Essays in applied microeconomics

(1)

Tilburg University

Essays in applied microeconomics

Ciccarelli, Nicola

Publication date: 2018

Document Version

Publisher's PDF, also known as Version of record

Link to publication in Tilburg University Research Portal

Citation for published version (APA):

Ciccarelli, N. (2018). Essays in applied microeconomics. CentER, Center for Economic Research.

General rights

Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain

• You may freely distribute the URL identifying the publication in the public portal

Take down policy

If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.

(2)

ESSAYS IN APPLIED MICROECONOMICS

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan Tilburg University op gezag van de rector magnificus, prof. dr. E.H.L. Aarts, in het openbaar te verdedigen ten overstaan van een door het college voor

promoties aangewezen commissie in de Ruth First zaal van de Universiteit op dinsdag 20 februari 2018 om 14.00 uur door

NICOLA CICCARELLI,

geboren te Loreto, Itali¨e.

(3)

PROMOTOR: Prof. dr. A.H.O. van Soest Prof. dr. B. Melenberg

PROMOTIECOMMISSIE: Dr. E. Bonsang

Prof. dr. ir. J.C. van Ours Dr. B.A. Vollaard

(4)

Acknowledgements

First and foremost, I would like to thank my supervisors, Arthur van Soest and Bertrand Melenberg, for their helpful comments and insights that improved my thesis. In addition, I want to thank the doctoral committee members, Eric Bonsang, Jan C. van Ours, and Ben A. Vollaard, for comprehensive and very helpful comments. It is an honor for me to draw from the expertise of such a distinguished group of researchers.

My time in Tilburg would have been less rewarding without Alaa and Sanja with whom I had an amazing time. I also would like to thank the following people: Bas, Cansu, Cynthia, Stefan, Victor, Yilong. I really hope that we will keep our friendship regardless of the distance. Finally, my heartfelt appreciation goes to my family.

Nicola Ciccarelli Tilburg, January 2018

(5)

This dissertation revolves around topics in Labor Economics, Health Economics and Economics of Ageing. In all papers included in this dissertation, we use linear panel data econometric methods that enable the researcher to control for time fixed effects and individual (or county) time-invariant unobserved heterogeneity. As shown in the research papers, individual time-invariant unobserved heterogeneity is often correlated with the explanatory variable and dependent variable that are used in analysis. Therefore, the results contained in this dissertation’s papers are not confounded by time-invariant and unobserved individual characteristics, such as ability or IQ of individuals. Moreover, we make use of instrumental variables to test (and eventually correct) for endogeneity of the main explanatory variable used in each study. The instrumental variables used in this dissertation’s papers are likely to be exogenous, especially since we control for individual (or county) fixed effects and time fixed effects. Furthermore, all research papers included in this dissertation are strongly related to policy making, and we provide relevant (and innovative) guidelines for policymakers. In the first study, we investigate the effects of rising health insurance premiums for employer-provided health insurance on wages, employment, and private health insurance coverage in the U.S. using a county-level panel dataset for the period 2005-2010. In the second study, we analyze the effect of job-related stress on smoking behavior of older U.S. workers using a longitudinal dataset for the period 1992-2010. In the third study, we examine the effects of informal care provision on employment, work hours and health of caregivers using a longitudinal dataset for older European individuals for the period 2004-2013.

(9)

2 Introduction

insurance coverage for counties with a high share of individuals aged 15-24.

In Chapter 3, we use data from the Health and Retirement Study to analyze the effect of job stress on smoking propensity and smoking intensity for U.S. workers. Using dynamic panel data models that pass tests for misspecification, we find that high job stress increases the smoking propensity and smoking intensity of workers. More specifically, we reach several conclusions in terms of smoking propensity and smoking intensity for U.S. workers. First, the effect of job stress on smoking propensity is contemporaneous (short run only). Second, the long run effect of a permanent increase in job stress on smoking intensity is larger than the short run effect of (a temporary increase in) job stress on smoking intensity. The latter result indicates that the effect of job stress on smoking intensity is dynamic rather than discrete. Third, restricting the analysis sample to smokers only, we find that a permanent increase in job stress exerts a statistically significant long run effect on smoking intensity for smokers, and the long run effect is significantly larger than the short run effect for this group of workers. This result indicates that the effect of job stress on smoking intensity of current smokers stretches throughout several years and is dynamic rather than discrete. Finally, in contrast to previous studies, we test for sample selection bias using variable addition tests for sample selection, and we test for exogeneity of job stress using heteroscedasticity-based instruments and respondent i’s firm size as the instrumental variables for job stress. We find that the data used in analysis do not suffer from sample selection bias, and we find that job stress can be treated as an exogenous variable.

In Chapter 4, using panel data on the age group 50-70 in 15 European countries, we analyze the effects of providing informal care to parents, parents-in-law, stepparents, and grandparents on employment, work hours and several measures of physical and mental health of caregivers. Using static and dynamic panel data models that pass tests for misspecification, we find that informal caregiving (and especially daily caregiving) neg-atively affects the labor market attachment and health of caregivers. First, we find a significant and negative effect of daily caregiving on paid work variables (i.e., employ-ment and work hours), and this effect is particularly strong in the case of women. Second, providing care at a weekly (or less than weekly) frequency does not significantly affect employment and work hours. Third, informal caregiving has a substantial negative effect on mental health, especially for individuals that provide daily caregiving. Fourth, the negative effect of informal caregiving on mental health is more pronounced in the case of women. Finally, the effect of informal caregiving on physical health is less substantial but still negative and significant. Since rates of informal care provision, employment rates, and self-reported health vary widely across European regions, we also analyzed whether the effects of informal caregiving on labor market attachment and health of caregivers are heterogeneous across European regions. We did not find evidence of heterogeneous effects of informal caregiving on labor market attachment and health of caregivers across European regions.

(10)

3

rise, we expect that employer health insurance mandates are likely to reduce employment, ceteris paribus, especially for counties with a high share of individuals aged 15-24. Based upon our analysis of the effect of job-related stress on smoking behavior of older U.S. workers, we reach an important conclusion for policymakers. Since stress management programs seem to be effective in reducing job-related stress (see Richardson and Rothstein (2008)), the implementation of stress management programs is likely to exert long-lasting positive effects on smoking behavior for older U.S. workers. Based upon our analysis of the effects of informal care provision on employment, work hours and health measures, we reach an important conclusion for policymakers. Cash benefits that can be used to outsource part of informal care tasks to specialized healthcare personnel are likely to increase the wellbeing of older European individuals.

(11)

CHAPTER 2 The Labor Market Effects of Health Insurance

Premiums: Evidence from Panel Data

Abstract

We analyze the effect of health insurance premiums on wages, employment, and private health insurance coverage in the U.S. using a county-level panel data set for the period 2005-2010. Using variation in medical malpractice payments and variation in medical malpractice legislation to identify the effect of health insurance premiums on the depen-dent variables, we estimate that a 10% increase in health insurance premiums reduces employment by 1.1 percentage points and implies an insignificant decrease of wages and private health insurance coverage. Since employers cannot reduce wages of minimum wage workers when health insurance premiums rise, and since most minimum wage work-ers are aged between 15 and 24, we conjecture that the effects of rising health insurance premiums on employment and private health insurance coverage are stronger for coun-ties with a high share of individuals aged 15-24. In line with our conjecture, we find that increases in health insurance premiums imply a stronger decrease of employment and private health insurance coverage for counties with a high share of individuals aged 15-24.

2.1 Introduction

Rising health insurance premiums for employer-provided health insurance may have sig-nificant labor market effects, including changes in wages, number of jobs, and (private) health insurance coverage. The effect of rising health insurance costs on employment and wages will depend on the elasticity of labor demand and labor supply, on institu-tional constraints on wages and compensation packages (e.g., minimum wage), and on how much workers value the increase in health insurance costs. If workers fully value health insurance benefits provided by the employer, and without institutional constraints on wages and compensation packages (e.g., the minimum wage is not binding for the worker), we expect that increases in the cost of providing health insurance will be fully offset by decreases in wages, with no change in employee utility or employment.1 How-ever, there may be two impediments to full adjustment through wages. First, employees may not fully value increases in the cost of health insurance benefits provided by the

1_{For example, Gruber (1994) shows that the introduction of mandated maternity benefits for women}

(12)

Introduction 5

employer, if rising costs of health insurance benefits are not accompanied by increases in the quality or quantity of benefits. Second, there are several reasons to believe that firms have limited ability to reduce wages in response to increases in the cost of health insur-ance benefits, especially for some groups of workers. For instinsur-ance, employers may not be able (are not able) to reduce wages of near-minimum-wage workers (minimum wage workers). Thus, rising costs of health insurance benefits may not be neutral in terms of employment. Moreover, since employers provide health insurance benefits voluntarily in the period under study (i.e., year 2005 to year 2010), increases in the cost of providing health insurance benefits may imply a reduction of employer-provided health insurance coverage,2 _{especially for workers who do not fully value health insurance benefits and for} workers for whom the minimum wage (or any other institutional constraint) is binding.3 Identifying the effects of rising costs of health benefits on wages, employment and health insurance coverage is challenging. Data on health insurance premiums, wages, employment status, and health insurance coverage are not jointly available at the in-dividual level. Moreover, most micro-data sets containing data on wages, employment status and health insurance coverage, such as the Current Population Survey (CPS), do not enable to control for time-invariant unobserved worker characteristics (e.g., ability) which are likely to be correlated with the health insurance premiums and the outcome variables (see Baicker and Chandra (2006) for more details). In contrast to previous stud-ies, which analyzed the effects of rising health insurance costs on labor market outcomes using repeated cross-sectional data sets (e.g., Cutler and Madrian (1998), Baicker and Chandra (2006)), this paper uses a newly constructed panel data set with data on health insurance premiums for employer-provided health insurance, labor market conditions (i.e., employment rate, annual wages) and private nonelderly health insurance coverage,4 _in order to study the effects of rising health insurance premiums on labor market outcomes and private health insurance coverage for the period 2005-2010. All variables used in analysis vary at the county level, except the health insurance premiums which vary at the state level. In this study we uncover the causal effects of health insurance premiums on labor market outcomes and private health insurance coverage by exploiting two plau-sibly exogenous sources of variation in the cost of providing health insurance: per capita medical malpractice payments, which vary over time and within states, and variation in medical malpractice legislation over time and within states. As will be discussed below, variation in medical malpractice payments affects malpractice insurance premiums and the cost of health insurance. Indeed, since the demand for health services is relatively inelastic (Baicker and Chandra (2005)), rising malpractice insurance premiums will have little effect on net physician compensation, and will be passed along to consumers of

2_{Even though employers are not obliged to offer health insurance coverage during 2005-2010, there is}

an important mechanism that decreases the likelihood that employers will drop health insurance coverage in response to rising health insurance premiums. Indeed, employer (and employee) contributions for employer-provided health insurance are not subjected to taxation in the U.S. (see Gruber (2011)), while wages are taxed both on the side of employers and on the side of employees, thus employers may find convenient to offer health insurance coverage even when premiums are rising.

3_{For instance, Baicker and Chandra (2006) argue that married women do not fully value health}

insurance benefits as they may have access to insurance though their husbands, and find that married women are more likely to lose health insurance coverage as premiums rise.

4_{Private nonelderly health insurance coverage is used as a proxy for employer-provided health}

(13)

6

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data health care in the form of higher health insurance premiums. Variation in medical mal-practice legislation also affects medical malmal-practice insurance premiums and the cost of health insurance. By limiting doctors’ malpractice liability, the implementation of med-ical malpractice tort reforms reduces medmed-ical malpractice insurance premiums, which in turn implies a decrease of health insurance premiums.

Using the two sources of identifying variation outlined above, and controlling for time-invariant unobserved factors, we find that increases in the cost of health insurance imply a significant decrease of the employment rate. Moreover, increases in the cost of health insurance imply an insignificant decrease of wages and private health insurance coverage. Since variation in medical malpractice payments and medical malpractice legislation do not enhance the quantity and quality of the health care provided by employers, employees may not fully value increases in the cost of health benefits that are driven by increases in medical malpractice payments and by variation in medical malpractice legislation. Thus, the final labor market equilibrium will be characterized by lower employment. We find that a 10% increase in the health insurance premiums implies a 1.1-percentage-points decrease of the employment rate, ceteris paribus. Since health insurance premiums increased by 11.1% in real terms between 2005 and 2010, our results indicate that the increasing cost of health insurance had a strong negative effect on employment in the period 2005-2010.

Since many individuals aged 15-24 earn the minimum wage,5 _{and since employers} cannot reduce wages of minimum wage workers when health insurance premiums rise, we conjecture that increases in the cost of health insurance have stronger negative effects on employment and possibly on employer-provided health insurance coverage for counties with a high share of individuals aged 15-24.6 _{To test this conjecture, we included in our} regression model the interaction between the health insurance premiums and the share of 15-24 year old persons living in each county, and we have found that the effect of rising health insurance premiums on the employment rate and private health insurance coverage is stronger for counties with a high share of individuals aged 15-24. These results provide suggestive evidence that the effect of health insurance premiums on the employment rate and private insurance coverage is stronger for individuals for which the minimum wage is binding. Finally, the results obtained in this study have important implications in terms of policy. By increasing the cost of employing workers, the introduction of employer health insurance mandates will reduce employment, especially for counties with a relatively large share of individuals aged 15-24.

This paper is organized as follows. In Section 2.2 we outline a conceptual framework for our analysis, and we summarize previous studies on the effects of health insurance pre-miums on labor market outcomes. Section 2.3 describes the data. Section 2.4 describes the identification strategy and the instrumental variables used for the health insurance premiums. In Section 2.5 we discuss the empirical results for the effects of health in-surance premiums on labor market outcomes and private health inin-surance coverage. In Section 2.6 we analyze whether the effect of increases in health insurance premiums on

5_{Bureau of Labor Statistics (2006): “Minimum wage workers tend to be young. About half of workers}

earning $5.15 [i.e., the minimum wage in 2005] or less were under age 25, and about one-fourth of workers earning at or below the minimum wage were age 16-19.”

6_{Data on share of minimum wage workers for each county is not available. Thus, we cannot directly}

(14)

Background 7

the employment rate and private health insurance coverage are stronger for counties with a high share of individuals aged 15-24. Section 2.7 concludes.

2.2 Background

In this section, we outline a conceptual framework for our analysis, and we summarize previous studies which analyzed the effects of health insurance premiums (or health care costs) on labor market conditions; see Section 2.2.1 and Section 2.2.2, respectively.

2.2.1 Conceptual Framework

The effect of increased health insurance premiums for employer-provided health insurance on wages and employment depends on several factors. Baicker and Chandra (2006) ex-amine the effects of increases in health insurance premiums for employer-provided health insurance on labor market outcomes, highlighting the importance of the employees’ valua-tion of the (employer-provided) health insurance benefit, and the presence of instituvalua-tional constraints that limit the firm’s ability to reduce wages in response to rising health insur-ance premiums. Following their framework, let Ld = fd(W + C) and Ls = fs(W + αC) be the labor demand and labor supply curves, respectively. The variable W represents wages, C represents the premium for the health insurance provided by the employer (which incorporates both the quantity of insurance and the price for that policy), αC represents the employees’ monetary valuation of that health insurance. Using this frame-work, the effect of an increase in the health insurance premium (C) on wages is given by

dW dC = −

ηd_{− αη}s

ηd_{− η}s , (2.1)

where ηd and ηs are the price elasticities of labor demand and labor supply, respectively. If there are no institutional constraints on wages and compensation packages, and if employees fully value the health insurance benefit (α = 1), then wages will decrease by the full cost of the increased health insurance premium. If employees do not value the health insurance benefit (α = 0), then the results are identical to those obtained for the incidence of a payroll tax assessed on employers.7 _{Additionally, the proportional change} in employment will be given by

dL L = η

d(W0− W1− dC) W0

, (2.2)

where W0 and W1 represent the initial and final levels of wages, and dC represents the change in the health insurance premium, which can be caused by a change in the quantity of health insurance or a change in the price of health insurance (i.e., medical inflation).

An increase in the health insurance premium implies an inward shift of the labor demand curve, lowering employment and wages, regardless of the actual value of the preference parameter α. The labor supply response to an increase in the health insurance premium depends on the employees’ valuation of the health insurance benefit (α). If employees do not value the health insurance benefit (α = 0), the labor supply curve will not shift in either direction after an increase in the health insurance premium, and the

(15)

8

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data final equilibrium will be characterized by lower employment and lower wages. If the cost of employer-provided health insurance increases, and if the preference parameter α is positive, workers will value the increase in employer expenditure on the health insurance benefit and will accept (even) lower wages in exchange for higher spending on the health insurance benefit -i.e., the labor supply curve will shift outwards.8

If employees value the health insurance benefit (i.e., α > 0), an increase in the cost of employer-provided health insurance thus implies a reduction in wages and an ambiguous change in employment.9 _{The lower (greater) is the employees’ valuation of the health} insurance benefit (α), the larger (smaller) will be the reduction in employment due to rising health insurance premiums; if workers fully value the health insurance benefit (α = 1), and if there are no institutional constraints limiting the employer’s ability to reduce wages in response to rising premiums, then there will be no reduction in employment.

There are, however, institutional constraints that may limit the firm’s ability to reduce wages for some groups of workers in response to rising health insurance costs. First, we might expect employment losses, rather than a reduction in wages, for workers earning slightly more than the minimum wage. Anyon (2005) reports that 8.9 percent of the U.S. workforce (9.9 million workers) earns the minimum wage in year 2004, and 18 percent of the U.S. workforce earns at most one dollar an hour more than the minimum wage in year 2004. Therefore, an increase in the cost of employer-provided health insurance may affect the employment status of a large share of the U.S. workforce. Second, there are anti-discrimination provisions of labor law, such as Internal Revenue Service non-discriminatory provisions, that restrict the ability of employers to offer different wages to different demographic groups (who may have heterogeneous demands for health in-surance), further limiting the firm’s ability to reduce wages in response to rising health insurance costs. For the latter group of workers, we also expect to see employment losses in response to rising health insurance costs. Third, the presence of labor unions may fur-ther limit the firm’s ability to reduce wages in response to rising premiums, since unions are likely to be opposed to a decrease in wages.

As reported above, the premium (C) incorporates both the quantity of health in-surance (provided by the employer) and the price of that health inin-surance. If increases in health insurance premiums are not caused by increases in the quantity or quality of health insurance – e.g., increases in premiums are caused by the practice of “defensive medicine”10 _{by doctors – we expect that workers will not (fully) value increases in the} health insurance premiums (see Section 2.4.2 for more details).

8_{We implicitly assume that the variation in the cost of health insurance in the case of the nongroup}

health (NH) insurance is similar to the variation in the cost of health insurance for employer-provided health (EH) insurance. Indeed, if the price of EH insurance increases while the price of NH insurance remains constant, workers could purchase health insurance on their own without using extra resources, and workers will not be willing to accept a reduction in wages. Since health insurance premiums depend on healthcare expenditure, which in turn depends on medical technology, aging, and other factors that are likely to be similar between the EH insurance and NH insurance markets, it may be reasonable to assume that variation in the price of health insurance is similar between the two health insurance markets.

9_{In the extreme case that the preference parameter (α) is larger than 1, the labor supply would}

increase to the point of raising total employment. However, the existing literature on the effect of the health insurance premiums on labor market conditions has found that the preference parameter (α) is significantly lower than 1 (see Baicker and Chandra (2006)). Therefore, this extreme case may be practically irrelevant.

(16)

Background 9

2.2.2 Literature Review

Cutler and Madrian (1998) study the effect of health insurance costs on hours worked for the period 1979-1992. The authors use data on (per-worker) spending on health insurance for each U.S. industry, and individual-level data on hours worked from the CPS (Current Population Survey). Using a pooled OLS regression model, and assigning health insurance costs to workers based on their industry, the authors find that rising health insurance costs in the 1980s increased the hours worked of employees covered by employer-provided health insurance by 3 percent. The authors explain this finding in the following way: since health insurance is a fixed cost for the employer, firms face an incentive to substitute hours per worker for employment when health insurance costs rise. Baicker and Chandra (2006) analyze the effect of health insurance premiums on wages, employment, work hours, and the distribution of part-time and full-time work for the period 1996-2002. The authors use state-level data on health insurance premiums for employer-provided health insurance from the Medical Expenditure Panel Survey, and individual-level data on labor market conditions from the CPS. Using a pooled OLS regression model with state fixed effects, and assigning health insurance premiums to workers based on their state of residence and on their family structure,11 the authors find a substantial negative effect of health insurance premiums on labor market outcomes. Indeed, a 10% increase in health insurance premiums reduces the probability of being employed by 1.2 percentage points, reduces hours worked by 2.4%, and increases the likelihood that a worker is employed only part time by 1.9 percentage points. On the other hand, the authors find that premiums do not significantly affect wages. Based on these results, the authors conclude that increases in health insurance premiums are not fully adjusted through (lower) wages.

Sood et al. (2009) analyze the effect of health care costs on employment, gross output and value added of U.S. industries for the period 1987-2005. The authors use data on health care costs for each U.S. industry from the Centers for Medicare and Medicaid Services’ National Health Expenditures Accounts database, and use data on employ-ment, gross output and value added for each U.S. industry from the Bureau of Economic Analysis’ Annual Industry Accounts database. The authors find a substantial negative effect of rising health care costs on industry-level employment, gross output and value added, and this effect is stronger for industries where large percentages of workers receive employer-provided health insurance. Based on these empirical results, the authors con-clude that increases in health care costs may force employers to reduce health benefits, cut employment, and raise prices.

Sommers (2005) hypothesizes that constraints on wage reductions - i.e., downward nominal wage rigidity - could limit the ability of employers to shift the cost of rising health insurance premiums to employees. Based on this hypothesis, Sommers (2005) constructs an economic model to investigate the effects of rising health insurance premiums on labor market conditions under the assumption of downward nominal wage rigidity. In the model by Sommers (2005), when employers face rising health insurance premiums, employers first try to recoup the added premium costs without decreasing nominal wages by allowing inflation to erode the real wage. If this strategy does not recoup all of the increased health insurance costs, employers turn to other strategies, such as dropping

11_{The premium for single (health insurance) coverage is assigned to single respondents, while the}

(17)

10

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data coverage or reducing employment. Sommers (2005) uses data from the CPS to test the predictions of the economic model described above. The author finds that less-educated workers in areas with low inflation – the group with the strongest binding constraint on nominal wage reductions – were more likely to experience unemployment and less likely to experience real wage reductions (than less-educated workers in areas with high inflation). Based on these empirical findings, the author concludes that increases in the cost of health insurance may decrease employment (rather than nominal wages) due to the presence of downward nominal wage rigidity.

2.3 Data

We use a panel of annual variables for counties in the 50 U.S. states plus the District of Columbia. All variables used in this study vary at the county-year level, except the health insurance premiums for employer-provided health insurance. County-level data on health insurance premiums for employer-provided health insurance are not available, and we use state-level data on health insurance premiums (see below). We focus our analysis on data for the period 2005-2010, since data for one dependent variable - private (nonelderly) health insurance coverage - are not available prior to 2005 (see discussion below), and since data for control variables are not available after 2010.12

Health Insurance Premiums

We use annual state-level data on health insurance premiums from the Medical Expen-diture Panel Survey-Insurance Component for 2005 to 2010 (see Online Appendix A.2 for more details on data sources). The main explanatory variable used in this study is the state-level average of annual health insurance premiums for single (health insurance) coverage.13 Data on health insurance premiums for year 2007 are missing for all U.S. states. As in Burtless and Mikusheva (2012), data for the health insurance premiums in year 2007 are imputed as the average of health insurance premiums in year 2006 and 2008 by state (see Online Appendix A.2 for more details on data imputation).

Using the Bureau of Labor Statistics’ CPI inflation calculator, we express the health insurance premiums in constant 2005 U.S. dollars. Moreover, we express the health insurance premiums in 1,000 U.S. dollars. As shown in Panel A of Table 2.1, the average of the health insurance premiums in real terms is 4.040 – i.e., 4,040 U.S. dollars (see Table 2.1, Panel A).

Dependent Variables Labor Market Outcomes

The labor market conditions used in this study are the employment rate and annual wages. Both variables vary across counties and over time. The employment rate is constructed as the ratio between number of employed persons and number of 15-64 year old persons

12_{Data on control variables are mainly obtained from the U.S. Census Bureau’s “USA Counties” data}

set. The U.S. Census Bureau has terminated its support for the “USA Counties” data set, and variables included in the “USA Counties” data set are missing for the years after 2010.

13_{Annual health insurance premiums for single coverage are the sum of the employee and employer}

(18)

Data 11

(by county). Data on number of employed persons for each county is obtained from the Bureau of Labor Statistics’ QCEW (Quarterly Census of Employment and Wages) data set. Data on number of 15-64 year old persons living in each county is obtained from the U.S. Census Bureau’s “USA Counties” data set (see Online Appendix A.3 for more details). The QCEW data set does not include self-employed workers in the number of employed persons - only employees are included in the number of employed persons in the QCEW data set - thus the average of the employment rate used in this study is lower than the average of the employment rate found in official statistics for the United States.14 _{The average of the employment rate is 0.51 (see Table 2.1, Panel A).}

County-level data on annual wages are obtained from the Bureau of Labor Statistics’ QCEW (Quarterly Census of Employment and Wages) data set (see Online Appendix A.3 for more details on data sources). In the QCEW dataset, annual wages are computed as the ratio between total annual wages and total annual employment (by county). Using the Bureau of Labor Statistics’ CPI inflation calculator, we express annual wages in constant 2005 U.S. dollars. The average of annual wages in real terms is almost 30,000 U.S. dollars for the year 2005 to 2010 (see Table 2.1, Panel A).

Private Nonelderly Health Insurance Coverage

County-level estimates for private nonelderly health insurance coverage are not available. Using county-level data on total nonelderly health insurance coverage, and county-level data on public nonelderly health insurance coverage, we computed estimates of private nonelderly health insurance coverage as the difference between the above-mentioned vari-ables. Data on total nonelderly health insurance coverage are obtained from the SAHIE (Small Area Health Insurance Estimates) data set, and data on public nonelderly health insurance coverage are obtained from the AHRF (Area Health Resource File) data set (see Online Appendix A.3 for more details on data sources). Our estimates for private (nonelderly) health insurance coverage represent the county-level share of nonelderly (0-64) individuals that are covered by either employer-provided health insurance or nongroup health insurance. Even though our estimates for private (nonelderly) health insurance coverage are not identical to employer-provided health insurance coverage (see above), it is worth mentioning that the vast majority of privately insured individuals are covered by employer-provided health insurance (see DeNavas-Walt et al. (2008)).15 _{Approximately,} 62.1% of nonelderly individuals are covered by private health insurance in the period 2005-2010 (see Table 2.1, Panel A).16

Control Variables

We control for an extensive set of variables that may be correlated to the dependent variables and the health insurance premiums. All control variables used in this study vary at the county-year level. Summary statistics for control variables are reported in Table 2.1 (Panel B).

14_{Retzer et al. (2013), p. 169: “The QCEW does not include self-employed workers, nor does it provide}

estimates by age, race or gender.”

15_{DeNavas-Walt et al. (2008) reports that approximately 88% of privately insured individuals are}

covered by employer-provided health insurance. Only 12% of privately insured individuals are covered by nongroup health insurance (DeNavas-Walt et al. (2008)).

16_{Summary statistics for total and public (nonelderly) health insurance coverage are reported in Table}

(19)

12

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data We control for socio-economic variables. Data on socio-economic variables are ob-tained from the US Census Bureau’s “USA counties” data set (see Online Appendix A.2 for more details on data sources). Socio-economic control variables include: (1) share of males; (2) three variables for the age distribution (i.e., share of 15-19 year old persons, share of 20-24 year old persons, share of 65+ year old persons); (3) population density (i.e., population per square mile); (4) eight variables for the racial (and Hispanic origin) composition of the county population. Individuals can be (a) White, (b) Black, (c) Asian, (d) of another race; moreover, individuals of any race can be either (I) Hispanic, or (II) non-Hispanic. We control for eight variables indicating the share of individuals belonging to a specific combination of race and Hispanic origin; the baseline variable is (share of) non-Hispanic Whites.

We control for share of small firms. The variable “share of small firms” is defined as the ratio between the number of small firms and total number of firms (by county), where a firm is defined as a small firm if it employs less than 500 employees. The variable “share of small firms” is computed using data from the U.S. Small Business Administration (see Online Appendix A.2 for more details on data sources). Since small (big) firms have low (high) bargaining power in extracting price concessions from health insurers (Stevens (2015), p. 74), it is likely that the variable “share of small firms” is correlated with the health insurance premiums.17

Panel Structure

The longitudinal dataset is almost balanced, and we use 15,366 observations in analysis (see Table A.2). For each year, our longitudinal dataset contains data on slightly more than 3,040 counties. This represents almost 97% of the total number of U.S. counties.18

17_{See also Lee (2002).}

(20)

Empirical Strategy 13

Table 2.1: Summary Statistics

PANEL A. Health insurance premiums and dependent variables

(1) (2) (3)

Observations Mean Std. Dev. Health insurance premium in 1,000 U.S. dollars† ₁₅₃₆₆ _4.040 _0.300

Employment rate†† ₁₅₃₆₆ _0.514 _0.204

Annual wages in U.S. dollars§ 15366 29781.62 6645.618 Private nonelderly health insurance coverage (share) 15366 0.621 0.117

PANEL B. Control variables

(1) (2) (3)

Observations Mean Std. Dev. Socio-Economic Controls Male (share) 15366 0.499 0.022 Age distribution Age (15-19) (share) 15366 0.072 0.011 Age (20-24) (share) 15366 0.062 0.025 Age (65+) (share) 15366 0.154 0.041

Population density (pop. per square mile) 15366 235.939 1702.654 Race/Hispanic origin

White non-Hispanic (share) 15366 0.791 0.196

White Hispanic (share) 15366 0.072 0.125

Black non-Hispanic (share) 15366 0.087 0.143

Black Hispanic (share) 15366 0.002 0.004

Asian non-Hispanic (share) 15366 0.011 0.023

Asian Hispanic (share) 15366 0.000 0.001

Other non-Hispanic (share) 15366 0.032 0.081

Other Hispanic (share) 15366 0.005 0.006

Share of small firms 15366 0.915 0.036

Notes: see Online Appendix A.2-A.3 for data sources for each variable used in this study.

†_{The health insurance premiums are expressed in real terms: we discount the nominal health insurance}

premiums back to year 2005 using the Bureau of Labor Statistics’ CPI inflation calculator, and we express the health insurance premiums in 1,000 U.S. dollars.

††_{The employment rate is constructed as the ratio between number of employed persons (QCEW data}

set) and number of 15-64 year old persons (“USA counties” data set). The average of the employment rate used in this study is lower than the average of the employment rate found in official statistics for the United States, since self-employed workers are not considered as employed in the QCEW data set.

§_{Annual wages are expressed in real terms: we discount nominal wages back to year 2005 using the}

Bureau of Labor Statistics’ CPI inflation calculator.

2.4 Empirical Strategy

(21)

14

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data

2.4.1 Methodology

As reported in Section 2.3, all dependent and explanatory variables used in this study vary at the county level and over time, except the health insurance premiums which vary at the state level and over time (see Section 2.3). Consider the following panel data model for county c (and state s) in year t:

yct = β1premiumsst+ controls>ctβ2+ Tt+ vct, where vct = αc+ ct. (2.3) Here, yct is the labor market outcome of interest (wages, employment rate) or private health insurance coverage. The expression controlsct measures county-level covariates, including socio-economic controls and share of small firms. The expression Tt represents year fixed effects, and vct is a composite error term which is composed by time-invariant county-level unobserved heterogeneity (αc) and a county-specific idiosyncratic error (ct). The expression premiumsst measures the health insurance premium in state s and year t.

As a first step, we use pooled ordinary least squares (OLS) to study the association between the health insurance premiums and the dependent variables. In the case of pooled OLS estimation, time-invariant county-level unobserved heterogeneity (αc) is treated as part of the error term, and is not eliminated from the regression model – i.e., the error term is the composite error term (vct) for pooled OLS estimation. Since residuals in eq. (3.2) may be heteroscedastic or may be clustered, we use cluster-robust standard errors, and the standard errors are clustered at the state level. As time-invariant and unobserved county characteristics might be correlated with the dependent variable and the health insurance premiums, the OLS estimates cannot be interpreted as causal.

Time-invariant unobserved heterogeneity can be eliminated from the panel data model (3.2) using within-group transformation or first-difference transformation of the data. In this study we rely on first-difference transformation, since the idiosyncratic error term in levels (ct) is positively serially correlated for most dependent variables used in analysis (i.e., annual wages, private health insurance coverage).19,20 The panel data model in first differences for county c (and state s) in year t takes the following form:

∆yct = β1∆premiumsst + ∆controlsct>β2+ ∆Tt+ ∆ct, (2.4) and the FD estimator is the OLS estimator applied to equation (2.4). Since residuals in eq. (2.4) may be heteroscedastic or may be clustered, we use cluster-robust standard errors, and the standard errors are clustered at the state level. Since the estimate of the main parameter of interest (β1) may be biased due to endogeneity problems (see Section 2.4.2 for more details), we will also analyze the effect of health insurance premiums on labor market conditions and private health insurance coverage using instrumental variables for the health insurance premiums.

2.4.2 Identification Strategy

Using state-level premiums as a proxy for county-level premiums may introduce a bias in the estimated coefficient on the health insurance premiums. The true (and unknown)

19_{If the idiosyncratic error term in levels (}

ct) is serially correlated, then the FD (first difference)

estimator is more efficient than the WG (within group) estimator (Wooldridge (2010), Section 10.7).

(22)

county-level premium can be thought of as the sum of the state-level premium and an error term which varies at the county level:

premiumsct = premiumsst+ mct, (2.5)

where s indicates the state, c indicates the county, and t indicates the year. Here, mct represents an error term that is idiosyncratic to county c in year t. It is probably the case that Cov((αc+ ct), mct) 6= 0 because counties with highly productive workers are also likely to be counties with the highest premiums (see Baicker and Chandra (2006), p. 616).21 _{Our main identification strategy estimates the parameters of interest using} esti-mation in first differences (see eq. (2.4)). Even though first-difference transforesti-mation of the data eliminates time-invariant heterogeneity (αc) from our regression model, reducing the likelihood that the measurement error term will be correlated with the error term of the structural equation, it may still be possible that the error term in differences (∆ct) is correlated with measurement error in differences (∆mct) – i.e., Cov(∆ct, ∆mct) 6= 0. An additional source of endogeneity may affect the estimate of the coefficient on health insurance premiums. Even though we control for several variables that may be correlated with the dependent variable and the health insurance premiums, a bias in the main co-efficient of interest may arise due to omitted (time-varying) variables that are correlated with the dependent variable and the health insurance premiums.

A solution to the problems reported above is to instrument for health insurance premi-ums using variables that are uncorrelated with ∆ct and ∆mctbut are correlated with the health insurance premiums. In our analysis we use state-level per capita medical malprac-tice payments as an instrument for premiums, and we use variation in medical malpracmalprac-tice legislation to generate two additional instrumental variables for premiums. Instrumental variables must satisfy two conditions. First, instrumental variables must affect the health insurance premiums. Second, it must also be the case that the instrumental variables are not correlated with measurement error and with omitted (time-varying) variables that are correlated with the dependent variable and premiums. In the next subsection, we explore the validity of these assumptions.

The Instrumental Variables

Per Capita Medical Malpractice Payments

Rising medical malpractice payments imply an increase of medical malpractice insurance premiums; indeed, insurer losses from increases in medical malpractice payments are thought to be the primary contributor to the growth of medical malpractice insurance premiums (see page 617 of Baicker and Chandra (2006) and references therein). More-over, since the demand for health services is relatively inelastic (Baicker and Chandra (2005)), rising medical malpractice insurance premiums will have little to no effect on net physician compensation, and will be passed along to consumers of health care in the form of higher health care prices. This in turn will lead to increases in health insur-ance premiums. Based on the previous arguments, it is likely that medical malpractice payments significantly affect the health insurance premiums. Finally, since increases in medical malpractice payments are not likely to increase the quantity or quality of health care, but are likely to increase the price of health care provided by employers, workers

21_{Baicker and Chandra (2006), p. 616: “More productive workers who can command large}

(23)

16

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data may not fully value increases in health insurance premiums that are caused by increases in medical malpractice payments.

We use data on per capita medical malpractice payments from the National Practi-tioner Data Bank (NPDB) for the period 2005-2010 to instrument the health insurance premiums (see Online Appendix A.5 for more details on data sources). Per capita medi-cal malpractice payments are the ratio between total malpractice payments in each U.S. state and the state population. Moreover, per capita medical malpractice payments are expressed in real terms – i.e., we express per capita payments in constant 2005 U.S. dollars. Summary statistics for medical malpractice payments are reported in Table A1. The variability of payments over time and within states is used to identify the effect of the health insurance premiums on the dependent variables. In line with the previous discussion, we find that medical malpractice payments are significantly correlated to the health insurance premiums at the 1% significance level, controlling for observable covari-ates and time fixed effects; see the first-stage estimate in column 1 of Table 2.2.22 The first-stage estimate suggest that when per capita malpractice payments double, health insurance premiums increase by 2.7%. This result is fairly similar to previous estimates which have found that a 100% increase of per capita malpractice payments leads to a 1%-2% increase of the health insurance premiums (see Baicker and Chandra (2006), pp. 622-624).

The use of per capita medical malpractice payments as the instrument for health insurance premiums is likely to remove the bias from any residual correlation between ∆ctand ∆premiumsst. Indeed, the instrument only picks up that part of the within-state variation in premiums that is attributable to the within-state variation in the malpractice climate. This implies that the instrumental variable “per capita medical malpractice payments” in differences (∆paymentsst) is unlikely to be correlated with measurement error in differences (∆mct). Moreover, as the dependent variables used in this study vary at the county-year level, it is also likely that time-varying omitted factors that affect the dependent variable also vary within counties and over time. Since our instrument only picks up that part of the within-state variation in premiums that is attributable to the within-state variation in the malpractice climate, it is likely that the instrument is uncorrelated with omitted factors that affect the dependent variable. Lastly, it may be tempting to argue that the correlation between premiums and malpractice payments is spurious because states with persistently high healthcare expenditure (and thus high premiums) may also be characterized by persistently high medical malpractice payments. However, since we control for time-invariant unobserved factors that may be correlated with the outcome variables and the premiums (see eq. (2.4)), this concern is not likely to be relevant in the case of this study.

Variation in Medical Malpractice Legislation

Medical malpractice tort reforms tend to decrease medical malpractice payments by lim-iting medical malpractice liability of doctors. Indeed, medical malpractice tort reforms usually set caps on the awards that the victim of malpractice can obtain from the negligent doctor (see Table A2 for an example of tort reform). Moreover, and more importantly, Kessler and McClellan (1996) show that medical malpractice tort reforms decrease the

22_{The regression results reported in column 1 of Table 2.2 are the estimates of equation (2.4), with}

(24)

likelihood that doctors practice “defensive medicine” – i.e., to perform excessive tests and procedures because of concerns about malpractice liability. Since the demand for health care is relatively inelastic (Baicker and Chandra (2005)), the decreased costs of medical malpractice (caused by the implementation of malpractice tort reforms) will be passed along to consumers of health care in the form of lower health care prices and potentially in the form of lower health insurance premiums, and will not benefit physicians in the form of higher compensation. In line with this theoretical argument, Avraham et al. (2012) find that the implementation of medical malpractice tort reforms reduces health insurance premiums for employer-provided health insurance.

Since state legislators are responsible for medical malpractice legislation in the U.S., we use variation in medical malpractice legislation at the state level (over time) to generate instrumental variables for the health insurance premiums. Several states enacted a cap on non-economic damage awards since 2005 (see Table A3);23 for example, a cap on non-economic damage awards was introduced in Georgia in year 2005, limiting non-non-economic damage awards to 350,000 U.S. dollars per healthcare provider. Non-economic damage awards are awards that the victim of malpractice can obtain from the negligent doctor for pain and suffering caused by medical malpractice. Based on the theoretical discussion outlined above, states that enacted caps on non-economic damage awards may also have experienced a decrease of health care prices and potentially a decrease of health insurance premiums over time. Thus, we construct an instrumental variable for premiums using the following procedure. First, we generate a dummy variable (caps on damagesst) which is equal to 1 for state s in the year of the reform and in successive years, 0 otherwise.24 Second, as health insurance contracts are typically finalized 3+ months prior to the calendar year in which they take effect (Avraham et al. (2012), p. 13), we use the lag of the dummy variable mentioned above (i.e., caps on damagess,t−1) as an instrumental variable for the health insurance premiums.25

Medical malpractice tort laws can also be repealed by state supreme courts if a medical malpractice tort law is incompatible with the state constitution, the federal constitution of the U.S., or with state or federal laws. Repeals of caps on non-economic damage awards may lead to higher health insurance premiums. Indeed, if a cap on non-economic damage awards is repealed in year t, we expect that non-economic damage awards paid to victims of malpractice will increase from year t onwards. In turn, this is likely to increase the price of health care, and possibly will increase the health insurance premiums. We construct an additional instrumental variable for the health insurance premiums using repeals of caps on economic damage awards. One state repealed the cap on non-economic damage awards during 2005-2010 (see Table A3). First, we generated a dummy variable (repealsst) which takes value 1 for state s in the year of the repeal of caps on non-economic damage awards and in successive years, 0 otherwise. Second, as health insurance contracts are typically finalized 3+ months prior to the calendar year in which they take effect, we use the lag of the dummy variable mentioned above (i.e., repealss,t−1) as an instrumental variable for the health insurance premiums.

In line with our previous discussion, we find that enactments of caps on non-economic damages are negatively (but insignificantly) correlated to the health insurance

premi-23_{See Online Appendix A.5 for details regarding data sources for medical malpractice tort reforms.} 24_{For example, caps on damages}

s,tis equal to 1 for Alaska in 2006-2010, and it is equal to 0 for Alaska

in 2005 (see Table A3).

(25)

18

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data ums, and we find that repeals of caps on non-economic damage awards are positively and significantly correlated to the health insurance premiums, controlling for observable covariates, individual fixed effects and time fixed effects; see the regression results in col-umn 1 of Table A.3.26 _{We find that repeals of caps on non-economic damage awards lead} to a 2.7% increase of the health insurance premiums. This result is relatively similar to the result reported in Avraham et al. (2012), as these authors find that repeals of tort reforms lead to a 2.1% increase of the health insurance premiums.

It seems plausible to assume that variation in medical malpractice tort legislation is exogenous with respect to the error term on labor market conditions and private health insurance coverage. Broadly speaking, instrumenting for premiums using variation in medical malpractice legislation at the state level is likely to remove the correlation between the error term of the structural equation (∆ct) and premiums (∆premiumsst). Indeed, the two instruments described above only pick up that part of the within-state variation in premiums that is attributable to within-state variation in the malpractice climate. More in detail, there are several reasons to believe that the instruments described above satisfy exogeneity requirements. First, the passage of tort reforms is a political process whose outcome depends on political and ideological factors that are unlikely to be related to labor market conditions and private health insurance coverage. The main opposition to the implementation of tort reforms is based on the argument that doctors will operate less carefully after the implementation of tort reforms (i.e., moral hazard). Moreover, it may be argued that states enacting tort reforms may attract the riskiest doctors from other U.S. states (e.g., see Seabury (2009)). Finally, victims of malpractice are likely to be opposed to the introduction of medical malpractice tort reforms, since tort reforms tend to reduce the rights of victims of malpractice. Therefore, political and ideological factors that are unrelated to labor market conditions and private health insurance coverage are the primary factors that determined the timing of the implementation of tort reforms. Second, since the procedure involved in passing tort reforms is long and complicated, and may be subjected to delays, we find it hard to believe that the timing of implementation of tort reforms is correlated to (the timing of) economic shocks that may affect the premiums and the dependent variable.

First-stage Estimates

Our preferred instrumental variable for the health insurance premiums is per capita med-ical malpractice payments, since medmed-ical malpractice payments vary over time and be-tween all U.S. states, while the instruments based on variation in medical malpractice legislation affect only few U.S. states. Therefore, in the following discussion we will an-alyze the effect of health insurance premiums on employment, wages and private health insurance coverage using two sets of instrumental variables: (a) per capita medical mal-practice payments (alone); (b) all three instrumental variables described above. The first-stage estimates for the two sets of instrumental variables are reported in column 1 and column 2 of Table 2.2, respectively.

The first-stage F -statistic is 7.966 in the case that we exclusively use per capita med-ical malpractice payments as the instrument for premiums.27 _{Since we were concerned}

26_{The regression results reported in column 1 of Table A.3 are the estimates of equation (2.4), with}

the health insurance premiums on the left hand side of equation (2.4), and with caps on damagess,t−1

and repealss,t−1 on the right hand side of equation (2.4).

(26)

Empirical Results 19

that the first-stage F -statistic of 7.966 could be indicative of a problem of weak instru-ments, following Stock et al. (2002) we also estimated the regression models using LIML (limited information maximum likelihood) methods. The second-stage LIML estimates were approximately identical to the standard TSLS (two-stage least squares) estimates.28 The first-stage F -statistic is larger than 10 (F -statistic=21.485) in the case that we use the full set of instrumental variables for the premiums, indicating that instruments have strong predictive power to predict the health insurance premiums in the first-stage.

Table 2.2: First-stage FDIV Estimates

(1) (2)

FDIV (first-stage) FDIV (first-stage) Dependent variable

HI Premium (in 1,000 $) HI Premium (in 1,000 $)

Medical malpractice payments 0.011*** 0.011***

(0.0038) (0.0038)

Caps on damages (enactments) - 0.015

- (0.048)

Caps on damages (repeals) - 0.125***

- (0.016)

F-statistic on the excluded instrument(s) 7.966 21.485

P-value 0.0069 0.0000

Observations 15,366 15,366

Adjusted R-squared 0.399 0.401

Year fixed effects? YES YES

County fixed effects?† YES YES

Controls?†† ALL ALL

Notes: *** p < 0.01, ** p < 0.05, * p < 0.10. Clustered (and robust) standard errors at the state level are reported in parenthesis. The outcome variable is health insurance premiums. Year FE (fixed effects) are included in the regression model. The first-stage (FDIV) estimates are obtained using the StataTM _{command “ivreg2, cluster(state identifier)” using the model in first differences.}

†_{County FE (fixed effects) are eliminated from the regression model using first-difference}

transformation of the regression model.

††_{Control variables include eight variables for race and Hispanic origin, share of males, three variables}

for the population distribution by age, population density, share of small firms.

2.5 Empirical Results

We report and discuss the empirical results for the effect of health insurance premiums on labor market outcomes and private health insurance coverage in Section 2.5.1 and Section 2.5.2, respectively. In Section 2.5.3 we explore the robustness of our identification strategy.

(F-statistic=623.390).

(27)

20

The Labor Market Effects of Health Insurance Premiums: Evidence from Panel Data

2.5.1 The Causal Effects of Health Insurance Premiums on

La-bor Market Outcomes

We begin with an examination of the effect of increases in health insurance premiums on the employment rate. Column 1 of Table 2.3 shows the results of the OLS estimation of equation (3.2). This regression includes year fixed effects and the county-level controls outlined above. Standard errors are clustered at the state level. Premiums do not sig-nificantly affect the employment rate, and the coefficient on premiums is positive. The OLS estimate cannot be interpreted as causal, since premiums and the employment rate may be correlated with unobserved county characteristics that do not vary over time.29 Once we additionally control for time-invariant unobserved county characteristics using the FD estimator (see eq. (2.4)), we find that increases in health insurance premiums lead to a statistically significant decrease of the employment rate (see column 2 of Table 2.3). The FD estimate suggests that a 10% increase in premiums leads to a 0.5% decrease (0.2-percentage-points decrease) in the employment rate, with an associated elasticity of -0.047. As discussed above, however, this estimate may be biased due to measurement errors in premiums (see Section 2.4.2). Moreover, positive (or negative) economic shocks that are not captured by our regression model may lead to upward bias in the coefficient on premiums. As a matter of fact, positive economic shocks are likely to increase the employment rate, and may increase spending on health care (e.g., see Acemoglu et al. (2013)) and health insurance premiums.

We use medical malpractice payments to instrument for health insurance premiums. Second-stage estimates from two-stage least squares estimation of equation (2.4) are shown in column 3 of Table 2.3.30 _{Even though we do not reject the null hypothesis} of exogeneity of premiums (Hausman test’s p-value=0.11), we find that the coefficient on premiums is more negative than in column 2. This result is in line with the argument that measurement error may lead to a bias towards zero in the coefficient on premiums. Moreover, local economic shocks (or other time-varying unobserved factors) may lead to an upward bias for the coefficient on premiums reported in column 2. The second-stage coefficient on premiums is statistically significant, and indicates that a 10% increase in premiums leads to a 2.2% decrease (1.1-percentage-points decrease) in the employment rate, with an associated elasticity of -0.220. This result is very similar to the results re-ported in a previous study (see Baicker and Chandra (2006)). We can use our estimates to study the economy-wide impact of the growth of health insurance premiums between 2005 and 2010. Since premiums increased by 11.1% in real terms between 2005 and 2010,31 _{rising premiums decreased the employment rate by 2.4% (1.2-percentage-points)} between 2005 and 2010, ceteris paribus.

Second-stage FDIV estimates using all three instrumental variables described in

Sec-29_{Since a direct comparison of OLS and FD (first differences) estimates could not be implemented,}

we compared OLS and FE (fixed effects) estimates, and we have found that OLS and FE estimates are systematically different. Moreover, we have found that RE (random effects) and FE estimates are systematically different. Thus, time-invariant county characteristics seem to be correlated with the explanatory variables and the employment rate.

30_{The corresponding first-stage (FDIV) estimates are reported in column 1 of Table 2.2.}

31_{Based on data for premiums in real terms (i.e., premiums in 2005 U.S. dollars), we find that the}

(28)

Empirical Results 21

tion 2.4.2 are reported in column 4 of Table 2.3.32 _{The second-stage coefficient on} pre-miums using three instrumental variables is approximately identical to the coefficient reported in column 3, in which case we use medical malpractice payments as the only excluded instrument. It is worth mentioning that the over-identifying restrictions are not rejected at the 5% level (Hansen test’s p-value=0.58), indicating that instruments are likely to be exogenous.

We also investigated the effect of health insurance premiums on wages. We report the OLS estimate for the effect of health insurance premiums on annual wages in column 5 of Table 2.3. The OLS estimate suggests that rising premiums lead to a statistically significant increase of wages (see column 5). This OLS result is likely to be (upward) biased by omitted time-invariant county characteristics. In fact, counties with highly productive workers are also likely to be characterized by high spending on health care and by high premiums (Baicker and Chandra (2006)).33 Controlling for time-invariant unobserved heterogeneity through the use of the FD estimator reverses the sign of the coefficient on premiums, and makes the coefficient statistically insignificant (see column 6 of Table 2.3).

Second-stage FDIV estimates for the effect of health insurance premiums on wages are reported in column 7-8 of Table 2.3.34 In the case of column 7, we exclusively use medical malpractice payments to instrument for the health insurance premiums. In the case of column 8, we use the full set of instrumental variables for the health insurance premiums. The second-stage coefficients on health insurance premiums using instrumental variables methods are more negative than the coefficient reported in column 6, in which case we assume that premiums are exogenous. This result is in line with our previous discussion on potential sources of bias that may affect the coefficient on premiums (see discussion above for the employment rate). As shown in column 7 and 8 of Table 2.3, we find that the coefficients on premiums are statistically insignificant at the 5% level, and over-identifying restrictions are not rejected at the 5% level (Hansen J test’s p-value=0.489), indicating that instruments are likely to be exogenous. Finally, based on the magnitude of the estimated coefficients on premiums reported in Columns 7-8 of Table 2.3, we find no evidence of full shifting of the increased price of health insurance onto wages.35

To conclude, the results for the effect of health insurance premiums on the employment rate and wages are consistent with the predictions of a model in which workers do not fully value health benefits or in which firms are constrained in their ability to reduce wages.36

32_{The corresponding first-stage (FDIV) estimates are reported in column 2 of Table 2.2.}

33_{Using the Hausman test, we compared OLS and FE (fixed effects) estimates, and we compared}

RE (random effects) and FE estimates. We have found that estimates were systematically different in both cases, indicating that time-invariant county characteristics are likely to be correlated with the health insurance premiums and the dependent variable. The Hausman test results are available from the author.

34_{The first-stage estimates corresponding to column 7 of Table 2.3 are reported in column 1 of Table}

2.2. The first-stage estimates corresponding to column 8 of Table 2.3 are reported in column 2 of Table 2.2.

35_{If increases in premiums are fully shifted onto wages, then the coefficient on premiums must be equal}

to -1000. This is due to the fact that premiums are expressed in 1,000 U.S. dollars, while annual wages are expressed in U.S. dollars.

(29)

22

Essays in applied microeconomics

ESSAYS IN APPLIED MICROECONOMICS

Acknowledgements

Contents

CHAPTER 1

Introduction

CHAPTER 2

The Labor Market Effects of Health Insurance

Premiums: Evidence from Panel Data

2.1

Introduction

2.2

Background

2.2.1

Conceptual Framework

2.2.2

Literature Review

2.3

Data

2.4

Empirical Strategy

2.4.1

Methodology

2.4.2

Identification Strategy

2.5

Empirical Results

2.5.1

The Causal Effects of Health Insurance Premiums on

La-bor Market Outcomes