Abstract
Beginning in the mid-1980s and extending through the early to mid-1990s, a substantial number of women and children in the United States gained eligibility for Medicaid through a series of income-based expansions. Using natality data from the National Center for Health Statistics, we estimate fertility responses to these eligibility expansions. We follow Currie and Gruber (2001) and measure changes in state Medicaid-eligibility policy by simulating the fraction of a standard population that would qualify for benefits in different states and different time periods. From 1985 to 1996, the fraction of women aged 15–44 who were eligible for Medicaid coverage for a pregnancy increased more than 20 percentage points. When we use a state and year fixed-effects model with a limited set of covariates, our estimates indicate that fertility increases in response to Medicaid expansions. However, after we include fixed effects for demographic characteristics, the estimated relationship diminishes substantially in size and is no longer statistically significant. We conclude that there is no robust relationship between Medicaid expansions and fertility.
Background
A substantial body of research investigates fertility responses to changes in social policy. This research is motivated by Becker’s early theoretical work suggesting that policies that lower the “price” of a child can lead to an increase in the total demand for children (Becker 1960, 1991). Although social policies that reduce the cost of a child may make parenthood more affordable, Becker and Lewis (1973) also acknowledged that there may be a quality-quantity tradeoff. Any policy that reduces the cost of a child will also increase parental resources available for investment in current children; thus, parents may elect to invest more in the children they do have rather than having additional children. Therefore, the change in the total demand for children created by a subsidy or an in-kind benefit for parents is often theoretically ambiguous.
The Medicaid program provides health insurance coverage to low-income women and their children, a group that constitutes roughly two-thirds of the Medicaid population (Gruber 2003). Although this group is a relatively healthy population compared with other populations that Medicaid covers, the program provides generous coverage for medical expenses associated with childbearing (prenatal, delivery, and postnatal care) and childhood health services that could otherwise represent substantial costs for low-income families. Medicaid has the potential to greatly reduce the financial costs associated with childbirth and, to a lesser extent, to reduce the financial cost of being a parent. Average prenatal care/delivery and health care expenses in the first year of a child’s life cost roughly $12,000 (based on Lewit and Monheit 1992 estimates and inflated to 2006 dollars); and because Medicaid generally prohibits cost sharing for children and pregnant women (Tritz et al. 2006), most, if not all, of the expenses would be covered by the Medicaid program.Moreover, Lino (2000) estimated that the direct economic cost of a low-income married couple’s child during his or her first 18 years is approximately $200,000 (in 2006 dollars). Thus, by some estimates, Medicaid covers an amount equal to at least 6% of the direct financial economic costs of having a child just by covering health care costs related to childbirth.1
Over time, an increasing number of families have become eligible to receive health insurance coverage for some or all of their members through the Medicaid program because of federal and state policy changes. Medicaid was originally designed to provide health care for single-parent families eligible for the Aid to Families with Dependent Children (AFDC) program and for low-income blind, elderly, and disabled individuals. The program underwent a series of important changes that are detailed elsewhere (e.g., Currie and Gruber 1996; Gruber 2003) beginning with the Deficit Reduction Act (DEFRA) of 1984, effective in 1985, when first-time pregnant women were covered by Medicaid if they were to qualify for AFDC, counting the fetus as a child. Although there were some smaller expansions of Medicaid prior to 1985, these were fairly minor (see Currie and Gruber 1996: Figure 2). Through further expansions in the Consolidated Omnibus Reconciliation Act (COBRA) of 1985, pregnant women who did not qualify for AFDC but who had similar financial circumstances were made eligible. The Omnibus Budget Reconciliation Act (OBRA) of 1986 permitted states to expand eligibility for children and pregnant women with incomes up to 100% of the federal poverty limit (FPL). The OBRA of 1987 allowed states to expand eligibility further to 185% of the FPL, and the OBRA of 1989 mandated that states expand eligibility to 133% of the FPL for pregnant women and children aged 5 or younger. Thus, by 1992, pregnant women and children aged 5 or younger with incomes below 133% of the poverty line qualified for the program; and in some states—such as California, Michigan, and Texas—the eligibility threshold was substantially higher. In total, from 1987 to 1992, both the number of children aged 18 or younger and the number of women between the ages of 15 and 44 who were eligible for Medicaid more than doubled (Cutler and Gruber 1996). Between 1993 and 1997, the federal government approved a series of waivers that allowed many states to further expand income eligibility for their Medicaid programs.
This article contributes to our understanding of the relationship between fertility and social welfare policies, specifically the Medicaid program. We use this expansionary period during the late 1980s and 1990s to estimate the relationship between Medicaid eligibility and fertility. As explained earlier, the expected fertility response created by an increase in eligibility for publicly provided health insurance coverage is theoretically ambiguous. In the case of Medicaid expansions, understanding the relationship is complicated further because Medicaid also subsidizes the use of contraception. A 1972 amendment to Medicaid required states to provide family planning services and supplies for qualified women of childbearing ages who desire such services.2 In fact, the federal government reimburses states for 90% of their costs for Medicaid family planning services (Gold et al. 2007), which is much higher than the 50%–75% federal reimbursement provided for the regular health insurance component of the Medicaid program. Given that Medicaid family planning services reduce fertility (Mellor 1998), one might be less likely to expect Medicaid expansions to increase fertility than otherwise.
Using birth data from the National Center for Health Statistics and a measure of state Medicaid eligibility generosity with a state and time fixed-effects model, our estimates indicate that Medicaid eligibility is positively related to births. This specification is similar to that which is used in much of the empirical literature investigating social policy and fertility. However, after we include fixed effects for demographic characteristics in our analysis, the size of the coefficient diminishes, the standard errors increase, and our point estimates are no longer statistically significant. We conclude that there is no robust relationship between Medicaid expansions and fertility.
Literature Review
A large literature has examined whether there is a fertility response to the financial incentives created by a variety of social programs in the United States. In perhaps the most prominent branch of this literature, researchers examined whether the AFDC program, now the Temporary Assistance for Needy Families (TANF) program, induces low-income women to have children. Several researchers have found positive effects of welfare policies on fertility (e.g., Horvath-Rose and Peters 2001; Kaestner et al. 2003; Lopoo and DeLeire 2006; Lundberg and Plotnick 1995), yet others find no relationship (e.g., An et al. 1993; Duncan and Hoffman 1990; Hao and Cherlin 2004; Hoffman and Foster 2000; Hoynes 1997; Kearney 2004; Mayer 1997). In summarizing this expansive literature, Moffitt (2003) wrote, “. . . although there is still considerable uncertainty in the literature and there remains a large number of studies reporting insignificant estimates, this reading of the literature leads to the conclusion that welfare is likely to have some effect on family structure” (p. 336).
Other researchers have focused on the effects of tax policy on the likelihood of childbearing. Whittington et al. (1990) used time series data from 1913 to 1984 and found that the magnitude of the personal exemption in the U.S. Internal Revenue Code is positively related to birth rates. Similarly, Dickert-Conlin and Chandra (1999) examined whether families expecting the birth of a child near the end of the year are more likely to manipulate their date of delivery (through inductions or Caesarian sections) to gain the income tax savings of a personal exemption. They show that those with the greatest potential for tax savings are the most likely to engage in strategic timing of their births.
The literature most closely related to the present research project is one that has documented how fertility-related decisions respond to having health insurance that covers the financial costs associated with pregnancy and childbirth. For one, the RAND Health Insurance Experiments showed that women who received free health care services were much more likely to have a birth than women who had to pay for at least a portion of their health care expenses (Leibowitz 1990). In addition, when states mandate the coverage of infertility treatments under private health insurance contracts, the use of infertility treatments and fertility increase (Bitler 2006; Bitler and Schmidt 2006; Bundorf et al. 2007; Hamilton and McManus 2005; Schmidt 2007).
To date, there has been very little research investigating the effect of Medicaid expansions on fertility. In the earliest work on this topic, Joyce et al. (1998) used data on birth rates over time among unmarried women aged 19–27 with 12 or fewer years of education from 15 states. They examined whether birth rates changed following several legislative changes that expanded Medicaid. Joyce et al. (1998) found that the expansions were associated with a 5% increase in birth rates for white women and no change for African American women. In a recent article, Bitler and Zavodny (2010) used natality data and found that Medicaid expansions were not consistently associated with birth rates in general, but they did find some evidence of a positive effect on birth rates among white females with less than a high school education.
Researchers have also investigated whether restrictions on Medicaid funding for abortions has influenced abortion and birth rates. Theoretically, one should expect reductions in funding for abortion to decrease the number of abortions, and empirical research seems to confirm this (Blank et al. 1996; Kane and Staiger 1996; Levine 2004; Levine et al. 1996). At the same time, the effect on births is ambiguous. Although a reduction in abortions should increase fertility rates when all else is held equal, one might anticipate that women will respond to these policy changes by increasing efforts at avoiding pregnancy (Levine et al. 1996). Empirical research on this question provides little evidence that birth rates fall in response to Medicaid funding reductions (Kane and Staiger 1996; Levine 2004; Levine et al. 1996). Another specific feature of the Medicaid program that two recent papers examined is the effect of Medicaid waivers for family planning services for women not eligible for regular Medicaid services. Lindrooth and McCullough (2007) and Kearney and Levine (2009) both found that extending family planning services reduces fertility. Kearney and Levine (2009) found that the waivers increased access to contraception and reduced teen births by about 4% and nonteen births by 2%.
A literature related to the current topic examines the impact of the expansions in Medicaid eligibility on health insurance coverage. Cutler and Gruber (1996) found that increases in Medicaid eligibility increased Medicaid coverage, but these increases were offset somewhat by reductions in private coverage. This estimate of “crowd-out” spurred a large literature that uses a variety of methods and data sets (e.g., Card and Shore-Sheppard 2004; Dubay and Kenney 1997; Yazici and Kaestner 2000). Estimates of crowd-out have been found to be sensitive to specification, particularly to how those children with both Medicaid and private coverage are treated. More recent research investigating the effect of State Child Health Insurance program (SCHIP) expansions has found large amounts of crowd-out and relatively low rates of take-up (e.g., Gruber and Simon 2008; Lo Sasso and Buchmueller 2004).3
Data and Methods
Data
We use a variety of data sources for our analysis. For birth counts, we use the NCHS natality data series from 1985 to 1997. Data in the NCHS natality series are compiled from birth certificates through the Vital Statistics Cooperative Program, which guarantees some uniformity in the information collected. Each state submits data electronically to the NCHS, and the resulting annual file contains a record for nearly all births that occur within the United States. The NCHS natality series also reports all information available on the U.S. Standard Certificate of Live Birth, including the mother’s age, race/ethnicity, state of residence, education level, and marital status. We restrict our analysis to births to white and African American women in the United States between the ages of 15 and 44, ages that are typically considered a woman’s fertile period in the research literature.
For our data set construction, we group women into 44 unique demographic cells in each state, each year (identical groups as those used by Currie and Gruber 2001). For each race, we count the number of births to teens aged 15–18. For women aged 19 and older, we create three age categories—19–24, 25–34, and 35–44—and separate them further by four categories of mothers’ education: high school dropout, high school graduate, some college, and at least a college education. Finally, for mothers with at least a high school education, we stratify by marital status. This procedure yields 22 cells for white mothers and 22 cells for African American mothers.4
Our key policy measure, an index of Medicaid eligibility rules, varies by quarter, year, state, and demographic cell from 1985 to 1996. It essentially captures the fraction of a certain standardized population that becomes eligible due to policy changes alone, thus being a suitable index that boils down the complex Medicaid eligibility determination process into one single policy variable. To create this variable, we combine four years of the March Current Population Survey (1991–1994) to obtain adequately sized nationally representative samples in each of the 44 age by education by marital status demographic cells. We then calculate the fraction of the women in each of these 44 demographic cells who would be eligible for Medicaid in each state, each quarter of each year from 1985 to 1996 based on their incomes and the policies in place in that state in that quarter of that year (adjusting for inflation), similar to the process used in past research (see, e.g., Currie and Gruber 1996 for details). For example, we take the nationally representative sample of white women aged 19–24 who are high school dropouts and calculate the fraction that would be eligible for Medicaid if they lived in Alaska in each quarter of 1985, in Alaska in each quarter of 1986 (with inflation adjustments), and so on for every state and every quarter and year. The use of a national sample for each cell is to account for the fact that a given income threshold change may have different ramifications depending on the income distribution and family structure of that demographic group.
Our simulated measure of eligibility can be interpreted as an exogenous index of the expansiveness of Medicaid eligibility. This measure represents an improvement in the measurement of the Medicaid expansions used in the fertility literature, but follows a standard method used in other research in the Medicaid literature looking at insurance outcomes. By using this identification strategy, we capture more of the variation in Medicaid eligibility policy at the state-by-quarter-by-year level than we would if we used a before-and-after approach that captures only the average effect.5 The measure is also an improvement over simply comparing those who are eligible for Medicaid to those who are not eligible (which we could not do in this article in any case because the natality data do not contain information on income to determine actual eligibility for Medicaid). Actual Medicaid eligibility is likely to be endogenous because it is partially determined by unobserved individual and family characteristics (reflected in income) that may be correlated with the demand for children and because women may reduce their hours of work (and their earnings) when they choose to have children. By contrast, the measure of Medicaid expansions that we use is exogenous to these concerns, and varies by state by quarter by year only to the extent that policy changes (Gruber 2003).
Methods
Ideally, to determine whether Medicaid has an influence on fertility, we would take a population of low-income women and randomly assign them to a treatment group that would be eligible for Medicaid, including prenatal care and health care for the child, and to a control group that would not be eligible to receive these benefits (similar to the RAND experiment in Leibowitz 1990). We could then compare the fertility patterns of the treatment and control groups to determine the effect of Medicaid expansions on fertility.
In addition to the state, year, and quarter indicators, we also include cell fixed effects in Model 3, which means that we identify the Medicaid coefficient using variation within demographic cells over time. The addition of the cell fixed effects allows us to control for unobserved factors that are invariant within a cell that also might be correlated with Medicaid eligibility and fertility.
In Model 4, we add quarter-by-cell fixed effects to the Model 3 specification; in Model 5, we add state-by-quarter fixed effects to the Model 3 specification; and in Model 6, we add state-by-cell fixed effects to the Model 3 specification. Finally, in Model 7, we include state-by-year, cell, cell-by-quarter, state-by-quarter, and state-by-cell fixed effects to the Model 3 specification, resulting in a fully saturated model. We compare the results from Models 3–7 with those reported in Models 1 and 2.
Results
Between 1985 and 1997, eligibility for the Medicaid program increased substantially, with much of the increase occurring between 1987 and 1992 (see Fig. 1).9 Table 1 reports the estimates for Models 1–7 for all women aged 15–44 by race. Among white women, Model 1 suggests that a 1-percentage-point increase in Medicaid eligibility is associated with a statistically significant 1.2% increase in births, all else being equal. Among African Americans, the point estimate is twice as large, at 2.4%. Including state-by-year effects in Model 2 does little to alter point estimates of the coefficient for Medicaid eligibility.
Next, we add cell fixed effects to the model. This should remove any bias created by unobserved demographic factors that are correlated with both Medicaid expansions and the fertility patterns across demographic cells. After the cell fixed effects are added in Model 3, the coefficient estimates decline considerably in size and are no longer statistically distinguishable from zero for white and African American women. In fact, the point estimates are nearly zero. In Model 4, we add cell-by-quarter fixed effects to the Model 3 specification. In Model 5, we add state-by-quarter fixed effects to the Model 3 specification. In neither instance does this alter the point estimate for Medicaid expansion. In Model 6, we include cell-by-state fixed effects to the Model 3 specification. We include this specification to allow for demographic differences by state. For example, the fertility rate of 25- to 34-year-old, unmarried, white women with less than a high school education may be different in Massachusetts than it is in Louisiana. We see no change in Model 6 or in Model 7, the fully saturated model, suggesting that after we include the cell fixed effects, there is little evidence of a fertility response to Medicaid expansions. This result is quite different from what we found by using the model specification (Model 1 and Model 2) that one often finds in the social science/policy literature on fertility responses to policy changes. It is also worth pointing out that the patterns we observe in Table 1 are repeated throughout this article regardless of sample construction and time period observed.
Given the changes that we observe over time in Medicaid eligibility for teenage women and high school dropouts in Figs. 2 and 3 in Online Resource 1, we next report results limited to this subsample of females in Table 2. In Models 1 and 2, we find that a 1-percentage-point increase in Medicaid eligibility is associated with a 1.2% or 1.3% increase in births for white women. However, we find trivial and statistically insignificant estimates for the African American women. After we control for cell fixed effects (including all Models 3–7), the point estimates for the white teens and high school dropouts decline by more than one-half and are no longer statistically significant. The coefficients estimates for Medicaid expansions for the African American women remain trivial in size and statistically insignificant for all models.
Tests of Robustness
We made a number of choices in our empirical analyses, particularly pertaining to the data used. If our findings are the result of these choices, one should be concerned about their validity. In this section, we rerun our models by using a variety of different samples with different decision rules governing data construction to determine whether the results reported in the previous section are robust. More specifically, we investigate the importance of our choice of log births as the dependent variable, the influence of missing data, Medicaid’s impact on family planning, and our choice of sample weights.
Log Births
Several demographic cells had no births in a given state in a given quarter of a given year. Because we take the natural log of births, the outcomes in these cells are undefined. Rather than dropping these cases as in the earlier studies, we assigned these cells a value of 1, implying a logged value of 0. To determine whether this choice influenced our results, we reran all the models, dropping the cells with zero births. The results (available upon request) are nearly identical to those reported earlier.
Missing Data
Although the natality data from the NCHS has many benefits, one of its weaknesses is missing data among some of the maternal characteristics that we used to define the demographic cells.10 In Table 3, we illustrate the impact missing data could potentially have on our results. The first column of Table 3 shows the number of births reported in the United States for women aged 15–44 each year from 1985 to 1997. Beginning in 1985, the number of births was increasing every year until 1990 when they began to fall until 1997, when our time series ends. The natality data have incomplete information on the mother’s education, race, and marital status, particularly from 1985 to 1988. For example, of the 3.7 million births in 1985, roughly 887,000 of them were missing the mother’s education value, about 12,000 were missing the race value, and around 1,200 were missing information on the marital status of the mother. Given all this missing information, we had complete information for only 2,859,632 observations in 1985 when we constructed our data cells.
One can also see several discontinuities in the time series for the educational attainment variable: in particular, one should notice a change in quantity of education values missing between the 1988 and 1989 series and between the 1991 and 1992 series. In 1989, the NCHS also began to impute race values as well as marital status. To maintain our sample size as closely as possible to the total births observed, we use the imputed values in our analyses.
Even with the imputed values, one observes a different trend in the actual births (Total Births column) from the trend in births observed among our complete cases (Total Cases with Complete Information column). Given this differential pattern, one might be concerned that the missing data is contributing to our findings. To address the importance of this issue, we ran a series of checks to determine the importance of missing data.
Restricting the Period Investigated
The period from 1985 to 1988 is differentially impacted by the missing data on the maternal characteristics, particularly maternal education. Therefore, in Table 4, we estimate the relationship between Medicaid expansions and births using data that commences in 1989 rather than 1985—that is, removing the part of the series that had the highest proportion of missing data. The removal of the first three years does little to alter our findings. Models 1 and 2 suggest that a 1-percentage-point increase in the proportion eligible for Medicaid is associated with a 1.2% increase in births for white mothers and a 2.4% increase for African American mothers. After we add cell fixed effects, however, the point estimates decline considerably and are no longer statistically significant. We see a similar pattern across all models, using a subsample composed of teens, high school dropouts of all ages, and unmarried women (see Table 10 in Online Resource 1) as well as teens and high school dropouts alone (Table 11, Online Resource 1).
The disadvantage of using only part of the time series is that a substantial proportion of the Medicaid expansions occurred prior to 1989. By using the post-1988 information, we are potentially losing important variation in our models. Figure 1 shows the annual proportion of white and African American women, collectively, who were eligible for Medicaid based on our simulated Medicaid eligibility measure. From 1985 to 1989, the eligibility index increased nearly 60% from 12.8% to 20.5%. From 1989 to 1996, the index increased from 20.5% to 35.4%, or by 72.7%. Thus, while using this portion of the time series is less than optimal, there is still considerable variation in Medicaid eligibility.
Balanced Panels Collapsing by Race, Marital Status, and Education
Next, we include all data from 1985 through 1997 and reconstruct our cell definitions to ignore some of the characteristics that are missing at high rates to incorporate a larger number of births in our analytic sample. For example, in Table 5, we continue to break up our cells by age, marital status, and education, but we no longer distinguish birth by racial categories. In other words, we pool both white and African Americans into one sample. By collapsing the racial categories into one group, we now have 22 cells instead of the 44 we used earlier (22 cells for whites and 22 cells for African Americans). Again, the advantage of this tactic is that missing data on the race of the individual will not preclude including that observation in our data.
After we collapse the data by race category, we find results consistent with our earlier findings: a 1-percentage-point increase in Medicaid eligibility is associated with a statistically significant 1.4% increase in births (see Models 1 and 2 in Table 5). After we include the cell fixed effects, however, the point estimates decline in magnitude, become negative, and are no longer statistically significant. Given the pattern observed in Tables 1 and 2 and that there are relatively few observations missing data on race, this finding is not surprising.
In Table 6, we report results using cells that do not separate married and unmarried mothers.11 In this series of models, the 13 cells are constructed based on the mother’s age and education level, and we report results separately for white and African American mothers. Among white mothers, although the point estimate for Medicaid eligibility is positive in Models 1 and 2, it is not statistically distinguishable from zero, and it changes very little in Models 3–7. Among African Americans, however, we see the same pattern observed before. In Models 1 and 2, the point estimate, 1.5%, is statistically significant. After we include cell fixed effects, however, the coefficients become less positive, and none is statistically significant.
In Table 7, we collapse the marital status categories as well as the education categories, but continue to show separate results by race. In these models, there are four cells for each racial group consisting of the four age categories: younger than 19, 19 to 24, 25 to 34, and 35 and older. Among white mothers, we see large and statistically significant estimates in Model 1 and Model 2. A 1-percentage-point increase in Medicaid eligibility is associated with a 5.6% increase in births in Model 1 and a 9.8% increase in Model 2. After we include the cell fixed effects, we find a much smaller (in magnitude), negative, and statistically significant result. For African American women, we find a positive and statistically significant result in Model 1 and Model 2. After we control for cell fixed effects, the coefficients estimates become much smaller, negative, and for Models 3–5, statistically significant. When we identify the Medicaid coefficient utilizing variation within a cell-state (Model 6), the effects are no longer statistically significant.
The results in Table 7 are quite different from the findings we have seen previously. The samples in Table 7 have the benefit of little data loss due to missing information on marital status and completed education for the mothers. At the same time, these models capture only the broad patterns of the relationship between births and Medicaid eligibility; that is, the models partition the data only by race and age and fail to account for some of the heterogeneity among potential Medicaid recipients. If this new result is due to the inclusion of the missing data and not our inability to account for marital status and education, then we have some evidence that Medicaid expansions may reduce fertility among the white population.
To investigate this issue in more detail, in Table 8, we report three sets of results for white women and three sets of results for African American women.12 Optimally, we would like to know the relationship between Medicaid expansions and fertility without any missing data, while including the demographic cell fixed effects. Unfortunately, this option is infeasible. In the first row of the panel for white women, we report results using the original data set (with all of the missing information) and use the same cell definitions as used in Table 7. In other words, the cell fixed effects are based on age alone. This set of results suffers from both missing data and does not include measures for the marital status and education of the mother. Although we cannot fix the missing data and include the demographic cells simultaneously, we will compare the results in the first row, a result suffering from both problems, with the results produced when we correct each issue alone.
Among whites, Model 1 and Model 2 (in the first row of Table 8) show that Medicaid eligibility is positively and statistically significantly related to births. Again, Model 3 only includes age in the cell fixed effects based on age. Inclusion of these cell fixed effects reduces the magnitude of the coefficient by 50%, but this result remains statistically significant. We find similar results for Models 4–7.
In the second row for whites, we show the same results reported in Table 7. The only difference between the results in the first row and the second is that we retrieve all the cases lost due to missing information on marital status and education. The point estimates are considerably larger in Models 1 and 2 with the additional cases. More interesting, however, is that with the addition of the missing cases, the point estimates for Medicaid eligibility are negative and statistically significant in Models 3–7. Our failure to add many of the missing cases for marital status—and, more likely, education—in the previous models likely led to point estimates that were upwardly biased (i.e., larger than appropriate).
In the third row, we include the original cell fixed effects (defined based on age, marital status, and education) but do not include the cases with missing information on marital status or education (i.e., the results reported in Table 1). In this case, again, the inclusion of the cell fixed effects substantially reduces the coefficient estimate relative to the results reported in the first row.
In sum, in the initial results our inability to obtain complete cases for the analysis led to point estimates for Medicaid eligibility that are too large (positive). Likewise, a failure to include demographic cell fixed effects also leads to point estimates that are too large (positive). What remains unknown is what would happen if we could add all the missing cases and control for the demographic cell fixed effects simultaneously. It is far from certain that the results would be negative and precisely measured, particularly given the size of the standard errors reported in the third row. Nevertheless, this set of results does suggest that the relationship is probably not positive as has been found in the extant literature.
Family Planning
We also consider another potential explanation for our results: the Medicaid variable could be capturing both the positive effect of the expansions due to the reduction in the cost of a child as well as a negative effect created by the increased access to the family planning services for which eligible families qualify. Together, these two countervailing influences may net one another out. To determine the importance of this potential explanation, we report results for first births from 1986 to 1997 (results available in Table 12 of Online Resource 1). Low-income women without children would not qualify for Medicaid prior to a pregnancy; therefore, they do not have the family planning services available to them. Thus, we should observe only the positive influence of Medicaid, assuming that it exists, on first births.
Regardless of the model chosen, we find small and statistically insignificant coefficients for Medicaid expansions among white mothers. The pattern observed for African American mothers is similar to that observed in most of the previous tables: positive and statistically significant relationships in the first two models and insignificant relationships in Models 3–7.
Statistical Weights
In all our analyses, we weight the observation by the population of women aged 15–44 in the state that year. This choice generates an equal weight for all of the births within a state in a given year, which could lead to incorrect inferences because there are an unequal number of females in each cell. For example, one might not want to weight the cell for the number of births to teens (aged 15–18) the same as the cell that represents the number of births to married women aged 25–34 with a high school education (or any other cell for that matter). Our current weighting scheme weighs each cell within a state-year equally.
It is difficult to obtain annual population data on females by race, age, marital status, and education from the U.S. Census Bureau, and no other potential data source has a sample size adequate to make these distinctions. To allow for different weights across the cells within a state in a given year, therefore, we used the 1990 5% Public Use Microdata Sample (PUMS), which has information on the age, marital status, education, and race. With the PUMS, we generated the proportion of all women 15 to 44 who fell into each of our 44 cells. We then multiplied these proportions by the state population of women aged 15–44 each year. Although not perfect, this reweighting scheme allows us to weight each cell by an estimate of the population that would be counted among that cell in a state in a given year.
We report results using the full sample (1985–1997), using the new weighting in Table 13 in Online Resource 1. Among white women, the results suggest small and statistically insignificant estimates for the Medicaid simulation variable. The results for African American women are similar to those reported earlier. A 1-percentage-point increase in the Medicaid eligible population increases births by 1.9% to 2% in Models 1 and 2. After we control for the cell fixed effects, however, none of the coefficients is statistically significant.
In Table 9, we report results using the new weights and control for the natural logarithm of the population, where population is measured as the product of the population of females aged 15–44 in that state in that year and the proportion of females who fall into that cell in the 1990 census. We find results for both whites and African Americans that are nearly identical to those reported earlier. In Model 1 and Model 2, we find a positive and statistically significant coefficient. After we control for cell fixed effects, however, the coefficients are much smaller in size and statistically insignificant.
Conclusion
We use natality data and a measure of state Medicaid generosity of eligibility to investigate the relationship between Medicaid expansions and fertility in the United States. We use several specifications to test for this relationship, including a state fixed-effects model with time effects, a model regularly used in the extant empirical literature. When using the state fixed-effects models, we consistently find a positive relationship between expansions and births, replicating the work of Joyce et al. (1998). However, after we use the variation that occurs within more narrowly defined demographic cells, none of the results is statistically significant and the coefficients are often trivial in size, leading us to conclude this is not a robust relationship.
Given the high costs of pregnancy, childbirth, and childhood-related health care, Medicaid receipt constitutes a considerable reduction in the cost of having a child. This fact, along with the finding of previous studies that fertility is responsive to financial incentives in general and to having health insurance in particular, makes the absence of a robust finding somewhat surprising.
This result has important policy implications for the Medicaid program, the SCHIP program, and even universal health care. All three of these programs potentially increase the health insurance coverage of many Americans. If an increase in health insurance coverage produced a concomitant change in fertility patterns, policymakers would be wise to consider the fertility implications of enacting or expanding health insurance policies. Our evidence suggests that that there is no robust Medicaid-fertility link. At minimum, our study suggests that more research into the area is needed.
Acknowledgement
We are grateful to Dan Black, Robert Kaestner, Melissa Kearney, Jeffrey Kubik, and seminar participants at Indiana University and RAND for helpful comments. We also thank Jonathan Gruber for sharing the Medicaid eligibility programs used in his previous research as well as Melissa Kearney for sharing her data on family caps, welfare waivers, and TANF implementation.
Notes
According to Espenshade (1977), the total cost of a child can be broken into two components: non-economic costs and economic costs. The non-economic costs are exceedingly difficult to measure. They include, among other things, the fatigue caused by abnormal sleeping patterns and the concern caused by having a sick child. The economic costs include both the direct financial costs as well as the opportunity costs. The estimates from Lino (2000) did not include the opportunity costs or the costs of prenatal care and delivery. Foster (2002) argued that the opportunity cost of a child is roughly the same as the direct financial economic cost.
In addition, 12 states made family planning services available to some women outside Medicaid eligibility (but typically just to former Medicaid moms in the postpartum period and/or women at less than 185% FPL) from 1994 to 2000 through Medicaid Section 1115 waivers. During 1994–1996, the years that overlap our study period, Delaware, Maryland, Rhode Island, and South Carolina implemented postpartum waivers, and Arizona implemented an income-based waiver (Kearney and Levine 2009; Lindrooth and McCullough 2007).
Even though the take-up rate among newly eligible children has been estimated to be rather low, we would expect nearly all pregnant women who are eligible for Medicaid (and not otherwise insured) to enroll in Medicaid at least to cover childbirth expenses. Nearly all births occur in hospitals, and hospitals have the ability and incentive to enroll eligible women in the program to be reimbursed for expenses. Despite this, Medicaid expansions likely will have a causal impact on fertility only to the extent that women know prior to conception that they will be covered by Medicaid; the percentage of women who end up having their child’s delivery expenses paid by Medicaid eligibility is likely an overestimate of the number of women who are aware, prior to pregnancy, that Medicaid will cover the costs of delivery, and only a subset of those will have altered their fertility behavior in response to their potential eligibility.
Currie and Gruber (2001) included an additional race category of “other,” thereby generating 66 cells. We chose not to report results using the “other” category because of the heterogeneity in the composition of this category across states. In preliminary results using this “other” category, however, the findings were identical to those we report for the white and African American populations.
Joyce et al. (1998) used a single dummy variable to record whether the state expanded eligibility to 100% of the poverty level by OBRA 1986 and a second dummy variable to record whether the state expanded eligibility under the 1987 and 1989 OBRAs from 100% of poverty to 185% of poverty.
Because we lagged several variables by three quarters, have birth data from 1985 to 1997, and have simulated Medicaid eligibility from 1985 to 1996, our actual time series runs from the last quarter of 1985 through the first three quarters of 1997.
For other examples, see Joyce et al. (1998), Kearney (2004), Levine et al. (1996), and Lopoo and DeLeire (2006).
Optimally, we would prefer to know the number of females that meet the definition for each cell: for example, females aged 25–29, married, with a college degree in a given state, in a given quarter of a given year. Data with this level of detail do not exist, to the best of our knowledge. One might have concerns because the weights are measured at a different level of aggregation than the fertility counts and Medicaid eligibility measures in our models. To determine whether this choice alters our findings, we construct a different weight using data from the 1990 5% PUMS and report results using this different weight in the Tests of Robustness section. Our results suggest that our choice of weight does not alter our findings.
Figure 2 (available in Online Resource 1) illustrates our measure of the proportion of the population of white women eligible for Medicaid by demographic cell, while Fig. 3 in Online Resource 1 illustrates the measure for African American women. The time series begins in the first quarter of 1985 (Time 0 in the graph) and terminates in the last quarter of 1996 (Time 48 in the graph). The graphs illustrate a couple of points. First, one can see the considerable heterogeneity in Medicaid eligibility across the cells. Secondly, as expected, the cells that have disproportionate shares of low-income women have greater increases in Medicaid expansions. For example, one sees considerable growth in eligibility among high school dropouts, particularly among white women. We investigate the relationship between Medicaid and fertility within some subsamples of the data, based in large measure on the evidence presented in these figures.
Data are missing for many reasons, including that some states as a matter of policy did not collect education information on birth certificates for several of the years in our time series.
Yelowitz (1998) found that Medicaid expansions increase the likelihood of marriage. If true, then our decision to use marital status in our cell definitions may be problematic. The results in Table 6 suggest that even when we do not separate the cells by marital status, our findings are largely the same.
We find nearly identical results for African Americans. To conserve space, we will focus on the white subsample, but the same conclusions apply for African Americans.
