Although teenage mothers have lower educational attainment and earnings than women who delay fertility, causal interpretations of this relationship remain controversial. Scholars argue that there are reasons to predict negative, trivial, or even positive effects, and different methodological approaches provide some support for each perspective. We reconcile this ongoing debate by drawing on two heuristics: (1) each methodological strategy emphasizes different women in estimation procedures, and (2) the effects of teenage fertility likely vary in the population. Analyses of the Child and Young Adult Cohorts of the National Longitudinal Survey of Youth (N = 3,661) confirm that teen pregnancy has negative effects on most women’s attainment and earnings. More striking, however, is that effects on college completion and early earnings vary considerably and are most pronounced among those least likely to experience an early pregnancy. Further analyses suggest that teen pregnancy is particularly harmful for those with the brightest socioeconomic prospects and who are least prepared for the transition to motherhood.
Since becoming a contentious U.S. political issue more than 40 years ago, researchers continue to dispute the socioeconomic consequences of teenage pregnancy and childbearing. Early evidence that teen childbearing substantially reduced educational attainment (Card and Wise 1978; Hayes 1987; Mott and Marsiglio 1985) implied that the costs of motherhood force women to sacrifice investments in their own education and training (Becker 1981; Manlove 1998). Scholars have since questioned whether these studies account for preexisting factors that affect women’s fertility decisions and attainment. Consequences may be negligible, for instance, if teen mothers have bleak socioeconomic prospects prior to childbearing (Furstenberg 2003; Geronimus 1997; Geronimus and Korenman 1992). Others have suggested that the responsibilities of motherhood could even serve as a positive turning point in the lives of troubled youth (Brubaker and Wright 2006; Edin and Kefalas 2005).
Scholars have attempted to resolve this debate by isolating the causal effect of teen childbearing. Efforts to reduce selection bias—that is, confounding factors that influence both early fertility and attainment—are the focal concern in such analyses. A variety of methodological strategies are employed to reduce bias, including regression, fixed effects, instrumental variables, and propensity score methods. More rigorous studies have found that teenage childbearing has negative effects on later education or wages (Ashcraft et al. 2013; Hoffman et al. 1993; Stange 2011), and although effects are smaller than originally suspected, they vary considerably across methods and samples. Concern with selection is clearly justified but is insufficient to resolve the ongoing debate.
We argue that the debate persists because the preoccupation with selection bias has come at the expense of ignoring effect heterogeneity. After all, the presence of selection itself suggests that young women have different predispositions for pregnancy and childbearing, and there is no reason to assume that this selection process is unrelated to the consequences of early fertility. Furthermore, scholars have rarely acknowledged that different methods identify treatment effects specific to subpopulations that differ in the likelihood of experiencing teenage fertility. If the effect varies among these groups, techniques will yield distinct causal estimates that are equally valid. What is most striking is that the range of documented effects aligns with theoretical perspectives after heterogeneity is considered.
This study, which deliberately focuses on teenage pregnancy, makes several contributions. First, we argue that key theories imply that the effect of teen pregnancy and childbearing systematically varies among women who have different predispositions toward early fertility, partly because of skills and resources acquired in childhood. Second, we demonstrate that the various strategies to identify causal effects emphasize subgroups of women who differ in their likelihood of teen pregnancy and childbearing. Given that methods presumed to be more rigorous emphasize women who are particularly likely to become pregnant as teens, we argue that the smaller effects in such studies are partly a consequence of effect heterogeneity rather than selection bias alone. By threading theory, methods, and empirical evidence together, we construct a more nuanced framework for thinking about the consequences of teen fertility.
Although prior literature has generally highlighted the consequences of teenage childbearing, teenage pregnancy is our treatment of interest. Casting pregnancy as the treatment reduces concerns with selection bias and better aligns with public policy efforts to reduce teen fertility. Our analysis uses inverse probability weighting methods to identify the effects of teen pregnancy on young women’s educational attainment and earnings across women with different risks of teen pregnancy. We also assess whether sources of effect heterogeneity are tied to young women’s personal attributes, fertility-related experiences, home environments and parental influences, other contextual influences, and sociodemographic characteristics. Our findings take a major step toward reconciling this contentious literature and informing social policy to improve the well-being of young women.
Consequences of Teenage Childbearing
Clearly, teen mothers have lower levels of socioeconomic attainment than their childless peers: they are less likely to complete high school, attend college, or earn a bachelor’s degree, and they tend to earn less and are more likely to experience poverty (Card and Wise 1978; Fergusson and Woodward 2000; Hofferth 1987; Holmlund 2005; Upchurch and McCarthy 1990; Waite and Moore 1978). These disparities also extend across generations: children of teenage mothers exhibit lower cognitive and noncognitive skills and are more likely to become parents as teenagers themselves (Campa and Eckenrode 2006; Furstenberg et al. 1987; Hao and Cherlin 2004). However, it is unclear whether such patterns are truly consequences of early fertility.
Does Teenage Fertility Cause Lower Attainment?
Scholars accepting a causal interpretation propose a number of mechanisms by which teen pregnancy or childbearing may lower attainment. Perhaps the most influential explanation is that of opportunity costs (e.g., Becker 1965). When a young woman has a child, her ability to acquire human capital is severely limited by caregiving and other responsibilities associated with motherhood (Becker 1981). Such time and energy constraints are likely to hinder educational attainment as well as employment opportunities; a young mother may also lack the necessary credentials or work experience to earn sufficient wages. Similarly, teen childbearing is associated with an increased need for material goods. Rather than investing in their own attainment, young mothers are forced to seek financial resources or childcare arrangements (Mollborn 2007).
Another perspective focuses on the stress associated with a rapid role transition. Specifically, an accelerated transition to motherhood may result in adverse psychological effects (Bacon 1974; Coleman 2006; Hagestad 1990) that impede human capital accumulation.
These accounts provide plausible causal explanations for the consequences of early fertility, but they also imply that the consequences should vary among women according to their personal attributes, skills, and resources. On the one hand, opportunity costs may be highest for young women with promising socioeconomic prospects, such as those who display higher intellectual skills or come from more socioeconomically advantaged families. On the other hand, opportunity costs could be lower for those who can acquire material and social resources to assist with childrearing while simultaneously furthering their own attainment. Adolescents from comparably disadvantaged families may receive less assistance, particularly if their parents are unable to provide social or emotional support (e.g., Barber et al. 1999). The stress of accelerated role transitions may also vary, being most severe in contexts where teen pregnancy or motherhood is rare, unexpected, or not socially sanctioned. We would thus expect greater consequences among young women who may be less prepared for motherhood or face greater stigma.
Selection Into Teenage Fertility
Causal interpretations of early fertility are often critiqued. Because teenage births are concentrated among those who are most disadvantaged, these disadvantages rather than early childbearing may be the direct cause of young mothers’ lower socioeconomic attainment (Geronimus and Korenman 1993; Hotz et al. 2005; Ribar 1994). In other words, the lower attainment of teenage mothers could be attributable to preexisting differences—such as skill deficits—that limit their socioeconomic prospects in the absence of a birth. Others have suggested that the family environment—including fertility-related attitudes in the home—is correlated with intergenerational patterns of teen fertility as well as lower attainment (Kahn and Anderson 1992; Manlove 1997).
This perspective also implies variation in effects rather than no effects whatsoever. Paralleling the causal explanation of opportunity costs, the selection perspective anticipates that the socioeconomic consequences should be lowest for young women who lack the skills and resources to achieve educational and economic success even in the absence of a teenage pregnancy or live birth.
Advantages for the Disadvantaged
From yet another perspective, early motherhood may hold benefits for some adolescents. Qualitative work by Edin and Kefalas (2005) suggests that some young women view motherhood as a positive turning point in their lives, motivating them to avoid delinquency, return to school, and search for employment. Similarly, others (e.g., Brubaker and Wright 2006; Frost and Oslak 1999; Zachry 2005) have found a heightened sense of responsibility and stability among those who experience an early birth. Such positive effects appear most likely among women experiencing struggles in their schools or homes. Although these studies follow small, particularly disadvantaged samples, they provide further insight into the notion that women differentially respond to motherhood.
An Empirical Puzzle
Scholars wading into the causal effect debate have largely ignored the effect heterogeneity implied by these perspectives. Although problematic, this omission is understandable given the immense challenges posed by selection bias. Researchers have thus adopted a variety of methods to adjust for preexisting factors that could confound the relationship between teen fertility and attainment. Given that effects decline under more rigorous methods, antecedent confounders are clearly a concern (Ashcraft and Lang 2006; Holmlund 2005; Upchurch and McCarthy 1990).
Depending on the analytic strategy employed, the magnitude of estimates ranges from 0 to a penalty of nearly 2 years for educational attainment (Fletcher and Wolfe 2009; Kane et al. 2013; Lee 2010). The evidence for earnings is slightly more nuanced. Some have argued that teenage mothers suffer a penalty in earnings compared with young women who did not experience a birth (Blackburn et al. 1993; Ermisch 2003), but others have found that they receive higher earnings well into early adulthood (e.g., Hotz et al. 2005). The effects of early childbearing on earnings appear largely indirect, operating through educational attainment and labor market attachment (Klepinger et al. 1999).
Although current estimates of teenage childbearing effects generated by scholars suggest that the magnitude is not as large as initially believed, the range of findings perpetuates debates about extent to which teenage fertility harms women’s life chances. Yet, the opposing perspectives and inconsistent evidence across studies are not as contradictory as they seem. If the effects of early fertility systematically differ—as most theories imply—we would expect estimates to vary across samples and methods. We argue that the consequences of early fertility will differ according to young women’s personal attributes, fertility-related experiences, home environment and parental influences, contextual influences, and sociodemographic characteristics—factors that also influence the likelihood of experiencing early fertility. Furthermore, we stress that different methodological approaches provide inconsistent results partly because they emphasize different types of young women.
Reconciling the Methodological Debate
Scholars have used a variety of approaches to estimate the casual effect of teen pregnancy and childbearing on later attainment. These strategies include ordinary least squares (OLS) regression, sibling fixed effects (FE), instrumental variables (IV) estimation, and propensity score matching (PS) techniques. Although rarely acknowledged in the teen childbearing literature, these techniques provide different approximations of the causal effect of interest. We highlight how each approach identifies a particular treatment effect from a larger distribution of effects. Consequently the majority of studies examining the costs of teen childbearing may be correct despite differences in the direction or magnitude of estimates.
We discuss four methods—OLS, FE, IV, and PS—borrowing language from the treatment effects literature. Specifically, the treatment refers to teen pregnancy; treated cases are those who become pregnant as teens; and control cases are those who do not become pregnant as teens. The average treatment effect (ATE) applies to typical young women in the population, and the treatment on the treated (TOT) and treatment on the control (TOC) effects apply to typical young women who do and who do not become pregnant as teens, respectively.
If the treatment effect is constant regardless of X, and teen pregnancy is conditionally independent of unmeasured influences on attainment (εi), the regression coefficient (δ) yields the ATE. Under heterogeneous treatment effects, however, this estimate differs from the arithmetic average across sample respondents (Elwert and Winship 2010). Instead, regression provides a variance-weighted average of the treatment effect, placing most weight on groups of observations where treatment variation is the largest; this variance is maximized when the probability of teen pregnancy reaches approximately .5, and it declines as the probability nears 0 and 1 (Angrist and Pischke 2009). Because vital statistics indicate that rates of teen pregnancy are much lower than .5 in the population, OLS weights observations with above-average probabilities of teen pregnancy most heavily, yielding a treatment effect tending toward the TOT.
Early OLS estimates have suggested substantial declines in educational attainment among teen mothers, ranging from 2 to 4 fewer years of schooling (Card and Wise 1978; Moore and Waite 1977; Mott and Marsiglio 1985; Waite and Moore 1978). However, this approach has been largely abandoned in favor of methods that relax functional form assumptions and better account for unobservable confounders.
Fixed Effects (FE)
One cost of exploiting between-sibling variation is reduced generalizability: the sample is restricted to families with multiple daughters. Another is that identification rests on families where the treatment varies. Hence, young women from larger families with higher rates of teen pregnancy than the population disproportionately contribute to these estimates. FE models can also be viewed as OLS models controlling for family identifiers; from this perspective, they provide variance-weighted average treatment effects with more weight on those from families with greater treatment variation. Hence, sibling FE models likely yield treatment effects approaching the TOT.
FE estimates in the literature are generally smaller than OLS estimates (e.g., Geronimus and Korenman 1992). The range of FE estimates falls between 0 and a 1-year reduction in schooling for teen mothers (Hoffman et al. 1993; Holmlund 2005) and suggests no significant differences in income (Ribar 1999).
Instrumental Variables (IV)
If heterogeneous treatment effects exist, IV estimates identify the local average treatment effect (LATE; Angrist et al. 1996). That is, they capture the average effect for compliers, or those induced to receive the treatment by the instrument. The effect does not necessarily apply to those who would or would not receive the treatment in the absence of the instrument. Hence, effect heterogeneity limits generalizability of IV estimates to specific subpopulations that are often difficult to define. It is also notoriously difficult to identify plausible instruments in many social science applications.
A number of studies have investigated the causal effect of teenage childbearing using an IV approach, with miscarriages being the most common instrument (Ashcraft et al. 2013; Ashcraft and Lang 2006; Fletcher and Wolfe 2009; Rindfuss et al. 1980).1 The assumption is that, on average, miscarriages are random with respect to socioeconomic outcomes—except through the reduction in fertility. This approach limits generalizability to those pregnant as teenagers, already a select population. Moreover, young women who would have another child following a miscarriage, or who would terminate the pregnancy or pursue adoption in the absence of a miscarriage, are not captured by the IV estimate. Considering the restriction to pregnant teens, estimates are likely to be close to TOT effects. IV studies tend to estimate modest effects on women’s subsequent education and earnings (e.g., Ashcraft et al. 2013; Fletcher and Wolfe 2009; Hotz et al. 2005). For instance, Ashcraft and Lang (2006) found positive but insignificant effects for earning a high school diploma as well as nontrivial increases in family income.
Propensity Score Matching
PS techniques make no functional-form assumptions about the outcome, so effect heterogeneity has different implications. Here the estimand depends on the success of the matching procedure (Rosenbaum and Rubin 1983). If all treated and control cases are matched (one-to-one), the ATE is identified and estimated by averaging the treatment and control differences across all matched pairs. Most often, however, the researcher attempts to find matches for all treated cases. If this is accomplished, the TOT is identified. If not, the population to which the estimate pertains is unclear. The same would hold for any effort to match all untreated cases in hopes of identifying the TOC.
A number of studies have used PS matching to estimate the effects of teen childbearing. Some have purportedly estimated the ATE (Chevalier and Viitanen 2003; Lee 2010), where teen mothers are nearly 40 % less likely to complete college. Others have pursued the TOT (Levine and Painter 2003; Sanders et al. 2007) and found comparably smaller effects on women’s subsequent schooling. Often, however, unmatched cases skew the estimate away from the desired estimand.
Summary of the Methodological Debate
Given the different methods employed in the literature, it is not surprising that such variation exists in causal estimates (Kane et al. 2013; Sanders et al. 2007; Upchurch and McCarthy 1990). One the one hand, each method provides different ways to address selection bias. From this perspective, the range of findings implies that some methods are more successful in eliminating bias than others. On the other hand, we have shown that each method provides effect estimates that apply to specific subsets of the population. If effects vary among women with different propensities for early fertility, discrepant findings would persist even if each approach successfully eliminated selection bias. We find it plausible that existing evidence supports all sides of the debate.
Figure 1 presents a hypothetical distribution of treatment effects that would reconcile much of this debate, plotting women’s probability of high school graduation against the probability of teenage pregnancy. The black line represents pregnant (treatment) teens, and the gray line represents nonpregnant (control) teens. At any given point on the x-axis, the vertical distance between the two lines captures the treatment effect. There are several noteworthy points.
First, both lines slope downward, indicating that the probability of graduating high school declines among women who are more likely to be pregnant teens. This is attributable to antecedent factors that heighten the risk of teen pregnancy and reduce educational attainment, which drive concerns about selection bias in this literature.
Second, the black line is beneath the gray line across much of the distribution, indicating negative teen pregnancy effects for most women. This finding supports the perspective that teen childbearing reduces educational attainment, presumably through opportunity costs, resource constraints, or stressful role transitions.
Third, the negative effect is greatest at the lowest probability of teen pregnancy and declines as the probability increases. Although not explicitly stated by advocates of the opportunity cost and selection arguments, these theories imply that the consequences of teen childbearing should be concentrated among those with the most promising socioeconomic prospects—who are least likely to experience a teen pregnancy—precisely because they have the most to lose.
Fourth, at the far right of the distribution, the effect becomes positive for those with the highest probability of a teen pregnancy. This pattern mirrors what has been documented in qualitative work, where higher rates of high school completion are reported among some young mothers. Proponents of this argument suggest that early childbearing may provide an increased sense of direction and responsibility among struggling adolescents.
We denote ATE, TOC, and TOT effects by dotted vertical lines placed at the average predicted probability of teen pregnancy for the population (.17), nonpregnant teens (.15), and pregnant teens (.39); these predicted probabilities come from our own data.2 We also illustrate the plausible range of the effect distribution highlighted by common methods: OLS regression, FE, IV, and PS. As shown in the prior section, OLS and PS methods target the ATE but often deviate toward the TOT, whereas IV and FE methods place the most weight on those who are most likely to become pregnant as teens, approximating or even surpassing the TOT. Extant literature has generally found OLS and PS estimates to be larger than FE and IV estimates. Thus, even if prior work succeeded in removing selection bias, the range of findings is consistent with positive selection, in which negative effects are less pronounced among women more likely to have a teenage pregnancy.
Not only could this figure reconcile the long-standing debate surrounding the consequences of teenage childbearing, but it also provides a new framework for thinking about the effects of teen pregnancy and childbearing. Specifically, we argue that negative, trivial, or positive effects could be simultaneously occurring among different types of women in the population. This perspective moves away from identifying the causal effect in favor of investigating whether there are unequal consequences of teen pregnancy among subpopulations of young women with different predispositions for early fertility.
Pregnancy as Treatment
As mentioned earlier, our analysis focuses primarily on respondents who experience teenage pregnancy—not necessarily those who had a live birth. This treatment effect captures how young women cope with pregnancy (which could result in an abortion, a live birth, or a miscarriage) and other consequences following this major life event. A number of reasons drove us to this decision.
First, this shifts the nonrandomness of childbearing following a pregnancy from a selection issue to a mechanism of the treatment effect. Evidence indicates that adolescents who terminate a pregnancy tend to come from advantaged families and score higher on achievement tests, and numerous other unobservable characteristics are likely to influence this decision (Ashcraft and Lang 2006; Coleman 2006; Fletcher and Wolfe 2009). A growing body of work also indicates that miscarriages are a nonrandom phenomenon. Although those who miscarry are less selected on cognitive ability and socioeconomic status (SES) than those who abort, they are more selected than those who carry to term (e.g., Ashcraft et al. 2013). Focusing on pregnancy avoids these issues as well as data limitations that prevent accurate reporting of miscarriages and abortions.
Second, these selection issues are magnified in an assessment of treatment effect heterogeneity because selection into childbearing differs across the risk of teen pregnancy. Those with a lower propensity to experience teenage pregnancy are advantaged in many respects, making them more likely to have an abortion or a miscarriage than comparatively disadvantaged women (i.e., those with a higher propensity). Although this selection process would produce additional bias in an analysis of childbearing effects, it suggests predictable heterogeneity in our analysis of pregnancy effects. For instance, more-advantaged teens may fare better after a teen pregnancy because they are less likely to have a live birth.
Finally, there are important theoretical and practical reasons to focus on pregnancy as opposed to live births. Even without a birth, teenage pregnancy is a pivotal event that brings social, psychological, and physiological changes likely to affect one’s well-being (Thompson 1986; Turner et al. 1990). Given that U.S. political rhetoric and policy emphasizes curtailing early pregnancy rather than preventing childbearing (e.g., Frost and Forrest 1995), and that those most critical of teenage pregnancy are likely to oppose abortion and stronger birth control measures, focusing on pregnancy is also more consistent with practical efforts to address this issue. Nonetheless, we supplement our analysis of teen pregnancy effects with a parallel analysis of teenage childbearing.
We assess how teen fertility impacts high school graduation, college attendance, college completion, and average earnings at ages 25–35, and we use two methods suited to examine systematic effect heterogeneity: smoothing-differencing and inverse probability weighting. Both methods begin with a logit model predicting teenage pregnancy using a set of covariates thought to influence early fertility and later socioeconomic outcomes. This step is virtually identical to creating a propensity score model. The resulting probabilities represent a composite that we use to identify comparable pregnant and nonpregnant teens.
The smoothing-differencing (S-D) method is an exploratory technique used to visualize treatment effects across the probability of treatment (Xie et al. 2012). This method incorporates the probability of treatment into nonparametric regressions and permits graphical illustrations similar to Fig. 1. After estimating the probability of teen pregnancy, we fit nonparametric regressions of each outcome on the predicted probability of teen pregnancy () separately for treatment and control groups.3 The difference between groups at any value of yields the corresponding treatment effect. If the treatment effect remains constant across predicted probabilities of teen pregnancy, it suggests women’s response to pregnancy does not systematically differ with respect to selection into treatment.
Although the S-D method is well suited to explore the effect heterogeneity obscured in most other approaches, it shares their failure to provide clearly interpretable estimands. For instance, it is not obvious how to derive ATE, TOT, or TOC estimates from such an analysis. Deriving these estimates is straightforward when using an inverse probability weighting (IPW) strategy described by Morgan and Todd (2008). Here the predicted probabilities of treatment are used to reweight the treated and control samples to mirror the overall population (ATE), the population of pregnant teens (TOT), and the population of those not pregnant as teens (TOC). Comparing the average outcomes of treated and control groups under these reweighting schemes yields the corresponding treatment effect. Weights (ψ) for each scheme and treatment group are shown in the following equations, where corresponds to the overall probability of each treatment condition (teenage pregnancy) in the sample.
When d = 0,
When d = 1,
Treatment effect estimates will differ if effect heterogeneity is systematically related to selection into treatment. Under positive selection, the effect will be more positive (or less negative) for those most likely to receive the treatment (i.e., TOT > TOC); the opposite will hold under negative selection.
Whereas regression-based methods seek to address bias by modeling the relationship between potential confounders and the outcome, PS-based methods address bias by removing the associations between confounders and the treatment. IPW is convenient for our purposes because it can be combined with regression to further adjust covariates when reweighting alone fails to balance potential confounders across treatment and control groups. We do this by using the IPW weights to perform weighted regressions of the outcome on the treatment along with main effects of all covariates used to construct the weights.
Another weakness of the S-D method and other PS-based approaches is the failure to provide insight to the sources of heterogeneity. The ability to combine IPW and regression is particularly useful here, allowing us to test theoretically motivated interaction terms in reweighted OLS regressions.
Data come from the Child and Young Adult Cohorts of the National Longitudinal Survey of Youth 1979 (NLSCYA). NLSCYA targets 11,502 children—5,630 females—born to more than 4,900 mothers (as of 2010) in a nationally representative cohort of women who were ages 14–21 in 1979 and were followed in the NSLY79 survey. As with all studies that use the NLSCYA, respondents in this sample were born to younger women than the average U.S. birth. Specifically, the NLSCYA women who were old enough to have reached young adulthood were born to relatively young mothers. Nevertheless, these data have unique strengths for this analysis, including several age-appropriate externally validated assessments of children’s skills, attitudes, and behaviors; information on their socioeconomic attainment as young adults; family social, demographic, and economic characteristics; and extensive information on their mothers, including socioeconomic background, skills, and other attributes during adolescence. We discuss nonrandom attrition and missing data later herein.
Our key predictor (treatment) is teen pregnancy, determined by whether the respondent’s first pregnancy occurred prior to age 18. This decision follows the age specifications of past work investigating the consequences of teenage fertility (i.e., Hotz et al. 2005; Kane et al. 2013). We exclude those respondents without reports of their age at first pregnancy (35 % of respondents) due to attrition or being too young to receive the young adult survey at the 2010 follow-up, leaving 3,661 subjects.4 Of these, 621 respondents (17 %) reported a teen pregnancy. Analytic samples are further restricted to those with complete outcome data. In our parallel analysis of teen childbearing, we examine whether the respondent gave birth by age 18. In total, 561 of 3,695 subjects with data (15.2 %) had a teen birth.
We follow a well-established literature in the social sciences that investigates consequences of teen childbearing on education completion and earnings by assessing whether respondents earned a high school diploma (N = 2,276), attended any college (N = 2,492), or earned a bachelor’s degree (N = 1,599) among those old enough to have completed each milestone. Specifically, we observe high school completion and college attendance among those reporting educational attainment after age 19, and bachelor’s degree completion among those reporting educational attainment after age 23. We also examine log-average earnings from salary and wages when respondents were between ages 25 and 35 (N = 1,527) for those reporting some earnings during the 10-year period. These four outcomes shed light on significant dimensions of attainment that are associated with the life chances for young women and their offspring.
Because our analysis uses observable data to address selection into teen pregnancy, it is crucial to have measures from a variety of domains related to pregnancy and socioeconomic attainment. We distinguish the following domains: personal attributes, fertility-related experiences, parental influence and home environment, contextual influences, and sociodemographic characteristics. In many instances, we create scales from factor analysis of multiple measures, in which case we denote maximal reliabilities (MR) based on factor loadings.5 All scales are standardized within the NLSCYA sample.
The data provide measures of several key skills and dispositions when female respondents were ages 11 to 14, prior to the vast majority (95.9 %) of first pregnancies. Intellectual skills is a scale (MR = .90) based on children’s scores on Peabody Individual Achievement Tests in reading recognition, reading comprehension, and math; the Peabody Picture Vocabulary Test; and the digit span assessment from the Wechsler Intelligence Scale for Children. A behavior problems (MR = .87) scale is based on maternal reports in six subscales of the Behavior Problems Index. General self-worth is a subscale of the Self-Perceptions Profile for Children provided by NLSCYA. Risk aversion (MR = .73) is constructed from six self-reported items assessing attitudes and behaviors related to risk-taking and planning for the future. Delinquency (MR = .93) is based on 13 self-reported items pertaining to delinquent behaviors and substance use. Educational expectations are obtained from self-reports of the highest level of schooling respondents think they will achieve (1 = less than a high school diploma, 2 = high school diploma, 3 = some college, 4 = bachelor’s degree, 5 = graduate or professional school). These measures are particularly relevant to potential heterogeneity related to the opportunity costs of early fertility, which may be highest among those with greater socioeconomic prospects.
Especially relevant to teen pregnancy are individual- and household-level experiences related to fertility and family formation (e.g., Barber 2001). We include self-reports of the age when the respondent began dating, the best age to have children, and whether the respondent ever talked to family members about sex. We also include whether the respondent had an older sister who experienced a teenage pregnancy, the mother’s age at her first child’s birth and at the focal child’s birth, as well as the mother’s expected age at marriage when she was an adolescent (<20, 20–24, 25–29, 30+, never).6 Young women for whom early fertility is expected (or in contexts where teen fertility is common) may face less-stressful role transitions or less stigma than those in other contexts.
We include several measures related to the home environment and parental influences, all measured at ages 11–14. Cognitive stimulation and emotional support scales come from the Home Observation Measurement of the Environment (HOME) inventory. Parental expectations for the respondent (MR = .63) are based on parent assessments of the child’s relative standing in school, expected educational attainment, and general prospects for the future. An additional eight-item scale captures the amount of activities the respondent participates in with her parents (MR = .90), and a nine-item scale captures the degree to which the respondent tells her parents about important events in her life (MR = .85). Young women who have stronger relationships with their parents or who have more-involved parents may be more likely to receive support after an early pregnancy, which could help buffer them from negative consequences.
Available measures of social context at ages 11–14 address a number of peer and school influences. These include the respondent’s report of whether she has no close friends; whether her closest friends are out of school or in a higher grade; and a five-item scale of negative peer pressure from friends, including pressure to skip class or use drugs or alcohol (MR = .96). School measures consist of whether the child attends private school, the number of minutes per week spent reading, and time spent on homework. These measures may serve as proxies for general attitudes about fertility, schooling, or social support outside the family—each of which could promote effect heterogeneity.
Socioeconomic and demographic factors include the respondent’s race (black, Hispanic, white), whether she lived with both biological parents at ages 11–14, average total income in her mother’s household from the respondent’s birth until age 18 (adjusted to 2011 dollars), and her mother’s highest level of educational attainment recorded after age 23 (no high school diploma, GED, high school diploma, some college, bachelor’s degree, or graduate school). We also include the average years of schooling completed by maternal aunts and uncles, which is a better predictor of teenage pregnancy than that of maternal grandparents in our analysis; this information was reported by mothers in 1983.
Missing Data and Attrition
We use multiple imputation with chained equations, separately by treatment, to impute missing items arising when children miss a wave of data collection or fail to complete an assessment.7 After excluding observations missing teen pregnancy or outcome data, we use IPW to adjust for sample selection due to attrition or children being too young to have outcome data at the 2010 follow-up. Specifically, we use imputed data to predict attrition with a logit specification, generate inverse probability weights from these predictions, and use these weights combined with the mother’s NLSY sampling weight to reweight the sample.8 This helps maintain the representativeness of the original sample and reduce bias due to nonrandom attrition. Our analytic samples are thus meant to be representative of young women who were born to the nationally representative cohort of women in the NLSY79.
Table 1 summarizes predictors and outcomes separately by teen pregnancy for the NLSCYA sample. As 11- to 14-year-old children, respondents who experience pregnancy as a teenager are more disadvantaged with respect to their personal attributes, parental influence and home environment, contextual factors, and sociodemographic background; and they differ in attitudes and behaviors related to fertility. As young adults, pregnant teens also have markedly lower educational attainment and earnings.
Specifically, pregnant teenagers exhibited more behavior problems and delinquency, as well as lower feelings of self-worth, intellectual skills, risk aversion, and educational expectations as children. At home, they received less cognitive stimulation and emotional support, they participated in fewer activities and shared fewer important events with their parents, and their parents had lower expectations for their future prospects. Similar patterns have been routinely documented in prior work (Fergusson and Woodward 2000; Levine and Painter 2003). Pregnant teens also reported spending less time reading and working on homework, and they were less likely to attend private schools. They felt more peer pressure to engage in negative behaviors, and their friends were more likely to be older or out of school.
Young women who became pregnant as teens also selected younger ages as the best age to have a child, were less likely to have talked to family members about sex, and started dating at earlier ages, possibly reflecting a greater duration of being “at risk” of pregnancy. Indicative of intergenerational similarities in fertility (Campa and Eckenrode 2006; Ribar 1994), pregnant teens were born to mothers an average of three years younger than those who did not become pregnant as teens; they were also more likely to have an older sister who experienced a teen pregnancy.
Finally, we note several socioeconomic and demographic differences between treatment and control groups. Those who experienced a teenage pregnancy were more likely to be identified as black, less likely to reside in a two-parent household, and had less-educated family members. These patterns echo earlier work finding that pregnant teenagers come from socioeconomically disadvantaged backgrounds (Hogan and Kitagawa 1985; Robbins et al. 1985).
Propensity Scores and Covariate Balance
We began with a propensity score model with main effects for all covariates, but further analyses indicated the reweighting adjustments failed to balance many covariates in the ATE and TOC analyses. To improve balance, we added quadratic and interaction terms to the model.9 Estimates from the final propensity score model are summarized in Table S1 in Online Resource 1. Overall teen pregnancy rates range from .16 to .23 across analytic samples, while average predicted probabilities range from .43 to .45 among treatment cases and .13 to .18 among control cases. Table 2 summarizes the covariates by treatment, as well as their standardized mean differences, unadjusted and under the three weighting schemes.10
Despite the improvements to the propensity score model, it was difficult to achieve balance on a number of covariates in the ATE and TOC analyses.11 In some cases, the residual imbalances continue to favor the control cases; but in others, they favor the treatment cases. Thus, many respondents in the control group lack similar counterparts in the treatment group, which leads the IPW weights to under-adjust some attributes and over-adjust others. Similar overlap problems arise in other studies that employ propensity-based techniques (e.g., Levine and Painter 2003). To the extent that educational attainment and early earnings depend on these variables, effect estimates will be biased. Balance was very good in the TOT analysis, suggesting that pregnant teens did have similar counterparts in the control group.
We supplement the IPW approach with regression to adjust for possible bias from the failure to achieve balance in the ATE and TOC analyses. That is, we run OLS regressions controlling for the covariates after reweighting the sample as discussed earlier. It is important to note that regression uses parametric assumptions (i.e., linearity) to extrapolate beyond the support of the data, which is necessary given problems with overlap in the ATE and TOC analyses. Intuitively, this means that because our sample does not contain sufficient numbers of nonpregnant teens who resemble pregnant teens on all covariates, the regression uses the estimated relations between covariates and the outcome to project the outcome for treated cases who resemble the untreated population.
High School Completion
Figure 2 uses the S-D method to plot the predicted probability of high school completion by the propensity of teenage pregnancy. This is similar to the stylized pattern in Fig. 1. First, the probability of high school graduation declines as the probability of teenage pregnancy increases among both groups, indicating that antecedent factors predicting teen pregnancy also depress the likelihood of graduating high school. Second, young women who experience a teen pregnancy have a lower probability of completing high school than nonpregnant teens across most of the probability distribution, suggesting that teen pregnancy has fairly pervasive negative effects on high school completion. Finally, the treatment effect appears most negative for respondents with a low to moderate propensity for teen pregnancy. In other words, teens who are most likely to experience a pregnancy experience smaller consequences in terms of high school completion.
The IPW regression analysis summarized in Table 3 suggests substantial negative effects of teen pregnancy but does not indicate systematic heterogeneity. Teen pregnancy reduces the probability of high school graduation about 14 percentage points for the average young woman, 15.5 points for the typical pregnant teen, and 14.7 points for the typical nonpregnant teen.
The difference between the S-D and IPW regression analysis likely results from the covariate imbalances in the reweighted samples for ATE and TOC. These imbalances suggest that the apparent effect in Fig. 2 is biased at lower probabilities of treatment, where the imbalances are concentrated. The regression analysis adjusts for remaining confounding and indicates pervasive negative effects on high school completion around 14–15 percentage points.
S-D plots for college attendance are shown in Fig. 3. The probability of college attendance for both treatment and control groups declines as the propensity of teenage pregnancy increases, but there is neither a clear treatment effect nor a pattern of effect heterogeneity. This revelation is verified in the estimates summarized in Table 3: there are no significant teen pregnancy effects. Hence, the factors that predict teen pregnancy rather than teen pregnancy itself deter college attendance. This finding is not shocking given that college attendance refers to attendance of any duration at any type of college, many of which do not require a high school diploma.
When it comes to college completion, we find large effects and substantial heterogeneity. The S-D plot for bachelor’s degree attainment is presented in Fig. 4. Again, the probability of earning a bachelor’s degree declines as the propensity for teenage pregnancy increases for both treatment and control cases, indicating that similar factors increase teen pregnancy and reduce college completion. This is especially true for young women who do not become pregnant as teenagers. Teen pregnancy also has a negative effect on bachelor’s attainment, although the effect is concentrated among those least likely to be pregnant and disappears as the probability of teen pregnancy approaches 0.5. The plot even indicates positive effect estimates for those most likely to become pregnant as teenagers, although these are not precisely estimated.
This effect heterogeneity is also supported by the estimates in Table 3. The analysis yields significant negative ATE (–.145), TOT (–.110), and TOC (–.198) effects, and the difference in TOT and TOC effects (.088) provides evidence of positive selection. Teenage pregnancy reduces the probability of graduating college almost 15 percentage points for the average female, 11 points for the typical pregnant teen, and 20 points for the typical nonpregnant teen. Thus, teen pregnancy appears most detrimental to those with otherwise fair prospects of completing college, but has less impact on those with characteristics that make college graduation unlikely.
Our final outcome of interest is log-average earnings when women are aged 25–35. The S-D plot again suggests stronger negative effects among those least likely to have a teenage pregnancy, with possibly positive effects for those with very high probabilities of early fertility (Fig. 5). The IPW analysis yields a similar pattern of positive selection. Exponentiating the coefficient produces the ratio of earnings among pregnant teens to nonpregnant teens. The ATE indicates a marginally significant 55 % reduction in earnings on average, whereas the TOT indicates a 23 % reduction among typical pregnant teens, and the TOC indicates a statistically significant 67 % reduction among typical nonpregnant teens. These differences continue to suggest that women who are most likely to become pregnant as teenagers fare better (less poorly) after a pregnancy than women who are least likely to become pregnant. This echoes the results for bachelor’s degree attainment and may very well be a consequence of it.
What Could Be Driving These Heterogeneous Effects?
Our findings that teenage pregnancy has homogenous negative effects on high school graduation are consistent with pervasive opportunity costs entailed by an early pregnancy, which precludes young women from earning a high school diploma. Our null findings for college attendance suggest that pregnant adolescents may instead pursue alternative routes to college, such as earning a high school equivalency certification—which does not yield similar socioeconomic returns as a diploma (e.g., Cameron and Heckman 1993)—or that women for whom early pregnancy deterred high school completion were unlikely to attend college regardless of their fertility. Conversely, the negative effects on college completion and early earnings are roughly twice as strong for women unlikely to become pregnant teens compared with those likely to become pregnant as teens. This pattern of positive selection is consistent with theoretical explanations of greater opportunity costs for women with greater skills and resources, and also supports the notion of stressful role transitions for those least likely to become young mothers. To better assess these explanations, we test interaction effects in regression models.
Table 4 summarizes interaction effects between teen pregnancy and covariates from OLS models weighted to reflect the average female in the sample. These analyses have limited statistical power because they include main effects and interactions for all covariates used to predict teen pregnancy, and the analytic samples are small relative to the large number of parameters. Because we find little evidence of systematic effect heterogeneity for high school completion or college attendance and few significant interactions for college completion (the smallest sample), we focus primarily on early earnings. Coefficients represent the change in the teen pregnancy effect given a one-unit change in each covariate.
With respect to earnings, teen pregnancy appears to have weaker negative effects on young women with a greater sense of self-worth, who have greater intellectual skills, and who are more risk-averse. These attributes may be protective against the challenges presented by early fertility; results do not support the notion that teen pregnancy is less consequential for young women lacking such skills. Adolescents with higher levels of behavior problems and delinquency also suffer weaker consequences of early fertility. This accords with the selection argument: these women’s behaviors likely limited their socioeconomic prospects regardless of a teen pregnancy. We find similar patterns for bachelor’s degree attainment.
The influence of individual- and household-level fertility experiences also emerges. Early fertility has stronger consequences on earnings for teens who were born to older mothers, and also has stronger consequences on earnings and bachelor’s degree attainment for women whose mothers were older at their first birth. Perhaps adolescents face less stigma (or a less stressful transition to motherhood) in contexts where early fertility is normative—or less unexpected—than adolescents who live in contexts where teen fertility is rare. It is unclear why the consequences on earnings would be less severe for those whose families have not talked to them about sex.
Teen pregnancy has weaker negative effects on the earnings of young women who share important life events with their parents. These women may have stronger, more positive relationships with parents, who may be more willing to provide social support. This interpretation is bolstered by the positive interaction with emotional support, although it is not significant. We see the opposite pattern for cognitive stimulation, which suggests that earnings consequences may be more severe in homes that have emphasized intellectual and educational investments in daughters. These patterns are not consistent, however, when bachelor’s degree attainment is the outcome.
We find no significant interactions for contextual variables or sociodemographic background characteristics when assessing earnings. One notable exception for bachelor’s degree attainment is maternal education, for which the negative consequences of teen pregnancy are strongest among offspring with more-educated mothers. Given the link between maternal and child education, this finding is consistent with the argument that early fertility has its strongest effects on young women with higher educational prospects.
Because certain skills mitigate the deleterious effects of early fertility whereas others exacerbate them, this analysis provides mixed evidence for opportunity costs as an explanation for the documented effect heterogeneity. Women with stronger and more supportive relationships with their parents seem to fare better, pointing to the importance of parental influence and home environment in coping with an early pregnancy. Young women in families with attitudes and experiences more conducive to early fertility also fare better, providing support for role transition theory. For the most part, however, these interaction analyses do not provide overwhelming support for one particular theory. We attribute this to our imperfect proxies for the constructs cited by such theories and the limited statistical power in our interaction analyses. Nevertheless, the substantial and variable effects of teenage pregnancy on college completion and early economic success present important issues for scholars and policy makers to consider.
Despite these striking findings, it is important to stress that existing work primarily documents the consequences of teen childbearing—not pregnancy. If the effects of teen pregnancy operate primarily through childbearing, childbearing effects should be greater than pregnancy effects. Our analyses of teen childbearing, presented in the supplement (see Tables S4–S7, Online Resource 1), suggest that this is the case for all outcomes. Earlier, we defended pregnancy as a treatment of interest, partly because of the additional challenges of selection from pregnancy to a live birth. We speculated that women who are least likely to have a teen pregnancy would also be less likely to carry a pregnancy to term, largely due the nonrandomness of abortion. Although we lack reliable data on miscarriages and abortions, Fig. S1 (Online Resource 1) indirectly shows that this pattern is apparent in our data: the probability of childbearing by age 18 increases with the predicted probability of pregnancy among pregnant teens.
Given that rates of childbearing are lower for those least likely to become pregnant, differences in the effects of childbearing and pregnancy should be most pronounced among this subpopulation. If this is true, our results for pregnancy will understate effect heterogeneity (i.e., positive selection) for childbearing. Our results substantiate this argument: the TOT-TOC difference for childbearing effects is more positive (less negative) than the difference for pregnancy effects for all outcomes (Table S6, Online Resource 1).12 Thus, our analysis of teenage pregnancy is a conservative approximation of effect heterogeneity for childbearing. Nonetheless, our focus on teenage pregnancy is most relevant to public policy and practical efforts targeted to improving young women’s well-being.
Discussion and Conclusion
Teenage pregnancy remains a contentious issue, depicted as everything from a pathological problem contributing to intergenerational disadvantage to a trivial event that diverts attention from greater barriers to attainment and mobility (Furstenberg 2003; Geronimus 1997). Various theories lend support to these opposing perspectives, and scholars have used a variety of methods to assess them. But the fixation on removing selection bias has led to uncertainty about which method is best suited to identify the true effect. Concern about bias is justified, but ignoring effect heterogeneity is not. When we acknowledge that the effects of teen pregnancy vary among women, and that different methods emphasize particular groups of women, the range of findings is predictable and aligns with seemingly opposing theories. As such, we develop a new framework for conceptualizing the effects of teen pregnancy and childbearing. Specifically, we argue that negative, trivial, or positive effects—each of which are documented in extant literature—could occur among different types of women in the population. Our framework moves away from isolating the “true” causal effect in favor of investigating variation in the consequences of teen fertility. This approach provides a more nuanced perspective on women’s life chances after a pregnancy or live birth.
Using rich longitudinal data, we use methods that address selection bias and explicitly examine effect heterogeneity. Although we find pervasive negative effects of teenage pregnancy on high school graduation, there is little evidence of systematic heterogeneity, and there are no effects on college attendance. We speculate that the costs associated with childrearing may encourage women to pursue other routes to college attendance (e.g., high school equivalence degrees), or that the women on the brink of high school graduation are unlikely to attend college regardless of early fertility. There is striking heterogeneity, however, for other outcomes. Teen pregnancy has strong negative effects on college completion and early earnings among women unlikely to become pregnant teens, and significantly weaker effects on those most likely to have a teenage pregnancy. These patterns are even more pronounced in our analyses of teenage childbearing.
In some ways, our findings align with theories that speak to the opportunity costs of childbearing, while in others, they align with theories pointing to social support or the stress of accelerated role transitions. For instance, teenage pregnancy is most detrimental among adolescents with more educated parents and low levels of delinquency and behavior problems. By the time they become pregnant, low-SES delinquent teens may already face bleak prospects and thus low opportunity costs. In other words, opportunity costs are concentrated among those with the most opportunity. The patterns expected under this explanation, however, were absent for other skills and attributes, which mitigated the consequences of early fertility. Women may also be buffered from the effects of teen pregnancy by parental influence and home environment: consequences are less severe among adolescents who receive more emotional support and share important life events with their parents. The more severe consequences for teens born to mothers who gave birth at older ages suggest the possibility of greater stigma or stress in contexts where early fertility is not normative.
Although the evidence is weak, the purported positive effects of early fertility may exist for some women. The graphical analysis suggests positive effects on college completion and early earnings for those who are especially likely to experience early fertility. However, these estimates are not very precise, and reweighting error or unmeasured confounders for these extreme cases may lead to bias.
Beyond the contributions to this scholarly debate, our results have important policy implications. By assessing effect heterogeneity, we can better identify who is hurt most by teenage pregnancy and what can be done to assist them. Based on our findings, efforts to prevent early pregnancies will have limited impacts on the disadvantaged teens most likely to experience early fertility. Ironically, such efforts would have stronger effects in comparably advantaged households and communities where early fertility is rare. Among disadvantaged women for whom early fertility is common, policies may be more effective if they promote the early skills and resources associated with both pregnancy and later attainment. Given the effect heterogeneity in this study, such efforts would also increase the effectiveness of teen pregnancy policies. Our findings imply that the deleterious consequences of teenage pregnancy may be buffered by certain skills and support from parents, a possibility worthy of further research. Ultimately, our analyses suggest that teen pregnancy prevention that is isolated from other factors that affect attainment is likely to be ineffective for disadvantaged women.
This study is not without limitations. First, the women of NLSY have only recently reached the end of their fertility, and many of their children are too young to have completed their schooling or entered the labor market. Thus, our sample likely draws on daughters born to younger mothers, on average, than typical in the population. Our attrition adjustments address this problem, but its impact in our study cannot be known until the full sample of children matures. It will also be useful to track this sample’s long-term attainment and other life outcomes as they age. Second,, the NLSCYA is not suited for methodological strategies requiring larger samples of pregnant young women. For instance, these data lack the power to effectively replicate existing work using FE or IV estimation. Without a greater number of sibling observations (or miscarriage reports), we are reluctant to draw on these methods for our own study. Nevertheless, this does not hinder the conclusions or contributions presented here.
The question also remains whether we adequately address selection into teenage pregnancy. Although we include a substantial number of background, personal, and contextual indicators in our propensity model, unobserved factors could introduce bias in our results; this is a weakness shared with other propensity-based studies (e.g., Chevalier and Viitanen 2003; Lee 2010; Sanders et al. 2007). Although no study can claim to remove all selection bias, our study makes a significant contribution even in the absence of our empirical findings. We demonstrate that effect heterogeneity is just as important as selection bias in understanding the effects of teenage pregnancy and may be vital in resolving longstanding theoretical and empirical debates.
The authors would like to acknowledge the Ford Foundation (Diaz, Pre-doctoral Fellowship #CHK-7020411); the Institute of Education Sciences, U.S. Department of Education (Fiel, Award #R305B090009); and the National Science Foundation Graduate Research Fellowship (Fiel, Award #DGE-0718123). This research was partially supported by the Center of Demography and Ecology at the University of Wisconsin–Madison. We thank participants of the Inequality & Methods Workshop at the Minnesota Population Center for providing valuable feedback, as well as Jenna Nobles for comments on earlier drafts. The authors are solely responsible for all content.
See Klepinger and colleagues (1995), Ribar (1994), and Walker and Zhu (2009) for other instruments.
The predicted probability for teen pregnancy in the population in our data is slightly higher than reported national rates. One reason is that our sample is not nationally representative of young women: it is representative of the children born to the 1979 cohort of women. In other words, our respondents represent offspring born to younger women (and who may be more disadvantaged) than the average female.
In all analyses presented here, we employ an Epanechnikov kernel with asymptotically optimal constant bandwidths, the default in Stata using the lpoly function.
All but 1 of the 1,969 cases excluded from our analytic sample were too young (not yet 18) to have reported teen pregnancy data as of 2010. Depending on the outcome, 10 % to 20 % of cases were dropped because of missing outcome data, with the remainder being too young to have met the age cutoff. See our discussion of attrition and missing data.
When items are categorical, factor analyses are performed on polychoric correlation matrices, and those tapping multiple factors use promax rotation to allow constructs to be correlated. Further details are available on request.
We tested alternative specifications that included mothers’ desired number of children and expected number of children when surveyed in 1979. Because these measures do not predict whether daughters go on to experience a teen pregnancy, we dropped these indicators from our prediction model to improve efficiency.
We create 25 imputed data sets separately by whether the respondent reported being pregnant before 18.
Attrition weights are multiplied by the IPW weights in reweighting analyses.
We tested all two-way interactions and quadratic terms and retained those that approached statistical significance. We then added a squared term for the predicted logit as a covariate to capture residual nonlinearities. (See Table S1, Online Resource 1.)
Table 2 contains balance for our high school graduation sample (among those old enough to complete high school). Results for our additional outcomes and treatments are similar and are provided in Tables S2–S3 in Online Resource 1.
Measures characterized with standardized differences exceeding .10 include self-worth, intellectual skills, risk aversion, delinquency, educational expectations, negative peer pressure, private school attendance, mother’s age at birth, mother’s expected age at marriage, family income, and maternal education.
See Tables S8–S10 in Online Resource 1 for additional sensitivity tests for both teen pregnancy and teen childbearing. Results are substantively similar across different treatment cutoff points.