Modeling the Labor Market Effects of the Affordable Care Act

Transcription

1 Modeling the Labor Market Effects of the Affordable Care Act Danielle Dobos 1 Department of Economics Stanford University, Stanford CA ddobos@stanford.edu Under the direction of Prof. Jayanta Bhattacharya May 5, 2016 Abstract This paper develops a simulation model to predict the labor market effects of the 2010 Affordable Care Act (ACA). Two provisions of the ACA the expansion of state Medicaid programs and the creation of public health insurance exchanges are likely to have significant effects on individual labor supply. This paper incorporates these new provisions into the budget constraint of the individual labor-leisure choice model in order to predict optimal work hours in response to the ACA. Using Central Population Survey (CPS) and premiums data to empirically estimate model parameters, I predict an -11.8% decrease in work hours after the ACA, and bunching before Medicaid and exchange subsidy eligibility thresholds. I also predict heterogeneous results for different demographic groups, with retirement age workers predicted to exhibit the largest decrease in work hours, and low-income workers previously eligible for Medicaid predicted to increase work hours. The model serves as a base for future policy simulation and complements a growing empirical literature base on the Affordable Care Act. Keywords: Affordable Care Act, Medicaid notch, simulation model, labor-leisure tradeoff, exchange subsidies 1 I am grateful to Professor Jayanta Bhattacharya for his extensive guidance, expertise, and generosity with his time as my honors thesis advisor. I am also grateful to Marcelo Clerici-Arias, director of the undergraduate honors program, for his support throughout the writing process.

2 1. Introduction The 2010 Affordable Care Act is the largest reform to the United States healthcare system since the implementation of Medicare and Medicaid in A collection of several major provisions, the main objective of the ACA is to extend coverage to the 47 million Americans who lacked health insurance prior to It broadly accomplishes these goals with a two-pronged approach: the legislation first imposes an individual mandate to carry some form of health insurance, and then provides an expanded set of options for individuals to obtain coverage. For the majority of Americans who receive health insurance through an employer, the ACA has very little effect, but many of its reforms did expand coverage significantly for the uninsured. The ACA allowed young adults to remain on their parents policies up to age 26 regardless of student status, provided states with the option to expand Medicaid to all individuals with incomes up to 138% of the Federal Poverty Line (FPL), and created a series of publicly regulated exchanges where individuals can purchase private insurance plans at a subsidized cost. Combined, these provisions are estimated to have reduced the uninsured rate by approximately 6% since 2013 (Witters, 2015). Despite these benefits, economic theory predicts that provisions of the ACA will impact United States labor markets. The primary rationale for these concerns is the expansion of state Medicaid programs to individuals with incomes below 138% of the Federal Poverty Line, as well as the creation of subsidized public health insurance exchanges. Theory predicts that meanstested programs in general reduce work incentives as workers reduce hours to stay below income eligibility thresholds, and there is a substantial literature base exploring this effect in relation to Medicaid. Yelowitz (1995) formalizes the impact of Medicaid by incorporating a Medicaid notch into the individual labor supply budget constraint. The Medicaid Notch model predicts DOBOS 2

3 that most individuals will choose not to work past the eligibility threshold for Medicaid and lose health insurance benefits, and some empirical studies corroborate this hypothesis (Yelowitz 1995; Dague, DeLeire, and Leininger 2013; Garthwaite 2014). The most recent empirical studies take advantage of abrupt policy changes in individual state Medicaid programs usually a sudden moratorium on Medicaid benefits in order to estimate the impact of Medicaid on the labor supply through differences-in-differences analysis. Results are mixed, with study estimates ranging from no significant effect on employment to a 25% increase in work hours as a result of a freeze on Medicaid (Garthwaite 2014). However, traditional empirical identification strategies may not work well for measuring the labor market effects of the ACA for several reasons. Firstly, states exhibited large preexisting differences in Medicaid generosity and employment trends, as well as prior state Medicaid expansions that may confound results. Secondly, the ACA included a host of other provisions that may affect the labor demand such as the employer mandate or the small business tax credit. A differences-in-differences estimator will capture both the supply and demand side effects of the Affordable Care Act, and make it difficult to isolate the effect of specific provisions of the ACA. Furthermore, differences-in-differences estimation will reveal the overall effect of the policy change and may mask competing effects among different subgroups of the population. Kaestner et al (2015) exploits the differential expansion of state Medicaid programs after the ACA by applying differences-in-differences estimation to Central Population Survey (CPS) and American Community Survey (ACS) data, but finds no significant impact on labor market outcomes one year after implementation. In light of these limitations, I take an alternative approach to modeling the labor market effects of the Affordable Care Act. I expand upon the Medicaid Notch model in Yelowitz (1995) DOBOS 3

4 by incorporating the new Medicaid eligibility thresholds and state exchange premiums into the individual budget constraint for the labor-leisure tradeoff. I derive a function for each worker s optimal choice of labor given his or her individual wage, Medicaid valuation, exchange plan premiums, and eligibility status to predict the optimal number of work hours before and after the ACA. I then input empirical data into the model and use binary search to derive individuals utility weightings to determine the final parameter of the model. Using these parameters, I then input national CPS employment, Medical Expenditure Panel Survey (MEPS), and state exchange premiums data to predict optimal work hours for each surveyed individual. I construct a subsample of employed workers within the CPS data from the years 2010 through 2015 within the age range eligible for Medicaid. I restrict my analysis to individuals residing in states and counties where I could obtain exchange premiums data and complete employment information through CPS for a final sample size of 17,359 respondents. Of these respondents, labor market outcomes were already quite high with workers working on average 37 hours per week for an average of $16.66 per hour. Respondents were evenly dispersed across counties where information was available, and faced a range in premium value of over $5,000 depending on their county of residence. I compare this value to each individual s valuation of Medicaid benefits, which I estimate through quadratic regression on key demographic characteristics from MEPS data. I find that Medicaid and the state exchange premiums have similar valuations for the average worker in the sample, suggesting a gentler cutoff than the characteristic Medicaid notch in the post-aca workforce. The main result the model predicts is a decrease in work hours of -11.8% as a result of the Affordable Care Act. This decrease is primarily a result of bunching around the exchange subsidy cutoffs rather than below the Medicaid cutoff in both expanding and non-expanding DOBOS 4

5 states, with 58% of workers opting to remain within 400% of the federal poverty line given their current wage scheme. This result suggests that Medicaid expansion may not be the primary driver behind decreases in labor force participation observed after the implementation of the ACA, which may have policy implications for non-expanding states still debating the merits of Medicaid reform. There is expected heterogeneity in results by population subgroup, with parents of dependent children and lower-income workers predicted to slightly increase work hours by 4.5% and 1.7%, respectively, as a result of the expansion; and other groups, such as individuals close to retirement, expected to decrease work hours by a magnitude of -17.3%. Since parents with dependent children and lower-income individuals were more likely to eligible for Medicaid benefits before the implementation of the ACA, an increase in work hours in response to higher eligibility thresholds is consistent with economic theory, and a large decrease in hours among elderly workers is consistent with this group s traditionally high labor market elasticity. I further test the labor market predictions of the model by running several simulations to predict the effect of future policy changes. Increasing the value of exchange subsidies by 20% or decreasing the value of Medicaid through cost-sharing of 20% has little effect on the model s estimates, but raising the federal minimum wage to $10 per hour does. The model predicts a -16.1% decrease in work hours as a response to the increase in wage, all other conditions remaining equal. This paper is organized in the following manner. Section 2 provides background on the Affordable Care Act and the provisions relevant to the labor supply model. Section 3 reviews the literature on the Medicaid Notch more generally, as well as the nascent literature base on the ACA. Section 4 develops the underlying model, and section 5 includes an overview of the data sources, summary statistics, and empirical methodology. Section 6 presents the simulation DOBOS 5

6 results, and section 7 concludes. All tables and figures are presented in section 9 at the end of the paper. 2. Background on the ACA The primary goal of the 2010 Affordable Care Act was to reduce the percentage of United States residents without health insurance. The ACA attempted to achieve this goal by expanding coverage through both public and private markets, and early research shows that the legislation has been largely successful in this domain. Prior to the implementation of the ACA, 47.3 million individuals in the United States did not have health insurance, but CDC and Census data report a decline in the uninsured rate from 15.7% in 2012 to 9.2% in 2015, and early research corroborates these results (Obamacare Facts; Kaestner et al., 2015). The ACA broadly accomplishes its goals by imposing an individual mandate to carry some form of health insurance, and then expanding the availability of coverage for traditionally underinsured groups. For the majority of working Americans already receiving group coverage through an employer, the ACA had very little effect. For uninsured US residents, the ACA now requires these individuals to carry some form of health insurance in 2016 or pay a tax penalty of $695 per adult and $ per child, up to $2,085 per family (ObamaCare Facts). 2 The ACA also includes an employer mandate to encourage firms to provide employer-sponsored health insurance for their employees. As of 2016, any business with 50 or more full-time equivalent employees (FTE) must provide health insurance for at least 95% of their employees and their 2 Or 2.5% of household income above the tax return filing status for an individual s filing status, whichever penalty is higher. Fees are prorated on a monthly basis. DOBOS 6

7 dependents (up to the age of 26) or face a $2,000 fee per FTE (ObamaCare Facts). 3 These penalties are meant to strongly incentivize individuals to carry some form of coverage, and incentivize firms to provide it. While the majority of Americans are covered through employer-sponsored health insurance, the ACA expands access to insurance through two other major provisions. The first of these provisions, and the primary focus of this paper, involves the expansion of state Medicaid programs and the second of these provisions establishes the state and federal health insurance exchanges. Combined, the interaction between these two programs is predicted to have a measurable effect on labor market outcomes in the United States. 2a. Medicaid expansion Medicaid is a jointly funded state-federal public health insurance scheme for low-income U.S. residents. The program is means-tested and completely free to residents who qualify, with effectively zero cost sharing for premiums and deductibles. Before the passage of the Affordable Care Act, Medicaid covered 49 million Americans, but there was significant variation in how it was implemented across states, leading to large gaps in care. The majority of states provided Medicaid to families with dependent children with household incomes up to 100% of the federal poverty line, but some states were more or less generous than others (Table 9.1). For example, states such as Massachusetts provided assistance for families up to 133% of the FPL and other states such as New York also extended benefits to childless adults (KFF, 2014). 4 Different 3 The 2,000 fee excludes the first 30 employees. A $3,000 fee is imposed per FTE if an employer provides insurance that does not meet minimum value standards, up to a total $2,000 per FTE. Fees are prorated on a monthly basis. 4 Most states also had different eligibility requirements for elderly, disabled, or pregnant individuals with or without dependent children. DOBOS 7

8 eligibility requirements, poverty rates, and the availability of funding created large disparities in Medicaid enrollment rates across states. The Affordable Care Act sought to expand Medicaid by standardizing eligibility requirements nationally. The ACA outlined a nationwide expansion that would increase eligibility levels to 138% of the FPL across states as of January However, in 2012 the Supreme Court ruled that states could opt out Medicaid expansion, and as of January 2014 only 26 states had elected to expand (Table 9.1). Combined, preexisting differences in program generosity along with the differential implementation of the ACA lead to large differences in enrollment, with expanding states such as California seeing a 32% increase in enrollment from 2013 to 2015 compared to a 4% increase in non-expanding states such as Texas during the same time period. This paper relies on this variation across state enrollment rates to explore the labor market effects of the Affordable Care Act among the newly insured. 2b. Health insurance exchanges The Affordable Care Act also created state and federal health insurance exchanges for individual coverage. The exchanges function as a publicly regulated private marketplace where individuals who are not covered by employers, Medicare, or Medicaid may purchase private insurance. The federal government provides subsidies to individuals making between 100% and 400% of the FPL, and calculates these subsidies by household income. Each person is free to choose whichever plan they wish to purchase through the exchanges, but they can only receive a subsidy up to a certain percentage of the cost of the second cheapest silver plan in their county. For example, a childless adult making 200% of the FPL would be expected to contribute up to 6.4% of his or her annual income toward the cost of health insurance premiums. If the second cheapest silver plan available in their county had an annual premium of $1,000, they would be DOBOS 8

9 expected to pay up to $64 toward whichever plan they select (not necessarily the second cheapest silver plan) and the federal government would provide up to $936, or the cost of the selected plan if it falls below this amount. Cost sharing subsidies for deductibles and copayments are also available for individuals between 100% and 200% of the federal poverty line to further subsidize healthcare costs. There is a substantial amount of variation in subsidies across states and counties. Some states elected not to set up their own state-specific insurance exchange but to use the federal insurance exchange instead. On the federal exchange, individuals are not eligible for subsidies below 100% of the FPL, leaving many without access to affordable coverage in states that have not expanded Medicaid or set up a state insurance exchange. Even among those who do receive subsidies, there is substantial variation in premium value by county. While this partly reflects differences in the cost of living, the range of the cost of annual premiums spans an annual price range of $5,000. In other words, an individual in the country s most expensive exchange would receive a subsidy value three times greater than an individual in the country s least expensive exchange, even if these two individuals earn the same income. It is still unclear how the state exchanges will affect the labor market in the United States. Like Medicaid, the program is means-tested and can potentially reduce labor supply, but the subsidies gradually decline with income rather than reach the sharp cutoff characteristic of Medicaid. In addition, the exchanges also include competing effects from the employer mandate and small business tax credit. While the employer mandate may raise costs and disincentivize firms from hiring 50 or more full time equivalents, the small business tax credit provides tax breaks for businesses with 25 FTEs or fewer that provide coverage to their employees, DOBOS 9

10 potentially increasing wages and employment. The exchanges thus introduce new labor market effects unique to the ACA that have not yet been studied extensively in the literature. 3. Literature Review There is a large and inconclusive literature base on the impact of Medicaid on labor market outcomes. Economic theory predicts that means-tested programs in general will create a disincentive for individuals near the income threshold to work, causing them to work fewer hours in order to continue receiving program benefits. As a program with a historically sharp income cutoff, Medicaid is one of the most well studied examples of the labor market distortions created by means-tested programs. Yelowitz (1995) describes this effect as the Medicaid Notch, where individuals face a discontinuity in their budget line at the income level in which they are no longer eligible for Medicaid (Figure 9.1). The Medicaid Notch model predicts fewer work hours among Medicaid recipients than non-medicaid recipients and sharp bunching before the cutoff. Yelowitz (1995) develops an empirical framework to test for such a discontinuity in individuals labor supply curves. He exploits a string of legislation passed by Congress from that increased the effective income cutoff for Medicaid recipients, and increased the age limit at which dependent children qualified for Medicaid. Using households with children on either side of the new age limit as experimental and control groups, Yelowitz employs a differences-in-differences method to estimate a 0.9% increase in labor force participation among single mothers in response to the expansion. These results suggest the existence of a small yet significant notch in Medicaid recipients labor market decisions. DOBOS 10

11 Other researchers contest these effects. Meyer and Rosenbaum (2000b) argue that other aspects of the 1980s policy change Yelowitz (1995) exploits confound his results, and indeed subsequent analyses of the national Medicaid expansion find little to no effect (Yacizi 1997; Shore Sheppard 2000). Gruber and Madrian (2002) review 50 papers on the link between health insurance and labor supply and conclude that there is little evidence that Medicaid disincentivizes labor force participation among current recipients. 5 However, they do acknowledge the limited number of studies specifically examining the Medicaid notch at the time of publication, and report that the last word has clearly not been written on this topic (Gruber & Madrian 2002, 19). A series of recent studies has reexamined the link between Medicaid and employment since the publication of Gruber & Madrian (2002). Garthwaite et al. (2013) take advantage of an abrupt discontinuation of TennCare, Tennessee s state Medicaid program, in 2005 that left 170,000 non-elderly adults without health insurance. The authors use the sudden change in Medicaid eligibility to estimate the labor market impact for childless adults, the group most affected by the policy change, and find large effects. Their results are consistent with an immediate increase in job seeking behavior and an estimated 25% increase in labor force participation among low-income, childless adults affected by the policy change. Garthewaite et al. acknowledge the large effect size and conclude that Medicaid expansion under the ACA may cause a substantial decrease in the labor force supply of low-income individuals. 5 Gruber and Madrian (2002) do not restrict analysis to papers exploiting the Medicaid expansions. They review studies employing one of four identification strategies: (i) exploiting differences in individual valuation of Medicaid on likelihood of labor force participation (ii) exploiting differences in state Medicaid generosity (iii) exploiting policy changes related to eligibiltiy and (iv) exploiting the differential generosity of employer-based insurance should an individual exit public health insurance. Small to zero effects are consistent across identification strategies. DOBOS 11

12 However, studies from Medicaid programs in different states provide conflicting evidence. Baicker et al. (2013) uses evidence from the Oregon health insurance experiment to conduct a randomized evaluation of Medicaid s effects on employment outcomes. In 2008, the state of Oregon selected 30,000 individuals on a waiting list for Medicaid through a randomized lottery to receive health insurance benefits. Baicker et al. analyze administrative records from 2009 and find no significant effect on labor market outcomes for lottery recipients. Dague, DeLeire, and Leininger (2014) find only modest effects from an enrollment cap on Wisconsin s Medicaid program in The cap was instituted five months after the program expanded public health insurance to childless adults below 200% of the FPL, who had previously been ineligible. Applicants who enrolled after the cutoff were placed on a waitlist until enrollment dropped, and ultimately did not receive program benefits. The authors employ both a regression discontinuity design and differences-in-differences estimator to estimate the effect of the policy reversal, and find that employment increased 2 10 percentage points up to 9 quarters after the freeze. These results are higher than Baicker et al. (2013) but significantly lower than Garthwaite et al. (2013), which the authors explain through variation in labor force opportunities between states. Unfortunately, the state-specific scope of each of these three studies makes it difficult to extrapolate results to a national policy change on the scale of the ACA. While still early, there is a nascent literature base directly evaluating the effects of the ACA. Antwi, Moriya, & Simon (2012) and Heim et al. (2015) evaluate the earliest insurance expansion provision of the ACA to go into effect: the extension of dependent coverage. The authors employ a triple differences estimator to find a significant uptake in parental coverage among the age group, as well as modest evidence of increased labor market mobility. The overall probability that 18 to 25 year olds were employed did not change after the ACA, but DOBOS 12

13 the probability of full-time employment and the number of work hours among this sample decreased by 5.8% and 3%, respectively. Heim, Lurie, and Simon (2015) perform a similar analysis on US tax data records and find no statistically significant impact of increased dependent coverage among young adults after the passage of the ACA. Kaestner et al. is one of the first studies to examine the full implementation of the ACA s expanded Medicaid program. Using differences-in-differences analysis, the authors exploit the differential implementation of Medicaid expansion by state to evaluate the labor market effects of Medicaid. They use Current Population Survey (CPS) and American Community Survey (ACS) data and restrict the sample size to adults with a high school education or less as a proxy for income. Among this group most likely to be affected by Medicaid expansion, the authors find no significant change in labor supply for total hours worked, full-time employment, or overall employment rates, despite a 4% uptake in public insurance. These preliminary results suggest that Medicaid does not have a significant impact on employment, but it is difficult to tease out causal effects from Medicaid expansion alone given the simultaneous implementation of state exchanges. In addition, differences-in-differences analysis may not be an appropriate experimental method in regard to state Medicaid expansion given the large preexisting differences in the generosity of state Medicaid programs, and the fact that many states expanded their Medicaid programs individually prior to 2014 (Table 9.5). These concerns, as well as the limited timeline of analysis, cast doubt on the applicability of Kaestner et al. s findings to the effect of Medicaid on labor supply. This paper contributes to the existing literature by expanding upon the Medicaid notch labor supply model formalized in Yelowitz (1995). The model incorporates the new budget constraints created by the state exchange subsidies in addition to the new eligibility requirements DOBOS 13

14 of state-specific Medicaid expansions. The use of this theoretical labor supply model provides an avenue for exploring more precise measurements of labor supply elasticity in response to the ACA not captured by empirical analysis. Differences-in-differences can estimate the overall effect of the ACA, but a labor supply model allows us to study the distribution of hours around the predicted optimum, and to predict and test heterogeneous effects for different population subgroups. These questions are particularly policy-relevant given the fact that many states have not yet expanded their Medicaid programs out of fiscal concerns, and labor market effects may play a role in these decisions. This paper provides a theoretical framework for such analysis as well as a review of the empirical findings reported in Kaestner et al. (2015) on the effects of the Affordable Care Act. 4. Model I follow the example of Yelowitz (1995) and expand upon the static labor supply model by incorporating tax credits, the value of state Medicaid, and exchange program subsidies into the individual budget constraint. Under this expanded model, individuals budget constraints vary by time period, state, and eligibility status. Figures 9.2 through 9.7 depict sample budget constraints for childless adults and parents of dependent children prior to the ACA, and how these budget constraints change in expanding and non-expanding states after While all individuals face a non-linear budget constraint after the implementation of the ACA, the budget constraint can be divided neatly into three cases i) The budget constraint without Medicaid or exchange subsidies ii) the budget constraint with Medicaid and iii) the budget constraint with exchange subsidies. All individuals fall into at least one of these three categories, making it possible to predict their optimum labor-leisure allocation before and after the ACA. DOBOS 14

15 I begin by assuming a Cobb-Douglas utility functional form, where individuals maximize income (Y) and leisure (l) according to their individual preferences α and β. max!,! α log Y + β log (l) (1) The simplicity of Cobb-Douglas utility provides us with a transparent base for modeling the effects of the ACA, and remains consistent with empirical estimates. In the linear case, the substitution and income effect from a change in wage exactly balance one another under Cobb- Douglas utility, which is consistent with historical employment trends. Average work hours in the United States have not changed substantially despite increases in the real wage over the past few decades, implying low labor supply elasticity. The model assumes that workers can vary the number of hours worked (h) but not hourly wage (w) in order to increase income, where income is a function of earned wages and Medicaid benefits (M), or earned wages and exchange plan benefits (P) less whatever portion of their income they must pay toward annual premiums (c i ). For illustration, I solve for the optimal allocation for an individual facing the nonlinear budget constraint in Figure 9.4. Their budget constraint is then Y = w T l + M h h! w c! w T l + P h!!! < h h! i 5 w(t l) h 5 (2) where T represents total time. It is important to note that I have simplified the budget constraint for exchange subsidies in the model above slightly. I average the percentage each worker is expected to pay of their annual income c i within each income tax bracket (Table 9.4), and ignore out-of-pocket costs associated with different exchange plans along with the cost sharing subsidies provide by the federal government for those earning between 100% and 250% of the FPL. In theory, these simplifications should not affect the optimal labor-leisure choice in an DOBOS 15

16 actuarially fair insurance market. Under actuarially fair insurance, insurers should set the price of premiums exactly equal to each individual s expected health expenditure, less any form of cost sharing. While information asymmetry and pooling contracts prevent perfect price discrimination in healthcare, I assume that we arrive at actuarially fair insurance in the aggregate, with premium prices reflecting individuals true valuation of health insurance. Returning to the model, individuals in the first case (h h 0 ) earn below 100% of the federal poverty line and receive Medicaid benefits in addition to earned income. Optimizing, workers let!! =!!" s. t Y = w T l (3) and substituting into the budget constraint find which yields an optimal allocation of Y = w T!!!! + M (4) Y! =!(!!!!)!!! l! =!!!!!!!!! (5) (6) They repeat a similar optimization in the second and third cases. In case 2 (h i-1 h h i for i 5), workers are between 100% and 400% of the FPL and ineligible for Medicaid, but receive exchange subsidies. The value of these subsidies are calculated as the value of the second cheapest silver plan premiums within their county less the percentage of their annual income they are expected to contribute toward health insurance. The optimal allocation is as follows Y! =![!!!!!!!!]!(!!!) l! =![!!!!!!!!]![(!!!! )(!!!) (7) (8) DOBOS 16

17 In case 3 (h h 5 ), workers earning over 400% of the FPL become ineligible for exchange subsidies and once again face a linear budget constraint. The optimal allocation under a static labor supply model is then Y! =!"!!!! l! =!!!!! (9) (10) and the worker s overall choice can be determined by! max!,! U! Y!, l!!!! (11) as a function of their individual wage, model parameters, and preferences for labor relative to leisure. 5. Data 5a. Data sources This paper uses information from three main data sources to estimate model parameters. 1) I obtain demographic data from United States monthly Central Population Surveys (CPS) from January 2010 December 2015 for 2.19 million individuals. The dataset contains detailed information on employment, income, household composition, and demographic data for a national cross-section of individuals; however, it does not include information on health insurance expenditures or health status. 2) I supplement this information with annual Medicaid expenditure data from the national Medical Expenditure Panel Survey (MEPS) from 2013, which includes state-reported annual Medicaid expenditures by respondent and self-reported health status for approximately 300,000 individuals. DOBOS 17

18 3) I obtain monthly exchange plan premiums by county for 2014 and 2015 from data.healthcare.gov. This dataset includes monthly premiums for different health insurance plans for individuals by metal level (bronze, silver, gold, and platinum) age group (ages 21, 30, 40, and 50) and household composition (individual, couple, or couple with children) across counties. The dataset includes full information for 33 states, and limited individual premium information for two states: California and Oregon. 6 I supplement this dataset with information on historical Medicaid income cutoffs by state (Table 9.1) from the Kaiser Family Health Foundation, and the new exchange subsidy income cutoffs from the informational site Obamacarefacts.com (Table 9.4). A summary of the model parameters and their respective data sources is included in Table b. Subsample Construction I) I construct a subsample of individuals for inclusion in the model estimation from CPS survey data. I begin with a sample of 2.19 million nationally representative individuals and narrow this number to respondents with a reported FIPS code for their county of residence, approximately 40% of the original dataset. I then narrow the subsample to workers who are in the age range for Medicaid (ages 26 to 64) with reported wage and weekly hours, and exclude respondents from states with missing premium information for a final sample size of 17,397 individuals. Of this sample, 11,448 respondents were surveyed before the implementation of the Affordable Care Act and 5,949 respondents were surveyed afterward, with roughly 3,000 6 Data was not available for Colorado, Connecticut, Hawaii, Idaho, Iowa, Kentucky, Maryland, Massachusetts, Minnesota, Nevada, New Mexico, NY, Oregon, Rhode Island, and Washington. Premium data was only available by age and not household composition for California and Oregon. Wisconsin was excluded from analysis. DOBOS 18

19 observations per calendar year. I use self-reported hourly wage data and self-reported usual weekly hours for employment measures in the model. 7 II) I supplement this model by creating a variable that represents each individual s private valuation of Medicaid using MEPS survey data. Restricting the sample size to only those continuously enrolled in Medicaid throughout the year and within age range, I regress total Medicaid expenditure for 2013 on key demographic traits that previous literature has demonstrated a strong correlation with health expenditure using the following formula: Total_Medicaid_Expen i = α 1 Age i + α 2 Age_Squared i + α 3 Male i + α 4 Bachelors i + ε i (12) where Total_Medicaid_Expen is the total state-reported Medicaid expenditure for individual i in 2013, Age and Age_squared represent the non-linear interaction between age and health care expenditure, Male is a dummy variable representing the gender of individual i, and Bachelors is a dummy variable indicating whether individual i has completed a bachelors degree or higher at an accredited four-year institution. 8 The regression results (Table 9.8) indicate a strong positive correlation between age and Medicaid expenditure significant at the 95% confidence level, and a negative correlation with being male or receiving a bachelor s degree, although these coefficients are not significantly different than zero. I use the regression coefficients in Table 9.8 and associated standard errors to then construct a Medicaid valuation variable within the CPS data set based on age, gender, and education data for each CPS respondent. Under this construction, I assume that MEPS constitutes a nationally representative sample of the population, and equate private Medicaid valuation with expected individual expenditure. 7 If usual weekly hours were not reported, the number of hours the respondent reported working in the last week was used instead. 8 Respondents with an associate s degree or some college were coded as not receiving a bachelor s degree. DOBOS 19

20 III) I then merge in exchange plan premiums for each individual by state and county, and calculate the second cheapest silver plan for each respondent given their age and household composition. I combine this information with wage and weekly hours to calculate earned annual income, and the percent of income that each individual is expected to contribute to health insurance premiums for those qualifying for the new state and federal exchanges. 9 The resulting dataset includes key variables on hourly wage, weekly hours, annual income, household composition, second cheapest silver plan annual premiums, expected exchange subsidies, and expected Medicaid expenditure for those who qualify. 5c. Summary statistics Table 9.9 includes an overview of summary statistics for key employment, demographic, education, and health expenditure variables used to calculate the model parameters. The first three rows of the table reveal relatively high levels of employment among the population subsample. Out of adults ages 26 to 64 who report wage and hours, respondents earn on average $16.66 an hour and work approximately 37 hours per week. Total average income across all four years is $31,327 well above the threshold for Medicaid eligibility, but below the maximum cutoff for exchange subsidies at 400% of the federal poverty line. The high level of employment in the sample is expected since respondents with missing wage or hours information were removed from the dataset, but it does indicate that workers estimated preferences for labor over leisure may provide an upper bound for labor preferences when compared to the general population. Figures 9.9 and 9.10 demonstrate trends in employment over time. Weekly hours and hourly wage decline for both expanding and non-expanding states prior to the passage of the 9 I assume 52 working weeks a year, with 50 total work weeks and 2 weeks of paid vacation. DOBOS 20

21 Affordable Care Act in 2012, but increase steadily after 2012 and the implementation of the Affordable Care Act in After 2012, individuals in non-expanding states work on average hours more per year than their counterparts in expanding states, but workers in expanding states earn 1.00 USD more per hour, significant at the 10% level and 1% level, respectively (Table 9.10). Figures 9.9 and 9.10 depict significant preexisting labor market outcomes between states, but a parallel trend after 2012 between expanding and non-expanding states. I also find preexisting heterogeneous labor market outcomes by demographic group, suggesting that the passage of the Affordable Care Act may have differential impact for various types of individuals. In this sample of employed workers, men work on average 195 more hours per year than women, and workers who have received a bachelor s degree or higher work 29 fewer hours per year than those who have not received a bachelors degree despite higher average wages significant at the 1% level (Table 9.11). Age has an expectedly high correlation with hours worked per week, with the number of weekly hours increasingly steadily with age until age 50, at which individuals closer to retirement decrease weekly hours (Table 9.12). There is no significant difference in hours worked between parents and childless adults in the aggregate. Geographically, there are no significant employment trends for different regions in the United States, although there is a large amount of variation between individual states. Figure 9.11 provides a breakdown of the states and counties that account for the largest number of observations in the subsample. As the United States most populous state, California constitutes 30.6% of the total number of observations in the data, followed by Florida, Pennsylvania, and Delaware, respectively, and also includes the largest number of counties. The majority of the counties included in the subsample are large population centers by design, as the CPS only DOBOS 21

22 reports county FIPS codes for respondents who live in states and counties large enough to avoid anonymity concerns. However, respondents are relatively dispersed between the 274 unique counties within the dataset, with the largest county, Los Angeles, only accounting for 3.5% of the total sample size. Despite its limitations, the availability of county level data provides an important source of variation for exchange premiums. Premiums vary greatly by county, with premiums for individuals in the age group ranging from $1,608 per year in Allegheny, Pennsylvania to $3,483 in Livingston, Louisiana. The range is even wider for older workers, with individuals in the age range facing second cheapest silver plan premiums from $4,367 per year in Allegheny to $10,452 in Imperial, California. While it was not possible to estimate differences in Medicaid expenditure by region given the anonymity of MEPS data, average Medicaid expenditures are close in value to average exchange premiums across age groups (Figure 9.8). Annual coverage through the exchanges second cheapest silver plan is slightly higher for the average worker, with Medicaid worth an estimated $5,624 per year compared to $5,974 through the exchange premiums. This is true for all age groups outside of the age group, although on average the difference between the two forms of coverage is $322, suggesting that the exchanges have decreased the sharp cutoff characteristic of the Medicaid notch. 5d. Model estimation I use the model developed in Section 4 to estimate predicted hours in the constructed subsample. I first separate the subsample into groups based on the shape of their budget constraint, where Groups 1 and 2 only include individuals surveyed before the implementation of the ACA ( ). Group 1 individuals are childless adults in states that did not offer DOBOS 22

23 Medicaid prior to the ACA who face the linear budget constraint depicted in Figure 9.2, while Group 2 includes parents of dependent children or childless adults who resided in states that offered Medicaid to childless adults below some level of the Federal Poverty Line. Group 2 s budget constraint is depicted in Figure 9.3. Conversely, Groups 3 through 5 only include respondents surveyed after the implementation of the Affordable Care Act ( ). Group 3 includes childless adults who reside in states participating in Medicaid expansion, Group 4 includes childless adults in nonexpanding states, and Group 5 includes parents with dependent children across both expanding and non-expanding states, with the budget constraints for each group depicted in Figures 9.4, 9.6, and 9.7, respectively. Table 9.13 provides summary statistics on employment, income, and demographic factors for each subgroup, with little significant variation across subgroups. I predict hours of work post-aca using equations 5 through 10 for each individuals optimal labor-leisure tradeoff. I assume that workers can vary the number of hours worked but not wage, and calculate individuals optimal labor choice for each discrete segment of their budget constraint. I let hours pred = the labor choice that provides workers with the highest level of utility across all optima, and then compare this number to empirical data in order to determine the utility weightings α and β. Specifically, I use binary search to minimize the sum of squared errors:!! (hours hours!"#$ )! (13) between predicted hours and the actual number of hours individuals report working in the subsample. In doing so, I assume that workers start with no endowment outside of their earned income and public health insurance benefits, and take wage as an exogenous parameter. I set the total number of hours to 8,760 the maximum number of hours in a year and normalize α and DOBOS 23

24 β to 1, so that β = 1 α. I estimate α as a constant value for different subgroups of the population, and then use the calculated α for different individuals to generate hours pred.the full code for the hours_pred function is included in Figure 9.14 of the appendix. 6. Results 6a. Labor market predictions The model predicts a moderate decrease in employment as a result of the implementation of the Affordable Care Act. Table 9.14 reports an overall expected -11.8% decrease in the number of hours worked after the ACA takes effect, and Table 9.15 provides a breakdown of these estimates by eligibility group. Individuals in Group 1 childless adults ineligible for Medicaid prior to the ACA work a predicted average of 41.6 hours per week, compared to the 33.8 hours per week predicted for parents or childless adults residing in states with more generous Medicaid programs. After the implementation of the ACA, newly eligible childless adults in Group 3 are predicted to decrease employment to 34.1 hours per week, while their counterparts in non-expanding states (Group 4) are predicted to work slightly longer hours at 35.3 hours per week. In this case, the exchanges lessen the labor market effects of income thresholds by creating a more gradual cutoff for individuals earning between 138% and 400% of the Federal Poverty Line. Parents with dependent children (Group 5) after the ACA continue to face the most generous eligibility cutoffs for Medicaid across groups, and are predicted to work the lowest average of 32.4 hours per week. The relative predicted hours between each group conforms to economic theory for means-tested programs, lending support to the validity of the overall labor market estimates. DOBOS 24

25 Furthermore, I observe predicted bunching around eligibility cutoffs. Figure 9.12 plots predicted hours against hourly wage and depicts significant bunching at the Medicaid eligibility threshold for individuals surveyed prior to the ACA, and bunching at each exchange subsidy eligibility threshold for individuals surveyed afterward. Table 9.16 reports the percentage distribution of hours around these cutoffs, with 17% of individuals below the Medicaid eligibility threshold before the ACA compared to 23% afterward. Among groups 3 through 5, 58% of individuals are eligible for exchange premium subsidies, with relatively smooth dispersion across the six subsidy brackets recognized by the federal government. Only 17% and 19% of workers opt to work enough hours to earn beyond Medicaid and exchange cutoffs, respectively, given their current hourly wage. These results remain consistent under different model specifications. I allow α and β to vary by demographic characteristics namely gender and educational background in order to obtain more precise individual utility weightings. I also specify models with different coverage estimates in order to test the robustness of the given Medicaid and exchange premiums data. I first test the model using different regression equations to determine each individual s Medicaid valuation (Table 9.18), and then test the model calculating exchange subsidy values from individual premiums, rather than family premiums adjusted for household size for parents of dependent children. The model predicts a decrease in the number of hours worked as a result of the ACA under each specification, with estimates ranging from a low of -10.4% to a high of 12.1% decrease in weekly hours. 6b. Applications One of the primary benefits of a theoretical model is the ability to explore heterogeneous effects by population subgroup, and simulate different policy scenarios. I use the developed DOBOS 25

26 model to explore these applications, beginning with the effect of the ACA for different population groups. While the model predicts that the ACA will depress employment in the aggregate, this is not true for every demographic. Population groups that were likely to be previously eligible for Medicaid such as parents of dependent children and workers in the bottom economic quartile are expected to increase work hours by 4.5% and 1.7%, respectively, as a result of the expansion. Other subgroups are predicted to work fewer hours as a result of the ACA, but the magnitude of the legislation s impact varies. For example, younger workers (age group 26 30) with relatively low labor market elasticity are expected to decrease the number of hours worked by only 7.3%, whereas older workers closer to retirement (age group 60 64) with traditionally high labor market elasticity are expected to decrease work hours by 17.3%. Table 9.17 summarizes predicted hours by different demographic and income groups. Table 9.19 explores several policy simulations. I test three exogenous shocks to the administration of the Affordable Care Act: 1) an increase in the federal minimum wage from $7.25 to $10.00 per hour 2) an increase in exchange plan subsidies by 20% for every income bracket eligible for public assistance on the exchanges, and 3) a decrease in the value of Medicaid by 20% with the addition of cost sharing component for all expenditures. Increasing exchange subsidies and decreasing the value of Medicaid both have negligible effects on employment, but increasing the minimum wage increases the magnitude of the ACA s negative employment effects. It increases the change in predicted work hours from -11.8% to -16.1%., suggesting that states with higher minimum wage requirements may experience larger labor market distortions after the implementation of the ACA. DOBOS 26

27 7. Conclusion I predict the effect of the Affordable Care Act on optimal hours of work using a simulation model. I find that overall, the ACA should result in an -11.8% decrease in work hours among the employed population. I also find bunching at the eligibility cutoff for the public health insurance exchanges, with 58% of post-aca individuals selecting hours of work to remain under 400% of the Federal Poverty Line and eligible for exchange subsidies. The amount of individuals under the Medicaid cutoff also increased after the ACA, but this percentage is small in comparison to the distribution of individuals within the exchanges. Combining this information, the model predicts that the state exchanges will have a larger impact on labor market outcomes than Medicaid expansion, a result that may be policy-relevant as other states consider expanding their Medicaid programs. However, it is important to place any results from this analysis in the context of the limitations of the model. As a simulation model, these results do not necessarily reflect empirical realities. I made a number of simplifying assumptions in developing the new labor supply budget constraint, and was limited by the scope of analysis. Future iterations of the model should include a more flexible utility functional form, such as constant elasticity of substitution (CES) utility, and employ maximum likelihood estimation to estimate individual utility weightings α and β as a function of various characteristics, such as age, gender, household composition, county, or self-reported health status. An additional avenue of research would be to test the model s simulation predictions against empirical policy changes, such as raising the state minimum wage, in order to assess the validity of the model as more post-aca data becomes available. Until then, this paper provides a starting point for a much richer model to estimate the labor market effects of the Affordable Care Act. DOBOS 27