Previous research has found a positive relationship between marriage and infant health, but it is unclear whether this relationship is causal or a reflection of positive selection into marriage. We use multiple empirical approaches to address this issue. First, using a technique developed by Gelbach (2009) to determine the relative importance of observable characteristics, we show how selection into marriage has changed over time. Second, we construct a matched sample of children born to the same mother and apply panel data techniques to account for time-invariant unobserved characteristics. We find evidence of a sizable marriage premium. However, this premium fell by more than 40 % between 1989 and 2004, largely as a result of declining selection into marriage by race. Accounting for selection reduces ordinary least squares estimates of the marriage premiums for birth weight, prematurity, and infant mortality by at least one-half.
Research has consistently found that marriage is associated with a number of positive health outcomes. Married people live longer, have fewer alcohol-related problems, and engage in fewer risky behaviors (Waite 1995). Studies also show that infants born to married parents are less likely to suffer from prematurity, low birth weight, and mortality than infants born to unmarried mothers (Bennett 1992; Bennett et al. 1994; Bird et al. 2000; Peacock et al. 1995). These infant health differences can be large, and vary with maternal characteristics such as race, age, or education (Jacknowitz and Schmidt 2008). Disparities in infant health are of particular concern because of the potential for large effects on long-term outcomes, including chronic illness, educational attainment, income, and the likelihood of physical disability (see Almond 2006; Almond et al. 2005; Barker 1995; Behrman and Rosenzweig 2004; Chay et al. 2009; Oreopoulos et al. 2008; Smith et al. 2008).
Despite the wealth of evidence of a positive relationship between marriage and infant health, it remains unclear whether there is a causal effect of marriage. A major challenge with interpreting these results as such is the possibility of selection into marriage. The observable characteristics of married and unmarried mothers are very different; they are likely different in unobservable ways as well. For example, a common concern is that healthier women may be more likely to marry and may also have healthier babies. And as Ribar (2004) noted, plausibly exogenous sources of variation in marriage have been difficult to find.
We contribute to the literature on the infant health–marriage premium by using several novel approaches to address the issue of selection into marriage. We begin by using birth certificate data from the Centers for Disease Control (CDC) to estimate a raw marriage premium of around 177 g and 0.28 weeks gestation; these magnitudes are similar to estimates of the effect of maternal smoking (Bardy et al. 1993; Ward et al. 2007). We then take advantage of the rich set of demographic and health controls available in the birth certificate data to account for selection on observable characteristics, using a strategy proposed by Gelbach (2009) to determine the relative importance of the included covariates. We also use the Gelbach (2009) procedure to determine how selection into marriage has changed in recent years.1
Next, to account for selection based on time-invariant unobserved characteristics, we exploit individual-level variation in marital status across births. We construct a unique matched sample of siblings from the 1980–1988 Natality Detail Files and use fixed-effects and first-differences methods to estimate the effects of transitioning into or out of marriage. Previous research on the effects of marriage using these techniques has been limited by small sample sizes (Aaronson 1998; Geronimus and Korenman 1993). In contrast, our matched sample has more than 620,000 sibling pairs. As both a check on our data and as a demonstration of the value of our matched sample, we supplement this analysis with data from the 1979 National Longitudinal Survey of Youth and the 1995, 2002, and 2006–2010 waves of the National Surveys of Family Growth.
Across specifications, we find that selection into marriage can account for more than 50 % of the observed infant health–marriage premium. We also show that demographic characteristics are particularly important—selection on race alone can account for about one-third of the gap in birth weights between infants born to married and unmarried mothers. However, selection on race fell between 1989 and 2004, contributing to a 40 % reduction in the overall premium over this period. Accounting for unobserved heterogeneity with our panel data strategies further reduces the estimates, but this earlier sample still appears to have a marriage premium—that is, infant health improves for women transitioning into marriage—while there is a decline of similar magnitude for women who transition out of marriage. But our evidence on the importance of selection in explaining the marriage–infant health premium has important implications for policy efforts to improve child outcomes by promoting marriage.2
Marriage and Infant Health: Theory
Most theories of marriage suggest that marriage should have a positive effect on the health of both married individuals and their children. Duncan et al. (2006) documented a number of these reasons. First, marriage facilitates easier monitoring of each other’s behavior. They noted that “people behave better when someone with power to reward or sanction is watching,” and marriage provides ample opportunity for such monitoring. Second, the institution of marriage itself may include the notion of “cleaning up your act,” and may come with expectations, obligations, and social sanctions against certain behaviors that are harmful to one’s health (or the health of a child). Third, marriage facilitates a wide net of social bonds involving the extended families and friends of both individuals in the marriage. Finally, marriage provides a couple with legal access to each other’s resources and a system in which each individual in the marriage can take advantage of economies of scale.
The Weiss and Willis (1985) model also provides insight into why marriage would be particularly beneficial for children. In their model, both the mother and father have a utility function that includes both their own consumption and the quality of their children. Thus, children are treated as a collective good by both parents. Marriage allows the couple to monitor and enforce each other’s investment in the collective good through proximity and trust. This allows the couple to overcome the free-rider problem inherent with all collective goods.
There are a few reasons that marriage may lead to worse outcomes for children. For example, marriage (and the economic interdependence that it creates) may tie women and children to an abusive relationship (Gelles 1976; Strube and Barbour 1983) or limit geographic mobility (Bartel 1979; Bielby and Bielby 1992). In cases where the husband has little income, the laws that ensure sharing of resources may draw resources away from the children (Edin 2000).
This theoretical work provides some intuition for how marriage might affect infant health specifically. Known inputs into infant health include the quality and timing of prenatal care (Abrevaya and Dahl 2008; Currie and Gruber 1996; Evans and Lien 2005; Joyce 1999), nutrition during pregnancy (Almond and Mazumder 2011), abstaining from smoking during pregnancy (Evans and Ringel 1999), and the stress level of the mother (de Weerth et al. 2003; Ponirakis et al. 1997). Consistent with the theoretical channels discussed earlier, marriage might affect access to care by providing legal access to the health insurance benefits and income of the spouse (Hahn 1993). Marriage may also reduce smoking or improve nutrition through the increased ability of a spouse to encourage and monitor good behavior (Laub et al. 1998; Umberson 1992). The economies of scale associated with marriage might also lead to better nutrition, and the emotional support that accompanies a good marriage might lead to lower levels of stress. Some of these channels could lead to worse health outcomes, however, if the spouse encourages harmful behaviors or if the marriage increases stress.
Of course, some of the channels for an effect of marriage on infant health might also exist in nonmarital relationships, particularly in cohabiting relationships. If, for example, cohabiting unmarried partners are equally able to monitor certain behaviors, such as smoking and nutrition, we will be less likely to find an effect of marriage on infant health. However, other channels, such as access to health insurance and income, are usually available only to legally married partners. Likewise, the stress-reducing benefits of marriage may be greater when the relationship is legally recognized (Stack and Eshleman 1998).
Marriage and Infant Health: Evidence
The preponderance of evidence across a number of studies indicates that infants born to married parents are less likely to suffer from prematurity, low birth weight, and mortality than infants born to unmarried mothers (Bennett 1992; Bennett et al. 1994; Bird et al. 2000; Peacock et al. 1995; Raatikainen et al. 2005). In a recent meta-analysis, Shah et al. (2011) concluded that unmarried women have higher unadjusted odds of having a child who had low birth weight (1.46), was premature (1.22), and was small for gestational age (1.45). When only adjusted odds estimates were included in the analysis, the ratios were attenuated but still large. A few studies failed to find a protective effect of marriage for some demographic groups. For example, Jacknowitz and Schmidt (2008) noted that children born to unmarried mothers who have at least a college degree do not suffer from negative birth outcomes. Bennett (1992) found the relationship between marital status and birth outcomes varies by maternal race and age, and also suggested that high infant mortality rates for married teen mothers challenge the notion that childbirth is protected by marriage per se.
A serious limitation of the research on the infant health–marriage premium is the difficulty in accounting for selection into marriage. In his review of empirical studies of the relationships between marriage and outcomes like health or children’s well-being, Ribar (2004: p.vi) stated that “selectivity appears to be more than a hypothetical concern” and that “studies in this area generally do not address issues associated with selection and omitted variables bias” (p. vii). When researchers have acknowledged potential selection bias, the usual approach has been to control for common cofounders, such as race, education, and age. Ribar asserted that this approach is “sensible but is only successful if the researcher knows which variables are missing and can find the corresponding measures” (p. v). Ribar suggested that panel data methods could provide some insight, but a lack of large panel data sets has precluded their use.3 He also suggested the use of instrumental variables, but good instruments for marriage are difficult to find.4
In this article, we address these issues in several ways. First, we take advantage of the rich set of observable characteristics available in the Natality Detail Files from 1989–2004 to account for selection on observables. Notably, we can include a measure of maternal health, which is a potentially important omitted variable in previous work. Second, we use an innovative statistical technique developed by Gelbach (2009) to determine which covariates are important in explaining the infant health–marriage premium and how much they matter. We apply this technique year by year to show how selection into marriage changed during this period. Third, we create a unique matched sample of more than 620,000 sibling pairs from the 1980–1988 Natality Detail Files, allowing us to implement fixed-effects and first-differences methods that account for time-invariant unobserved characteristics.
Our primary analysis uses data from U.S. birth certificates for the years 1980–2004, from the Centers for Disease Control (CDC) Natality Detail Files. As of 1985, all states report 100 % of their birth certificate data, accounting for more than 99 % of all births.5 These data include information on characteristics of the mother (age, race, education, state of residence, and marital status), the infant (gender, birth order), and infant outcomes (birth weight and gestation). A major revision to the birth certificate data in 1989 added indicators for maternal risk factors, such as anemia and diabetes. Approximately 4 million records exist for each year. We restrict our sample to mothers over age 18 because they are eligible to be married in all states and years.
For most states, the birth certificate includes information about the marital status of the mother, although some states impute that a mother is unmarried if the surname of the father is missing or does not match that of the mother. The number of states that impute the marital status has dropped from nine states in 1980 to only two states in 2004. One potential limitation of the birth certificate data is that they do not indicate when the mother was married. In the years just before the start of our sample (1975–1979), 49 % of white women and 11 % of black women who had a premarital conception experienced a “shotgun” wedding (O’Connell and Rogers 1984).6 If marriage has a causal effect on infant health, having mothers in our sample who get married shortly before the birth will bias our estimates toward zero because the likely channels for the effect might not have been in place during the prenatal period.
Another limitation of these data is that we observe only whether the mother is married—not whether single women are cohabiting at the time of birth. Cohabitation rates have been increasing during the period that we include in our analysis. If cohabitation confers many of the same benefits as marriage, our estimates of the marriage premium in infant health should decrease over time. We discuss this issue in more detail in the upcoming results section.
Jacknowitz and Schmidt (2008) showed that the effect of marriage on infant health is heterogeneous, and so we also conduct the analysis separately for white and black women.7 Black mothers are an important population, given that their rates of nonmarital childbearing and adverse outcomes (such as low birth weight) are especially high. Moreover, our results suggest that there is significant selection into marriage by race; estimating results separately for blacks and whites allows us to examine the degree of selection along other characteristics within these groups.
Figure 1 depicts the change in nonmarital birth rates from 1980–2004, for all mothers over age 18, and for blacks and whites. The overall nonmarital birth rate for mothers over age 18 has increased from about 14 % in 1980 to about 32 % in 2004. Rates are much higher for blacks, for whom nearly two-thirds of births in 2004 were nonmarital. Since 1994, however, nonmarital childbearing appears to have stabilized for blacks, while rates for whites have continued to rise.
Summary statistics by marital status are presented in Table 1. Here, the sample is limited to births after the 1989 birth certificate revision. We use six different measures of infant health: birth weight in grams, an indicator for low birth weight (<2500 g), weeks of gestation, an indicator for prematurity (<37 weeks), an indicator for low Apgar score (<7), and an indicator for infant mortality within the first year.8 These outcomes are widely used measures of infant health, and each measure has strengths and weaknesses. Birth weight in grams and gestation in weeks are appealing because they measure health across the entire distribution. Low birth weight, prematurity, and low Apgar scores measure adverse infant health outcomes but are defined by arbitrary (although widely accepted) cutoffs. Apgar scores are correlated with neonatal death but may not be good predictors of longer-term outcomes (Committee on Fetus and Newborn and Committee on Obstetric Practice 1996). Finally, infant mortality is an important but extreme outcome; there are fewer than 5 infant deaths per 1,000 births for this sample. Although no measure is perfect, our use of the six in combination should give a comprehensive picture of the relationship between marriage and infant health.
For all samples, married mothers have babies with higher birth weight and greater gestation. The gap in birth weight is 177 g for the full sample, and the gap in gestation is 0.28 weeks. These gaps are comparable to estimates of the effects of maternal smoking: Ward et al. (2007) found that smoking is associated with a gap of 168 g, and Bardy et al. (1993) estimated a gap of 0.24 weeks gestation for infants exposed to tobacco smoke. Infants born to married mothers are less likely to have low birth weight (4.1 percentage points), be premature (4.7 percentage points), or have a low Apgar score (0.7 percentage points). Infant mortality rates are 59 % higher for single women versus married women. The estimated premiums are generally smaller for black and white subsamples, which we would expect if selection into marriage by race is responsible for some of the full-sample premium.
Table 1 also shows demographic characteristics of mothers by marital status. Married mothers are, on average, older, more educated, and more likely to be white than are unmarried mothers. Married mothers are less likely to experience a medical risk factor (26 % vs. 30 %). For the full sample, pregnancy-related hypertension is the most common condition (3.6 %), followed by diabetes (3.0 %) and anemia (2.3 %). For each measure, married women appear to be positively selected on characteristics that are associated with improved infant health outcomes.
The Marriage Premium and Observable Characteristics
OLS estimates of Eq. (1) without the terms Xist, riskist, , and will yield our baseline (unadjusted) estimates of the marriage premium. We will then add these controls for observable maternal and infant characteristics to account for some types of selection into marriage. The vector Xist includes the following maternal and infant demographic characteristics: single year-of-age dummy variables, single year-of-education dummy variables, birth-order dummy variables, indicators for black and other race, and the fraction female in the cell. The next term, riskist, is the fraction with a health-related risk factor in the cell (defined earlier) to address the issue of the selection of healthier women into marriage. Finally, is a vector of state-of-residence fixed effects, and is a vector of year dummy variables that control nonparametrically for national trends in infant health outcomes. Because maternal risk factors are available only after 1989, we restrict our sample to the 1989–2004 period for this analysis. We further limit our sample to women for whom education is observed (more than 95 % of the full sample).
A comparison of the estimated marriage premium in the basic and full specifications indicates how much of the unadjusted premium is due to selection along observable characteristics. However, we are interested in not only how much the covariates change the estimate but also which covariates are responsible for the change. Is selection along demographic characteristics more important or less important than selection on health, for example? One common strategy would be to see how the premium changes as covariates are added sequentially; however, the results from this approach can be very sensitive to the order of their addition (Gelbach 2009). Gelbach provided a method for estimating the contribution of various sets of covariates to the change in the coefficient that is conditional on all covariates and invariant to the order in which they are added. Intuitively, the mean differences between married and unmarried mothers in (for example) demographic characteristics or health are scaled by their infant-health effect conditional on all other covariates.10 We implement this “Gelbach decomposition” to identify the important dimensions along which there is selection into marriage. We then extend this work by conducting the analysis year by year in order to explore how selection into marriage is changing over time.
The Marriage Premium and Unobservable Characteristics
The methods in the previous section allow us to gauge the extent to which marriage is related to infant health through measurable channels. To address the issue of selection into marriage based on unobservable characteristics, we use an estimation strategy that exploits individual-level variation in a woman’s marital status across births. We use data from the 1980–1988 Natality Detail Files and exploit information on the month and year of the previous birth. This allows us to match mothers who are having a second or higher birth with the natality record from the previous birth. We match both on the month and year of the previous birth as well as characteristics of the mother, including her own state of birth, state of residence, race, and the year when she was born. Because some mothers with these characteristics could still potentially match with many other previous records, we keep only the births that generate a unique match. We are able to identify a unique match for 2.5 % of the sample, or more than 620,000 sibling pairs.11
The natality data allow us to create a very large sample of sibling pairs, but the matched sample is not likely to be nationally representative given the way matches are identified. In Table 2, we show the results of a probit regression in which the dependent variable is equal to 1 if we can uniquely identify a younger sibling in the data. We are more likely to find a unique match for infants who are part of a small group, such as a minority race or a small state. Also, we find more matches for infants who are more likely to have a younger sibling by 1988—those born earlier in the decade or who are lower in birth order. Infants with mothers in their 30s are more likely to be matched (as part of a smaller group), but infants with mothers in their 40s are less likely to be matched (because of a lower probability of a subsequent birth by 1988). Although these results show that our sample has some distinctive characteristics, there does not appear to be clear positive or negative selection on characteristics that affect infant health.12
Variables are defined as earlier, but j indexes the mother, and represents the mother-specific fixed effect. Observations are also at the individual level rather than the cell level, and characteristics that do not vary across births are omitted from Xijt. The coefficient of interest is , which shows the relationship between marriage and infant health for women who change marital status between births.
The Marriage Premium and Observable Characteristics
Table 3 provides our estimates of the relationship between marital status and infant health for the full sample. The first column provides the baseline marriage premium in birth weight, gestational age, infant mortality, as well as probability of low birth weight, prematurity, and low Apgar score. For all six measures, marriage is associated with better infant health. The magnitudes are identical to the differences in means for the two groups described earlier.
In the next column of Table 3, we show the estimate of the marriage premium obtained after including the full set of controls in Eq. (1). The third column gives the difference between the baseline (unadjusted) and full (adjusted) estimates. For all measures except the Apgar score, the marriage premium falls by more than one-half when the additional covariates are included, indicating that selection along observables contributes significantly to the estimate of the unadjusted premium. For Apgar scores, the premium declines by 46 %.
The results of the Gelbach decomposition are presented in the remaining columns of Table 3 and indicate which sets of covariates are important in accounting for the marriage premium. In all cases, most of the reduction in the premium comes from the demographic controls. For birth weight, demographic controls account for 52 % of the raw marriage premium (92.03 / 176.70). Maternal health and state and year fixed effects combined, on the other hand, account for less than 6 %. Demographic characteristics account for 63 % of the marriage premium in infant mortality (–1.52 / –2.42), while maternal health has very little effect. For the other measures of infant health, demographic characteristics account for between one-third and one-half of the raw premium. Thus, the results of the Gelbach decomposition show that selection into marriage along demographic characteristics (age, education, race, birth order, and gender) drives the infant health–marriage premium much more than selection along other observable characteristics.
The two panels in Table 4 provide these same specifications for whites and blacks separately. We omit results for weeks of gestation and low Apgar score for brevity, but results are consistent with those for the four measures shown. For whites, estimates of the marriage premium are generally smaller than those for the full sample, but adding the controls still reduces the estimate significantly. For blacks, however, there appears to be much less selection on observables. For low birth weight and prematurity, adding the controls reduces the estimate of the premium by only 4 %. This is driven by the addition of the demographic controls, which actually serve to increase the estimate of the premium, suggesting that for blacks, the characteristics that are associated with marriage are correlated with higher rates of low birth weight and prematurity.
In Fig. 2, we explore how the marriage premium for birth weight is changing over time.13 Both the raw and adjusted premiums fell substantially between 1989 and 2004. The raw premium went from 226 g to 135 g (a 40 % drop), and the adjusted premium went from 105 g to 55 g (a 48 % drop). Thus, the fraction of the gap explained by observable characteristics increased slightly during this period. In results not shown here, we confirmed that this dramatic drop in the premium was a result of both an increase in birth weights for unmarried mothers and a decrease in birth weights for married mothers.
In the natality data, we are unable to distinguish between unmarried mothers who are single and unmarried mothers who are cohabiting. Because cohabiting relationships may provide many of the same protective effects as marriage, the infant health of children born to cohabiting mothers will likely be more similar to children born to married mothers than are the outcomes of children born to noncohabiting single mothers (Duncan et al. 2006 and Wu et al. 2003). During most of the period that we examine in our analysis, the fraction of children born to cohabiting couples has followed an increasing trend (Raley 2001). This rise in the cohabiting rates among the single mothers likely contributes to the decrease in the marriage premium that we document in Fig. 2.
In Fig. 3, we further investigate this fall in the premium by again using the Gelbach decomposition to determine how selection along observables contributes to the marriage premium. But here, we calculate the decomposition year by year so that we can see how this selection changes over time. We also disaggregate the demographic covariates so that we can see the extent of selection along child characteristics (birth order and gender) and the mother’s age, race, and education separately. The figure shows the contribution of the indicated characteristic to the raw birth weight premium, in grams.
First, note that selection into marriage by race is the most important factor in explaining the marriage premium. In 1989, 80 of the 226 g difference between birth weights for married and unmarried mothers (35 % of the gap) is accounted for by race. By 2004, race is still the most significant contributor, but both the contribution in grams and the percentage of the premium accounted for by race have fallen significantly (to 39 g and 29 %, respectively). Mother’s education is the second most important factor for most of the period, while mother’s health, age, child characteristics, and state fixed effects contribute fewer than 10 g to the premium in any given year. Mother’s age actually increases the gap in the beginning of the period but attenuates it in later years.
Figure 3 provides some insight into how and why the infant health–marriage premium changed between 1989 and 2004. Rates of nonmarital childbearing stabilized for blacks during the late 1990s and early 2000s, while rates for whites increased. Rates for other racial groups were relatively stable. Thus, selection into marital childbearing according to race declined, contributing to a reduction in the marriage premium (given that birth weights are higher for white women, on average). At the same time, married women were having babies much later; we see a nearly two-year increase in both the mean and median age at birth for married mothers over this period but little change for unmarried mothers. Given that older women have higher birth weight babies, on average, these trends led to an increase in the marriage premium and the importance of age in explaining it.14 But on net, the declining selection into marital childbearing according to race dominated, so that the raw marriage premium fell dramatically over this short period.
Before turning to our panel data results, we conduct one additional exercise to explore the possible role of selection on unobservable characteristics in explaining the marriage premium. Using the method developed by Altonji et al. (2005), we estimate the ratio of selection on unobservables to selection on observables that would be required in order to attribute the entire infant health–marriage premium to selection bias. The included observables are the same as those used to create the adjusted estimate in Table 3, with the exception of race and state and year fixed effects.15 The Altonji et al. method is valid only under the assumption that no single characteristic dominates the distribution of the endogenous or dependent variable; because the results in Fig. 3 suggest that this assumption may be violated for the full sample, we conduct this analysis separately for blacks and whites.
For birth weight, the implied ratio for whites is 0.49 and for blacks is 0.18, suggesting that the marriage premium would be completely explained by selection bias if the amount of selection on unobservables was at least 49 % and 18 % as large as the amount of selection on observables, respectively. For rates of prematurity, the ratio is 0.19 for whites and 0.25 for blacks; for low birth weight, the ratios are 0.22 and 0.28, respectively. These implied ratios suggest that estimates of the marriage premium for 1989–2004 could be entirely attributable to selection if there is even a moderate amount of selection on unobservables.
The preceding results show that much of the marriage premium for infant health is accounted for by selection, but there is still scope for a small causal impact. To investigate the importance of some of the mechanisms for a causal effect discussed in the Background section, we now add controls for smoking, prenatal care, household income, and insurance coverage. We view these variables as endogenous to marriage, although there may be selection along these characteristics as well.16 We continue to use the 1989–2004 Natality Detail Files for this analysis because the data include information on both maternal smoking and prenatal care. Our measure of prenatal care is an indicator for receiving care in the first trimester.
Household income and insurance coverage are not available in the natality data. To include them as controls in our model, we use data on women with children younger than 6 from the 1989–2004 Current Population Survey (CPS). We collapse the data to cells according to state, year, age, education, race, and marital status, and then merge income and insurance characteristics to corresponding cells in the natality data.17 Seventy-six percent of the observations in the natality data are matched to CPS data; cells with older women or women who are neither white nor black are less likely to be matched. Although we expect some measurement error with this approach, it will allow us to investigate whether these channels are likely to be important. There is reason to believe that access to insurance, in particular, could contribute to the infant health–marriage premium given that the decline in the premium during the 1990s coincides with rising rates of insurance coverage for single women as a result of Medicaid expansion (Aizer and Grogger 2003). On the other hand, there is less evidence that the expansion of coverage led to improved birth weight outcomes (Dubay et al. 2001).
Results of this exercise are in Table 5. We include all controls from Eq. (1) and then successively add controls for the fraction in the cell who are smokers, the fraction who received prenatal care in the first trimester, the average household income for the cell, and the fraction covered by insurance in the cell. For both birth weight and low birth weight, controlling for smoking reduces the coefficient on the indicator for married by 32 %. Thus, smoking behavior plays an important role in explaining the remaining marriage premium, although we cannot distinguish between a selection mechanism (smokers are less likely to marry) and a causal mechanism (marriage makes people less likely to smoke). Controlling for prenatal care also reduces the marriage coefficient, by about 10 % in each case. Our controls for household income and insurance coverage, on the other hand, have very little effect—likely because the variables themselves have little relationship to our measures of infant health. When all controls are included, marriage is associated with a birth weight premium of about 40 g and a 1 percentage point decrease in the incidence of low birth weight.
The Marriage Premium and Unobservable Characteristics
The results in the previous section indicate that much of the infant health–marriage premium is due to observable characteristics, but that in most cases a meaningful premium remains after these controls are added. We now turn to methods that allow us to address issues of unobserved heterogeneity. First, in Table 6 we present estimates of the marriage premium for our matched sample from the 1980–1988 natality data. Compared with Tables 3 and 4, which use data from 1989–2004, we find even larger estimates of the baseline marriage premiums. This is consistent with the evidence of a declining infant health–marriage premium presented earlier.
Just as before, adding controls for observable characteristics substantially decreases estimates of the premium. When we include the mother fixed effects, the marriage gap drops even further, so that our fixed-effects estimates of the premiums for the full sample are about 40 % of the estimates of the unadjusted premium. However, a statistically significant premium remains; for birth weight, the fixed-effect estimate (96 g) is more than half of a standard deviation of the birth weight across the full sample. The fixed effects also show that being married is associated with a 2.7 percentage point decrease in the likelihood of low birth weight and a 3.0 percentage point decrease in the likelihood of prematurity. Results are similar for both whites and blacks.
In the bottom two rows of Table 6, we show the results from a replication of our analysis using data from two panel data sets: the 1979 National Longitudinal Survey of Youth (NLSY79) and the National Surveys of Family Growth (NSFG 1995, 2002, and 2006–2010). Results using these data are interesting for two reasons. First, the NLSY79 and the NSFG are much smaller than our matched sample, and a comparison of the precision of the estimates demonstrates the value of our large sample. Second, these data sets are nationally representative and allow for siblings to be matched perfectly. If our point estimates are similar to those from the NLSY79 and the NSFG, it suggests that our results are not an artifact of the unique properties of our sample.
Particularly for birth weight and low birth weight, the OLS point estimates are very similar to those for the full sample from the natality data. Fixed-effects estimates of the premium are smaller in the NLSY79 and the NSFG. In fact, the NLSY estimate for birth weight is negative, and the NSFG estimates for low birth weight and prematurity are positive. The standard errors are very large, however, so that large beneficial marriage premiums for the fixed-effects parameters cannot be rejected. That the estimates from the NLSY and the NSFG are similar to those in the matched natality sample but are much less precise highlights the value of our approach.
The results in Table 7 use data from the same matched natality sample but allow the effects of marriage to vary by whether the mother transitioned into or out of a marital relationship. The results are striking. First, estimates of the premium are quite close to those from the fixed-effects specification. Moreover, in most cases, the effect of moving into the married state is very similar in magnitude to the effect of moving out of it. For example, when using birth weight as the outcome variable, we find that women who enter marriage experience an increase of 98 g between adjacent births relative to women who are married for both, while women who exit marriage experience a 108 g relative decrease. For six of the nine estimates in Table 7, we fail to reject the hypothesis that the absolute values of the coefficients are different at the 5 % level. One concern about the fixed-effects estimates is that there are unobserved time-varying characteristics that affect both marital transitions and infant health. But there is no reason to expect that a spurious relationship resulting from time-varying characteristics or direct effects would be symmetric; the fact that they are symmetric gives us some confidence that what we are capturing is largely the effect of marriage.
We might also be concerned that adverse infant health outcomes could affect marital transitions given the ample evidence that prematurity and NICU stays cause stress and anxiety (Miles et al. 1991, 2007). Were that the case, divorce and infant health would appear to be positively related because the pre-divorce baby is of especially poor health. In our switching results in Table 7, however, we find the opposite: parents who transition out of marriage have worse infant health outcomes in the second birth. We therefore feel that our fixed-effects results are not driven by a causal effect of infant health on marriage/divorce.
Finally, it is possible that the transition itself has a direct effect on infant health.18 There is some support for this in the results: where the estimates are statistically different, the effect of switching out of the marriage state is larger in absolute value terms than the effect of switching in, which may reflect adverse changes in stress or resources associated with ending a marriage.19 Thus, it is important not to interpret the fixed-effects results as identifying the causal effect of marriage. Rather, the results show that accounting for selection on observable and time-invariant characteristics significantly reduces estimates of the marriage premium, but the possibility of a meaningful causal effect for the 1980–1988 period remains.
Using birth certificate data from the 1989–2004 Natality Detail Files, we find large health disparities between babies who are born to married and unmarried parents. In fact, the marriage gaps for birth weight and gestation are as large as the gaps between mothers who do and do not smoke.20 We use a decomposition approach developed by Gelbach (2009) to shed light on the extent to which the observed relationship between marriage and infant health can be explained by selection according to observable characteristics, and how that selection varies over time.21 Adding a rich set of demographic and maternal health controls reduced the estimated premiums substantially for the full sample—selection along demographic characteristics alone could account for more than one-half of the birth weight and infant mortality gaps. Race is particularly important, accounting for about one-third of the birth weight gap. We also find some evidence of negative selection into marriage along demographic characteristics for black mothers. And between 1989 and 2004, the raw marriage premium fell by more than 40 %, largely driven by declining selection into marriage by race. Our results implementing the Altonji et al. (2005) method suggest that the marriage premium in more recent years could be due entirely to selection if there is even a moderate amount of selection on unobservables.
We add to this analysis by constructing a unique matched sample of more than 620,000 sibling pairs from the 1980–1988 Natality Detail Files, allowing us to estimate fixed-effects and first-differences specifications to account for heterogeneity in time-invariant unobserved characteristics. Doing so further reduces estimates of the marriage premium, and results are similar when using comparable samples of mothers from the NLSY and the NSFG. We do find that a meaningful premium remains in these specifications using the 1980–1988 data: for the full sample, transitioning into marriage is associated with an increase in birth weights of about 100 g, and there is a similar-in-magnitude decline for women who transition out of marriage.
Taken together, our results find little scope for a large causal effect of marriage on infant health, particularly for recent years. This finding is relevant to a number of important public policy reforms that have been viewed as opportunities for increasing marriage rates, including welfare reform, reducing requirements for a marriage license, or changing the way taxes penalize marriage. Our results suggest that efforts to improve child outcomes by increasing the marriage rate may have limited impact. An important caveat is that our results assess the importance of a mother’s marital status, rather than the quality of the relationship. They do not, therefore, speak directly to policies designed to improve the quality of marriages, such as those that promote marital counseling or marriage education.
This article has benefitted from the research assistance of Alan Gelder, Phillip Manwaring, Angie Otteson, Craig Palsson, and Kristy Parkinson. We are thankful for comments from Bill Evans, Dan Hungerman, Lucie Schmidt, and participants in seminars at the University of Washington, University of Notre Dame, Brigham Young University, University of Miami, Baylor University, and the 2011 Southern Economics Association Meetings.
Throughout the article, we use the term “selection” to refer to the marriage decision. There may also be selection into the sample if the likelihood of having a live birth conditional on a pregnancy is correlated with the mother’s characteristics. Women may select out of the sample through abortion; rates of abortion fell for both single and married women between 1980 and 2004 (Jones et al. 2009); and between 1994 and 2008, the fraction of women receiving abortions who were married declined from 18.4 to 14.8 (Jones et al. 2002, 2010). It is unclear, however, how selection into abortion varies by marital status, and we are therefore unable to determine how this type of selection would affect our results.
For example, the Administration for Children and Family’s Healthy Marriage Initiative is motivated by the 1996 Congressional finding that “marriage is an essential institution of a successful society which promotes the interests of children.”
Although not the focus of their studies, Royer (2004) and Abrevaya and Dahl (2008) provided fixed-effects estimates of the effect of marriage on infant health using state-level data linking birth certificates.
Two notable papers that have used an instrumental variables specification for marriage are Finlay and Neumark (2010) and Dahl (2010). Finlay and Neumark used incarceration rates as an instrument for marriage, and Dahl used state variation in minimum age requirements for marriage. Interestingly, both studies found that for women whose decision to marry is affected by these instruments (and who are generally low socioeconomic status), marriage has negative effects on outcomes.
Prior to 1985, a few states reported only 50 % of the birth certificate data.
A “shotgun” wedding occurs when a couple is forced to marry to avoid the embarrassment from a nonmarital pregnancy.
We have also produced results that stratify the sample by education and maternal age. Because the results are qualitatively similar to those for the full sample, we omit them here for brevity.
We use the five-minute Apgar score, which is an assessment of the infant’s overall health 5 minutes after birth, using a 10-point scale. All data are from the 1989–2004 Natality Detail Files, with the exception of infant mortality, for which we use the Vital Statistics linked infant death/birth certificate data from 1989–1991 and 1995–2002. The NCHS did not produce these data for 1992–1994.
Cells are defined by single year of age, single year of education, birth order, race, state, birth year, and marital status.
See Buckles and Hungerman (forthcoming) for a more detailed explanation.
Two previous studies have used similar approaches to create longitudinal data set using the Natality Detail Files. Currie and Moretti (2002) used first and second births from the 1970–1999 files to estimate the effect of education on infant health. Abrevaya (2006) matched mothers for a restricted subset of state pairs (using smaller states) for 1990–1998 to estimate the effects of maternal smoking.
Matched infants are more likely to be black, which is associated with worse infant health, but their mothers are more likely to be in their 20s and 30s, which is associated with better infant health. We have no reason to think that being from a smaller state would be systematically related to infant health.
We use birth weight as our measure of infant health in all figures because it is a widely used measure that captures trends in the entire distribution of infant health. Results are similar using the alternative measures.
The relationship between mother’s age and birth weights is actually quadratic, with a peak occurring at around age 32. The average age for married mothers increased from 27.7 in 1989 to 29.4 in 2004, which is still in the range where increasing maternal age is associated with higher birth weights.
We omit the fixed effects because the Altonji et al. method requires that the observables are drawn randomly from the full set of characteristics that determine the outcome. Conceptually, we think that state and year fixed effects might violate this assumption; practically, it makes almost no difference because these covariates explain very little of the variation in infant health.
In the earlier analysis, we treat mother’s health as exogenous because many of the health conditions identified are chronic (such as chronic hypertension, renal and cardiac disease, and some diabetes). However, the mother’s health is also a mechanism through which marriage could affect infant health. Because the aforementioned results consider mother’s health separately, one could easily approximate the effect of treating health as a mechanism rather than as a channel for selection.
To increase the number of cells in the natality data with a match in the CPS data, we use five-year of age cells and education cells defined by degree status.
Wu and Hart (2002) found that transitions out of marriage are associated with decreases in physical and mental health.
The coefficient on the dummy variable indicating that the woman was unmarried for both births is small in magnitude and statistically insignificant for whites but is statistically significant for blacks. Black women who were unmarried for both births saw a greater decrease in infant health than those who were married for both, which may indicate that time-varying factors that affect infant health worsen over time for single black women.
For another comparison, nonmarital childbearing is associated with a 67 % increase in the rate of low birth weight, and interpregnancy intervals of 3 rather than 18 months are associated with a 49 % increase (Conde-Agudelo et al. 2006).
Demographers should find the Gelbach approach useful for accounting for selection bias in a variety of other settings as well.