UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Essays on labor and technology in asset pricing Knesl, Jiri 2019

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2019_september_knesl_jiri.pdf [ 3.24MB ]
Metadata
JSON: 24-1.0379796.json
JSON-LD: 24-1.0379796-ld.json
RDF/XML (Pretty): 24-1.0379796-rdf.xml
RDF/JSON: 24-1.0379796-rdf.json
Turtle: 24-1.0379796-turtle.txt
N-Triples: 24-1.0379796-rdf-ntriples.txt
Original Record: 24-1.0379796-source.json
Full Text
24-1.0379796-fulltext.txt
Citation
24-1.0379796.ris

Full Text

Essays on Labor and Technology inAsset PricingbyJiri KneslB.Sc., Vienna University of Economics and Business, 2009M.Sc., Vienna University of Economics and Business, 2012M.Sc., Vienna University of Economics and Business, 2013A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate and Postdoctoral Studies(Finance)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)July 2019c© Jiri Knesl, 2019   The following individuals certify that they have read, and recommend to the Faculty of Graduate and Postdoctoral Studies for acceptance, the dissertation entitled:  Essays on Labor and Technology in Asset Pricing  submitted by Jiri Knesl  in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Finance  Examining Committee: Adlai Fisher Co-supervisor Lorenzo Garlappi Co-supervisor  Murray Carlson Supervisory Committee Member C Keith Head University Examiner Giovanni Gallipoli University Examiner   Additional Supervisory Committee Members: Jan Bena Supervisory Committee Member Jesse Perla Supervisory Committee Member  iiAbstractTechnological innovations are important for economic growth but they are also a source ofvarious risks. This thesis is a collection of three self-contained essays in which I study howtechnology shocks create risk for firms and households, and affect stock prices. Overall, thethesis helps us better understand the role of labor in the transmission of these shocks to stockprices. The first essay examines the asset pricing implications of technological innovationsthat allow capital to displace labor: automation. I develop a theory in which firms withhigh share of displaceable labor are negatively exposed to such technology shocks due tocompetition that makes technology adoption appear profitable but in equilibrium erodes theexpected profits. Empirically, I develop a firm-level measure of displaceable labor share,based on detailed job classifications from the O*NET database, and find that firms with highdisplaceable labor share have negative exposure to technology shocks. A long-short portfoliobased on this new measure is highly correlated with macroeconomic measures of technologyshocks. I further show that firms with negative exposure to these technology shocks earn a4% per year return premium. At the firm level, I provide support for the hypothesis of costlyautomation following technology shocks. In the second essay, I study how investment shocksaffect different types of labor. I construct panel data sets of geographical areas, manufacturingindustries and individual workers to examine the effects of investment shocks at three differentlevels of observations. I utilize the cross-sectional variation in routine intensity of occupationsacross these three panel data sets. I show that investment shocks are an important sourceof job displacement and labor income risk. The third essay examines how a firm’s capitalintensity can affect the measurement of firm’s exposure to investment shocks by a popularmeasure, the IMC portfolio. I show that this measure suggests a considerable premium foran exposure to investment shocks when applied in a sample of capital-intensive firms butalmost no premium for the same exposure when applied in labor-intensive sample. I extenda model from previous literature by capital intensity to provide a possible explanation.iiiLay SummaryTechnological innovations are an important source of economic growth, but also a source ofrisk. This thesis is a collection of three essays in which I examine different effects of technol-ogy shocks on firms, labor, and stock prices. The first essay shows that technology shocksthat bring new or cheaper capital affect predominantly firms with a high number employeesworking in automatable occupations. These effects are reflected in the firms financial perfor-mance and their stock prices. The second essay examines the impact of technology shocks ondifferent types of labor. It shows that cheaper capital can lead to capital-labor substitutionand increase the risk of losing job or relative wage decrease. The third essay investigates howfirms use of capital vs. labor can affect the impact of technology shocks especially on theirgrowth opportunities. In summary, this thesis helps us better understand how technologyshocks create risk in the economy.ivPrefaceThis dissertation, including the formulation of the research questions, construction of theo-retical models, and empirical investigation, is an original intellectual product of the author,Jiri Knesl. The author received valuable advice and guidance from his dissertation commit-tee members on both the theoretical modeling and empirical execution components of hisresearch.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Automation and the Displacement of Labor by Capital: Asset Pricing The-ory and Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2.1 Firms, Technology, and Automation . . . . . . . . . . . . . . . . . . . 82.2.2 Households . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.2.3 Valuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.2.4 Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2.5 Implications for Firms Risk Exposure and the Dynamics of Stock Re-turns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2.6 Implications for Stock Return Premium . . . . . . . . . . . . . . . . . 162.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.3.1 Measuring Firms’ Share of Displaceable Labor, The Automation Po-tential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.3.2 Financial Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.3.3 Measuring Technology Shocks Embodied in Capital . . . . . . . . . . 202.3.4 I-shock as an Income State Variable . . . . . . . . . . . . . . . . . . . 222.4 Technology Shocks and Displaceable Labor . . . . . . . . . . . . . . . . . . . 242.4.1 Co-movement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.4.2 Firms’ Exposure to Technology Shocks . . . . . . . . . . . . . . . . . 252.5 Exposure to Technology Shocks and Return Premium . . . . . . . . . . . . . 26vi2.5.1 Return Premium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.5.2 Time Variation and Predictability of the Return Premium . . . . . . 292.5.3 Firms’ Responses to Technology Shocks . . . . . . . . . . . . . . . . . 302.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322.7 Equilibrium and Market Clearing Conditions . . . . . . . . . . . . . . . . . . 332.8 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 Job Displacement through Technology, How Technology Shocks AffectHouseholds’ Income and Employment . . . . . . . . . . . . . . . . . . . . . . 533.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.2 Data and Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.2.1 Measures of Investment Shocks . . . . . . . . . . . . . . . . . . . . . . 573.2.2 Routine Task Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . 613.2.3 Geographical Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.2.4 Individual Workers Data . . . . . . . . . . . . . . . . . . . . . . . . . 643.2.5 Manufacturing Industries . . . . . . . . . . . . . . . . . . . . . . . . . 643.3 Model and Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663.4.1 Local Labor Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . 663.4.2 Individual Workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.4.3 Manufacturing Industries . . . . . . . . . . . . . . . . . . . . . . . . . 693.4.4 Growth in Wages and Salaries . . . . . . . . . . . . . . . . . . . . . . 713.5 The Time Profile of the Estimates in Later Years . . . . . . . . . . . . . . . 723.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.7 Derivation of the Model Implications . . . . . . . . . . . . . . . . . . . . . . 763.8 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774 Capital Intensity and Investment Shocks: Implications for Stock Returns 1014.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.2 Data and Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034.2.2 βIMC Sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044.2.3 Capital Intensity and βIMC Sorting . . . . . . . . . . . . . . . . . . . 1054.3 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064.3.1 The Cash Flow of Consumption-Goods Firms . . . . . . . . . . . . . 1064.3.2 Firms’ Optimal Capital and Labor Decisions . . . . . . . . . . . . . . 1094.3.3 Valuation of Consumption-Goods Firms . . . . . . . . . . . . . . . . . 1094.3.4 Stock Returns of Consumption-Goods Firms . . . . . . . . . . . . . . 1104.3.5 Valuation of the Investment-Goods Firm . . . . . . . . . . . . . . . . 1114.3.6 Expected Excess Return of the Investment-Goods Firm . . . . . . . . 1134.3.7 Expected Excess Return on the IMC Portfolio . . . . . . . . . . . . . 1134.4 Empirical Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.4.1 βIMCf,t as a Measure of Firm’s Exposure to the Investment Shocks . . 1154.5 Model Calibration and Simulation . . . . . . . . . . . . . . . . . . . . . . . . 1184.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1194.7 Derivation of Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121vii4.8 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1245 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135AppendicesA Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140A.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140A.2 The O*NET Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141A.3 Validity and Robustness of the RTI Measure . . . . . . . . . . . . . . . . . . 142A.4 Filtering the Stock Return Data . . . . . . . . . . . . . . . . . . . . . . . . . 144A.5 Measuring the Technology Shocks Embodied inCapital, Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145A.6 Additional Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146B Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174B.1 Definition of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174B.2 A Note on Estimation with the Linear Time Trend . . . . . . . . . . . . . . 176viiiList of Tables2.1 Summary Statistics and Variance Decomposition of the AP measure . . . . . 392.2 Technology Shocks and Investment, Employment, and Income Growth . . . . 402.3 Correlation between the I-shock Measure and the Long-Short (KML) PortfoliosBased on the between- and within-industry Automation Potential . . . . . . . 412.4 Exposure of Portfolio Returns to Technology Shocks (I-shocks) . . . . . . . . 422.5 Properties of Portfolios Sorted on βKML . . . . . . . . . . . . . . . . . . . . . 432.6 Mean Portfolio Returns and Alphas of Portfolios Sorted on βKML . . . . . . 442.7 Panel Regressions of Annual Stock Returns on βKML and Firm Characteristics 452.8 Mean Portfolio Returns for Conditionally Double-Sorted Portfolios . . . . . . 462.9 Long-Horizon Predictability of the Return Premium . . . . . . . . . . . . . . 472.10 Firms’ One-Year Response to Technology Shocks . . . . . . . . . . . . . . . . 492.11 Firms’ Three-Years Response to Technology Shocks, Highly Competitive ver-sus Less Competitive Industries . . . . . . . . . . . . . . . . . . . . . . . . . . 513.1 I-shock: Summary Statistics 1960-2012 . . . . . . . . . . . . . . . . . . . . . . 803.2 Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Investment Shocks . . . . . . . . . . . . . . . . . . . . . . . 813.3 Summary Statistics: Share of Routine-Intensive Labor and Other Labor Mar-ket Variables in Metropolitan Statistical Areas . . . . . . . . . . . . . . . . . 833.4 Summary Statistics: Individual Workers Microdata . . . . . . . . . . . . . . . 833.5 Summary Statistics: Industries in the U.S. Manufacturing Sector . . . . . . . 843.6 The Immediate Impact of the IST Shock and Share of Routine-Intensive Laboron Employment and Unemployment Rate in Geographical Areas . . . . . . . 853.7 The Immediate Impact of the IST Shock on the Job Creation and Job De-struction Rate across Geographical Areas . . . . . . . . . . . . . . . . . . . . 863.8 The Impact of the IST Shock on the Probability of Job Loss for Routine andNon-routine Occupations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873.9 The Immediate Impact of the IST Shock and Routine Task Intensity on Em-ployment, Investments and Capital Intensity in the U.S. Manufacturing Indus-tries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 893.10 The Immediate Impact of the IST Shock and Share of Routine Intensive Laboron the Growth in Total and Average Wages and Salaries in Geographical Areas 913.11 Summary Statistics: Aggregate Variables . . . . . . . . . . . . . . . . . . . . 923.12 The Time Profile of the Impact of the IST Shock and Routine Intensity on theUnemployment Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 933.13 The Time Profile of the Impact of the IST Shock and Routine Intensity on theJob Creation Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95ix3.14 The Time Profile of the Impact of the IST Shock and Routine Intensity on theJob Destruction Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973.15 The Time Profile of the Impact of the IST Shock and Routine Intensity on theGrowth in the Total Wages and Salaries . . . . . . . . . . . . . . . . . . . . . 994.1 Factors: Correlation Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 1244.2 Portfolio Return Properties of All Firms in the Consumption-Goods Sec-tor Sorted by βIMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1244.3 Summary of Firms in Consumption and Investment Goods Sectors . . . 1254.4 Summary Statistics of Capital-Intensive and and Labor-Intensive Firmsin each Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1264.5 Portfolio Return Properties of Capital- and Labor-intensive Firms in theConsumption-Goods Sector Sorted by βIMC . . . . . . . . . . . . . . . . . 1274.6 Exposure of Firm-Value Components to Both Types of Shocks and to theIMC Portfolio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1284.7 Panel Regressions of Firms βIMC on Firms’ Market-to-Book Ratio, Cap-ital, and Operating Leverage . . . . . . . . . . . . . . . . . . . . . . . . . . 1284.8 Portfolio Return Properties of Capital- and Labor-intensive Firms in theConsumption-Goods Sector Sorted by βIMC+controls . . . . . . . . . . . . . 1294.9 Parameter Values for Model Calibration . . . . . . . . . . . . . . . . . . . 1314.10 Simulated Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1324.11 Simulated Stock Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132A.1 Parameters for an Illustrative Model Solution . . . . . . . . . . . . . . . . . . 140A.2 Number of Occupations with RTI measure 1998 - 2016 . . . . . . . . . . . . . 151A.3 The Underlying O*NET Variables for Constructing the Occupations’ RoutineTask Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152A.4 Summary Statistics of the Routine Task Intensity and the Underlying Tasksat the Occupational Level in Three Releases of the O*NET Database . . . . . 153A.5 Examples of Routine and Non-routine Occupations . . . . . . . . . . . . . . . 153A.6 The Underlying O*NET Variables for Constructing the Occupations’ Technology-Use Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154A.7 Ranking of Industries with Highest and Lowest AP . . . . . . . . . . . . . . . 155A.8 Share of Routine Labor and Changes in Information and CommunicationEquipment per Employee at Industry Level . . . . . . . . . . . . . . . . . . . 156A.9 Routine Task Intensity and Changes in Technology Score within Occupations,2003-2016 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157A.10 I-shock Measure: Summary Statistics 1960-2015 . . . . . . . . . . . . . . . . . 158A.11 Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Technology Shocks . . . . . . . . . . . . . . . . . . . . . . . 159A.12 Exposure of Portfolio Returns to Technology Shocks (I-shocks) . . . . . . . . 161A.13 Summary Statistics of Firm Characteristics at the Portfolio Level . . . . . . . 162A.14 Mean Portfolio Returns and Alphas of Portfolios Sorted on βKML (Constituentsof S&P500) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163A.15 Mean Portfolio Returns and Alphas of Portfolios Sorted on βKML . . . . . . 164xA.16 Panel Regressions of Annual Stock Returns on βKML and Firm Characteristics(Constituents of S&P500) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166A.17 Mean Portfolio Returns for Conditionally Double-Sorted Portfolios (Constituentsof S&P500) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167A.18 Number of Industries in the OES Industry-Specific Surveys, 1988-2016 . . . . 168A.19 The Underlying Occupational Employment Statistics and O*NET File Re-leases for the Construction of the AP measure . . . . . . . . . . . . . . . . . . 169A.20 Correlation between I-shock Measure and the Long-Short (KML) Portfolio (byDifferent Samples and Specifications) . . . . . . . . . . . . . . . . . . . . . . . 170A.21 Firms’ Three-Years Response to Technology Shocks . . . . . . . . . . . . . . . 171A.22 Firms’ One- and Three-Years Responses in Profit Margins to Technology Shocks173B.1 Definition of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174xiList of Figures2.1 Households’ Possible States. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2 Three Types of Sector Equilibria. . . . . . . . . . . . . . . . . . . . . . . . . . 342.3 Technology Shocks, Equilibrium Automation, and Profits. . . . . . . . . . . . 352.4 Firms’ Exposure to Technology Shock and Expected Returns. . . . . . . . . . 352.5 Idiosyncratic Labor Productivity  and the Risk Premium for Technology Shocks. 362.6 Technology Shocks (I-shock) and Low AP minus High AP Firms. . . . . . . . 372.7 The I-shock Measure and Following 3-Year Cumulative Return of the Low-HighβKML Portfolio. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.1 Equipment-per-Worker Indexes and Routine-Task Intensity of Employed andUnemployed Workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.2 Investment Shocks: 1980-2012 . . . . . . . . . . . . . . . . . . . . . . . . . . . 783.3 Investment Shock and the Aggregate Labor Market . . . . . . . . . . . . . . . 79A.1 Equilibrium Automation and Profits. . . . . . . . . . . . . . . . . . . . . . . . 146A.2 Technology Shocks, Equilibrium Automation, and Profits at Different Margins. 146A.3 Number of Firms and Correlation between ln(LK)and LroutineLtotal. . . . . . . . . 147A.4 RTI at Occupations Level in 2003 and 2016. . . . . . . . . . . . . . . . . . . . 148A.5 Technology Shocks (I-shock) and Low AP minus High AP Firms (by DifferentSamples of Firms). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149A.6 Correlation between Automation Potential (AP) and βKML over Time. . . . 150xiiAcknowledgementsI would like to express my extreme gratitude to my supervisors, Lorenzo Garlappi and AdlaiFisher. Lorenzo and Adlai guided me through the PhD program and spent countless hoursdiscussing my research. They encouraged me and always helped me find a way how I can makea step forward and what to focus on. I am grateful for having them around and learning fromthem. I would like to thank my committee members, Jan Bena, Murray Carlson and JessePerla. They provided me with important guidance and countless feedback. Discussing myresearch with them helped me develop my economic thinking and formulate research questionsfrom different perspectives. They asked the right questions, which made me thinking andadvance in my research. Although not on my committee, Jack Favilukis and Carolin Pfluegerwere very engaging and supportive. I learned from them and I am thankful for their help.I benefited from countless informal discussions with many faculty members in the Financedivision. They gave me not only important feedback but also helped me see broader picturebeyond the research questions in my research. I am also thankful to senior students CharlesMartinaeu, Sheng-Jun Xu, Ting Xu, and Kairong Xiao. They introduced python and clustercomputing to me and shared lots of advice even after they finished their study. I also thankmy fellow students in the PhD program. I appreciate the financial support through multipleUBC scholarships and fellowships. Sally Bei and Elaine Cho were invaluable in taking careof many administrative matters. I am very thankful to my brothers and parents. Althoughbeing far geographically, they care about my wellbeing and I am really happy to care abouttheirs. Above all, I want to thank my partner, Jisun. Thank you for going through this withme. You are the most important part of my life.xiiiDedicationTo my parents, Veˇra a Ladislav KneslovixivChapter 1IntroductionTechnological innovations affect the economy in multiple ways. Firms invest in new tech-nologies to improve their productivity and competitiveness over their competitors or to startventures that were not technologically feasible before. New technologies can also affect work-ers in many different ways. They can make some workers obsolete as new machines can carryout their duties faster, cheaper or with higher precision. New technologies can simplify theduties in many other jobs and make them accessible to workers with less specialized skills.They also create new, often highly complex, jobs and hence bring new opportunities for thosewith the necessary skills. Thus, technological innovations can create risk and opportunitiesfor both firms and households. One can easily imagine that technological innovations wouldbe reflected also in the prices of financial assets, such as stocks. One reason for this wouldbe that households face labor income risk, or opportunities, brought about by technologicalinnovations when they trade stocks of firms whose profitability and income are can be affectedby the very same innovations. My thesis is a collection of three essays at the intersection offinance and labor economics. Although each essays focuses on a different research question,they share the common objective, to gain understanding of how technology shocks embodiedin different forms of capital, shortly investment shocks, create risk in the economy.In the first essay, I examine the asset pricing implications of technological innovationsthat allow capital to displace labor: automation. I develop a theory in which firms with highshare of displaceable labor are negatively exposed to such technology shocks. In the model,firms optimally adopt technology to gain competitive advantage in the product market, butin equilibrium competition erodes profits and decreases firm value. Empirically, I developa firm-level measure of displaceable labor share, based on detailed job classifications fromthe O*NET database, and find that firms with high displaceable labor share have negativeexposure to technology shocks. A long-short portfolio sorted on this new measure is highlycorrelated with existing macroeconomic measures of technology shocks that are based onprices of capital goods. I further study the stock return premium arising from this risk expo-sure and investment, employment and financial performance of firms in different competitiveenvironment following technology shocks.In the second essay, I study how an investment shock affects different types of labor.To answer the question, I construct panel data sets of geographical areas, manufacturingindustries and individual workers. I examine the effects of investment shocks on labor marketvariable such as employment, unemployment rate, job create and destruction rate, wagegrowth and likelihood of job loss as well as the effects on investment in different types ofcapital. I show that investment shocks are an important source of job displacement and laborincome risk. This result provides empirical support for the modeling choice of householdslabor income risk in the first essay.In the third essay I focus on the role of capital intensity for measuring firms exposure toinvestment shocks using a measure suggested by previous literature, the return differences be-1tween producers of capital goods and producers of consumption goods, shortly and hereafterthe IMC portfolio. Since this measure has been applied in a framework with firms growthopportunities, I apply similar framework but extend it to include labor as a production factor.This model abstracts from the possibility of capital-labor substitution, which underlies themanifestation and highlights the importance of the investment shocks for labor in the firsttwo essays. Instead, it focuses on the importance of investment shocks for firms growth op-portunities and hence describes a different economic mechanism. I show that capital-intensivestocks sorted by the exposure to the IMC portfolio generate a highly significant average re-turn premium of up to 5% annually, but labor-intensive firms do not have such pattern. Ishow that this difference can arise from a measurement problem and potentially differentexposure of growth opportunities of capital-intensive and labor-intensive firms to investmentshocks. A similar return premium is present in the sub-sample of capital-intensive firms butabsent among labor-intensive firms, while the exposures to the IMC portfolio are similar inboth sub-samples.Although all three essays share the broad theme of technological innovations, each essaysinvestigates a different research question. For this reason, chapters were designed to beself-contained. Thus, I leave a more exhaustive discussion of the research question andcontribution to the introduction specific to each chapter.2Chapter 2Automation and the Displacementof Labor by Capital: Asset PricingTheory and Empirical Evidence2.1 IntroductionTechnological advances over the last four decades have led to adoption of many new tech-nologies.1 An important subset of these innovations, including for example robots and soft-ware, focus on automation that displaces human labor in favor of machine capital (see, e.g,Acemoglu and Restrepo (2017) and Autor et al. (2003)). The massive adoption of such tech-nologies by businesses, the associated large-scale displacement of routine-task labor, and theresulting society-wide impacts have been termed ‘’The Second Machine Age” (Brynjolfssonand McAfee, 2014). The importance of this wave of automation for the economy naturallyraises the question of how automation has affected, and continues to affect, asset prices.In this paper, I study whether firms’ share of displaceable labor that can be automated bycapital affects firms’ value and their exposure to technology shocks. The amount of routinetasks which can be automated varies across occupations and thereby across industries.2 I findthat firms with high share of displaceable labor have more negative exposure to technologyshocks that facilitate automation. I demonstrate that a long-short portfolio of firms sortedon their share of displaceable labor replicates the dynamics of price changes in capital goods,thus creating a strong connection between a macroeconomic variable and the dynamics ofstock returns. I further show that firms with a negative exposure to technology shocks earnan average return premium of 4% per year. The premium varies over time, being high inperiods of strong technological progress, and is positively predicted by decreases in pricesof capital goods. These properties of the premium are consistent with a compensation forhouseholds’ time-varying job displacement risk.To understand these findings, I develop a general equilibrium model of optimal technologyadoption with heterogeneous firms that compete in product markets, and heterogeneoushouseholds that experience uninsurable idiosyncratic labor income shocks. In the model, allfirms within a given industry have the same opportunity to adopt a technology. I show thata technology shock that facilitates automation decreases profits and market value of firmsthat can replace labor by capital. Although additional economic mechanism such as marketentry, operating leverage or inability of firms to adopt new technology can deliver a similarresult, this model points out that only a realistic assumption of product market competitionis necessary to obtain this result. The model further highlights that the price of technology1See, e.g., Jovanovic and Rousseau (2005).2See Autor and Dorn (2013). This paper provides similar results.3shocks depends on households’ idiosyncratic labor income when some of the households arereplaced by capital.In the model, a technology shock embodied in capital allows firms to use cheaper capital-based, instead of labor-based, production. Automation, which creates an optimal displace-ment of labor by capital, leads to a lower production cost but entails technology adoptioncost. In a competitive environment, a firm’s profit depends on its product price relative tothe product prices of its competitors. As all firms in a given industry have the same shareof displaceable labor, a symmetric Nash equilibrium of optimal automation exists. Becauseautomation lowers a firm’s product price relative to its competitors and increases its rev-enues, upon the occurrence of a technology shock each firm in the industry has an incentiveto automate despite the technology adoption cost. These benefits deteriorate as the numberof automated firms in the industry increases and lower product prices become the industry-wide standard. As a consequence, automation does not provide an advantage but is rathera costly necessity to stay competitive. I show that in equilibrium, technology shocks lead tolower firms’ profits. Firms with a displaceable labor share have negative exposure to suchtechnology shocks.An appealing feature of the model is its flexibility to generate both positive and negativeprice of technology shocks. The sign of the price of risk depends on the dispersion of individualworkers’ future labor productivity. Technology shocks create risk for households. The adop-tion of capital-based production reallocates workers from old to new occupations. Althoughon average all workers will be more productive in their new occupations, individual workersexperience either increase or decrease in their productivity. This labor productivity shocksare purely idiosyncratic and realized only once the worker moves to the new occupation. Sincethe occurrence of these idiosyncratic shocks is triggered by the aggregate technology shocksand markets are incomplete, the idiosyncratic risk affects the premium of the aggregate tech-nology shock. If the dispersion of the individual labor productivity shocks is low, householdsdo not fear technology shocks that facilitate automation and the technology shocks generategood states for households. If the dispersion of individual workers’ productivity increases,households dislike technology shocks as some of them may lose parts of their labor income.This result highlights the importance of imperfect risk sharing as described in Constantinidesand Duffie (1996) and its relationship to technological progress as described in Kogan et al.(2017), and Gaˆrleanu and Panageas (2017).Empirically, I measure a firm’s share of displaceable labor by a ratio of labor that can beeasily substituted by capital and the firm’s current capital stock. To construct this measure,I classify employees’ occupations by the routine intensity of their tasks using a comprehensiveset of occupational characteristics from the O*NET database and merge the occupational leveldata with firm data for employment and capital stock from Compustat.3 I find that the shareof displaceable labor is predominantly an industry-level characteristic. The variation betweenindustries explains 70% of the total cross-sectional variation in the share of displaceable labor.I use changes in quality-adjusted relative prices of capital goods as an empirical measure3O*NET database is currently the primary source of occupational information in the U.S. It provides a com-prehensive and updated set of occupational characteristics such as required skills and abilities or importance ofvarious activities and tasks. O*NET database is administered by the U.S. Department of Labor/Employmentand Training Administration through the O*NET Resource Center. The predecessor of the O*NET databaseis the Dictionary of Occupational Titles whose data are incorporated in the first releases of the O*NETdatabase.4of technology shocks.4 The theoretical foundation of this measure has been shown in previousliterature (e.g., Cummins and Violante (2002); Hornstein and Krusell (1996)) and is based onthe idea that improvements in the production of capital goods such as machines and softwareare reflected in lower prices of these goods adjusted for changes in quality. I empirically verifythis assumption using data from the NBER-CES Manufacturing Industries database and theInput-Output Accounts from the U.S. Bureau of Economic Analysis (BEA).5 Decreases inquality-adjusted prices of capital goods are indeed associated with improvements in industriesproducing capital goods.I first show that portfolios of firms sorted on their share of displaceable labor have large dif-ferences in their exposure to technology shocks. Firms with the highest share of displaceablelabor have strongly negative exposure to technology shocks. The magnitude of the exposuredecreases monotonically across the portfolios toward firms with low share of displaceable la-bor. A long-short portfolio based on this measure successfully approximates the time-seriesof technology shocks with a correlation of 0.6 and a strong overlap in all periods of largetechnology shocks. This result is present among both manufacturing and non-manufacturingfirms. I show that the differences in the risk exposure to technology shocks are also not drivenby firms’ differential exposure to other risk factors.Second, I investigate whether firms’ risk exposure to technology shocks earns a returnpremium in the cross section of stocks. I first show that technology shocks measured by thechanges in quality adjusted relative prices of capital are a state variable indicating the riskof labor income for households. Assets that co-move with this state variable are expected toearn a premium. I find that firms with strongly negative exposure earn a return premium of4% per year. The average returns monotonically decrease across portfolios toward firms withweak risk exposure. Firms’ size or growth opportunities might be related to firms’ share ofdisplaceable labor and correlate with the stock returns in the cross section. I first show thatthe Capital Asset Pricing Model (CAPM) and the size (SMB) and value (HML) factors donot explain the return premium. I next show that standard cross-sectional return predictorsdo not explain the return premium in firm-level panel regressions. I use conditional doublesorts and show that the premium is also not driven by these return predictors in a non-linearway.Third, I investigate the time variation and predictability of the return premium. If theestimated return premium is associated with labor income risk, it will follow a similar patternover time. The labor income risk is high in states of high technology shocks when the pricesof capital goods decrease. I find that the return premium is predictable by the changes inthe prices of capital goods. Periods of positive technology shocks, which decrease the pricesof capital goods, are associated with high return premium. In a companion paper (Knesl,2018), I show that decreases in the prices of capital goods increase households’ labor incomerisk and job displacement risk as cheaper and more productive capital substitutes for humanlabor. This evidence further rationalizes the return premium as a compensation of laborincome risk.Last, I examine whether firms with different exposure to technology shocks differ alsoin their behavior upon the occurrence of technology shock. Firm’s differential behavior and4See Greenwood et al. (1997, 2000); Fisher (2006).5NBER-CES Manufacturing Industries database from Becker et al. (2013) provides data on productivityand other variables at a detailed industry level. The Input-Output Accounts from the BEA provide MAKEand USE tables that allow to categorize industries into industries producing capital and consumption goods.5financial performance following technology shocks provide a valuable indication for the un-derlying economic mechanism. I find that when the relative prices of capital goods decrease,firms with strongly negative exposure decrease employment and increase capital more thanfirms with weak exposure. These firms also experience large decrease in sales, return oncapital and return on equity. This evidence directly supports the conjecture that firms’risk exposure to technology shocks reflects firms’ financial underperformance associated withchanges in production technology following a technology shock. I document that these ef-fects are stronger for firms in highly competitive industries, a finding that underlines theimportance of product market competition.In this paper, I contribute to two streams of literature. First, a recent literature hasfocused on the implications of technology shocks embodied in capital for firms’ growth op-portunities (e.g., Papanikolaou (2011); Kogan and Papanikolaou (2014); Kogan et al. (2017)),the pricing of cross-sectional return anomalies (e.g., Kogan and Papanikolaou (2013); Gar-lappi and Song (2016)), and the relationship with market power (Garlappi and Song, 2017).While the general focus of these studies is on technological advances that increase the pro-ductivity of capital with limited effects on labor, I examine technology shocks where capitalreplaces labor. I show that these technology shocks are important for firms’ adoption oflabor-replacing technologies and drive households’ job displacement risk.6 A useful benefitof my approach is the construction of a zero-cost portfolio based on firms’ displaceable laborshare that highly correlates with a macroeconomic measure of technology shocks.Second, I contribute to literature on the asset pricing when markets are incomplete, specif-ically by the inability to insure households’ job displacement and labor income risk. The assetpricing implications of uninsurable idiosyncratic labor income or consumption risk on assetprices have been addressed in several earlier studies (e.g., Constantinides and Duffie (1996);Constantinides and Ghosh (2017); Schmidt (2016)). Recent papers examine the effects oftechnological growth in the presence of imperfect risk sharing (e.g., Gaˆrleanu et al. (2012);Gaˆrleanu and Panageas (2017)). Using overlapping generations models, these papers showthat although innovations can especially benefit new generations of agents, innovations havean important displacement effect if the benefits and risk cannot be shared within or acrossgenerations of agents. I adopt this attractive feature of technological innovation creatingdisplacement risk, but I directly model job displacement as arising from capital displacinghuman workers in their occupations. Empirically, I show that the time-varying and pre-dictable return premium of technology shocks that facilitate automation is consistent withtime-varying job displacement risk that is driven by automation. This finding underlines theimportance of state-dependent displacement risk as the singular element in this stream ofliterature.A particular aspect of my paper, the riskiness of firms with displaceable labor, has beenstudied in a recent paper by Zhang (2018). He suggests that routine labor lowers firmsriskiness by providing a real option hedge. In his empirical analysis, firms with a higherroutine labor share relative to industry average have lower expected returns. I arrive ata different conclusion that is supported both by my theory and empirical evidence. Thetheoretical difference arises due to three main factors. First, differently from Zhang’s partial6An important aspect of my model is technology adoption, which has been studied in different settings ine.g., Pastor and Veronesi (2009), Atkeson and Kehoe (2007), Greenwood and Jovanovic (1999), and Hobijnand Jovanovic (2001).6equilibrium setting that focuses exclusively on a single firm’s riskiness, I develop a generalequilibrium model to examine how the process of automation is priced on capital markets bycreating risk for both firms and households. This highlights the link between automation andjob displacement risk that is important to understand the pricing of automation on capitalmarkets and the time-series properties of the return premium. Second, my model includesproduct market competition, whose relevance is supported empirically but not considered inZhang’s model. Third, I explicitly show that firms’ decision to displace labor by capital isdirectly related to prices of capital goods. I support all these features by empirical evidence.To reconcile the empirical differences between our papers, one has to consider that I usefirms’ share of displaceable labor relative to capital, a variable that varies primarily betweenindustries, while Zhang examines a firm’s deviation in routine labor share from an industryaverage.7 To the best my knowledge, I am the first to show how the automation process ispriced on capital markets by affecting both firms and households. The key part of this is todocument the strong co-movement between changes in prices of capital goods and returns ofstocks sorted by the share of displaceable labor.This paper relates to several other streams of literature. Many of these papers study therelationship between asset prices and different aspects of labor. Wage rigidities and operatingleverage have been addressed in Chen et al. (2011), and Favilukis and Lin (2015, 2016).Organization capital stemming from the key talent and labor mobility have been studied inEisfeldt and Papanikolaou (2013) and Donangelo (2014), respectively. Labor market frictionsare examined in Belo et al. (2014, 2017), and Kuehn et al. (2017). Other studies examineindustry competition (e.g., Hou and Robinson (2006); Loualiche et al. (2014); Corhay (2017);Bustamante and Donangelo (2017), and Corhay et al. (2017)) and offshoring (e.g., Bretscher(2017)). The potential of capital to substitute some types of labor and complement others isat the center of many economic models that study the effects of technological change (e.g.,Acemoglu and Autor (2011); Acemoglu and Restrepo (2018)) and more broadly (e.g., Zeira(1998); Krusell et al. (2000)).The rest of the paper is organized as follows. In the next section, I develop a general equi-librium model of an optimal technology adoption that highlights how automation affects stockprices. Section 2.3 provides the data sources, construction and validity of the measures ofdisplaceable labor share and technology shocks. I examine the exposure to technology shocksof firms with different share of displaceable labor in section 2.4. I describe the zero-costportfolio that approximates the empirical measure of technology shocks in this section. Sec-tion 2.5 documents the cross-sectional return premium and examines its properties includingtime-series predictability. This section also shows that firms’ exposure to technology shocksis associated with differential investment and employment policies. I conclude in section 2.6.2.2 The ModelI develop a two-period general equilibrium model that provides a clear picture of how technol-ogy shocks affect asset prices through stimulating labor displacement and technology adop-tion. The model consists of two sectors with firms competing in monopolistic fashion and7The percentage of the between-industry variation in the firms’ share of displaceable labor depends on theparticular industrial classification. I show in table 2.1 that 71.7% of variation in this variable is due to thebetween-industry variation when industries are defined at SIC 3-digit and NAICS 4-digit level.7households supplying sector specific labor. Technology shocks decrease the price of capitaland motivate firms to automate by adopting capital-based production. The model showsfirms’ optimal decision to adopt the capital-based production and its effects on the compo-sition and risk exposure of firms within a given sector as well as the effect on the income ofhouseholds supplying labor to the sector.2.2.1 Firms, Technology, and AutomationThe economy exists for two periods. The production side of the economy consists of twosectors s ∈ {1, 2}. Sector s = 1 initially consists of labor-based firms and sector s = 2 ofcapital-based firms. Firms in sector s = 1 have an opportunity to automate by adoptingcapital-based production upon the occurrence of a technology shock in the second period asdescribed in more detail in below.Within each sector s, there are N firms that produce firm-specific differentiated goodsYf,s,t at time t and compete in monopolistic fashion. The number of firms is large enough sothat each firm can abstract from the consequences of its choices for the sector level aggregates.Firms produce their goods using either a labor-based, Y lf,s,t = Lf,s,t, or a capital-based linearproduction function, Y kf,s,t = Kf,s,t. A firm with capital-based production buys sector-specificcapital Kf,s,t at price PIs,t. I assume that firm’s capital is active and fully depreciates in theperiod of investment. The cash flow of a capital-based firm isCF kf,s,t = Pkf,s,tYkf,s,t − P Is,tKf,s,t, (2.1)where P kf,s,t denotes the price of firm-specific good produced in a capital-based production.A firm with labor-based production hires sector-specific labor Lf,s,t at sector-specific wageWs,t. Labor-based firms have an option to automate by adopting capital-based productionand paying automation cost κ. Firms will optimally do so when the price of capital investmentP Is,t becomes low enough. The cash flow of a firm with labor-based production isCF lf,s,t = Plf,s,tYlf,s,t −Ws,tLf,s,t if not automatedCF kf,s,t = Pkf,s,tYkf,s,t − P Is,tKf,s,t − κ if automated,(2.2)where P lf,s,t and Pkf,s,t are the prices of the firm-specific good.Sector-specific capital, Ks,t, is produced completely competitively in a linear productionfunction with an aggregate technology level At and sector-specific labor Lks,t that is employedin producing capital:Ks,t = AtLks,t. (2.3)I further assume that installation of capital is subject to decreasing returns to scale at thetotal amount of sector-specific capital as described in equation 2.4. The assumption of strictlydecreasing returns to scale 0 < α < 1 is important for the existence of a mixed-strategy8equilibrium described below, but not for the final implications for stock returns.8Is,t = Kαs,t where Is,t =N∑f=1Kf,s,t (2.4)Hence capital production and installation imply that individual firms face the following priceof capital:P Is,t =1AtWs,tI1−ααs,t . (2.5)Equation 2.5 shows that an increase in technology At decreases the effective price of capitalinvestment. This effect is partially offset by higher sector-level investment if α < 1. Thetechnology At and the total amount of sector-specific investment determine the marginalcost of capital-based production relative to the marginal cost of labor-based production.Automation becomes favorable when the marginal cost of capital-based production decreasesrelative to the marginal cost of labor-based production.In the first period, sector s = 1 consists of firms that use labor-based production and havethe option to automate. Sector s = 2 consists entirely of firms that already use capital-basedproduction. The level of technology At is subject to a technology shock that is realized atthe beginning of the second period. A technology shock increases the level of technologyAt. Specifically, I assume an initial level of technology ALow that can increase to AHigh withprobability pH as described in 2.6.A1 = ALow, A2 ={ALow, with probability 1− pHAHigh, with probability pH(2.6)A technology shock ALow → AHigh decreases the price of capital investment P Is,2 in bothsectors s = 1, 2. The assumption of constant or decreasing prices of capital investment issupported by the empirical evidence as the quality adjusted relative prices of capital goodsexperienced a secular decrease over the last four decades (see, Fisher (2006); Cummins andViolante (2002)).9The firm-specific goods are combined to sector-specific goods using a constant elasticityof substitution (CES) function:Ys,t = 2∑f=1Yν−1νf,s,t νν−1 , (2.7)where ν is the elasticity of substitution between the firm-specific goods. I assume ν > 1which implies that the firm-specific goods are imperfect substitutes to each other. The prices8The assumption of decreasing returns to scale for installation of capital at the sector level implies thatindividual firms do not internalize the effect of their capital investment on the marginal cost of capital invest-ment. It further implies that the amount of capital used in a sector is always equal or greater than the sumof capital investments of individual firms. This assumption is equivalent to adjustment or installation cost ofcapital at the sector level.9I document an average decrease in quality adjusted relative prices of capital goods of 4 percent per yearsince 1960 (see, table A.10)9of the sector-specific goods can be expressed asPs,t = N∑f=1P 1−νf,s,t 11−ν . (2.8)The sector goods are combined to an aggregate final consumption good using a CES functionwith a different elasticity of substitution, ρ10YC,t =(2∑s=1Yρ−1ρs,t) ρρ−1. (2.9)I choose the aggregate consumption goods as a numeraire by setting its price PC,t ≡ 1, andassume that ν−1ν >ρ−1ρ . This assumption implies that firm-specific goods within a givensector are closer substitutes than sector-specific goods. For example, individual products infood industry substitute each other more readily than all food products substitute productsof a different industry such as paper mills.Firms within each sector compete in a monopolistic fashion. Given the production type,firms maximize dividends by setting the price of their goods:Dlf,s,t = maxP lf,s,t{P lf,s,tY lf,s,t −Ws,tLf,s,t} where Y lf,s,t =(P lf,s,tPs,t)−νYs,t (2.10)Dkf,s,t = maxPkf,s,t{P kf,s,tKf,s,t − P Is,tKf,s,t} where Y kf,s,t =(P kf,s,tPs,t)−νYs,t, (2.11)where Y lf,s,t and Ykf,s,t are the demand functions for the firms’ goods. The first order conditionsof firms’ optimization in 2.10 and 2.11 imply that firms charge constant mark-ups over themarginal production cost:P lf,s,t =ν1− νWs,t and Pkf,s,t =ν1− ν PIs,t. (2.12)Upon the realization of the technology shocks at the beginning of t = 2, firms in thelabor-based sector s = 1 decide whether to automate the production. Automation changesthe labor-based production to capital-based production by incurring automation cost κ:Df,1,2 = max{Dlf,1,2, Dkf,1,2 − κ} (2.13)Automation changes the firm composition as well as the use of labor within the sectorwithin the sector s = 1. δ =∑Nf=1 1{Automate=Y es}fN denotes the percentage of firms in sectors = 1 that decide to automate. Automation shifts labor from the direct production of firms’goods to the production of the sector-specific capital. Ll1,2 =∑Nf=1 Lf,1,21{Automate =No}f is sector 1′s labor that is employed directly in labor-based firms.Lk1,2 =1A2(∑Nf=1Kf,1,21{Automate = Y es}f) 1αis sector 1′s labor that is employed in theproduction of the sector-specific capital. The total amount of labor in sector s = 1 is L1,2 =Ll1,2 + Lk1,2.10An equivalent aggregation can be achieved when the sector-specific goods are consumed directly by house-holds with CES preferences and elasticity of substitution ρ.102.2.2 HouseholdsThere is a large number, Nh, of ex-ante identical households. Each household consists of twospouses, each of which supplies labor effort to a different sector. Households derive utilityfrom consumption ch,t and disutility from labor effort Hh,1,t and Hh,2,t according to a utilityfunction in the spirit of Greenwood et al. (1988):U (ch, Hh,1, Hh,2) = E2∑τ=tβτ−t(ch,τ − χH1+θh,1,τ1+θ − χH1+θh,2,τ1+θ)1−γ1− γ , (2.14)where β is households’ subjective discount factor, χ governs the disutility of labor effort, γis the coefficient of risk aversion, and 1θ is the Frisch elasticity of labor supply.Now I introduce an idiosyncratic labor productivity risk. The labor effort Hh,s,t trans-forms to Lh,s,t = Hh,s,t effective hours of labor unless the spouse shifts from working di-rectly in a labor-based firm to working in the production of sector-specific capital. If thespouse has to move to the production of sector-specific capital, the labor effort transforms toLh,1,2 = (1 + ˜h)Hh,1,2 effective hours of labor. ˜h is the spouses’ idiosyncratic random vari-able that determines their labor productivity when they move to the new job. It is importantto mention that the job movements are purely intra-sectoral. When firms in sector s = 1automate by adopting capital-based production, a corresponding number of spouses will bedisplaced from these firms and employed in the production of sector-specific capital withinthe same sector s = 1. The spouses cannot observe their idiosyncratic labor productivity ˜huntil they are employed in the new job. I specify the distribution of the idiosyncratic laborproductivity by:˜h ={− with probability p+ with probability 1− pwhere  > 0. (2.15)It is important to emphasize that markets are incomplete in the sense that households cannottrade an asset that would insure them against the realization of the idiosyncratic laborproductivity. This setup is analogous to Constantinides and Duffie (1996).To maximize their utility, each household solves following problem:maxch,t,Hh,1,t,Hh,2,tU (ch,t, Hh,1,t, Hh,2,t) s.t. ch,t = Dh,t + Lh,1,tW1,t + Lh,2,tW2,t, (2.16)where Dh,t =1Nh(∑Nf=1Df,1,t +∑Nf=1Df,2,t)is the households capital income. The specificform of the utility function 2.14 ensures that households do not take the wealth effect of theidiosyncratic labor productivity shock ˜h into account when they decide the optimal laborsupply. The first order condition determines the optimal labor effort as:Hh,s,t =(Ws,tχ) 1θs ∈ {1, 2}. (2.17)The total supply of labor to sector s at time t is Ls,t =∑Nhh=1 Lh,s,t. The symmetry ofthe distribution of ˜h ensures that the labor effort of spouses working in capital produc-tion in sector s = 1 at t = 2 is transformed to effective hours on average in the same11way as the labor effort of spouses working directly in labor-based firms in the same sec-tor, E[Lh,1,2] = E[(1 + ˜h)Hh,1,2] = Hh,1,2.The utility optimization gives a household-specific stochastic discount factor of the form:Λh = βch,2 − χH1+θh,1,21+θ − χH1+θh,2,21+θch,1 − χH1+θh,1,11+θ − χH1+θh,2,11+θ−γ , (2.18)Since all households are identical ex ante, the stochastic discount factors are identical acrossall households. Each single household prices assets in the same way. At the same time,aggregate consumption and labor efforts are not sufficient statistics for pricing firms as theyomit the idiosyncratic component of households consumption due to the realization of theidiosyncratic labor productivity shock.2.2.3 ValuationThe value of firms in labor-based sector s = 1 at t = 1 is given by:V lf,s,1 = Dlf,s,1 + E [ΛhDf,s,2] , (2.19)The value of firms in the capital-based sector s = 2 is defined equivalently as:V kf,s,1 = Dkf,s,1 + E[ΛhDkf,s,2]. (2.20)The expectation operator in the valuation equations 2.19 and 2.20 is defined from householdh′s perspective over both aggregate and idiosyncratic states. The realization of all states andthe corresponding probabilities are depicted in figure 2.1. It is important to note that thestochastic discount factor Λh takes a different value in each possible realization including therealization of idiosyncratic labor productivity risk ˜h. Hence the households’ idiosyncraticrisk affects the valuation of assets. The idiosyncratic risk is relevant only when AHigh isrealized. The importance of the idiosyncratic risk is then given by the percentage of firmsthat automate δ∗. Hence the idiosyncratic risk is state dependent as in Constantinides andGhosh (2017).12Figure 2.1: Households’ Possible States.A21 − pH pHALow AHighFirm compositionin sector 11 − δ∗ δ∗Not automated Automated˜h1 − p p− +2.2.4 EquilibriaThe equilibrium is defined by the solution of the households’ optimization problem 2.16,firms’ optimization 2.10, 2.11, and 2.13 and market clearing conditions on all markets. Themarket clearing conditions for product markets at the level of firms and sectors as well as forcapital and labor markets are stated in section 2.7.In the second period, t = 2, there is a further condition for a Nash-type equilibrium forthe equilibrium percentage δ∗ of firms that automate within sector s = 1. Each firm mustprefer its choice of the production type given the choices of the remaining firms in the sector.There is a unique equilibrium of one of the following three types.Equilibrium 1. None of the firms in the sector prefers capital-based production, δ∗ = 0.The sector remains homogeneous with labor-based firms only. In this equilibrium, inequal-ity 2.21 a holds. The left-hand side shows the payoff when the firm switches to capital-basedproduction. This payoff consists of dividends and an automation cost. The right-hand sideis the dividend of a firm that keeps labor-based production. If the price of capital is highenough, no firm will have an incentive to automate as it leads to lower sales and requiresautomation cost κ.11This equilibrium is shown in the first column (no shock) in figure 2.2. Panel A shows themarket-clearing prices of goods produced by each type of firm, labor-based and automated,as a function of the percentage of firms that automate δ. The blue line depicts the pricesof goods produced by labor-based firms and the red line depicts the equivalent functionfor automated firms. I select the model parameters so that the lines intersect at δ = 0if the technology A2 remains constant at ALow. The goods prices of automated firms arepredominantly increasing in δ. The increase reflects the decreasing returns to scale of the11This reasoning abstracts from the effects on the sector level variables.13capital installation function 2.4.a. νP kf,1,2(P kf,1,2P1,2 (δ∗))−νY1,2 (δ∗)− κ ≤ νP lf,1,2(P lf,1,2P1,2 (δ∗))−νY1,2 (δ∗) , δ∗ = 0b. νP kf,1,2(P kf,1,2P1,2 (δ∗))−νY1,2 (δ∗)− κ = P lf,1,2(P lf,1,2P1,2 (δ∗))−νY1,2 (δ∗) , 0 < δ∗ < 1c. νP kf,1,2(P kf,1,2P1,2 (δ∗))−νY1,2 (δ∗)− κ ≥ νP lf,1,2(P lf,1,2P1,2 (δ∗))−νY1,2 (δ∗) , δ∗ = 1,whered. P1,2 (δ∗) =(P kf,1,21−νδ∗N + P lf,1,21−ν(1− δ∗)N) 11−νe. Y1,2 (δ∗) =(Y kf,1,2ν−1ν δ∗N + Y lf,1,2ν−1ν (1− δ∗)N) νν−1(2.21)Panel B shows the corresponding demands for the goods of each firm type as functions ofδ. The demand for goods of labor-based firms is above the demand for goods of automatedfirms for δ > 0 as the prices of labor-based firms’ goods are lower in this region. This followsfrom the demand functions in equations 2.10 and 2.11.Panel C shows the dividends of each type of firm as functions of δ. The dividends oflabor-based firms are above the dividends before automation cost of automated firms forδ > 0. This follows from the price elasticity of the demand functions for firm-specific goods−ν < −1. The dividends of automated firms are even smaller when the automation cost κ istaken into account. Hence the equilibrium percentage of automated firms in sector s = 1 isδ∗ = 0.Equilibrium 2. A percentage 0 < δ∗ < 1 of firms in the sector automates and paysautomation cost. This equilibrium can be thought of as a mixed-strategy equilibrium wherefirms choose an optimal probability δ∗ to automate and probability 1− δ∗ to keep the labor-based production. The equilibrium is defined by the equality 2.21 b. The left-hand side showsthe payoff of a firm that automates. The right-hand side is the dividend of a firm that keepslabor-based production. Abstracting from the effects at the sector level, the equality holdswhen the gain from higher sales due to lower goods prices(Pkf,1,2P1,2(δ∗))−νexactly counterweightsthe automation cost κ. Firms that adopt capital-based production extract higher rents atthe cost of firms that keep labor-based production.The middle column of figure 2.2 shows this type of equilibria. I assume an incrementalimprovement in the technology level AHigh. In panel A, the prices of goods from automatedfirms shift downward and intersect the price function of the labor-based firms. Panel B showsthat the changes in prices lead to a higher demand for goods of automated firms relative tolabor-based firms if the percentage of automated firms is low enough. Panel C shows theequilibrium percentage of firms δ∗ in sector s = 1 that optimally automate production. Inthe region δ < δ∗, labor-based firms have an incentive to automate as they will benefit fromhigher demand. In the region δ∗ < δ < δˆ, firms do not want to automate as the slightlyhigher demand for their products does not compensate the automation cost κ. In the region14δ∗ < δ, the demand for goods of automated firms is already below the demand for goodsof labor-based firms. Firms are indifferent between keeping the labor-based production andautomation only at δ∗.Equilibrium 3. All firms in the sector automate, δ∗ = 1. The sector is homogeneouswith capital-based firms only, and inequality 2.21 c holds. This equilibrium arises when theprice of capital P I1,2 is low enough. In this case, a firm will experience dramatically lowersales when it keeps the labor-based production given all other firms have automated. Henceall firms optimally adopt the capital-based production despite the automation cost to avoidmore severe punishment through lower sales. Abstracting from the effects at the industrylevel, all firms keep the same rent but pay the automation cost κ.This equilibrium is shown in the right-hand side column in figure 2.2. I assume a largeincrease in the technology level AHigh. In this case the prices of goods of automated firms arealways below the prices of labor-based firms. This is reflected in the demand function of theautomated firms that is always above the demand function of the labor-based firms. If theimprovement in the technology level A2 is large enough, the net dividend of automated firmswill be above the dividend of labor-based firms regardless of the percentage of automatedfirms in the sector. The equilibrium percentage of automated firms is δ∗ = 1. Even in thiscase the net dividend can be below the original level. All firms in the sector optimally decideto automate even if automation leads to a lower net dividend. Firms which had stuck with thelabor-based production would have experienced much lower revenues that is more penalizingthan the automation cost κ. This situation highlights that adopting a relatively cheaperproduction technology can become a pure necessity to stay competitive and the automationbenefits are passed to the final consumers in the form of lower product prices.Panel A in figure 2.3 presents the relationship between the magnitude of the technologyshock AHigh, conditional on the occurrence of this state, and the equilibrium percentage ofautomated firms δ∗. This relationship is strictly positive. A higher level of technology AHighdecreases the prices of capital investment P I1,2 and makes capital-based production morefavorable despite a higher amount of invested capital I1,2. I select the model parameters suchthat the first type of equilibria, δ∗ = 0, exists when A2 = ALow < 0.85. The second type ofequilibria, 0 < δ∗ < 1 emerges when A2 = AHigh and 0.85 < AHigh < 1.06. The third type ofequilibria, δ∗ = 1 establishes if the technological change is large enough, A2 = AHigh > 1.06.2.2.5 Implications for Firms Risk Exposure and the Dynamics of StockReturnsIn this section I show that a positive technology shock that facilitates automation can lead toan adverse outcome for firms in a sector that is undergoing at least partial automation. Theadverse effect of such a shock is equally present among all firms within the sector regardlessof whether the particular firm automates or not.Panel B of figure 2.3 shows the equilibrium dividends of firms in sector s = 1 as a functionof the magnitude of the technology shock AHigh, conditional on realization of this state. Inequilibrium, the dividends are identical for all firms in the sector regardless of whether theparticular firm automates or not. An increase in the technology shock leads to a lowerdividend of every firm in this sector. This relationship is weak for low levels of AHigh butbecomes stronger for higher values of the technology. This reflects the convex relationship15between the equilibrium percentage of automated firms δ∗ and AHigh shown in panel A.12A larger technology shock leads to a higher percentage of firms that automate δ∗ andcharge lower product prices. This decreases the price of the sector-aggregate good P1,2 (δ∗)as shown in formula 2.21 d. Since firm’s relative revenues depend on the price of its goodrelative to the sector-aggregate good(Pf,1,2P1,2(δ∗))−νwith ν > 1, it becomes more prohibitive tokeep the labor-based production that produces goods at higher relative prices. At the sametime, the rents of automated firms decrease as the prices of their goods and the price of thesector-aggregate good come closer to each other. Automation becomes a costly necessity tostay competitive.13Panel C shows the effect of technology shocks on dividends of firms in sector s = 2. Thisrelationship is strictly positive. Since this sector already consists of firms with capital-basedproduction, firms do not undergo the costly process of automation. Hence firms in sectors = 2 benefit directly from lower prices of capital investment P I2,2 brought by a higher levelof technology.In this model, the effect of technology shock on dividends translates directly to firms’ riskexposure to the shock. I plot the risk exposure of firms in each sector as functions of AHigh inpanel A of figure 2.4. Firms in sector s = 1 have negative exposure to technology shocks. Themagnitude of the exposure is low for low values of AHigh but increases dramatically. Hencea technology shock decreases the firm value of firms in this sector. These firms experiencenegative return on the impact of technology shock. Firms in sector s = 2, which consistsentirely of capital-based firms, have positive exposure and experience positive stock returnupon the occurrence of technology shock.The results in this section show that competition forces firms to undergo costly automationprocess although its benefits are passed through to the final consumers. Firms with laborthat can be automated by capital then exhibit negative risk exposure to technology progressthat facilitates such automation.2.2.6 Implications for Stock Return PremiumIn this section, I show that the differences in firms’ exposure to technology shocks are asso-ciated with differences in expected stock returns as the risk of technology shocks is priced bythe households.First, I emphasize that the model is able to general both negative and positive premium forthe risk of technology shock that facilitate automation. The sign of the premium is determinedby the dispersion of the idiosyncratic labor productivity shock  the households face uponthe occurrence of technology shock and the associated possible job movement. Figure 2.5shows the relationship between the dispersion of the idiosyncratic labor productivity  andthe risk premium of an asset E[RI]−Rf that has β = 1 exposure to the technology shock A2.Households require a positive premium for this asset if the idiosyncratic labor productivityrisk  is low. In this case, the higher average labor productivity of spouses working inthe capital production due to AHigh dominates the idiosyncratic labor productivity risk.12I plot the firms’ dividends as a function of the equilibrium percentage of automated firms δ∗ in appendixfigure A.1. This relationship becomes generally linear.13This statement abstracts from inter-sectoral demand shifts. The inter-sectoral demand shifts are consid-erably smaller as the sector-level goods are weaker substitutes.16As  increases the idiosyncratic labor productivity risk dominates and the return premiumbecomes negative. In this case, households dislike the occurrence of technology shock as theymay experience a severe decrease in their labor income. The same asset becomes valuable asit outperforms upon the occurrence of the technology shock.For the following analysis, I assume the case of negative premium. This assumptionwill align with the empirical evidence in later sections. The relativity of the return pre-mium described in figure 2.5, however, highlights that the implications of technology shockfor expected stock return depend on the particular economic environment, in this case thehouseholds’ ability to stay productive in new occupations.Panel B in figure 2.4 shows the expected stock returns of firms in each sector as functionsof the magnitude of the technology improvement AHigh. Firms in sector s = 1 earn higherexpected return than firms in sector s = 2. This follows directly from the firms’ exposure andthe assumption of negative risk premium. Firms in sector s = 1 have a negative exposure(see panel A) to a negatively priced risk which results in high expected returns. The positiveexposure of firms in sector s = 2 to the same risk leads to lower expected returns.The difference in the expected returns between firms in sector s = 1 and s = 2 increaseswith the magnitude of AHigh as shown in panel C. This follows from the changes in theexposure of firms in sector s = 1 as shown in panel A. The exposure of firms in sector s = 1becomes more negative as AHigh increases, while the exposure of firms in sector s = 2 remainsalmost constant. The increasing differences in the exposure of firms in each sector lead tosimilarly increasing differences in expected return.2.3 Data2.3.1 Measuring Firms’ Share of Displaceable Labor, The AutomationPotentialThe main input in my analysis is a measure of firms’ share of displaceable labor, the AutomationPotential (AP). I use the occupational characteristics from the O*Net database, the com-position of labor force from Occupational Employment Statistics (OES) and the relativeimportance of labor and capital to calculate the AP measure at the firm level. The APmeasures the relative importance of labor that can be easily replaced by capital.I measure firm j′s AP at time t as:APj,t = ln(Lroutinej,tKj,t)(2.22)The numerator of the AP formula, Lroutinej,t measures the number of employees in routineintensive occupations, i.e. labor that can potentially be substituted by capital. This numberis a combination of firm-level and industry-level data, where the number of employees isdefined at firm level and the percentage of employees in routine occupations is imposed oneach firm from industry-level OES. To assess the relative importance of routine labor in theproduction process, I divide the labor by firms’ capital.14 I carefully examine a range of14The term production process is used widely in this context. It refers not only to physical production butalso to all operations of a given firm including firms in service sectors that do not necessarily produce physicalproducts.17alternative specifications of the AP formula and discuss them in detail below.I measure the potential of replacing human workers by capital in each occupation by theRoutine Task Intensity measure (RTI). This measure was originally developed by Autor andDorn (2013) and measures the importance of routine, abstract and manual tasks in eachoccupation. High RTI scores indicate high importance of routine tasks in a given occupation.Routine tasks are defined as tasks that can be accomplished by following a precise set of rulesand potentially be executed by machines programmed to follow this set of rules.15 Low RTIscores indicate low importance of routine tasks and high importance of abstract and manualtasks, which cannot easily be performed by machines.The original measure was calculated for 330 occupations based on data from the Dic-tionary of Occupational Titles (DOT) 1977. I develop a new and more detailed version ofthe RTI measure based on occupational characteristics from the majority of the O*NETdatabase releases. I describe the structure and the data collection procedure of the O*NETdatabase in appendix A.2. This approach significantly improves the RTI measure in severaldimensions. First, I use multiple releases of the O*NET database, which allows the newRTI measure to capture the changes in the occupational characteristics within occupationsover time. Technological progress and automation are likely to change the nature of someoccupations over time. Autor et al. (2003) show that occupations that increased the use ofcomputers significantly reduced the input of routine tasks and became more non-routine.16This evidence stresses the importance to account for changes in occupational characteristicsover time. Second, I calculate the RTI measure for a larger number and more detailed occu-pations. Depending on the release of the O*NET database, I calculate the RTI measure for654-731 occupations.17 Third, I use larger set of underlying occupational characteristics (30instead of 5) to measure the importance of routine, abstract and manual task inputs. Thewide range of the characteristics provides a robust measurement of the importance of eachtype of tasks. It allows for changes in task inputs within occupations to be easily detectedover time as some occupations may change only one characteristic while keeping others con-stant. This approach also addresses a possible bias in selecting suitable characteristics tomeasure the importance of tasks inputs.18To calculate the RTI measure, I select 30 occupational descriptors and group them intodescriptors measuring abstract, routine and manual tasks. I report the full list of the descrip-tors including the corresponding scales of measurement in appendix table A.3. I calculatethe routine, abstract and manual task inputs for each occupation by averaging the scores of15Autor et al. (2003) developed the original underlying data for the RTI measure and define the routinetasks as follows: ‘In our usage, a task is routine if it can be accomplished by machines following explicitprogrammed rules. Many manual tasks that workers used to perform, such as monitoring the temperatureof a steel finishing line or moving a windshield into place on an assembly line, fit this description. Becausethese tasks require methodical repetition of an unwavering procedure, they can be exhaustively specified withprogrammed instructions and performed by machines.’16I find supportive evidence for changes in the RTI measure within occupations in appendix figure A.4.17I report the exact number of occupations in each O*NET release for which I calculate the RTI measurein table A.218Acemoglu and Autor (2011) also construct RTI measure based on the O*NET database version 14.0. Theapproach and the selection of the underlying descriptors is similar and my RTI measure strongly correlateswith the RTI measure in this paper. The main distinction is that I utilize the updates of the O*NET databaseas well as the original data from the Dictionary of Occupational Titles included in the first releases of theO*NET database. This allows me to account for the changes within occupations over time and avoid carryingover possible measurement errors to other years.18the corresponding descriptors as followsT jocc,t =1Nj∑i∈NjDescriptorocc,t,ij ∈ {Routine, Manual, Abstract}, (2.23)where T jocc,t is the input of task j in occupation occ at time t and Nj is the set of descriptorsfor j′s task input. Using the task inputs, I calculate the RTI for each occupation asRTIocc,t = ln(TRocc,t)− ln (TMocc,t)− ln (TAocc,t) (2.24)Occupations with a high RTI score predominantly consist of routine tasks and hence havehigh potential to be automated by machines. Occupations with a low RTI score require stronginput of abstract or manual tasks, such as critical thinking, active learning, negotiation oroperating vehicles and hence cannot easily be automated. I report the summary statistics ofthe RTI scores and the underlying task inputs at the level of occupations in appendix tableA.4. To provide examples of the most routine and non-routine occupations, I list represen-tative routine and non-routine occupations in appendix table A.5. These occupations areselected to represent the typical occupations with low and high RTI scores.19 It is importantto note that the routine task intensity is not the same dimension as labor skill dimension.Routine occupations are generally concentrated in the middle skill-level occupations withaverage labor income (see e.g. Autor and Dorn (2013)).20The AP measure can vary between industries, as well as across firms within the sameindustry. The majority of the variation, however, is at the between industry level. Table 2.1reports the summary statistics and variance decomposition of the AP measure into withinand between industry variation. Panel A shows that there is a considerable variation in theshare of displaceable labor measured by the AP measure at the firm level across firms inall industries and the results in panel B show that similar variation is present also whenfirm level observations are aggregated at the industry level, i.e. variation between industries.Panel C then shows formally that the majority of the variation in the AP measure is due tobetween industry variation, as it accounts for about 70% of the total variance. This resultindicates that the automation potential measured by the ratio of routine labor to capital isa predominantly industry-specific characteristic.21 I list the industries with the lowest andhighest automation potential, as defined by the AP measure in appendix table A.7. I establishthe validity and robustness and discuss potential shortcomings of the new RTI measure inappendix A.3.I next establish the relationship between the AP measure and two other potential mea-sures. First, I use the total number of all employees instead of the total capital in thedenominator of the AP measure, i.e. Routine intensityj,t = ln(Lroutinej,tLtotalj,t). Second, I use19Since the RTI score changes for each occupation over time, I select representative routine and non-routineoccupations with high and low RTI score, respectively, in multiple years. Full lists and detailed descriptionsof the most routine and non-routine occupations in O*NET 98, O*NET 5.1 and O*NET 20.1 are in appendix.20Another interesting fact shown in Abis (2017) is the ongoing pervasion of computing technology andalgorithmic trading relative to discretionary human-based trading in fund management industry.21Although the majority of the variation in AP is at the industry level, there are still firms within the sameindustry that differ from the industry peers. The within industry variation in share of routine labor relativeto industry peers has been studied by Zhang (2018).19the number of all employees instead of employees in routine occupations in the numerator.This measure is in essence a firm’s labor intensity, i.e. Labor intensity = ln(Ltotalj,tKj,t). Thesetwo measures are positively correlated with each other. Appendix figure A.3 shows that thecorrelation is strong except for the 1990s when the sample was inflated by a large number ofnew dot-com firms that lacked typical production facilities. The strong relationship betweenthese two measures has re-established again since 2005.22 Due to this strong relationship,both measures sort firms and industries cross-sectionally in a very comparable way in mostof the years examined.I also examine the importance of different types of capital in the denominator of theAP measure. Specifically, I use only equipment or equipment and intellectual products andproperty instead of a firm’s total capital to address a potential concern that the total capitalof some firms is dominated by structures. I also account for the changes in the composition ofequipment over time on the basis of the real stock of 64 different types of equipment.23 Thisaddresses the potential concern that some industries significantly increased the real stock ofequipment, such as ICT equipment, whose nominal prices decreased over time and hence theincrease in real stock would not be reflected in firm’s total nominal capital. None of thesevariations are essential for the subsequent analysis. The results in the following sectionsare robust to the use of these alternative AP measures and results based on alternativespecifications of AP measure are available upon request.2.3.2 Financial DataI focus on common stocks of firms listed on NYSE, NASDAQ, and AMEX from 1970 to 2015and use monthly data from the Center for Research in Security Prices (CRSP). To calculatethe AP measure, I require the firms to have underlying data for the number of employeesand total net capital in the Standard and Poors Compustat database. I exclude financialsSIC 6000-6799, utilities SIC 4900-4999, and public sector companies SIC≥9000. Hou et al.(2017) find that the a large number of previously documented cross-sectional return anomaliesare strongly driven by micro-cap stocks and these anomalies are hardly replicable when thesample excludes micro-cap stocks. To address this concern, I exclude small and illiquid stocksas well as stocks with short history from my sample. I describe the applied filters in detailin appendix A.4. None of these filters are essential for the documented results. The finalsample contains a total of 8,150 unique firms. I also redo my analysis on the basis of S&P500 stocks and find quantitatively and qualitatively very similar results. The results basedon the constituents of S&P 500 are presented in appendix.2.3.3 Measuring Technology Shocks Embodied in CapitalI measure the technology shocks embodied in capital by the changes in quality adjustedrelative prices of investment goods. The formal theoretical validity of this measure has beenderived in previous literature, e.g. Greenwood et al. (1997); Cummins and Violante (2002);Hornstein and Krusell (1996) (see appendix A.5 for a parsimonious theoretical foundation22The pattern is robust to the use of RTI measures based on each O*NET or DOT data.23I use the BEA classification for the different types of equipment.20based on these papers).24 The intuition behind the theoretical foundation of this measure isthat investment-specific technology shocks allow for production of new capital goods, suchas machines, software or equipment, at lower cost or higher quality. Hence the changes inthe quality adjusted prices of capital goods relative to aggregate prices are informative aboutthe technological progress. I construct the measure of technology shocks embodied in capital(I-shock measure) by subtracting the log changes in the quality-adjusted prices of capitalgoods from the log changes in the aggregate price level. An increase in the I-shock measureindicates a positive technology shock and is associated with higher quality and/or lower pricesof capital goods.The underlying price indexes are from the U.S. Bureau of Economic Analysis (BEA). Imeasure the aggregate price level by the price index of the personal consumption expendi-tures for nondurable goods. The quality-adjusted prices of capital goods are based on theprice index for private fixed investment in nonresidential equipment. This price index is acombination of price indexes of 25 different types of equipment. I use annual frequency forboth time-series and calculate the I-shock measure from 1970 to 2015.Since the I-shock measure is based on quality adjusted prices of capital goods, it is im-portant to account for the changes in quality of goods in the underlying price indexes. Theincrease in the horsepower of cars and increases in the processing power and memory of com-puters over recent decades are prominent examples of large quality changes of goods. Ideally,quality-adjusted price indexes take the changes of quality into account and express the priceper efficiency unit. Several earlier papers, e.g. Griliches (1961); Gordon (1990), point outthe lack of an appropriate quality adjustment of the NIPA price indexes of equipment. TheBEA has addressed the problem of quality adjustment over the last three decades and devel-oped price indexes that account for quality changes using various techniques such as hedonicregressions.25 The price indexes of fast improving types of equipment such as computers andperipherals have used hedonic-based quality adjustments since the 1980s. I further accountfor possible quality changes by using the quality adjustment from Cummins and Violante(2002). This adjustment accounts for the measured quality difference in the underlying com-ponents of the price index of capital goods from Gordon (1990) until 1983 and extrapolatesthe adjustment to recent years.Consistent with previous literature, e.g. Fisher (2006), I document a positive mean ofthe I-shock measure, indicating a secular trend of decreasing prices and increasing qualityof capital goods. I provide a full set of summary statistics of the I-shock measure in ap-pendix table A.10. The I-shock measure correlates strongly with the TFP shock of industriesproducing the capital goods. The correlation with TFP shock of industries producing con-sumption goods is practically zero. This result is in line with the theoretical foundation ofthis measure. The correlation with the aggregate TFP shock is negligible. This reflects thefact that industries producing the capital goods comprise a relatively small part of the U.S.economy. The I-shock measure has a positive correlation with the GDP growth. This isconsistent with previous papers showing that prices of capital goods are an important driver24Various versions of these measure has been successfully used in previous studies focusing on the effectsof the investment-specific technology shock on economic growth, e.g. Greenwood et al. (1997) and businesscycle, e.g. Greenwood et al. (2000); Fisher (2006)25Wasshausen et al. (2006) provide a brief overview of the use of hedonic methods by the BEA to measurereal GDP; Fox and McCully (2016) provide detailed information about the price indexes for all types ofequipment and discusses the quality adjustments employed by BEA.21of economic growth, Fisher (2006); Greenwood et al. (2000).The dashed blue line in figure 2.6 plots the normalized time-series of the I-shock mea-sure from 1970 to 2015.26 The most prominent and pervasive technology shock during thisperiod was from 1997 to 2001. During these years, the normalized I-shock measure recordeda cumulative increase of 5.71 standard deviations. This increase was driven mainly by alarge drop in the quality-adjusted prices of information processing equipment. The prices ofthis equipment decreased by 40 percent, while the prices of personal expenditures for non-durables increased by 8 percent. Since information processing equipment accounted for onethird of total private investment, the aggregate price index of capital goods decreased by 15percent. The massive drop in the prices of information processing equipment can be trackedback to a significant productivity growth (TFP shock) in a single manufacturing industry.The electronic computer manufacturing industry recorded an increase in productivity of 216percent. The remaining manufacturing industries, including consumption goods producers,experienced only moderate productivity growth of 5.33 percent.27I link all other major shocks in the I-shock measure to the differential productivity growthof a small number of capital goods producing industries in appendix table A.11.28 I show inappendix table A.11 that the whole time-series of the I-shock measure can be approximatedby differential productivity growth of capital producing industries. This evidence providesvalidity to the theoretical foundation of the I-shock measure as an appropriate measure oftechnology shocks embodied in capital goods.2.3.4 I-shock as an Income State VariableI show in a companion paper Knesl (2018) that technology shocks embodied in capital,measured by the changes in relative prices of capital goods, are an important source of jobdisplacement and labor income risk. Lower prices and higher quality of capital goods allowsubstituting capital for labor in occupations that have the potential to be automated. Hence,technology shocks embodied in capital are an important state variable indicating the risk inlabor income of households. This result is broadly in line with other papers investigating theimpact of innovation and technological progress on labor income, e.g. Kogan et al. (2017,2018). I provide a concise illustration of this result below. I refer the reader to the papersmentioned above for detailed results.Panel A of table 2.2 reports the results of regressing contemporaneous and future changesin personal income, private fixed investment in equipment and employment on the I-shockmeasure and aggregate productivity shock. Specifically, I estimate the regression of this form1K + 1K∑k=0yt+k − yt+k−1 = α+ γIshock × Ishockt + γTFP × TFPt + t+k, (2.25)where yt is the log level of nominal personal income, private fixed investment, or total non-farm employees.29 Ishockt is the I-shock measure representing the change from t − 1 to t.26The solid red line is a factor mimicking portfolio that is describe in sections 2.4.27The calculations are based on the data from the NBER-CES Manufacturing Industries database Bartles-man and Gray (1996) and the Input-Output Accounts data from the U.S. Bureau of Economic Analysis28See also Greenwood and Yorukoglu (1997) for a detailed discussion of the positive technology shock in1974.29I use nominal personal income to avoid a potential mechanical effect of deflating as the I-shock measure22The TFPt is the change in the aggregate productivity over the same time period. I estimatethe regression at annual frequency from 1970 to 2015 for time horizons K = 0, 1, and 2. Theresults are robust to controlling for lagged dependent variable and alternative specifications.The first three columns show that a positive technology shock embodied in capital (I-shock)is associated with an immediate increase in investment in equipment that fades away in thesubsequent years. The estimates in columns (4)-(6) indicate that the I-shock leads to asignificantly lower employment in the first and second year after the shock. The last threecolumns then show that a positive I-shock is also associated with lower personal incomegrowth.Looking only at the aggregate average income growth can hide potentially important ef-fect of technology shocks on individual personal income growth. To examine the effect ofthe I-shock on higher moments of cross-sectional distribution of individual income, I esti-mate the regression equation 2.25 with changes in the first three central moments of thecross-sectional distribution of individual income growth as the dependent variable using thedata from Guvenen et al. (2014).30 Panel B of table 2.2 presents the estimates. The firstthree columns verify that the I-shock has a negative effect on the average income growth asdocumented in panel A. Columns (4)-(6) show the standard deviation increases only slightlyfollowing a positive I-shock. The last three columns show that a decrease in prices of capitalgoods, a positive technology shock, has the most negative effect on the skewness of individualincome growth. A positive technology shock embodied in capital significantly increases theleft skewness of the income growth distribution. Individuals are more likely to experiencelarge drops in personal income in years following a positive I-shock.This result is in line with Knesl (2018) in showing that a positive I-shock increases theprobability of losing job. Large negative idiosyncratic shocks to income growth such as jobloss are then pronounced in higher moments of income growth distribution as shown above.This support the assumption of the state dependent idiosyncratic labor productivity risk inthe theoretical model presented earlier. Higher moments of the cross-sectional distributionof individual income growth play a central role in pricing assets in incomplete markets whenagents face uninsurable income risk. Constantinides and Duffie (1996) derive a theoreticalmodel of incomplete markets with heterogeneous agents and uninsurable idiosyncratic incomerisk where the second moment of the cross-sectional distribution of individual labor incomeenters the stochastic discount factor. Constantinides and Ghosh (2017) show that the thirdcentral moment of individual income growth drives asset prices and risk-free rate.31 Schmidt(2016) provides evidence that the idiosyncratic tail risk of income growth is state-dependentand predicts stock returns, and further shows that a model with state-dependent, time-varying idiosyncratic disaster risk and incomplete markets matches the key asset pricingmoments.contains changes in prices of consumption goods in its numerator. The estimation results based on nominalpersonal income represent lower bound.30Estimating the regression equation (2.25) with levels of the central moments as dependent variablesprovides similar results. The results are similar when I control for lagged dependent variable.31The theoretical model in Constantinides and Ghosh (2017) assumes that households consumption equalshouseholds labor plus dividend income, where the labor income is subject to uninsurable idiosyncratic shock.The model is estimated using individual consumption data from the Consumer Expenditure Survey232.4 Technology Shocks and Displaceable Labor2.4.1 Co-movementI sort stocks into five portfolios based on their share of displaceable labor (AP) to examine thetime-series and cross-sectional properties of stock returns associated with firms’ automationpotential. I report the summary statistics of firm characteristics in each portfolio in appendixtable A.13. The first portfolio consists of stocks with the highest AP. These firms also havehigh average labor intensity and a high percentage of employees in routine occupations. Theaverage AP measure then decreases from -2.9 in the first portfolio to -5.9 in the fifth portfolio.The portfolios are rebalanced annually at the end of June and portfolio returns are value-weighted unless stated otherwise.Figure 2.6 plots the I-shock measure and the annual stock return difference between firmswith low and high displaceable labor share, i.e. long-short portfolio with a long positionin the bottom quintile of firms ranked by the AP measure and a short-position in the topquintile, the KML portfolio. These two time-series co-move very strongly and overlap in allspikes. The correlation coefficient is 0.60. This strong co-movement is a robust feature of thedata and it is present across various samples of firms. It holds when I exclude firms producingcapital goods from the sample and it is present independently among both, manufacturing andnon-manufacturing firms. The co-movement is also not driven by firms’ operating leverageor size (see figure A.5 and table A.20 in the appendix for the long-short portfolio consistingof firms in these samples as well as portfolios controlling for operating leverage and size).This results indicates a very strong connection between macroeconomic variable measuringchanges in capital goods prices and the dynamics of stock returns.I first examine whether the co-movement of the KML portfolio with the I-shock measureis driven by ranking individual firms relative to industry peers or by ranking whole indus-tries across the economy. Table 2.3 reports the correlation coefficients between the I-shockmeasure and the long-short portfolios based on the within- and between-industry variationin displaceable labor share, AP. The correlation is strong, 0.53-0.59, for long-short portfoliosbased on ranking whole industries. It is weak for long-short portfolios of firms ranked relativeto industry peers. This is in agreement with the dominance of the between-industry varia-tion in the AP measure discussed above (see panel C in table 2.1). Both pieces of evidenceunderline the role of displaceable labor share as an industry characteristic.I next examine how the exposure to technology shocks varies across the five portfolios.Panel A of table 2.4 reports the results from regressing annual portfolio returns on a constantand the I-shock measure. The first portfolio, consisting of stocks with the highest share ofdisplaceable labor, has the strongest and most negative exposure to technology shocks. Theestimated coefficients are negative for all five portfolios. They increase monotonically acrossall portfolios from -3.4 for the first portfolio to -1.0 for the fifth portfolio. The estimatedcoefficients are economically large and statistically significant for the first four portfolios. Thisresult indicates that the strong co-movement of the KML portfolio with the I-shock measureis generated by stocks with high share of displaceable labor. These stocks react negativelyto technology shocks. The KML portfolio, i.e. Low-High AP, then exhibits positive andsignificant exposure to technology shocks of 2.3.Panel B in table 2.4 repeats the analysis controlling for the aggregate market return.The first portfolio has still the strongest and most negative exposure to technology shocks.24The pattern of the estimated coefficients remains monotonic across the portfolios. A notabledifference is that all coefficients of the I-shock measure are shifted up so that the fifth portfoliohas a positive and significant exposure. The coefficients of the I-shock measure across thefirst three portfolios remain negative and statistically and economically significant. The KMLportfolio has a positive and significant exposure to technology shocks of same magnitude, 2.4.The major part of this exposure, 1.6, comes again from the short position in the first portfolioconsisting of firms with high share of displaceable labor.Regressions in panel C control additionally for the size (SMB) and value (HML) factors asdefined by Fama and French (1993). The results are very similar. The estimated coefficientsof the I-shock measure increase monotonically across the portfolios from -1.4 for the firstportfolio to 0.56 for the fifth portfolio. These results show that standard factors cannotexplain the differential exposure to technology shocks across portfolios containing firms withlow and high displaceable labor share. I show in appendix table A.12 that the results are alsonot driven by the profitability and investment factors of Fama and French (2015), the returndifference between capital goods and consumption goods producers (IMC) of Papanikolaou(2011) and aggregate economic growth.The findings in this section document that technology shocks embodied in capital have astrong and negative impact on firms with high share of displaceable labor. These firms aregenerally labor-intensive with relatively low capital but high use of labor that can be easilyautomated. Firms with low displaceable labor share usually have high capital or not easilyreplaceable labor and weak exposure to technology shocks. The results on firms’ exposurepresented in this section are consistent with the predictions of the theoretical model.2.4.2 Firms’ Exposure to Technology ShocksThe previous section shows that the stock return differences between firms with low andhigh displaceable labor share, the KML portfolio, strongly correlate with the time-seriesof technology shocks, the I-shock measure. Hence, the KML portfolio is a reasonable, stockreturn-based measure capturing the time-series variation of technology shocks. I use the KMLportfolio to estimate firms’ exposure to technology shocks directly from the stock returns.Specifically, for each firm I estimate the exposure to technology shock, βKMLi , from a rollingregression:ri,t = αi + βKMLi rKML,t + i,t, (2.26)where ri,t denotes firm i′s monthly excess return and rKML,t is the monthly return on theKML portfolio. I estimate this regression for every stock in the sample each year at the endof June using monthly returns over the last 60 months.32 This method allows the βKMLi varyover time. The results reported below are robust to including market excess return and thesize, and value factors of Fama and French (1993) as control variables in regression equation2.26.33Technology shocks are a significant source of labor income and job displacement risk. TheI-shock measure is a state variable indicating the labor income and job displacement risk,32I require the stock to have at least 30 observations in the last 60 months.33The results are also robust when the KML portfolio is orthogonalized to the factors of Fama and French(1993).25especially the risk of large adverse idiosyncratic events such as job loss that are pronouncedin the higher moments of income growth distribution. The βKMLi measures stock’s exposureto the labor income risk directly from the covariances of the returns with the state variable.Hence, sorting stocks on the βKMLi is conceptually different from sorting stocks on the shareof displaceable labor, AP. Although these two variables are related and correlate strongly,βKMLi captures stock i′s co-movement with a state variable that is highly relevant for investorsmarginal utility. The differences between sorting on characteristics and sorting on covarianceshas been discussed in number of previous studies, e.g. Daniel and Titman (1997).I next examine the properties of the portfolios sorted by the βKML. Each year at the endof June, I sort stocks into five portfolios by their βKML and hold the stocks in a given portfoliofrom July of year t to June of year t+1. Panel A of table 2.5 reports the post-ranking exposureto the I-shock measure across the portfolios. The exposure is estimated from regressing eachportfolio’s annual value-weighted return on a constant and the I-shock measure. The firstportfolio consisting of stocks with low βKML has strongly negative and significant exposureto the I-shock measure. The coefficients are then increasing, absolute values are decreasing,across the portfolios almost monotonically. The sole exception is the second portfolio, whichhas slightly more negative exposure to the I-shock than portfolio one. This result shows thatβKML successfully estimates firms’ exposure to the underlying state variable, the I-shockmeasure. Firms’ exposure to the KML portfolio predicts the exposure to the I-shock measureat horizon of at least one year. Panel B shows additional properties of stocks within eachportfolio. The fifth portfolio contains stocks with relatively low share of displaceable laborcompared to stocks in portfolio one. This is consistent with low displaceable labor-share firmshaving weaker exposure to technology shocks.34 The book-to-market ratio and market betasare slightly decreasing across the portfolios, while firm size is somewhat increasing. Othercharacteristics are flat across the portfolios or do not exhibit a clear pattern.2.5 Exposure to Technology Shocks and Return PremiumIn this section, I investigate whether a stock’s expected return is related to the exposureof its returns to technology shocks, βKML. I estimate the stock return premium using thecross section of stocks. Then I examine the time variation and predictability of the premium.Finally, I show that stocks exposure implies differential employment and investment behaviorconsistent with the mechanism in the model.2.5.1 Return PremiumTo examine whether differences in stocks’ exposure to technology shocks are associated withdifferences in their average returns, I sort stocks into five portfolios according to their βKMLas described in the previous section. For each portfolio, I calculate a time-series of monthlyportfolio returns as value-weighted average returns (unless stated otherwise) across all stocksin the portfolio in a given month. The sample period is from 1975 to 2015.3534I show the time-series of the contemporaneous correlation between the share of displaceable labor AP andthe βKML at firm and industry level in appendix figure A.6.35The first five years 1970-1975 are used to estimate the βKML for portfolios formed in June 1975.26Panel A of table 2.6 reports the average annualized portfolio excess returns. The firstportfolio has the highest average excess return, 12.45%. This portfolio consists of stockswith the most negative βKML. The average excess returns decrease monotonically across theportfolios. The long-short, High-Low βKML, portfolio earns a negative premium of -6.85%per year. This premium is statistically and economically significant.The results are very similar for equally-weighted portfolio returns, reported in panel B.The average excess returns decrease monotonically across the portfolios, although the returndifferences are smaller. This results in a somewhat smaller return premium of the High-LowβKML portfolio, -4.29% per year. The return premium is still statistically and economicallysignificant.I next examine whether the return differences can be explained by different exposures toaggregate market risk. Panel C reports the results of regressing the portfolio monthly returnson a constant and the market excess return. The estimated intercepts, alphas, exhibit similarpattern as average excess returns. The first portfolio has the highest alpha of 2.97% per year.The alphas decrease monotonically across portfolios. The High-Low βKML earns a significantand negative alpha of -5.39%. Although the estimated coefficients of the market returnsdecrease somewhat across the portfolio, the differences in exposure to aggregate market riskcannot explain the differences in average returns.Panel D reports estimates of regressions that include also the size (SMB) and value (HML)factors of Fama and French (1993) to control for potential exposure to these factors. Theestimated alphas are slightly smaller in their magnitude when compared to panel C, but theoverall result is similar. The alphas decrease across the portfolios almost monotonically. Theonly exception is the second portfolio, whose alpha marginally exceeds the alpha of the firstportfolio. The alpha of the High-Low βKML portfolio is -3.6% and is significant at the 10%level. Although the coefficients of the SMB factor also decrease across the portfolios, thedifferences in the exposure SMB factor do not explain the return premium. The pattern inthe SMB coefficients is consistent with the evidence in table 2.5 indicating increasing stocksize across the portfolios.I repeat this analysis for the constituents of the S&P500 index, which are the largest andmost liquid stocks in the U.S. stock market. The results are very similar. They indicate areturn premium for High-Low βKML portfolio between -7.16% and -4.75% (see table A.14in the appendix). Hence, the results do not depend on the presence of small and illiquidstocks.36I next use panel regressions with firms’ annual stock returns to differentiate the returnpremium for stocks’ exposure to technology shocks, βKML, from well-known cross-sectionalreturn predictors. Specifically, I run panel regressions at the firm level:Ri,t+1 = α+ γβKMLi× βKMLi,t +∑nγn ×Xi,n,t−1 + Y eart + i,t+1, (2.27)where Ri,t+1 is firm i′s annual stock return from July of year t to June of year t+ 1, βKMLi,tis a firm’s exposure to technology shock at the end of June of year t, Xi,n,t−1 are firm-levelcontrol variables at the end of year t− 1, and Y eart is year fixed effect.3736The results are also very similar when the stocks are sorted into portfolios based on βKML that is estimatedfrom multivariate rolling regressions including the size (SMB) and value (HML) factors of Fama and French(1993) as control variables (see table A.15 in the appendix).37The control variables Xi,n,t−1 that are based on data from firms’ financial statements, i.e. book-to-market27Table 2.7 presents the results of estimating the panel regression 2.27. The specificationin column (1) includes only a constant and βKML. The estimated coefficient is negative andstrongly significant. The coefficient implies an average return differential of -8.71% per yearbetween firms with high and low βKML.38 Including the year fixed effect, column (2), leadsto a lower coefficient that implies a return premium of -3.64%. Columns (3)-(10) show resultsfor regressions that include βKML, year fixed effect and one of the control variables Xi,n,t.Each of these control variables has only a negligible impact on the estimated coefficient ofβKML. The coefficients are negative and significant across the specifications with differentcontrol variables. The implied return difference between firms with high and low βKMLranges from -3.29% to -4.62% per year. The regression in the last column controls jointlyfor all variables Xi,n,t and the year fixed effect. The estimated coefficient of βKML remainsnegative and statistically significant. The slightly higher magnitude implies an annual returnpremium of -4.8%.The estimated coefficients of βKML imply a similar return premium for firms’ exposureto technology shocks as the portfolio analysis described above. I show in appendix table A.16that the estimation results are similar for the constituents of the S&P500. The coefficient ofβKML is also robust to controlling for additional firm- and industry-level variables.The panel regressions assume a strictly linear relationship between stock returns and thecross-sectional return predictors. To control for a potentially non-linear relationship, I useconditional double sorts. Specifically, in the first step I sort stocks into three baskets basedon the value of the control variable. In the second step, I sorts the stocks within each ofthese baskets again into three additional baskets according to the βKML. This creates a totalof 9 baskets. I then create three portfolios by pooling all stocks in the βKML-based basketswith the same rank. This procedure creates portfolios sorted on βKML that are relativelyindependent of the return predictors used in the first step.39Table 2.8 reports the average value-weighted portfolio returns. I first report the aver-age portfolio returns for unconditional sorting on βKML, column (1). The average returndifferential based on unconditional sorting into three portfolios is -6.6% per year. Column(2) shows results for βKML-sorted portfolios conditional on the book-to-market ratio. Theportfolio with the low βKML stocks has the highest average return of 10.83% per year. Theaverage return decreases monotonically across the remaining two portfolios. The conditionalHigh-Low βKML portfolio earns a statistically and economically significant return premiumof -5.23% per year. This return premium is relatively independent of the differences in firms’book-to-market ratio. Columns (3)-(9) show that conditioning the portfolio sorts on othercross-sectional return predictors provides very similar results. The average portfolio returnsratio, cashflow, financial leverage, price-to-cost margin, profitability, and size, are calculated from financialstatements for fiscal year ending in year t− 1. This allows a lag of 6-18 months for the information to becomepublicly available. Market betat−1 is estimated over the last 60 months ending in June of year t. Turnover iscalculated over the calendar year t− 1.38I calculate the average return differential by multiplying the estimated coefficient of βKML, -3.829, withthe average difference in βKML between the 95th and 5th percentile, 2.275. The average difference in βKML iscalculated as time-series average of contemporaneous differences in βKML between the 95th and 5th percentile.The implied return premium for the average difference in βKML between the 75th and 25th percentile is-4.826%.39In an alternative approach, I first calculate time-series of value-weighted portfolio returns for each of thenine conditionally-sorted portfolios. I then create three portfolios by calculating a simple average across theβKML-sorted portfolios with the same rank. The results based on this approach are very similar.28decrease across the portfolios. The premium for the High-Low βKML portfolio ranges be-tween -4.21% and -5.43% per year and is statistically significant and economically large.Results based on the constituents of S&P500 are very similar (see appendix table A.17). Thereturn premium is also robust to conditioning on other firm- and industry-level variables.Fora comparison, the last column shows the premium for the book-to-market ratio conditionalon firms’ size. The conditional value premium is 2.15% but not statistically significant.The portfolio return analysis, panel regressions and conditional double sorts imply similarreturn premium for firms’ exposure to technology shocks. The latter two methods show thatthe return premium is not driven by linear or non-linear relationship with well-known cross-sectional return predictors.2.5.2 Time Variation and Predictability of the Return PremiumIn this section, I examine whether the return premium varies over time and is predictable.As mentioned above, technology shocks are an important source of labor income and job dis-placement risk and the I-shock measure predicts higher moments of individual income growthdistribution. Hence the I-shock measure can be considered a state variable for households’marginal utility. This evidence gives rise to the consideration that a return premium of anasset with a strong exposure to the technology shock is predictable by this state variable.I use the I-shock measure to forecast the return premium of the Low-High βKML portfolio.I focus on the Low-High βKML instead of High-Low βKML portfolio to facilitate the inter-pretation of the return premium. This portfolio has negative exposure to technology shocksand earns a positive return premium (see table 2.6). Specifically, I run forecasting regressionsof the cumulative log return premium for the Low-High βKML portfolio,∑ys=1 rβKMLt+s , overy years, on the I-shock measure Ishockt:y∑s=1rβKMLt+s = a(y) + b(y)Ishockt + e(y)t+y. (2.28)Panel A of table 2.9 reports the estimation results of the forecasting regression for theone- to five-year cumulative log return premium of the Low-High βKML portfolio. Theestimated coefficients have the expected positive sign; a decrease in relative prices of capitalgoods - positive technology shock increases the return premium. The estimated coefficientsare marginally statistically significant for return horizons of two, three and four years. Themagnitude of the coefficient is economically important. For example, a positive technologyshock of one standard deviation increases the expected two-year return by 2.4%. Panel Bshows the results of the forecasting regressions that include additional forecasting variables.The estimated coefficient of the I-shock have similar magnitude and remain statisticallysignificant for the same return horizons.Figure 2.7 plots the I-shock measure and the subsequent realized three-years cumulativelog return premium of the Low-High βKML portfolio. The return premium varies over timetogether with the I-shock measure. Both time-series overlap relatively closely. Periods ofsubstantial decreases in the relative prices of capital goods predict a high return premiumfor the Low-High βKML portfolio.As discussed above, labor income and job displacement risk are high in states of largepositive technology shocks. This evidence suggests that the time-varying return premium can29be compensation for this type of risk. Households will seek a return premium for assets thathave negative exposure to states of high labor income risk. The time variation of the returnpremium between the states of high and low technology shocks supports this explanation.2.5.3 Firms’ Responses to Technology ShocksThe evidence in earlier sections shows large differences in firms’ risk exposure to technologyshocks. In this section, I investigate whether firms with different exposure to technologyshocks differ also in their employment, investment, sales and financial performance when therelative prices of capital goods decrease. Costly employment and capital adjustments, andchanges in firms’ sales following technology shocks can affect firms’ financial performance andcontribute to firms’ risk exposure. To explore this mechanism, I estimate panel regressionsin the form:∆yi,t+s = b0 Ishockt + b1 IβKMLi,t + b2 Ishockt × IβKMLi,t (2.29)+controlsi,t + TFPt + Trendt + ei,t+s,where ∆yi,t+s is the firm i′s percentage change in employment, capital, capital-labor ratio,sales, percentage points change in return on capital, or return on equity between years tand t + 3, Ishockt is the I-shock measure in year t, IβKMLi,t is an index between 0 and 9indicating the decile of cross-sectionally ranked firms’ exposure to technology shock, βKMLi,t ,and Ishockt × IβKMLi,t is their interaction term. Controlsi,t is a set of firm-specific controlvariables consisting of firms’ capital Ki,t, age Agei,t, market capitalization Sizei,t, firm fixedeffects, and a lagged dependent variable at one-year horizon t−1 to t. The regressions includeTFP shock TFPt and deterministic trend Trendt as aggregate control variables.Panel A of table 2.10 reports the results of estimating regression equation (2.29) forchanges in employment, capital and capital-labor ratio. The first column shows that a posi-tive technology shock on average decreases firms’ employment. The coefficient of the I-shockmeasure is negative, statistically significant and economically large. The effect of the tech-nology shock differs significantly across firms with different βKMLi,t . The positive coefficientof the interaction term Ishockt × IβKMLi,t indicates that the effect is stronger for firms withnegative exposure to technology shocks. In the next two columns, I split the sample in halvesbased on firms βKML and estimate the panel regression (2.29) without the interaction term.Firms in the first half have strongly negative exposure to technology shocks, while firms in thesecond half have only mild exposure that is much closer to zero (see table 2.5). The resultsshow that firms with strongly negative exposure decrease employment following a positivetechnology shock. At the same time, employment is on average constant in firms with mildexposure over the same time horizon. The following three columns show that technologyshocks increase firms’ capital. The effect is approximately equal for firms with strongly neg-ative and mild exposure. This result is consistent with evidence discussed above (see table2.2) and indicates that lower quality-adjusted relative prices of capital goods lead to higherinvestment. The last three columns show that technology shocks increase firms’ capital-laborratio. The effect is again stronger for firms with strongly negative exposure.As described earlier, firms with strongly negative exposure to technology shocks are gen-erally firms with very high share of displaceable labor (see table 2.5). The results indicate30that these firms indeed adjust, potentially automate, their production processes by decreasingthe number of employees and investing in capital when a technology shock occurs.Panel B reports the results of the percentage changes in sales and percentage point changesin return on capital and return on equity. The first column shows that a technology shock hasnegative impact on firms’ sales. The estimated coefficient is statistically and economicallysignificant. The following two columns show that the effect is stronger for firms with stronglynegative exposure. The estimated coefficient of the I-shock measure is more than twice aslarge for firms with strong exposure compared to firms with mild exposure. The next threecolumns show that lower relative prices of capital goods decrease firms’ return on capital.This effect is again larger for firms with strong risk exposure. The estimates in the lastthree columns indicate also negative and strongly differential effect of technology shock onthe return on equity. Firms with strongly negative exposure experience decreases in returnon equity almost twice as large as firms with mild exposure. The estimated coefficients arestatistically significant and economically large.I repeat the analysis for different time horizons. I report the results for a time horizon ofthree years in appendix table A.21. The results are qualitatively very similar. The estimatedcoefficients are generally larger as they refer to changes compounded over three years. Theseresults indicate that the differential effects of technology shocks between firms with strongand mild risk exposure are not driven by potential lead-lag differences in firms’ responses.Further I examine the effects of technology shocks on firms’ profit margin and how theseeffects differ between firms with a strongly negative exposure to technology shocks and thosewith a mild exposure. I report these results in appendix table A.22. The results indicatethat profit margins of firms with a strongly negative exposure to technology shocks decreasemore following a technology shock than profit margins of firms with a mild exposure.The theoretical model presented earlier highlights the importance of competition for theeconomic mechanism how technology shocks affect firm value through costly technology adop-tion. I focus on this mechanism by solving the model for different degrees of elasticity ofsubstitution for firm-specific goods, which governs the intensity of the within-industry com-petition. I plot the outcomes in appendix figure A.2. This figure shows that the negativeeffects of technology shocks on income of firms with automation potential will be larger inhighly competitive industries. I next investigate this mechanism empirically. I split the sam-ple into two halves by the degree of competition and estimate the panel regression (2.29)without the interaction term for firms with strong and mild exposure within each half sepa-rately. I measure the degree of industry competition by the Herfindahl-Hirschman index ofsales at the level of SIC 3 digit industries.Panel A of table 2.11 reports the results for percentage changes in employment and capital.The estimated coefficients have the same sign as discussed earlier. A positive technology shockdecreases firms’ employment and increases capital. For firms with strongly negative exposure,these effects are stronger in highly competitive industries. For firms with mild exposure, theeffect on capital is stronger in highly competitive industries, but the effect on employment isweaker.Panel B reports the results for percentage changes in capital-labor ratio and sales. Theestimated coefficients of the I-shock measure have uniformly higher magnitude in highlycompetitive industries for both firms with strong and mild exposure. Firms with stronglynegative exposure increase their capital-labor ratio upon the occurrence of a technology shock31more if they are in highly competitive industries. These firms experience also more severedecrease in sales in comparison to firms in less competitive industries. This pattern holdsequally for firms with mild exposure.In panel C, I report the results for percentage point changes in return on capital andreturn on equity. The effects of the technology shocks on these two variables are generallymore pronounced in highly competitive industries. Firms experience a stronger decrease inthe return on capital and in the return on equity if they are in more competitive industries.The findings in this section document that firms with differential exposure to technol-ogy shocks have different employment and capital adjustment in years following a decreasein prices of capital goods, while financial performance is also affected. Firms with stronglynegative exposure have generally high share of displaceable labor and tend to increase theircapital-labor ratio after technology shocks. During this process they experience lower sales,return on capital and return on equity. This evidence indicates that the risk exposure totechnology shocks reflects firms’ financial underperformance after the shocks. Industry com-petition generally increases the observed effects.2.6 ConclusionThe potential of capital to substitute some types of labor and complement others plays acentral role in economic models that study the effects of technological change. I find thatfirms’ potential to automate the production process is also an important determinant of firms’exposure to technology shocks. Technology shocks embodied in capital are an importantdriver of stock returns with a differential impact over the cross section of stocks.I show theoretically that the firms’ share of displaceable labor that can be automatedby capital can increase firms’ riskiness in a competitive environment. Firms in industrieswith high displaceable labor share have negative exposure to technology shocks that facil-itate automation. Automation becomes a costly necessity for firms to stay competitive ascompetition eliminates its benefits for firms and passes them to the consumers in form oflower product prices. This result is strongly supported by empirical evidence.I show empirically that firms with a high share of displaceable labor have very strongand negative exposure to technology shocks. I document that a macroeconomic measure oftechnology shocks based on prices of capital goods can be successfully approximated by stockreturns of a zero-cost portfolio based on firms’ share of displaceable labor and this portfolioshas captured all major technology shocks over the last four decades. I further show that firmsexposure to technology shocks that facilitate automation earn a robust and time-varyingreturn premium that can be empirically and theoretically rationalized as a compensation forhouseholds’ job displacement and labor income risk. The exposure to technology shocks isassociated with firms’ adjustment of the capital-labor ratio and underperformance after adecrease in relative prices of capital goods consistent with the model prediction.My model abstracts from other interesting economic aspects such as market entry, dif-ferential ability to adopt new technology or operating leverage, which are certainly relevantin some real world scenarios and worth exploring both empirically and theoretically. Futureresearch can also focus on a multiperiod fully dynamic setting of my model with endogenousarrival of technology shocks. Such model can deliver richer implications for asset prices aswell as implications for differential growth between heterogeneous industries.322.7 Equilibrium and Market Clearing ConditionsThe CES aggregator technology at the sector implies optimal demand functions for the firm-specific goods of the form:Yf,s,t =(P lf,s,tPs,t)−νYs,t and Yf,s,t =(P kf,s,tPs,t)−νYs,t for s = 1, 2 and t = 1, 2 (2.30)The CES aggregator technology at the aggregate level implies optimal demand function forthe sector-specific goods of the form:Ys,t =(Ps,tPA,t)−ρYA,t (2.31)Market clearing condition for the aggregate good is defined asNh∑h=1ch,t = YA,t (2.32)Market clearing on the labor markets is defined as:Ll1,1 =Nh∑h=1Lh,1,1Ll1,2 + Lk1,2 =Nh∑h=1Lh,1,2Lk1,t =Nh∑h=1Lh,2,t for t = 1, 2(2.33)where the Lls,t is labor used directly in goods production in sector s and Lks,t is labor used inproduction of capital in sector s. The market clearing on the investment capital market is:N∑f=1Kf,1,21{Automated = Y es}f,1,2 = Kα1,tN∑f=1Kf,2,t = Kα2,t for t = 1, 2.(2.34)332.8 Tables and FiguresFigure 2.2: Three Types of Sector Equilibria.This figure plots the relationship between the prices of goods (panel A), demand for goods(panel B), and dividends (panel C) of labor-based and automated firms and the percentageof automated firms δ within the sector. Plots in the first column show a state of no shock.The parameter z2 is chosen so that the good prices of each type of firm are same when nofirm automates (δ∗ = 0). Plots in the middle column show a situation with an incrementaltechnology shock when a percentage of firms (0 < δ∗ < 1) automates. Plots on the right-hand side show a situation of large disruptive technology shocks when all firms in the sectorautomate despite automation cost κ, (δ∗ = 1). The black horizontal line depicts the originaldividend level. The model parameters are: ρ = 0.6, ν = 0.95, χ = 1, θ = 2, N = 1000,Nh = 100, ALow = 0.85, AHigh = 0.95, 1.05, β = 0.85, γ = 21,  = 0.39, pH = 0.15, p = 0.5,and κ = 8× 10−5. Demand fo firms’ goods and firms’ dividends are scaled by a factor 100.34Figure 2.3: Technology Shocks, Equilibrium Automation, and Profits.This figure plots the equilibrium relationship between the magnitude of the technology shockAt and percentage of firms that automate within the sector 1 (panel A), the dividends of firmsin sector 1 (panel B) and the dividends of firms in sector 2 (panel C). Model parameters arestated in the description of figure 2.2. Dividends are scaled by a factor 100.Figure 2.4: Firms’ Exposure to Technology Shock and Expected Returns.This figure plots the relationship between the magnitude of the technology shocks and firms’exposure (panel A), the relationship between the magnitude of the technology shock andexpected stock returns (panel B), and the differences in expected stock returns between firmsin sector 1 and sector 2 (panel C). Model parameters are stated in the description of figure 2.2.35Figure 2.5: Idiosyncratic Labor Productivity  and the Risk Premium for Technology Shocks.This figure plots the risk premium for technology shocks E[RI] − Rf as a function of theidiosyncratic labor productivity risk . RI is a return on an asset that has β = 1 to technologyshock A2. The magnitude of the technology shock AHigh is chosen so that all firms in sectors = 1 automate (δ∗). Model parameters are stated in the description of figure 2.2.36Figure 2.6: Technology Shocks (I-shock) and Low AP minus High AP Firms.This figure plots the I-shock measure (dashed blue line) and the annual return of the zero-cost(KML) portfolio (solid red line). The zero-cost portfolio has a long position in firms withlow share of displaceable labor , i.e. the bottom quintile of firms ranked by the AP measure,and a short position in firms with high share of displaceable labor, i.e. the top quintile. Theportfolio returns within the short and long position of the KML portfolio are value weighted.The KML portfolio in this figure consists of stocks from S&P500 except for financial andutilities firms. Both time-series are plotted at annual frequency and are normalized to meanzero and standard deviation of one. The sample period is from 1970 to 2015.37Figure 2.7: The I-shock Measure and Following 3-Year Cumulative Return of the Low-HighβKML Portfolio.This figure plots the I-shock measure (dashed blue line) and the subsequent realized three-year cumulative log return of the zero-cost Low-High βKML portfolio (solid red line). Thezero-cost portfolio has a long position in stocks with low βKML and short position in stockswith high βKML. The portfolio returns within the short and long position are value-weighted.The portfolio is rebalanced annually at the end of June. Both time-series are plotted at annualfrequency and are normalized to mean zero and standard deviation of one. The sample periodis from 1970 to 2015.38Table 2.1: Summary Statistics and Variance Decomposition of the AP measureThis table reports the summary statistics of the AP measure and its main components, ln(KL)and LroutineLtotal. (-)AP denotes ln(KLroutine)to avoid negative numbers and facilitate readability.The reported results are time-series averages of cross-sectional statistics. Panel A shows thesummary statistics at the firm level and panel B at the industry level. Panel C reports thevariance decomposition of the 1/AP measure and its components into within and betweenindustry variance. The number in panel C denote the percentage of total variance. Byconstruction, LroutineLtotalhas 100% variance at industry level. Industries are defined at SIC 19873-digit and NAICS 2002 4-digit level.Mean Median St. dev. Min MaxPanel A. Summary statistics at the firm level(-)AP 4.379 4.186 1.274 -1.689 9.442ln(KL)3.363 3.221 1.156 -2.428 8.142Panel B. Summary statistics at the industry level(-)AP 4.452 4.229 1.204 1.191 8.682ln(KL)3.413 3.248 1.064 0.306 7.172LroutineLtotal0.407 0.407 0.187 0.043 0.814Panel C. Decomposition of variance into within and betweenindustry components(-)AP ln(KL)LroutineLtotalWithin industries 28.33 34.87 0Between industries 71.67 65.13 10039Table 2.2: Technology Shocks and Investment, Employment, and Income GrowthThis table reports the results of regressing contemporaneous and future changes in invest-ment, employment and income, panel A, and change in moments of income growth distri-bution, panel B, on the I-shock measure and TFP shock. The regression is of the form1K+1∑Kk=0 yt+k − yt+k−1 = α + γIshock × Ishockt + γTFP × TFPt + t+k, for time horizonsK = 0, 1, and 2. Ishockt is the I-shock measure representing the change in quality-adjustedrelative prices of capital goods between t − 1 and t. TFPt is the change in aggregate pro-ductivity between t− 1 and t. The regressions are estimated at annual frequency. The timeperiod in panel A is from 1970 to 2015 and in panel B from 1978 to 2010. *** Significant atthe 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Source: Income, investment and employment: U.S. Bureau of Economic Analysis; Momentsof income growth distribution: Guvenen et al. (2014).Panel A. Impact of technological shock on income, investment and employmentInvestment Employment Personal incomet t+1 t+2 t t+1 t+2 t t+1 t+2(1) (2) (3) (4) (5) (6) (7) (8) (9)Constant -0.026 0.038 0.065*** 0.005 0.026*** 0.026*** 0.06*** 0.122*** 0.094***(0.024) (0.028) (0.019) (0.006) (0.009) (0.006) (0.01) (0.018) (0.014)Ishock 1.296*** 0.516 -0.201 0.081 -0.184* -0.272*** -0.204* -0.472** -0.453***(0.348) (0.474) (0.298) (0.063) (0.103) (0.077) (0.113) (0.202) (0.139)TFP 2.383*** 4.775*** 2.457*** 0.662*** 1.156*** 0.748*** 0.289 0.223 0.183(0.802) (1.42) (0.732) (0.194) (0.339) (0.194) (0.425) (0.719) (0.469)R2 41.05 29.2 20.47 26.3 25.36 36.16 7.95 10.12 18.61Panel B. Impact of technological shock on distribution of income growthMean Standard deviation Skewnesst t+1 t+2 t t+1 t+2 t t+1 t+2Constant 0.024** 0.042*** 0.018** -0.008** -0.015*** -0.009** 0.168* 0.279*** 0.119**(0.011) (0.009) (0.008) (0.004) (0.005) (0.005) (0.087) (0.088) (0.052)Ishock -0.524** -0.785*** -0.319* 0.116 0.253** 0.156 -4.814** -6.545*** -2.839**(0.254) (0.167) (0.163) (0.104) (0.115) (0.102) (1.932) (1.575) (1.055)TFP 0.117 -0.487 -0.378* 0.012 -0.094 -0.069 6.594** 2.247 0.461(0.393) (0.361) (0.21) (0.165) (0.152) (0.103) (2.473) (2.594) (1.565)R2 17.33 36.2 26.11 6.4 15.82 15.98 37.66 33.51 22.3140Table 2.3: Correlation between the I-shock Measure and the Long-Short (KML) PortfoliosBased on the between- and within-industry Automation PotentialThis table reports the correlation coefficients between the I-shock measure and the long-shortportfolio (KML) constructed by the AP measure based on within or between industry vari-ations. The long-short portfolios in the ’Between’ row are constructed by ranking industriesby the industry-level average share of displaceable labor, AP. The long position consists of in-dustries in the bottom quintile ranked by the AP measure and the short position of industriesin the top quintile. The long-short portfolios in the ’Within’ row are constructed by rankingindividual firms by firms’ deviation from the industry average share of displaceable labor.The long position consists of firms in the bottom quintile ranked deviation from industry’saverage AP and the short position of firms in the top quintile. Industries are defined at SIC4-digit level in column 1, SIC 3-digit level in column 2, NAICS 6-digit level in column 3, andNAICS 5-digit level in column 4.(1) (2) (3) (4)SIC 4 SIC 3 NAICS 6 NAICS 5Between industries 0.586 0.583 0.545 0.529Within industries 0.266 0.235 0.24 0.31241Table 2.4: Exposure of Portfolio Returns to Technology Shocks (I-shocks)This table reports the results from regressing portfolio annual excess returns on a constant,the I-shock measure and return factors. Regressions in panel A include a constant and theI-shock measure. The aggregate market return (CRSP value-weighted index) is added to theregressions in panel B, and the size (SMB) and value (HML) factors of Fama and French(1993) are added to the regressions in panel C. Stocks are assigned to one of five portfoliosbased on the share of displaceable labor, AP. Portfolios are rebalanced annually at the end ofJune. Portfolio returns are value-weighted. The sample period is from 1970 to 2015. Neweyand West (1987) standard errors are in parentheses. *** Significant at the 1 percent level.** Significant at the 5 percent level. * Significant at the 10 percent level.High AP 2 3 4 Low APLow-High APKMLPanel A. Exposure to I-shock measureIshock -3.381*** -2.714*** -2.443*** -1.957** -1.04 2.341***(0.952) (0.685) (0.811) (0.809) (0.808) (0.466)R2 29.19% 28.21% 20.53% 15.31% 4.43% 30.15%Panel B. Exposure to I-shock measure and aggregate marketIshock -1.646*** -1.197*** -0.691*** -0.264 0.725*** 2.37***(0.396) (0.338) (0.246) (0.226) (0.258) (0.453)Market 0.955*** 0.835*** 0.965*** 0.932*** 0.971*** 0.016(0.091) (0.069) (0.036) (0.036) (0.058) (0.119)R2 81.79% 88.47% 92.74% 93.67% 91.63% 30.18%Panel C. Exposure to I-shock measure, aggregate market,and factors of Fama and French (1993)Ishock -1.365*** -1.068*** -0.751*** -0.386* 0.564** 1.929***(0.228) (0.264) (0.249) (0.21) (0.252) (0.314)Market 0.94*** 0.861*** 0.934*** 0.905*** 0.959*** 0.02(0.047) (0.041) (0.035) (0.03) (0.036) (0.059)SMB 0.526*** 0.117 0.014 -0.102** -0.22*** -0.746***(0.098) (0.074) (0.085) (0.046) (0.059) (0.142)HML 0.033 0.109 -0.102 -0.109 -0.079 -0.112(0.137) (0.118) (0.073) (0.065) (0.082) (0.152)R2 89.96% 89.62% 93.08% 94.7% 94.31% 66.84%42Table 2.5: Properties of Portfolios Sorted on βKMLThis table reports the results from regressing portfolio annual excess returns on a constant andthe I-shock measure in panel A and time-series averages of additional characteristics acrossthe stocks within each portfolio in panel B. Stocks are assigned to one of five to portfoliosbased on their exposure to the KML portfolio, βKMLi . Portfolios are rebalanced annuallyat the end of June. Portfolio returns are value-weighted. Newey and West (1987) standarderrors are in parentheses. (-)AP is the negative of firm’s displaceable labor share measuredby (−1)× ln(LroutineK), BM is the book-to-market ratio, βmarket is the regression coefficientof market excess return from rolling time-series regressions of firm excess return onto marketexcess return and a constant, CF denotes cash flow, Lev denotes financial leverage, PCMdenotes the price-to-cost margin, Profitability is the ratio of sales minus cost of goods sold,interest expenses, and selling, general, and administrative expenses to book equity, Size isthe natural logarithm of the market capitalization in thousands, and Turnover is the fractionof shares traded to the total shares outstanding. The sample period is from 1975 to 2015.*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant atthe 10 percent level.Low βKML 2 3 4 High βKMLPanel A. Exposure to I-shock, postranking betasIshock -2.319** -2.495*** -1.744*** -1.701*** -1.18(1.109) (0.888) (0.606) (0.582) (0.754)R2 14.22% 25.4% 13.05% 12.72% 6.32%Panel B. Additional properties(-)AP 4.047 4.068 4.099 4.187 4.623BM 0.677 0.664 0.668 0.655 0.619βmarket 1.431 1.166 1.077 1.052 1.180CF 0.089 0.095 0.096 0.094 0.092Lev 0.330 0.322 0.314 0.316 0.310PCM 0.310 0.321 0.317 0.319 0.331Profitability 0.249 0.256 0.255 0.247 0.230Size 5.576 5.909 6.063 6.119 6.240Turnover 0.109 0.086 0.080 0.081 0.10143Table 2.6: Mean Portfolio Returns and Alphas of Portfolios Sorted on βKMLThis table reports the time-series averages of value-weighted portfolio excess returns in panelA, equally-weighted portfolio excess returns in panel B, results of regressing monthly value-weighted portfolio excess returns on a constant and market excess return in panel C andresults of regressing monthly value-weighted portfolio returns on a constant, market excessreturn, and the size (SMB) and value (HML) factors of Fama and French (1993) in panelD. Newey and West (1987) standard errors are reported in parentheses. Stocks are assignedto one of five portfolios based on their exposure to the KML portfolio, βKML, is estimatedfrom rolling regressions of stock’s monthly excess return on a constant and monthly returnof the KML portfolio over the last 60 months. Portfolios are rebalanced annually at the endof June. The average excess returns and standard errors in panel A and B are annualizedaverages of monthly excess returns. The alpha estimates and their standard errors in panelC and D are annualized. The sample period is from 1975 to 2015. *** Significant at the 1percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Low βKML 2 3 4 High βKMLHigh-LowβKMLPanel A. Value-weighted portfolio excess returnExcess return 12.45*** 11.54*** 10.03*** 7.45*** 5.6** -6.85***(3.2) (2.76) (2.66) (2.65) (2.65) (2.14)Panel B. Equally-weighted portfolio excess returnExcess return 12.95*** 12.57*** 11.1*** 10.45*** 8.66*** -4.29**(3.48) (3.16) (2.96) (2.94) (3.24) (1.7)Panel C. CAPMAlpha 2.966* 2.985** 2.028** -0.79 -2.423** -5.389**(1.745) (1.419) (0.96) (0.817) (1.188) (2.475)Market 1.16*** 1.047*** 0.979*** 1.009*** 0.981*** -0.179*(0.06) (0.044) (0.036) (0.017) (0.054) (0.104)R2 78.48% 85.02% 86.28% 90.22% 80.17% 3.56%Panel D. Three-factor model of Fama and French (1993)Alpha 2.263 2.304** 1.737* -0.42 -1.385 -3.648*(1.584) (1.015) (0.968) (0.791) (0.942) (1.989)Market 1.11*** 1.036*** 0.989*** 1.014*** 0.974*** -0.137(0.054) (0.035) (0.034) (0.015) (0.046) (0.084)SMB 0.312*** 0.144*** -0.001 -0.075** -0.106** -0.419***(0.108) (0.04) (0.041) (0.031) (0.046) (0.123)HML -0.001 0.068 0.056 -0.038 -0.152** -0.151(0.116) (0.075) (0.052) (0.039) (0.074) (0.168)R2 80.82% 85.79% 86.39% 90.46% 81.27% 12.35%44Table 2.7: Panel Regressions of Annual Stock Returns on βKML and Firm CharacteristicsThis table reports the results of regressing annual stock returns on a constant, βKML, other stocks characteristics and year fixedeffect as indicated. Standard errors are clustered at the firm level and reported in parentheses. βKML is estimated from rollingregressions of stock’s monthly excess return on a constant and monthly return of the KML portfolio over the last 60 months. BMis the book-to-market equity ratio, βmarket is the regression coefficient of market excess return from rolling time-series regression offirm excess return onto market excess return and a constant, CF denotes cash flow, Lev denotes financial leverage, PCM denotesthe price-to-cost margin, Profitability is the ratio of sales minus cost of goods sold, interest expenses, and selling, general, andadministrative expenses to book equity, Size is the natural logarithm of the market capitalization in thousands, and Turnover isthe fraction of shares traded to the total shares outstanding. The sample period is from 1975 to 2015. *** Significant at the 1percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)βKML -3.829∗∗∗ -1.599∗∗∗ -1.531∗∗∗ -2.032∗∗∗ -1.544∗∗∗ -1.690∗∗∗ -1.653∗∗∗ -1.444∗∗∗ -1.902∗∗∗ -1.558∗∗∗ -2.106∗∗∗(0.249) (0.296) (0.307) (0.297) (0.292) (0.297) (0.297) (0.324) (0.299) (0.295) (0.333)BM 1.207∗∗∗ 2.005∗∗∗(0.314) (0.393)Beta -4.728∗∗∗ -2.649∗∗∗(0.331) (0.399)CF 18.39∗∗∗ 11.01∗∗(3.563) (4.736)Lev -4.471∗∗∗ -4.495∗∗∗(0.881) (1.032)PCM 3.279∗∗∗ 4.033∗∗∗(1.099) (1.244)Profitability 0.397∗∗ 0.360∗∗(0.162) (0.148)Size 0.799∗∗∗ 0.909∗∗∗(0.118) (0.139)Turnover -29.38∗∗∗ -23.80∗∗∗(2.320) (2.576)Year fixed effect No Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 57959 57959 54810 57959 57906 57769 57958 49348 57959 57959 49234R2 0.00528 0.154 0.154 0.158 0.156 0.155 0.154 0.157 0.155 0.159 0.16645Table 2.8: Mean Portfolio Returns for Conditionally Double-Sorted PortfoliosThis table reports the average excess returns for conditionally double-sorted portfolios. In the first step, I sort stocks into threebaskets based on the control variable indicated in each column. In the second step, I sort stocks within each of these three basketsinto three additional baskets based on βKML resulting in nine portfolios in total. I then create three portfolios by pooling thestocks in the βKML-based baskets with the same rank. I report the annualized average value-weighted returns for portfolios withconditionally low-, medium-, and high-βKML stocks as well as for the High-Low βKML portfolio. Newey and West (1987) standarderrors are in parentheses. βKML is estimated from rolling regressions of stock’s monthly excess return on a constant and monthlyreturn of the KML portfolio over the last 60 months. BM is the book-to-market equity ratio, βmarket is the regression coefficient ofmarket excess return from rolling time-series regression of firm excess return onto market excess return and a constant, CF denotescash flow, Lev denotes financial leverage, PCM denotes the price-to-cost margin, Profitability is the ratio of sales minus cost ofgoods sold, interest expenses, and selling, general, and administrative expenses to book equity, Size is the natural logarithm of themarket capitalization in thousands, and Turnover is the fraction of shares traded to the total shares outstanding. The first columnshows average excess return for unconditional sorting based on βKML. The last column reports average excess return for sortingon BM conditional on Size. The sample period is from 1975 to 2015. *** Significant at the 1 percent level. ** Significant at the 5percent level. * Significant at the 10 percent level.Uncond BM βmarket CF Lev PCM Profitability Size TurnoverBMcondSize(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Low βKML 12.72 10.83 10.53 10.68 10.41 11.02 10.78 9.64 10.24 6.22 9.56 7.42 7.26 7.99 8.38 7.84 8.43 7.7 8.39 7.33High βKML 6.12 5.6 5.94 5.44 5.42 5.66 5.35 5.43 5.53 8.35High-Low βKML -6.6*** -5.23*** -4.59*** -5.24*** -5.0*** -5.36*** -5.43*** -4.21*** -4.7*** 2.15(1.83) (1.51) (1.55) (1.64) (1.55) (1.64) (1.61) (1.59) (1.56) (1.81)46Table 2.9: Long-Horizon Predictability of the Return PremiumThis table reports the results of forecasting regressions of the y-year cumulative annual logreturn premium of the Low-High βKML portfolio on the I-shock measure in panel A. Panel Bshows the results of forecasting regressions that include also the aggregate economic growth(GDP), a proxy for the consumption-wealth ratio (cay) of Lettau and Ludvigson (2001),and the price-earnings ratio (CAPE P/E10) of Campbell and Shiller (1988). The horizon ofthe cumulative log return is indicated in columns. Low-High βKML portfolio consists of along position in stocks with low βKML and a short position in stocks with high βKML. Theportfolio is rebalanced annually at the end of June. The portfolio returns are value-weighted.Newey and West (1987) standard errors are reported in parentheses and Hansen and Hodrick(1980) standard errors in brackets. The sample period is from 1970 to 2015. *** Significant atthe 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Source: The data for aggregate economic growth: U.S. Bureau of Economic Analysis; theproxy for the consumption-wealth ratio: Martin Lettau’s website; the price-earnings ratio:Robert Shiller’s website.1-1 1-2 1-3 1-4 1-5Panel A. Predicting the Low-High βKML portfolio returns by I-shockConstant 0.05 0.056 0.119 0.216 0.319(0.041) (0.084) (0.104) (0.11)* (0.116)***[0.041] [0.091] [0.113] [0.117]* [0.107]***Ishock 0.722 2.413 2.731 2.474 2.058(0.738) (1.334)* (1.599)* (1.574) (1.6)[0.586] [1.199]* [1.526]* [1.374]* [1.232]R2 2.5% 13.9% 13.42% 8.54% 5.35%47Table 2.9 Long-Horizon Predictability of the Return Premium: Continued1-1 1-2 1-3 1-4 1-5Panel B. Predicting the Low-High βKML portfolio returns by I-shock andadditional predictorsConstant 0.12 0.132 0.128 0.172 0.247(0.062)* (0.114) (0.152) (0.171) (0.191)[0.063]* [0.132] [0.175] [0.182] [0.189]Ishock 1.327 3.062 2.822 2.151 1.544(0.924) (1.5)** (1.615)* (1.501) (1.45)[0.834] [1.307]** [1.289]** [0.769]*** [0.98]GDP -2.453 -1.584 -0.068 2.022 2.95(1.257)* (1.5) (1.528) (1.897) (1.72)*[1.302]* [1.455] [1.332] [1.223] [1.379]**cay 0.647 0.594 -0.384 -1.955 -3.365(1.141) (2.013) (2.281) (2.056) (2.035)[1.18] [2.007] [1.874] [1.254] [1.569]**CAPE P/E10 -0.001 -0.003 0 0 0.001(0.003) (0.004) (0.006) (0.007) (0.007)[0.002] [0.004] [0.007] [0.007] [0.008]R2 14.88% 17.9% 13.56% 12.26% 13.84%48Table 2.10: Firms’ One-Year Response to Technology ShocksThis table reports the results of regressing firm-level percentage changes in employment,capital and capital-labor ratio in panel A, and sales, percentage point changes in return oncapital and return on equity in Panel B, between years t and t + 1 on the I-shock measurein year t, Ishockt, firm i′s exposure to technology shock in year t measured by the deciles’index, IβKMLi,t , of cross-sectionally ranked βKMLi,t , and their interaction term, Ishockt×IβKMLi,t .The regressions include aggregate TFP shock, TFPt, and deterministic trend, Trendt, asaggregate control variables and firm i′s capital, Ki,t, age, Agei,t, market capitalization, Sizei,t,lagged dependent variable for one-year horizon t − 1 to t, and firm fixed effects as firm-specific control variables. Columns βKML ≤ median and βKML > median show results ofthe regressions without the interaction term estimated for firms with below and above medianβKML, respectively. Standard errors are clustered at SIC 4 digit industry level and reportedin parentheses. The sample period is from 1975 to 2015. *** Significant at the 1 percentlevel. ** Significant at the 5 percent level. * Significant at the 10 percent level.Panel A. Changes in employment, capital and capital-labor ratioEmployment Capital Capital-labor ratioFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianIshockt -0.830∗∗∗ -0.560∗∗∗ 0.008 1.038∗∗∗ 1.138∗∗∗ 1.100∗∗∗ 2.047∗∗∗ 1.837∗∗∗ 1.174∗∗∗(0.185) (0.146) (0.106) (0.135) (0.103) (0.110) (0.170) (0.141) (0.130)IβKMLi,t -0.039 0.080 0.032 -0.080∗ -0.011 -0.130 -0.063 -0.204∗∗ -0.209∗∗(0.043) (0.102) (0.092) (0.043) (0.092) (0.106) (0.039) (0.094) (0.101)Ishockt × IβKMLi,t 0.102∗∗∗ 0.002 -0.110∗∗∗(0.031) (0.027) (0.029)TFPt 1.229∗∗∗ 1.442∗∗∗ 1.070∗∗∗ 0.392∗∗∗ 0.652∗∗∗ 0.009 -1.160∗∗∗ -1.155∗∗∗ -1.307∗∗∗(0.097) (0.146) (0.131) (0.082) (0.112) (0.107) (0.097) (0.130) (0.151)Ki,t -4.999∗∗∗ -5.087∗∗∗ -5.299∗∗∗ -4.927∗∗∗ -4.993∗∗∗ -5.370∗∗∗ 0.446∗∗ 0.598∗∗ 0.086(0.223) (0.342) (0.333) (0.218) (0.324) (0.352) (0.224) (0.297) (0.347)Agei,t 0.025 -0.141 0.168 0.107 0.281∗ 0.065 0.029 0.258∗∗ -0.118(0.085) (0.137) (0.190) (0.071) (0.148) (0.220) (0.065) (0.110) (0.104)Sizei,t 0.028∗∗∗ 0.049 0.023∗∗ 0.035∗∗∗ 0.064 0.031∗∗∗ 0.016∗∗∗ 0.021 0.016∗∗∗(0.010) (0.033) (0.010) (0.010) (0.039) (0.010) (0.005) (0.015) (0.006)Trend 0.078 0.199 -0.006 -0.164∗∗ -0.374∗∗ -0.079 -0.287∗∗∗ -0.519∗∗∗ -0.122(0.086) (0.137) (0.193) (0.074) (0.145) (0.225) (0.067) (0.108) (0.111)Firm FE Yes Yes Yes Yes Yes Yes Yes Yes YesN 44109 22515 21607 44109 22515 21607 44109 22515 21607R2 0.234 0.309 0.285 0.311 0.397 0.337 0.147 0.217 0.21149Table 2.10 Firms’ One-Year Response to Technology Shocks: ContinuedPanel B. Changes in sales, return on capital, and return on equitySales Return on capital Return on equityFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianIshockt -1.420∗∗∗ -1.364∗∗∗ -0.591∗∗∗ -0.872∗∗∗ -0.786∗∗∗ -0.515∗∗∗ -1.767∗∗∗ -1.509∗∗∗ -0.816∗∗∗(0.275) (0.237) (0.179) (0.086) (0.071) (0.109) (0.194) (0.157) (0.202)IβKMLi,t -0.131∗∗ 0.083 -0.210 -0.004 0.024 -0.003 0.005 0.102 -0.037(0.051) (0.104) (0.148) (0.017) (0.035) (0.051) (0.036) (0.090) (0.097)Ishockt × IβKMLi,t 0.073 0.046∗∗∗ 0.128∗∗∗(0.047) (0.016) (0.035)TFPt 2.084∗∗∗ 2.182∗∗∗ 1.862∗∗∗ 0.466∗∗∗ 0.437∗∗∗ 0.447∗∗∗ 0.783∗∗∗ 0.761∗∗∗ 0.681∗∗∗(0.171) (0.204) (0.202) (0.059) (0.058) (0.102) (0.114) (0.136) (0.182)Ki,t -4.426∗∗∗ -4.310∗∗∗ -4.684∗∗∗ -1.659∗∗∗ -1.815∗∗∗ -1.781∗∗∗ -2.796∗∗∗ -2.944∗∗∗ -3.058∗∗∗(0.236) (0.339) (0.407) (0.091) (0.133) (0.139) (0.208) (0.312) (0.338)Agei,t 0.002 -0.100 0.103 0.026 0.028 0.010 0.058 0.025 0.131(0.077) (0.151) (0.149) (0.028) (0.044) (0.026) (0.069) (0.114) (0.118)Sizei,t 0.019∗ 0.049 0.015 -0.009∗∗∗ -0.004 -0.011∗∗∗ -0.016∗∗∗ -0.023 -0.019∗∗∗(0.011) (0.034) (0.011) (0.003) (0.005) (0.004) (0.006) (0.018) (0.007)Trend -0.031 0.037 -0.114 0.100∗∗∗ 0.120∗∗∗ 0.116∗∗∗ 0.145∗∗ 0.205∗ 0.086(0.076) (0.151) (0.151) (0.029) (0.044) (0.029) (0.071) (0.115) (0.122)Firm FE Yes Yes Yes Yes Yes Yes Yes Yes YesN 44109 22515 21607 44044 22476 21581 40760 20714 20057R2 0.237 0.342 0.257 0.168 0.252 0.215 0.183 0.258 0.24250Table 2.11: Firms’ Three-Years Response to Technology Shocks, Highly Competitive versusLess Competitive IndustriesThis table reports the results of regressing firm-level percentage changes in employment andcapital in panel A, capital-labor ratio and sales in panel B, and percentage point changes inreturn on capital and return on equity in Panel C, between years t and t+ 3 on the I-shockmeasure in year t, Ishockt, and firm i′s exposure to technology shock in year t measuredby the deciles index, IβKMLi,t , of cross-sectionally ranked βKMLi,t . The regressions includeaggregate TFP shock, TFPt, and deterministic trend, Trendt, as aggregate control variablesand firm i′s capital, Ki,t, age, Agei,t, market capitalization, Sizei,t, lagged dependent variablefor one-year horizon t − 1 to t, and firm fixed effects as firm-specific control variables. Theregressions are estimated for firms with below and above median βKML in industries with highand low competition as indicated. Industries with high and low competition are industrieswith above and below average Herfindahl-Hirschman index of sales at the SIC 3 digit industrylevel. Standard errors are clustered at SIC 4 digit industry level and reported in parentheses.The sample period is from 1975 to 2015. *** Significant at the 1 percent level. ** Significantat the 5 percent level. * Significant at the 10 percent level.Panel A. Changes in employment and capitalEmployment CapitalHigh competition Low competition High competition Low competitionβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianIshockt -2.810∗∗∗ -1.730∗∗∗ -2.595∗∗∗ -2.196∗∗∗ 0.983∗∗ 1.281∗∗∗ 0.841∗∗∗ 0.538∗(0.454) (0.383) (0.333) (0.297) (0.398) (0.380) (0.315) (0.283)IβKMLi,t -0.409 0.203 0.202 0.183 -0.590∗ -0.018 0.184 -0.239(0.346) (0.261) (0.318) (0.327) (0.329) (0.397) (0.305) (0.337)Controls Yes Yes Yes Yes Yes Yes Yes YesFirm FE Yes Yes Yes Yes Yes Yes Yes YesN 8139 9785 11672 9647 8135 9782 11667 9646R2 0.593 0.544 0.547 0.548 0.648 0.557 0.600 0.55151Table 2.11 Firms’ Three-Years Response to Technology Shocks, Highly Competitive versusLess Competitive Industries: ContinuedPanel B. Changes in capital-labor ratio and salesCapital-labor ratio SalesHigh competition Low competition High competition Low competitionβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianIshockt 4.210∗∗∗ 3.398∗∗∗ 3.693∗∗∗ 2.711∗∗∗ -3.410∗∗∗ -2.800∗∗∗ -3.039∗∗∗ -2.743∗∗∗(0.557) (0.360) (0.327) (0.349) (0.642) (0.565) (0.428) (0.365)IβKMLi,t -0.175 -0.241 -0.161 -0.531∗ -0.531 -0.558 -0.192 -0.283(0.344) (0.338) (0.296) (0.319) (0.414) (0.534) (0.333) (0.374)Controls Yes Yes Yes Yes Yes Yes Yes YesFirm FE Yes Yes Yes Yes Yes Yes Yes YesN 8135 9782 11667 9646 8138 9782 11671 9644R2 0.463 0.446 0.422 0.433 0.586 0.488 0.548 0.542Panel C. Changes in return on capital and return on equityReturn on capital Return on equityHigh competition Low competition High competition Low competitionβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianβKML ≤medianβKML >medianIshockt -1.289∗∗∗ -1.016∗∗∗ -0.932∗∗∗ -0.541∗∗∗ -2.431∗∗∗ -1.911∗∗∗ -1.842∗∗∗ -0.797∗∗∗(0.141) (0.249) (0.102) (0.093) (0.319) (0.448) (0.265) (0.226)IβKMLi,t -0.233∗∗ 0.052 -0.058 0.043 -0.330 0.130 -0.005 -0.134(0.112) (0.134) (0.080) (0.110) (0.273) (0.275) (0.215) (0.264)Controls Yes Yes Yes Yes Yes Yes Yes YesFirm FE Yes Yes Yes Yes Yes Yes Yes YesN 8136 9777 11650 9634 7373 8865 10417 8731R2 0.436 0.384 0.369 0.405 0.447 0.414 0.368 0.44552Chapter 3Job Displacement throughTechnology, How TechnologyShocks Affect Households’ Incomeand Employment3.1 IntroductionDo machines create or steal jobs? Labor together with machines and other forms of capital arethe major inputs in any production process. The price of capital has decreased over time, e.g.Fisher (2006); Karabarbounis and Neiman (2013), and firms’ use of capital has increased atthe same time. Both the price and quality of capital are subject to shocks, called investmentspecific technology shocks, or IST shocks in short. Recent literature has analyzed the assetpricing implications of IST shocks, e.g. Garlappi and Song (2017); Kogan and Papanikolaou(2014), while earlier studies focused on the implications for economic growth, e.g. Greenwoodet al. (1997); Fisher (2006). Although IST shocks directly affect capital, it is not clear howthe relationship between capital and labor propagates this shock to labor. Since labor is animportant source of consumers’ income it is important to understand whether an IST shockcan be the underlying source of labor income risk and job displacement risk in the economy.I study how IST shocks affect the labor market and labor income and how it depends onthe complementarity or substitutability of capital for different types of labor. I measure thedegree of substitutability of a job by the routine intensity of the tasks performed in the job.To investigate the effects of the IST shock on the labor market variables, I utilize the cross-sectional variation of different types of labor at three different levels ranging from local labormarkets through industries to individual worker observations. I analyze the immediate effecton employment, unemployment, wages, investment, and the likelihood of job loss within oneyear after the investment shock as well as the effect over multiple years.To motivate the analysis, it is helpful to compare the trends of the use of different types ofequipment and the characteristics of employed and unemployed workers in figure 3.1. PanelA plots the equipment-per-worker indexes for various types of equipment. The use of eachtype of equipment has been constantly increasing. The information processing equipmentexperienced the strongest increase. Especially since 1980, the use of information processingequipment per worker has increased at higher speed than the use of other types of equipment.This evidence indicates a wide adoption of this type of capital over the last four decades.4040The indexes are based on current cost valuation. Indexes based on real cost valuation have even strongerincrease for each type of equipment. Since the quality-adjusted prices of information processing equipmentdecreased relative to the prices of other types of equipment, the difference between the index for this type of53Panel B plots the average routine-task intensity of employed and unemployed workers.I define the routine-task intensity in detail below. For now, it can be understood as anindex expressing, for each occupation, the importance of repetitive tasks that can potentiallybe carried out by programmable machines.41 Although the routine-task intensity of theoccupations of employed and unemployed workers was very similar and almost constant inthe 1970s, it has shown differential trends for each group since 1980. It has decreased at highspeed for employed workers indicating that the nature of occupations is changing towardless routine. Workers who experienced unemployment have generally been in more routineoccupations. The gap between these two groups has increased considerably. The trend inthe use of equipment per worker and the increasing gap in the routine-task intensity betweenemployed and unemployed workers suggests that adoption of technology in form of equipmentcapital may have interesting implications for the cross section of workers.I show that investment shocks are an important source of job displacement and laborincome risk. The impact of these shocks strongly differs between routine-intensive and non-routine types of labor. While routine-intensive labor experiences negative effects, capturedin multiple labor market variables, e.g., unemployment rate or wage growth, following aninvestment shock, non-routine labor is almost unaffected. I further document that investmentshocks are followed by higher investment in equipment, which is especially pronounced inroutine-intensive industries, which can easily replace labor by capital. This evidence suggeststhat investment shocks operate, to a large extent, through substituting capital for labor inoccupations that can easily be automated.The analysis at the level of local labor markets utilizes the geographical variation of laborroutine intensity. I apply the concept of commuting zones and metropolitan areas both ofwhich are geographic areas meaningfully defined by economic ties. I use this set up to measurethe differential response of local labor markets to the IST shock based on the routine intensityin each locale. This allows me to show how IST shocks affect different types of labor, as well aslabor income. The analysis at the level of industries is based on detailed U.S. manufacturingindustries. I use the changes in employment and in the real stock of different types of capitalto illuminate how investment shocks affect capital-labor substitution. In the third approach,I construct a longitudinal micro data set of individual worker observations. I link the routinetask intensity to each employed worker by his or her occupation and observe the employmentstatus in the subsequent year. This approach allows me to quantify the probability of losing ajob, i.e., the job displacement risk, in response to an IST shock and also how this probabilityvaries with routine intensity of the jobs.I measure IST shocks by the changes in quality adjusted relative prices of investmentgoods. This measure has a theoretical foundation, e.g., Cummins and Violante (2002); Horn-stein and Krusell (1996), and is based on the idea that productivity improvements in indus-tries producing capital goods are reflected in lower quality-adjusted prices of these goods.First, I empirically verify the link between the productivity improvements in the capitalgoods-producing industries and the prices of the capital goods. I show that all major changesin the quality adjusted relative prices of investment goods can be linked to productivityequipment and all other types is even larger. Hence, the current cost valuation represents a lower bound.41The average routine-task intensity of employed workers is based on routine-task intensity of workers’current occupations averaged across the sample of employed workers in each time period. For unemployedworkers, it is based on the routine-task intensity of the occupations in their last job averaged across the sampleof unemployed workers.54shocks in a very small number of manufacturing industries producing these capital goods.For example, the major decrease in the quality adjusted prices of capital goods in the secondhalf of the 1990s can be linked to productivity improvements in a single manufacturing in-dustry; the electronic computer manufacturing. This evidence is important as it empiricallymotivates the use of the changes in the prices of investment goods as an appropriate measureof the IST shock.I show that the routine task intensity determines how the IST shock affects all thesevariables. A positive IST shock immediately decreases employment in local labor marketswith high share of routine jobs but does not affect employment in geographical areas withlow share of routine labor. This is reflected in the local unemployment rate as it rises inareas with high share of routine labor but stays unaffected in the local labor markets whereroutine occupations constitute only a small fraction of all jobs.Changes in the unemployment rate can result from the destruction of existing jobs orthe creation of new jobs. I show that the IST shock affects the unemployment rate throughboth the job creation and job destruction rate. On aggregate, the job creation rate decreasesfollowing a positive IST shock and the job destruction rate increases. These effects change inthe cross-section of different jobs depending on the routine intensity. A positive IST shocksignificantly lowers the job creation rate in areas with high share of routine jobs. In contrast,it increases the job creation rate in areas with low share of routine jobs. This differentialeffect is especially pronounced among existing firms, while the job creation rate through newfirm entrants does not seem to be affected by the share of routine jobs in the particular area.The reaction of the job destruction rate is opposite. A positive IST shock accelerates the rateof job destruction in areas with a high share of routine jobs, while it considerably mitigatesor even stops the destruction of jobs in areas with mostly non-routine jobs. This differentialresponse to an IST shock is similar among both existing firms and firm closures.I next examine how an IST shock affects the job displacement risk of different occupations.I construct longitudinal micro data of individual workers and track the workers employmentstatus, occupation and personal characteristics over two subsequent years. This allows me toestimate the workers’ probability of job loss conditional on the workers’ employment status inthe previous year. Hence this estimation goes beyond a pure analysis of workers’ employmentstatus and it illuminates how an IST shock affects the job displacement risk from workers’perspective. I show that a positive IST shock increases the probability of job loss for workersin routine intensive occupations over the following five years. The effect of workers in non-routine occupations is mixed.From the analysis of the U.S. manufacturing industries, I observe that an IST shock op-erates through the substitution of capital for labor. A positive IST shock lowers employmentin industries whose jobs consist of routine tasks. At the same time, investment in equipmentincreases in these industries following a positive IST shock. Industries with non-routine laborare affected significantly less. The positive IST shock also increases investment in structures,but this effect is same across all industries. Taken together, these results indicate that in-dustries acquire equipment capital to perform routine tasks that had previously been carriedout by human workers. Since the overall production capacity increases following a positiveIST shock, these results provide evidence against the potentially confounding effect of out-sourcing. I also control for the industry specific productivity to rule out the concern that theeffects of IST shocks could be driven by the industries’ low productivity.55An IST shock does not need to necessarily lead to a complete substitution of equipmentfor human labor in occupations performing routine intensive tasks. Employees in routineintensive occupations may accept lower wages to avoid losing their job completely. I confirmthis intuition and show that positive IST shocks lead to a lower growth in nominal wages inlocal labor markets with high share of routine jobs starting from the second year after theshock, although the immediate response is mildly positive. Labor markets with a low shareof routine labor benefit by a faster growth in nominal wages.I next show that the effects of the IST shock spread over multiple years and have inter-esting and differential dynamics among local labor markets and industries with high and lowshare of routine labor. For example, a positive IST shock increases unemployment in thefirst three to four years. In the subsequent years the unemployment rate tends to decreaseagain. This dynamic is especially pronounced in areas with a high share of routine labor. Idocument similar results for other variables, such as job creation and destruction rates, wagegrowth and industries’ investments.These empirical results have several important implications. First, this analysis showsthat shocks to capital prices considerably shape the labor market and increase the job dis-placement risk. Cheaper machines or other forms of innovation embedded in capital goodscan destroy routine intensive jobs by substituting capital for human workers and makingthem obsolete for these jobs. On the other hand, jobs which require performing non-routinetasks are affected much less or can even slightly benefit in terms of both employment andwages. These results suggest that machines can substitute human workers in some jobs, andcomplement them in other jobs. Second, the effects on employment and wage growth indi-cate that an IST shock is a source of labor income risk even if it does not necessarily leadto a complete job displacement. Third, the relatively long time-profile of the effects on allvariables suggests that IST shock trigger changes which are spread over multiple years. Itsuggests that companies need some time to carry out the processes that are involved in thesubstitution of capital for human workers such as investment in new technology, adjustmentof the existing production structures and renegotiating or dissolving employment contracts.My paper is related to three streams of literature. Asset pricing literature has focusedon role of job displacement and labor income risk for asset prices. Constantinides and Duffie(1996) study the implication of idiosyncratic labor income shocks for aggregate asset prices.Gaˆrleanu et al. (2012) derive the asset pricing implications of displacement risk caused by in-novation in an overlapping generation model. Kogan, Papanikolaou, Schmidt and Song (2017)show that innovation is followed by increases in labor income risk and shifts in the distributionof workers earnings growth. Kogan et al. (2017) show that an unequal distribution of inno-vation benefits can generate a high equity premium. Literature on investment shocks studiesthe macroeconomic and asset pricing implications of these shocks. For example, Greenwoodet al. (1997) studies the importance of these shocks for economic growth and Greenwoodet al. (2000); Fisher (2006) examine their role in business cycle variations. Papanikolaou(2011) focuses on the asset pricing implications in a two-sector economy. Kogan and Pa-panikolaou (2014, 2013) focus on the relationship to firms’ growth opportunities and assetpricing anomalies. Garlappi and Song (2017) investigate the importance of capital utilizationand market power for the asset pricing implications of investment shocks. Garlappi and Song(2016) examine the effect on equity returns across various portfolios and time periods. Theprior literature has also focused on the changes in the characteristics and task content of56labor. Autor et al. (2003) find that, over a long horizon, computerization is associated withreduced labor input of routine tasks and increased labor input of non-routine cognitive tasks.Autor and Dorn (2013), who are the first to use the share of routine task-intensive labor incommuting zones, illuminate the long-horizon relationship between non-neutral technologicalprogress, the demand for low-skill service activities, and the polarization of employment andwages. Krusell et al. (2000) show that capital-skill complementarity is important for explain-ing the increase in the skill premium in the presence of the observed increase in the stock ofequipment and the relative quantity of skilled labor.I contribute to the above literature in two ways. First, I show that investment shocks havelarge and differential effects on different types of labor. Prior literature focused especially onthe direct effect of these shocks on capital, where investment shocks change the efficiency ofcapital investment. In such framework, investment shocks are more impactful for firms withlarge growth opportunities in a cross-section of different firms. I document large effects ofthese shocks on labor, which operate, to large extent, through capital-labor substitution. Myresults indicate that in the cross-section of firms, investment shocks are especially relevant forfirms with a high share of displaceable/automateable labor. Second, I show that investmentshocks are an important source of job displacement and labor income risk. The long-termshifts in the task content and other characteristics of labor documented in the above literatureare not smooth over short horizons. In fact they are associated with disturbances such ashigher unemployment and lower wage growth and create substantial risk for households.Importantly, these changes seem to be driven by the dynamics of investment shocks withimmediate but also long-lasting impact. These results are relevant for asset pricing literaturesince they show that investment shocks systematically affect both, households income andfirms investment and employment decisions.In the next section, I describe the construction and characteristics of the measure ofan IST shock, routine task intensity and the data sets. In section III, I outline a simpletheoretical model and derive the empirical implications. I test the empirical predictions overa one year horizon in section IV and over multiple years horizon in section V. I conclude insection VI.3.2 Data and Measures3.2.1 Measures of Investment ShocksI measure the IST shock by the changes in the quality adjusted relative price of invest-ment goods42. The theoretical foundation of this measure is provided in e.g. Greenwoodet al. (1997); Cummins and Violante (2002); Hornstein and Krusell (1996). Intuitively, theinvestment-specific technology shock allows to produce new equipment at lower cost or in-creases its quality.I formally describe this intuition in a simple two-sector model from Cummins and Violante(2002). Final goods producers competitively produce final goods xt at price pct . The finalgoods can be either consumed or used as an input for production of investment goods. The42A similar measure has been successfully used in previous studies focusing on the effects of the investment-specific technology shock on economic growth Greenwood et al. (1997) and business cycle Greenwood et al.(2000); Fisher (2006)57investment goods sector can produce it efficiency units from xt units of consumption goodsaccording to the production function it = qt xt. qt captures the level of technology in theinvestment goods sector. Prices in the investment goods sector are set competitively so thatpitit = pctxt. Combining this result with the production function leads to pit/pct = 1/qt andhence ∆qt = ∆pct −∆pit. Accordingly, the changes in the prices of investment goods relativeto consumption goods measure the investment specific technology shock qt. Based on thismodel, I construct the measures of investment shocks by subtracting the annual changes in thequality-adjusted prices of investment goods from the quality-adjusted prices of consumptiongoods. Accordingly, an increase in the I-shock measure indicates a positive investment shock.Since the I-shock measure is based on quality-adjusted prices of investment goods, Iaddress the adjustment of quality changes in price indexes on the next few lines. The qualityadjustment of price indexes captures the idea of expressing prices for goods of constant quality.Both the prices and the quality of goods change over time. The increase in the horsepowerof cars and trucks and the increases in the processing power and memory of computers overlast several decades are illustrative examples of substantial changes in the quality of capitalgoods. Quality-adjusted or constant-quality price indexes try to take the changes in qualityinto consideration and express the prices of goods, e.g., equipment, adjusted for increasesor decreases in their quality. Accordingly, quality-adjusted price indexes seek to express theprice per unit of efficiency. The question of an appropriate quality adjustment of investmentgoods price indexes goes back to Griliches (1961) who used hedonic techniques to adjustfor quality improvements of some types of goods and points out that a significant part ofthe price increases can be explained by improvements in the quality of the particular goods.Gordon (1990) points out the lack of an appropriate quality adjustment of the NIPA priceindex of equipment and develops quality-adjusted price indexes for several types equipmentuntil 1983. These quality-adjusted price indexes were extrapolated by Cummins and Violante(2002) until later years. The most significant improvements in the quality of investment goodsin the last three decades were in the information processing equipment. Fortunately, the U.S.Bureau of Economic Analysis (BEA) started using hedonic-based quality-adjusted index forcomputers and peripherals in 1980s. In general, BEA has addressed the problem of qualityadjustment over the last three decades and developed price indexes which account for qualitychanges using various techniques such as hedonic regression.43I use the price index for nonresidential equipment and the price index for nondurableconsumption goods from the BEA to measure the quality-adjusted prices of equipment andconsumption, respectively. Since I focus on the time period from 1980 and the BEA hasemployed hedonic-based quality adjustment for fast improving types of equipment, a possiblebias in prices due to inappropriate quality adjustment is a minor concern. I construct twomeasures of investment shock. The first measure, price shock, is the log change in prices ofnondurable consumption goods minus the log change in prices of nonresidential equipment.This measure completely relies on the quality adjustment in the underlying price indexes. Thesecond measure, I-shock, extrapolates the quality adjustment from Cummins and Violante(2002) until 2012.I report the properties of the investment shock measures in table 3.1. I-shock measure43Wasshausen et al. (2006) provides a brief overview of the use of hedonic methods by BEA to measure realGDP; Fox and McCully (2016) provides detailed information about the price indexes for all types of equipmentand discusses the quality adjustments employed by BEA.58has higher mean and higher standard deviation than the price shock measure. Both measuresare slightly autocorrelated and their mutual correlation coefficient is almost one. The factthat both measures have almost identical dynamics but the I-shock measure has higher meanand standard deviation comes from the nature of extrapolating the quality adjustment inCummins and Violante (2002), which basically relies on the price dynamics of the underlyingprice series and some other macroeconomic variables.Panel B of table 3.1 shows that both measures of the investment shock are stronglycorrelated with the productivity (TFP) shock of the manufacturing producers of capitalgoods. The correlation coefficient of the I-shock and price shock measure is 0.418 and 0.427,respectively. At the same time both measures have almost zero correlation with the TFPshock in the manufacturing industries producing consumption goods. This result is in linewith the model above showing that the investment specific shocks originate as productivityshocks of the capital goods producers and can be observed in the changes of quality-adjustedrelative price of investment goods. This also addresses a concern of potential endogeneityof these measures. Although prices are endogenous variables, the I-shock and price shockmeasures are based on the changes in price differences between investment and consumptiongoods and their relationship to aggregate productivity is not straightforward. Both measureshave a slightly positive correlation with the aggregate GDP growth, which is consistent withthe results in Greenwood et al. (1997), and almost a zero correlation with the aggregateproductivity (TFP) shock.To illuminate the origin of the investment shock and support its theoretical foundation, Iwill show that all big investment shocks since 1980 are associated with productivity shocks ina very small number of manufacturing industries producing capital goods. The manufacturingindustries producing capital goods are a relatively small part of the manufacturing sector andeven a smaller part of the U.S. economy as their total output accounts for about 6 to 7 percentof the total U.S. output.44Figure 3.2 shows the time-series of the I-shock measure from 1980 to 2012, a time periodof the main focus in this paper. I standardize the I-shock measure to mean zero and standarddeviation of one to facilitate the interpretation of the particular shocks. The I-shock measureis quite volatile and records four investment shocks of large scale since 1980. Three rela-tively short but strong negative investment shocks occurred in years 1981-1983, 1985-1986and 2008-2009. A prominent and long-lasting positive investment shock occurred in years1997-2001. It might be helpful to emphasize that these negative investment shocks indicatea slowdown in the secular decrease in the relative (quality-adjusted) prices of capital goods,or a mild increase. As mentioned earlier, the I-shock measure exhibits a mean of 4.123 anda standard deviation of 2.937. Hence, when the normalized I-shock measure hits approxi-mately zero in 1993 in the figure 3.2, the relative quality-adjusted prices of capital goodsdecreased approximately by the average 4.123 percent. The negative shock in 1982 indicatesan increase in the relative quality-adjusted prices of capital goods by 1.46 percent.45 In orderto understand the origin of these shocks, I report the productivity shocks, the changes in44This estimate is based on shipments of the producers of capital goods in manufacturing sector to final usesin 2010. The shipments of the capital goods producers to fixed investments comprise about one third of thetotal shipments of the whole manufacturing sector. Manufacturing sector itself accounts for approximately 19percent of the U.S. total output in 2010.45This estimate can be obtained when the negative shock in 1982 of -1.9 standard deviation is multipliedwith the estimate of standard deviation, 2.937, and added to the long-term mean, 4.123.59shipment prices and changes in real shipments of various manufacturing industries duringthe corresponding years in table 3.2. I quantitatively describe these shocks in the remainderof this sub-section to address the questions whether these shocks appear to be driven bypotential demand shocks.The first investment shocks marks a negative shock which was reflected in a significantincrease in the quality adjusted relative prices of equipment in the period from 1981 to 1983.On the standardized scale, the I-shock measure recorded a cumulative decrease by 2.35 stan-dard deviations over two years. This shock is associated with a negative productivity shockin the manufacturing industries that produce capital goods. Although the manufacturingindustries exhibit productivity increase of 1.48 percent on average, the productivity of thecapital goods producers decreased by 1.49 percent as shown in panel A of table 3.2. Atthe same time the consumption goods producers increased their productivity by 2.5 percent.The TFP shocks is reflected in the shipment prices as the prices of capital goods shipmentsincreased by almost 6 percent while the prices of shipments from consumption goods pro-ducers grew by only 4.17 percent. The correlation coefficient between the TFP shock andgrowth in the shipment prices across the manufacturing industries is -0.5 during this time pe-riod indicating strong relationship between the productivity shocks and the shipment prices.The lower growth in the total shipments of capital goods producers relative to consumptiongoods producers alleviates the concern that the increase in shipment prices could be drivenby shocks to the demand for capital goods. During the second investment shock, the I shockmeasure decreased by 1.62 standard deviation. Similarly to the first shock, the productivityof the capital goods producers decreased by almost -0.4 percent while the consumption goodsproducers exhibited productivity increase as shown in panel B.The third event captures the massive investment shock from 1997 to 2001. During thistime period, the standardized measure of I-shock increased by 5.71 standard deviations. Thisincrease was driven mainly by a large drop in prices of information processing equipment.The prices of information processing equipment decreased by 40 percent, while prices of in-dustrial, transportation and other equipment recorded growth in the range of 3 to 6 percent.Since the information processing equipment accounted for approximately one third of thetotal private investments during this time period, the aggregate price index of equipment de-creased by 15 percent. At the same time, the prices of personal expenditures for nondurablegoods increased by 8 percent.46 The massive drop in the prices of information processingequipment can be traced back to a significant productivity growth in a single manufacturingindustry producing this type of equipment. The capital goods producers recorded an averageincrease in productivity by 19.61 percent from 1997 to 2001, while the average productivitygrowth in the whole manufacturing sector was only 5.33 percent. The producers of consump-tion goods recorded even lower productivity growth of 4.17 percent. The productivity growthof the capital goods producers was reflected in both lower shipment prices and higher totalshipments. The prices of capital goods shipments decreased by 9.68 percent while the totalreal shipments increased by 66.29 percent. At the same time, the shipments of the consump-tion goods producers grew by only 7.05 percent and their prices stagnated. The tremendousproductivity growth of the capital goods producers was driven solely by electronic computermanufacturing industry, which recorded an increase in productivity by 216 percent and a46I use the U.S. Bureau of Economic Analysis price indexes to calculate the growth in prices of equipmentand nondurable goods.60decrease in shipment prices by 83 percent. The remaining capital goods producers actuallyrecorded negative productivity growth of -7.58% and stagnating prices. This event providesa strong evidence that a large-scale investment shock can originate from a productivity shockin a single industry producing a specific type of capital goods.The fourth significant investment shock, which occurred in 2008-2009, coincides with thegreat recession. Even in this case, however, the investment shock can be traced back todifferential productivity shocks between capital goods and consumption goods manufactur-ers. During this time period, the manufacturers of capital goods suffered the most severenegative productivity shock in the manufacturing sector. Their productivity dropped by 8.17percent, while the productivity of the consumption goods producers decreased only by 5.72percent. These different productivity shocks were also reflected in the prices of shipments.The shipment prices of consumption goods producers decreased by more than 12 percent,but the prices of capital goods shipments stagnated.While the empirical evidence above suggests that the I-shock measure correctly capturesthe investment shock as defined in the theoretical model, other economic forces may alsoplay a role. The most prominent alternative is an investment shock due to changes in importprices of investment goods. This mechanism is outside of the model described above. Since theBEA incorporates the import price indexes when calculating the price indexes for investmentgoods, the I-shock measure is taking this source of investment shock into account.3.2.2 Routine Task IntensityA crucial aspect in my analysis is the relationship between labor and capital. Some types oflabor are more substitutable by capital, while other types of labor have more complementaryrelationship to capital. Following Autor and Dorn (2013), I measure the substitutability oflabor by capital at the level of job occupations by the routine task-intensity measure (RTI).The idea of this measure is to describe each occupation in terms of how much it is involved inperforming routine vs. non-routine tasks. Routine tasks can be performed by following a setof precise rules and hence they can easily be performed by a machine. Computer NumericalControl (CNC) machines are typical example of machines that are taking over a large partof routine tasks previously performed by various manufacturing operators. In contrast, non-routine tasks cannot be easily described by a precise set of rules and they require high degreeof creativity and improvisation. Occupation that predominantly consists of very routine tasks(i.e. high routine task-intensity) have a high potential to be substituted by capital.Since I focus on the relationship between capital and labor at various levels of observationsranging from individual workers to local geographical areas and manufacturing industries, Iuse two RTI measures, each of which is suitable for a different level of observations. I usethe RTI measure from Autor and Dorn (2013) to measure the routine task intensity atthe individual worker and local geographical area level. This RTI measure is based on the1977 Fourth Edition of the Dictionary of Occupational Titles (DOT) matched to the CensusOccupational Categories (COC). It measures the use of routine, manual and abstract tasksin each occupation. The RTI measure is then calculated asRTIk = ln(TRk)− ln (TMk )− ln (TAk ) , (3.1)where TRk , TMk and TAk are, respectively, the routine, manual and abstract task inputs for eachoccupation k. By this construction, occupations for which routine tasks play an important61role have high RTI value while occupations consisting mostly of abstract or manual taskshave low RTI value. This measure is calculated for 330 time-consistent occupations whichcover all COC occupations.47To measure the routine task intensity at individual workers level, I match this RTI measureto the Current Population Survey (CPS) sample using appropriate crosswalks. To transformthe routine task intensity from occupational level to an appropriate measure at the levelof geographical units, I follow the approach of Autor and Dorn (2013). First, I rank theoccupations by their RTI score and identify the top employment-weighted third routine task-intensive occupations as of 1980. Then, for each geographical unit I calculate the routineemployment share asRSHj =(K∑kLjkt × 1[RTIk > RTI661980])( K∑kLjkt)−1, (3.2)where Ljkt denotes the employment in occupation k in geographical unit j at time t and1[RTIk > RTI661980]indicates the routine intensive occupations. I calculate this measure foreach decade based on the occupational employment counts from the most recent Census.I construct the second RTI measure based on the occupational descriptions from theO*NET database. O*NET is the successor of the DOT and provides a standardized anddetailed description of a large number of occupations covering the U.S. economy. Similarly tothe approach above, I focus on occupational characteristics that describe the routine, abstractand manual tasks and combine their importance according to equation (3.1). I provide a fulllist of the characteristic for each of these three categories as well as the underlying databaserelease in the appendix. A significant advantage of my new RTI measure based on theO*NET data over the old DOT-based measure is the detailed occupational classification. TheO*NET database is based on the Standard Occupational Classification (SOC) and allows tocalculate the RTI measure for 740 detailed occupations, which is more than twice as many asfor the DOT-based measure.48 This detailed and comprehensive occupational classificationallows to transform the routine task-intensity measure from the occupational level to thelevel of manufacturing industries classified at the 4-digit 1987 level without a significant lostof the variation across the industries. I calculate the routine task-intensity, RTIi, for eachmanufacturing industry i as the employment-weighted average:RTIi =(K∑kRTIk ∗ Li,k)L−1i , (3.3)where RTIk is the routine task-intensity of occupation k, Li,k is the employment in occupationk in industry i and Li is the total employment in industry i.Although the DOT-based and O*NET-based RTI measures are each constructed from adifferent data source, they identify the routine intensity in a comparable way. When aggre-gated up to the common occupational classification the measures have a positive correlationcoefficient above 0.5. Routine intensive occupations identified by each of these measures47The Census classification (COC) changed over time and the number of occupations varied. Autor andDorn (2013) create a system of 330 time-consistent occupations which cover all occupations in the COCbetween 1980 and 2000.48The total number of detailed 2000 SOC detailed occupations in O*NET 14.0 is 749.62strongly overlap. The overlap is strong also for occupations which are identified as non-routine by each of these measures. The fact that these measures are constructed from differ-ent data but still succeed to identify the occupational routine task intensity in a very similarway increases the validity of the measures.3.2.3 Geographical AreasI use two types of geographical units, commuting zones and metropolitan statistical areas(MSA). In order to properly utilize the geographical variation in routine intensity of labor andother economic variables, the geographical areas should represent local labor markets basedon economic ties. Commuting zones and MSAs adhere to this requirement. The MSAs consistof core urbanized areas with at least 50,000 population and adjacent counties with strongsocio-economic ties to the core based on commuting ties. Although the MSAs do not spreadacross the total area of United States, they cover the high population areas with adjacentcounties and represent over 80% of the total population of U.S. A potential issue are thechanges in the boundary delineations of MSAs over time, which respond to the developmentand urbanization in the regions. This problem does not apply to the Business DynamicsStatistics data used in this paper since U.S. Census constructs this data from various datasources based on the 2009 definition of MSA by U.S. Office of Management and Budget forall years. The commuting zones were originally proposed by Tolbert and Killian (1987);Tolbert and Sizer (1996) and later used by Autor and Dorn (2013); Dorn (2009). Commutingzones are created by bundling counties with strong commuting ties into zones such that thecommuting ties are maximized within each zone and minimized between the zones.I use the unemployment data reported in the Local Area Unemployment Statistics (LAUS)by the U.S. Bureau of Labor Statistics. This data is reported for individual counties at annualfrequency from 1990 to 2015. The estimates consist of annual averages of the number ofemployed and unemployed persons as well as the total number of persons in the labor force.I match the records for individual counties to commuting zones and MSAs and calculate theunemployment rate in these geographical units. For the wage and employment, I use theregional data on local personal income and employment provided by the Bureau of EconomicAnalysis. I use the county level data from 1980 to 2015 and match it to commuting zones andMSAs. The wages and salaries are recorded by the place-of-work principle and represent thewages and salaries accrued during a year. The wage and salary employment measures full-time and part-time jobs in each area by place of work for which wages and salaries are paid.The employment estimates are annual averages of monthly observations for the year. Thejob creation and job destruction data is from the U.S. Census Bureau’s Business DynamicsStatistics. I use the job creation and destruction data for the new, incumbent and closedestablishments at the level of metropolitan statistical areas from 1980 to 2015. This data ismeasured in March-to-March cycle.I report the summary statistics for MSAs in Table 3.3. Panel A shows that there isconsiderable cross-sectional variation in the share of routine task-intensive labor (RSH).The cross-sectional standard deviation is between 2.55 and 3.82 percent. The average RSHis quite persistent over time and stays close to 30 percent in all Census years.49 The individual49By construction, the employment-weighted average RSH should be exactly 33 percent in 1980. The smalldifference arises from the fact that the summary statistics is based on simple average and that MSA do notcover the entire U.S. population63MSAs, however, changed their share of routine task-intensive labor between the Census years.The standard deviation of the changes in RSH ranges from 1.52 to 2.4 percent. Panel B showsthe summary statistics for labor market variables. The average changes in unemploymentrate in MSAs are close to zero, but their standard deviation is 1.16%. The average growth intotal and average nominal wages is 5.06 and 3.47 percent, respectively. The job creation rateis higher for incumbents than for new firms and similar pattern holds between incumbentsand closed establishments for the job destruction rate.3.2.4 Individual Workers DataI construct my micro-data sample from the IPUMS Current Population Survey (CPS) datafrom 1980 to 2015. The CPS is a monthly U.S. household survey conducted by the U.S. CensusBureau and the Bureau of Labor Statistics to measure unemployment, income as well as otherindividual and household characteristics. Specifically, I use the Annual Social and EconomicSupplement (ASEC) of the CPS which is conducted alongside the March CPS every year.ASEC provides a rich set of variables including the individual income, employment status,occupational classification and demographic characteristics referring to the year prior to thesurvey. Based on 4-8-4 sampling scheme of the CPS, each selected household is interviewedfor initial four months, not interviewed in the subsequent eight months and added back tothe sample for the final four months. According to this scheme, every selected householdis included in the interviewed sample in the same months in two subsequent years. I usethis design and match the records on individuals in two subsequent years to create a paneldata set. This longitudinal feature of the data set allows me to estimate the probability oflosing job, i.e. the job displacement risk, following an investment shock at individual workerlevel conditional on the worker’s employment status. Summary statistics for this sample arereported in table 3.4.3.2.5 Manufacturing IndustriesI investigate the effects of the investment shock at the industry level using data for the U.S.manufacturing industries from the NBER-CES Manufacturing Industry Database (Bartles-man and Gray (1996); Becker et al. (2013)). This data set is a balanced panel of 457 man-ufacturing industries defined at the 4-digit 1987 SIC codes. To keep consistency with theother data sets, I focus on the time period from 1980 to 2011. During this time period themanufacturing sector produced between 33 and 18 percent of the U.S. total output.The industry level data allows me to extend the analysis of the investment shock in severaldimensions. First, the data set contains annual observations of each industry’s employment,real equipment and real structures. Comparing the responses of these variables to the invest-ment shock can help illuminate the mechanism behind the effects of the investment shock.I will show below that each of these variables react to the investment shock differently andthat the reactions vary in the cross-section of industries dependent on industry’s routine taskintensity. Second, the industry level analysis provides also evidence against potentially con-founding effects of outsourcing. Third, I control for the factor-neutral (TFP) shock at the levelof each industry. This addresses the potential concern that the investment shocks coincidewith factor-neutral productivity shocks among routine task-intensive industries. Summarystatistics for the U.S. manufacturing industries are reported in table 3.5.643.3 Model and ImplicationsI derive the formal implications of the investment shocks for capital and different typesof labor in a simple model of an economic unit i such as an industry or a geographicalarea with a production facility. Since I focus on the effects of investment shock operatingthrough the capital-labor substitution, I use a production function with a constant elasticityof substitution (CES). This type of a production function allows to model various degrees ofcapital-labor substitutability in a concise way. The economic unit has a market for a singleconsumption good with iso-elastic downward sloping aggregate demand function, Q, of theform:Qi = P1ξ−1c,i , (3.4)where Pc,i is the price of the consumption good and11−ξ is the price elasticity of demand.For simplicity, I assume that this consumption good is produced by a monopolist with a CESproduction function of the form:Yi = (Kσii + Lσii )1σi , σi < 1, (3.5)where Ki denotes capital and Li labor and11−σi is the elasticity of substitution betweencapital and labor. Accordingly a high σi indicates a high elasticity of substitution betweencapital and labor.50 I assume that capital is supplied perfectly elastically at an exogenousaggregate price Pq, which can be subject to an exogenous investment-specific technologyshock as described earlier. Labor is supplied perfectly elastically at a constant aggregatewage W . The elasticity of substitution between capital and labor implies:δlnLiδlnPq=11− σi > 0,d δlnLiδlnPqdσi=1(σi − 1)2> 0,δlnKiδlnPq< 0. (3.6)The expressions in formula 3.7 allow to state the empirical implications of the model for across-section of economic units with different σi. It is also useful to state the derivative of anoptimal labor demand w.r.t. the price of capital Pq in equilibrium as:dL∗dPq= −ξ−σσ−1L(ξ − 1)(1 +(PqW) σ1−σ)Pq(3.7)If I further impose the restriction that the price elasticity of the demand for the consump-tion good is below the elasticity of substitution between capital and labor, 11−ξ <11−σ , thederivative in the formula 3.7 becomes positive. Under this additional assumption, a decreasein the price of capital, Pq, will result in strictly lower demand for labor.This simple theoretical model provides testable empirical implications for cross sectionsof economic units with different elasticities of capital-labor substitution, 11−σi . To test theseimplications, I use a panel data sets of geographical areas and manufacturing industries.Although σi, which governs the elasticity of the capital-labor substitution, is not directly50Please note that I do not require σi to be positive which would imply that capital and labor are substitutes.65observable in the data, the natural proxy for it is the share of routine task-intensive laborin case of geographical units and the average routine task-intensity of employees in case ofmanufacturing industries. Using this set up, the model predicts that an aggregate positiveinvestment shock will:1. Decrease the employment in geographical areas with high share of routine task-intensivelabor relative to the employment in areas with low share of routine task-intensive labor.2. Decrease the employment in routine task-intensive manufacturing industries relative tonon-routine task-intensive industries.3. Increase investment in equipment in routine task-intensive manufacturing industriesrelative to their non-routine task-intensive counterparts.If I further impose the restriction that the price elasticity of the demand for consumptiongood is below the elasticity of capital-labor substitution, the model predicts also an absolutedecrease in employment in the predictions 1 and 2, regardless of the routine task intensity ofthe specific industry or the share of routine task-intensive labor of the specific geographicalarea.3.4 Empirical Results3.4.1 Local Labor MarketsI begin the empirical analysis by testing the first empirical prediction of the model. Theeffect of a positive investment shock on employment is expected to vary with the share ofroutine task-intensive labor across the geographical areas and have the most negative effectin areas with high share of routine labor. To test this prediction, I estimate panel regressionof the form:∆Yt+1,z,i = β0 + β1ISTt + β2RSHt,z,i + β3ISTt ×RSHt,z,i+ψ Trendt + γi + et+1,z,i, (3.8)where ∆Yt+1,z,i is the growth in the number of employed persons in each MSA z in state ibetween year t and t+ 1, ISTt is the investment shock measured by the standardized I-shockmeasure, RSHt,z,i is the share of routine task-intensive labor in each MSA at time t andISTt × RSHt,z,i is the interaction term between the last two variables. I also include thestate fixed effect γi and the linear time trend Trendt. The results are not sensitive to theinclusion of the time trend and I discuss its use in appendix. I estimate this regression at thelevel of MSAs from 1980 to 2012. The results are presented in table (3.6).First, I estimate regression (3.8) only with the investment shock as the independentvariable. The estimates corresponds to an average effect across the MSAs. The first columnsin panel A shows that on average a positive investment shock decreases employment. Theestimate is significant at the 5 percent level. Since a one standard deviation investmentshock decreases employment by almost 0.3 percent it has also high economic significance.This result holds when I include the share of routine labor RSHz,t. Column 4 shows theestimates for the full specification of the regression equation (3.8). The effect of investmentshock on employment varies with the share of routine labor as implied by the theoretical66model. The interaction term ISTt×RSHt,z,i is statistically significant at the 1 percent leveland has high magnitude of −6.576 percent. According to this estimate, a positive investmentshock affects the employment in the routine labor-intensive areas more severely than in areaswith low share of routine labor. In fact, employment in areas with very low share of routinelabor may even increase due to the positive coefficient of the IST shock.I also estimate the regression without the interaction term in a sub-sample of MSAs withbelow-median share of routine labor and report the estimates in column 5. Column 6 presentsthe estimates from the complementary sub-sample of MSA with above-median RSHt,z,i. Theestimates in these two columns are consistent with the estimate of the interaction term inthe fully specified regression. A positive investment shock with a magnitude of one standarddeviation significantly decreases unemployment in the above-median MSAs but does notaffect employment in the below-median MSAs. Comparing directly the coefficients of ISTtestimated from different sub-samples of MSAs in regression without the interaction term asin columns 5 and 6 five also eliminates any potential concern about multicollinearity due tothe highly time-persistent share of routine labor RSHt,z,i.I also investigate whether the effects of the investment shock on employment are reflectedin the unemployment rate. I define the dependent variable in regression (3.8), ∆Yt+1,z,i, as thepercent change in the unemployment rate in MSA z between years t and t+ 1 and report theestimates in panel B of table (3.6). These results show that, on average, a positive investmentshock significantly increases the unemployment rate and the effects strongly vary in the cross-section of MSAs dependent on the share of routine task-intensive labor. Following a positiveinvestment shock, MSAs with high share of routine labor suffer from higher unemploymentrate, while MSAs with very low share of routine labor are almost unaffected. The results arein line with the estimates in panel A. They also suggest that workers do not immediatelyrespond to investment shock by migrating to other areas.The effects on employment and unemployment rate may stem from both higher job de-struction or lower job creation. To analyze these two effects in detail, I estimate regression(3.8) for changes in job destruction and job creation rates. I also distinguish between jobcreation by new establishments and job creation by incumbents and in case of job destructionbetween destruction through establishment closures and job destruction by ongoing estab-lishments.I report the estimates in table 3.7. Panel A shows that positive investment shocks stim-ulate job creation through new establishments. This effect is independent of the share ofroutine labor and is present in approximately the same magnitude among MSAs with lowshare of routine labor as well as among MSAs with high share of routine labor. This resultsupports the findings in previous literature (e.g. Greenwood et al. (1997)) which identifiesthe investment shock as an important driver of economic growth. In case of existing es-tablishments, a positive investment shock decreases the job creation rate and the effect issignificantly stronger in MSAs with high share of routine labor.Panel B presents the findings for job destruction rate. Investment shock significantly in-creases the job destruction rate and the effect intensifies in MSAs with high share of routinelabor. The estimates are similar in magnitude among both, closing and ongoing establish-ments. These results suggest that the effect of the investment shock on employment andunemployment operates through both job creation and job destruction. Although invest-ment shock is able to stimulate new job openings through new establishments, it lowers the67net job creation rate at existing establishments, which account for the majority of jobs.51It also increases job destruction through establishment closures so that the total effect onemployment is negative. The fact that the effect on the net job creation rate at the ongoingestablishments increases with the share of routine jobs suggests that the effects of investmentshocks operate through capital-labor substitution. Accordingly, a positive investment shock,i.e., a decrease in the quality-adjusted prices of capital goods, would stimulate a higher useof capital in performing routine tasks that used to be carried out by human workers. I willprovide evidence supporting this interpretation in the industry level analysis below.3.4.2 Individual WorkersNow I investigate how the investment shock affects individual workers. Since the theoreticalmodel predicts that lower quality-adjusted prices of capital goods favor the use of capitalespecially if labor is easily substitutable by capital, workers in routine task-intensive occupa-tions are expected to face a higher job displacement risk. To test this prediction, I estimatea logit model of the form:prob (yi,t+1 = 1) =11 + exi,t,τxi,t,τ = −(β0 + β1,t−τISTt−τ + β2RTIi,t + β3,t−τISTt−τ ×RTIi,t+γΛi,t + ψTrendt + ei,t+1), (3.9)where yi,t+1 is an indicator function and equals 1 if worker i experiences unemployment forat least one week in year t + 1. ISTt−τ is the investment shock at time t − τ measuredby the standardized I-shock measure, RTIi,t the routine task-intensity of worker i′s currentoccupation at time t and ISTt−τ × RTIi,t is the interaction term between these variables.The vector Λi,t includes individual control variables such as gender and age. I estimate thismodel with various time lags for the investment shock ranging from τ = 0 to τ = 5.Table (3.8) presents the estimates of the logit model. Panel A shows the raw estimatesof the logit model. I transform these estimates to the expected levels of probabilities of jobloss at various levels of the investment shock and routine task intensity in panel B. Low andhigh RTI correspond to occupation’s routine task intensity value of -2.0 and 4.5, respectively.Negative IST represents an investment shock of -1.5 standard deviation, while positive IST aninvestment shock of +1.5 standard deviations. The probabilities range from 7.16% to 10.6%,which is higher than the average rate of unemployment of 6.4% for the period 1980-2012.The probabilities of job loss are generally higher than the unemployment rate since theyrepresent the probability of experiencing unemployment for at least one week conditional onbeing employed before.Panel C shows how the investment shock affects the probability of job loss at the twodifferent levels of routine task intensity. The IST effect at low RTI shows how the probabilityof job loss changes between positive and negative investment shock for occupations withlow RTI score. The IST effect at high RTI represents the same effect for occupations withhigh RTI score. A positive investment shock decreases the probability of job loss for bothtypes of occupations by about 1.7 percentage points in the first year. From the second yearonwards, each type of occupation experiences a very different effect. A positive investment51Net job creation rate is defined as the job creation rate minus the job destruction rate68shock increases the probability of job loss for workers in routine occupations. The increase issignificant in third, fourth and fifth year following an investment shock. The strongest effectis in the third year, when the investment shock increases the probability by 2.13 percentagepoints. Occupations with low RTI score experience generally lower probability of job loss inall years following a positive investment shock except in third year.To highlight the importance of the routine task intensity for the effects of the investmentshock on the probability of losing job, I calculate the difference between the IST effect at thelow and high RTI in panel D. This difference has a similar interpretation as the interactionterm of the investment shock with the routine task intensity. The differences are positive fromthe second year onwards and significant in the third, fourth and fifth year. This indicatesthat the effect of the investment shock on the job displacement risk depends significantly onthe routine intensity of the particular occupation. Routine intensive occupations are exposedto a relatively higher job displacement risk after a positive investment shock than non-routineoccupations.3.4.3 Manufacturing IndustriesNow I investigate the responses to the investment shocks at the level of detailed industriesin the U.S. manufacturing sector. As shown above, investment shocks have differential ef-fects on various types of labor. While routine task-intensive occupations experience loweremployment, higher rate of unemployment and higher risk of losing job following a positiveinvestment shock, non-routine task-intensive occupations are affected significantly less in allthese dimensions. These results suggest that routine task-intensive labor is substituted bycapital following a positive investment shock. To test this mechanism directly, I analyze theinvestment and employment reactions to the investment shocks across the industries in theU.S manufacturing sector. This analysis allows me to test whether industries with more rou-tine labor not only decrease the number of employees following a positive investment shockbut also invest more in capital. Specifically, I estimate panel regression models of the form:∆Yt+1,j = β0 + β1ISTt + β2RTIj + β3ISTt ×RTIj+ζTFPt,j + ψ Trendt + et+1,j , (3.10)where the dependent variable ∆Yj,t+1 is the change in the number of employees or the changein the real stock of a specific type of capital, either equipment or structures, in industry j be-tween year t and t+ 1. I use the distinction between equipment and structures to investigatewhether industries substitute employees’ labor by investment in equipment following a posi-tive investment shock or whether they expand their overall production capacity by investingin additional structures. The aggregate investment shock ISTt is measured by the standard-ized I-shock measure, RTIj is the employment-weighted average routine task-intensity ofemployees in industry j and ISTt×RTIj is the interaction term between these two variables.I further control for industry j′s total-factor productivity shock TFPt,j in year t to addresspotential concerns that the effects could be driven by industry’s productivity. I also include alinear time trend in the model, but the results are not sensitive to the inclusion of this term.I estimate the model using panel data for manufacturing industries from 1980 to 2011.To establish the link to the results in the previous section, I first estimate equation(3.10) for the changes in the number of employees. I present the estimates in panel A of table693.9. Consistent with the results in the previous section, a positive investment shock decreasesemployment and the effect is stronger among manufacturing industries with very routine task-intensive labor. On average, a positive investment shock of one standard deviation decreasesindustry’s employment by more than half a percent. The coefficient of the interaction termbetween the aggregate investment shock and industry’s routine task intensity is significantlynegative indicating that the employment reduction following a positive investment shock isstronger among industries with routine task-intensive labor. I also split the sample into twohalves, one consisting of industries with non-routine labor, i.e. below-median RTIj and oneconsisting of industries with routine labor, i.e. above-median RTIj , and estimate the modelwithout the interaction term separately for each half of the sample. The comparison of theestimated coefficients of the aggregate investment shock in each half of the sample confirmsthat the investment shock has much stronger effect on employment in industries with routinelabor. While the employment in industries with routine labor decreases by more than onepercent following an investment shock of one standard deviation, the employment in theindustries with non-routine labor is almost unaffected.Now I complement the picture by estimating the model (3.10) for changes in the realstock of equipment. The results are presented in panel B of table 3.9. The estimates of thecoefficient of the aggregate investment shock in the first four columns show that a positiveinvestment shock on average increases investment in equipment by approximately one percent.This is opposite to how the investment shock affects employment. Industries reduce thenumber of employees following a positive investment shock but increase the real stock ofequipment. Taken together, these two effects indicate that the negative effect of investmentshock on employment is not due to a reduction of production capacity but rather due tosubstituting capital for labor. To provide more support to this hypothesis, I estimate howthe effect of the investment shock on equipment depends on the routine task intensity of theemployees in each industry. Column 5 reports the estimates of the model with the interactionterm between the investment shock and industry’s RTIj . The coefficient of the interactionterm is positive and significant showing that the positive effect of the investment shockon industry’s stock of equipment increases with the routine task intensity of the industry’semployees. Splitting the sample into industries with non-routine and routine labor andestimating the model without the interaction term in each half of the sample provides thesame result.Since these results show that the investment shock affects investment in equipment, Ifurther investigate whether it also affects investment in structures in the same manner. Sinceequipment capital encompasses machines that can perform specific tasks by themselves, butstructures represent rather the overall production capacity, the comparison of investment inthese two types of capital helps further reveal whether investment shocks lead exclusively tocapital-labor substitution or also adjustment of the production capacity. The estimates inpanel C show that on average a positive investment shock increases investment in structures.This effect is, however, much weaker than in case of equipment as the estimated coefficientis 0.41% for structures but 1.1% for equipment. The effect also does not differ much be-tween industries with non-routine labor and industries with routine labor. Coefficient ofthe interaction between investment shock and the routine task intensity is insignificant andconsiderably smaller than in case of equipment. This result suggests that investment shockeffects operate through capital-labor substitution by reducing the number of employees and70increasing the use of equipment capital. At the same time, the mild but still positive impacton the investment in structures indicates an expansion of the production capacities acrossall manufacturing industries following a positive investment shock. This is additional evi-dence suggesting that the estimated effect of investment shock on employment is not due toa reduction of production capacities.A combination of these two effects, reduction of employment and expansion of the stockof equipment following a positive investment shock, should increase the capital intensity ofthe industries and the increase is expected to be higher for industries with routine labor. Toconfirm this intuition, I estimate the equation (3.10) with capital intensity as the dependentvariable, where industry j′s capital intensity is the ratio of its stock of equipment and thenumber of employees. The results are presented in panel D and fully support this intuition.The effects of investment shocks are reflected in industries’ capital intensity. As the relativeprice of capital decreases, industries shift from employing human workers towards usingmore capital. Industries with workers performing very routine tasks make much higher use ofthe substitution of capital for labor than industries with workers primarily engaged in non-routine tasks. The opposite effects of the investment shock on employment and the stockof equipment together with the changing magnitude of these effects across the industrieswith different routine task intensity suggest that once the relative price of capital decreases,industries acquire more capital to perform routine tasks previously carried out by humanworkers.3.4.4 Growth in Wages and SalariesNow I study the effect of the investment shock on wages and salaries. The results in theprevious section show that the effect of the investment shock operates through substitutionof capital for labor. Laying off employees who perform routine intensive tasks and investingin machines that can perform these tasks does not have to be the only manifestation of apositive investment shock. Some employees can accept lower wages instead of losing thejob completely. Then an investment shock will create a downward pressure on wages andsalaries in occupations that are subject to a potential replacement by capital. To investigatethis effect, I estimate the panel regression (3.8) for the growth in average and total nominalwages and salaries in the local labor markets as the dependent variables.Table (3.10) shows the results of this estimation. The investment shock has an immediatepositive average effect on the growth of both the total and the average wages and salaries.The estimates in column 1 and 7 indicate an increase in the growth of total and averagewages and salaries by 0.37 and 0.28 percentage points, respectively. The positive sign ofthese two coefficient is surprising and deserves explanation. First, the wages and salaries areexpressed in nominal values. Estimates based on real wages and salaries are likely to deliverlower coefficients. Second, I will show in later section that the average effect is significantlynegative and with stronger magnitude in the subsequent years.The estimates of the full specification with the interaction term of the investment shockand the share of routine intensive jobs are presented in columns 4 and 10. The coefficientsof the interaction term are strongly negative and indicate that a positive investment shockincreases the growth in wages and salaries in areas with a very low share of routine taskintensive jobs. The coefficients of the investment shock in columns 5 and 6 confirm thisresult. A positive investment shock immediately increases the growth in the total nominal71wages and salaries by 0.79 percentage points in geographical areas with low share of routinelabor but only by 0.295 percentage points in areas with high share of routine jobs. Theestimates for the growth in average wages in columns 11 and 12 are similar. This resultsindicate that the investment shock has also differential effect on wages and salaries. Incontrast to unemployment, the immediate effect on wages and salaries is concentrated inareas with low share of routine jobs. I will show in the next section that areas with highshare of routine labor are affected more strongly in subsequent years following a positiveinvestment shock and decrease the growth in wages and salaries.In summary, the estimates in this section support the predictions of the theoretical model.Following a positive investment shock, employment immediately decreases in areas and in-dustries with high share of routine jobs, but does not change in areas and industries withlow share of routine labor. At the same time, industries with high share of routine jobsincrease investment in equipment. The focus of this section was on the immediate effectswithin one year after the investment shock. I next investigate the dynamics of these effectsin the following years.3.5 The Time Profile of the Estimates in Later YearsThe theoretical model predicts differential response in employment and investment in locallabor markets and industries, but it is silent about when these responses begin and forhow long they last. The previous section shows that most of the responses are observableimmediately in the year following the investment shock. Now I focus on how long theseresponses last and how they change in the subsequent years after the investment shock.Before I look at the differential responses across local labor markets and manufacturingindustries, I show how the IST shock affects the aggregate macro variables.Specifically, I analyze how personal income, unemployment rate, duration of unemploy-ment, personal expenditures evolve over the next 20 quarters following an investment shock.52I also use the data from Davis et al. (2006) and analyze the effects of the investment shockon the aggregate job creation and job destruction rate. I report the summary description ofthese aggregate variables in table 3.11. I estimate the following regression equation.∆Yt+s = βXt + t+s (3.11)∆Yt+s is the change in the macro labor variable between t+ s and t+ s− 1. Specifically, it isa percentage change in the the nominal disposable personal income (per capita), a change inpercentage points for the civilian unemployment rate, job creation and job destruction rateand a change in the number of weeks for the average duration of unemployment. I focus onthe lagged effect of the IST shock on this labor variables and hence I use s ∈ [0, 20]. TheXt is the measure of the IST shock. In order to focus on larger IST shocks, I sum up the(normalized) shocks to the I-shock measure over the previous 8 quarters (i.e. 2 years).52Personal income is the nominal disposable personal income. I use the nominal time series of personaldisposable income since I want to avoid the mechanical effect on the real income. Since the IST measureis constructed from the differences in relative prices, an increase in prices of consumption goods would bereflected in both, in a positive IST shock and in lower real income.72I also analyze the cumulative effect on the macro labor variables and estimate the regres-sion equation adjusted in the following way.t+s∑i=0∆Yt+i = βXt + t+s (3.12)The estimated coefficients for all time lags are in figure 3.3. A shock to the IST measure de-creases the nominal disposable income. The coefficients are statistically significant from thirdto twelve-th quarter. The estimated coefficients for quarters 3 to 12 are between - 0.049 and-0.082 hence a (two-year sum of) one standard deviation shock to the IST measure decreasesthe nominal income by ca. 0.05 to 0.08 % per quarter. Considering that between 1965 and2012 there are 47 (normalized) shocks of a magnitude of 5.0 or more, these coefficients alsohave economic significance. For example, an investment shock of with a magnitude of 5.0changes the average nominal income by ca 3% (0.05%*12*5.0) over the subsequent 3 years(12 quarters).The effect on the unemployment rate is positive with significant coefficients for quarters 1to 8. A shock of 5.0 magnitude changes the unemployment rate by 1.17% over the next threeyears. Considering that the average unemployment rate between 1965 and 2012 was 6.1%,this shock can change the unemployment rate by 19%. A positive IST shock also prolongs theaverage duration of unemployment. The coefficients are positive and statistically significantfor quarters 4 to 11. The average duration of unemployment is 15.8 weeks and a 5.0-magnitudeshock changes the average duration by 2.8 weeks or 17.6%. The results for job creation andjob destruction rates are in line with the estimates for the unemployment rate. A positiveinvestment shock decreases the job creation rate and increases the job destruction rate, whichresults in higher rate of unemployment. The results are robust to including the TFP shockin the regression equation.Now I focus on how these effects differ across the local labor markets. I estimate avariation of the panel regression (3.8). Specifically, I estimate the regression in this form:∆Yt+1,z,i = β0 + β1ISTt−τ + β2RSHt,z,i + β3ISTt−τ ×RSHt,z,i+ψ Trendt + γi + et+1,z,i, (3.13)where τ ∈ [0, 5] is the time lag in number of years.53Table 3.12 presents the estimates for the unemployment rate. To relate the results tothe aggregate effect discussed above, I first estimate the panel regression (3.13) without theinteraction term. I report these estimates in columns 7 to 12. The coefficients are positiveand significant for the first four years and become significantly negative in year five and six.This is consistent with the evidence on the aggregate rate of unemployment and it showsthat the investment shock is followed by a dynamic changes in the unemployment rate.The estimates of the panel regression with the interaction terms, in columns 1 to 6,show that these changes significantly differ between areas with low and high share of routinelabor. The interaction term is significantly positive in the first two years which indicatesthat unemployment rate increases especially in areas with high share of routine jobs over this53Estimation of the regression (3.13) with share of routine labor lagged by τ years, i.e. RSHt−τ,z,i providessimilar results. This is due to the fact that the share of routine labor is relatively persistent at the level oflocal labor markets.73time horizon. In the years four, five and six, the coefficient of the interaction term becomessignificantly negative, so that the areas with high share of routine labor are expected todecrease the unemployment rate relative to their non-routine counterparts. I confirm thisresult by estimating the panel regression without the interaction term in two sub-samples, oneconsisting of areas with non-routine labor and other consisting of areas with routine labor.The estimates in panel B show that zones with routine labor experience both a strongerincrease in unemployment rate in the first two years and a stronger decrease in the lateryears.Tables (3.13) and (3.14) report the results for job creation and job destruction rates.These two tables show that the observed differential dynamics of the effects of the investmentshocks on the unemployment rate is due to both, job creation and job destruction. Areaswith routine labor suffer from lower job creation and higher job destruction in the first twoyears following the investment shock. But they benefit from mostly higher job creation andlower job destruction in the later years.The estimated coefficients for the growth in total wages also exhibit differential dynamicsacross the local labor markets. Areas with non-routine labor experience higher growth in totalwages in the first year, but do not show a clear pattern in the later years. Areas with highshare of routine jobs have only a mild growth in the first year, but suffer from significantlylower wage growth in the following two years. The effect becomes again positive only inthe fifth year. This result suggests that a positive investment shock creates a significantand relatively long lasting labor income risk especially in markets with high share of routineintensive occupations.I conduct an equivalent analysis at the level of the U.S. manufacturing industries. Theseestimates, which are not reported here, exhibit similar differential dynamics and supportthe result in this section. I also estimate various alternations of the panel regressions (3.8)and (3.13) to investigate the robustness of the results. I estimate the changes in industries’employment, equipment and capital intensity during each of the major investment shocksbetween 1980 and 2011. This exercise allows to exclude the aggregate investment shockfrom the regression and focus on whether the sign of the coefficient of industries’ routinetask intensity changes between the periods of positive and negative investment shocks. Theresults of this approach, which are not reported here, support the robustness of the estimatedcoefficients of the interaction term presented in this and previous section. I also estimate thepanel regression (3.13) with all lags together. The results are similar to those presentedabove.3.6 ConclusionThis paper shows that the investment shock is an important source of job displacement andlabor income risk. I analyze the responses to the investment shock at three different levels,local labor markets, U.S. manufacturing industries and individual workers. I show that theeffects of the investment shock on labor operate through substitution of capital for laborin tasks that are highly routine intensive and hence easily automatable. These effects areobservable immediately after the investment shock and can last for multiple years.These results highlight the importance of the labor heterogeneity for the analysis of theinvestment shock. While workers in routine intensive occupations experience higher probabil-74ity of losing job and lower growth in wages when a positive investment shock occurs, workersin non-routine occupations are mostly unaffected or even benefit from higher wage growth.Job displacement risk and labor income risk can significantly affect households’ income.Hence my results are relevant for studying households’ consumption and investment decisions.At the same time, the effects of the investment shocks are manifested by firms’ decision toshift from the use of labor towards the use of capital. These results can also speak to firms’optimal investment and employment decision and to the firms’ growth dynamics.753.7 Derivation of the Model ImplicationsThe monopolist’s is maximizing the profit given the iso-elastic demand function and the CESproduction function:maxK,LΠ = maxK,L{PcQ− PqK −WL}= maxK,L{Qξ − PqK −WL}= maxK,L{(Kσ + Lσ)ξσ − PqK −WL}. (3.14)The first-order conditions for optimal choice of capital and labor state:(Kσ + Lσ)ξσ−1Kσ−1 = Pq, (3.15)and(Kσ + Lσ)ξσ−1 Lσ−1 = W. (3.16)Combining equations 3.15 and 3.16 and rearranging gives an implicit function of the form:f (Pq, L) =((PqW) σσ−1+ 1) ξ−σσLξ−1 = W. (3.17)The total derivative of the implicit function f (Pq, L) allows to express the derivative of theequilibrium optimal labor choice of labor w.r.t. the price of capital as:dLdPq= −((PqW) σσ−1+ 1)−1ξ−σσ−1P1σ−1q Wσ1−σ(ξ − 1)L−1 . (3.18)763.8 Tables and FiguresFigure 3.1: Equipment-per-Worker Indexes and Routine-Task Intensity of Employed and Unemployed WorkersPanel A plots the employment-per-worker indexes for information processing equipment, industrial equipment, transportationequipment, other equipment and total equipment. The indexes are calculated by dividing the current stock (in current U.S.dollars) of each equipment type by the total number of civilian employees and normalized to one at the beginning of each time-series. Panel B plots the average routine-task intensity of the current occupations of employed and most recent occupations ofunemployed workers. Routine-task intensity is calculated at the level of occupations as described by formula 3.1 below. Thesamples of employed and unemployed workers is from the Current Population Surveys. The straight lines with break in 1982 areestimated linear trends of each time-series. The break in 1982 is selected based on the evidence from Fisher (2006). The dashedlines are 5th-95th confidence intervals calculated from bootstrap. The sample period in both panels is from 1970 to 2015.77Figure 3.2: Investment Shocks: 1980-2012This figure shows the time series of the I-shock measure from 1980 to 2012. The I-shockmeasure is standardized to mean zero and standard deviation of one. 1st Shock 2nd Shock 3rd Shock 4th Shock 78Figure 3.3: Investment Shock and the Aggregate Labor MarketRegression coefficients and the 95% confidence intervals in panels A, B, C, G, H, and I are from∆Yt+s = βXt + t+s and in panels D, E, F, J, K and L from∑t+si=0 ∆Yt+i = βXt + t+s. Bothregressions are estimated from quarterly data at lags of 0 to 20 quarters. The independentvariable Xt is the change in the IST measure cumulated over 8 quarters. Data cover theperiod from 1966 to 2012, except for job creation and destruction rate, which is only until2005. Job creation and destruction data are from Davis et al. (2006), remaining variables arefrom the U.S. Bureau of Economic Analysis, retrieved from FRED, Federal Reserve Bank ofSt. Louis; https://fred.stlouisfed.org.Panel A. Personal income, single Panel B. Unemployment rate, single Panel C. Dur. of unemploy., singlePanel D. Personal income, cum. Panel E. Unemployment rate, cum. Panel F. Dur. of unemployment, cum.Panel G. Personal expenditures, single Panel H: Job creation rate, single Panel I: Job destruction rate, singlePanel J. Personal expenditures, cum Panel K: Job creation rate, cum. Panel L: Job destruction rate, cum.79Table 3.1: I-shock: Summary Statistics 1960-2012I-Shock is the change in the quality-adjusted relative price of investment goods. It is calcu-lated by subtracting the log change in the prices of nondurable consumption goods from thethe log change in the quality-adjusted prices of equipment. Price shock is calculated equiv-alently using changes in the raw price series. Both measures are at annual frequency. TFPshock of capital goods producers is a weighted average TFP shock of manufacturing indus-tries producing investment goods, with weights based on their relative shipments to privatefixed investments. TFP shock of consumption goods producers is calculated equivalently. Allstatistics are calculated from 1960 to 2012. Mean and standard deviations are in percent.Data sources: U.S. Bureau of Economic Analysis, Bureau of Labor Statistics, NBER-CESManufacturing Industry DatabasePanel A. Time-series properties of investment shock measuresMean St. dev.SerialCorrelationCorrelationI-Shock -Price shockI-shock 4.123 2.937 0.201 0.993Price shock 1.963 2.658 0.207 0.993Panel B. Correlation CoefficientsGDPgrowthAggTFPshockTFP shock ofcapital goodsproducersTFP shock ofconsumptiongoods producersI-shock 0.228 -0.031 0.418 -0.015Price shock 0.197 -0.015 0.427 -0.01580Table 3.2: Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Investment ShocksThe sample in panel A includes 457 4-digit 1987 SIC industries in the manufacturing divi-sion and sample in panel B consists of 473 6-digit NAICS industries in the manufacturingdivision. Capital goods producers are industries which actively supply to private fixed invest-ments based on the BEA Input-Output tables. Consumption goods producers are definedequivalently based on positive supply to Personal Consumption expenditures. There are 145capital goods producers and 290 consumption goods producers in panel A and 115 capitalgoods producers and 196 consumption goods producers in panel B. The Electronic ComputerManufacturing in Panel C is one industry defined by 6-digit NAICS code, 334111. The TFPshock is the growth in the TFP index from 1981 to 1983. All growth variables are valueweighted averages weighted by the industry’s shipments to the corresponding type of thefinal use. The growth variables are in percent. Data sources: NBER-CES ManufacturingIndustry Database and U.S. Bureau of Economic Analysis, Benchmark Input-Output Data.Industries TFP ShockShipmentPrices,growthShipments,real growthPanel A. 1st Investment Shock: 1981-1983All ManufacturingIndustries1.48 3.55 -1.18Capital GoodsProducers-1.98 6.56 -2.40Consumption GoodsProducers2.69 4.18 3.29Panel B. 2snd Investment Shock: 1985-1986All ManufacturingIndustries0.21 -1.93 1.20Capital GoodsProducers-0.75 0.58 -0.40Consumption GoodsProducers0.57 -1.41 1.7181Table 3.2 Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Investment Shocks: ContinuedIndustries TFP ShockShipmentPrices,growthShipments,real growthPanel C. 3rd Investment Shock: 1997-2001All ManufacturingIndustries5.33 -0.09 22.63Capital GoodsProducers19.61 -9.68 66.29Consumption GoodsProducers4.17 0.71 14.51Electronic ComputerManufacturing215.83 -80.56 383.75Other CapitalGoods Producers-7.58 0.14 22.30Panel D. 4th Investment Shock: 2008-2009All ManufacturingIndustries-4.05 -8.28 -9.13Capital GoodsProducers-8.17 -0.07 -14.87Consumption GoodsProducers-5.72 -12.51 -4.2182Table 3.3: Summary Statistics: Share of Routine-Intensive Labor and Other Labor MarketVariables in Metropolitan Statistical AreasShare of routine intensive jobs in a given area is the number jobs in routine intensive occu-pations divided by the total number of jobs in that area. Routine-intensive occupations aredefined as the occupation in the highest third by the routine task-intensity, RTI, weightedby the employment share in 1980. All statistics in panel A are from the cross-section of 363metropolitan statistical ares (MSAs) in the particular year. Statistics in panel B are froma balanced panel of 363 MSAs from 1980 to 2012 with 11,979 records. All numbers are inpercent. Employment and population data is from Census 1 percent sample for 1970 andCensus 5 percent samples for 1990 and 2000; Routine task-intensity data is from Autor andDorn (2013); Unemployment is from Local Area of Unemployment Statistics from Bureau ofLabor Statistics; Wages and salaries are from U.S. Bureau of Economic Analysis; Job creationand destruction rates are from Business Dynamics Statistics from U.S. Census Bureau.Mean Median St. dev. Min MaxPanel A. Share of routine intensive labor, RSH, in metropolitan areasYear 1980 31.09 31.05 3.82 20.63 40.17Year 1990 30.08 30.19 2.88 22.04 37.03Year 2000 30.22 30.41 2.55 21.86 35.96Difference 1990 - 1980 -1.01 -1.01 2.03 -8.24 6.42Difference 2000 - 1990 0.14 0.1 1.52 -4.75 5.45Difference 2000 - 1980 -0.87 -0.9 2.4 -7.91 5.42Panel B. Labor market variables in metropolitan areasUnemployment rate, change 0.07 -0.14 1.16 -13.31 9.29Total wages and salaries, growth 5.06 5.08 3.51 -9.14 18.53Average wage and salary, growth 3.47 3.34 2.12 -3.56 12.31Job creation rate, new firms 6.19 5.8 2.43 1.5 52.9Job creation rate, incumbents 9.99 9.79 2.54 3.82 67.22Job destruction rate, closures 4.99 4.7 1.84 1.1 27.4Job destruction rate, incumbents 9.43 9.1 2.59 3.16 94.83Table 3.4: Summary Statistics: Individual Workers MicrodataSample includes persons 14+ years old. A precise definition of all variables is in table B.1in appendix. Average RTI is the average routine task-intensity in each of the sub-sample.Data sources: Current Population Survey, U.S. Census Bureau and Integrated Public UseMicrosample from 1980 to 2012. RTI measure is from Autor and Dorn (2013).N Average age Average RTIPotential labor force 2,291,703 45.45 1.06Labor force 1,564,524 41.05 1.05Working, at least part of the year 1,539,360 41.14 1.05Not working, looking for work 25,164 35.93 1.31Unemployed for part of the year 142,814 36.12 1.2Unemployed for part or whole year 167,978 36.09 1.2183Table 3.5: Summary Statistics: Industries in the U.S. Manufacturing SectorThe sample includes 457 4-digit 1987 SIC industries in the manufacturing division. The statis-tics is based on time period from 1980 to 2011. The growth variables are in percent. Numberof employees, equipment, shipments and structures growth is from the NBER-CES Manufac-turing Industry Database; RTI ONET is based on Author’s calculations using O*NET andOES data.Mean Median St. dev. Min MaxNumber of employees, growth -2.21 -1.78 8.74 -40.26 40.44Equipment, growth 1.64 1.18 5.18 -17.69 60.69Shipments, growth 2.93 3.26 11.78 -44.8 58.32Structures, growth 0.04 -0.44 3.13 -36.27 48.82RTI ONET, log -1.15 -1.15 0.19 -1.98 -0.54RTI ONET, level -2.93 -2.85 0.92 -7.18 -0.4484Table 3.6: The Immediate Impact of the IST Shock and Share of Routine-Intensive Laboron Employment and Unemployment Rate in Geographical AreasThe sample is a balanced panel of 363 metropolitan statistical areas from 1980 to 2012in panel A and from 1990 to 2012 in panel B. The estimates are from panel regressions∆Yt+1,z,i = β0 + β1ISTt + β2RSHt,z,i + β3ISTt × RSHt,z,i + ψ Trendt + γi + et+1,z,i. Thedependent variable in panel A is the growth (%) in the number of employed persons in eachMSA z between year t and t + 1. The dependent variable in panel B is the annual change(%) in the unemployment rate in each MSA z between year t and t + 1. The independentvariables are the aggregate IST shock ISTt measured by the I-shock measure, the shareof routine intensive jobs in the MSA z RSHz,t at time t and their interaction term. Allregressions also include intercept, linear trend and state fixed effect γi. The aggregate ISTshock is standardized at mean zero and standard deviation of one. Results in column 5 areestimated from a subset of MSA with below-median share of routine task-intensive labor,while results in column 6 are estimated from the above-median complementary subset ofMSA. All regressions are weighted by the share of population in each MSA at the beginningof the decade. The standard errors are clustered at the state level. *** Significant at the 1percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Panel A. Change in employment(1) (2) (3) (4) ≤ RSH50 > RSH50ISTt -0.295∗∗ -0.304∗∗ 1.813∗∗ 0.0132 -0.387∗∗∗(-2.32) (-2.46) (2.16) (0.11) (-2.80)RSHz,t -6.654∗∗ -6.992∗∗ -3.796 -6.215∗ -9.963∗∗(-2.05) (-2.25) (-0.96) (-1.78) (-2.65)ISTt ×RSHz,t -6.576∗∗(-2.37)Trend -0.0524∗∗∗ -0.0614∗∗∗ -0.0603∗∗∗ -0.0593∗∗∗ -0.0725∗∗∗ -0.0604∗∗∗(-4.43) (-5.78) (-5.72) (-5.72) (-5.62) (-5.04)State FE Yes Yes Yes Yes Yes YesN 7964 7964 7964 7964 3982 3982R2 0.112 0.109 0.117 0.120 0.0990 0.128Panel B. Change in unemployment rate(1) (2) (3) (4) ≤ RSH50 > RSH50ISTt 0.160∗∗∗ 0.161∗∗∗ -1.880∗∗∗ -0.0414 0.212∗∗∗(3.23) (3.26) (-4.92) (-0.57) (4.71)RSHz,t 0.966∗∗ 1.145∗∗∗ -1.936∗∗ 0.393 2.009(2.54) (3.04) (-2.35) (0.46) (1.42)ISTt ×RSHz,t 6.341∗∗∗(5.16)Trend 0.0180∗∗∗ 0.0199∗∗∗ 0.0193∗∗∗ 0.0183∗∗∗ 0.0215∗∗∗ 0.0198∗∗∗(9.65) (11.23) (10.34) (9.80) (10.48) (6.87)State FE Yes Yes Yes Yes Yes YesN 7964 7964 7964 7964 3982 3982R2 0.0251 0.0175 0.0256 0.0331 0.0197 0.032685Table 3.7: The Immediate Impact of the IST Shock on the Job Creation and Job Destruction Rate across Geographical AreasThe sample is a balanced panel of 363 MSAs from 1980 to 2012. The estimates are from panel regressions ∆Yt+1,z,i = β0+β1ISTt+β2RSHt,z,i + β3ISTt × RSHt,z,i + ψ Trendt + γi + et+1,z,i. The dependent variables are the annual March-to-March job creationand job destruction rates (%) of new firms, incumbents and closures in each metropolitan area z between year t and t + 1. Theindependent variables are the aggregate IST shock ISTt measured by the normalized I-shock measure, the share of each MSA’sroutine intensive jobs RSHz,t at time t and their interaction term. All regressions include intercept, linear trend and state fixedeffect γi. All regressions are weighted by the share of population in each MSA. The standard errors are clustered at the state level.*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Panel A. Change in job creation rateChange in job creation rate, new firms Change in job creation rate, incumbents(1) (2) (3) (4) ≤ RSH50 > RSH50 (7) (8) (9) (10) ≤ RSH50 < RSH50ISTt 0.272∗∗∗ 0.273∗∗∗ 0.299 0.292∗∗∗ 0.265∗∗∗ -0.328∗∗∗ -0.322∗∗∗ 0.487∗ -0.205∗∗∗ -0.346∗∗∗(10.91) (11.19) (1.49) (6.30) (13.68) (-8.69) (-8.52) (1.95) (-5.31) (-8.06)RSHz,t 0.894 1.086 1.101 5.460∗ 1.598 9.798∗∗∗ 9.571∗∗∗ 10.03∗∗∗ 10.43∗∗∗ 11.42∗∗(0.36) (0.43) (0.43) (1.81) (0.35) (4.94) (4.87) (5.27) (3.84) (2.36)ISTt ×RSHz,t -0.0805 -2.470∗∗∗(-0.14) (-3.05)Trend -0.0786∗∗∗ -0.0726∗∗∗ -0.0779∗∗∗ -0.0779∗∗∗ -0.115∗∗∗ -0.0707∗∗∗ -0.0692∗∗∗ -0.0699∗∗∗ -0.0637∗∗∗ -0.0636∗∗∗ -0.0888∗∗∗ -0.0574∗∗∗(-8.94) (-7.72) (-8.07) (-8.07) (-10.42) (-6.41) (-8.10) (-7.99) (-7.15) (-7.16) (-7.43) (-5.64)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11584 11584 11584 11584 5793 5793 11584 11584 11584 11584 5793 5793R2 0.488 0.470 0.488 0.488 0.452 0.525 0.370 0.365 0.387 0.388 0.319 0.419Panel B. Change in job destruction rateChange in job destruction rate, firm closures Change in job destruction rate, incumbents(1) (2) (3) (4) ≤ RSH50 > RSH50 (7) (8) (9) (10) ≤ RSH50 > RSH50ISTt 0.322∗∗∗ 0.325∗∗∗ -0.581∗∗∗ 0.218∗∗∗ 0.346∗∗∗ 0.472∗∗∗ 0.476∗∗∗ -0.531 0.331∗∗∗ 0.507∗∗∗(15.42) (15.11) (-3.44) (6.38) (17.04) (9.97) (10.02) (-1.22) (5.33) (10.79)RSHz,t 4.607∗∗∗ 4.835∗∗∗ 4.325∗∗∗ 3.578 7.617∗∗∗ 6.294∗∗∗ 6.630∗∗∗ 6.062∗∗∗ -5.629 8.574(4.23) (4.39) (4.05) (1.53) (3.01) (3.49) (3.54) (2.96) (-1.26) (1.55)ISTt ×RSHz,t 2.767∗∗∗ 3.077∗∗(5.59) (2.32)Trend -0.0612∗∗∗ -0.0521∗∗∗ -0.0584∗∗∗ -0.0584∗∗∗ -0.0908∗∗∗ -0.0495∗∗∗ -0.0437∗∗∗ -0.0306∗∗∗ -0.0399∗∗∗ -0.0399∗∗∗ -0.0501∗∗∗ -0.0349∗∗∗(-11.70) (-9.13) (-10.56) (-10.50) (-15.86) (-8.75) (-8.95) (-5.89) (-8.15) (-8.09) (-6.60) (-5.71)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11580 11580 11580 11580 5790 5792 11584 11584 11584 11584 5793 5793R2 0.415 0.381 0.423 0.427 0.421 0.452 0.192 0.149 0.200 0.203 0.176 0.21786Table 3.8: The Impact of the IST Shock on the Probability of Job Loss for Routine andNon-routine OccupationsSample consists of persons who were in the CPS survey for two con-secutive years and were employed in the first year. The estimatesare from logit regressions prob (yi,t+1 = 1) =11+exi,t,τ, where xi,t,τ =− (β0 + β1,t−τISTt−τ + β2RTIi,t + β3,t−τISTt−τ ×RTIi,t + γΛi,t + ψTrendt + ei,t+1).The dependent variable is one if the person was unemployed for at least one week in hersecond year in the survey. The independent variables are the normalized aggregate ISTshock at lags τ , the routine intensity score of the person’s occupation and their interactionterm. The vector Λi,t includes control variables for person’s gender and age. All modelsinclude a linear time trend. Coefficients of the logit model are in the panel A. The predictedprobabilities in panel B are at covariates’ levels -2, 4.5, -1.5 and 1.5 for Low RTI, High RTI,neg. IST and pos. IST, respectively. The differences in probabilities in panel C are definedas Low RTI pos. IST minus Low RTI neg. IST and High RTI pos. IST minus High RTI neg.IST for IST effect at low RTI and IST effect at high RTI, respectively. The difference in ISTeffect in panel D is defined as the IST effect at high RTI minus the IST effect at low RTI.*** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant atthe 10 percent level.(1) (2) (3) (4) (5) (6)Panel A. Logit coefficientsISTt -0.0748∗∗∗(-13.09)ISTt ×RTIi,t 0.00200(0.82)ISTt−1 -0.0131∗∗(-2.22)ISTt−1 ×RTIi,t 0.00366(1.46)ISTt−2 0.0668∗∗∗(8.89)ISTt−2 ×RTIi,t 0.00426(1.35)ISTt−3 -0.0185∗∗∗(-2.89)ISTt−3 ×RTIi,t 0.0103∗∗∗(3.91)ISTt−4 -0.00149(-0.30)ISTt−4 ×RTIi,t 0.00582∗∗∗(3.07)ISTt−5 -0.00793∗(-1.83)ISTt−5 ×RTIi,t 0.00258(1.46)87Table 3.8 The Impact of the IST Shock and Routine Intensity on the Probability of Job Loss:Continued(1) (2) (3) (4) (5) (6)Panel B. Levels of probabilitiesLow RTI neg. IST 0.0902 0.0818 0.0716 0.0829 0.0804 0.0809High RTI neg. IST 0.106 0.0950 0.0827 0.0915 0.0926 0.0950Low RTI pos. IST 0.0735 0.0786 0.0853 0.0778 0.0794 0.0789High RTI pos. IST 0.0890 0.0959 0.104 0.0986 0.0989 0.0960Panel C. Differences in probabilitiesIST effect at low RTI -0.0168 -0.00324 0.0138 -0.00516 -0.000953 -0.00200(-11.57) (-2.24) (7.81) (-3.29) (-0.80) (-1.91)IST effect at high RTI -0.0170 0.000867 0.0213 0.00708 0.00629 0.000939(-6.82) (0.35) (7.16) (2.72) (3.21) (0.54)Panel D. Difference in IST effectDifference in IST effect -0.000240 0.00411 0.00756 0.0122 0.00724 0.00294(-0.08) (1.35) (2.06) (3.84) (3.15) (1.37)N 674565 674565 674565 674565 674565 674565PseudoR2 0.0303 0.0298 0.0301 0.0298 0.0298 0.029888Table 3.9: The Immediate Impact of the IST Shock and Routine Task Intensity on Employ-ment, Investments and Capital Intensity in the U.S. Manufacturing IndustriesThe balanced panel consists of 457 U.S. manufacturing industries defined at 4-digit 1987 SICcode from 1980 to 2011. The estimates are from the panel regressions ∆Yt+1,j = β0+β1ISTt+β2RTIj +β3ISTt×RTIj + ζTFPt,j +ψ Trendt+et+1,j . The panels present estimates for thefollowing dependent variables: changes in employment (panel A), changes in the real stockof equipment (panel B), changes in the real stock of structures (panel C), and changes inthe capital intensity (panel D). Each dependent variable is the change (percent) at industryj between year t and t + 1. The independent variables are the aggregate IST shock ISTtmeasured by the I-shock measure, the employment-weighted average routine task intensityof employees in industry j RTIj and their interaction term. The regression also includesindustry-specific productivity shock TFPt,j in year t and linear time trend Trendd. Columns6 and 7 present estimates for industries with below-median RTIj and above-median RTIj ,respectively. NBER-CES Manufacturing Industry Database is due to (Bartlesman and Gray(1996); Becker et al. (2013)).All U.S. manufacturing industriesIndustrieswith nonroutinelaborIndustrieswith routinelabor(1) (2) (3) (4) (5) (6) (7)Panel A. Changes in employmentISTt -0.562∗∗∗ -0.562∗∗∗ -0.698∗∗∗ -2.651∗∗∗ -0.151 -1.246∗∗∗(-3.39) (-3.39) (-4.25) (-4.36) (-0.73) (-5.98)RTIj,t 0.683∗∗∗ 0.681∗∗∗ 0.638∗∗∗ 0.790∗∗∗ -0.768∗∗∗ -6.037∗∗∗(60.99) (62.88) (50.10) (16.72) (-75.66) (-70.55)ISTt ×RTIj,t -0.654∗∗∗(-3.32)TFPj,t 25.58∗∗∗ 25.61∗∗∗ 23.33∗∗∗ 29.06∗∗∗(8.57) (8.57) (5.78) (7.49)Trendt -0.130∗∗∗ -0.149∗∗∗ -0.130∗∗∗ -0.122∗∗∗ -0.122∗∗∗ -0.0755∗∗∗ -0.167∗∗∗(-7.40) (-6.93) (-7.40) (-7.06) (-7.04) (-3.00) (-8.37)Industry FE Yes Yes Yes Yes No Yes YesN 14057 14057 14057 14057 14057 6895 7162R2 0.0954 0.0924 0.0954 0.131 0.135 0.101 0.172Panel B. Changes in the real stock of equipmentISTt 1.107∗∗∗ 1.107∗∗∗ 1.104∗∗∗ 2.109∗∗∗ 0.865∗∗∗ 1.343∗∗∗(12.75) (12.75) (12.48) (4.24) (8.07) (11.45)RTIj,t -0.0826∗∗∗ -0.0797∗∗∗ -0.0809∗∗∗ -0.159∗∗∗ -1.537∗∗∗ -3.543∗∗∗(-7.31) (-7.02) (-7.76) (-4.39) (-150.39) (-43.92)ISTt ×RTIj,t 0.337∗∗(2.00)TFPj,t 0.698 0.685 1.014 0.216(0.53) (0.51) (0.52) (0.15)Trendt -0.144∗∗∗ -0.108∗∗∗ -0.144∗∗∗ -0.144∗∗∗ -0.144∗∗∗ -0.144∗∗∗ -0.144∗∗∗(-6.71) (-4.96) (-6.71) (-6.76) (-6.76) (-3.86) (-6.74)Industry FE Yes Yes Yes Yes No Yes YesN 14057 14057 14057 14057 14057 6895 7162R2 0.321 0.291 0.321 0.321 0.323 0.346 0.28589Table 3.9 The Immediate Impact of the IST Shock and Routine Task Intensity on Employ-ment, Investments and Capital Intensity in the U.S. Manufacturing Industries: ContinuedAll U.S. manufacturing industriesIndustrieswith nonroutinelaborIndustrieswith routinelabor(1) (2) (3) (4) (5) (6) (7)Panel C. Changes in the real stock of structuresISTt 0.414∗∗∗ 0.414∗∗∗ 0.414∗∗∗ 0.808∗∗∗ 0.339∗∗∗ 0.490∗∗∗(12.35) (12.35) (12.07) (3.39) (6.64) (13.18)RTIj,t 0.165∗∗∗ 0.166∗∗∗ 0.166∗∗∗ 0.136∗∗∗ -0.515∗∗∗ -1.293∗∗∗(29.36) (29.54) (33.56) (7.71) (-110.44) (-32.35)ISTt ×RTIj,t 0.132(1.63)TFPj,t 0.0715 0.0664 0.820 -1.147(0.07) (0.07) (0.61) (-1.08)Trendt -0.0539∗∗∗ -0.0403∗∗∗ -0.0539∗∗∗ -0.0539∗∗∗ -0.0540∗∗∗ -0.0605∗∗∗ -0.0478∗∗∗(-5.13) (-3.73) (-5.13) (-5.21) (-5.21) (-3.34) (-4.62)Industry FE Yes Yes Yes Yes No Yes YesN 14057 14057 14057 14057 14057 6895 7162R2 0.362 0.351 0.362 0.362 0.363 0.448 0.259Panel D. Changes in the capital intensityISTt 1.611∗∗∗ 1.611∗∗∗ 1.768∗∗∗ 4.961∗∗∗ 0.950∗∗∗ 2.588∗∗∗(6.79) (6.79) (7.49) (7.19) (3.44) (8.52)RTIj,t -0.801∗∗∗ -0.797∗∗∗ -0.747∗∗∗ -0.995∗∗∗ -0.468∗∗∗ 3.667∗∗∗(-63.84) (-65.45) (-57.75) (-20.13) (-48.37) (43.81)ISTt ×RTIj,t 1.070∗∗∗(4.95)TFPj,t -29.42∗∗∗ -29.47∗∗∗ -26.70∗∗∗ -33.62∗∗∗(-7.26) (-7.24) (-4.69) (-7.23)Trendt 0.00710 0.0603∗∗ 0.00710 -0.00241 -0.00279 -0.0519∗ 0.0447∗∗(0.35) (2.51) (0.35) (-0.13) (-0.15) (-1.76) (2.06)Industry FE Yes Yes Yes Yes No Yes YesN 14057 14057 14057 14057 14057 6895 7162R2 0.0815 0.0643 0.0815 0.115 0.121 0.125 0.12290Table 3.10: The Immediate Impact of the IST Shock and Share of Routine Intensive Labor on the Growth in Total and AverageWages and Salaries in Geographical AreasPanel consists of 363 metropolitan areas from 1980 to 2012. The estimates are from panel regressions ∆Yt+1,z,i = β0 + β1ISTt +β2RSt,z,i + β3ISTt ×RSt,z,i + ψ Trendt + γi + et+s,z,i. The dependent variables are the annual growth (%) in the total wages andsalaries and in the average wage and salary in each metropolitan area z between year t and t + 1. The independent variables arethe aggregate IST shock ISTt, the share of routine intensive jobs in the metropolitan area RSt and their interaction term. Allregressions also include intercept, linear trend and state fixed effect γi. The aggregate IST shock is standardized at mean zero andstandard deviation of one. All regressions are weighted by the share of population in each metropolitan area at the beginning ofthe decade. The standard errors are clustered at the state level. *** Significant at the 1 percent level. ** Significant at the 5percent level. * Significant at the 10 percent level.Growth in total wages and salaries Growth in average wage and salary(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 0.374∗∗∗ 0.373∗∗∗ 4.589∗∗∗ 0.790∗∗∗ 0.295∗∗∗ 0.280∗∗∗ 0.279∗∗∗ 2.622∗∗∗ 0.499∗∗∗ 0.244∗∗∗(3.44) (3.43) (5.66) (7.08) (2.82) (3.88) (3.84) (5.67) (9.22) (3.14)RSHz,t 1.613 1.379 3.618 8.467∗ -7.714 1.684 1.510 2.754 -6.006∗∗ -3.177(0.36) (0.31) (0.82) (1.88) (-0.86) (1.03) (0.91) (1.56) (-2.40) (-0.77)ISTt ×RSHz,t -12.91∗∗∗ -7.176∗∗∗(-5.25) (-4.87)Trend -0.178∗∗∗ -0.166∗∗∗ -0.177∗∗∗ -0.178∗∗∗ -0.150∗∗∗ -0.190∗∗∗ -0.130∗∗∗ -0.120∗∗∗ -0.129∗∗∗ -0.129∗∗∗ -0.0901∗∗∗ -0.140∗∗∗(-19.47) (-13.74) (-15.99) (-16.24) (-6.57) (-13.98) (-17.41) (-13.56) (-18.44) (-18.39) (-11.26) (-20.17)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11520 11520 11520 11520 5761 5761 11520 11520 11520 11520 5761 5761R2 0.230 0.223 0.230 0.238 0.172 0.255 0.296 0.287 0.296 0.302 0.158 0.33491Table 3.11: Summary Statistics: Aggregate VariablesAll data is quarterly. All series are from 1966 to 2012 except for job creation, destructionand net job creation rates, which is from 1966 to 2005. IST measure cum. 8 quarters isdefined as the changes in the I-shock measure cumulated over the past eight quarters. Theunderlying I-shock measure is standardized to mean zero and standard deviation of one.Personal income, number of employees, unemployment rate, duration of unemployment andpersonal expenditures are from U.S. Bureau of Economic Analysis, retrieved from FRED,Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org. Job creation, destruction andnet job creation rates are from Davis et al. (2006).Mean Median St. dev. Min MaxPersonal income ∆% 1.452 1.35 0.978 -1.50 5.70Unemployment rate ∆% 0.020 0.00 0.344 -0.90 1.70Dur. of unemployment ∆ weeks 0.144 0.00 0.827 -1.30 3.70Personal expenditures ∆% 1.455 1.35 0.854 -2.80 4.00Job creation rate % 5.690 5.56 0.950 3.73 8.98Job destruction rate % 5.810 5.61 1.060 3.69 9.53IST measure, cum. 8 quarters 0.800 1.31 3.750 -11.86 9.5492Table 3.12: The Time Profile of the Impact of the IST Shock and Routine Intensity on the Unemployment RateData in panel A is a balanced panel consisting of 722 commuting zones from 1990 to 2012. Data in panel B are the top andbottom thirds of zones ranked by share of routine intensive labor. Estimates in panel A are from panel regressions Yt+1,z,i =β0 + β1ISTt−τ + β2RSHz,t + β3ISTt−τ ×RSHz,t +ψ Trendt + γi + et+1,z,i. Regressions in panel B omit the interaction term. Thedependent variable is the annual change (%) in the unemployment rate in each zone z between year t and t+ 1. The independentvariables are the aggregate IST shock ISTt−τ at lags τ , the share of routine jobs in the zone RSHz,t and their interaction term.All regressions include intercept, linear trend and state fixed effect γi and are weighted by the share of population in each zone.The standard errors are clustered at the state level. *** Significant at the 1, ** at the 5, and * at the 10 percent level.Panel A. Interaction and base effectChange in unemployment rate, interaction effect Change in unemployment rate, base effect(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt -1.647∗∗∗ 0.217∗∗∗(-4.50) (5.20)ISTt ×RSHz,t 5.793∗∗∗(4.92)ISTt−1 -1.937∗∗∗ 0.292∗∗∗(-7.25) (11.42)ISTt−1 ×RSHz,t 6.946∗∗∗(8.18)ISTt−2 0.459∗∗ 0.248∗∗∗(2.16) (6.73)ISTt−2 ×RSHz,t -0.657(-0.90)ISTt−3 0.857∗∗ 0.207∗∗∗(2.62) (7.09)ISTt−3 ×RSHz,t -2.028∗∗(-2.04)ISTt−4 1.764∗∗∗ -0.283∗∗∗(4.94) (-6.12)ISTt−4 ×RSHz,t -6.209∗∗∗(-5.69)ISTt−5 1.966∗∗∗ -0.0645∗(3.65) (-1.71)ISTt−5 ×RSHz,t -6.179∗∗∗(-3.66)RSHz,t -1.904∗∗ -2.750∗∗∗ 1.113∗∗ 1.669∗∗ 2.621∗∗∗ 2.539∗∗∗ 0.812∗∗ 0.765∗ 0.815∗∗ 0.768∗ 0.314 0.477(-2.46) (-4.49) (2.09) (2.52) (4.39) (2.84) (2.03) (1.97) (2.04) (1.89) (0.70) (1.13)Trend 0.00798∗∗∗ 0.00720∗∗∗ 0.00840∗∗∗ 0.00946∗∗∗ 0.0183∗∗∗ 0.0102∗∗∗ 0.00889∗∗∗ 0.00790∗∗∗ 0.00830∗∗∗ 0.00921∗∗∗ 0.0186∗∗∗ 0.0103∗∗∗(3.14) (3.14) (3.61) (4.28) (8.02) (4.47) (3.47) (3.44) (3.58) (4.18) (7.95) (4.38)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 8326 8326 8326 8326 8326 8326 8326 8326 8326 8326 8326 8326R2 0.0277 0.0410 0.0251 0.0199 0.0310 0.0148 0.0215 0.0324 0.0251 0.0192 0.0230 0.0076593Table 3.12 The Time Profile of the Impact of the IST Shock and Routine Intensity on the Unemployment Rate: ContinuedPanel B. Base effect in sub-samplesChange in unemployment rate, zones with non-routine labor Change in unemployment rate, zones with routine-intensive labor(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 0.0369 0.261∗∗∗(0.58) (6.72)ISTt−1 0.148∗∗∗ 0.325∗∗∗(5.12) (11.99)ISTt−2 0.222∗∗∗ 0.252∗∗∗(6.80) (6.15)ISTt−3 0.222∗∗∗ 0.200∗∗∗(4.34) (7.56)ISTt−4 -0.211∗∗∗ -0.305∗∗∗(-3.76) (-6.59)ISTt−5 0.0337 -0.0917∗∗(1.14) (-2.06)RSHz,t 0.304 0.402 0.401 0.217 0.285 0.237 0.694 0.413 0.512 0.243 -1.567 -0.927(0.38) (0.50) (0.47) (0.25) (0.42) (0.30) (0.51) (0.32) (0.41) (0.19) (-1.29) (-0.76)Trend 0.00768∗∗∗ 0.00749∗∗∗ 0.00787∗∗∗ 0.00899∗∗∗ 0.0153∗∗∗ 0.00645∗∗ 0.00897∗∗ 0.00756∗∗ 0.00809∗∗∗ 0.00869∗∗∗ 0.0173∗∗∗ 0.00964∗∗∗(3.03) (3.15) (3.43) (4.30) (8.10) (2.20) (2.68) (2.62) (2.82) (3.17) (5.47) (3.07)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 4163 4163 4163 4163 4163 4163 4163 4163 4163 4163 4163 4163R2 0.00683 0.0124 0.0198 0.0195 0.0143 0.00665 0.0301 0.0408 0.0277 0.0203 0.0279 0.010394Table 3.13: The Time Profile of the Impact of the IST Shock and Routine Intensity on the Job Creation RatePanel consists of 363 MSAs from 1980 to 2012. The estimates are from panel regressions Yt+1,z,i = β0 + β1ISTt−τ + β2RSHz,t +β3ISTt−τ × RSHz,t + ψ Trendt + γi + et+1,z,i. The dependent variables are the annual March-to-March job creation rate (%) ofnew firms (panel A) and incumbents (panel B) in each metropolitan area z between year t and t + 1. The independent variablesare the aggregate normalized IST shock ISTt at lags τ , the MSA’s share of routine intensive jobs RSHz,t and their interactionterm. All regressions include intercept, linear trend and state fixed effect γi and are weighted by the share of population in eachMSA. The standard errors are clustered at the state level. *** Significant at the 1, ** at the 5, and * at the 10 percent level.Panel A.Interaction and base effectJob creation rate, incumbents, interaction Job creation rate, incumbents, base effect(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 0.523∗∗ -0.313∗∗∗(2.10) (-8.39)ISTt ×RSHz,t -2.554∗∗∗(-3.17)ISTt−1 0.0398 -0.264∗∗∗(0.19) (-7.91)ISTt−1 ×RSHz,t -0.924(-1.35)ISTt−2 -0.166 0.0398∗∗(-0.95) (2.15)ISTt−2 ×RSHz,t 0.621(1.15)ISTt−3 -0.532∗ 0.247∗∗∗(-1.87) (7.70)ISTt−3 ×RSHz,t 2.357∗∗(2.57)ISTt−4 -0.282∗ -0.00659(-1.89) (-0.42)ISTt−4 ×RSHz,t 0.833∗(1.78)ISTt−5 0.104 0.0257(0.74) (1.50)ISTt−5 ×RSHz,t -0.237(-0.56)RSHz,t 9.969∗∗∗ 9.566∗∗∗ 9.638∗∗∗ 9.642∗∗∗ 9.630∗∗∗ 9.769∗∗∗ 9.470∗∗∗ 9.404∗∗∗ 9.725∗∗∗ 9.403∗∗∗ 9.681∗∗∗ 9.741∗∗∗(5.34) (5.09) (4.96) (4.85) (4.92) (5.07) (4.90) (4.87) (4.97) (4.85) (4.97) (4.98)Trend -0.0592∗∗∗ -0.0609∗∗∗ -0.0669∗∗∗ -0.0775∗∗∗ -0.0653∗∗∗ -0.0666∗∗∗ -0.0592∗∗∗ -0.0609∗∗∗ -0.0671∗∗∗ -0.0786∗∗∗ -0.0655∗∗∗ -0.0665∗∗∗(-6.77) (-7.11) (-7.64) (-8.82) (-7.59) (-7.63) (-6.75) (-7.10) (-7.75) (-9.16) (-7.69) (-7.69)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946R2 0.378 0.371 0.356 0.374 0.356 0.356 0.376 0.370 0.356 0.372 0.356 0.35695Table 3.13 The Time Profile of the Impact of the IST Shock and Routine Intensity on the Job Creation Rate: ContinuedPanel B. Base effect in sub-samplesJob creation rate, incumbents, MSA with non-routine labor Job creation rate, incumbents, MSA with routine-intensive labor(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt -0.192∗∗∗ -0.338∗∗∗(-4.79) (-8.00)ISTt−1 -0.173∗∗∗ -0.282∗∗∗(-6.46) (-6.96)ISTt−2 0.0685∗∗ 0.0377∗(2.30) (1.81)ISTt−3 0.217∗∗∗ 0.250∗∗∗(4.92) (6.79)ISTt−4 -0.0244 -0.00146(-1.13) (-0.08)ISTt−5 0.0565∗∗∗ 0.0247(2.70) (1.16)RSHz,t 9.186∗∗∗ 9.140∗∗∗ 9.155∗∗∗ 8.131∗∗∗ 9.042∗∗∗ 9.350∗∗∗ 11.20∗∗ 11.03∗∗ 12.04∗∗ 11.21∗∗ 11.94∗∗ 12.09∗∗(3.18) (3.19) (3.14) (2.75) (3.14) (3.21) (2.31) (2.26) (2.44) (2.28) (2.43) (2.44)Trend -0.0830∗∗∗ -0.0837∗∗∗ -0.0890∗∗∗ -0.0966∗∗∗ -0.0856∗∗∗ -0.0888∗∗∗ -0.0531∗∗∗ -0.0551∗∗∗ -0.0611∗∗∗ -0.0733∗∗∗ -0.0598∗∗∗ -0.0605∗∗∗(-7.05) (-7.06) (-7.08) (-7.52) (-7.33) (-7.19) (-5.29) (-5.59) (-6.09) (-7.12) (-5.96) (-6.04)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974R2 0.311 0.310 0.305 0.313 0.305 0.306 0.409 0.401 0.384 0.401 0.384 0.38496Table 3.14: The Time Profile of the Impact of the IST Shock and Routine Intensity on the Job Destruction RatePanel consists of 363 MSAs from 1980 to 2012. The estimates are from panel regressions Yt+1,z,i = β0 + β1ISTt−τ + β2RSHz,t +β3ISTt−τ ×RSHz,t +ψ Trendt + γi + et+1,z,i. The dependent variables are the annual March-to-March job destruction rate (%) offirm closures (panel A) and incumbents (panel B) in each metropolitan area z between year t and t+ 1. The independent variablesare the aggregate normalized IST shock ISTt at lags τ , the MSA’s share of routine intensive jobs RSHz,t and their interactionterm. All regressions include intercept, linear trend and state fixed effect γi and are weighted by the share of population in eachMSA. The standard errors are clustered at the state level. *** Significant at the 1, ** at the 5, and * at the 10 percent level.Panel A. Interaction and base effectJob destruction rate, incumbents, interaction Job destruction rate, incumbents, base effect(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt -0.653 0.461∗∗∗(-1.50) (9.74)ISTt ×RSHz,t 3.402∗∗(2.55)ISTt−1 -0.160 0.595∗∗∗(-0.56) (18.55)ISTt−1 ×RSHz,t 2.303∗∗(2.43)ISTt−2 -0.336 0.297∗∗∗(-1.37) (7.39)ISTt−2 ×RSHz,t 1.906∗∗(2.46)ISTt−3 -0.492∗∗ -0.292∗∗∗(-2.12) (-8.41)ISTt−3 ×RSHz,t 0.604(0.77)ISTt−4 0.203∗∗ -0.0547∗∗(2.68) (-2.60)ISTt−4 ×RSHz,t -0.781∗∗∗(-3.39)ISTt−5 -0.289∗∗∗ -0.105∗∗∗(-2.79) (-6.65)ISTt−5 ×RSHz,t 0.556∗(1.73)RSHz,t 6.012∗∗∗ 6.593∗∗∗ 6.366∗∗∗ 6.755∗∗∗ 6.349∗∗∗ 6.072∗∗∗ 6.676∗∗∗ 6.997∗∗∗ 6.634∗∗∗ 6.694∗∗∗ 6.300∗∗∗ 6.140∗∗∗(3.04) (3.38) (3.59) (3.85) (3.64) (3.51) (3.70) (3.82) (3.73) (3.88) (3.63) (3.55)Trend -0.0477∗∗∗ -0.0491∗∗∗ -0.0478∗∗∗ -0.0227∗∗∗ -0.0364∗∗∗ -0.0347∗∗∗ -0.0478∗∗∗ -0.0491∗∗∗ -0.0484∗∗∗ -0.0230∗∗∗ -0.0361∗∗∗ -0.0349∗∗∗(-10.06) (-9.54) (-7.97) (-3.83) (-7.11) (-7.01) (-10.13) (-9.59) (-8.00) (-3.76) (-7.05) (-7.06)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946 11946R2 0.211 0.240 0.178 0.185 0.163 0.166 0.208 0.239 0.177 0.185 0.163 0.16697Table 3.14 The Time Profile of the Impact of the IST Shock and Routine Intensity on the Job Destruction Rate: ContinuedPanel B. Base effect in sub-samplesJob destruction rate, incumbents, MSA with non-routine labor Job destruction rate, incumbents, MSA with routine-intensive labor(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 0.313∗∗∗ 0.491∗∗∗(4.88) (10.60)ISTt−1 0.449∗∗∗ 0.627∗∗∗(17.96) (17.70)ISTt−2 0.249∗∗∗ 0.306∗∗∗(4.49) (7.14)ISTt−3 -0.325∗∗∗ -0.287∗∗∗(-9.83) (-7.27)ISTt−4 0.0195 -0.0679∗∗∗(0.89) (-3.16)ISTt−5 -0.0401 -0.118∗∗∗(-1.64) (-6.59)RSHz,t -4.259 -4.237 -3.827 -2.662 -4.056 -4.279 8.511 9.471∗ 8.274 8.284 7.229 6.727(-0.97) (-0.97) (-0.89) (-0.62) (-0.92) (-0.97) (1.58) (1.75) (1.55) (1.59) (1.38) (1.28)Trend -0.0579∗∗∗ -0.0595∗∗∗ -0.0607∗∗∗ -0.0369∗∗∗ -0.0527∗∗∗ -0.0504∗∗∗ -0.0432∗∗∗ -0.0439∗∗∗ -0.0434∗∗∗ -0.0179∗∗∗ -0.0309∗∗∗ -0.0301∗∗∗(-7.31) (-7.24) (-6.35) (-4.40) (-6.06) (-5.85) (-7.43) (-7.17) (-6.63) (-2.75) (-5.44) (-5.57)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974 5974R2 0.182 0.199 0.174 0.186 0.166 0.167 0.224 0.259 0.185 0.192 0.170 0.17498Table 3.15: The Time Profile of the Impact of the IST Shock and Routine Intensity on the Growth in the Total Wages and SalariesPanel consists of 722 commuting zones from 1980 to 2012. The estimates are from panel regressions Yt+1,z,i = β0 + β1ISTt−τ +β2RSHz,t+β3ISTt−τ ×RSHz,t+ψ Trendt+γi+et+1,z,i. The dependent variable are the annual growth (%) in the total wages andsalaries and in the average wage and salary in each zone z between year t and t+ 1. The independent variables are the aggregatenormalized IST shock ISTt−τ at lags τ , the zone’s share of routine intensive jobs RSHz,t and their interaction term. All regressionsinclude intercept, linear trend and state fixed effect γi and are weighted by the share of population in each zone. The standarderrors are clustered at the state level. *** Significant at the 1, ** at the 5, and * at the 10 percent level.Panel A. Interaction and base effectGrowth in total wages and salaries, interaction effect Growth in total wages and salaries, base effect(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 4.234∗∗∗ 0.339∗∗∗(5.71) (3.31)ISTt ×RSHz,t -11.93∗∗∗(-5.29)ISTt−1 3.868∗∗∗ -0.445∗∗∗(6.01) (-4.80)ISTt−1 ×RSHz,t -13.20∗∗∗(-6.61)ISTt−2 0.923∗ -0.660∗∗∗(1.78) (-9.57)ISTt−2 ×RSHz,t -4.835∗∗∗(-2.72)ISTt−3 1.317∗∗∗ -0.0364(3.95) (-0.66)ISTt−3 ×RSHz,t -4.124∗∗∗(-3.90)ISTt−4 -0.0436 0.315∗∗∗(-0.08) (5.33)ISTt−4 ×RSHz,t 1.081(0.66)ISTt−5 -0.777 -0.412∗∗∗(-1.60) (-6.57)ISTt−5 ×RSHz,t 1.104(0.79)RSHz,t 3.290 4.657 1.968 2.169 1.628 2.072 1.268 1.113 1.024 1.445 1.780 1.960(0.82) (1.12) (0.49) (0.53) (0.40) (0.48) (0.31) (0.27) (0.25) (0.35) (0.43) (0.47)Trend -0.174∗∗∗ -0.156∗∗∗ -0.151∗∗∗ -0.164∗∗∗ -0.175∗∗∗ -0.143∗∗∗ -0.174∗∗∗ -0.156∗∗∗ -0.151∗∗∗ -0.164∗∗∗ -0.175∗∗∗ -0.143∗∗∗(-17.84) (-14.55) (-13.70) (-15.80) (-16.60) (-10.60) (-17.63) (-14.61) (-13.69) (-15.80) (-16.62) (-10.91)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 11880 11880 11880 11880 11880 11880 11880 11880 11880 11880 11880 11880R2 0.259 0.268 0.273 0.247 0.251 0.258 0.252 0.257 0.271 0.246 0.251 0.25899Table 3.15 The Time Profile of the Impact of the IST Shock and Routine Intensity on the Growth in the Total Wages and Salaries:ContinuedPanel B. Base effect in sub-samplesGrowth in total wages and salaries, zones with non-routine labor Growth in total wages and salaries, zones with routine-intensive labor(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)ISTt 0.737∗∗∗ 0.265∗∗(7.43) (2.65)ISTt−1 0.185∗ -0.576∗∗∗(1.91) (-6.72)ISTt−2 -0.267∗∗∗ -0.742∗∗∗(-7.79) (-8.73)ISTt−3 0.127 -0.0751(1.61) (-1.30)ISTt−4 0.355∗∗∗ 0.302∗∗∗(5.71) (4.39)ISTt−5 -0.519∗∗∗ -0.387∗∗∗(-4.63) (-6.69)RSHz,t 6.762 9.431∗∗ 9.465∗∗ 9.282∗∗ 9.697∗∗ 11.60∗∗∗ -6.916 -7.905 -8.106 -6.731 -5.666 -5.351(1.65) (2.43) (2.44) (2.36) (2.40) (2.93) (-0.84) (-0.97) (-1.00) (-0.81) (-0.68) (-0.64)Trend -0.149∗∗∗ -0.137∗∗∗ -0.129∗∗∗ -0.136∗∗∗ -0.146∗∗∗ -0.110∗∗∗ -0.185∗∗∗ -0.167∗∗∗ -0.163∗∗∗ -0.176∗∗∗ -0.188∗∗∗ -0.157∗∗∗(-7.78) (-7.09) (-6.39) (-6.95) (-7.01) (-4.44) (-14.87) (-12.89) (-12.51) (-14.08) (-15.12) (-10.59)State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 5941 5941 5941 5941 5941 5941 5941 5941 5941 5941 5941 5941R2 0.200 0.176 0.178 0.175 0.179 0.191 0.274 0.289 0.302 0.271 0.274 0.281100Chapter 4Capital Intensity and InvestmentShocks: Implications for StockReturns4.1 IntroductionTechnological innovations have been identified as the main driver of economic growth, Green-wood et al. (1997).54 A large part of technological innovations is embodied in the formationof new capital and has been labeled as investment-specific technology shocks (IST), or moreconcisely, investment shocks. Recent finance literature, e.g. Papanikolaou (2011), suggeststhat investment shocks can also affect expected stock returns. Hence, measuring firm’s ex-posure to these shocks is crucial for understanding the relationship between firms’ expectedstock returns and the associated sources of risk.In this paper, I study how firm’s capital intensity relates to the measurement of firms expo-sure to these shocks using a popular measure, stock return differences between capital-goodsand consumption-goods producers, shortly and hereafter the IMC portfolio. An intuitionbased on previous literature suggests that technological innovations embodied in new capitalare expected to be relevant especially for capital-intensive firms as capital is the key produc-tion factor for these firms. In contrast, labor-intensive firms are expected not to be directlyaffected by such innovations.55I show that the pattern documented by previous literature, e.g. Kogan and Papaniko-laou (2014), of decreasing abnormal stock return of stocks sorted by the exposure to theIMC portfolio is present among capital-intensive firms, but it is almost absent among labor-intensive firms. This suggests that the documented pattern is driven mainly by firms in thefirst sub-sample. Specifically, I divide the cross-section of firms into capital-intensive andlabor-intensive firms. I use the IMC portfolio as an empirical measure of investment shocksbased on the return spread between investment-goods and consumption-goods firms, as pro-posed by Papanikolaou (2011). While sorting firms by their exposure to the IMC portfoliogives a statistically significant abnormal return of 7% among capital-intensive firms, the samesorting among labor-intensive firms leads to insignificant abnormal return of low magnitude.Interestingly, the estimated exposures to the IMC portfolio span approximately the samerange in capital-intensive and labor-intensive firms. This result is puzzling since the sameexposure to the IMC portfolio seems to be priced differently among each types of firms.54Greenwood et al. (2000); Fisher (2006); Justiniano et al. (2010) also identify the investment-specifictechnology shocks as a major source of business-cycle fluctuation.55This statement is based on the assumption that firms do not substitute between capital and labor andthe price (and productivity) of labor is not affected by investment-specific technology shocks.101I provide a potential explanation for these observations. The IMC portfolio is intended tomeasure the IST shocks, a risk that is assumed to carry negative premium. At the same time,the IMC portfolio has an empirically strong and positive exposure to the market risk andthe size factor (SMB), both of which are priced positively. Labor-intensive firms, in general,use only limited capital and their exposure to the IMC portfolio does not arise due to theirexposure to the investment shocks but due to other reasons, e.g., an exposure to the marketand size factors. In contrast, capital-intensive firms use large amounts of capital and hencetheir exposure to the IMC portfolio can arise due mainly to their exposure to the investmentshocks. In such setting, the IMC portfolio can be a reasonable proxy for measuring theexposure to the investment shocks among capital-intensive firms but fails to capture thisexposure among labor-intensive firms. Since the IMC portfolio is an easily available measureof investment shocks that is available at high frequency it is important to understand itscapability.I extend the existing models of the investment-specific technology shocks in a simple andtractable fashion to illustrate the economic mechanisms and to analyze the quantitative as-pects. I study the asset pricing implications of a firm’s capital intensity in a framework withboth the total productivity (disembodied) technology shock and the investment-specific (em-bodied) technology shock, building on the two-sector model from Kogan and Papanikolaou(2014). This model consists of consumption-goods and investment-goods producers and dif-ferentiates between the factor neutral and the investment-specific shocks. The novel aspectsof my model are firms with different capital (or labor) intensities, their potential to resaleobsolete capital, and wage rigidity.I focus on two sub-samples of firms, one consisting of firms with a high capital-laborratio and one of firms with a low capital-labor ratio. Capital intensity creates a differencein the exposure of firms’ growth opportunities to the investment shocks between capital-intensive and labor-intensive firms. In the model, capital-intensive firms use larger amountsof capital in their (potential) production and hence their growth opportunities are exposed tothe investment shocks, which change the price of capital. Wage rigidity generates operatingleverage and makes a firm more exposed to market risk. Since labor-intensive firms tendto optimally choose higher use of labor, they have the capacity to be highly levered and beexposed to the aggregate risk more than capital-intensive firms.In the model, I allow the IMC portfolio to be positively correlated with both the aggregaterisk and investment shocks as documented by empirical evidence. In such setting, bothcapital- and labor-intensive firms can have high exposure to the IMC portfolio while differentexposures to the investment shock. The model can generate high abnormal return for capital-intensive stocks sorted on their exposure to the IMC portfolio and a considerably smallerabnormal return for labor-intensive firms.This paper contributes to two streams of literature. First, a large part of macroeconomicliterature studies the implications of investment shocks for growth and other macroeconomicvariables. Representative papers in this area are Greenwood et al. (1997), Greenwood et al.(2000) and Fisher (2006). They show that investment shocks account for a large part ofeconomic growth as well as for variations in output and other macroeconomic variables.Justiniano et al. (2010) show that investment shocks are the main driver for business cyclefluctuations. Greenwood et al. (1988) investigates the role of capacity utilization for businesscycles and positive correlation between macroeconomic variables in a model with investment-102specific shocks. Second, investment shocks have become an active research area in financialeconomics. Papanikolaou (2011) is the first to study the implications of investment shocks forasset prices both on aggregate and in the cross-section. Garlappi and Song (2017) study theimplications of capital utilization and market power for pricing assets exposed to investmentshocks. Kogan and Papanikolaou (2013) and Kogan and Papanikolaou (2014) focus on theimplications of investment shocks for growth options, investments and several return anoma-lies. Garlappi and Song (2016) examine empirically how various measures of the investmentshocks price a range of cross-sectional return anomalies. Garlappi and Song (2018) use dataon firms’ investment to infer firms’ exposure to the investment shocks.I differ from this literature in that I introduce to the prior asset pricing models withinvestment shocks a novel dimension, firms’ capital intensity. In such extended model, thecapital intensity determines the exposure of a firm’s growth opportunities to investmentshocks and hence also the riskiness of the whole firm. I further allow the measure of investmentshocks, the IMC portfolio, to correlate with market returns as observed in the data as wellas partially rigid wages. These extensions to the model are helpful as they allow the modelto further examine the role of investment shocks for firms’ growth opportunities and theestimation of their exposure to these shocks based on the IMC portfolio.The paper proceeds as follows. Section 4.2 describes the data and first empirical results.Section 4.3 shows the extended model. In section 4.4 I derive the empirical implications andprovide supportive empirical evidence. I describe the calibration and simulation results insection 4.5 and conclude in section 4.6.4.2 Data and Empirical Evidence4.2.1 DataThe data on stock prices are from CRSP. I use the universe of ordinary common stocks(shrcd=10, 11) of firms traded on NYSE, AMEX and NASDAQ (exchcd= 1, 2, 3) in the timeperiod from 1950 to 2015. I exclude financial firms (SIC 6000-6799) and utilities (SIC 4900-4949). In order to categorize the firms into investment-goods producers and consumption-goods producers, I follow the previous literature (Garlappi and Song (2016), Gomes et al.(2009) and Papanikolaou (2011)) and use the NIPA Input-Output tables from 1987 andcategorize the firms into investment-goods and consumption-goods producers based on theircontribution to each sector. Accounting data are from Compustat. I measure firm’s capitalintensity by the number of employees over property, plant and equipment (empf,tppegtf,t).Table 4.3 shows the summary statistics of firms categorized into investment-goods andconsumption-goods sector. The investment-goods sector is smaller than the consumption-goods sector.56 The firms in the consumption-goods sector are similar to firms in theinvestment-goods sector in terms of book-to-market equity ratio and cash flow-to-assets ratio,but differ slightly in operating leverage, capital, number of employees and capital-labor ratio.These differences, however, are rather small compared to the 10th and 90th percentiles ofthese variables.I construct the IMC portfolio following the methodology in Kogan and Papanikolaou56Papanikolaou (2011) uses 1997-NIPA Input-Output tables based on NAICS code and identifies even highernumber of firms in the consumption-goods sector.103(2014) and Garlappi and Song (2016). First, I calculate the value-weighted return for port-folios consisting of investment-goods firms (I-portfolio) and consumption-goods firms (C-portfolio), respectively. Then, I create the IMC (Investment Minus Consumption), consistingof long position in the I-portfolio and short position in the C-portfolio. Since this measureof the IST shock is based on stock returns, it is available at the frequency of stock returns.For further analysis, I use monthly stock returns.I estimate the exposure of each consumption-goods firm to the IMC portfolio by estimatingthe βIMCf,t from following regression equation:Rf,t −Rriskfree = αf,t + βIMCf,t ×RIMCt + f,t. (4.1)I use a rolling and overlapping window of monthly returns over the last 60 months. Accord-ingly, firm f ′s βIMCf,t at time t is estimated from monthly returns ranging from t − 60 tot− 1.57 The betas are updated annually at the end of June.I sort the stocks by their estimated βIMCf,t into 5 portfolios annually at the end of June.The return of each portfolio is the weighted average return of the stocks in that portfolio. Iconstruct the time series of the portfolios from 1970 to 2015. Following the existing literature,e.g., Kogan and Papanikolaou (2013, 2014), I focus on the universe of consumption-goodsstocks.Table 4.1 shows the pairwise correlation coefficients between the IMC portfolio and themarket excess return, and the SMB and HML factors. Th IMC portfolio is negativelycorrelated (-0.25) with the value factor HML which is consistent with the intuition in Koganand Papanikolaou (2014), which is described later in the text. The IMC portfolio, however,is also positively correlated with the two other factors, namely the market return and theSMB with correlation coefficient of 0.45 in both cases. This suggests that sorting stocksby their βIMCf,t is likely to capture exposure to these two factors as well. Stocks which havegenerally higher exposure to market and SMB will tend to have a higher βIMCf,t .4.2.2 βIMC SortingFirst I focus on the return pattern of portfolios sorted by their exposure to the investmentshock. Table 4.2 presents the return characteristics of portfolios sorted by the exposure tothe IMC portfolio based on the βIMC estimated from a univariate regression, which is thestandard in previous literature. Panel A shows that sorting stocks by the βIMCf,t creates onlya very weak decreasing pattern in average excess returns. The estimates in panel B showthat the abnormal return α decreases more strongly across the portfolios due mostly to theincreasing loadings on the market return. The difference in α between the portfolios at eachend amount to statistically significant -6.48%. Papanikolaou (2011); Kogan and Papanikolaou(2014) first document this pattern and show that the decreasing returns coincide with adecreasing exposure to the HML portfolio creating a possible relationship between firms’growth opportunities and the exposure to the IMC portfolio.In Panel C, I regress the portfolio returns onto the Fama and French (1993) three factors,the market, SMB, and HML. The loadings on the HML factor are not significant for thefirst four portfolio but strongly negative for portfolio five indicating that firms with high57Kogan and Papanikolaou (2014) show that it is possible to use also weekly returns which highlights theadvantage of a high frequency measure.104exposure to the IMC portfolio have more growth opportunities. The HL portfolio, then, hasa significant and negative exposure to the HML factor. The loadings on the market and theSMB factors are increasing across the portfolios, so that the HL portfolio has a significantlypositive exposure to these two factors. This is consistent with the evidence of a positivecorrelation between the IMC portfolio and these two factors shown in table 4.1 suggestingthat sorting by the βIMC will result in sorting on the exposure to these two factors as well.It is interesting to observe that the pattern of the abnormal returns, α, is also presentwhen the market, SMB, and HML factors are included. Although the original model ofKogan and Papanikolaou (2014) creates a potential link between firm’s exposure to investmentshocks and its growth opportunities, the extended model later in this text shows that a firmcan have a differential exposure to investment shocks due to the potential resale of its capitalregardless of its growth opportunities.Panel D shows the median sorting βIMC of each portfolio together with the portfolios’estimated post-sorting exposure to the IMC portfolio. The post-sorting exposures are esti-mated from an univariate regression with monthly portfolio returns. The clearly increasingpattern of both, the sorting and post-sorting exposures, shows that sorting stocks into port-folios based on βIMC translates the exposure to the IMC portfolio to subsequent portfolioreturns.4.2.3 Capital Intensity and βIMC SortingI divide the universe of the consumption-goods firms into a sub-sample of capital-intensivefirms and a sub-sample of labor-intensive firms to analyze the patterns of the abnormal returnsin each sample separately. The motivation for this approach is based on the intuition that thegrowth opportunities of capital-intensive firms would be more sensitive to technology shocksembodied in capital goods. I measure the capital-intensity of each firm by the ratio of itscapital to the number of employees,ppegtf,temployeesf,t, where ppegtf,t stands for firm f′s property,plant and equipment in year t. I allocate the firm into capital-intensive sub-sample if itscapital-intensity is above the cross-sectional median in year t and into labor-intensive if it isbelow.Table 4.4 shows the summary statistics of the capital-intensive and labor-intensive sub-samples for firms in consumption-goods, investment-goods sector as well as both sectorstogether. Capital-intensive firms are generally bigger as their market capitalization is abouttwo to three times higher than the market capitalization of labor-intensive firms. Obviously,capital-intensive firms differ from labor-intensive firms in their capital-labor ratio. Interest-ingly, these differences are driven especially by the numerator of this ratio, capital, as bothtypes of firms tend to have comparable number of employees. It is important to observethat the book-to-market ratio and its ranges are comparable for capital-intensive and labor-intensive firms. In contrast, the operating leverage and the difference between its 90th and10th percentile are higher for labor-intensive firms than for capital-intensive firms.Next, I sort the firms in each sub-sample into quintile portfolios by their βIMCf,t as inprevious sections. Table 4.5 reports the return characteristics of the portfolios sorted byβIMC in each sub-sample. Panel A shows a stronger pattern of decreasing excess returnsacross the portfolios of capital-intensive firms than observed above. In contrast, the averageexcess returns are increasing across the βIMC sorted portfolios for labor-intensive firms. Theestimates in panel B show that the abnormal return. α, is exhibits very strong and decreasing105pattern for capital-intensive firms which is even slightly stronger than the pattern observed inthe full sample. The α of the long-short, HL, portfolio amounts to -7.56% at per year. Suchpattern is, however, considerably weaker among labor-intensive firms. The α of the long-short portfolio is only -2.87% per year in this sub-sample. This suggests that the pattern inα identified in the previous section is driven by capital-intensive firms.Panel C shows the estimated loadings on the market, SMB and HML factors. Theloadings of the portfolios on the HML factor decrease across the portfolios more in the labor-intensive than in the capital-intensive sub-sample. At the same time, capital-intensive firmsexhibit somewhat higher differences in the loadings on the SMB factor. This observationmay suggests that a relatively low growth opportunities of the capital-intensive long-shortportfolio (a negative exposure to the HML factor of only -0.31) is sufficient to result in a largeabnormal return α despite a large exposure to the SMB factor. In contrast, labor-intensiveHL portfolio seems exhibit much higher growth opportunities (negative loading to the HMLfactor of -0.59) but earns considerably smaller α.Panel D then shows portfolios in both sub-samples have almost the same pattern in theirexposure to the IMC portfolio. This result is puzzling for two reasons. First, the sameloadings of the HL capital- and labor-intensive portfolios on the IMC portfolio result in verydifferent α. Second the capital- and labor-intensive HL portfolios have also very differentexposure to the HML factor despite having almost the same exposure to the IMC portfolio.The loading of the HL portfolio on the market risk is comparable for both sub-samples, butthe capital-intensive firms with high exposure to the IMC portfolios earn lower average andabnormal returns than capital-intensive firms with low exposure to investment shock. Themodel in the subsequent section will seek to reconcile these discrepancies by the differencesin firms’ capital intensity.4.3 The ModelMotivated by the empirical evidence, I build on the partial-equilibrium model of Kogan andPapanikolaou (2014) (KP) and extend it by introducing capital intensity and the potentialresale of firms’ existing assets to see whether these two dimensions can reconcile the empiricalobservations. KP model the cross-section of consumption-goods firms, while the investment-goods firm is modeled in a simplified reduced form so that the model is able to generate theIMC portfolio as a potential measure of an investment shock. The extensions I add to themodel allow three things. First the IMC portfolio can be exposed not only to the investmentshocks but also the a factor-neutral aggregate productivity shock. Second, they further allowto analyze the potential bias in the βIMC as a measure of a firm’s exposure to investmentshocks. Third, they allow to examine how capital-intensity can alter the link between a firm’sgrowth opportunities and its exposure to the investment shocks.4.3.1 The Cash Flow of Consumption-Goods FirmsThe universe of the consumption-goods firms consists of two sectors s ∈ {L,H}, where L andH denotes the low and high capital intensive firms, respectively. Each of these two sectorsconsists of set of firms Fs. Each firm consists of an individual number of projects enumeratedby j ∈ Jft . Firms create projects by investment in productive capital and by hiring labor106when a new project opportunity arrives. Project j owned by firm f in sector s generatesoutput equal to:yf,j,t = f,tuj,txtKαsj L1−αˆsj (4.2), where f,t is a firm-specific shock affecting all project owned by firm f , uj,t is a project-specific shock affecting only project j, and xt is an aggregate shock affecting all projects of allfirms. The firm- and project-specific shocks are governed by mean-reverting processes, whilethe aggregate shock evolves as geometric Brownian motion to simulate aggregate growth:df,t = −θ (f,t − 1) dt+ σ√f,tdBf,t (4.3)duj,t = −θu (uj,t − 1) dt+ σu√uj,tdBj,t (4.4)dxt = µxxtdt+ σxxtdBx,t (4.5)αs and αˆs determine the capital intensity of the firms in each sector s. Holding αˆs−αs fixedacross both sectors, αs is higher for capital-intensive sector. Moreover, αs + (1− αˆs) < 1 tosuffice decreasing returns to scale. New capital Kj,t can be acquired at price xtz−1t , where ztrepresents the investment shock and is governed by geometric Brownian motion:dzt = µzztdt+ σzztdBz,t (4.6)The projects expire randomly according to Poisson process with a constant arrival rate δ.When the project expires, the capital will be re-sold at the current price of capital to otherfirms demanding capital.The total cash flow of the project consists of three components, (i) cash inflow generatedby production CFIj,t, (ii) cash outflow due to labor cost CFOj,t and (iii) cash inflow fromcapital re-sale RSj,t when the project expires. The value of each of this components is derivedlater in the text.The stochastic discount factor is defined exogenously and is motivated by Papanikolaou(2011):dpitpit= −r dt− γx dBx,t − γz dBz,t (4.7)This specification includes two priced shocks, the aggregate shock xt with price γx > 0 andthe investment shock zt with price γz < 0, where the sign of the price of each shock is basedon the assumption in Kogan and Papanikolaou (2014).Time t value of cash inflow generated by an existing project j is:CFIj,t (f,t, uj,t, xt, w,Kj , Lj) = Et[∫ ∞te−δ(s−t)pispitf,suj,sxsKαsj L1−αˆsj ds]= A (f,t, uj,t)xtKαsj L1−αˆsj (4.8), whereA =[1r + γxσx + δ − µx +f,t − 1r + γxσx + δ − µx + θ+uj,t − 1r + γxσx + δ − µx + θu +(uj,t − 1) (− 1)r + γxσx + δ − µx + θu + θ](4.9)107I assume inelastic labor (i.e., infinite supply of labor for a given wage). The wage is givenexogenously and has the same dynamics as the aggregate shock, so that the wage isWt = w ∗ xt, (4.10)where w is a positive constant. This assumption is reasonable for this type of partial-equilibrium model.58 I assume that a fraction of the hired labor force, v, has flexible wage,i.e. their wage evolves stochastically over the project lifetime as specified in formula (4.10).The remaining fraction 1 − v has a rigid wage, i.e. the wage of this labor force is locked tothe wage level at the arrival of the project (Tj) and stays so for the project’s lifetime. Theparameter v allows to model a degree of wage rigidity in a tractable way without a timedependence.In this model, the wage rigidity creates an operating leverage and helps to differentiatebetween the riskiness of firms in each sector s. In the data, capital-intensive and labor-intensive firms differ. Labor-intensive firms tend to be smaller, have higher volatility ofreturns, and higher exposure to market return and to the SMB factor. All this tends toincrease the riskiness of firm and their exposure to positively priced factors. For the sake ofsimplicity of the model, I use operating leverage as the only source of a potentially differentexposure of labor-intensive firms to the aggregate shock xt. The time t value of labor cost ofan existing project j is:CFOj,t = Et[∫ ∞tve−δ(s−t)pispitxswLj + (1− v)e−δ(s−t)pispitxTjWsLjds]=[vxt(Bflex)−1+ (1− v)xTj(Brig)−1]wLj , (4.11)where(Bflex)−1= 1r+γxσx+δ−µx ,(Brig)−1= 1r+δ and xTj is the level of aggregate productivityat the time of project j′s arrival, so that the wage for the project j′s (1− v) fraction of laborforce is constant at xTjw.The expected time t value of the cash flow from the capital re-sale is:RSj,t = Et[∫ ∞tδe−δ(s−t)pispitxsz−1s Kj]= xtz−1t MKj , (4.12)where M = δr+δ+−µx+µz−σ2z+γxσx−γzσz .The time t value of project j is the sum of all cash flow components, cash inflow, re-sale,and cash outflow, generated by the project:p (f,t, uj,t, xt, zt, w,Kj , Lj) = CFIj,t +RSj,t − CFOj,t= A (f,t, uj,t)xtKαsj L1−αˆsj + xtz−1t MKj−[vxt(Bflex)−1+ (1− v)xTj(Brig)−1]wLj . (4.13)New projects arrive to each firm randomly according to a Poisson process with a firm-specific arrival rate λf,t. The firm-specific arrival rate itself is a random variable:λf,t = λf × λ˜f,t, (4.14)58In general equilibrium, the wage would be determined by supply of labor from household and demand oflabor from the firms.108where λf is a firm-specific constant and λ˜f,t underlies two-state Markov process with valuesλ˜f,t ∈ {λH , λL} and with transition probability matrix (between t and t+dt):P =(1− µLdt µLdtµHdt 1− µHdt)(4.15)4.3.2 Firms’ Optimal Capital and Labor DecisionsEach project j arrives with project-specific productivity at the long-term mean uj,t = 1.When a project j arrives, the firms f chooses labor Lj and capital Kj to maximize NPV:NPV = A (f,t, 1)xtKαsj L1−αˆsj + xtz−1t MKj − z−1t xtKj−[vxt(Bflex)−1+ (1− v)xTj(Brig)−1]wLj (4.16)The first order condition for Lj gives:L∗j =(vxtBflex+(1− v)xTjBrig)− 1αˆsA (f,t, 1)1αˆs Kαsαˆsj(1w) 1αˆsx1αˆst (1− αˆs)1αˆs (4.17)Note that at the project j′s arrival at t = Tj , xt and xTj cancel out of this formula. Usingthis expression in the NPV formula (4.16) and taking the first order condition for Kj givesoptimal investment:K∗j =(αˆsαs) αˆsαs−αˆs (z−1t (1−M)) αˆsαs−αˆs A (f,t, 1)−1αs−αˆs D (αˆs, w)−αˆsαs−αˆs , (4.18)where D (αˆs, w) =(v(Bflex)−1+ (1− v) (Brig)−1) αˆs−1αˆs ((1−αˆsw ) 1−αˆsαˆs − w (1−αˆsw ) 1αˆs).4.3.3 Valuation of Consumption-Goods FirmsThe time t value of a firm is the sum of the values of all existing projects, i.e., value of assetsat place, and the sum of the positive values of projects that are expected to arrive in future,i.e., firm’s growth opportunities. The time t value of firm f ′s existing projects is:V APf,t =Jf∑j∈Jfp (f,t, uj,t, xt, zt, w,Kj , Lj) =Jf∑j∈JfCFIj,t +RSj,t − CFOj,t= CFIf,t +RSf,t − CFOf,t. (4.19)The present value of firm f ′s growth opportunities, PV GOf,t, is the sum of the net presentvalues, NPVf,t, of all future projects:NPVf,t = xtA (f,t, 1)−1αs−αˆs(1−Mzt) αsαs−αˆsD (αˆs, w)−αˆsαs−αˆs[(αˆsαs) αsαs−αˆs −(αˆsαs) αˆsαs−αˆs].(4.20)109The expression for the value of firm f ′s growth opportunities can be concisely written as:PV GOf,t = xtzαsαˆs−αst G (f,t, λf,t, αs, αˆs, w) , (4.21)where G (f,t, λf,t, αs, αˆs, w) is defined in appendix.The firm f ′s total value is then:Vf,t =Jf∑j∈Jfp (f,t, uj,t, xt, w,Kj , Lj) + xtzαsαˆs−αst G (f,t, λf,t, αs, αˆs, w) (4.22)4.3.4 Stock Returns of Consumption-Goods FirmsThe expected excess return on firm f in the consumption-goods sector is:1dtEt [Rf,t]− rf = −cov(d Vf,tvf,t,d pitpit). (4.23)Explicit closed-form expression for the expected excess return can be derived if the expectedexcess return is calculated as weighted average expected excess return of the particular com-ponents of the firm value Vf,t, namely CFIf,t, CFOf,t, RSf,t and PV GOf,t. The expectedreturn of the first three components are:1dtEt[RCFIt]− rf = −cov(dCFItCFIt,dpitpit)= σxγx (4.24)1dtEt[RRSt]− rf = −cov(dRStRSt,dpitpit)= σxγx − σzγz (4.25)1dtEt[RCFOt]− rf = −cov(dCFOtCFOt,dpitpit)= vσxγx. (4.26)Accordingly, the expected excess return of the firm’s value at place V APf,t is:1dtEt[RV APt]− rf = ( 1dtEt[RCFIt]− rf) CFIf,tCFIf,t +RSf,t − CFOf,t+(1dtEt[RRSt]− rf) RSf,tCFIf,t +RSf,t − CFOf,t−(1dtEt[RCFOt]− rf) CFOf,tCFIf,t +RSf,t − CFOf,t= σxγxCFIf,t +RSf,t − vCFOf,tV APf,t− σzγzRSf,tV AP. (4.27)It is obvious from the formula (4.27) that expected return of the value of assets at placedepends on the exposure to both underlying risks xt and zt. This is different from the KPmodel, where the expected return of assets in place depends only on its exposure to theaggregate productivity shock xt. The first term in formula (4.27) is a levered claim on the110aggregate productivity shock xt. The leverage arises from the rigidity of the wage. Whilethe output fluctuates with the aggregate productivity shock xt, the fraction (1− v) of thelabor force has a constant wage and hence results in constant labor cost, which is reflected inoperating leverage. The operating leverage is determined by the parameter v. If v = 1, theCFIf,t+RSf,t−vCFOf,t = V APf,t and the firm is unlevered. In contrast v = 0 correspondsto the maximum possible leverage where CFIf,t+RSf,t−vCFOf,t = CFIf,t+RSf,t > V APf,t.The second term in the formula reflects the exposure of assets at place to the investmentshocks zt. This exposure arises from the capital re-sale when a project expires.The expected excess return on the growth option is:1dtEt[RPV GOf,t]− rf = −cov(dPV GOf,tPV GOf,t,dpitpit)= σxγx +αsαˆs − αsσzγz. (4.28)Accordingly, the expected excess return on the whole firm is a weighted average of the ex-pected return on the firm f ′s assets at place, V AP , and growth opportunities, PV GO:1dtEt [Rf,t]− r = σxγxVf,t + (1− v)CFOf,tVf,t+ σzγz( αsαˆs−αsPV GOf,t −RSf,tVf,t)= σxγxβxf,t + σzγzβf,tz. (4.29)It is also useful to derive an expression for a firm’s exposure to the investment shock zt:βzf,t =δ ln Vf,tδ ln zt=αsαˆs−αsPV GOf,t −RSf,tVf,t. (4.30)Finding a reliable empirical counterpart for firm’s true exposure to the investment shocks,βzf,t, is important for measuring the effects of investment shocks on firms. Below, I derivethe formula for the empirical proxy of the investment shock suggested by previous literature,the IMC portfolio, which allows for the IMC portfolio to be exposed also to the aggregateproductivity shocks xt. This allows to discuss how reliably a firm’s exposure to such IMCportfolio, βIMCf,t reflects the firm’s growth opportunities, PV GOf,t, and its true exposure tothe investment shock, βzf,t. At this point, it is obvious that even simply approximating firm’sexposure to the investment shock by firm’s growth opportunities might be problematic fortwo reasons. First, βzf,t depends not only on firm’s growth opportunities, PV GOf,t, but alsoon the potential capital re-sale, RSf,t. Second, the growth opportunities are multiplied byαsαˆs−αs , which captures a firm’s capital intensity and varies strongly in the cross section offirms.4.3.5 Valuation of the Investment-Goods FirmInvestment firm is modeled in a simplified form to get an appropriate counter-part forconsumption-goods firms. I assume that the investment-goods firm produces exactly thecapital demanded by the consumption-goods firms less the capital that is re-sold among theconsumption-goods firms by themselves. This assumption would correspond to a marketclearing condition on the market for capital in a general equilibrium setting.The total expected demand for capital from sector s is λ¯∫FsK∗f,tdf , where K∗f,t is theoptimal capital for newly arrived projects of firm f as described above. The expected capitalre-sold by the consumption-goods firms in sector s consists of two parts. First, the already111existing capital is∫fsKf,tdf = Ks,t and has a probability to be re-sold in future. Second,the capital which will be demanded in future for newly arrived project will be re-sold whenthese projects expire later. The profit of the investment-goods firm is the total amount ofthe capital sold by the investment firm (i.e., total demanded capital less the capital re-soldamong the consumption-goods firms) multiplied with the profit margin φ.The total expected amount of capital demanded by consumption-goods firms in eachsector s at each point in time is:λ¯∫FsK∗f,tdf = λ¯(αˆsαs) αˆsαs−αˆs (z−1t (1−M)) αˆsαs−αˆs D (αˆs, w)αˆsαˆs−αs∫FsA (f,t, 1)1αˆs−αs df.(4.31)The present value of the total capital demand is sector s is:PDVs,t = Et[∫ ∞txsz−1spispitλ¯s(αˆsαs) αˆsαs−αˆs (z−1s (1−M)) αˆsαs−αˆs D (αˆs, w)αˆsαˆs−αs(∫FsA (f, t, 1)1αˆs−αs df)ds]. (4.32)The present value of the re-sale of the existing capital is:NDV Ps,t = Et[∫ ∞tδe−δ(s−t)pispitxsz−1s(∫FsKf,tdf)ds]= xtz−1t MKs,t. (4.33)The present value of the re-sale of the capital of projects that are expected to arrive in futureis:NDV Fs,t =1Qλ¯Mxtz−1t(∫FsK∗f,tdf)(4.34), where Q = r− µx + µz − σ2z + σxγx − σzγz. The present value of the demand which will besupplied by the investment firm is:IDt =∑s∈{L,H}PDVs,t −NDV Ps,t −NDV Fs,t= xtzαLαˆL−αLt∑s∈{L,H}Γs (1−M)αˆsαs−α¯s − xtzt(1Qλ¯sM(∫FsK∗f,tdf)−MKs,t).(4.35)The value of the investment-goods firm is then:V It = φIDt (4.36)1124.3.6 Expected Excess Return of the Investment-Goods FirmThe expected excess return of the investment-goods firm can be calculated as the weightedaverage expected excess return of the individual demand components as:1dtEt[RIt]− r = ∑s∈{L,H}(1dtEt[RPDVst]− r)φPDVs,tV It−∑s∈{L,H}(1dtEt[RNDV Pst]− r)φNDV Ps,tV It−∑s∈{L,H}(1dtEt[RNDV Fst]− r)φNDV Fs,tV It. (4.37)Expected excess return of the positive demand component (PDVs,t) is:1dtEt[RPDVst]− r = σxγx + αsαˆs − αsσzγz. (4.38)Expected excess return of the negative demand components (NDV Ps,t and NDV Fs,t) is:1dtEt[RNDV Pst]− r = σxγx − σzγz, (4.39)and1dtEt[RNDV Fst]− r = σxγx + αsαˆs − αsσzγz. (4.40)The expected excess return of the investment-goods firm is:1dtEt[RIt]− r = σxγx + σzγz ∑s∈{L,H}αsαˆs−αs (PDVs,t −NDV Fs,t) +NDV Ps,tV It.(4.41)The second term in this formula shows the exposure of the investment-goods firm to theinvestment shock zt. In this setting, an investment-goods firm will a positive exposure to theinvestment shock if∑s∈{L,H}αsαˆs−αs (PDVs,t −NDV Fs,t) +NDV Ps,t > 0, which is satisfied inthis model. The first term captures the exposure fo the firm to the aggregate productivityshock xt. Since I do not explicitly model the capital or operating structure of the investment-goods firms, the first term is an unlevered claim on the underlying aggregate shock xt. Alevered exposure of the investment-goods firm to the aggregate shock xt would, however,affect the exposure of the IMC portfolio to this shock and I introduce this to the modeldirectly when I construct the IMC portfolio below.4.3.7 Expected Excess Return on the IMC PortfolioThe closed-form expressions for the expected returns of the consumption-goods and investment-goods firms allow to express also the expected returns of the IMC portfolio, i.e., a long-short portfolio with a long position in the investment-goods firm and a short position in113the consumption-goods firms. The expected excess return on each of the cross-section ofconsumption-goods firms is the value-weighted average of expected returns across the firmsin each consumption-goods sector s:1dtEt[Rs,Ct]− r =∫Fs(1dtEt [Rf,t]− r)Vf,t∫FsVv,tdvdf= σxγx∫Fs1 + (1− v)CFOf,tdf∫FsVf,tdf+ σzγz∫Fsαsαˆs−αsPV GOf,t −RSf,tdf∫FsVf,tdf.(4.42)To simplify the notation, I define V Ct =∑s∈{L,H}∫FsVf,tdf , which denotes the total value ofall consumption-goods firms in both, the capital-intensive and labor-intensive sector. Theexpected excess return of the whole consumption-goods sector can be written on as:1dtEt[RCt]− r = σxγx ∑s∈{L,H}∫Fs(1 + (1− v)CFOf,t) dfV Ct+σzγz∑s∈{L,H}∫FsαLαˆL−αLPV GOf,t −RSf,tdfV Ct. (4.43)The expected excess return of the IMC portfolio can be calculated by subtracting theexpected return of the consumption-goods sector from the expected return of the investment-goods sector. If both sectors have exactly the same exposure to the aggregate shock xt,these exposures will cancel each other and the resulting IMC portfolio will span purely thedimension of the investment shock zt. In reality both, the investment-goods and consumption-goods firms differ in their exposure to aggregate risk factors. Table 4.1 shows that theempirical IMC portfolio is positively correlated with the market and the SMB factors, bothof which carry a positive return premium. In order to allow the theoretical IMC portfolio tocorrelate not only with the negatively priced investment shocks zt, but also with positivelypriced shocks, I introduce the parameter LDIFF . In the narrow sense, this parameterexpresses the difference in the operating leverage between consumption-goods and investment-goods firms. More broadly, it can be thought of as any difference between investment-goodsand consumption-goods firms which results in a different exposure to the positively pricedaggregate risk factor xt.59 Using this modeling shortcut, the expected return of the IMC59It is possible to model the operating (or other) leverage of the investment-goods firm in the specificationof the firm’s cash flow in formula 4.35. Since the return of the IMC portfolio and not the return of theinvestment-goods firm itself is of immediate interest, the above specification provides an equivalent modelingsolution.114portfolio can be written as:1dtEt[RIt −RCt]= −σxγx∑s∈{L,H}∫Fs(1− v)CFOf,tdfV CtLDIFF+σzγz∑s∈{L,H}αsαˆs−αs (PDVs,t −NDV Fs,t) +NDV Ps,tV It−σzγz∑s∈{L,H}∫Fsαsαˆs−αsPV GOf,t −RSf,tdfV Ct. (4.44)The realized return of the IMC portfolio is the expected return and the corresponding stochas-tic part. For the sake of simplicity, I define CFOt =∑s∈{L,H}∫Fs(1−v)CFOf,tdfV Ct,INVt =∑s∈{L,H}αsαˆs−αs (PDVs,t−NDV Fs,t)+NDV Ps,tV ItandPV GORSt =∑s∈{L,H}∫Fsαsαˆs−αs PV GOf,t−RSf,tdfV Ct. The realized return of the IMC portfolio canbe written as:RIt −RCt = −CFOt × LDIFF (σxγxdt+ σx dBxt)+ (INVt − PV GORSt) (σzγzdt+ σz dBzt) . (4.45)The exposure of the IMC portfolio to the aggregate shock xt depends on the similarity ofthe investment and consumption firms, CFOt × LDIFF . The exposure to the investmentshocks zt, depends on the relative value of growth opportunities to total firm value of theconsumption firm as well as on how much of the demanded capital will be supplied by theinvestment firm vs. by re-sale, INVt − PV GORSt.4.4 Empirical Implications4.4.1 βIMCf,t as a Measure of Firm’s Exposure to the Investment ShocksIn this section I analyze the IMC portfolio as an empirical proxy for investment shocks. Thepurpose is to understand how well the IMC portfolio can capture firm’s exposure to theinvestment shock. In the ideal case when the IMC portfolio is exposed only to investmentshock zt, firm’s exposure to the IMC portfolio will map one-to-one to firm’s exposure toinvestment shock zt. However, in reality the IMC portfolio correlates strongly with themarket and SMB factors, whose theoretical counterpart is the aggregate shock xt. Theclosed-form expression of firm’s exposure to the IMC portfolio, βIMCf,t , allows to analyze thepitfalls of estimating firm’s exposure to investment shocks from its exposure to the IMCportfolio and, at least partially, reconcile the empirical observations.115Firm f ′s exposure to the IMC portfolio is:βIMCf,t =cov(Rf,t, RIMCt)var(RIMCt)= −σ2x(Vf,t+(1−v)CFOf,tVf,t)CFOt × LDIFFσ2x (CFOt × LDIFF )2 + σ2z (INVt − PV GORSt)2+σ2z( αsαˆs−αs PV GOf,t−RSf,tVf,t)(INVt − PV GORSt)σ2x (CFOt × LDIFF )2 + σ2z (INVt − PV GORSt)2. (4.46)Firm f ′s exposure to the IMC portfolio depends on the particular components of firm f ′stotal value as well as on the IMC portfolio’s exposure to the shocks xt and zt. Motivated bythe empirical observation of a positive correlation between the IMC portfolio and the marketreturns (and the SMB factor), I assume LDIFF < 0, which delivers the correspondingcorrelation between the aggregate shock xt and the IMC portfolio. Further, I assume INVt−PV GORSt > 0 which means that investment-goods firms are more exposed to the IST shockthan consumption-goods firms, which is the main motivation for using this portfolio as anempirical proxy for the IST shock as described in Papanikolaou (2011).Table 4.6 summarizes how the particular value components affect firm’s exposure to theshocks xt and zt as well as its exposure to the IMC portfolio. Firm’s operating leveragearising from (1− v) CFOf,tVf,t increases firm’s exposure to aggregate shock xt so that also βIMCf,tincreases. The fraction of firms value arising from growth opportunitiesPV GOf,tVf,tincreasesfirms exposure to the investment shock and to the IMC portfolio by a factor αsαˆs−αs . Thisfactor depends on a firm’s capital intensity and hence will be different for capital-intensiveand labor-intensive firms. The value of potential re-sale RSf,t is negatively correlated withthe shock zt and hence it decreases firm’s exposure to both, the investment shock and theIMC portfolio.This analysis shows that firm’s βIMCf,t is not a precise measure of firm’s exposure to theIST shocks and the quality of this measure depends both, on the components of firm’s totalvalue and on the exposure of the IMC portfolio to the aggregate shock xt. For instance,a capital-intensive firm with large growth opportunities relative to existing assets will havea high exposure to the IMC portfolio due to its exposure to the investment shock, zt. Incontrast, highly levered labor-intensive firm will have high exposure to the IMC portfolio dueto its exposure to the aggregate shock xt. Accordingly, similar magnitudes of the βIMCf,t mayhave a different meaning in terms of describing the exposure to the underlying shocks forcapital-intensive and labor-intensive firms.I test this empirical prediction in table 4.7. I use market-to-book ratio as an approx-imation of a firm’s growth opportunities and the ratio of a firm’s capital (property, plantand equipment) to market capitalization as an approximation for of firm’s potential capitalresale. Operating leverage is measured as a ratio of cost of goods sold and selling, general andadministrative expenses to firm’s market capitalization. The first column shows that βIMCf,tis positively related to firm’s growth opportunities as well as to firm’s operating leverage aspredicted by the model. Contrary to the prediction of the model, firm’s exposure to the IMCportfolio is also positively related to firm’s potential capital resale. Potential reason for thisresult is that the ratio of firm’s capital to market capitalization is a poor measure of firm’s116capital resale.Second and third column show the estimates for capital-intensive and labor intensive firmsseparately. Consistent with the prediction of formula (4.46), the link between firm’s growthopportunities and its exposure to the IMC portfolio is much stronger among capital-intensivefirms, but almost absent among labor-intensive firms. These two columns also show that thelink between firm’s operating leverage and the exposure to the IMC portfolio is significantfor labor intensive firms, but of lower magnitude and insignificant for capital-intensive firms.I repeat the estimation with an alternative measure of operating leverage that accounts onlyfor selling, general and administrative expenses. The motivation for this alternation is thatthese costs are likely to be more rigid than the costs of goods sold and can better approximatefirm’s operating leverage. The last two columns show that the results based on this measureare similar.The model also allows to derive cross-sectional asset pricing implications of shock, growthopportunities and capital-intensity. The formula (4.29) shows that in this model, firm’sexpected return is a linear function of firms exposure to the aggregate shock xt and theexposure to the investment shock zt. The expression for a firm’s βzf,t in formula (4.30) showsthat firm’s exposure to the investment shock depends on firm’s growth opportunities andfirm’s capital intensity (and firm’s potential capital resale that does not seem to be empiricallyrelevant). According to this model, we should expect a linear relationship between firm’sexpected returns and βf,tz. Moreover, under the assumption that the growth opportunitiesspan approximately the same range among both, capital-intensive and labor-intensive firms(see the 10th and 90th percentiles of the book-to-market ratio in table 4.4), portfolios ofcapital-intensive firms sorted by βzf,t should exhibit larger differences in expected returns andin the exposure to the investment shock than the portfolios consisting of labor-intensive firms.While the results in 4.5 seem to support the prediction of stronger return differencesamong capital-intensive firms, they also show almost identical patterns in the exposure tothe IMC portfolio for both types of firms. Comparing the expressions for βIMCf,t and βzf,t informulas (4.46) and (4.30) suggests that this result can arise due to a potential bias in βIMCf,tas a measure of βzf,t.In table 4.8, I repeat the analysis from table 4.5 but I sort the stocks based on βIMC+controlsf,tthat is estimated from multivariate regression of this form:Rf,t −Rriskfree = αf,t + βIMC+controlsf,t ×RIMCt+βmarketf,t ×Rmarkett + βSMBf,t ×RSMBt + f,t. (4.47)I include the returns on market and the SMB factors, so that βIMC+controls is relativelyunbiased by the exposure to these two factors and hence may come closer to the theoreticalideal of βzf,t.Panels A, B and C confirm the results for average excess and abnormal returns fromtable 4.5. Capital-intensive firms exhibit a slightly decreasing pattern of average excessreturns and a strong decreasing pattern of α across the portfolios, while labor-intensive firmsdo not. Panel D, however, shows that also the exposure to the IMC portfolio differ. Whilethe exposure of the capital-intensive long-short portfolio, HL, has an exposure to the IMCportfolio of 0.97, the labor-intensive counterpart has exposure of only 0.64. Panel E showsthe exposures to the IMC portfolio and the loadings on the market return and the SMB117factor estimated from a multivariate regression that controls for these two factors. Theseresults also show large differences in the portfolios’ exposures to the IMC portfolio betweencapital-intensive and labor-intensive firms. Although it is not clear whether the difference in αbetween the capital-intensive and labor-intensive long-short portfolios can be fully explainedby their different exposure to the IMC portfolio, the results support the intuition of thetheoretical model.It is worthwhile mentioning that also the relatively small exposure of the labor-intensivelong-short portfolio to the IMC portfolio is associated with much stronger (negative) exposureto the HML factor. This also seems to be consistent with the theoretical prediction ofthe model that large growth opportunities of labor-intensive firms will be associated withonly moderate exposure to the investment shock. At the same time, relatively low growthopportunities of capital-intensive firms will be linked to a high exposure to the investmentshock.4.5 Model Calibration and SimulationI calibrate the model with most of the parameters having same or similar values as in Koganand Papanikolaou (2014) to allow a comparison with this benchmark model. These parame-ters were originally selected to match moments of aggregate dividend and investment growth,asset returns, accounting ratios, the IMC portfolio properties and capital intensity of firms.60The additional parameters in my extended model, αL, αH , αˆL, αˆH , w, v, and LDIFF , governthe capital-labor ratio of capital- and labor-intensive firms, wage rigidity, operating leverage,the exposure of firms’ growth opportunities to investment shock and correlation of the IMCportfolio returns with aggregate market excess returns, aggregate shock xt and investmentshock zt. I select these parameters to approximately match the relative capital-labor ratio,i.e., capital-labor ratio of capital-intensive firms relative to the same ratio of labor-intensivefirms, the correlation between the IMC portfolio returns and the excess market return, andthe different patterns of abnormal returns between capital-intensive and labor-intensive firms.The parameters are provided in table 4.9.The goal of this exercise is to provide a quantitative result in support of the asset pric-ing implication of the model and show that capital-intensity matters for βIMCf,t sorted stocksas found in the data. At the same time, it is important to note that not all moments canbe matched precisely as the model abstracts from other potential factors such as compe-tition, or capital utilization, which can affect the investment behavior of firms and theirriskiness. I simulate a cross-section of consumption-goods firms and the necessary counter-part of investment-good firm. The cross-section of consumption-goods firms consists of twosub-samples, namely capital-intensive and labor-intensive firms. Each sub-sample consists of300 firms. I simulate the model 100 times for 100 years and in each simulation I use only thesecond half (year 51-100) for estimating the moments. I report the median of the momentsacross individual simulations. The cross-sectional distribution of firm’s project arrival ratesλf = E [λf,t] is given as:λf = µλδ − σλδlog (Xf ) (4.48)60I use a lower parameter for the growth of the investment shock, µz = 0.001 instead of 0.005, as it helps thesimulated relative capital-labor ratio, and some return properties come closer to their empirical counterparts.118, where Xf ∼ N (0, 1).Table 4.10 shows the moments from the simulated data from my extended model, theKogan and Papanikolaou (2014) model (indicated by KP model) and the empirical coun-terparts. Most of the moments from the extended model are relatively close their empiricalor KP counterparts. The biggest differences are for investment growth and investment rate.This is caused by the fact that the extended model is based on a Cobb-Douglas productionfunction with capital and labor. Investment in this model is driven, to large extent, by in-vestment shocks zt that affect directly capital but not labor. I calibrate the model so thatthe correlation between the IMC portfolio and aggregate market return is somewhat higherthan observed in the data. The return of the empirical IMC portfolio correlates not onlywith return of the market portfolio, but also with the SMB factor. Since the model doesnot include the SMB factor, a higher correlation with market return can, to some extent,compensate for the missing factor and help match the cross-sectional patterns of portfolioreturns. The higher correlation between the IMC portfolio and market return can be achievedby larger differences in operating leverage between consumption-goods and investment-goodsfirms, LDIFF . Higher magnitude of this parameter then results also in higher volatility ofthe IMC portfolio return than observed in the data.Table 4.11 reports the properties of the portfolio returns of firms sorted by their exposureto the IMC portfolio, βIMC , for capital-intensive and labor-intensive firms simulated in myextended model. It is apparent that the average excess returns of all labor-intensive portfoliosare higher than their counterparts for capital-intensive returns. Although we can see in thetable (4.5) that the average excess returns of labor-intensive portfolios are somewhat higherthan for capital-intensive portfolios, the differences are not as high as those in simulated data.This discrepancy is due to labor-intensive firms having in the model, on average, a higherexposure to positively priced aggregate risk and only a limited exposure to negatively pricedinvestment shock, which is apparent when individual capital- and labor-intensive portfoliosare compared.Apart from this, the simulated returns show two important patterns. First, sortingcapital-intensive firms by the exposure to the IMC portfolio generates much higher α thanthe same sorting among labor-intensive firms. This is consistent with the observed patternsin the data. Second, the differences in firms’ true exposure to the investment shock, βz, aremuch higher among capital-intensive firms. This seems to be consistent with the patternin βIMC in panel E of table 4.8. One has to, however, acknowledge that observing βz ischallenging in the data and hence the comparison with βIMC even if it controls for otherfactors has to be taken cautiously.4.6 ConclusionIn this paper, I exploit the intuition that in a model with a Cobb-Douglas production func-tion and vintage capital, positive investment shocks, i.e., shocks to the formation of newcapital, are expected to benefit prominently capital-intensive firms with large growth op-portunities. I extend the Kogan and Papanikolaou (2014) model by introducing labor withpartially rigid wage as a production factor, which leads to firms’ operating leverage. Thisextended model shows that firm’s capital intensity can affect both, the exposure of firm’sgrowth opportunities to the investment shocks and the measurement of firm’s exposure to119these shocks by its exposure to an IMC portfolio which correlates also with other factors.Differentiating between capital-intensive and labor intensive firms provides some empiricalsupport for the implications of the model. Although the IMC portfolio does not seem toexplain the strongly negative exposure of firms with displaceable labor to technology shocksas mentioned in the first essay, the results based on the exposure to the IMC portfolio suggestthat capital-intensive growth opportunities are more affected by investment shocks than theirlabor-intensive counterparts.1204.7 Derivation of FormulasPresent value of growth opportunities:PV GOf,t = Et[∫ ∞tpispitxs(z−1s (1−M)) αsαs−αˆs A (f,t, 1)−1αs−αˆs×D (αˆs, w)−αˆsαs−αˆs[(αˆsαs) αsαs−αˆs −(αˆsαs) αˆsαs−αˆs]λf,sds]= xtzαsαˆs−αst C (αs, αˆs, w) Et[∫ ∞te−ρs(s−t)λf,sA (f,s, 1)1αˆs−αs ds]= xtzαsαˆs−αst G (f,t, λf,t, αs, αˆs, w) (4.49), where ρs = r + γxσx − µx − αsαˆs−αs(µz − γzσz − 12σ2z)− 12 ( αsαˆs−αs)2 σ2zand C (αs, αˆs, w) = D (αˆs, w)αˆsαˆs−αs (1−M) αsαs−αˆs[(αˆsαs) αsαs−αˆs −(αˆsαs) αˆsαs−αˆs].G (f,t, λf,t, αs, αˆs, w) = C Et[∫ ∞teρ(s−t)λf,sA (f,t)αsαˆs−αs ds]=λf(G1 (f,t) +µLµL+µH(λH − λL)G2 (f,t)), λ˜f,t = λHλf(G1 (f,t)− µHµL+µH (λH − λL)G2 (f,t)), λ˜f,t = λL(4.50), where G1 and G2 are defined as:G1 (f,t) = XEt ∞∫te−ρ(s−t)A (f,s)αsαˆs−α ds (4.51)G2 (f,t) = XEt ∞∫te−(ρ+µL+µH)(s−t)A (f,s)αsαˆs−α ds (4.52)and satisfy the ordinary differential equations:C ×A (f,t)αsαˆs−αs − ρG1 (f,t)− θ (− 1) ddG1 (f,t) +12σ2 d2d2G1 (f,t) = 0(4.53)C ×A (f,t)αsαˆs−αs − (ρ+ µH + µL)G2 (f,t)− θ (− 1) ddG2 (f,t) +12σ2 d2d2G2 (f,t) = 0.(4.54)The present value of the re-sale of the capital of projects that are expected to arrive in future121is:NDV Fs,t = Et[∫ ∞tpiτpit(∫ ∞τδe−δ(u−τ)piupiτxτz−1τ λ¯(∫FsK∗f,tdf)du)dτ]= Et[∫ ∞tpiτpitδλ¯δ + r − µx + µz − σ2z + σxγx − σzγzxτz−1τ(∫FsK∗f,tdf)dτ]=∫ ∞te−Q(τ−t)λ¯Mxτz−1τ(∫FsK∗f,tdf)dτ=1Qλ¯Mxtz−1t(∫FsK∗f,tdf)(4.55)Expected excess return on the positive demand component (PDVs,t):1dtEt[RPDVst]− r = −cov(dPDVs,tPDVs,t,dpitpit)= −cov(dxtxt+ααˆs − αdztzt,−γxdBxt − γzdBzt)= −cov(σxdBxt +ααˆs − ασzdBzt ,−γxdBxt − γzdBzt)= σxγx +αsαˆs − αsσzγz (4.56)Expected excess return on the positive demand components (NDV Ps,t and NDV Fs,t):1dtEt[RNDV Pst]− r = −cov(dNDV Ps,tNDV Ps,t,dpitpit)= σxγx − σzγz (4.57)1dtEt[RNDV Fst]− r = −cov(dNDV Fs,tNDV Fs,t,dpitpit)= −cov(dxtxt+αsαˆs − αsdztzt,−γxdBxt − γzdBzt)= −cov(σxdBxt + σzαsαˆs − αsdBzt ,−γxdBxt − γzdBzt)= σxγx +αsαˆs − αsσzγz (4.58)The expected excess return on the investment-goods firm is:1dtEt[RIt]− r = σxγx(PDVL,t + PDVH,t −NDVL,t −NDVH,tV It)+σzγz(αLαˆL − αLPDVL,t −NDV FL,tV It+αHαˆH − αHPDVH,t −NDV FH,fV It−NDV PL,t +NDV PH,tV It)= σxγx +σzγzαLαˆL−αL (PDVL,t −NDV FL,t) +αHαˆH−αH (PDVH,t −NDV FH,t)V It+σzγzNDV PL,t +NDV PH,tV It(4.59)122, where I define NPVs,t = NDV Ps,t +NDV Fs,t for sake of simplicity.The exposure of the IMC portfolio to each of the aggregate shocks x and IST shock z is:βIMC,zt =cov(RIMCt ,dztzt)var(dztzt) = cov (σzdBzt (INVt − PV GORSt) , σzdBzt)var (σzdBzt)= INVt − PV GORSt (4.60)βIMC,xt =cov(RIMCt ,dxtxt)var(dxtxt) = cov (−σxdBxtCFOt × LDIFF, σxdBxt)var (σxdBxt)= −CFOt × LDIFF (4.61)Firm f ′s exposure to the IMC portfolio is:βIMCf,t =cov(Rf,t, RIMCt)var(RIMCt)=cov(σxdBxtCFIf,t+RSf,t+PV GOf,tVf,t,−σxdBxtCFOt × LDIFF)var(RIMCt)+cov(σzdBxt(αsαˆs−αsPV GOf,tVf,t− RSf,tVf,t), σzdBz,t (INVt − PV GORSt))var(RIMCt)= −σ2x(CFIf,t+RSf,t+PV GOf,tVf,t)CFOt × LDIFFσ2x (CFOt × LDIFF )2 + σ2z (INVt − PV GORSt)2+σ2z(αsαˆs−αsPV GOf,tVf,t− RSf,tVf,t)(INVt − PV GORSt)σ2x (CFOt × LDIFF )2 + σ2z (INVt − PV GORSt)2(4.62)1234.8 TablesTable 4.1: Factors: Correlation MatrixThe table shows the pairwise correlation coefficients between monthly returns of the IMCportfolio, market portfolio and the Fama and French (1993) SMB and HML factors. Thetime period is from 1970 to 2015.IMC Market SMB HMLIMC 1.00 0.45 -0.25 0.45Market 0.45 1.00 -0.32 0.29HML -0.25 -0.32 1.00 -0.24SMB 0.45 0.29 -0.24 1.00Table 4.2: Portfolio Return Properties of All Firms in the Consumption-Goods SectorSorted by βIMCThe table shows the average excess returns (panel A) and the estimates of the CAPM (panelB) and Fama and French (1993) model (panel C) for portfolios sorted by βIMC . Panel Dshows the median sorting βIMC of the five portfolios as well as the portfolios’ post-rankingexposures to the IMC portfolio. Stocks are sorted into five portfolios at the end of each Junebased on βIMC estimated from a univariate regression of monthly stock returns over the last60 months. Portfolios are value-weighted. HL indicates a long-short portfolio with a longposition in portfolio 5 and a short position in portfolio 1. The time period is from 1970 to2015. *** Significant at 1 percent level. ** Significant at 5 percent level. * Significant at 10percent level.1 2 3 4 5 HLPanel A. Aveage excess returnMean 8.71*** 8.11*** 7.67*** 6.93*** 7.66** -1.06(1.87) (2.2) (2.41) (2.26) (3.48) (2.68)Panel B. CAPMα 3.42*** 1.56** 0.28 -1.79 -3.06 -6.48***(0.98) (0.71) (0.99) (1.23) (1.92) (2.43)Market 0.73*** 0.91*** 1.02*** 1.21*** 1.48*** 0.75***(0.03) (0.02) (0.02) (0.02) (0.05) (0.07)R2 73.18 87.8 83.4 86.18 73.28 28.67Panel C. Fama and French (1993)α 3.94*** 1.54** 0.8 -1.92 -1.46 -5.39***(0.88) (0.77) (1.04) (1.17) (1.74) (2.06)Market 0.79*** 0.93*** 1.0*** 1.16*** 1.3*** 0.51***(0.02) (0.02) (0.03) (0.03) (0.04) (0.06)SMB -0.33*** -0.12*** 0.02 0.22*** 0.54*** 0.87***(0.03) (0.03) (0.08) (0.04) (0.07) (0.08)HML -0.01 0.04 -0.1* -0.04 -0.46*** -0.45***(0.04) (0.03) (0.06) (0.05) (0.09) (0.11)R2 79.65 88.58 83.78 87.53 81.62 52.89Panel D. Sorting and post-sorting βIMCsortingβIMC-0.16 0.58 1.02 1.56 2.8IMC 0.1 0.5*** 0.7*** 1.04*** 1.5*** 1.4***(0.06) (0.08) (0.1) (0.09) (0.1) (0.08)124Table 4.3: Summary of Firms in Consumption and Investment Goods SectorsThe table shows the summary statistics of firms in the consumption goods and investment goods sector together (panel A) and forfirms in each sector separately (panels B and C). Market capitalization is the market price for all outstanding shares (millions).Operating leverage is based on market equity. Capital indicates firms’ gross book value (millions) of property, plant and equipment.Number of employees is stated in thousands. The statistics are time-series averages of the annual percentiles. The time period isfrom 1965 to 2015.Panel A:Both sectorsPanel B:Consumpion goods sectorPanel C:Investment goods sector10thpercentilemedian90thpercentile10thpercentilemedian90thpercentile10thpercentilemedian90thpercentileNumber of firms 4018.96 2603.76 2603.76Market capitalization 13.09 187.75 3206.26 13.86 196.91 3521.42 12.24 182.20 2976.23Book-to-market 0.22 0.67 1.66 0.21 0.67 1.69 0.23 0.68 1.61Cashflow to assets -4.53 0.55 6.18 -3.55 0.53 5.65 -6.32 0.57 6.89Operating leverage 0.28 1.45 6.09 0.32 1.64 7.27 0.23 1.23 4.36Capital 3.87 68.69 1728.56 3.73 68.40 1644.63 4.60 78.55 2019.74Number of employees 0.17 1.64 19.96 0.20 1.92 22.99 0.15 1.32 15.99Capital-labor ratio 12.76 44.99 352.13 10.92 38.97 232.72 19.62 58.51 794.41125Table 4.4: Summary Statistics of Capital-Intensive and and Labor-Intensive Firms in each SectorThe table shows the summary statistics of the capital-intensive (panel A) and labor-intensive firms (panel B) in the consumption-goods, investment-goods, and both sectors. Capital(labor)-intensive firms are defined as firms with above(below)-median capital-labor ratio in the given sector. Market capitalization is the market price for all outstanding shares (millions). Operating leverageis based on market equity. Capital indicates firms’ gross book value (millions) of property, plant and equipment. Number ofemployees is stated in thousands. The statistics are time-series averages of the annual percentiles. The time period is from 1965to 2015.Consumption goodssectorInvestment goodssectorBoth sectors10thpercentilemedian90thpercentile10thpercentilemedian90thpercentile10thpercentilemedian90thpercentilePanel A. Capital-intensive firmsMarket cap 23.89 433.75 9326.40 19.93 334.63 5441.10 21.90 395.97 7916.19Book-to-market 0.23 0.69 1.62 0.25 0.71 1.74 0.23 0.69 1.67Operating leverage 0.25 1.26 5.64 0.15 1.03 3.93 0.20 1.15 4.91Capital 10.49 277.60 7762.07 15.05 286.27 5069.78 11.63 277.94 6753.01Number of employees 0.12 1.57 25.49 0.07 1.09 18.94 0.10 1.39 23.16Capital-labor ratio 54.83 114.40 1150.59 75.47 175.18 3035.60 58.67 129.59 1470.43Panel B. Labor-intensive firmsMarket cap 12.72 158.95 1999.32 9.59 131.43 1797.31 11.53 150.07 1935.94Book-to-market 0.21 0.70 1.76 0.21 0.65 1.54 0.21 0.68 1.69Operating leverage 0.44 2.06 8.86 0.35 1.42 4.94 0.39 1.77 7.31Capital 1.33 26.93 463.88 2.34 27.50 448.75 1.60 26.92 454.95Number of employees 0.13 1.40 18.47 0.12 0.84 10.15 0.12 1.16 15.75Capital-labor ratio 6.82 23.29 41.49 15.58 36.28 58.50 8.70 27.00 47.30126Table 4.5: Portfolio Return Properties of Capital- and Labor-intensive Firms in the Consumption-Goods Sector Sorted byβIMCThe table shows the average excess returns (panel A) and the estimates of the CAPM (panel B) and Fama and French (1993)model (panel C) of portfolios sorted by βIMC for capital-intensive and labor-intensive. Panel D shows the median sorting βIMC ofthe five portfolios as well as the portfolios’ post-ranking exposures to the IMC portfolio. Capital(labor)-intensive firms are definedas firms with above(below)-median capital-labor ratio. Stocks are sorted into five portfolios at the end of each June based onβIMC estimated from a univariate regression of monthly stock returns over the last 60 months. Portfolios are value-weighted. HLindicates a long-short portfolio with a long position in portfolio 5 and a short position in portfolio 1. The time period is from 1970to 2015. *** Significant at 1 percent level. ** Significant at 5 percent level. * Significant at 10 percent level.Capital-intensive firms Labor-intensive firms1 2 3 4 5 HL 1 2 3 4 5 HLPanel A. Average excess returnMean 9.06*** 7.89*** 7.7*** 5.76*** 6.63** -2.44 8.18*** 8.05*** 9.86*** 11.15*** 10.8*** 2.63(1.91) (2.1) (2.55) (2.2) (3.2) (2.69) (2.24) (2.43) (2.49) (2.93) (4.12) (3.22)Panel B. CAPMα 3.93*** 1.63* 0.38 -2.52* -3.63* -7.56*** 2.36 0.76 2.26* 1.88 -0.5 -2.87(1.02) (0.92) (1.06) (1.38) (1.93) (2.61) (1.69) (1.14) (1.35) (1.71) (2.51) (2.81)Market 0.71*** 0.87*** 1.01*** 1.14*** 1.42*** 0.71*** 0.8*** 1.01*** 1.05*** 1.28*** 1.56*** 0.76***(0.03) (0.02) (0.03) (0.02) (0.05) (0.06) (0.04) (0.02) (0.03) (0.04) (0.07) (0.1)R2 68.9 81.94 78.99 82.76 73.96 27.55 59.54 79.67 78.05 72.89 63.71 23.27Panel C. Fama and French (1993)α 4.49*** 1.62* 1.18 -2.78** -2.78 -7.28*** 2.68 0.3 1.5 2.29* 1.8 -0.88(0.91) (0.9) (1.15) (1.39) (1.71) (2.21) (1.65) (1.14) (1.3) (1.31) (2.53) (2.67)Market 0.77*** 0.91*** 1.0*** 1.13*** 1.26*** 0.49*** 0.83*** 0.98*** 1.0*** 1.12*** 1.34*** 0.51***(0.02) (0.02) (0.04) (0.02) (0.05) (0.06) (0.04) (0.02) (0.02) (0.03) (0.06) (0.08)SMB -0.37*** -0.2*** -0.07 0.09* 0.56*** 0.93*** -0.18*** 0.22*** 0.35*** 0.65*** 0.57*** 0.75***(0.03) (0.03) (0.07) (0.05) (0.07) (0.08) (0.06) (0.04) (0.07) (0.06) (0.11) (0.1)HML -0.01 0.06 -0.13** 0.03 -0.32*** -0.31*** -0.01 0.02 0.05 -0.26*** -0.6*** -0.59***(0.04) (0.04) (0.06) (0.06) (0.08) (0.1) (0.06) (0.05) (0.05) (0.07) (0.13) (0.15)R2 76.89 84.17 79.56 82.98 81.67 53.17 60.82 81.32 81.76 83.35 72.67 41.85Panel D. Sorting and post-sorting βIMCsortingβIMC-0.15 0.55 0.99 1.51 2.73 -0.15 0.61 1.06 1.61 2.85IMC 0.09 0.45*** 0.69*** 0.93*** 1.49*** 1.4*** 0.13 0.64*** 0.69*** 1.26*** 1.49*** 1.36***(0.06) (0.08) (0.1) (0.09) (0.1) (0.08) (0.08) (0.09) (0.09) (0.09) (0.14) (0.11)127Table 4.6: Exposure of Firm-Value Components to Both Types of Shocks and to theIMC PortfolioThe table shows how the firm’s value components affect firm’s exposure to the shocks xt,zt, and to the IMC portfolio. The value components are in rows and the shocks (the IMCportfolio) are in columns.xt zt IMCCFOf,tVf,t(1-v) none (1-v)PV GOf,tVf,tnone(αsαˆs−αs) (αsαˆs−αs)RSf,tVf,tnone -1 -1Table 4.7: Panel Regressions of Firms βIMC on Firms’ Market-to-Book Ratio, Capital,and Operating LeverageThe table shows the results of panel regression βIMCf,t = constant + ξ1market-to-book +ξ2capitalmarketcap + ξ3Operating leverage + f,t for all firms in the consumption goods sector andfor capital- and labor-intensive firms separately. The last two columns use the ratio of theselling, general and administrative expenses to market capitalization as an alternative mea-sure of operating leverage. The panel regressions include industry and year fixed effectswith industries defined at 4-digit SIC code. Standard errors are clustered at 3-digit SICcode and account for heteroscedasticity. The time period is from 1970 to 2015. *** Signif-icant at 1 percent level. ** Significant at 5 percent level. * Significant at 10 percent level.All firmsCapital-intensive firmsLabor-intensive firmsCapital-intensive firmsLabor-intensive firmsMarket-to-book 0.005∗∗ 0.009∗∗∗ 0.003 0.009∗∗∗ 0.004(0.002) (0.003) (0.003) (0.003) (0.003)capitalmarket cap 0.021∗∗∗ 0.032∗∗∗ 0.001 0.034∗∗∗ -0.002(0.007) (0.007) (0.011) (0.007) (0.009)Operatingleverage0.009∗∗∗ 0.006 0.009∗∗(0.003) (0.005) (0.003)xsgamarket cap 0.026 0.051∗∗∗(0.020) (0.012)Industry FE Yes Yes Yes Yes YesYear FE Yes Yes Yes Yes YesN 120374 57693 62729 57693 62729R2 0.311 0.347 0.315 0.347 0.315128Table 4.8: Portfolio Return Properties of Capital- and Labor-intensive Firms in the Consumption-Goods Sector Sorted byβIMC+controlsThe table shows the average excess returns (panel A) and the estimates of the CAPM (panel B) and Fama and French (1993)model (panel C) of portfolios sorted by βIMC for capital-intensive and labor-intensive. Panel D shows the median sorting βIMCof the five portfolios as well as the portfolios’ post-ranking exposures to the IMC portfolio estimated from a univariate regression.Panel E shows the loadings of the Market, SMB, and IMC factors estimated from a model with a constant and these threefactors. Capital(labor)-intensive firms are defined as firms with above(below)-median capital-labor ratio. Stocks are sorted intofive portfolios at the end of each June based on βIMC+controls estimated from a regression with the IMC portfolio, and the marketand SMB factors using monthly stock returns over the last 60 months. Portfolios are value-weighted. HL indicates a long-shortportfolio with a long position in portfolio 5 and a short position in portfolio 1. The time period is from 1970 to 2015. *** Significantat 1 percent level. ** Significant at 5 percent level. * Significant at 10 percent level.Capital-intensive firms Labor-intensive firms1 2 3 4 5 HL 1 2 3 4 5 HLPanel A. Average excess returnMean 8.5*** 8.68*** 8.01*** 7.12*** 6.49** -2.0 7.89*** 8.46*** 9.28*** 10.57*** 7.94** 0.05(2.09) (1.84) (2.34) (2.29) (2.74) (2.46) (2.72) (2.24) (2.27) (2.9) (3.3) (2.45)Panel B. CAPMα 2.78** 2.95*** 1.77* 0.03 -1.8 -4.58* 0.4 1.71 2.21* 2.46 -2.03 -2.43(1.2) (0.84) (1.08) (1.09) (1.94) (2.44) (1.83) (1.34) (1.21) (1.5) (2.18) (2.42)Market 0.79*** 0.79*** 0.86*** 0.98*** 1.15*** 0.36*** 1.03*** 0.93*** 0.98*** 1.12*** 1.38*** 0.34***(0.03) (0.03) (0.02) (0.02) (0.03) (0.05) (0.05) (0.03) (0.03) (0.03) (0.05) (0.08)R2 66.83 78.56 80.87 83.96 74.55 11.3 63.67 71.64 76.3 74.28 67.32 8.77Panel C. Fama and French (1993)α 3.53*** 3.3*** 1.81* 0.29 -1.36 -4.89** 0.71 1.78 1.56 2.62* -0.74 -1.45(1.16) (0.76) (1.01) (1.1) (1.95) (2.35) (1.77) (1.37) (1.25) (1.39) (1.99) (2.48)Market 0.81*** 0.84*** 0.91*** 0.99*** 1.1*** 0.29*** 0.93*** 0.93*** 0.97*** 1.02*** 1.2*** 0.27***(0.03) (0.02) (0.02) (0.02) (0.04) (0.05) (0.06) (0.03) (0.04) (0.03) (0.05) (0.08)SMB -0.23*** -0.28*** -0.2*** -0.07 0.12 0.35*** 0.39*** 0.02 0.15* 0.41*** 0.56*** 0.17(0.04) (0.03) (0.03) (0.05) (0.08) (0.1) (0.08) (0.04) (0.08) (0.06) (0.09) (0.15)HML -0.08 0.01 0.05 -0.03 -0.12* -0.04 -0.17** -0.02 0.08 -0.14** -0.4*** -0.23(0.06) (0.03) (0.04) (0.05) (0.06) (0.07) (0.08) (0.06) (0.06) (0.06) (0.11) (0.14)R2 69.3 82.8 83.1 84.14 75.31 16.27 69 71.67 77.1 79.66 75.8 11.8129Table 4.8 - continued, Portfolio Return Properties of Capital- and Labor-intensive Firms in the Consumption-Goods SectorSorted by βIMC+controlsCapital-intensive firms Labor-intensive firms1 2 3 4 5 HL 1 2 3 4 5 HLPanel D. Sorting and post-sorting βIMCsortingβIMC-1.38 -0.44 -0.02 0.43 1.5 -1.55 -0.55 -0.09 0.39 1.57IMC 0.16** 0.24*** 0.45*** 0.66*** 1.14*** 0.97*** 0.67*** 0.37*** 0.54*** 0.95*** 1.31*** 0.64***(0.07) (0.07) (0.08) (0.09) (0.09) (0.06) (0.12) (0.09) (0.1) (0.1) (0.12) (0.14)Panel E. Market, SMB and IMCα 1.28 2.09*** 1.65 0.17 -0.19 -1.47 -1.35 0.02 0.98 2.04 -2.09 -0.75(1.27) (0.67) (1.15) (1.1) (1.87) (2.32) (1.73) (1.38) (1.12) (1.5) (1.94) (2.57)Market 0.95*** 0.92*** 0.93*** 0.99*** 1.0*** 0.05 1.04*** 1.04*** 1.02*** 1.03*** 1.21*** 0.17**(0.03) (0.02) (0.02) (0.03) (0.04) (0.05) (0.06) (0.02) (0.03) (0.03) (0.04) (0.09)SMB -0.04 -0.16*** -0.17*** -0.07 -0.04 0.01 0.54*** 0.18*** 0.23*** 0.42*** 0.56*** 0.02(0.04) (0.03) (0.04) (0.05) (0.07) (0.08) (0.09) (0.05) (0.09) (0.06) (0.09) (0.13)IMC -0.47*** -0.32*** -0.11** 0.01 0.47*** 0.93*** -0.29*** -0.43*** -0.27*** 0.05 0.22** 0.51***(0.05) (0.03) (0.05) (0.04) (0.06) (0.08) (0.08) (0.05) (0.06) (0.08) (0.09) (0.13)R2 76.21 86.8 83.39 84.11 78.81 39.98 69.91 76.27 78.62 79.25 74.17 16.29130Table 4.9: Parameter Values for Model CalibrationParameter Symbol ValueAggregate shocksMean growth rate of agg. productivity shock µx 0.01Volatility of agg. productivity shock σx 0.12Mean growth rate of the IST shock µz 0.001Volatility of the IST shock σz 0.033Idiosyncratic shocksPersistence of the firm-specific shock θ 0.35Volatility of the firm-specific shock σ 0.20Persistence of the project-specific shock θu 0.50Volatility of the project-specific shock σu 1.50Project arrival and depreciationProject depreciation rate δ 0.10Arrival rate parameter 1 µλ 2.00Arrival rate parameter 2 σλ 2.00Transition probability into high-growth state µH 0.075Transition probability into low-growth state µL 0.160Project arrival rate in the high-growth state λH 2.35Stochastic discount factorRisk-free rate r 0.03Price of risk of the aggregate productivity shock γx 0.69Price of the IST shock γz -0.35Capital and labor-intensity (production function)Capital-intensity of capital-intensive firms αH 0.775Capital-intensity of labor-intensive firms αL 0.575Labor intensity of capital-intensive firms 1 − αˆL 0.075Labor intensity of labor-intensive firms 1 − αˆH 0.275OtherProfit margin of the investment sector φ 0.07Aggregate wage factor w 0.007Labor force with flexible wage v 0.25Leverage difference LDIFF -10131Table 4.10: Simulated MomentsThe table shows the empirical moments, the calibrated moments from the Koganand Papanikolaou (2014) model (KP) and my extended model (Capital-intensitymodel). Most of the empirical moments are from Kogan and Papanikolaou (2014).Relative capital-labor ratio is the capital-labor ratio of capital-intensive firms rel-ative to the same ratio of labor-intensive firms. IQR indicates the interquartilerange. Relative capital labor ratio is the capital-labor ratio of capital intensivefirms divided by capital-labor ratio of labor-intensive firms. The moments for theextended model (Capital-intensity model) are medians across simulations of the model.Data KP model Capital-intensity modelAggregate momentsAgg. dividend growth, mean 0.025 0.017 0.017Agg. dividend growth, std. 0.118 0.150 0.206Agg. investment growth, mean 0.047 0.041 0.020Agg. investment growth, std. 0.157 0.171 0.200Asset pricing momentsMean excess return of market portfolio 0.059 0.056 0.073Volatility of market portfolio return 0.161 0.164 0.142Mean return of the IMC portfolio -0.014 -0.039 -0.015Volatility of the IMC portfolio return 0.113 0.115 0.312Correlation between the IMC and market return 0.45 NA 0.549Correlation between the IMC return investment shock NA 1 0.834Correlation between the IMC return aggregate shock NA 0 0.397Cross-sectional momentsFirms investment rate, median 0.112 0.121 0.156Firms investment rate, IQR 0.157 0.168 0.092Cash flows-to-capital, median 0.160 0.249 0.170Cash flows-to-capital, IQR 0.234 0.222 0.151Tobin’s Q, median 1.412 1.988 1.714Tobin’s Q, IQR 2.981 1.563 0.600Relative firm size, median 0.200 0.701 0.652Relative firm size, IQR 0.830 0.882 0.936Correlation betweem Tobin’s Q and relative firm size 0.160 -0.369 -0.287Capital- vs. labor-intensive firmsRelative capital-labor ratio, median 4.912 NA 4.959Table 4.11: Simulated Stock ReturnsThe table shows the properties of the portfolio returns of firms sorted by βIMC forcapital-intensive (panel A) and labor-intensive firms (panel B) simulated in the ex-tended model. Stocks are sorted into 10 portfolios based on their βIMC estimatedfrom a univariate regression. The average sorting βIMC within each portfolio are re-ported in the corresponding row. α is the annualized average excess return estimatedin a univariate regression with market excess return. β denotes the portfolios’ expo-sure to the market excess return. βIMC is the portfolios’ post-ranking exposure tothe IMC portfolio estimated from a univariate regression. βz denotes the true expo-sure to the investment shock zt. The reported results are medians across the simulations.p1 p2 p3 p4 p5 p6 p7 p8 p9 p10 HLPanel A. Capital-intensive firmsAverage excess return 8.052 7.846 7.741 7.659 7.308 7.010 6.706 6.481 5.510 4.309 -6.708α -2.734 -3.123 -3.349 -3.490 -3.903 -4.387 -4.797 -5.358 -6.497 -8.162 -5.269β 0.895 0.894 0.903 0.910 0.921 0.927 0.943 0.958 0.986 1.030 0.129sortingβIMC0.149 0.234 0.267 0.299 0.329 0.360 0.396 0.444 0.514 0.672βIMC 0.188 0.185 0.198 0.206 0.222 0.236 0.259 0.289 0.350 0.446 0.238βz 0.613 0.633 0.714 0.820 0.956 1.095 1.329 1.588 2.110 2.944 2.309Panel B. Labor-intensive firmsAverage excess return 11.638 11.644 11.305 11.345 11.171 11.182 10.784 10.108 9.532 8.731 -6.219α -0.168 -0.635 -0.754 -0.772 -0.899 -1.080 -1.516 -2.025 -2.678 -3.871 -3.695β 0.993 0.992 0.997 0.993 1.002 1.003 1.001 1.006 1.003 1.021 0.021sortingβIMC0.090 0.190 0.230 0.263 0.293 0.322 0.352 0.388 0.443 0.603βIMC 0.193 0.179 0.188 0.195 0.195 0.202 0.216 0.233 0.272 0.349 0.142βz 0.443 0.333 0.377 0.384 0.489 0.557 0.679 0.853 1.207 1.807 1.383132Chapter 5ConclusionThe thesis is a collection of three essays on finance and labor economics. Each essays studies aspecific question about technological innovations to better understand how technology shocksgenerate risk in the economy and affect stock prices.The first essay documents a very strong co-movement between prices of capital goods andstock prices of firms with a high share of displaceable labor. This is a new empirical resultthat establishes a very strong link between a macroeconomic variable and the dynamics ofstock returns. I develop a theoretical model of optimal technology adoption of competitivefirms and heterogeneous household. Using this model, I show that optimal adoption of costlytechnology such as replacement of routine labor by new and cheaper machines can decreasethe value of firms as in equilibrium competition erodes firms profits. This is likely to occurin the event of technology shock that decreases the price of new technology, which can createa co-movement between stock prices and prices of capital goods as observed in the data. Ifurther show that firms with negative exposure to these technology shocks earn a 4% per yearreturn premium. The premium is positively predicted by decreases in the cost of capital goodsand appears to be related to job displacement and labor income risks that arise from capital-labor substitution induced by technology shocks. At the firm level, I confirm that a largenegative exposure to technology shocks predicts lower employment and profitability followingtechnology shocks, and these effects are amplified by higher within-industry competition.The second essay examines how an investment shock affects different types of labor. I showthat investment shocks are an important source of job displacement and labor income risk.I use geographical variation in labor routine intensity and show that investment shock havedifferential and long lasting effects on routine and non-routine jobs. Areas with a high shareof routine labor experience higher unemployment and lower wage growth following a positiveinvestment shock. In contrast, areas with a low share of routine labor are affected significantlyless or even benefit. Using detailed industry level observations, I show that an investmentshock operates through substituting capital for labor in occupations that have the potentialto be automated. I use individual worker micro data to measure the job displacement risk. Ishow that a positive investment shock significantly increases the probability of job loss overmultiple years.The third essay examines the role of capital intensity for measuring firms exposure toinvestment shock using a measure and framework applied to study the exposure of firmsgrowth opportunities to these shocks. I show that capital-intensive stocks sorted by theexposure to this measure generate a highly significant average return premium of up to 5%annually. A similar return premium is present in the sub-sample of capital-intensive firmsbut absent among labor-intensive firms, while the exposures to the IMC portfolio are similarin both sub-samples. I show that this puzzle can be, to some extent, resolved if the originalmodel applied in prior literature is extended by labor and allows the IMC portfolio to beexposed also to other risk factors.133In the future work, I plan to extend my thesis in several dimensions. For instance, the firstessay is based on a static two-period general equilibrium model. It would be interesting toextend it to a fully dynamic infinite horizon model. Such framework would allow to study theevolution of firms exposure to technology shocks over time. A departure from the assumedDixit-Stiglitz imperfect competition to a more general setting of imperfect competition wouldalso allow to study the evolution of industry and firm size. In the second essay, combining theobservation units from multiple levels of observations, e.g., a worker employed in a certainindustry or living in a certain economic area, would further allow to study the importance ofthe economic environment of the effects of investment shocks.134BibliographyAbis, S. (2017). Man vs. machine: Quantitative and discretionary equity management. Un-published working paper. Columbia Business School.Acemoglu, D. and Autor, D. (2011). Skills, tasks and technologies: Implications for employ-ment and earnings. Handbook of labor economics, 4:1043–1171.Acemoglu, D., Autor, D., Dorn, D., Hanson, G. H., and Price, B. (2016). Import competitionand the great us employment sag of the 2000s. Journal of Labor Economics, 34(S1):S141–S198.Acemoglu, D. and Restrepo, P. (2017). Robots and jobs: Evidence from us labor markets.Acemoglu, D. and Restrepo, P. (2018). The race between man and machine: Implicationsof technology for growth, factor shares, and employment. American Economic Review,108(6):1488–1542.Atkeson, A. and Kehoe, P. J. (2007). Modeling the transition to a new economy: lessonsfrom two technological revolutions. American Economic Review, 97(1):64–88.Autor, D. and Dorn, D. (2013). The growth of low-skill service jobs and the polarization ofthe us labor market. The American Economic Review, 103(5):1553–1597.Autor, D. H., Levy, F., and Murnane, R. J. (2003). The skill content of recent technologicalchange: An empirical exploration. The Quarterly journal of economics, 118(4):1279–1333.Bartlesman, E. and Gray, W. B. (1996). The nber manufacturing productivity database.Becker, R., Gray, W., and Marvakov, J. (2013). Nber-ces manufacturing industry database:Technical notes. NBER Working Paper, 5809.Belo, F., Li, J., Lin, X., and Zhao, X. (2017). Labor-force heterogeneity and asset prices:The importance of skilled labor. The Review of Financial Studies, 30(10):3669–3709.Belo, F., Lin, X., and Bazdresch, S. (2014). Labor hiring, investment, and stock returnpredictability in the cross section. Journal of Political Economy, 122(1):129–177.Bretscher, L. (2017). From local to global: Offshoring and asset prices.Brynjolfsson, E. and McAfee, A. (2014). The second machine age: Work, progress, andprosperity in a time of brilliant technologies. WW Norton & Company.Bustamante, M. C. and Donangelo, A. (2017). Product market competition and industryreturns. The Review of Financial Studies, 30(12):4216–4266.135Campbell, J. Y. and Shiller, R. J. (1988). Stock prices, earnings, and expected dividends.The Journal of Finance, 43(3):661–676.Chen, H. J., Kacperczyk, M., and Ortiz-Molina, H. (2011). Labor unions, operating flexibility,and the cost of equity. Journal of Financial and Quantitative Analysis, 46(1):25–58.Constantinides, G. M. and Duffie, D. (1996). Asset pricing with heterogeneous consumers.Journal of Political economy, 104(2):219–240.Constantinides, G. M. and Ghosh, A. (2017). Asset pricing with countercyclical householdconsumption risk. The Journal of Finance, 72(1):415–460.Corhay, A. (2017). Industry competition, credit spreads, and levered equity returns.Corhay, A., Kung, H., and Schmid, L. (2017). Competition, markups and predictable returns.Cummins, J. G. and Violante, G. L. (2002). Investment-specific technical change in the unitedstates (1947–2000): Measurement and macroeconomic consequences. Review of Economicdynamics, 5(2):243–284.Daniel, K. and Titman, S. (1997). Evidence on the characteristics of cross sectional variationin stock returns. the Journal of Finance, 52(1):1–33.David, H. (2013). The” task approach” to labor markets: an overview. Technical report,National Bureau of Economic Research.Davis, S. J., Faberman, R. J., and Haltiwanger, J. (2006). The flow approach to labormarkets: new data sources and micro–macro links. The Journal of Economic Perspectives,20(3):3–26.Donangelo, A. (2014). Labor mobility: Implications for asset pricing. The Journal of Finance,69(3):1321–1346.Dorn, D. (2009). Essays on inequalitiy, spatial interaction, and the demand for skills. na.Eisfeldt, A. L. and Papanikolaou, D. (2013). Organization capital and the cross-section ofexpected returns. The Journal of Finance, 68(4):1365–1406.Fama, E. and French, K. (2015). A five-factor asset pricing model.[fama-miller workingpaper]. 2014. Journal of Financial Economics, 116(1):1–22.Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks andbonds. Journal of financial economics, 33(1):3–56.Favilukis, J. and Lin, X. (2015). Wage rigidity: A quantitative solution to several assetpricing puzzles. The Review of Financial Studies, 29(1):148–192.Favilukis, J. and Lin, X. (2016). Does wage rigidity make firms riskier? evidence fromlong-horizon return predictability. Journal of Monetary Economics, 78:80–95.Fisher, J. D. (2006). The dynamic effects of neutral and investment-specific technologyshocks. Journal of political Economy, 114(3):413–451.136Fox, D. R. and McCully, C. P. (2016). Concepts and methods of the u.s. national incomeand product accounts.Garlappi, L. and Song, Z. (2016). Can investment shocks explain the cross section of equityreturns? Management Science, 63(11):3829–3848.Garlappi, L. and Song, Z. (2017). Capital utilization, market power, and the pricing ofinvestment shocks. Journal of Financial Economics, 126(3):447–470.Garlappi, L. and Song, Z. (2018). An exposure to investment shocks: a new approach basedon investment data. Working paper, University of British Columbia.Gaˆrleanu, N., Kogan, L., and Panageas, S. (2012). Displacement risk and asset returns.Journal of Financial Economics, 105(3):491–510.Gaˆrleanu, N. and Panageas, S. (2017). Finance in a time of disruptive growth.Gomes, J. F., Kogan, L., and Yogo, M. (2009). Durability of output and expected stockreturns. Journal of Political Economy, 117(5).Gordon, R. J. (1990). The measurement of durable goods prices. University of Chicago Press.Greenwood, J., Hercowitz, Z., and Huffman, G. W. (1988). Investment, capacity utilization,and the real business cycle. The American Economic Review, pages 402–417.Greenwood, J., Hercowitz, Z., and Krusell, P. (1997). Long-run implications of investment-specific technological change. The American Economic Review, pages 342–362.Greenwood, J., Hercowitz, Z., and Krusell, P. (2000). The role of investment-specific techno-logical change in the business cycle. European Economic Review, 44(1):91–115.Greenwood, J. and Jovanovic, B. (1999). The information-technology revolution and thestock market. American Economic Review, 89(2):116–122.Greenwood, J. and Yorukoglu, M. (1997). 1974. In Carnegie-Rochester conference series onpublic policy, volume 46, pages 49–95. Elsevier.Griliches, Z. (1961). Hedonic price indexes for automobiles: An econometric of quality change.In The Price Statistics of the Federal Goverment, pages 173–196. NBER.Guvenen, F., Ozkan, S., and Song, J. (2014). The nature of countercyclical income risk.Journal of Political Economy, 122(3):621–660.Hansen, L. P. and Hodrick, R. J. (1980). Forward exchange rates as optimal predictors offuture spot rates: An econometric analysis. Journal of Political Economy, 88(5):829–853.Hobijn, B. and Jovanovic, B. (2001). The information-technology revolution and the stockmarket: Evidence. American Economic Review, 91(5):1203–1220.Hornstein, A. and Krusell, P. (1996). Can technology improvements cause productivityslowdowns? NBER Macroeconomics annual, 11:209–259.137Hou, K. and Robinson, D. T. (2006). Industry concentration and average stock returns. TheJournal of Finance, 61(4):1927–1956.Hou, K., Xue, C., and Zhang, L. (2017). Replicating anomalies. Technical report, NationalBureau of Economic Research.Jovanovic, B. and Rousseau, P. L. (2005). General purpose technologies. In Handbook ofeconomic growth, volume 1, pages 1181–1224. Elsevier.Justiniano, A., Primiceri, G. E., and Tambalotti, A. (2010). Investment shocks and businesscycles. Journal of Monetary Economics, 57(2):132–145.Karabarbounis, L. and Neiman, B. (2013). The global decline of the labor share. TheQuarterly Journal of Economics, 129(1):61–103.Knesl, J. (2018). Job displacement through technology, how technological shocks affect house-holds’ income and unemployment.Kogan, L. and Papanikolaou, D. (2013). Firm characteristics and stock returns: The role ofinvestment-specific shocks. The Review of Financial Studies, 26(11):2718–2759.Kogan, L. and Papanikolaou, D. (2014). Growth opportunities, technology shocks, and assetprices. The journal of finance, 69(2):675–718.Kogan, L., Papanikolaou, D., Schmidt, L., and Song, J. (2018). Technological innovation andthe distribution of labor income growth. Society for Economic Dynamics.Kogan, L., Papanikolaou, D., and Stoffman, N. (2017). Winners and losers: Creative de-struction and the stock market. Technical report, National Bureau of Economic Research.Krusell, P., Ohanian, L. E., R´ıos-Rull, J.-V., and Violante, G. L. (2000). Capital-skill com-plementarity and inequality: A macroeconomic analysis. Econometrica, 68(5):1029–1053.Kuehn, L.-A., Simutin, M., and Wang, J. J. (2017). A labor capital asset pricing model. TheJournal of Finance, 72(5):2131–2178.Lettau, M. and Ludvigson, S. (2001). Consumption, aggregate wealth, and expected stockreturns. the Journal of Finance, 56(3):815–849.Loualiche, E. et al. (2014). Asset pricing with entry and imperfect competition. Unpublishedmanuscript, Sloan School of Management, MIT.Newey, W. K. and West, K. D. (1987). A simple, positive semi-definite, heteroskedasticityand autocorrelation consistent covariance matrix. Econometrica (1986-1998), 55(3):703.Papanikolaou, D. (2011). Investment shocks and asset prices. Journal of Political Economy,119(4):639–685.Pastor, L. and Veronesi, P. (2009). Technological revolutions and stock prices. AmericanEconomic Review, 99(4):1451–83.138Schmidt, L. (2016). Climbing and falling off the ladder: Asset pricing implications of labormarket event risk.Spitz-Oener, A. (2006). Technical change, job tasks, and rising educational demands: Lookingoutside the wage structure. Journal of labor economics, 24(2):235–270.Tolbert, C. M. and Killian, M. S. (1987). Labor market areas for the united states. washing-ton, dc: Us department of agriculture. Economic Research Service.Tolbert, C. M. and Sizer, M. (1996). Us commuting zones and labor market areas: A 1990update.Wasshausen, D., Moulton, B. R., et al. (2006). The role of hedonic methods in measuringreal gdp in the united states. Technical report, Bureau of Economic Analysis.Zeira, J. (1998). Workers, machines, and economic growth. The Quarterly Journal of Eco-nomics, 113(4):1091–1117.Zhang, M. B. (2018). Labor-technology substitution: Implications for asset pricing. Journalof Finance, forthcoming.139Appendix AAppendix to Chapter 2A.1 ParametersTable A.1: Parameters for an Illustrative Model SolutionParameter Valueρ 0.6ν 0.95χ 1θ 2N 1000Nh 100ALow 0.85AHigh 0.85-1.05β 0.86γ 21 0.39pH 0.15p 0.5κ 8× 10−5140A.2 The O*NET DatabaseThe O*NET database provides a comprehensive set of characteristics describing occupationsin a number of dimensions such as workers’ abilities, skills, knowledge and experience, workvalues, work styles, work activities, etc. The O*NET collects the underlying data in a two-stage design. First, O*NET selects a statistically random sample of businesses that areexpected to employ targeted occupations. Second, it selects a random sample of workers inthe targeted occupations within those businesses and collects the data by surveying the jobincumbents using questionnaires. Additional variables such as variables describing skills andabilities are developed from the workers’ responses by occupational analysts. O*NET con-tinuously updates the occupational characteristics by ongoing surveys and provides updatedoccupational characteristics in new database releases quarterly/semi-annually starting from2003. This procedure allows to track the changes in occupational characteristics. The firstreleases of the O*NET database before 2003, however, are based on the data from Dictionaryof Occupational Titles (DOT), which is a predecessor of O*NET. The underlying data inDOT refers to years 1991 (revised 4th edition) and 1977 (4th edition). The number of occu-pations in the O*NET database releases varies between 812 and 1122 occupations dependingon the occupational classification used in each database release. The occupations are cur-rently classified by the 2010 O*NET - Standard Occupational Classification (O*NET SOC)that is a refined classification based on 2010 Standard Occupational Classification (SOC). 6161The previous releases of the O*NET database are based on previous versions of O*NET-SOC, namely2000, 2006, and 2009. The first release of the O*NET database in 1998 uses O*NET 98 OU classification thatis based on the Occupational Employment Statistics classification (OES classification).141A.3 Validity and Robustness of the RTI MeasureThe construction of the RTI measure involves some degree of subjectivity in selecting theunderlying descriptors as well as in the aggregation of the task inputs into one number ex-pressing the routine task intensity. To address the potential concerns stemming from thefreedom of selecting the underlying descriptors, I constructed alternative RTI measures fromsubsets and extended sets of descriptors and found that alternative measures identify routineand non-routine occupations in a very similar way. An aggregation of the underlying taskinputs without logarithmic transformation as in formula 2.24 also generates very comparableRTI measure. I also compare my RTI measure with the original RTI measure from Autor andDorn (2013). When matched to the common occupational classification, the measures havestrong positive correlations across the occupations. Routine intensive occupations identifiedby each of these measures overlap and the overlap is also strong for non-routine occupations.The fact that these measures are constructed from different data sources and using differentoccupational characteristics but still succeed to identify the occupational routine task inten-sity in a very similar way increases the validity of these measures. Despite these validity androbustness checks, the use of the RTI measure still requires caution. First the underlyingdata can still be subject to some measurement errors and the selected descriptors may notnecessarily describe all dimensions of routine intensity or automation potential of a givenoccupation. Second the RTI measure is defined only up to the occupational level and henceomits any variation in the routine task intensity across workers in the same occupation orof the same occupation across different firms. Given these concerns, the RTI measure hasto be considered as an imprecise measure of the routine tasks and automation potential.62Second, the given specification of the RTI measure is still time- and technology-dependent.Although, I use multiple data sources that update the occupational characteristics over time,the choice of the underlying descriptors reflects the current and past state of the technology.For instance, ’Operating vehicles, mechanized devices, or equipment’ is used as a descriptorof non-routine manual task input. Using this measure, occupations such as bus, truck, andtaxi drivers or commercial pilots are ranked as non-routine which reflects the state of thetechnology in most of the 20th century. The advances in technology, however, change thecharacter of these occupations towards more routine intensive and potentially automatable.For example, various version of autopilots are widely used in commercial airplanes and theircapabilities to assist during different phases of the flight have been constantly increasing.Similarly, a number of tasks involved in driving a car or truck, such as manually changing thegears or navigation, have been automated or significantly facilitated. Currently a number ofcompanies both within and outside of car manufacturing are developing and testing systemsfor autonomous/automated vehicles. Given the ongoing changes in the available technology,this descriptor, ’Operating vehicles, mechanized devices, or equipment’, may be re-classifiedto measure routine manual input in near future.I match the RTI measured at the occupational level to firm-level observations by theOES conducted by Bureau of Labor Statistics. OES provides detailed tables of occupationalcomposition of the labor force at a detailed industry level starting from 1988. The industriesare classified at SIC 3-digit or NAICS 4-digit level across the whole U.S. economy. The number62David (2013) discusses the challenges of measuring task inputs in more detail.142of industries ranges between 290 and 378.63 I extrapolate the occupational composition fromthe first OES releases to the years before 1988 to extend my sample to 1970. To translatethe RTI measure from occupational level to industry level, I calculate for each industry thepercentage of employees employed in routine intensive occupations. I define routine intensiveoccupations as occupations above the top employment-weighted tercile of occupations rankedby RTI score in a given year.64 I report the exact match of the underlying OES file and thecorresponding release of the O*NET database in the appendix table A.19. I transform the RTIfrom industry to firm level by imposing the percentage of employees in routine occupationsin a given industry to all firms in that industry. Since the main focus of this paper is on thebetween-industry variation, the omission of the intra-industry variation in the RTI is not alimitation. Moreover, I will show in the subsequent text that there is a strong relationshipbetween labor-intensity and routine task intensity.To validate the AP measure, I examine whether it can predict future automation. Au-tomation, generally, refers to the substitution of machines or equipment for human workersin performing specific tasks. This can be reflected in various adjustments such as higher useof machines, reduction of employment and/or changes in the use of technology by specificoccupations. Table A.8 in appendix shows that industry’s share of routine labor predicts theuse of information and communication equipment (ICT) per worker over the following twodecades. This relationship is especially strong for manufacturing industries and is robust toother factors such as offshorability and import penetration. I next examine whether the RTIalso predicts the use of technology within occupations. For each occupation, I calculate atechnology-use index measuring the use of various forms of technology such as computers andelectronics and the importance of technology-relevant skills such as programming or control-ling machines. I report the full set of the underlying variables in appendix table A.6. TableA.9 in appendix shows that occupational routine task intensity predicts the use of technologyat occupational level. Occupations with relatively higher RTI score experienced relativelyhigher increase in the technology-use index in the following years. These two results indicatethat the AP measure and its component, the RTI measure, indeed predict automation atindustry and occupational level, respectively.6563I report the exact number of industries in each OES release in appendix, table A.1864The results in the subsequent sections are very similar if I use different thresholds such as 70-th percentilesor 80-th percentile.65Other studies also provide supportive evidence. For instance, a companion paper Knesl (2018) shows thatmanufacturing industries with high share of routine labor substitute equipment capital for human workers fol-lowing technology shocks. Autor and Dorn (2013) document a differential reduction of employment of routineoccupations in commuting zone with originally high share of routine labor. Spitz-Oener (2006) documents adifferential increase in requirements for complex skills in rapidly computerizing occupations.143A.4 Filtering the Stock Return DataThe sample consists of common stocks listed at the NYSE, NASDAQ, and AMEX. I excludefinancials SIC 6000-6799, utilities SIC 4900-4999, and public sector companies SIC≥9000. Toavoid the results to be driven by small and micro-cap firms, I exclude the bottom 20% offirms ranked by market capitalization. I also exclude the bottom 10% of firms ranked by theshare price. The results are similar when firms with share price below 2 dollars are excluded.I further exclude firms in the bottom 5% by trading volume. Results are similar when morerestrictive percentiles are applied. I further exclude firms with less than 500 employees ortotal net property, plant and equipment less than 5 millions. Estimation of firm’s betasrequires the firm to have monthly data over at least last 30 months. The sample of S&P500firms consists of the continuously updated index constituents constituents. I also excludefinancials and utilities.144A.5 Measuring the Technology Shocks Embodied inCapital, DetailsI formally derive the validity of the I-shock measure in a simple two-sector model fromCummins and Violante (2002). Final goods producers competitively produce final goods xtat price pct . The final goods can be either consumed or used as an input for productionof capital goods. The capital goods sector can produce it efficiency units from xt unitsof consumption goods according to production function it = qt xt. qt captures the level oftechnology in the capital goods sector. Prices in the capital goods sector are set competitivelyso that pitit = pctxt. Combining this result with the production function leads to pit/pct = 1/qtand hence ∆qt = ∆pct −∆pit. Accordingly, the changes in the prices of capital goods relativeto consumption goods measure the investment specific technology shock qt. Based on thismodel, I construct the measures of the technology shock by subtracting the annual log changesin the quality-adjusted prices of capital goods from the log changes in prices of consumptiongoods. Accordingly, an increase in the I-shock measure indicates a positive technology shockembodied in capital.The prices of the capital goods are from the price index of private fixed investment in non-residential equipment provided by the U.S. Bureau of Economic Analysis. This price index isa composite of price indexes of 25 different types of equipment that are broadly categorizedas information processing equipment, industrial equipment, transportation equipment, andother equipment. I measure the aggregate prices by the BEA’s price index of personal con-sumption expenditures for nondurable goods. The sampling frequency of both time series isannual. I use the data from 1970 to 2015. To account for possible bias due to lack of qualityadjustment in the price indexes of capital goods, I use the underlying data from Cummins andViolante (2002) and extrapolate the quality adjustment until 2015. The data in Cumminsand Violante (2002) directly accounts for the quality changes in the 24 types of equipmentthat comprise the aggregate price index of capital goods. The quality adjustment is basedon the original data from Gordon (1990). To construct the I-shock measure, I subtract thelog change in the quality-adjusted prices of capital goods from the log change in prices ofnondurable consumption goods. I also calculate an alternative measure without using thequality adjustment from Cummins and Violante (2002). This measure relies completely onthe quality adjustment in the underlying price indexes from BEA. The results in this paperare very similar when I use the alternative measure. The time-series variation of both mea-sures is very similar. The major difference between the quality-adjusted and the alternativemeasure is the long-term average. The average is higher for the quality-adjusted measure asit reflects the improvements in quality of the capital goods over time.145A.6 Additional ResultsFigure A.1: Equilibrium Automation and Profits.This figure plots equilibrium relationship between the magnitude of the technology shock Atand the equilibrium percentage of firms that automate within the sector 1 (panel A), theequilibrium percentage and the dividends of firms in sector 1 (panel B), and the equilibriumpercentage and the dividends of firms in sector 2 (panel C). Dividends are scaled by a factor100.Figure A.2: Technology Shocks, Equilibrium Automation, and Profits at Different Margins.This figure plots the equilibrium relationship between the magnitude of the technology shockAt and percentage of firms that automate within the sector 1 (panel A), the dividends of firmsin sector 1 (panel B) and the dividends of firms in sector 2 (panel C) at three different levelsof profit margins (high, medium, and low). High profit margins are defined by elasticityof substitution ν = 0.90, medium profit margins by ν = 0.93 and low profit margins byν = 0.95. Other model parameters are stated in the description of figure 2.2. Dividends andthe magnitude of the technology shock At are normalized to allow comparison between modelsolutions at different parameter values.146Figure A.3: Number of Firms and Correlation between ln(LK)and LroutineLtotal.Panel A shows the number of firms in the sample from 1970 to 2015. Panel B showsthe correlation between labor intensity ln(LK)and routine intensity LroutineLtotalacross indus-tries. Industries are defined at NAICS 4-digit level and each industry consists of at least5 firms. Blue line depicts the correlation between ln(LK)and LroutineLtotal, each of whichis the industry mean across all firms in each industry. Solid red line is based on in-cumbent firms, i.e. firm age > 5 years, and excludes all entrants between 1985 and2000. Dashed red line is based solely on entering firms, i.e. firms with age <= 5 years.147Figure A.4: RTI at Occupations Level in 2003 and 2016.This figure plots the RTI score for each occupation (single dot) in 2003 (x-axis) and in 2016(y-axis). The blue line is a 45-degree line indicating a constant RTI score over time. The redline is fitted by OLS to describe the changes in occupations’ RTI score between 2003 and 2016.148Figure A.5: Technology Shocks (I-shock) and Low AP minus High AP Firms (by DifferentSamples of Firms).This figure plots the I-shock measure, dashed blue line, and the annual return of the zero-cost portfolio (KML), solid red line. The zero-cost portfolio has a long position in firms withlow automation potential, i.e. the bottom quintile of firms ranked by the AP measure, anda short position in firms with high automation potential, i.e. the top quintile. The stockreturns within the short and long position of the KML portfolio are value weighted. The KMLportfolio consists of all firms in the full sample in panel A and it excludes all firms in industriesproducing capital goods in panel B. The KML portfolio consists of non-manufacturing firms,SIC<2000 or SIC≥4000 in panel C and of manufacturing firms 2000≤SIC<4000 in panelD. Both time-series are plotted at annual frequency and are normalized to mean zero andstandard deviation of one. The sample period is from 1970 to 2015.149Figure A.6: Correlation between Automation Potential (AP) and βKML over Time.This figure plots the time-series of contemporaneous correlation between automation po-tential (AP) and the βKML. The black line shows the cross-sectional correlation cal-culated at the firm level. Red line shows cross-section correlation calculated at thelevel of NAICS 4-digit industries. Industries’ automation potential (AP) and βKMLare the average AP and βKML across firms within each industry, respectively. Theblue line is the weighted correlation across NAICS 4-digit industries, where the in-dustry observations are weighted by the number of firms within each industry. Thetime-series are plotted at annual frequency and show the moving average of correla-tion coefficients over the last three year. The sample period is from 1975 to 2015.150Table A.2: Number of Occupations with RTI measure 1998 - 2016This table reports the number of occupations in releases of the O*NET databases from1998 to 2016 for which the RTI measure. The number of occupations with calcu-lated RTI measure is lower than then total number of occupations in each of theO*NET database releases as the detailed O*NET-SOC occupations are matched withthe occupations in the corresponding version of the Standard Occupational Classifica-tion (SOC) that is the underlying classification in the Occupational Employment Statis-tics (OES). Since the O*NET-SOC is slightly more detailed than the SOC, some de-tailed occupations from the O*NET SOC are aggregated to a single SOC occupation.O*NET 98 O*NET 5.1 O*NET 7.0 O*NET 9.0 O*NET 9.0Number of occupations 724 654 688 712 723O*NET 12.0 O*NET 13.0 O*NET 14.0 O*NET 15.1 O*NET 16.0Number of occupations 725 730 739 730 731O*NET 12.0 O*NET 13.0 O*NET 14.0 O*NET 15.1 O*NET 16.0Number of occupations 725 730 739 730 731151Table A.3: The Underlying O*NET Variables for Constructing the Occupations’ RoutineTask IntensityThis table shows the descriptors of occupations used to construct the routine task inten-sity measure. Panel A lists the descriptors for abstract tasks, panel B for routine tasksand panel C for manual tasks. Columns Category and ID provide information for locat-ing the descriptors in the O*NET Content Model. Scale reports the type of scaled used.Type of thetaskCategory Variable ID ScalePanel A. Abstract tasksAnalytical Abilities Fluency of ideas 1.A.1.b.1 IMOriginality 1.A.1.b.2 IMDeductive reasoning 1.A.1.b.4 IMInductive reasoning 1.A.1.b.5 IMInformation ordering 1.A.1.b.6 IMCategory flexibility 1.A.1.b.7 IMWork values Independence 1.B.2.f ENSkills Critical thinking 2.A.2.a IMActive learning 2.A.2.b IMComplex problem solving 2.B.2.i IMJudgment and decision making 2.B.4.e IMSystems analysis 2.B.4.g IMSystems evaluation 2.B.4.h IMSocial Skills Active listening 2.A.1.b IMSocial Perceptiveness 2.B.1.a IMPersuasion 2.B.1.c IMNegotiation 2.B.1.d IMWorkactivitiesInterpreting the meaning of infor-mation for others4.A.4.a.1 IMCommunication with supervisors,peers, or subordinates4.A.4.a.2 IMProvide consultation and advice toothers4.A.4.b.6 IMPanel B. Routine tasksRoutinegeneralWork context Importance of being exact and ac-curate4.C.3.b.4 CXPace determined by speed of equip-ment4.C.3.d.3 CXRoutinecognitiveWork context Importance of repeating same tasks 4.C.3.b.7 CXRoutinemanualSkills Quality control analysis 2.B.3.m IMWorkactivitiesHandling and moving objects 4.A.3.a.2 IMWork context Spend time using your hands tohandle, control, or feel objects toolsor controls4.C.2.d.1.g CXSpend time making repetitive mo-tions4.C.2.d.1.i CXPanel C. Manual tasksNon-routinemanual andsocialAbilities Multilimb coordination 1.A.2.b.2 IMSpacial coordination 1.A.1.f.1 IMWorkactivitiesOperating vehicles, mechanized de-vices, or equipment4.A.3.a.4 IM152Table A.4: Summary Statistics of the Routine Task Intensity and the Underlying Tasks atthe Occupational Level in Three Releases of the O*NET DatabaseThis table reports the summary statistics of the RTI measure, panel A, and theunderlying task inputs, panels B, C, and D. The reported summary statistics arebased on three different releases of the O*NET database. O*NET 98 that is en-tirely based on data from the Dictionary of Occupational Titles, O*NET 5.1 that isbased on newly collected data until 2003 and O*NET 21.0 that incorporates (mul-tiple) updates of the underlying data for the majority of occupations since 2003.Mean Median St. dev. Min MaxPanel A. Routine task intensityONET 98 -0.388 -0.428 1.111 -3.988 2.518ONET 5.1 -0.836 -0.922 1.01 -3.941 2.125ONET 21.0 -1.366 -1.405 0.654 -3.453 1.311Panel B. Abstract tasksONET 98 3.862 3.202 2.316 0.35 10ONET 5.1 4.382 4.216 2.359 0.311 10ONET 21.0 5.506 5.626 2.105 0.336 10Panel C. Routine tasksONET 98 5.629 5.796 1.654 0.57 10ONET 5.1 4.692 4.724 1.707 0.288 10ONET 21.0 5.944 6.081 1.674 0.114 10Panel D. Manual tasksONET 98 2.875 2.595 1.55 0.288 10ONET 5.1 3.136 2.835 1.655 0.135 10ONET 21.0 4.657 4.497 1.492 1.315 10Table A.5: Examples of Routine and Non-routine OccupationsThis table lists representative routine and non-routine occupations. These occu-pations are selected to represent typical occupations with low RTI (non-routine)and high RTI (routine) in multiple years. The full lists of the 15 most rou-tine and 15 most non-routine occupations based on O*NET 98, O*NET 5.1 andO*NET 20.1 including a detailed description of each occupation are in appendix.Routine (high RTI) Non-routine (low RTI)Data keyers Residential councelorsWood machinists ClergyHand sewers Kindergarden teachersFile clerks Sales managersFast food cooks Commercial pilotsMachine feeders and offbearers Chief executivesPayroll and timekeeping clerks Travel guidesTax preparers Construction managers153Table A.6: The Underlying O*NET Variables for Constructing the Occupations’ Technology-Use IndexThis table shows the descriptors of occupations used to construct the technology-use index. Columns Category and ID provide information for locating the de-scriptors in the O*NET Content Model. Scale reports the type of scaled used.Category Variable ID ScaleSkills Technology design 2.B.3.b IMEquipment selection 2.B.3.c IMInstallation 2.B.3.d IMProgramming 2.B.3.e IMOperation monitoring 2.B.3.g IMOperation and control 2.B.3.h IMEquipment maintenance 2.B.3.j IMKnowledge Computers and electronics 2.C.3.a IMEngineering and technology 2.C.3.b IMMechanical 2.C.3.e IMWork activities Controlling machines and processes 4.A.3.a.3 IMInteracting with computers 4.A.3.b.1 IMRepairing and maintaining elec-tronic equipment4.A.3.b.5 IMWork context Degree of automation 4.C.3.b.2 IM154Table A.7: Ranking of Industries with Highest and Lowest APThis table shows industries with the lowest and highest automation potential measured by AP.Industries are defined at SIC 1987 3-digit level. ln(KLroutine)= −AP to facilitate readability.High AP industries Low AP industriesSIC Industry title ln(KLroutine)SIC Industry title ln(KLroutine)7360 Personnel Supply Services 1.622330 Women’s, Misses’, and Juniors’Outerwear2.262300 Men’s and Boys’ Suits, Coats,and Overcoats2.273140 Footwear, Except Rubber 2.322320 Men’s and Boys’ Furnishings,Work Clothing, and Allied Gar-ments2.332250 Knitting Mills 2.532210 Broadwoven Fabric Mills, Cot-ton2.747340 Services To Dwellings and OtherBuildings2.755910 Drug Stores and ProprietaryStores2.813910 Jewelry, Silverware, and PlatedWare2.832510 Household Furniture 2.872450 Wood Buildings and MobileHomes2.913650 Household Audio and VideoEquipment, and Audio Record-ings2.982200 Broadwoven Fabric Mills, Cot-ton2.993940 Dolls, Toys, Games and Sportingand Athletic3.022530 Public Building and Related Fur-niture3.055410 Grocery Stores 3.063050 Gaskets, Packing, and SealingDevices and Rubber3.062010 Meat Products 3.088720 Accounting, Auditing, andBookkeeping Services3.081310 Crude Petroleum and NaturalGas7.544920 Gas Production and Distribution 6.984400 Deep Sea Foreign Transportationof Freight6.964520 Air Transportation, Nonsched-uled6.864700 Arrangement of PassengerTransportation6.564010 Railroads 6.54890 Communications Services, notelsewhere classified6.381220 Bituminous Coal and LigniteMining6.197510 Automotive Rental and Leasing,Without Drivers6.182910 Petroleum Refining 6.154810 Telephone Communications 6.14840 Cable and Other Pay TelevisionServices6.077900 Dance Studios, Schools, andHalls5.978300 Individual and Family SocialServices5.937350 Miscellaneous Equipment Rentaland Leasing5.914510 Air Transportation, Scheduled,and Air Courier5.881040 Gold and Silver Ores 5.825170 Petroleum and Petroleum Prod-ucts5.754830 Radio and Television Broadcast-ing Stations5.751000 Iron Ores 5.68155Table A.8: Share of Routine Labor and Changes in Information and Communication Equipment per Employee at Industry LevelThe coefficients are from OLS regression ∆(ICTemp)i,t+1= constant+γ1×Shroutinei,t +γ2×offshorei,t+ ∆import penetrationi,t+1 +i,t+1. The dependent variable is the change in the information and communication equipment per worker in industry i betweenyear t and t + 1. The independent variables are Shroutinei,t the share of employees in routine occupations in industry i at thebeginning of the time period t, offshorei,t the industry i′s offshorability score at t, and ∆import penetrationi,t+1 the change inChina’s import penetration in industry i between t and t+ 1. The data for China’s import penetration are from Acemoglu et al.(2016) and available only for manufacturing industries between 1991 and 2011. The industries are defined at SIC 1987 2-digit level.The sample is restricted to industries that have at least 5 firms with Compustat data. Standard errors are in parentheses. ***Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.1970 - 19901970 - 1990manufacturingindustries1970 - 1990non-manufacturingindustries1990 - 20101990 - 2010manufacturingindustries1990 - 2010non-manufacturingindustriesConstant 2.623*** 1.865*** 2.284*** 1.371*** 0.365 1.175***(0.306) (0.486) (0.574) (0.155) (0.29) (0.238)Share of routine labor 1.504** 2.285** 3.988** -0.168 1.134** 1.07(0.619) (0.934) (1.612) (0.365) (0.553) (0.767)Offshoring potential -1.186*** -0.643 -1.46*** -0.04 0.188 -0.084(0.317) (0.452) (0.483) (0.157) (0.337) (0.189)∆ Import penetration 0.004(0.005)R2 0.153 0.088 0.401 0.002 0.137 0.027Observations 104 67 37 142 59 83156Table A.9: Routine Task Intensity and Changes in Technology Score within Occupations,2003-2016The coefficients are from OLS regression ∆Techocc,t+1 = constant + γ × RTIocc,t + occ,t+1.The dependent variable is the change in the occupation occ′s technology score between t andt + 1. The independent variable is the occupation occ′s RTI score at the beginning of thetime period t. The sample is restricted to occupations that have a unique SOC code in bothtime periods to avoid potential inaccuracy due to multiple matching of occupational codes indifferent time periods. The sample is further restricted to occupations whose characteristicshave been updated during the examined time period. First column reports estimates basedon O*NET 5.0 that is predominantly based on DOT data and O*NET 22.0 is based on jobincumbents data. To avoid a potential inaccuracy due to different underlying data sourcesat the beginning and end of the time period, the second column reports estimates based onO*NET 11.0 and O*NET 22.0 both of which are entirely based on job incumbents data.Standard errors are in parentheses. *** Significant at the 1 percent level. ** Significant atthe 5 percent level. * Significant at the 10 percent level.O*NET 5.0 vs. O*NET 22.02003 - 2017O*NET 11.0 vs. O*NET 22.02006 - 2017Constant -0.1*** -0.168***(0.015) (0.027)RTI 0.026** 0.061***(0.011) (0.017)R2 0.009 0.023Observations 607 579157Table A.10: I-shock Measure: Summary Statistics 1960-2015This table reports the time-series properties of the I-shock measure in panel A and corre-lations between the I-shock measure and other macroeconomic variables in panel B. TheI-shock measure is the change in the quality-adjusted relative price of investment goods. Itis calculated by subtracting the log change in the prices of nondurable consumption goodsfrom the the log change in the quality-adjusted prices of equipment. The GDP growth andAgg TFP shock are aggregate measures of economic growth and factor neutral productiv-ity shock, respectively. TFP shock of capital goods producers is a weighted average TFPshock of manufacturing industries producing investment goods, with weights based on theirrelative shipments to private fixed investments. TFP shock of consumption goods producersis calculated equivalently. The sampling frequency is annual. The sample period is from1960 to 2015. Source: I-shock measure: author’s calculation using data from U.S. Bureauof Economic Analysis and Cummins and Violante (2002); GDP growth and aggregate TFPshock: U.S. Bureau of Economic Analysis; TFP shock of capital goods producers and TFPshock of consumption goods producers: author’s calculation using data from U.S. Bureauof Economic Analysis and NBER-CES Manufacturing Industries database Bartlesman andGray (1996).Panel A. Time-series properties of the Ishock measureMean Standard deviation Serial correlationI-shock 4.008 3.297 0.232Panel B. Correlation with macroeconomic time-seriesTFP shock ofcapital goodsproducersTFP shock ofconsumptiongoods producersGDPgrowthAggTFPshockI-shock 0.42 0.006 0.256 -0.007158Table A.11: Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Technology ShocksThis table reports the productivity growth, changes in shipment prices and changes in thereal quantity of shipments during the major technology shocks from 1970 to 2015. Thesample in panels A-D includes 457 4-digit 1987 SIC industries in the manufacturing divi-sion and sample in panel E and F consists of 473 6-digit NAICS industries in the man-ufacturing division. Capital goods producers are industries which actively supply to pri-vate fixed investments based on the BEA Input-Output tables. Consumption goods pro-ducers are defined equivalently based on positive supply to Personal Consumption expen-ditures. There are 145 capital goods producers and 290 consumption goods producersin panels A-D and 115 capital goods producers and 196 consumption goods producers inpanels E and F. The Electronic Computer Manufacturing in Panel E is one industry de-fined by 6-digit NAICS code, 334111. The TFP shock is the cumulative growth in theTFP index during over each time period. All growth variables are value weighted averagesweighted by the industry’s shipments to the corresponding type of the final use. The growthvariables are in percent. Source: NBER-CES Manufacturing Industry Database Bartles-man and Gray (1996); U.S. Bureau of Economic Analysis, Benchmark Input-Output Data.Industries TFP shockShipment prices,growthShipments,real growthPanel A. 1st shock: 1972-1974All ManufacturingIndustries0.76 27.03 6.63Capital GoodsProducers4.02 13.83 11.63Consumption GoodsProducers0.53 23.03 1.25Panel B. 2nd shock: 1974-1976All ManufacturingIndustries-0.13 16.83 0.53Capital GoodsProducers0.77 19.97 -0.56Consumption GoodsProducers1.67 14.31 6.70Intermediate GoodsProducers-1.19 17.16 -1.73Panel C. 3rd shock: 1981-1983All ManufacturingIndustries1.48 3.55 -1.18Capital GoodsProducers-1.98 6.56 -2.40Consumption GoodsProducers2.69 4.18 3.29159Table A.11 Productivity Shocks of Capital Goods and Consumption Goods Manufacturersduring Significant Technology Shocks: ContinuedIndustries TFP shockShipment prices,growthShipments,real growthPanel D. 4th shock: 1985-1986All ManufacturingIndustries0.21 -1.93 1.20Capital GoodsProducers-0.75 0.58 -0.40Consumption GoodsProducers0.57 -1.41 1.71Panel E. 5th shock: 1997-2001All ManufacturingIndustries5.33 -0.09 22.63Capital GoodsProducers19.61 -9.68 66.29Consumption GoodsProducers4.17 0.71 14.51Electronic ComputerManufacturing215.83 -80.56 383.75Other CapitalGoods Producers-7.58 0.14 22.30Panel F. 6th shock: 2008-2009All ManufacturingIndustries-4.05 -8.28 -9.13Capital GoodsProducers-8.17 -0.07 -14.87Consumption GoodsProducers-5.72 -12.51 -4.21160Table A.12: Exposure of Portfolio Returns to Technology Shocks (I-shocks)This table reports results from regressing portfolio annual excess returns on a constant, the I-shock measure and return factors of Fama and French (2015) in panel A. Regressions in panelB include also the IMC portfolio of Papanikolaou (2011) that has a long position in producersof capital goods and short position in producers of consumption goods and aggregate growthin GDP. Stocks are assigned to one of five portfolios based on automation potential, AP.Portfolios are rebalanced annually at the end of June. Portfolio returns are value-weighted.The sample period is from 1970 to 2015. Newey and West (1987) standard errors are inparentheses. *** Significant at the 1 percent level. ** Significant at the 5 percent level. *Significant at the 10 percent level.High AP 2 3 4 Low APLow-High APKMLPanel A. Exposure to I-shock measure, aggregate market,and Fama and French (2015) factorsIshock -1.46*** -1.128*** -0.779*** -0.372* 0.574** 2.034***(0.244) (0.244) (0.245) (0.216) (0.246) (0.335)Market 1.038*** 0.921*** 0.964*** 0.89*** 0.949*** -0.089(0.073) (0.053) (0.039) (0.034) (0.043) (0.09)SMB 0.485*** 0.091 0.004 -0.098** -0.216*** -0.702***(0.084) (0.068) (0.083) (0.047) (0.055) (0.118)HML -0.054 0.07 -0.156* -0.066 -0.066 -0.012(0.093) (0.101) (0.08) (0.078) (0.088) (0.126)CMA 0.272* 0.138 0.136 -0.098 -0.035 -0.308*(0.136) (0.101) (0.112) (0.07) (0.062) (0.176)RMW 0.401*** 0.294** 0.037 0.023 -0.033 -0.434***(0.132) (0.125) (0.081) (0.062) (0.08) (0.137)R2 92.36% 91.38% 93.31% 94.86% 94.35% 73.01%Panel B. Exposure to I-shock measure, aggregate market,Fama and French (2015) factors, IMC portfolio and GDP growthIshock -1.075*** -1.059*** -0.582** -0.309 0.318 1.393***(0.255) (0.222) (0.243) (0.273) (0.251) (0.283)Market 1.077*** 0.908*** 0.982*** 0.911*** 0.922*** -0.155(0.087) (0.057) (0.035) (0.044) (0.041) (0.105)SMB 0.458*** 0.067 -0.024 -0.102** -0.238*** -0.696***(0.078) (0.058) (0.063) (0.044) (0.051) (0.105)HML 0.089 0.148 -0.024 -0.012 -0.06 -0.149(0.132) (0.091) (0.076) (0.072) (0.073) (0.125)CMA 0.176 0.019 0.016 -0.146** 0.033 -0.143(0.145) (0.115) (0.087) (0.065) (0.086) (0.211)RMW 0.415*** 0.27** 0.033 -0.001 0.108 -0.307**(0.122) (0.105) (0.111) (0.078) (0.089) (0.151)IMC 0.023 0.013 0.045 0.008 0.188** 0.165(0.116) (0.083) (0.077) (0.069) (0.076) (0.145)GDP -0.012*** -0.008** -0.011*** -0.004 0.005 0.016***(0.003) (0.004) (0.003) (0.004) (0.003) (0.005)R2 92.93% 91.83% 94.16% 95.56% 95.15% 77.8% 161Table A.13: Summary Statistics of Firm Characteristics at the Portfolio LevelThis table reports time-series averages of annual median firm characteristics at portfoliolevel. Portfolios are constructed from firms sorted on their automation potential, AP, andare rebalanced annually. Panel A consists of all firms in the sample and panel B includesonly the constituents of S&P500 index without financials and utilities. I report negative AP,i.e. (−)AP = ln ( KLroutine)to avoid negative numbers and facilitate readability. ln(KL)iscapital intensity, i.e. (−1)×labor intensity, LroutineLtotalis the percentage of employees in routineoccupations, BM is the book-to-market equity ratio, βmarket is the regression coefficient ofmarket excess return from rolling time-series regression of firm excess return onto marketexcess return and a constant, CF denotes cash flow, Lev denotes financial leverage, PCMdenotes the price-to-cost margin, Profitability is the ratio of sales minus cost of goods sold,interest expenses, and selling, general, and administrative expenses to book equity, Size isthe natural logarithm of the market capitalization in thousands, and Turnover is the fractionof shares traded to the total shares outstanding. The sample period is from 1970 to 2015.Source: The stock market data are from CRSP, data from financial statements are fromStandard and Poor’s Compustat, data on industries’ occupational employment compositionare from Occupational Employment Statistics (BLS) and occupation level data are from theO*NET database.(-)AP ln(KL)LroutineLtotalBM βmarket CF Lev PCM Profitability Size TurnoverPanel A. Portfolio characteristics, full sampleHigh AP 3.040 2.293 0.546 0.713 1.139 0.090 0.271 0.289 0.246 5.131 0.0772 3.688 2.909 0.470 0.659 1.139 0.093 0.288 0.316 0.255 5.547 0.0833 4.164 3.258 0.410 0.637 1.156 0.094 0.291 0.330 0.248 5.806 0.0904 4.789 3.713 0.349 0.615 1.157 0.093 0.331 0.334 0.248 6.182 0.097Low AP 6.038 4.820 0.312 0.638 1.137 0.097 0.415 0.331 0.245 6.700 0.096Panel B. Portfolio characteristics, S&P500 firmsHigh AP 3.400 2.720 0.564 0.555 1.043 0.105 0.307 0.318 0.311 7.502 0.0932 4.056 3.225 0.440 0.500 1.075 0.107 0.308 0.326 0.317 7.985 0.0923 4.598 3.618 0.387 0.469 1.046 0.109 0.334 0.390 0.317 8.060 0.1034 5.245 4.245 0.377 0.538 1.047 0.109 0.342 0.394 0.283 8.220 0.096Low AP 6.617 5.331 0.327 0.619 1.003 0.105 0.403 0.337 0.265 8.420 0.098162Table A.14: Mean Portfolio Returns and Alphas of Portfolios Sorted on βKML (Constituentsof S&P500)This table reports time-series averages of value-weighted portfolio excess returns in panelA, equally-weighted portfolio excess returns in panel B, results of regressing monthly value-weighted portfolio excess returns on a constant and market excess return in panel C andresults of regressing monthly value-weighted portfolio returns on a constant, market excessreturn, and the size (SMB) and value (HML) factors of Fama and French (1993) in panel D.Newey and West (1987) standard errors are reported in parentheses. Stocks are assigned toone of five portfolios based on their exposure to the KML portfolio, βKML. βKML is estimatedfrom rolling regressions of stock’s monthly excess return on a constant and monthly returnof the KML portfolio over the last 60 months. Portfolios are rebalanced annually at the endof June. The average excess returns and standard errors in panel A and B are annualizedaverages of monthly excess returns. The alpha estimates and their standard errors in panel Cand D are annualized. The sample consists of contemporaneous constituents of the S&P500excluding financials and utilities. The sample period is from 1975 to 2015. *** Significant atthe 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Low βKML 2 3 4 High βKMLHigh-LowβKMLPanel A. Value-weighted portfolio excess returnExcess return 12.85*** 8.91*** 9.24*** 7.79*** 5.69** -7.16***(2.9) (2.59) (2.59) (2.72) (2.48) (2.01)Panel B. Equally-weighted portfolio excess returnExcess return 12.57*** 10.99*** 10.77*** 9.4*** 7.82*** -4.75**(2.88) (2.73) (2.59) (2.55) (2.87) (2.06)Panel C. CAPMAlpha 4.309*** 1.338 1.698 0.004 -1.86 -6.168***(1.36) (1.596) (1.099) (1.054) (1.255) (2.131)Market 1.045*** 0.926*** 0.923*** 0.952*** 0.924*** -0.121(0.049) (0.068) (0.034) (0.025) (0.049) (0.088)R2 78.28% 75.29% 81.44% 83.1% 77.01% 1.86%Panel D. Three-factor model of Fama and French (1993)Alpha 4.156*** 0.423 1.95* 0.603 -0.702 -4.858***(1.136) (1.377) (1.073) (0.827) (1.034) (1.645)Market 1.033*** 0.973*** 0.939*** 0.999*** 0.932*** -0.1(0.044) (0.051) (0.035) (0.018) (0.041) (0.072)SMB 0.075 -0.081 -0.105** -0.281*** -0.193*** -0.268***(0.056) (0.055) (0.044) (0.05) (0.044) (0.087)HML -0.004 0.21* -0.003 0.008 -0.137** -0.134(0.082) (0.117) (0.05) (0.057) (0.066) (0.137)R2 78.45% 77.09% 81.88% 86.1% 78.98% 6.35%163Table A.15: Mean Portfolio Returns and Alphas of Portfolios Sorted on βKMLThis table reports time-series averages of value-weighted portfolio excess returns in panel A, equally-weighted portfolio excessreturns in panel B, results of regressing monthly value-weighted portfolio excess returns on a constant and market excess returnin panel C and results of regressing monthly value-weighted portfolio returns on a constant, market excess return, and the size(SMB) and value (HML) factors of Fama and French (1993) in panel D. Newey and West (1987) standard errors are reportedin parentheses. Stocks are assigned to one of five portfolios based on their exposure to the KML portfolio, βKML. βKML isestimated from rolling regressions of stock’s monthly excess return on a constant, monthly return of the KML portfolio, marketexcess return, and the size (SMB) and value (HML) factors of Fama and French (1993) over the last 60 months. Portfolios arerebalanced annually at the end of June. The average excess returns and standard errors in panel A and B are annualized averagesof monthly excess returns. The alpha estimates and their standard errors in panel C and D are annualized. The sample period isfrom 1980 to 2015. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Low βKML 2 3 4 5 6 7 8 9 High βKMLHigh-LowβKMLPanel A. Value-weighted portfolio excess returnExcess return 10.41*** 9.2*** 9.21*** 8.73*** 7.32*** 7.52*** 8.38*** 6.59** 6.19** 5.05 -5.36*(2.81) (2.82) (2.7) (2.65) (2.81) (2.81) (2.89) (2.93) (3.08) (3.6) (3.19)Panel B. Equally-weighted portfolio excess returnExcess return 10.39*** 9.5*** 9.41*** 9.19*** 9.32*** 9.53*** 10.13*** 9.6*** 9.14*** 8.24** -2.15(3.41) (3.23) (3.16) (3.21) (3.11) (3.09) (3.23) (3.33) (3.52) (4.09) (2.36)Panel C. CAPMAlpha 2.925* 1.913 1.908 1.569 -0.216 0.211 0.976 -0.917 -2.034 -3.683* -6.608**(1.633) (1.198) (1.527) (1.096) (1.057) (0.929) (1.172) (1.292) (1.769) (2.218) (3.327)Market 0.966*** 0.94*** 0.942*** 0.924*** 0.973*** 0.944*** 0.956*** 0.969*** 1.062*** 1.127*** 0.161(0.059) (0.049) (0.052) (0.043) (0.033) (0.037) (0.028) (0.044) (0.064) (0.073) (0.123)R2 72.88% 80.45% 80.51% 82.16% 83.79% 83.27% 82.92% 76.97% 75.35% 63.96% 1.66%164Table A.15 Mean Portfolio Returns and Alphas of Portfolios Sorted on βKML: ContinuedLow βKML 2 3 4 5 6 7 8 9 High βKMLHigh-LowβKMLPanel D. Three-factor model of Fama and French (1993)Alpha 2.984** 1.399 1.487 1.274 -0.025 0.446 0.952 -0.57 -0.826 -2.614 -5.598*(1.489) (1.19) (1.161) (0.845) (1.043) (0.888) (1.194) (1.239) (1.634) (2.146) (3.057)Market 0.967*** 0.966*** 0.958*** 0.936*** 0.97*** 0.943*** 0.971*** 0.942*** 1.023*** 1.084*** 0.117(0.055) (0.036) (0.037) (0.03) (0.03) (0.035) (0.026) (0.041) (0.053) (0.059) (0.104)SMB -0.021 -0.038 0.002 -0.004 -0.023 -0.051 -0.092** 0.089 -0.043 0.016 0.037(0.081) (0.051) (0.042) (0.033) (0.077) (0.046) (0.036) (0.066) (0.055) (0.091) (0.157)HML -0.009 0.107 0.084 0.059 -0.035 -0.04 0.016 -0.08 -0.235*** -0.215 -0.206(0.094) (0.082) (0.098) (0.091) (0.063) (0.037) (0.049) (0.066) (0.077) (0.131) (0.215)R2 72.89% 80.95% 80.76% 82.3% 83.85% 83.41% 83.25% 77.49% 76.81% 64.9% 2.82%165Table A.16: Panel Regressions of Annual Stock Returns on βKML and Firm Characteristics (Constituents of S&P500)This table reports results of regressing annual stock returns on a constant, βKML, other stocks characteristics and year fixedeffect as indicated. Standard errors, clustered at the firm level, are reported in parentheses. βKML is estimated fromrolling regressions of stock’s monthly excess return on a constant and monthly return of the KML portfolio over the last60 months. BM is the book-to-market equity ratio, βmarket is the regression coefficient of market excess return from rollingtime-series regression of firm excess return onto market excess return and a constant, CF denotes cash flow, Lev denotes fi-nancial leverage, PCM denotes the price-to-cost margin, Profitability is the ratio of sales minus cost of goods sold, inter-est expenses, and selling, general, and administrative expenses to book equity, Size is the natural logarithm of the marketcapitalization in thousands, and Turnover is the fraction of shares traded to the total shares outstanding. The sample con-sists of contemporaneous constituents of the S&P500 excluding financials and utilities. The sample period is from 1975 to2015. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)βKML -4.265∗∗∗ -2.649∗∗∗ -2.383∗∗∗ -2.843∗∗∗ -2.651∗∗∗ -2.639∗∗∗ -2.727∗∗∗ -2.540∗∗∗ -2.672∗∗∗ -2.590∗∗∗ -2.621∗∗∗(0.446) (0.485) (0.506) (0.476) (0.482) (0.486) (0.489) (0.505) (0.506) (0.477) (0.534)BM 0.609 0.541(0.521) (0.610)Beta -4.749∗∗∗ -4.589∗∗∗(0.609) (0.778)CF 6.694 4.420(4.572) (5.987)Lev 0.711 0.110(1.372) (1.655)PCM 3.460∗∗ 5.202∗∗∗(1.635) (1.790)Profitability 0.594 0.307(0.399) (0.402)Size 0.0454 -0.567∗(0.258) (0.293)Turnover -17.70∗∗∗ -7.082(3.826) (4.696)Year fixed effect No Yes Yes Yes Yes Yes Yes Yes Yes Yes YesN 14402 14402 13346 14402 14385 14365 14402 12096 14402 14402 12079R2 0.00838 0.231 0.235 0.236 0.232 0.232 0.232 0.236 0.231 0.234 0.243166Table A.17: Mean Portfolio Returns for Conditionally Double-Sorted Portfolios (Constituents of S&P500)This table reports the average excess returns for conditionally double-sorted portfolios. In the first step, I sort stocks into three bas-kets based on the control variable indicated in each column. In the second step, I sorts stocks within each of these three baskets intothree additional baskets based on βKML resulting in nine portfolios in total. I then create three portfolios by pooling the stocks inthe βKML-based baskets with the same rank. I report the annualized average value-weighted returns for portfolios with condition-ally low-, medium-, and high-βKML stocks as well as for the High-Low βKML portfolio. Newey and West (1987) standard errors arein parentheses. βKML is estimated from rolling regressions of stock’s monthly excess return on a constant and monthly return of theKML portfolio over the last 60 months. BM is the book-to-market equity ratio, βmarket is the regression coefficient of market excessreturn from rolling time-series regression of firm excess return onto market excess return and a constant, CF denotes cash flow, Levdenotes financial leverage, PCM denotes the price-to-cost margin, Profitability is the ratio of sales minus cost of goods sold, interestexpenses, and selling, general, and administrative expenses to book equity, Size is the natural logarithm of the market capitalizationin thousands, and Turnover is the fraction of shares traded to the total shares outstanding. The first column shows average excessreturn for unconditional sorting based on βKML. The last column reports average excess return for sorting on BM conditional onSize. The sample consists of contemporaneous constituents of the S&P500 excluding financials and utilities. The sample period isfrom 1975 to 2015. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.Uncond BM βmarket CF Lev PCM Profitability Size TurnoverBMcondSize(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Low βKML 10.95 9.48 9.62 9.7 9.39 9.48 9.48 8.35 9.59 6.412 8.9 7.44 7.19 7.89 8.24 7.64 8.08 7.91 8.12 6.75High βKML 6.38 5.93 6.27 5.69 5.59 6.01 5.72 5.67 5.75 8.13High-Low βKML -4.57*** -3.55** -3.35** -4.01** -3.8** -3.47** -3.76** -2.69* -3.85** 1.71(1.71) (1.55) (1.55) (1.62) (1.57) (1.64) (1.53) (1.46) (1.58) (1.68)167Table A.18: Number of Industries in the OES Industry-Specific Surveys, 1988-2016This table reports the number of industries in the OES industry-specific surveys from 1988 to2016. In years 1988-1995, the OES surveys contain only one third of all industries each year.During these years, the industry counts are based on a combination of three subsequent OESindustry-specific surveys. Since the OES survey is not available in year 1996, I complementthe surveys in years 1994 and 1995 with the survey from year 1993. The industries are clas-sified by 3-digit 1987 SIC codes in years 1988-2001, 4-digit 2002 NAICS codes in 2002-2006,4-digit 2007 NAICS codes in 2007-2011 and 4-digit 2012 NAICS codes from 2012. Source:Bureau of Labor Statistics, U.S. Department of Labor, Occupational Employment Statistics.Year1988-19901990-19931993-19951997 1998 1999Number of industries 378 378 378 378 378 378Year 2000 2001 2002 2003 2004 2005Number of industries 378 378 294 294 295 295Year 2006 2007 2008 2009 2010 2011Number of industries 295 295 291 291 291 291Year 2012 2013 2014 2015 2016Number of industries 290 290 290 290 290168Table A.19: The Underlying Occupational Employment Statistics and O*NET File Releasesfor the Construction of the AP measureThe table shows the underlying Occupational Employment Statistics and O*NET filereleases used for constructing the routine task-labor ratio. In years 1988-1995, theEOS surveys contain only one third of all industries each year. In these years, thecross-section of all industries in OES is described by a combination of three subse-quent OES surveys. Since the OES survey is not available in year 1996, I use thecombination of the OES surveys from 1993-1995. Source: Bureau of Labor Statis-tics, U.S. Department of Labor, Occupational Employment Statistics; National Cen-ter for O*NET Development, Database Release Archive, O*NET Resource Center.TimeperiodYear of theEmploymentOccupationStatisticssurveysIndustryclassificationO*NET databaseupdate andrelease dateYear of theO*NET-SOCOccupationalclassification1965-1990 1988-1990 1987 SIC O*NET 98, August 2000 20001991-1993 1991-1993 1987 SIC O*NET 98, August 2000 20001994-1996 1993-1995 1987 SIC O*NET 98, August 2000 20001997 1997 1987 SIC O*NET 98, August 2000 20001998 1998 1987 SIC O*NET 98, August 2000 20001999 1999 1987 SIC O*NET 98, August 2000 20002000 2000 1987 SIC O*NET 98, August 2000 20002001 2001 1987 SIC O*NET 98, August 2000 20002002 2002 2002 NAICS O*NET 5.1, November 2003 20002003 2003 2002 NAICS O*NET 5.1, November 2003 20002004 2004 2002 NAICS O*NET 7.0, December 2004 20002005 2005 2002 NAICS O*NET 9.0, December 2005 20002006 2006 2002 NAICS O*NET 11.0, December 2006 20062007 2007 2002 NAICS O*NET 12.0, June 2007 20062008 2008 2007 NAICS O*NET 13.0, June 2008 20062009 2009 2007 NAICS O*NET 14.0, June 2009 20092010 2010 2007 NAICS O*NET 15.1, February 2011 20092011 2011 2007 NAICS O*NET 16.0, July 2011 20092012 2012 2012 NAICS O*NET 17.0, July 2012 20102013 2013 2012 NAICS O*NET 18.0, July 2013 20102014 2014 2012 NAICS O*NET 19.0, July 2014 20102015 2015 2012 NAICS O*NET 20.1, October 2015 20102016 2016 2012 NAICS O*NET 21.1, November 2016 2010169Table A.20: Correlation between I-shock Measure and the Long-Short (KML) Portfolio (byDifferent Samples and Specifications)This table reports correlation coefficients between the I-shock measure and the annual re-turn of the long-short portfolio (KML). The long-short portfolio has a long position infirms with low automation potential, i.e. the bottom quintile of firms ranked by the APmeasure, and a short position in firms with high automation potential, i.e. the top quin-tile. The KML portfolios in panel A consist of firms from different samples. Firms inindustries producing capital goods are excluded in column 1 and manufacturing firms areexcluded in column 2. The sample consists only of manufacturing firms in column 3.The KML portfolios in panel B are constructed in alternative specifications. Stocks arepre-sorted by operating leverage (OPL) into five baskets and within each of these bas-kets sorted by the AP measure in column 1. Portfolio in column 2 is constructed simi-larly with pre-sorting on Size. Portfolio in columns 3 is constructed by an alternative APmeasure that uses only equipment type of capital in the denominator of the AP formula.(1) (2) (3)Panel A. Correlation between I-shock measure and KML portfolioby different samples of firmsWithoutcapitalproducersWithoutmanufacturingsectorWithinmanufacturingsectorCorrelation 0.595 0.485 0.521Panel B. Correlation between I-shock measure and KML portfolioby alternative specificationsPresorton OPLPresorton Sizeln(LroutineEquipment)Correlation 0.504 0.58 0.595170Table A.21: Firms’ Three-Years Response to Technology ShocksThis table reports results of regressing firm-level percentage changes in employment, capitaland capital-labor ratio in panel A, and sales, percentage point changes in return on capitaland return on equity in Panel B, between years t and t + 3 on the I-shock measure inyear t, Ishockt, firm i′s exposure to technology shock in year t measured by the deciles’index, IβKMLi,t , of cross-sectionally ranked βKMLi,t , and their interaction term, Ishockt×IβKMLi,t .The regressions include aggregate TFP shock, TFPt, and deterministic trend, Trendt, asaggregate control variables and firm i′s capital, Ki,t, age, Agei,t, market capitalization, Sizei,t,lagged dependent variable for one-year horizon t − 1 to t, and firm fixed effects as firm-specific control variables. Columns βKML ≤ median and βKML > median show results ofthe regressions without the interaction term estimated for firms with below and above medianβKML, respectively. Standard errors are clustered at SIC 4 digit industry level and reportedin parentheses. The sample period is from 1975 to 2015. *** Significant at the 1 percentlevel. ** Significant at the 5 percent level. * Significant at the 10 percent level.Panel A. Changes in employment, capital and capital-labor ratioEmployment Capital Capital-labor ratioFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianIshockt -3.406∗∗∗ -2.828∗∗∗ -2.045∗∗∗ 0.329 0.806∗∗∗ 0.845∗∗∗ 4.173∗∗∗ 3.953∗∗∗ 3.126∗∗∗(0.376) (0.280) (0.237) (0.352) (0.248) (0.223) (0.370) (0.299) (0.233)IβKMLi,t -0.085 -0.273 0.181 -0.210 -0.272 -0.173 -0.122 -0.072 -0.428∗(0.122) (0.246) (0.196) (0.138) (0.235) (0.249) (0.104) (0.226) (0.234)Ishockt × IβKMLi,t 0.167∗∗∗ 0.068 -0.130∗∗(0.064) (0.068) (0.058)TFPt 2.251∗∗∗ 2.464∗∗∗ 2.264∗∗∗ 2.042∗∗∗ 2.786∗∗∗ 0.921∗∗∗ -0.739∗∗∗ -0.234 -1.813∗∗∗(0.179) (0.278) (0.284) (0.191) (0.281) (0.280) (0.216) (0.278) (0.293)Ki,t -14.739∗∗∗ -15.076∗∗∗ -14.389∗∗∗ -15.981∗∗∗ -15.992∗∗∗ -16.284∗∗∗ -0.628 -0.065 -1.832∗(0.676) (0.977) (0.887) (0.688) (0.937) (1.140) (0.636) (0.775) (0.957)Agei,t 0.068 -0.275 0.556 0.260 0.308 0.305 0.105 0.344∗∗ -0.252(0.211) (0.337) (0.422) (0.183) (0.286) (0.512) (0.124) (0.166) (0.204)Sizei,t 0.003 -0.024 0.003 0.044 0.049 0.044∗ 0.055∗∗∗ 0.077 0.057∗∗∗(0.026) (0.053) (0.022) (0.028) (0.062) (0.026) (0.016) (0.047) (0.017)Trend 0.431∗∗ 0.662∗ 0.040 -0.256 -0.397 -0.239 -0.776∗∗∗ -1.008∗∗∗ -0.394∗(0.213) (0.339) (0.433) (0.191) (0.284) (0.529) (0.126) (0.162) (0.220)Firm FE Yes Yes Yes Yes Yes Yes Yes Yes YesN 38813 19649 19177 38800 19640 19173 38800 19640 19173R2 0.440 0.520 0.504 0.484 0.585 0.515 0.308 0.395 0.390171Table A.21 Firms’ Three-Years Response to Technology Shocks: ContinuedPanel B. Changes in sales, return on capital, and return on equitySales Return on capital Return on equityFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianIshockt -3.369∗∗∗ -3.325∗∗∗ -2.916∗∗∗ -1.202∗∗∗ -1.110∗∗∗ -0.776∗∗∗ -2.352∗∗∗ -2.139∗∗∗ -1.367∗∗∗(0.479) (0.380) (0.324) (0.108) (0.085) (0.131) (0.264) (0.204) (0.241)IβKMLi,t -0.525∗∗∗ -0.494∗ -0.476 -0.022 -0.097 0.034 -0.074 -0.105 -0.095(0.154) (0.273) (0.312) (0.033) (0.062) (0.081) (0.074) (0.158) (0.182)Ishockt × IβKMLi,t 0.006 0.059∗∗∗ 0.135∗∗∗(0.085) (0.022) (0.050)TFPt 4.214∗∗∗ 4.271∗∗∗ 3.698∗∗∗ 0.671∗∗∗ 0.428∗∗∗ 0.859∗∗∗ 1.353∗∗∗ 1.054∗∗∗ 1.496∗∗∗(0.278) (0.425) (0.325) (0.082) (0.100) (0.115) (0.171) (0.225) (0.218)Ki,t -15.939∗∗∗ -16.130∗∗∗ -15.795∗∗∗ -1.742∗∗∗ -1.939∗∗∗ -1.691∗∗∗ -2.619∗∗∗ -2.545∗∗∗ -2.935∗∗∗(0.749) (1.085) (1.085) (0.184) (0.245) (0.330) (0.451) (0.596) (0.795)Agei,t 0.137 0.034 0.386 0.051 0.073∗∗ 0.013 0.057 -0.075 0.271(0.210) (0.312) (0.429) (0.035) (0.032) (0.033) (0.113) (0.170) (0.277)Sizei,t -0.048 -0.063 -0.047 -0.034∗∗∗ -0.054∗ -0.032∗∗∗ -0.057∗∗∗ -0.172∗∗∗ -0.051∗∗∗(0.035) (0.050) (0.032) (0.007) (0.028) (0.006) (0.013) (0.058) (0.013)Trend -0.031 0.008 -0.281 0.109∗∗∗ 0.131∗∗∗ 0.128∗∗∗ 0.186∗ 0.375∗∗ -0.021(0.212) (0.316) (0.443) (0.035) (0.033) (0.037) (0.112) (0.168) (0.277)Firm FE Yes Yes Yes Yes Yes Yes Yes Yes YesN 38805 19647 19171 38767 19624 19156 35019 17643 17386R2 0.422 0.523 0.470 0.271 0.356 0.356 0.296 0.360 0.393172Table A.22: Firms’ One- and Three-Years Responses in Profit Margins to Technology ShocksThis table reports results of regressing firm-level percentage point changes in profit marginsbetween years t and t + 3 on the I-shock measure in year t, Ishockt, firm i′s exposure totechnology shock in year t measured by the deciles’ index, IβKMLi,t , of cross-sectionally rankedβKMLi,t , and their interaction term, Ishockt× IβKMLi,t . The regressions include aggregate TFPshock, TFPt, and deterministic trend, Trendt, as aggregate control variables and firm i′scapital, Ki,t, age, Agei,t, market capitalization, Sizei,t, lagged dependent variable for one-year horizon t − 1 to t, and firm fixed effects as firm-specific control variables. ColumnsβKML ≤ median and βKML > median show results of the regressions without the interactionterm estimated for firms with below and above median βKML, respectively. Standard errorsare clustered at SIC 4 digit industry level and reported in parentheses. The sample periodis from 1975 to 2015. *** Significant at the 1 percent level. ** Significant at the 5 percentlevel. * Significant at the 10 percent level.Panel A. Profit margins,1-year responsePanel B. Profit margins,3-year responseFullsampleβKML ≤medianβKML >medianFullsampleβKML ≤medianβKML >medianIshockt -0.385∗∗∗ -0.322∗∗∗ -0.208∗∗∗ -0.425∗∗∗ -0.347∗∗∗ -0.259∗∗∗(0.045) (0.031) (0.044) (0.067) (0.051) (0.048)IβKMLi,t 0.007 0.046∗∗ 0.004 -0.011 0.006 -0.077(0.009) (0.021) (0.029) (0.022) (0.044) (0.064)Ishockt × IβKMLi,t 0.027∗∗∗ 0.027∗∗(0.009) (0.012)TFPt 0.041 0.019 0.061∗ -0.008 -0.024 0.031(0.027) (0.032) (0.036) (0.045) (0.059) (0.060)Ki,t -0.374∗∗∗ -0.417∗∗∗ -0.393∗∗∗ -0.644∗∗∗ -0.685∗∗∗ -0.633∗∗∗(0.048) (0.062) (0.090) (0.114) (0.135) (0.218)Agei,t 0.013 0.057∗∗ -0.006 0.008 0.056∗∗ -0.043(0.011) (0.025) (0.011) (0.032) (0.024) (0.028)Sizei,t -0.003∗ -0.004 -0.003∗ -0.012∗∗ -0.018 -0.012∗∗(0.002) (0.005) (0.002) (0.005) (0.016) (0.006)Trend 0.018 -0.025 0.041∗∗∗ 0.056∗ 0.009 0.108∗∗∗(0.011) (0.024) (0.011) (0.031) (0.025) (0.030)Firm FE Yes Yes Yes Yes Yes YesN 43839 22416 21423 38574 19567 19007R2 0.125 0.214 0.174 0.240 0.350 0.308173Appendix BAppendix to Chapter 3B.1 Definition of VariablesTable B.1: Definition of VariablesVariable DefinitionRTI - Routine taskintensityRTI is a measure of the routine task-intensity of individ-ual occupations developed by Autor and Dorn 2013. RTIis based on the routine, manual and abstract task inputsin each occupation as of 1980. RTI of an occupation isincreasing in the occupation’s routine intensity.RS - Share of routineintensive laborShare of routine intensive jobs in a given area is the num-ber jobs in routine intensive occupations divided by thetotal number of jobs in that area. Routine intensive oc-cupations are defined as the occupation in the highestthird by the routine-intensity index weighted by the em-ployment share in 1980.Total wages andsalariesaThe total wages and salaries is the total amount accruedduring a specific year in specific geographical area. Wagesand salaries are measured before deductions and includecommissions, tips, bonuses, voluntary employee contribu-tions to deferred compensation plans and employee gainsfrom exercising stock options.Average wage andsalaryaThe average wage and salary is defined as total wages andsalaries divided by the number of wage and salary jobs.Unemployment rateb Unemployment rate is the annual average rate of unem-ployment in a given area reported by Local Area Unem-ployment statistics by BLS.Wage and salaryemploymentaWage and salary employment are all full-time and part-time jobs in each area by place of work for which wagesand salaries are paid.Job creation ratec, e Job creation is defined as the sum of all jobs cre-ated through expanding establishment including start-ups from year t-1 to t. Job creation rate is defined asthe job creation divided by the average number of jobsin years t-1 and t and expressed in percent. Yeat t endson March 12 and the rate refers to the March-to-Marchcycle.174Table B.1— Definition of Variables: (continued)Variable DefinitionJob destructionrated, eJob destruction is defined as the sum employment lossesfrom contracting establishments and establishment clo-sures. The job destruction rate is defined analogously tojob creation rate.Potential labor force Potential labor force is defined as all persons who workedeither part-time or full-time or during the prior year andpersons who didn’t work but were of age 15+ years in theprior year.Labor force Labor force is defined as all persons who either workedor actively looked for work in the prior year.Working, at leastpart of the yearAll workers who worked at least one week.Not working, lookingfor workPersons who did not work at all but looked for work forat least one week.Unemployed for partof the yearPersons who were unemployed for part of the year andemployed for other part of the year.Unemployed for partor whole yearAll persons who were unemployed for at least part of theyear.a - Growth The growth in the variable is defined as the percentagechange from the previous year.b - Change The change in the unemployment rate is defined as thepercentage points difference from the previous year.c - New firms The job creation rate of new firms is the number of jobscreated by new firms divided by the average number ofjobs in years t and t-1.d - Closures The job destruction rate of closures is the number of jobsdestroyed by firm closures divided by the average numberof jobs in years t and t-1.e - Incumbents The job creation (destruction) rate of incumbents is thenumber of jobs created (destroyed) by incumbents di-vided by the average number of jobs in years t and t-1.175B.2 A Note on Estimation with the Linear Time TrendWhy I include linear trend in the regression: I include a linear trend starting at thebeginning of the time period and increasing by one every year. The IST shock, the changesin the unemployment rate and job creation and destruction rates as well as the growth inwages and salaries are stationary by definition, which is also empirically true over long timehorizons. At the same time, both the IST shock and the changes or growth in the dependentvariables exhibit a very mild and opposite trend over the time period of the approximatelythree decades for which the panel data is available. While IST shocks have a slightly positivetrend starting from 1980, the dependent variables exhibit negative time trends. For example,the growth in nominal wages and salaries was higher in 1980s than in first decade after 2000.A regression specification with the linear time trend allows estimating the impact of therelatively short-term IST shocks on the variables clean of any relatively long-term (negative)co-movement between the IST shock and the dependent variable.176

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}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0379796/manifest

Comment

Related Items