," in Kevin R. Hopkins, ed. Welfare Dependency: Behaviour, Culture and Public Policy, prepared for the U.S. Department of Health and Human Services. Alexandria, VA: Hudson Institute. Kalbfleisch, J.D. and RX.Prentice (1980). The Statistical Analysis of Failure Time Data, New York: Wiley. 152 Katz, L.F. and B.D.Meyer (1990). "The impact of the potential duration of unemployment benefits on the duration of unemployment", Journal of Public Economics, v.41, 45-72. Kiefer, N.M. (1988). "Economic duration data and hazard functions", Journal of Economic Literature, 26, 646-679. Lancaster, T. (1990). The Econometric Analysis of Transition Data, Cambridge: Cambridge University Press. MaCurdy, Thomas (1989). "Measures of Welfare Dependency: An Evaluation," Hoover Institution Working Paper E-89-10. MaCurdy, Thomas (1989). "A Synthesis of the Behavioral Aspects of Welfare Dependency and An Assessment of the Findings," Hoover Institution Working Paper E-89-11. Manton, K.G., E.Stallard and J.W.Vaupel (1986). "Alternative Models for Heterogeneity of Mortality Risks Among the Aged", Journal of the American Statistical Association, v81(4), 635-644. Meyer, B.D. (1986) "Semiparametric estimation of hazard models", Mimeo, MIT. Meyer, B.D. (1990). "On Unemployment Insurance and Unemployment Spells", Econometrica, v58(4), pp.757-782. Moffit, R. (1983). "An Economic Model of Welfare Stigma", American Economic Review, v73(5), pp. 1023-1035. Moffit, R. (1985). "Unemployment Insurance and the Distribution of Unemployment Spells", Journal of Econometrics, 28, 85-101. Moffit, R. (1992). "Incentive Effects of the U.S. Welfare System: A Review", Journal of Economic Literature, v3 0( 1), pp. 1 -61. Mortensen, D.T. (1977). "Unemployment insurance and job search decisions", Industrial and Labor Relations Review, 30, 505-517. Murray, Charles and Deborah Laren (1986). According to Age: Longitudinal Profiles of AFDC Recipients and the Poor by Age Group. Prepared for Working Seminar on the Family and American Welfare Policy. National Council of Welfare (1987). Welfare Incomes in Canada: The Tangled Safety Net, Ottawa, Minister of Supply and Services. National Council of Welfare (1990). Welfare Incomes 1989, Ottawa, Minister of Supply and Services. 153 National Council of Welfare (1991). Welfare Incomes 1990, Ottawa, Minister of Supply and Services. National Council of Welfare (1992). Welfare Incomes 1991, Ottawa, Minister of Supply and Services. National Council of Welfare (1993a). Welfare Incomes 1992, Ottawa, Minister of Supply and Services. National Council of Welfare (1993b). Incentives and Disincentives to Work, Ottawa, Minister of Supply and Services. Olsen, R.J., A.Smith and G.Farkas (1986). "Structural and Reduced-Form Models of Choice among Alternatives in Continuous Time: Youth Employment under a Guaranteed Jobs Program", Econometrica, v54(3), pp.375-394. ONeill, June, Douglas A. Wolf, Laurie J. Bassi and Michael T. Ffannan (1984). "An Analysis of Time on Welfare." Final report to the U.S. Department of Health and Human Services. Washington, D.C.: The Urban Institute. O'Neill, J.A., L.J.Bassi and D.A.Wolf (1987). "The Duration of Welfare Spells", Review of Economics and Statistics, v.69(l), 241-248. Plotnick, Robert (1983). "Turnover in the AFDC Population: An Event History Analysis," Journal of Human Resources, v. 18, pp. 65-81. Prentice, R and L.Gloeckler (1978). "Regression analysis of grouped survival data with an application to breast cancer data", Biometrics, 34, 57-67. Walker, James R (1993). "Migration Among Low-Income Households: Helping the Witch Doctors Reach Consensus," Institute for Research on Poverty, University of Wisconsin-Madison, Discussion Paper No. 1031-94. 

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:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0087822/manifest

Comment

Related Items