The Effects of Medicaid Earnings Limits on Earnings Growth Among Poor Workers. Sarah Hamersma University of Florida 9/3/07

Transcription

1 The Effects of Medicaid Earnings Limits on Earnings Growth Among Poor Workers Sarah Hamersma University of Florida 9/3/07 PRELIMINARY PLEASE DO NOT CITE I. Introduction Labor force participation among single mothers has risen dramatically since many aspects of public assistance were reformed in the mid-1990s. Wage subsidy programs and child care assistance have made it increasingly attractive for women to undertake market work, and welfare work requirements have made it less feasible for them to remain out of the labor force for long periods of time. However, earnings remain low for this population, as many do not work full-time and few have jobs with high wages. This may be in part because many public assistance benefits phase out as earnings increase, limiting the gains associated with increased earnings. In this paper I examine the work disincentives of a key public assistance program for which the loss of benefits can be particularly costly: public health insurance. Medicaid is a federal health insurance program available for needy families and children, subject to eligibility guidelines that are determined primarily at the state level. Since 1996, states have been free to vary their Medicaid eligibility criteria independently of any changes in other public programs. The result is substantial variance across states and over time in the earnings limits to which workers must adhere to maintain their Medicaid eligibility. In some states, earnings from a part-time job are likely to exceed the threshold; other states continue to provide Medicaid coverage for those with many more hours of work. My goal in this study is to identify the effects of these varying earnings limits on labor force behavior of low-income single mothers. The existence of a Medicaid earnings limit introduces a potentially complicated work disincentive for workers with low earnings. Rather than discouraging work entirely (as some former cash assistance programs have done), or creating lower effective wages than the market wage (due to I am grateful for the generous support of the Smith Richardson Foundation and for excellent research assistance from Katie Sherron, Doug Webber, and Daniel Weinberg. I would also like to thank Chris Bollinger, David Figlio, Jonathan Hamilton, Larry Kenny, Christine Piette, Lara Shore-Sheppard, and Jim Ziliak for productive discussions about this work. I also appreciate the University of Kentucky Center for Poverty Research for contributing to this work by inviting me to visit as part of their Emerging Scholars Program. All errors are mine. 1

2 phasing out benefits as earnings increase), the Medicaid earnings limits create a notch in the potential budget of working women. When women who were on Medicaid exceed the earnings limit by a small amount, they typically lose Medicaid entirely. 1 It is clear, then, that the added earnings from working additional hours or accepting a promotion with higher wages may actually have negative effects on overall income when the value of Medicaid is taken into account. I test two hypotheses about the potential effects of the Medicaid notch on the earnings levels of low-income women. First, I expect that women who would typically have earnings near the Medicaid earnings limit would intentionally avoid going over this limit, either by avoiding increased hours of work or by choosing not to advance to jobs with higher wages. If this is the case, there should be a disproportionate number of women with earnings just below the limit and proportionally fewer women with earnings just above the limit. I test this hypothesis on earnings levels by using the Current Population Survey (CPS) to examine the distributions of earnings for single mothers in states with differing Medicaid earnings limits. I then expand the analysis using a regression strategy with state fixed effects. The CPS analysis provides some suggestive, but not definitive, evidence that Medicaid expansions may contribute to increased earnings due to bunching below the threshold. I then perform the same analysis using data from the 1996 and 2001 Surveys of Income and Program Participation (SIPP) to corroborate the evidence. This regression analysis does not indicate a strong systematic relationship between Medicaid thresholds and earnings. My second hypothesis relates to the potential effects of changes in Medicaid earnings limits on earnings growth over time. If some women s earnings had reached a plateau under an existing limit, an increase in that limit should result in a more positive earnings trajectory since the old limit is no longer binding. At the same time, some previously non-medicaid workers may be induced to reduce earnings to get below the new increased limit, since the needed reduction in earnings may now be small relative to the value of Medicaid. These opposing trends may confound one another to generate the apparent lack of relationship between Medicaid thresholds and earnings levels, even if workers earnings trajectories are in fact responding to the thresholds. Since this trajectory analysis requires data on individuals over time, I utilize data from the 1996 and 2001 panel of the SIPP. The econometric model allows differential effects of the Medicaid threshold changes on women in different parts of the earnings distribution (relative to the original threshold). Even in this 1 There is a Transitional Medical Assistance program that is implemented to varying degrees in each state. This allows some limited transition time for a family that has lost eligibility to maintain coverage. The Food Stamp program also has a small notch due to its gross income standard. 2

3 more flexible model, I find little evidence that the Medicaid threshold affects earnings patterns. This suggests that the program may provide less of a work disincentive than is commonly thought. II. Medicaid Income Threshold and Incentives The Medicaid program has gone through fundamental changes in the last 10 years. Prior to 1996, adult eligibility for Medicaid was tied directly to eligibility for cash assistance (Aid to Families with Dependent Children, or AFDC). 2 The two programs were de-linked in the Personal Responsibility and Work Opportunity Reconciliation Act of 1996 (also know as welfare reform). States were allowed to either maintain their 1996 eligibility guidelines for Medicaid or to expand eligibility as they saw fit, which the majority of states have since done (Aizer and Grogger, 2003). The result is a program with widely varying earnings eligibility guidelines across states and over time. Many states have also made more children eligible for coverage even when their parents income exceeds the Medicaid earnings limit, through the State Children s Health Insurance Program (SCHIP) that was introduced in Medicaid is in particular need of careful study because it has become an increasingly prominent feature of public assistance. From the late 1970s until about 1990, Medicaid coverage remained steady at about 20 million participants. 3 Since 1990, coverage has more than doubled in conjunction with more generous eligibility guidelines for children and, later, for adults. The current Medicaid caseload exceeds 45 million people. In contrast, cash assistance (AFDC/TANF) caseloads fell dramatically in the mid- to late-1990s, from a peak of about 14 million recipients to just 5 million in recent years. The fact that nearly 90 percent of Medicaid recipients do not receive AFDC/TANF cash assistance makes the examination of Medicaid interesting beyond its connection with other assistance programs. There are a number of empirical studies of Medicaid and its recent expansions, but very few have focused on labor supply issues. 4 Moffitt and Wolfe (1992), Yelowitz (1995), and Ham and Shore-Sheppard (2005) are three of the key papers in this small literature, and each of them 2 There were some state-level expansions for childrens coverage in years prior to 1996, but larger expansions in family coverage have occurred in more recent years since Medicaid eligibility was formally separated from cash assistance at the federal level. Yelowitz (1995) exploits the pre-1996 variation to examine labor supply and welfare participation among single mothers. 3 The information on caseloads in this paragraph is drawn from a brief from the Congressional Budget Office (2005). 4 For example, Aizer and Grogger (2003) examine the effects of eligibility expansions on health coverage and crowd-out and Gruber and Yelowitz (1999) examine the role of Medicaid asset limits on savings behavior. A detailed review of other Medicaid research, as well as additional information about Medicaid s legislative history, can be found in Gruber (2000). 3

4 addresses the effects of Medicaid on both labor supply and welfare participation, reflecting the often tight link between Medicaid and cash assistance prior to welfare reform. Moffitt and Wolfe (1992) develop a family-specific proxy for the valuation of Medicaid benefits and find that Medicaid has the expected negative effect on labor supply and a positive effect on welfare participation. Yelowitz (1995) exploits state and year differentials in the Medicaid eligibility guidelines for young children during the period 1989 to 1992 to examine the same questions. Consistent with Moffitt and Wolfe (1992), he finds that more generous income limits are associated with increased participation in the labor force and decreased participation in cash welfare. However, Ham and Shore-Sheppard (2005) demonstrate that some key assumptions in Yelowitz (1995) are inappropriate for the Current Population Survey data that were used. 5 Using a more flexible functional form with the same data, they find no evidence of a relationship between Medicaid income limits and welfare or labor force participation. There are a few limitations of these key papers that I address in this project. First, all three studies generate only cross-sectional estimates of program effects, either with one cross section of a panel (Moffitt and Wolfe, 2002) or with repeated cross sections (Yelowitz, 1995; Ham and Shore- Sheppard, 2005). 6 The authors therefore do not address the potential effects of Medicaid policy on individual earnings growth over time. The current project uses both repeated cross sections and panel data to address the effects of Medicaid earnings limits. Second, these papers use data prior to the recent expansions in adult Medicaid earnings limits. There is now much more variation in Medicaid earnings limits than in the years examined in this earlier work; a variety of changes in eligibility standards for both children and adults (which reach higher into the income distribution) make research based on more recent data particularly interesting. Although this study will focus primarily on changes in adult eligibility, I will also incorporate changing child eligibility standards. The role of these two separate sets of standards is particularly interesting as the public considers targeted Medicaid cuts or expansions in coming years. Finally, and perhaps most importantly, the papers I described focus on labor supply decisions on the extensive margin (the choice of whether to work or not) with less attention paid to the intensive margin (the decision of how much to work). The recent substantial increase in Medicaid coverage for people who are working makes this 5 The key issue is that Yelowitz (1995) examined the effects of Medicaid thresholds using a variable called GAIN% defined as the percent difference between the Medicaid threshold and the AFDC (welfare) threshold. This effectively forced the Medicaid and AFDC thresholds to have effects that were equal in magnitude and opposite in sign. Ham and Shore-Sheppard (2005) test whether this assumption holds in the data, and reject the assumption. They instead allow AFDC and Medicaid thresholds to enter separately. 6 Moffitt and Wolfe (1992) utilize the Survey of Income and Program Participation, which is a panel data set, but they use the panel information primarily to assemble variables related to past health care for use in a cross-sectional analysis. 4

5 distinction particularly important in the current policy regime. Meyer (2002) reports the puzzling finding that the extensive margin appears to react strongly to policy (in his case, the Earned Income Tax Credit) but the intensive margin appears quite unresponsive to the incentives created by higher benefit reduction rates. In addition, Saez (2002, QJE) establishes that programs with phaseouts along the intensive margin are most effective when labor supply elasticities are small on the intensive margin, and he cites several studies confirming that these elasticities are small. It is not clear whether the same unresponsiveness exists with a more discrete benefit reduction (loss of Medicaid). This project concentrates primarily on the intensive margin, since labor force participation among single mothers has increased dramatically since the 1996 welfare reform but total earnings remain low. Beyond the Medicaid literature, this project is also related to the literature on effective marginal tax rates. This literature examines the impact of benefit reduction rates on workers incentives to increase labor supply. For example, recent papers by Hepner and Reed (2004) and Romich (2005a) map out the effective marginal tax rates for workers in several states (accounting for several different assistance programs), noting the often very small increases in total resources obtained when low-income workers increase their work hours and/or earnings. Liebman and Zeckhauser (2004) and Romich (2005b) investigate potentially confusing tax and transfer policies, in hopes of better understanding workers reactions to (among other things) benefit reductions that occur at some marginal rate that changes across earnings levels. I am not aware of any paper in this area that has explicitly focused on the Medicaid notch, so this study fills an important gap in this literature as well, helping to clarify whether workers appear to be aware of the Medicaid threshold and whether they respond in the expected way. III. Data and Methods for Assessing the Effects of Medicaid Income Thresholds For my empirical analysis, I focus on two primary questions: First, do workers near the earnings limit appear to respond to that limit by keeping earnings levels low enough to retain eligibility for Medicaid? Second, do workers experience changes in earnings growth when the Medicaid earnings limit increases? My analytical approaches will make use of the fact that many states have substantially increased their Medicaid earnings limits in the last decade, and the extent of these expansions has varied widely across states. The resulting mix of policies allows me to compare earnings patterns in states with large changes to those with small changes or no changes at all. Of course, other differences across states can be important factors in determining outcomes, so this 5

6 work also takes these differences into account in an attempt to isolate the effects of the Medicaid earnings limit. I address the first question using microdata from the Current Population Survey s Merged Outgoing Rotation Groups (CPS MORG) for November 1999 December These data provide rich demographics and recent earnings information for a nationally-representative crosssection of about 30,000 individuals each month. I then address the second question and corroborate my findings on the first using the Survey of Income and Program Participation (SIPP). I use the SIPP for this purpose because it contains detailed longitudinal data on labor force and public program participation of a sample of the lower income population. Though it is a smaller sample than the CPS, an important feature of the SIPP is that it follows the same individuals for three to four years, gathering monthly data based on three interviews per year. I will focus on the most recent two panels: 1996 (4 years of interviews March February 2000) and 2001 (three years of interviews ). In both data sets, I restrict the sample to a group of women who are likely to be affected by Medicaid policy: single mothers ages with no more than a high school degree. I merge each of these data sets with the appropriate Medicaid thresholds for each mother s state of residence (by month and family size). Gathering data on monthly Medicaid thresholds for every state over several years is a surprisingly difficult task. There have been a few periodic surveys done by private foundations (such as the Kaiser foundation) during this time period, and any additional state data must be gathered on a state-by-state basis from the Center for Medicare and Medicaid services and other sources, all of which are listed in the Appendix. None of these sources list thresholds for families of any size except three; however, limiting the analysis to this family size would be overly restrictive. Therefore, three tactics are used for establishing the thresholds for families of two and four. First, Medicaid thresholds for any state with thresholds tied to the federal poverty line can be appropriately established for families of two and four using the poverty lines for these family sizes. Second, any state continuing to use the original AFDC limits can be investigated via the Urban Institute s Welfare Rules Database (WRD), which lists AFDC payment standards separately by family size. Finally, those states that have more unusual rules for determining Medicaid thresholds can be dealt with on a case-by-case basis, often involving the use of the WRD and state-specific websites. The final result is close to complete coverage of every state-month from 1996 through Table 1 shows a condensed version of the finalized earnings threshold data. The range of thresholds is perhaps surprisingly large, ranging from $254 per month in Louisiana to 6

7 $2544 per month in Washington, D.C. Note that while some states do not change Medicaid earnings thresholds during this time period, several have substantial changes. My final samples for the CPS and SIPP analysis include single mothers with 1, 2, or 3 children for whom state of residence is identifiable and state Medicaid threshold is available and for whom earnings are not imputed. 7 Some key descriptive statistics for the two samples are given in Table 2. The two samples have very similar characteristics. Average earnings and Medicaid thresholds are a bit lower in the SIPP, perhaps because the SIPP sample contains only 41 states (vs. the 47 available for the CPS) and these states have above-average Medicaid thresholds. The SIPP sample is also slightly older and less urban. All of the covariates in the table will be controlled for in the analysis to follow, which begins with the CPS analysis of earnings levels, followed by a comparison to earnings levels in the SIPP and an analysis of earnings growth. IV. Earnings Levels: Results of Kernel Density and Regression Estimation using CPS I approach my first research question (on earnings levels) with the following hypothesis: women who would typically have earnings near the Medicaid earnings limit may intentionally avoid going over this limit, either by avoiding increased hours of work or by choosing not to advance to jobs with higher wages. If this is the case, the data should reflect a disproportionate number of women with earnings just below the limit and proportionally fewer women with earnings just above the limit. To investigate this directly, I use the CPS data to generate a simple kernel density estimate of monthly earnings relative to Medicaid thresholds (in all states) for single mothers. I begin by using each mother s earnings and applicable Medicaid threshold to establish her distance from the threshold. Earnings are estimated using the woman s report of earnings last week at main job multiplied by the average number of weeks in a month (4.33). While it is possible that some women have more than one job (and thus higher total earnings), only a small percentage of the sample reports that their work hours last week at all jobs were much higher than their 7 The SIPP data do not separately identify the state for those living in ME, VT, ND, SD, and WY, but instead group them together. Since Medicaid thresholds vary by state, these states are dropped from the SIPP sample. I drop OR and MN from both data sets due to their implementation of state health insurance programs in the early 1990s that complicate the role of Medicaid. I also drop MO and NE due to legal conflicts that led to odd time patterns in their thresholds. The remaining missing state-months in the Medicaid data are all in ND, SD, and VT. These states are already eliminated from the SIPP analysis by necessity, and are also eliminated from the CPS analysis for those months that are missing. The sample restrictions also result in only one observation available from HI in the SIPP, so it is dropped as well. 7

8 usual work hours at their main job. 8 The distribution of the difference between estimated earnings and the Medicaid threshold is reported in a kernel density. This allows an examination of bunching below the threshold even across states with different (and changing) thresholds, since the earnings are normalized relative to the relevant threshold. However, identifying this bunching may be difficult since the shape of the underlying distribution near the threshold may differ across states; for instance, states with a very low threshold are likely on an upward-sloping portion of the earnings distribution in the region of the threshold, so that the bunching may be less apparent than in a flatter distribution. To help distinguish the patterns of bunching in different parts of the distribution, I divide the observations into two groups: those with sample median earnings of more than twice the state Medicaid threshold in a given month (where the threshold is likely on an upward-sloping part of the earnings distribution) and those with sample median earnings less than twice the threshold (more likely on a downward-sloping part of the earnings distribution near the threshold). To maintain a large enough data set near the threshold, I do not restrict the education levels of the single mothers in this sample. The kernel density estimates of these distributions are shown in Figures 1a and 1b, which use data from all working single mothers whose monthly income was within $400 of the monthly Medicaid threshold. The vertical line on each figure indicates the distinction between those with earnings below the threshold (on the left) and above the threshold (on the right). The prediction of bunching below the threshold is not strongly supported here, particularly in the states with relatively high thresholds. While one could argue that the distribution is consistent with a small effect for states with relatively low thresholds, it is also clear that there are many irregularities in the distribution unrelated to the Medicaid threshold. In any case, this very basic analysis is hardly definitive; it simply suggests that a more detailed statistical analysis of the hypothesis may be interesting. To improve the analysis of earnings levels with the CPS, I continue with a more careful examination of the possible role of the earnings limits including controls for other determinants of earnings. Because earnings distributions tend to be skewed, I use the log of individual earnings as the dependent variable. 9 The econometric model is as follows: 8 Less than 6 percent of women in the sample report have worked more than 10 additional hours relative to their usual hours at the main job. Some of these were likely working extra hours at their main job, but some may have second jobs. Unfortunately, the MORG data do not provide earnings last week at all jobs, which would be ideal, so I proxy for this with earnings at the main job. 9 I use the variable pternwa which is total wage and salary earnings for the individual. Mothers without earnings are not included, since the focus here is on earnings changes on the intensive margin. Box-Cox estimates support the use of 8

9 ln(earnings) ist = α s + φ t + β*x ist + γ*medthresh st + ε ist where i indicates an individual, s indicates a state, and t indicates a time period. Earnings ist is the measure of earnings already described, α s is a state-specific determinant of earnings (the same in all months), φ t is a time-specific determinant of earnings (the same across all individuals in a given month), X ist is a set of individual characteristics relevant to earnings, MedThresh st is the Medicaid threshold in state s in month t, and ε ist is an error term representing the remaining unexplained variation in earnings. This approach is similar to that taken by Yelowitz (1995) and Ham and Shore- Sheppard (2005) in their modeling of employment and welfare participation. Some of the past literature would suggest that γ > 0, i.e. higher Medicaid thresholds should be associated with higher earnings because they allow continued Medicaid coverage to be maintained at higher earnings levels. However, this framework requires the assumption that the Medicaid threshold affects workers in the same way regardless of their location in the earnings distribution. 10 Theory would predict that workers far above the threshold (and perhaps far below it) would be relatively unaffected by it, while those near the threshold are likely to take it into account when making hours/earnings decisions. I address this further in the earnings growth analysis; here, I simply avoid including some of the workers far above the threshold by restricting my regression sample to women with no more than a high school degree. I use individuals in all of the states listed in Table 1 to assemble the sample of unmarried female household heads for a cross-sectional regression analysis modeling the log of earnings levels. The regression includes controls for the number of children under 6, marital status (never married, widowed, or divorced), age, age squared, education level, race, and urban/rural residence. I also include fixed effects for each state (and the District of Columbia) represented in the sample. This is important because there may be a spurious correlation between earnings and Medicaid thresholds within states even if there is no causal relationship. Finally, I include fixed effects for each calendar year and each calendar month. The results are reported in Table 3. the logged dependent variable with the level of Medicaid threshold as an independent variable (resulting in an estimate of a partial elasticity). 10 It also does not fully incorporate the incentives of higher-earning poor workers, who may reduce earnings upon an increase in the Medicaid threshold in order to become eligible (once it requires only a small reduction to do so). If there are many of these workers, it will (appropriately) decrease the estimated value of γ. This estimate is thus a net effect and not simply the effect on workers below the threshold. The earnings growth model estimated later in the paper addresses this issue. 9

10 The regression results suggest that the Medicaid threshold may play a small role in affecting wages. The coefficient of.006 is statistically significant at the 95 percent level, suggesting that a $100 increase in the monthly Medicaid threshold is associated with roughly a 0.6 percent increase in monthly earnings. The other coefficients suggest that the regression specification is reasonable. For instance, education is positively correlated with earnings and additional children are correlated with lower earnings (perhaps because fewer hours are worked). Racial minorities have slightly lower wages than whites, which is also consistent with the literature. The positive coefficient on age and negative coefficient on age squared indicate a concave age-earnings profile. There are no unusual results that would suggest the regression has fundamental flaws. This result is somewhat sensitive to some reasonable adjustments in the specification. Specifications using lagged levels of the Medicaid threshold in place of the concurrent threshold (in case it takes some time for workers to learn of policy changes) fail to produce a statistically significant effect of the policy change when I use a one-month or three-month lagged threshold. However, using a six-month threshold suggests about the same effect and significance as the reported (no-lag) results. All of the lagged coefficients, regardless of the length of the lag, fall within the confidence interval of the reported point estimate on the concurrent threshold. In addition, results using year-by-month indicators (instead of year and month indicators) and estimates maintaining the use of imputed earnings both continue to produce small positive, statistically significant estimates of the effect of the Medicaid threshold. However, if I check robustness to outliers by eliminating the top 1% of earners from the sample, the estimated effect is similar in magnitude but no longer statistically significant. In addition, adjusting standard errors for multiple observations per individual reduces the t-statistic on the estimated effect to just Finally, I check whether higher education levels may increase threshold responsiveness by interacting the education dummy variable (which indicates a high school graduate rather than high school dropout) with the Medicaid threshold. The interaction term is not statistically significant (suggesting little role for education in workers responses to the Medicaid threshold), but adding this variable reduces the size and significance of the overall estimated effect. 11 Individuals are in the CPS during two consecutive years, so some may appear in my data twice. I use the most straightforward matching algorithm to identify individuals who appear in two consecutive years, taking advantage of CPS household and person identifiers that when combined should identify individuals. Based on this matching algorithm, only about ¼ of the individuals appear in the sample twice. Another ¼ are unmatched by construction due to being in the early or late part of the sample (so that their other interview is not included in my time period of interest). The remaining sample appears to consist of separate individuals, although there may be some missed matches due to coding problems in the CPS identifiers. Since the individual-level clustered standard errors made little difference in the results, I move on to the SIPP data - which is truly longitudinal - rather than further investigating this matching issue. 10

11 These regression results suggest that an increase in one s state Medicaid income limit may have a positive influence on earnings levels, a result which is consistent with the pattern in the kernel density analysis, but the estimate is fairly small and quite fragile. In addition, the lack of longitudinal data on these workers prevents me from controlling for unobserved heterogeneity that may bias the estimates. These limitations suggest that a more thorough examination with panel data would be valuable. 12 V. Earnings Levels: Comparing SIPP and CPS Estimates To corroborate the results on earnings levels obtained with the CPS, I perform the same analysis as in the previous section (using only state fixed effects) with my other data set, the SIPP. I follow this with a richer analysis using individual fixed effects to see whether individual heterogeneity may be contaminating the earlier results. I extend this analysis by conditioning on additional state-level policy variables. After comparing these to the CPS results, I move on to addressing the effects of the Medicaid earnings limit on earnings growth rather than earnings levels. My sample is defined as unmarried mothers aged with one, two, or three children under age 18. I also limit the sample to women who, at some point in their 3- or 4-year panel, had no more than a high school degree. These women must be the designated parent/guardian of children in the household and they must appear at least twice in the survey to be in my sample. Although my SIPP sample is smaller than the CPS sample used earlier, it has the advantage of repeated interviews with the same people. Each SIPP interview includes separate questions about each of the last 4 months, but I use only information provided for the most recent month to avoid both recall bias and seam bias. 13 I use the woman s own earnings as the measure of earnings, and in this case the measure is monthly (rather than weekly as in the CPS). The first column of Table 4 shows the results of SIPP estimation in which I use the same specification as the CPS regression in Table 3. While many of the coefficients are very similar to those estimated with the CPS data, and the explanatory power of the regression is similar, the Medicaid coefficient is tiny (about 1/3 the size of that estimated with the CPS) and statistically 12 I did much of the same CPS analysis using the annual earnings measures in the March CPS rather than the CPS MORG. Results were less precisely estimated, which may be due to using an annual measure of earnings and Medicaid eligibility. The CPS MORG is a better sample in this respect, as well as being much larger, which is why I exclude the March CPS estimates here. 13 Seam bias is the estimation problem that results from individuals commonly reporting their information in 4-month blocks, such that transitions (ex. between jobs) appear to happen disproportionately on the seam between surveys. The use of just the most recent month from each interview is a standard method of dealing with this measurement problem. 11

12 insignificant. When I include fixed effects at the individual level, I am able to control for timeinvariant individual heterogeneity. The estimated Medicaid effect, shown in the second column of Table 4, remains small and imprecisely estimated. The other coefficients change in ways that would be expected (for example, the education effect ceases to be positive and significant when individual fixed-effects are added to the model, which would be consistent with ability bias). Because other policies for low-income workers were also changing during the time of my sample, earnings may be affected by policies beyond the Medicaid threshold. I include several policy variables in the last column of Table 4. One relevant variable is the level of welfare benefits, which varies substantially across states. 14 Another is the threshold for children s medical insurance coverage. Children s public health insurance can be provided through the Medicaid program directly (often with expansions in eligibility targeted toward children), through the State Children s Health Insurance Program, or through some combination of the two. I include the value of the income eligibility threshold for coverage of the youngest child in the family. Benefits vary for infants (age 0), small children (ages 1-5), and older children (ages 6-18), so families experience changes in this threshold as their children age as well as when policies change. 15 Finally, I include the minimum wage that applies in each state for each year, the federal maximum Earned Income Tax Credit by family size and year (which is effectively a wage subsidy), and an indicator for the availability of a state Earned Income Tax Credit each year. When these variables are included in the individual fixed-effect regression, only the minimum wage is statistically significant, and I continue to find no statistically significant effect of the Medicaid threshold. This finding of no effect is robust to adding an education interaction, including imputed values, or using year-by-month indicators. VI. Earnings Growth: Examining Heterogeneous Effects using SIPP I approach my second research question (on earnings growth) with the hypothesis that increases in Medicaid earnings limits may affect earnings growth in systematically different ways across workers. Specifically, it may be difficult to identify the effects of Medicaid on earnings levels if different workers experience varying labor incentives at the time of a Medicaid expansion. Any effect on earnings levels (or lack thereof) will reflect only an average effect, which may be of limited interest in a situation with a great deal of heterogeneity. For example, for some workers, 14 I use the maximum benefit, by year, by family size. These data, along with those on state minimum wage levels and state EITC programs, were provided by the University of Kentucky Center for Poverty Research. 15 Threshold values are assigned month-by-month, and were assembled using a variety of sources available from the author upon request. 12

13 earnings growth that had been stunted by the previous limit may be free to increase upon the relaxing of the limit. However, earnings may stagnate or even decline for those workers who were above the old limit but are now qualified or will qualify with a small earnings reduction for Medicaid. Since these effects are in opposite directions for two potentially large groups of workers, panel data are necessary in order to distinguish the effects for each group rather than pool together these effects (as in the above analysis). Figure 2 illustrates these incentives in the usual labor supply context. Note that the change in the budget constraint reflects an increase in a Medicaid threshold that was already above the welfare threshold. I illustrate this particular scenario for two reasons. First, the discussion of an initial expansion is already carefully described in Ham and Shore-Sheppard (2005), and motivates their focus on increased labor force participation as a possible response to early Medicaid expansions (though, interestingly, they do not find evidence of such a response). Second, modeling a further expansion such as that shown is increasingly relevant, particularly with the recent data I use in this study. During the time period of 1996 through 2004, 11 states in my sample had no change in their threshold, 13 had exactly one change, and 18 had multiple changes. As states continue to increase thresholds, it is appropriate to consider the incentives created by this type of further expansion. Perhaps the most important difference between modeling an initial expansion and a further expansion is the resulting difference in the incentives for non-workers. In Ham and Shore- Sheppard s model, the initial expansion creates an incentive to join the labor force; in my model of a further expansion, there is no change in incentives for non-workers. This follows from the lack of an income effect (as is always the case at the extensive margin) and because there is no change in the price of leisure, since it was already possible to work for a positive effective wage and receive Medicaid (in the region beyond welfare). In other words, any worker who would be interested in working at some level above the initial Medicaid threshold (and below the new one) would have already been in the labor force, with earnings near the initial threshold. Similarly, I do not expect any effect of an expansion on people who were in the labor force but already choosing to earn well below the Medicaid limit; if the expansion were to make them more interested in earning beyond the initial threshold, it suggests that they were bound by it, in which case we would expect their initial earnings to be fairly near the limit One might argue that these workers do not choose their earnings, in the sense that their wage is not their choice. However, in the range of workers with positive earnings meaningfully below the Medicaid threshold, low earnings are at least partly driven by few hours worked. (Full-time work would put people near or above the Medicaid threshold in most states, even at low wages.) 13

14 My model predicts that the expansion will increase hours of work only for those near the initial earnings threshold, i.e. who were bound by the initial constraint. Those with initial earnings between the old and new thresholds will have an incentive to decrease hours of work (a pure income effect) and those with initial earnings above both the old and new thresholds may also benefit from reducing hours to obtain eligibility. However, those whose incomes far exceed even the higher threshold are unlikely to gain utility by substantially reducing hours in order to obtain Medicaid coverage. Like those at the bottom of the earnings distribution, there is no reason to expect an effect of Medicaid threshold changes on high-earning workers. To check the appropriateness of my focus on the intensive margin, where my model suggests all response to the threshold will occur, I first verify that there is minimal response on the extensive (labor force participation) margin. Since there was no such response found by Ham and Shore-Sheppard (2005), who were using data from a time when most expansions did provide incentives on that margin, it would be very surprising for me to find an effect on the extensive margin. Nevertheless, if I did find such an effect in the more recent SIPP data, it might suggest a mistaken focus on those with positive earnings. When I estimate the effects of the Medicaid expansions on the extensive margin (an employment indicator) using the final fixed-effect regression specification from the previous section, I find a statistically insignificant coefficient so small that its confidence interval rules out any economically meaningful effects. 17 This confirms that my focus on the intensive margin is appropriate. I proceed with two distinct approaches to estimating the possibly heterogeneous effects of changes in the Medicaid threshold on earnings and hours. First, I suggest a very simple model in which changes in the threshold can have different effects on workers above and below the former threshold. I estimate two versions of this model, based on two ways of categorizing workers. Second, I extend the model to allow the size of the threshold effect to vary continuously depending on how close each worker is to the threshold, with the expectation that those workers nearest the threshold are the most likely to respond to changes. A. Simple Model with Heterogeneous Effects Clearly, a worker s initial earnings level has a strong effect on the incentives she faces when Medicaid earnings thresholds increase. Because of this I test the effects of the Medicaid threshold 17 The table of results is available from the author upon request. The upper-bound on the confidence interval for Medicaid effects would indicate an increase in labor force participation of about 0.5 percentage points for a $100 increase in the Medicaid threshold. The lower bound would suggest a 0.4 percentage point decrease in participation. 14

15 on earnings growth in a way that allows for different responses from different parts of the earnings distribution. I approach this with two different regression specifications, using all observations for which both the current and previous period have positive earnings. I also examine the change in hours worked across periods, which I will call hours growth to make clear the parallel with earnings growth. In the most basic specification, I simply allow for a change in the Medicaid threshold to generate different responses from people above and below the previous period s Medicaid threshold. Consider the following specification: Y it - Y i,t-1 = α + λ t + β*x it + γ 1 *((MedicaidThreshold it - MedicaidThreshold i,t-1 ) * I(MedicaidThreshold i,t-1 - Earnings i,t-1 > 0)) + γ 2 *((MedicaidThreshold it - MedicaidThreshold i,t-1 ) * I(MedicaidThreshold i,t-1 - Earnings i,t-1 0)) + δ * I(MedicaidThreshold i,t-1 - Earnings i,t-1 > 0) + ε it where I(.) is an indicator variable for a worker s location above or below the Medicaid threshold in the previous period. The outcome Y is either earnings or hours. As noted earlier, theory predicts a positive effect on earnings growth for those previously below the threshold, so I expect γ 1 > 0. For those above the threshold, there is some incentive to move below it, so I expect γ 2 < 0. If there are people with very low or very high earnings who are unaffected by the policy change, this will bias the coefficients toward zero, but the prediction of opposite signs remains reasonable. The model also allows for separate intercepts for the two groups (α and α+δ). The predictions are the same when I examine hours growth. A variation on this specification allows more heterogeneity in the effects of the Medicaid threshold on earnings growth. Instead of classifying workers simply as above or below the previous threshold (2 categories), I classify them into 4 categories far above, close but above, close but below, and far below. I define far here as $300, which is the difference in earnings that a person could have if she worked for 10 additional hours per week at about $7.00 per hour. (Results aren t substantively different for definitions of $200 or $400). The prediction in this case is for there to be no effect for the two groups that are far from the threshold, but the same predicted effects as before for those who are close above or close below. Results of this estimation are provided in Table 5. The first two columns of results relate to earnings growth, while the last two relate to hours growth. Predicted signs are shown on the left side of the table for ease of interpretation. Neither earnings specification provides evidence that 15

16 workers respond to changes in the Medicaid threshold in the expected way. Indeed, both of the key interaction variables have the wrong sign, in the sense that those below the former earnings limit are predicted to have decreases in earnings upon an expansion of the Medicaid threshold, while those above the former limit are predicted to have increases in earnings upon expansion of the threshold (though only the estimate for higher-earning workers is significant). Even the second specification, with the wider set of income categories, fails to provide support for the basic theoretical predictions. This puzzling result seems unlikely to reflect the true causal effect of Medicaid expansions, and may instead reflect a lack of sufficient additional covariates for establishing trajectories for these groups. It is interesting that the indicator for having been below the Medicaid threshold in the previous period is strongly positive and significant (as are its parallel set of indicators in the second specification). This seems to suggest that those with low observed earnings tend to see larger increases in earnings into the next month than those starting with higher observed earnings. This could partly be by construction, since whenever a new job begins anytime other than the first of the month, we will see low earnings with a large increase the following month. The more positive earnings growth of lower-earning workers may also reflect some sort of mean-reverting process; if low earnings reflect a negative temporary shock, we should not be surprised that lower earnings would predict high growth afterward. The estimated effects of changes in the Medicaid threshold on hours (rather than earnings) are a bit more consistent with the theory. The first specification (third column of results in Table 5) has the "right" signs on the Medicaid threshold interactions, but the magnitudes are tiny and statistically insignificant. The second specification has the expected pattern of negative estimates for those just above the threshold, and almost no effects for those far above the threshold, but none of the estimates are very precise or large. For example, the point estimate suggests that a $100 increase in the Medicaid threshold would reduce monthly hours worked by about two-thirds of an hour (0.68) for those near (but above) the threshold. The estimates for those under the threshold are practically zero. Given the weak nature of these results in terms of precision and total explanatory power (the overall R 2 never exceeds 4%), it would appear that a richer specification is needed. This richer specification can help establish whether the data are truly inconsistent with the basic theory of responsiveness typically utilized in this literature. 16

17 B. Model with Heterogeneous Effects by Absolute Distance from Threshold The previous model involved quite a coarse definition of who should be expected to respond to changes in the Medicaid threshold. In this section, I take a different approach. Instead of looking for the effect of the size of the change in threshold, I introduce an effect of distance from the new threshold on earnings growth. (This implicitly uses the threshold changes for identification, though the threshold change is no longer a variable in the model.) I include a set of squared terms as well, to allow different effects for those far from the threshold. These distance measures are interacted with an indicator for whether the individual's earnings would place her above or below the current threshold if she stayed at last period's earning level. Y it - Y i,t-1 = α + λ t + β*x it + γ 1 * MedicaidThreshold it - Earnings i,t-1 * I((MedicaidThreshold it - Earnings i,t-1 ) > 0) + γ 2 * MedicaidThreshold it - Earnings i,t-1 * I((MedicaidThreshold it - Earnings i,t-1 ) 0) + γ 3 *(MedicaidThreshold it - Earnings i,t-1 ) 2 * I((MedicaidThreshold it - Earnings i,t-1 ) > 0) + γ 4 *(MedicaidThreshold it - Earnings i,t-1 ) 2 * I((MedicaidThreshold it - Earnings i,t-1 ) 0) + ε it where I(.) is an indicator variable for whether the worker s past earnings exceed the current threshold. If the worker s earnings last period were below the current threshold, i.e. (MedicaidThreshold it - Earnings i,t-1 ) > 0, there is room for increasing earnings without losing Medicaid, which would imply γ 1 > 0. If they were above the threshold, the worker may potentially decrease earnings in order to get under the threshold, which suggests γ 2 < 0. While there is more space for movement toward the threshold for those more distant from it (hence the use of a distance measure), it would be surprising if the largest effects on earnings occurred for those furthest from the threshold, as would be implied by these linear terms. Indeed, theory suggests those very far from the threshold will have less sensitivity to it. The addition of two squared terms allows the distinction that must be made between two types of workers: (a) those who are less convinced to increase income because theirs is so low that the threshold isn t at all binding and (b) those who are less convinced to lower their income because theirs is so high they would not consider the drop needed to qualify for Medicaid. I expect γ 3 to be negative to counteract the positive term γ 1 that suggests increased earnings among the lowest earners. Similarly, I expect γ 4 to be positive to counteract the negative term γ 2 that indicates decreased earnings among the highest earners, since 17