Abstract
As college-going among women has increased, more women are going to college from backgrounds that previously would have precluded their attendance and completion. This affords us the opportunity and motivation to look at the effects of college on fertility across a range of social backgrounds and levels of early achievement. Despite a substantial literature on the effects of education on women’s fertility, researchers have not assessed variation in effects by selection into college. With data on U.S. women from the National Longitudinal Survey of Youth 1979, we examine effects of timely college attendance and completion on women’s fertility by the propensity to attend and complete college using multilevel Poisson and discrete-time event-history models. Disaggregating the effects of college by propensity score strata, we find that the fertility-decreasing college effect is concentrated among women from comparatively disadvantaged social backgrounds and low levels of early achievement. The effects of college on fertility attenuate as we observe women from backgrounds that are more predictive of college attendance and completion.
Introduction
Higher education influences women’s career aspirations, labor market involvement and experiences, familial roles, and fertility (Becker 1981; Rindfuss et al. 1980). Scholars characterize the wide, and widening, differences in family patterns by educational attainment in the United States as a demographic divide between the more and less educated (McLanahan 2004). Women’s significantly increasing educational attainment (Buchman and DiPrete 2006) motivates further attention to the impact of higher education, particularly among women who have a low likelihood of attaining a college education. As college attendance expands, more women go to college from backgrounds that had hitherto made college attendance unlikely. Are the effects of college the same for women who have a low likelihood of attaining a college education as for women who come from more advantaged backgrounds and high levels of achievement? Despite a substantial literature on the effects of college on women’s fertility, researchers have not assessed variation in these effects by selection into college. Our study fills this gap, examining the effects of college on women’s fertility by the probability that women attend and complete college.
Education and Fertility
Average Effects of Education on Fertility
Educated women delay the onset of childbearing and have fewer children overall compared with less-educated women (Blossfeld and Huinink 1991; Brewster and Rindfuss 2000; Caucutt et al. 2002; Martín-García and Baizán 2006; McCall 2000; Rindfuss et al. 1980, 1996; Spain and Bianchi 1996). There are several, potentially congruent, explanations for why more education is associated with lower fertility. One explanation involves “opportunity costs” (Becker 1981; Blossfeld and Huinink 1991). When making family and career decisions, women weigh competing demands on their time, energy, and commitment. Demands compete because of role incompatibility between childrearing and participating in economically productive work in modern industrial society (Brewster and Rindfuss 2000; Goldin and Katz 2000). Women with higher educational attainment and higher potential wages encounter higher opportunity costs to childbearing, yielding fertility differences by education (Ellwood and Jencks 2004).1,2
Cultural norms offer a second explanation for educational differentials in fertility. Women from different socioeconomic backgrounds and access to college assemble different notions of personal success. Educational and career attainments characterize advantaged young women’s early goals, whereas childbearing marks young disadvantaged women’s social identity and achievement (Edin and Kefalas 2005; Furstenberg 1976; Wilson 1987). The difference between advantaged and disadvantaged young women’s goals relates to opportunity costs because motherhood offers a valid social role for women who perceive “little access to the academic degrees, high status marriages, and rewarding professions that provide many middle- and upper-class women with gratifying social identities” (Edin and Kefalas 2005:171).
A third explanation involves the decoupling between marriage and fertility among less-educated women in contrast to the strong link between marriage and fertility among educated women. In recent decades, family patterns of women at the top of the socioeconomic order have diverged from those of women at the bottom (Lundberg and Pollack 2007). Highly educated women postpone parenthood as well as marriage while less-educated women postpone only marriage. Given patterns of educational homogamy (Schwartz and Mare 2005), less-educated women are appreciably affected by the deterioration of less-educated men’s employment stability and economic prospects (Autor et al. 2006; Cherlin 2009; Wilson 1987). When “marriageable” men are relatively scarce, (nonmarital) childbearing is a potential family-formation strategy among economically disadvantaged women (Edin and Kefalas 2005; Furstenberg 2001, 2003; Oppenheimer 1988, 1994; South and Lloyd 1992; Upchurch et al. 2002; Willis 1999). Accordingly, nonmarital births have increased dramatically among disadvantaged less-educated women (Ellwood and Jencks 2004). Furstenberg (2003:36) contended that “marriage has become something of a luxury good” reserved for advantaged women. And as advantaged and educated women are reluctant to engage in nonmarital childbearing, marital delay is associated with fertility delay (Caucutt et al. 2002). Moreover, with increasing availability and acceptability of childcare (Rindfuss and Brewster 1996), married women who are advantaged, are educated, and have firmly established careers can use their higher incomes to buffer some of the time and energy costs of raising children (Martin 2000; Spain and Bianchi 1996). Insofar as employers strive to keep their most valuable employees, high-ability educated women may also effectively negotiate maternity leaves without deflecting their career trajectories.
Pre-Treatment and Treatment Effect Heterogeneity and College Effects on Fertility
An alternative explanation for the observed relationship between education and fertility is “pre-treatment heterogeneity bias” or “selection bias” (Upchurch et al. 2002). If unobserved factors are correlated both with selection into higher education and fertility patterns, estimates of college effects based on fertility comparisons between women with and without higher education will be biased. In addition to the concern over factors such as unobserved ability and motivation influencing educational attainment and fertility, the relationship between education and fertility is the result of a complex, potentially reciprocal process (Ahituv and Tienda 2004; Rindfuss et al. 1980). A woman who intends to go to college may lower her fertility preference and take precautions to avoid pregnancy (Marini 1984), whereas a woman who has low aspirations and expectations regarding her educational prospects may begin a family early in her life course (Upchurch et al. 2002).
Individuals differ not only in background attributes but also in their behavioral response to a college education. Although scholars have recognized pre-treatment heterogeneity in the relationship between education and fertility, they have not attended to treatment effect heterogeneity, or systematic variation in the effect of education on women’s fertility by the likelihood that women go to college. An estimate of the effect of education on fertility is a weighted average of heterogeneous effects, a quantity that depends on population composition and may vary along some predictable dimensions (Angrist and Krueger 1999; Brand and Xie 2010; Card 1999; Heckman et al. 2006; Morgan and Winship 2007).
One approach to studying treatment effect heterogeneity is to locate empirical patterns of treatment effects as a function of observed covariates. A few studies have modeled variation in college effects on fertility by examining the interaction between college and specific covariates influencing college attainment, such as race. For example, Musick et al. (2009) and Yang and Morgan (2003) found a greater effect of college on fertility for black than for white women. Although such studies offer some evidence as to variation in college effects on women’s fertility, examining interactions with individual covariates quickly exhausts degrees of freedom as more covariates are considered and, more importantly, misses their implications for selection bias. For the question of comparing college effects between women who go to college and those who do not, the most consequential interaction is between college attainment and the propensity for college attainment. We focus on variation in effects of college on fertility by the probability that a woman obtains a college education.
Theories on Heterogeneous College Effects
If we allow for variation in college effects on fertility by the propensity for college, what theoretical predictions can we make as to the pattern in effects? Prior studies addressing college effect heterogeneity have focused on the economic impact, formulating theoretical postulations accordingly. One theory is that individuals who have the highest economic returns to college are rationally the most likely to select into college (Becker 1964; Carneiro et al. 2010; Mincer 1974; Willis and Rosen 1979). It is less clear what this comparative advantage theory predicts as to heterogeneity in effects of college on fertility, but seemingly women with a high propensity for college secure high economic returns via low rates of intermittent labor force participation and low levels of fertility.
The comparative advantage perspective is not the only theory guiding research on heterogeneous college effects. A prominent tradition in sociological research on the determinants of higher education is that many non-economic factors predict college attainment because not only rational cost-benefit analyses but also cultural and social norms and circumstances govern college-going behavior (Boudon 1974; Bourdieu 1977; Brand and Xie 2010; Parsons 1937; Sewell and Shah 1968; Willis 1981). For individuals in socially advantaged positions, college is a culturally expected outcome and thus less exclusively and intentionally linked to economic gain than it is for individuals in less-advantaged groups, for whom college is a novelty that may well demand economic justification (Brand and Xie 2010). Moreover, earnings prospects are particularly bleak for disadvantaged low-skilled workers, who increasingly face limited labor market opportunities, yielding such individuals an acutely significant benefit to college. By contrast, individuals with more-advantaged social backgrounds, in the absence of college, can still rely on their advantaged social and cultural capital and ability. Brand and Xie (2010) found empirical support for heterogeneity in the economic returns to college, and evidence that individuals with the lowest probability of completing college benefit the most, rather than the least, from college. They label this a pattern of negative selection. Other studies similarly have found evidence suggesting negative selection in heterogeneous effects of schooling (Brand and Halaby 2006; Bryk et al. 1993; Card 1995, 2001; Goldthorpe and Jackson 2008; Hoffer et al. 1985; Morgan 2001).
We hypothesize that women with a lower propensity for college attendance have larger fertility-decreasing effects of college than women with a higher propensity for college attendance. Because low-propensity women are resource constrained, they are acutely interested in an economic payoff from college if they attend. Strong economic motives associated with college-going also dictate delayed childbearing. In contrast, low-propensity women who do not attend college have poor labor market prospects and attendant low economic opportunity costs coupled with cultural norms of success via motherhood; these factors motivate relatively high fertility. College attendance, thus, should generate a large fertility-decreasing effect for low-propensity women. We expect the fertility differences between college and noncollege women to decline as we observe women from backgrounds more predictive of college attendance.
Analytic Strategy
Our analysis proceeds in four steps. First, we estimate the probability of two distinct treatment conditions based on a set of observed covariates: (1) the probability that a woman attends college by age 19; and (2) the probability that a woman completes college by age 23. The control groups consist of women who completed at least high school by age 19 but who did not attend college by age 19 for the college attendance analysis and did not complete college by age 23 for the college completion analysis. We restrict the sample to women with a high school diploma, as women without a high school diploma are not “at risk” of a college education. The most disadvantaged women, who are among those women with the highest fertility levels (Astone and Upchurch 1994), are thus excluded. Because of the complex and potentially reciprocal process of education and fertility, we examine timely college attendance and completion. The more time we allow for women to attend and to complete college, the more likely it is that a woman’s fertility affects her educational attainment. Although this approach aims at a causal interpretation, it may not map onto the demography of many women’s lives. A variety of college pathways increasingly characterize the route to a bachelor’s degree, particularly among relatively disadvantaged students (Goldrick-Rab 2006). Moreover, education and fertility could pose competing risks in a woman’s life rather than one purely causing the other.
represents the Level 2 intercept (i.e., the predicted value of the effect of college for the lowest propensity women), and γ represents the Level 2 slope (i.e., the change in the effect of college on fertility with each one-unit change to a higher propensity score stratum).5We use variance-weighted least squares to estimate Eq. 5, and thus we do not assume homogeneity of variances across the δs. Variances across the δs come from two sources: sampling variation (due to different sample sizes by group) and true population variance (heteroskedasticity). When we account for varying precision of Level 1 slopes estimated within strata due to sampling variation, the Level 2 slope estimate is more efficient. Heteroskedasticity is more substantively important, as it suggests that the uncertainty of treatment effects may vary across groups (Raudenbush and Bryk 2002). Disentangling the two sources is beyond the scope of this article.
Our multilevel approach is similar to propensity score matching as propensity scores characterize women’s observed fertility differences. The two methods differ in how comparisons are constructed. Using propensity score matching, comparison by treatment status is first made on an individual basis and then averaged over a population. Using our approach, comparison by treatment status is first constructed for a relatively homogeneous group based on propensity scores and then examined across different groups of similar propensity scores through a multilevel model.
Fourth, we conduct auxiliary analyses. We report stratum-specific descriptive statistics of potential mechanisms to help interpret our main results.
A focus on variation in treatment effects by observed covariates is limited because we overlook heterogeneity in effects attributable to unobserved variables. Alternative models for heterogeneous treatment effects, such as switching regression or marginal treatment effects, however, depend on strong parametric or exclusion assumptions about unobservable variables. We considered a sensitivity analysis with Rosenbaum bounds (Rosenbaum 2002) to determine how strongly a hypothetical unmeasured variable must influence the selection process to undermine the Level 1 slopes. Rosenbaum bounds on the Level 1 slopes are conceptually and analytically straightforward. In the multilevel setting, however, we are principally concerned with what would undermine the implications of the Level 2 slope, for which such bounds are less conceptually and analytically straightforward. Although we can identify potential unmeasured variables that might impact selection into college and fertility, such as career and family orientation, there must be differential selection bias across propensity score strata to render spurious the trend in effects. We will proceed as if ignorability holds and discuss alternative conclusions under violations of ignorability.
Data, Measures, and Descriptive Statistics
We use panel data from the National Longitudinal Survey of Youth 1979 (NLSY), a nationally representative sample of 12,686 respondents who were 14–22 years old when they were first interviewed in 1979. NLSY women represent the late baby boom cohort, characterized by women largely carrying the brunt of housework and childcare while increasingly simultaneously working in the paid labor force (Spain and Bianchi 1996). NLSY respondents were interviewed annually through 1994 and are currently interviewed biennially. We use data gathered from 1979 through 2006. NLSY data have been used extensively for studying access to and the socioeconomic impact of college (e.g., Brand and Xie 2010; Carneiro et al. 2010) and for studying the relationship between college and fertility (e.g., Musick et al. 2009; Upchurch et al. 2002).
We restrict our sample to women (n = 6,283) who were 14–17 years old at the baseline survey in 1979 (n = 2,736), who had completed at least the 12th grade when they were 19 years old (n = 2,090), and who did not have missing data on college attainment (n = 2,013). We set these sample restrictions so that all measures we use are pre-college, particularly ability, and to compare college-educated women with women who completed at least a high school education. We impute missing values for our set of pre-treatment covariates. For most variables, values are missing for 1%–2% of the sample. Only two variables are missing for more than 5% of the sample: parents’ income (355 cases) and high school college-preparatory program (135 cases).6 We lose cases in Poisson models of fertility by age 41 (n = 1,512), mainly the result of attrition. The women we lose to attrition tend to be from less-advantaged family backgrounds and levels of achievement. All analyses are weighted by the probability of sample selection and nonresponse.
Our treated group of college attendees consists of women who completed at least the first year of college by age 19, and our treated group of college completers consists of women who completed college by age 23. Our control group of non–college attendees consists of women who completed high school by age 19 but did not attend college by age 19, and our control group of non–college completers consists of women who completed high school by age 19 but did not complete college by age 23. Of those women with some college by age 19, roughly one-half completed college by age 23, and two-thirds completed college by their early 40s. About 40% of non–college attendees attended college later, although less than 14% completed college and less than 12% of non–college completers by age 23 completed college later. “Non–college attendees” who attend college at some future point represent a distinct treatment group who are, on average, more disadvantaged than timely college attendees (Rosenbaum et al. 2006).7 We do not restrict the control group to women who never attend college; we follow Brand and Xie (2007) in this regard and collapse all future paths when assessing the treatment at a particular time. That is, we focus on whether a college education occurs at a particular time and remain agnostic about future educational acquisition, allowing the reference to be a composite of future counterfactual paths.
We describe the observed pre-college covariates in Table 1. These measures have figured prominently in studies of educational attainment, and their measurement is mostly straightforward. Parents’ income is measured as total family income in 1979 dollars. Friend’s plans indicate the highest level of schooling a respondent reported that his or her friend planned to obtain in 1979. Parents’ encouragement indicates whether the most influential person in the respondent’s life (in more than two-thirds of cases, a parent) would disapprove if the respondent did not go to college. In 1980, 94% of the NLSY respondents were administered the Armed Services Vocational Aptitude Battery (ASVAB), comprising 10 intelligence tests measuring knowledge and skill in areas such as mathematics and language. We residualize separately by race and ethnicity each of the ASVAB tests on age at the time of the test; standardize the residuals to mean 0 and variance 1; and construct a scale of the standardized residuals (α = .92) with a mean of 0, a standard deviation of 0.75, and a range of −3 to 3 (Cawley et al. 1997). We evaluate women with comparable fertility histories prior to treatment status by including an indicator for whether a woman had a child by age 18 for the college attendance models, and indicators for whether a woman had a child by age 18 and had a child by age 22 (i.e., before and during college) for the college completion models.
We report descriptive statistics of all pre-college variables and fertility in Table 1. The likelihood of college varies by race and ethnicity with non-Hispanic whites more likely to attend and complete than blacks and Hispanics. College goers are more likely to have families with high incomes, highly educated parents, intact families, and few siblings than noncollege women. Additionally, women with higher levels of secondary school academic success and higher levels of cognitive ability are more likely to attend and complete college. Students who received high levels of encouragement from parents to attend college and had friends with high educational aspirations are more likely to attend and complete college. Finally, women who had children in their teens or early 20s are less likely to attend and complete college. These descriptive statistics suggest that many non-economic factors figure in the educational attainment of young women.
Results
Propensity Score Models
We first derive estimated propensity scores for each woman in the sample by using probit regressions of college attendance and college completion on the set of pre-college covariates described in Table 1.8 The results reported in Table 2 support the literature on the determinants of college attainment. Race predicts college completion, although not college attendance, net of socioeconomic background. Parents’ income is also significantly associated with timely college completion. Ability and achievement in high school strongly predict timely college attendance and completion, as do parents’ encouragement and friend’s plans. For instance, the predicted probability of attending college is about .3 for college-preparatory women and .2 for women who were not in a college-preparatory track, with other independent variables held at their mean. Ability, academic achievement, and parents’ encouragement are among the strongest predictors of college attainment, as is pre-college fertility, demonstrating the reciprocal relationship between women’s schooling and childbearing. The predicted probability of attending college is .24 for women who did not have a child by age 18 and .04 for women who had a child by age 18.
Homogenous College Effects on Fertility
We report the estimated effects of timely college attendance and completion on number of children using Poisson regression models and on the probability of first birth using discrete-time event-history models controlling for the estimated propensity score under an assumption of treatment effect homogeneity in Table 3. The units of analysis are persons for the Poisson regression models and person-years for the event history models. Risk sets range from age 19 to 41 for the college attendance event history model and from age 23 to 41 for the college completion event history model. Women who had a first birth prior to age 19 for the attendance models and prior to age 23 for the completion models are no longer at risk of a first birth and are excluded from the respective analyses, limiting the sample to more-advantaged women.
The results suggest a significant 12% average decrease in the number of children women bear by age 41 associated with college attendance and a nonsignificant 8% average increase associated with timely college completion. Controlling for college attainment, the number of children decreases as women’s propensity for college increases. We also find a large significant effect on the probability of first birth associated with attending college, but not with completing college. The college attendance effect on the probability of first birth peaks when women are in their mid- to late 20s and approaches zero in the late 30s. The small, positive average college completion effect on a first birth remains constant from age 23 to 41 (figures available from the authors upon request).
Heterogeneous College Effects on Fertility
Average effects of college on fertility reported in Table 3 conceal underlying systematic college effect heterogeneity shaped by the population composition of college goers. To assess college effect heterogeneity, we first generate balanced propensity score strata. We estimate propensity scores from the probit regressions described earlier.9 Balance is satisfied when within each interval of the propensity score, the average propensity score and the means of each covariate do not significantly differ between treated (college) and control (noncollege) units (Becker and Ichino 2002). The frequency distributions for the college and noncollege women run in opposite directions: for college-educated women, the frequency count increases with the propensity score, whereas for non-college-educated women, the count decreases. There is, however, overlap within each stratum (i.e., for each propensity score stratum, there are both college- and non-college-educated women).10
Table 4 provides covariate means by propensity score strata and college attendance.11 These statistics demonstrate the characteristics of a typical woman within each stratum. Women whose parents are high school dropouts, and who have four siblings, have low ability, are enrolled in a non-college-preparatory track, and have friends who do not plan to go to college are characteristic of Stratum 1. By contrast, women who have parents with some college, have two siblings, have high ability, are enrolled in a college-preparatory track, and have friends who plan to go to college are characteristic of Stratum 6. Roughly one out of five low-propensity women who attend college have had a child by age 18, whereas we observe very few high-propensity teenage mothers.12
is the sample mean, and S2 is the sample variance for the treated and control groups as indexed by d = (1,0). The standardized difference is clearly larger in some strata than in others for some covariates, suggesting that what differentiates who attends college and who does not differs between more- and less-advantaged women. Bias between college and non–college goers’ ability, parents’ education, and residence are largest in Stratum 1, while bias between college and non–college goers’ parents’ income and family structure are largest in Stratum 6.Table 5 reports our main results, our multilevel models of heterogeneous effects of college on fertility.13 The Level 2 slope for the college attendance Poisson model indicates a significant reduction in the fertility-decreasing effect of college, a difference of 0.10, for each unit change in propensity score rank. Level 1 estimates range from a 65% decrease (incidence rate ratio) in the number of children for women with a low propensity to attend college (Stratum 1), to an 18% decrease in Stratum 2, to an 8% increase in the number of children for women with a high propensity to attend college (Stratum 6).14 Results from the college attendance multilevel event-history model complement the Poisson model, indicating a large significant decline in the probability of first birth among low-propensity women, as well as a significant reduction in the fertility-decreasing effect as the propensity for college increases.
The Level 2 slope for the college completion Poisson model indicates a significant 0.17 reduction in the fertility-decreasing college effect for each unit change in propensity score rank. Level 1 estimates range from a 48% decrease in the number of children for women with a low propensity to complete college (Stratum 1) to a 42% increase in the number of children for women with a high propensity to complete college (Stratum 6). Results from the multilevel event-history model also suggest a large college effect on the probability of first birth among low-propensity women and a significant trend in effects by propensity score strata, complementing the pattern in effects we observe in the Poisson model. We observe a stronger fertility-decreasing impact for low-propensity college attendees relative to low-propensity completers, suggesting that attending even some college produces a particularly consequential alternative life path for disadvantaged women. That advantaged women, by both ascribed and achieved statuses, are likely to have social and economic support for childbearing may explain the fertility-increasing college completion effect among high-propensity women.
Figure 1 for college attendance and Fig. 2 for college completion summarize the multilevel Poisson results in Table 5. The dots in Figs. 1 and 2 represent point estimates of Level 1 slopes, stratum-specific Poisson regression effects of college on number of children by age 41. The linear plots in the figures are the Level 2 variance-weighted least squares slopes. We reverse the y-axis scale to emphasize the fertility-decreasing effect of college. The results for college attendance and completion are similar in that the Level 2 slopes indicate a comparable significant decline in the fertility-decreasing college effect as the propensity for college increases. The figures depict the close correspondence between the Level 1 college effect estimates and the Level 2 linear slopes. A notable exception is the atypically large effect of college attendance on fertility for low-propensity women who attend college.
Figure 3 for college attendance and Fig. 4 for completion illustrate the propensity stratum-specific college effects on the probability of first birth using discrete-time event-history models, offering a depiction of heterogeneous age patterns in the onset of childbearing. Again, the risk set is age 19 through 41 for the college attendance models and age 23 through 41 for the completion models. Figs. 3 and 4 confirm Figs. 1 and 2 in suggesting systematic variation in the college effect on the probability of first birth, reinforcing and elaborating those results. We find that the large fertility-decreasing impact of college attendance for the most disadvantaged women concentrates in their early to mid-20s and flattens in their early to mid-30s, as fewer women remain at risk of a first birth. Women with a moderate propensity to attend college have a mid-range fertility-decreasing college effect, which flattens in women’s early to mid-30s. And women with the highest propensity to attend college have no effect of college attendance on fertility—a flat nil effect from age 19 through 41. Figure 4 for college completion is like Fig. 3 in that there is systematic variation in college effects on the probability of first birth, with the fertility-decreasing college effect concentrated among lower-propensity women. Figure 4 differs from Fig. 3 in that the fertility-decreasing effect for low-propensity women is smaller. As we note earlier, a critical portion of the fertility-decreasing college effect occurs prior to age 23. Women with a moderate propensity to complete college have no effect of college on first birth age 23 through 41. Although college completion increases the probability of first birth for high-propensity women, particularly in their 30s, college attendance alone does not.
Auxiliary Analyses
To further elaborate the mechanisms underlying the heterogeneous effects we observe, we present mean number of children, cumulative years employed, and cumulative years married by propensity score strata in Table 6. The mean number of children by age 41 for non-college-educated women decreases as the propensity score increases (from 2.1 to 1.6 for non–college attendees and from 2.3 to 1.3 for non–college completers), but the mean number of children increases as the propensity score increases for college-educated women (from 0.8 to 1.7 for college attendees, and from 1.1 to 1.7 for college completers). These statistics support our hypothesis: marked high fertility among disadvantaged noncollege-educated women stands in contrast to low fertility among disadvantaged college-educated women. Relatively low fertility undifferentiated by college status among more-advantaged women generates smaller effects of college as the propensity for college increases.
We also hypothesize that low-propensity college-educated women are especially economically motivated, and this may help explain lower overall fertility. We would expect then that such women spend more time employed. Table 6 reports the proportion of years that women reported they were employed between ages 25 and 41. Results lend support for our hypothesis. Low-propensity educated women have the highest levels of employment. College appreciably increases the time employed for disadvantaged women: we observe a large discrepancy (17% difference) between low-propensity women with and without college education. Marriage patterns offer another mechanism for differential fertility between low- and high-propensity college-educated women. College is associated with lower levels of nonmarital childbearing (Mincieli et al. 2007), and we find that the number of years spent married between ages 25 and 41 generally increases with the propensity for college.
Summary and Discussion
In the general population, adjusting for background factors that dispose women to go to college and affect fertility, college attendance postpones and decreases overall fertility, but college completion has no significant effect on fertility. However, these are averages of varying effects across groups of women classed by their precollege social backgrounds and levels of early achievement. We use multilevel models to estimate Level 1 regression slopes for propensity score stratum-specific effects of college on number of children and probability of first birth and Level 2 slopes for the trend in these effects across strata. We find a statistically significant reduction in the fertility-decreasing effect of college attendance and completion as women’s propensity for college increases. That is, disaggregating the effects of college shows that comparatively disadvantaged women who attend or complete college have lower and delayed fertility than similar women who did not attend or complete college. The effects of college attenuate as we consider women from backgrounds more predictive of college attendance. If we focus on college completion, rather than college attendance, the effect may even reverse. A woman from the kind of background that disposes one to complete college is likely to have more children than a woman from the same background who, for one reason or another, did not complete college. Multilevel propensity score stratum-specific event-history models show that the college effect on first birth among low-propensity women occurs largely in women’s early 20s. That the fertility-decreasing effect of college concentrates among young women from disadvantaged backgrounds is an important finding for the literature on education and fertility.
Our empirical results support sociological theory. Educated women from disadvantaged social backgrounds seemingly use college for economic gain and consequently limit fertility, while less-educated women from disadvantaged backgrounds have particularly poor labor market prospects and deem motherhood their means to personal fulfillment. These hypothetical mechanisms are consistent with our finding of a large fertility differential between more- and less-educated women from disadvantaged backgrounds. Moreover, although educated women with disadvantaged social backgrounds may perceive and encounter high role incompatibility between market work and fertility, educated women with advantaged social backgrounds are more likely to have a sense of personal efficacy, egalitarian gender role attitudes, employers willing to adjust to their family needs, and husbands who make good money (McLanahan and Adams 1987; Waite and Goldsheider 1992). Such financial and social resources translate into domestic assistance and childcare, making it possible to have children without worrying about whether they will have support, can afford it, or will lose their jobs in the process.
Our findings yield a new result that is empirically robust and theoretically coherent. Nevertheless, this study has several limitations. First, results such as ours are always subject to the possibility that some important omitted variables differentiate college goers from non–college goers. If the effects of college on fertility among low-propensity women were more strongly subject to unobserved heterogeneity than effects for high-propensity women, the pattern in effects by strata would be flatter than that which we observe. But given the systematic pattern of observed effects, there would need to be sizable biases across strata to render our results invalid. Second, we focus on the heterogeneous effects of timely college attendance and completion. We impose this age restriction because of the potentially reciprocal process between education and fertility and because our analytic strategy is sufficiently complex. Nevertheless, future research should develop a dynamic model of college effect heterogeneity on fertility with time-varying treatments and outcomes that works to address potential reciprocity between education and childbearing. Third, by focusing on the effects of college, we necessarily omit study of some of the most disadvantaged women, those who do not complete high school, who are central to many debates on social class differentiation in family patterns. Finally, fertility is but one aspect of family life. We plan to consider heterogeneous effects of college on marriage and cohabitation in future research.
Recent research on economic returns to college suggests that individuals who are least likely to obtain a college education benefit most from college (Brand and Xie 2010). We offer evidence for a corresponding result for fertility: women who are least likely to obtain a college education experience the largest effect of college on fertility. Whether or not we describe this result as low-propensity women benefiting the most from college, however, is more dubious than in the case of earnings. On the one hand, college may alter the characteristic path these women would have journeyed, marked by single motherhood in young adulthood and correlated socioeconomic adversity. On the other hand, upward mobility for women from disadvantaged social backgrounds comes with its own opportunity cost: fewer or no children. Relative to a high-propensity educated woman’s life course, a low-propensity educated woman remains disadvantaged with respect to family formation. Thus, even equalizing educational attainment, polarization of family formation patterns by socioeconomic origins persists.
Acknowledgement
Financial support for this research was provided by the National Institutes of Health, Grant 1 R21 NR010856-01; a UCLA Faculty Research Grant; and by the California Center for Population Research at UCLA, which receives core support from the National Institute of Child Health and Human Development, Grant R24 5R24HD041022. Versions of this paper were presented at the Department of Sociology and the Institute for Policy Research at Northwestern University, Center for Poverty and Inequality at Stanford University, and the Population Association of America 2009 Annual Meeting. We thank Paula England, Ben Jann, Yana Kucheva, Robert Mare, Kelly Musick, Fabian Pfeffer, Hiromi Ono, Judith Seltzer, Yu Xie, and several anonymous reviewers from Demography for helpful suggestions. The ideas expressed herein are those of the authors.
Notes
In addition to the lower likelihood of labor force participation, Budig and England (2001) found that employed mothers suffer an average per-child wage penalty of approximately 5%, possibly resulting from employer discrimination against mothers (Correll et al. 2007). Amuedo-Dorantes and Kimmel (2005) found, however, that educated women who delay fertility do not experience a motherhood wage penalty.
Musick et al. (2009) argued that the effect of college on fertility is largely the result of unintended births and thus questioned whether opportunity costs explain fertility differences by education. Still, unintended fertility differences may be associated with opportunity costs if disadvantaged women lack a strong sense of efficacy and positive future outlook because of fewer economic and traditional family prospects and are thus less likely to take precautions to avoid pregnancy (Edin and Kefalas 2005; Mirowsky and Ross 2007).
We use Poisson rather than negative binomial models because we did not find evidence of overdispersion (i.e., the variance of the outcome is not greater than the mean of the outcome).
We explored several representations of age and decided that a squared term adequately represented the observed curvature.
Normality of ε is assumed for inference but not for summary slope estimation.
Given the additional complexity of combining multiple imputed data sets with our heterogeneous treatment effects analyses, we use single imputations. We note, however, that single imputations typically suffer from the problem of over-fitting and do not fully represent the uncertainty in the procedure.
Roughly two-fifths of timely college attendees began at a community college compared with two-thirds of those who attended college after age 19. Less than one-fifth of timely college completers began at a community college.
In order to achieve balanced propensity score strata, the covariates are slightly different in our two model specifications.
Because we use regression models to generate propensity scores, different specifications result in different classifications of individuals to strata. We tried several specifications before we decided on our model reported here. Our results were robust to these alternative specifications.
We did not initially, however, have a sufficient number (roughly 20 cases) of non–college goers in the final stratum for the multilevel model. We therefore collapsed the final two strata and adjusted for the estimated propensity score in Level 1 analyses for the college attendance and completion models. We also did not initially have a sufficient number of college completers in the first stratum, and we therefore also collapsed the first two strata and adjusted for the estimated propensity score for the college completion model.
Hispanic was not balanced in Stratum 1, as shown in Table 4, for the college attendance model and was thus added as a covariate in our Level 1, Stratum 1 model.
Comparing Table 4 with an analogous table for college completion (results available upon request), we observe that women with a college degree and their matched controls are more advantaged than women with some college and their matched controls.
To facilitate implementation of our method, we use the Stata module –hte– (Jann et al. 2010; available for public use).
In results not shown, we also explored heterogeneous effects of college attendance by age 20, age 21, and age 22 on number of children by age 41. Substantive conclusions are comparable with analyses based on college attendance by age 19.
