Premarital first births: The influence of the timing of sexual onset versus post-onset risks in the United States

Abstract

Motivated by long-standing debates between abstinence proponents and skeptics, we examine how socioeconomic factors influence premarital first births via: (1) age at first sexual intercourse and (2) the risk of a premarital first birth following onset. Factors associated with an earlier age at first intercourse will imply more premarital first births due to increased exposure to risk, but many of these same factors will also be associated with higher risks of a premarital first birth following onset. Our analyses confirm previous findings that women from disadvantaged backgrounds are younger at first intercourse and have higher premarital first birth risks relative to those from more advantaged backgrounds. However, differences in onset timing have a strikingly smaller influence on premarital first birth probabilities than do differences in post-onset risks. Our findings thus suggest that premarital first births result primarily from differences in post-onset risk behaviors as opposed to differences in onset timing.

Special thanks are due to Larry Bumpass and Andrew Cherlin for earlier work that inspired the questions leading to this paper. We thank: the editors, anonymous reviewers, Paula England, Herbert Smith, and Megan Sweeney for their comments; and Yeh-chen Chen, Hassan El Menyawi and Robert Wihr Taylor for their research assistance. Earlier versions of this paper were presented at the Institute for Research on Poverty, University of Wisconsin-Madison; the Initiative in Population Research, Ohio State University; the Population Research Center, University of Maryland; the Center for Advanced Social Science Research, New York University; the Population Studies Center, University of Pennsylvania; the 2010 Annual Meetings of the American Sociological Association; the Minnesota Population Center; the 2012 Annual Meetings of the Population Association of America; and the Center for Demography and Ecology, University of Wisconsin-Madison. Research funding from NICHD (R01 HD 29550) is gratefully acknowledged. Direct all correspondence to Lawrence L. Wu, Department of Sociology, 295 Lafayette Street, Puck Building, 4th floor, New York University, New York, NY 10012-9605, lawrence.wu@nyu.edu.

Nonmarital fertility comprises a substantial proportion of all U.S. fertility, with over 40% of all U.S. births now occurring outside of formal marriage (Hamilton et al. 2013). Numerous past studies in both Europe and the U.S. have documented that women who have a first birth outside of formal marriage are drawn disproportionately from disadvantaged backgrounds. In this study, we provide greater insight into this issue by examining how disadvantage influences: (1) a woman’s age at onset of sexual activity and (2) her risk of a premarital first birth in the period following onset.

Demographers and policy-makers have long been concerned with the link between sexual activity and nonmarital fertility. The decoupling of sex from marriage and the resulting near-universal levels of premarital sexual activity in many countries have been argued to be core elements of the second demographic transition (see, e.g., van de Kaa 1987; Lesthaeghe and Surkyn 1988; Lesthaeghe 2010). The U.S. Personal Responsibility and Work Opportunity Reconciliation Act not only transformed welfare provisions for single mothers and their children, but also explicitly cast the link between sexual activity and nonmarital fertility as an acute policy concern, asserting that “abstinence from sexual activity is the only certain way to avoid out-of-wedlock pregnancy” and that “having children out-of-wedlock is likely to have harmful consequences for the child, the child’s parents, and society.”

The policy debate concerning abstinence has continued but remains deadlocked, as evidenced by two recent federal bills—the Abstinence Education Reallocation Act of 2013 (H.R. 718), which advocated “teaching the skills and benefits of sexual abstinence as the optimal sexual health behavior for youth” and the Real Education for Health Youth Act of 2013 (S. 372), which would have prohibited funding of programs that were “insensitive and unresponsive to the needs of sexually active youth.” What is obscured in these debates is the fact that the arguments of abstinence proponents and skeptics are neither logically inconsistent nor empirically incompatible. The argument by abstinence proponents mirrors a core demographic insight—that when onset is delayed, exposure to risk will decrease, thus resulting in fewer premarital first births. But post-onset factors such as poorer contraceptive use or knowledge will also lead to more premarital first births, leading skeptics to argue that reductions in teen and premarital births would be best achieved by focusing on factors that reduce post-onset risks.

We contribute to this debate as well as to the larger literature on nonmarital fertility in several ways. First, we employ an innovative sequential hazard framework that explicitly links two key processes—an onset process by which some women become sexually active while never-married, and a post-onset process during which never-married women are at risk of a premarital first birth. Second, our analyses provide greater insight into the trajectories that do (or do not) lead to a premarital first birth by documenting how onset timing and post-onset risks differ for women from advantaged and disadvantaged backgrounds. Third, we decompose the probability of a premarital first birth into components reflecting differences in exposure to risk, as generated by earlier or later sexual onset, and differences in risks following onset.

We analyze data from the 1979 National Longitudinal Survey of Youth (NLSY79), a nationally representative sample of U.S. women aged 14–21 in 1979. Our results confirm previous findings that women from disadvantaged backgrounds become sexually active at younger ages and have higher premarital first birth risks relative to those from more advantaged backgrounds. Nevertheless, differences in exposure to risk, as generated by earlier or later onset, have a negligible influence on the probability of a premarital first birth; by contrast, differences in risks following onset are large in magnitude. Thus, one contribution of this study is to move beyond a sole focus on which social disadvantages matter by also asking when it is that disadvantage exerts its influence.

More generally, we take seriously the view that greater insight into various demographic processes might be obtained by analyzing their proximate determinants (Davis and Blake 1956; Bongaarts 1978). Previous studies in this vein have analyzed aggregate-level outcomes showing, for example, how proximate factors affect aggregate-level measures of fertility in developed (Smith and Cutright 1986; Smith et al. 1996) and developing countries (Bongaarts and Potter 1983). Thus, yet another contribution of this study is to show how questions involving proximate determinants can be addressed using individual-level data within a continuous-time hazard framework.

THEORY

We begin by discussing factors found to influence onset risks, many of which will also influence premarital first birth risks. We next review how onset timing might itself influence women’s premarital first birth risks in the period following onset. We then contrast arguments by abstinence proponents and skeptics on how teen and nonmarital births might best be reduced. Throughout, we maintain the sharp but standard demographic distinction between the terms risk and probability as in the statement “all else being equal, an earlier age at onset will increase exposure to risk, thereby increasing the probability of a premarital first birth.”

Factors influencing onset and premarital first birth risks

Socioeconomic disadvantage, such as nonintact family structure or membership in a disadvantaged racial and ethnic minority group, has been shown to be strongly associated both with sexual onset and premarital first births. Numerous studies have documented a strong association between disadvantage and earlier sexual onset (see, e.g., Dorius et al. 1993; Stanton et al. 1993; Upchurch et al. 1998; Paul et al. 2000; Wu and Thomson 2001; Cavanaugh 2004; Longmore et al. 2004; Duper et al. 2008; Cavasos-Rehg et al. 2010; Madkour et al. 2010). Similarly, previous studies have documented strong associations between disadvantage and the higher risk of a teen or premarital first birth (see, e.g., An, Haveman, and Wolfe 1993; Wu and Martinson 1993; Wu 1996; Powers and Hsueh 1997; Michael and Joyner 2001; Fomby et al. 2010; Hofferth and Goldscheider 2010; England et al. 2011).

Why, theoretically, might disadvantage be linked to sexual onset and premarital first births? One set of theories sees these links as arising from limited marital prospects reflecting various structural conditions (see, e.g., Wilson 1987; Anderson 1991; Geronimus 1991; Willis 1999; Edin and Kefalas 2005) or as choice behaviors reflecting lower opportunity costs (see, e.g., Becker 1981; Akerlof et al. 1996; Michael and Joyner 2001; Hotz 2008). Other theories see adolescent preferences and behavior as influenced by religious institutions (see, e.g., Burdette and Hill 2009; Regnerus and Uecker 2011; Murray 2012), by offspring modeling of parental attitudes and sexual behaviors (see, e.g., Newcomer and Udry 1984; Thornton and Camburn 1987; Barber 2001; East et al. 2007), by parental supervision and monitoring of offspring sexual risk-taking (see, e.g., Dornbusch et al. 1985; Pearson et al. 2006; Brauner-Otto and Axinn 2010), or by the greater stress and instability experienced by disadvantaged youth (see, e.g., Wu and Martinson 1993; Capaldi et al. 1996; Wu 1996; Wu and Thomson 2001; Fomby et al. 2010).

Studies of onset have often asserted that differences in onset timing will influence nonmarital fertility; similarly, studies of nonmarital fertility have invariably noted the importance of sexual behavior. Yet only a handful of studies have provided any empirical evidence of possible linkages between these factors. Michael and Joyner (2001) presented a conceptual model that, like ours, assumes that women will not be at risk of a birth until they become sexually active. However, their empirical results, based on separate logistic regressions for onset before age 18 and a first birth before age 18, did not account for differential exposure to risk from variations in onset timing. Finer and Philbin (2013) presented descriptive analyses of individual-level onset data and aggregate data on teen pregnancies. Fomby et al. (2010) estimated separate Cox hazard regressions to examine the effects of family instability, race/ethnicity, and other background variables on age at first sexual intercourse and age at a nonmarital first birth. However, their model of nonmarital first births also did not account for differential exposure to risk. England et al. (2011) conducted separate logistic regressions for several outcomes including sexual activity within the past 12 months, contraceptive use, pregnancies, and teen births. Their analysis revealed strong effects of parental education across these outcomes, but their results also provided at best indirect evidence concerning any link between sexual onset and a premarital first birth. Hofferth and Goldscheider (2010) specified a dummy variable for onset before age 14 in modeling early parenthood but did not otherwise account for differential exposure to risk due to onset timing. The study closest in spirit to ours is Miller and Heaton (1991), whose model assumed, like ours, that a woman is not at risk of a premarital birth prior to onset. However, they did not model onset risks and hence their analyses did not consider the possibility that covariates might affect both onset timing and premarital first birth risks.

Onset timing as a factor influencing premarital first birth risks

We now turn to theoretical arguments and behavioral mechanisms by which onset timing, as a distinct factor, might influence premarital first birth risks. A first argument is that later onset will entail, on average, greater emotional and behavioral maturity, which in turn will lower premarital first birth risks. Adolescence, in particular, is a period of unusually rapid physical, cognitive, and emotional change, with the development of a positive sexual and emotional sense of self, it is argued, posing unique developmental challenges (Freud 1961 [1925]; Adams et al. 1996; Horne and Zimmer-Gembeck 2005; Diamond 2006; Halpern 2010; Regnerus and Uecker 2011). This perspective thus views “typical” ages at onset as a normal but also critical component in the healthy socioemotional development of adolescents and young adults. Nevertheless, later onset will, ceteris paribus, pose fewer negative risks and more potential benefits as individuals gain in emotional maturity with age (Houlihan et al. 2008; Kim and Rector 2008; Sabia and Rees 2008).

A second set of arguments sees “off-time” ages at onset—particularly onset at very early or very late ages—as posing serious risks. Women with especially early ages at onset are more likely to have experienced sexual abuse, incest, and non-consensual intercourse, but even in their absence, a very early onset is argued to pose emotional and developmental traumas that may influence subsequent behaviors (Musick 1993; Newcomb 1996; Joyner and Laumann 2001; Wellings et al. 2001; Staff et al. 2004; Kaestle et al. 2005; Sandfort et al. 2008). Likewise, an “especially late” age at onset could carry negative consequences if those experiencing late onset are seen by peers as less socially or physically desirable (Donnelly et al. 2001; Sandfort et al. 2008).

Pre- and post-onset expectations of abstinence proponents and skeptics

In many respects, the question of whether and how onset timing and premarital first births might be linked lies at the core of ongoing debates between abstinence proponents and skeptics. Abstinence skeptics have long argued that reductions in teen and nonmarital births are best achieved by providing comprehensive and timely knowledge about human sexuality and by encouraging effective contraception among those who are sexually active (Furstenberg et al. 1985; Marsiglio and Mott 1986; Mauldon and Luker 1996; Darroch et al. 2000; Lieberman et al. 2000; O’Donnell et al. 2001; Lindberg et al. 2006; Furstenberg 2007). By contrast, abstinence proponents argue that “[d]elaying the initiation of or reducing early sexual activity among teens can decrease their overall exposure to risks of unwed childbearing” and that “abstinence [is] therefore crucial to efforts at reducing unwed childbearing” (Kim and Rector 2006). These expectations are mirrored in the 1996 Personal Responsibility and Work Opportunity Reconciliation Act (PRWORA). In addition to transforming U.S. welfare provisions to single mothers and their children, PRWORA held that the “prevention of out-of-wedlock pregnancy and reduction in out-of-wedlock birth are very important Government interests.” Among specific PRWORA provisions intended to address the “crisis” of nonmarital childbearing were directives to states to “establish goals and take action to prevent and reduce the incidence of out-of-wedlock pregnancies” and to establish numerical goals for reducing the illegitimacy ratio (Section 402), with funds specifically allocated to encourage states to establish “abstinence education … with a focus on those groups likely to bear children out of wedlock.”

A study evaluating sites that randomly assigned students to courses that followed abstinence-only guidelines in PRWORA found no difference between treatment and controls four to six years after treatment for a wide array of outcomes, including whether the student had initiated sexual activity, age at onset, recent sexual activity, STDs, pregnancy, and whether the student had given birth or fathered a child (Trenholm et al. 2007). Although these findings have renewed skepticism among critics, proponents have countered that such conclusions are premature in that interventions were short in duration and administered in early grades, thus targeting students at ages before most become sexually active. These findings have thus had the effect of shifting the terms of debate from short-term interventions to what might persistently influence the fertility-related behaviors of adolescents and young adults, with abstinence proponents holding that delayed onset will be most strongly affected by more persistent influences and skeptics seeing such persistent influences as most likely to affect risk behaviors following onset.

Our analyses focus on two broad sets of questions about factors thought to be persistent influences on adolescent and young adult outcomes. First, by how much do such factors hasten or delay onset, and how many more or fewer premarital first births are implied by the resulting differences in exposure to risk? Second, by how much might such factors increase or decrease premarital first birth risks following onset, and how many more or fewer premarital first births are implied by these influences?

MODEL

We modeled premarital first births in terms of two sequential events, the transition to sexual activity and the transition to a premarital first birth. Let T₁ and T₂ denote a never-married woman’s age at first sexual intercourse and a premarital first birth, respectively. Most studies to date have ignored T₁ when modeling T₂, using, for example, a conventional proportional hazard specification

$${\mathit{r}}_{2\mathit{ci}}(t|{\mathbf{x}}_i)=q_{2c}(t)\exp ({\mathbf{b}}_{2c}{\mathbf{x}}_i),$$ (1)

where q_2c(t)denotes the age-specific baseline for T₂, x_i a set of observed covariates for woman i, b_2c the corresponding set of estimated coefficients, i indexes women, and the subscript “c” signals the use of a “conventional” specification. Note that (1) implicitly assumes that a woman is at risk of a birth both before as well as after she initiates sexually activity.

Although one might be tempted to specify onset timing t_1i as a right-hand-side covariate in (1)

$$r_{2\mathit{ci}}^{\ast }(t|{\mathbf{x}}_i)=q_{2c}^{\ast }(t)\exp (\alpha t_{1i}+{\mathbf{b}}_{2c}^{\ast }{\mathbf{x}}_i)$$ (2)

the difficulty is that (2) assumes knowledge of the future—that is, that onset timing is known at all ages, including t <t_1i (Kiernan and Hobcraft 1997). That is, consider a woman who initiates sexual activity at 17; then note that (2) models her birth risks at age 14 with knowledge that onset will occur at 17, three years in the future. One might also seek to modify (1) by including an age-varying dummy variable d_i(t) contrasting women who are and are not sexually active at age t However this approach cannot imply zero birth risks prior to onset unless the estimated contrast equals | ∥ | for d_i(t) = 1 prior to onset and 0 otherwise.

To avoid these problems, we instead modeled two transitions—the transition 0 →T₁ for the onset of premarital sexual activity, and the transition T₁→ T₂ for a woman’s age-specific risk of a premarital first birth among never-married women who have initiated sexual activity. This model thus explicitly assumes that no woman is at risk of a premarital first birth until she becomes sexually active.

We modeled the 0→ T₁ transition using a conventional proportional hazard specification

$$r_{1i}(t|{\mathbf{x}}_i)=q_1(t)\exp ({\mathbf{b}}_1{\mathbf{x}}_i),$$ (3)

where q₁(t) denotes the T₁ baseline for how onset risks vary with age. We then specified the T₁ → T₂ transition conditional on t_1i, woman i’s observed age at onset as

$$r_{2i}(t|{\mathbf{x}}_i,t_{1i})=q_{21}(t|t_{1i})q_{22}(u)\exp (\alpha t_{1i}+{\mathbf{b}}_2{\mathbf{x}}_i),$$ (4)

where u = t−t₁ denotes duration since onset and q₂₁(t|t₁) and q₂₂(u)denote the baseline functions modeling how T₂ risks vary with age and duration, respectively. Our sequential model thus differs from (1) and (2) by assuming that no woman is at risk of a birth prior to onset. It also assumes that T₁ and T₂ are not jointly determined and so can be modeled sequentially, with this assumption violated, for example, if women sought to conceive and to take the pregnancy to term when having sex for the first time.

As noted above, differences in onset timing will, ceteris paribus, imply differences in the numbers, proportions, and probability of a premarital first birth via its influence on exposure to risk. These intuitions can be both formalized and quantified in that (3) and (4) carry implications for both onset timing and for the probability of having a premarital first birth conditional on onset timing.

Under (3), the probability that woman i initiates sexual activity by age t is given by

$$\begin{array}{c}\mathrm{Pr}(T_{1i}\le t|{\mathbf{x}}_i)=1-\mathrm{Pr}(T_{1i}>t|{\mathbf{x}}_i)\\ =1-S_1(t|{\mathbf{x}}_i)\\ =1-\exp [-{\int }_0^t{r}_1(s|{\mathbf{x}}_i)\mathit{ds}]\end{array}.$$ (5)

where S₁ denotes the so-called survivor function. Similarly the probability that woman i has a premarital first birth by age t conditional on onset at t_1i is given by

$$\begin{array}{c}\mathrm{Pr}(T_{2i}\le t|{\mathbf{x}}_i,T_1=t_{1i})=1-\mathrm{Pr}(T_{2i}>t|{\mathbf{x}}_i,T_1=t_{1i})\\ =1-S_2(t|\mathbf{x},t_{1i})\\ =1-\exp [-\exp (\alpha t_{1i}+{\mathbf{b}}_2\mathbf{x}){\int }_{t_{1i}}^t{q}_{21}(\nu )\mathrm{d}\nu {\int }_0^u{q}_{22}(w)\mathrm{d}w]\end{array}.$$ (6)

with the timing of entry into T₂ risk reflected in the lower limit of the first integral in (6).

Because a covariate x may influence both the 0 → T₁ and T₁ → T₂ transitions, there will be both direct and indirect effects of x on the probability in (6) in ways that roughly mirror direct and indirect effects in linear regression. The issues are more complicated in a hazard setting, so we consider them in turn. To fix ideas, consider comparing women from intact and nonintact families. A first issue is that, all else being equal, earlier onset will necessarily imply more premarital first births via increased exposure to risk. As noted above, a robust finding in past research is that, all else equal, women from nonintact families have an earlier age at onset than those from intact families. Let the median age at onset, as predicted by our model, be denoted by t₁ and t₁ + Δt₁ for women from nonintact and intact families, respectively, with Δt₁ thus giving the difference in predicted median ages and the (median) difference in exposure to risk for women from nonintact and intact families. (We focus on the median of T₁ because of censoring.) Then for a covariate x such as the dummy variable contrasting women from nonintact and intact families, we refer to this exposure consequence of x as the “indirect effect of exposure” for the probability in (6), with indirect referring to the fact that x influences the probability of a premarital first birth indirectly via x’s influence on exposure to risk.

A second issue is that covariates that hasten or delay onset will also typically raise or lower premarital first birth risks following onset. As our analyses below show, women from nonintact families have, holding other factors constant, both earlier onset and higher risks following onset, relative to women from intact families. Elevated risks in turn imply a higher probability of a premarital first birth and we refer to this as the “direct effect of x ” on the probability in (6).

A final issue is that t₁ appears as an ordinary right-hand-side covariate in the T₁ → T₂ equation in (4), as is appropriate for hypotheses positing a causal effect of onset timing on premarital first birth risks following onset, as discussed above or, alternatively, if T₁ were to be correlated with unobservables that in turn affect T₂ risks following onset. Including women’s age at onset as a right-hand-side covariate in (4) also parallels the logic of models of status attainment, in which family background affects both an individual’s years of schooling completed and occupational attainment, and in which years of schooling completed also affects occupational attainment. Our example supposes that the median ages at onset are t₁ and t₁ + Δt₁ for women from nonintact and intact families, respectively; hence, the difference Δt₁ will also affect the probability in (6) because we have included t₁ as a right-hand-side covariate in (4). We refer to this as the “indirect covariate effect of x ” on the probability in (6).

We used a das-Gupta decomposition to quantify how these three components affect the probability of a premarital first birth in (6). Let Pr(T₂ ≤ t) = G(t) = g(γ₁,γ₂,γ₃) be a function of three terms—a direct effect γ₁, an indirect effect of exposure γ₂, and an indirect covariate effect γ₃. Then consider the difference in the probability in (6) for two groups, A and B

$$G_A(t;t_{1A},{\mathbf{x}}_A)-G_B(t;t_{1B},{\mathbf{x}}_B)=\mathrm{Pr}(T_2\le t|t_{1A},{\mathbf{x}}_A)-\mathrm{Pr}(T_2\le t|t_{1B},{\mathbf{x}}_B).$$ (7)

To decompose (7), we used a das-Gupta (das Gupta 1983) technique that evaluates switching from group A to B across all possible permutations of γ₁, γ₂, and γ₃. For example, one set of permutations will be

$$\begin{array}{l}g({\gamma }_{1B},{\gamma }_{2A},{\gamma }_{3A})=1-\exp [-\exp (\alpha t_{1A}+\mathbf{b}{\mathbf{x}}_B){\int }_{t_{1A}}^t{q}_{21}(s)\mathit{ds}{\int }_0^{t-t_{1A}}{q}_{22}(u)\mathit{du}]\\ g({\gamma }_{1A},{\gamma }_{2B},{\gamma }_{3A})=1-\exp [-\exp (\alpha t_{1A}+\mathbf{b}{\mathbf{x}}_A){\int }_{t_{1B}}^t{q}_{21}(s)\mathit{ds}{\int }_0^{t-t_{1B}}{q}_{22}(u)\mathit{du}]\\ g({\gamma }_{1A},{\gamma }_{2A},{\gamma }_{3B})=1-\exp [-\exp (\alpha t_{1B}+\mathbf{b}{\mathbf{x}}_A){\int }_{t_{1A}}^t{q}_{21}(s)\mathit{ds}{\int }_0^{t-t_{1A}}{q}_{22}(u)\mathit{du}]\end{array}$$ (8)

proceeding similarly for the remaining possible permutations g(γ_1A, γ_2A, γ_3A), …, g(γ_1B, γ_2B, γ_3B). For additional details, see Wu and Martin (2009).

DATA

We analyzed data from the 1979–93 waves of the National Longitudinal Survey of Youth (NLSY79), a household-based national probability sample of persons aged 14–22 at the 1979 and 28–36 at 1993 waves;, with virtually all premarital first births having occurred within this age range.

Data on age at first sexual intercourse were gathered in the 1984–86 NLSY waves. In 1984, age at onset was obtained to the nearest year. In 1985, the calendar month and year of onset was asked of all female respondents and repeated in 1986 for all 1984–85 non-respondents. When calendar month of onset was missing or not obtained, we flagged such cases and imputed calendar month using a hot-deck procedure. We censored a woman’s onset history if she reported never having had sex at her age at 1984, 1985, or 1986 interview, depending on the wave providing onset data. We likewise censored a woman’s onset history at her age at first marriage if she reported onset on or after first marriage. The quality of these data appears to be reasonably high (Wu et al. 2001).

We constructed a woman’s age (in months) at a premarital first birth from the calendar month and year of a first birth and a first marriage. We censored a woman’s premarital first birth history at either her age at first marriage or at her age at last interview if she did not give birth prior to first marriage or last survey. We included women with censored onset histories when modeling onset, but excluded them when analyzing premarital first birth risks conditional on onset.

The covariates we examined were dummy variables for race and ethnicity (white, black, Hispanic, and other), measures of family background (mother’s education, number of siblings, catholic religion, intact family structure at age 14, mother’s age at first birth, and a composite index for reading materials at age 14), cognitive ability as measured by the respondent’s age-normalized score on the Armed Forces Qualifying Test (AFQT), and a set of dummy variables for missing data on mother’s education, mother’s age at first birth, family structure at age 14, reading materials, AFQT, and calendar month at first sexual intercourse.

For our analyses of the 0 → T₁ transition, we also controlled for sexual maturity using a time-varying dummy variable equal to 1 at ages following first menses. For the T₁ → T₂ transition, we included a linear term for age at onset (measured to the nearest month) and dummy variables for early and late onset (onset earlier than age 15 and onset at age 21 or later).

Of the 6,283 women present at the initial 1979 interview, we excluded those: (1) with missing data on race and ethnicity (n=45); (2) who reported not knowing their biological mother (n=9); (3) with missing data on the timing of first menstruation (n=254); (4) with missing data on number of siblings (n=9); and (5) with missing first intercourse, first birth, or first marriage histories (n=371). The resulting sample contained n=5,595 women (weighted n=5563.4).

RESULTS

Table 1 summarizes transitions between three statuses—the onset of premarital sexual activity, a premarital first birth, and first marriage—by focusing on the sequence of these events, thus ignoring the ages at which these events occurred. Table 1 shows that 4.1% of NLSY79 women report never having initiated sexual activity, 8.3% report that they were not sexually active prior to a first marriage, and 9.7% report premarital sexual activity but no subsequent marriage or birth. Another 16.3% of women report premarital sexual activity followed by a premarital first birth, while 61.6% report premarital sexual activity followed by a first marriage. Table 1 thus shows that for U.S. women born between 1958 and 1965, the vast majority—87.6% or 7 out of 8 women—reported having initiated sexual activity while never-married.

Table 1.

Frequency distribution of transitions between the statuses of premarital sexual onset, a premarital first birth, and first marriage.

Percent	Transition
4.1	0
8.3	0 → first marriage
9.7	0 → onset
16.3	0 → onset → premarital first birth
61.6	0 → onset → first marriage
100.0	Total

Source: 1979-93 National Longitudinal Survey of Youth

Table 2 presents life table estimates (Kaplan and Meier 1958) of the percent of never-married NLSY79 females initiating sexual activity by selected ages. Onset occurs within the space of a highly condensed set of ages. By age 14, fewer than 4% reported initiating sexual activity, but onset then increases rapidly, with roughly 9% reporting onset by age 15, 75% by age 19, and 90% by age 22. Thus, about 4 in 5 never-married NLSY79 females reported onset between age 15 and 22.

Table 2.

Kaplan-Meier estimates of the percent of never-married females who have initiated sexual activity by selected ages.

Age	10.00	10.25	10.50	10.75	11.00	11.25	11.50	11.75	12.00	12.25	12.50	12.75
Percent	0.13	0.23	0.27	0.30	0.37	0.38	0.44	0.51	0.55	0.74	0.94	1.10
Age	13.00	13.25	13.50	13.75	14.00	14.25	14.50	14.75	15.00	15.25	15.50	15.75
Percent	1.25	1.91	2.37	2.80	3.62	5.26	6.40	7.30	9.03	11.87	13.78	15.62
Age	16.00	16.25	16.50	16.75	17.00	17.25	17.50	17.75	18.00	18.25	18.50	18.75
Percent	19.23	24.66	28.72	32.44	37.23	42.81	47.69	51.28	55.52	61.61	65.77	69.77
Age	19.00	19.25	19.50	19.75	20.00	20.25	20.50	20.75	21.00	21.25	21.50	21.75
Percent	72.81	75.61	77.55	79.41	81.34	82.78	84.04	85.36	86.45	87.51	88.47	89.46
Age	22.00	22.25	22.50	22.75	23.00	23.25	23.50	23.75	24.00	24.25	24.50	24.75
Percent	89.99	90.46	91.00	91.48	91.90	92.25	92.52	92.72	93.00	93.18	93.29	93.36
Age	25.00	25.25	25.50	25.75	26.00	26.25	26.50	26.75	27.00
Percent	93.54	93.67	93.80	93.82	93.88	93.90	93.92	93.92	93.96

Source: As in Table 1

Table 3 presents estimated coefficients from our hazard regressions. We specified age dependence in the 0 → T₁ and T₁ → T₂ transitions using a splined piecewise Gompertz model that specifies logq(t) as a piecewise linear spline with nodes at ages 16, 18.5 and 20. We specified duration dependence in the T₁ → T₂ transition using a piecewise constant specification for durations 0 to 7, 7 to 14, 14 to 36, and 36+ months. Estimated coefficients from these models closely resemble those from a Cox (1972) model; see Appendix Table 1.

Table 3.

Estimates from conventional and sequential hazard regression models for: (a) 0 →T₂, the unconditional transition to a premarital first birth; (b) 0 →T₁, the transition to first sexual intercourse; and (c) T₁ → T₂, the transition to a premarital first birth conditional on age at onset.

	Conventional	Sequential
	0 →T2	0 →T1	T1 →T2
Race and ethnicity
White	----	----	----
Black	0.88***	0.12*	0.76***
Hispanic	0.14	−0.39***	0.54**
Other	0.06	0.04	0.02
Family background
Mother’s education	−0.05***	−0.01	−0.05***
Number of siblings	0.06***	0.01	0.06***
Catholic	−0.20*	0.03	−0.19*
Reading materials	−0.12**	−0.02	−0.12**
Intact family at age 14	−0.62***	−0.34***	−0.34***
Mother’s age at first birth	−0.05***	−0.03***	−0.04***
Ability
AFQT	−0.49***	−0.15***	−0.44***
Sexual maturation and onset
1 if reached menarche by age t	1.12*	0.95***
Age (in months) at onset			−0.009**
Early onset (onset < 15)			0.01
Intermediate onset (15 ≥ onset < 21)			-----
Late onset (onset ≥ 21)			0.39
Age baseline
Intercept (at 0 months)	−27.67***	−16.85***	−14.24***
Slope, t ∈(0,192]	0.113***	0.066***	0.064***
Slope, t ∈(192,222]	0.022***	0.035***	−0.007*
Slope, t ∈(222,240]	−0.001	−0.012***	−0.007*
Slope, t ∈(240, ∞]	−0.005***	−0.033***	−0.005**
Duration baseline
u ∈(0,7]			−0.91***
u ∈(7,14]			----
u ∈(14,36]			−0.20
u ∈(36, ∞]			−0.30

Note: All models also include dummy variables for missing values of: mother’s education, mother’s age at first birth, family structure at age 14, reading materials, AFQT, calendar month of first menstruation, and calendar month of first sexual intercourse. See text for additional details.

^*p < .05

^**p < .005

^***p < .0005 (two-tailed tests)

Source: As in Table 1

The first column of Table 3 presents results ignoring onset timing, with findings consistent with prior studies. Estimates suggest substantially higher risks for blacks, but no significant difference for Hispanics or other race/ethnicities, compared to whites. Mother’s education, number of siblings, Catholic religion, the index of reading materials, and mother’s age at first birth have statistically significant effects in the expected directions. Family structure at age 14 and AFQT have large and statistically significant effects.

The next two columns present results for the transitions to first sexual intercourse and to a premarital first birth conditional on onset. Relative to whites, blacks have significantly higher onset risks as well as significantly higher premarital first birth risks post-onset. However, the Hispanic/white contrasts are opposite in sign for the two transitions, with significantly lower onset risks but significantly higher premarital first birth risks following onset. This suggests that Hispanic women in the NLSY79 delayed onset but had substantially higher premarital first birth risks following onset relative to whites. There are no significant differences in onset or premarital first birth risks following onset for whites and the residual group of other race/ethnicities.

Results from the conventional and sequential hazard models again provide a different picture of the roles of mother’s education, number of siblings, religion, and the index for reading materials. In our sequential model, these variables have small and statistically insignificant effects on onset, but large and statistically significant effects on premarital first births following onset. Our sequential model thus shows that these factors play a major role in post-onset risks but have small effects on onset timing.

The next three rows present estimates for family structure, mother’s age at first birth, and AFQT. Estimates from our sequential model show that these factors have statistically significant effects in the expected directions for both onset and premarital first birth risks. However, estimates are uniformly smaller in magnitude in our sequential models compared to the conventional specification, a pattern most pronounced for family structure. The conventional specification suggests that women from intact families have a 46% lower risk (.46 = exp[−.62]) of a premarital first birth than women from nonintact families, whereas our sequential approach suggests that, conditional on onset, risks are 29% lower (.29 = exp[−.34]) in the period following onset.

The next four rows report estimates for sexual maturation and onset timing. For onset, risks increase substantially following first menses. For premarital first births, later onset, specified using a linear term for age at onset, lowers risks significantly, but net of this, risks for those with especially early or especially late onset do not differ significantly from those with more usual ages at onset.

The T₁ → T₂ equation also contains age and duration baselines. Estimates for the age baseline show that risks increase rapidly at early ages but are relatively flat at later ages. Estimates for the duration baseline show that risks are low in the first seven months following onset, consistent with typical gestational durations. At later durations, risks are higher but flat, with estimated levels at later durations not differing significantly from one another.

We now turn to our decomposition analyses. Panel A of Table 4 shows that the estimated median ages at onset are 17.51 and 17.73 (210.1 and 212.7 months) for black and white women, respectively, holding other factors at their mean values. The 2.6 month difference in onset timing is small but is implied by the modest black/white onset contrast of 0.12 in Table 3. See also Upchurch et al. (1998), who find that black/white difference in onset decline sharply with controls.

Table 4.

Decomposition of differences in the percent of never-married women predicted to have a premarital first birth at 60 and 90 months following onset. White and black women.

Panel A	Age in months
Predicted median age at first intercourse
Blacks	210.1
Whites	212.7
Panel B	Months of exposure
	60	90
Predicted percentage, premarital first birth
A: Blacks	21.82	28.90
B: Whites	10.15	13.97
Predicted difference, premarital first birth
C: A – B	11.68	14.93
Decomposition of C
D: Direct component	10.59	13.80
E: Indirect component, differential age at onset	0.33	0.42
F: Indirect component, differential exposure	0.76	0.70

Note: Predicted percentages are calculated holding all other covariates at their respective means.

Source: As in Table 1

Panel B provides an answer to the question, “What is the probability that a black woman will have a premarital first birth were she to remain never-married and thus exposed to risk for 60 or 90 months after the onset sexual activity?” Because the black median age at onset is 210.1, 60 and 90 months of exposure following onset corresponds to ages 17.5 to 22.5 (210.1 to 270.1 months) and 17.5 to 25.1 (210.1 to 300.1), respectively. Panel B shows that 21.8% and 28.9% of black women are predicted by our model to have a premarital first birth in the 60 and 90 months following black median onset, respectively. Because median onset is 2.6 months later for whites than blacks, our decompositions consider exposures of 57.4 months and 87.4 months for whites. This yields a 10.15% and 13.97% probability of a premarital first birth for whites and thus black/white differences of 11.68% and 14.93% at 60 and 90 months following black median onset, respectively.

We now decompose these 11.68% and 14.93% differences into three components. The first (“D”) reflects black/white differences in premarital first birth risks following onset; the second (“E”) reflects an indirect covariate effect of differences in onset timing that follow from including onset as a covariate in the T₁ → T₂ equation; and the third (“F”) reflects an indirect effect of exposure as generated by an earlier or later age at onset.

The results are striking. A first is a negative finding: Net of controls, the greater probability for blacks is not due to earlier black onset. Black/white differences in onset timing and exposure to risk account for only 0.33 (“E”) and 0.76 (“F”) percentage points, respectively, of the 11.68% black/white difference at 60 months. By contrast, black/white differences in post-onset risks account for 89.9% (10.50/11.68) of the black/white difference in the probability of a premarital first birth in the 60 months following black median onset. Results for 90 months are similar.

Table 5 provides a summary of our decompositions. The largest differences in the probability of a premarital first birth are for blacks and whites (11.7%) followed by women with low and high AFQT (10.5%) and women from nonintact and intact families (6.3%). The decomposition follow a clear pattern in which direct components dominate the indirect components. The decompositions in which the indirect components account for large shares of the total differences are those comparing women from intact and nonintact families, low and high mother’s age at first birth, white and Hispanic women, and white women and women from other race/ethnicities. In the nonintact/intact decompositions, the direct component accounts for 3.9% percentage points of the total nonintact/intact difference of 6.3%, with the two indirect components accounting for 2.5% percentage points or just under two-fifths of the total difference. In the decompositions involving mother’s age at first birth, the direct component accounts for 3.9% percentage points of the total difference of 5.7%, with the two indirect components accounting for 1.8% percentage points, or 31.5% of the total difference. In the white/Hispanic decompositions, the combined indirect components are roughly half the size of the direct component. But for the other seven variables we examined, direct components dominate their corresponding indirect components. (In results not presented, we obtain a very similar pattern for decompositions obtained from estimates that let covariates have nonproportional effects. These results are available upon request.)

Table 5.

Summary of: total, direct, and indirect decomposition components for differences in the percent of never-married women predicted to have a premarital first birth during the first 60 months following the onset of sexual activity; median age at onset for baseline group; and difference in median age at onset between comparison and baseline groups.

	Total	Direct		Indirect
	Exposure Onset age				t1	Δt1
Black vs. white	11.7	10.6	0.8	0.3	210.1	2.6
AFQT ∓ 1.0 sd	10.5	8.5	1.4	0.6	209.7	6.0
Nonintact vs. intact family at age 14	6.3	3.9	1.7	0.8	207.6	7.5
Early vs. late mother’s age at first birth	5.7	3.9	1.2	0.5	210.0	5.4
High vs. low number of siblings	3.7	3.3	0.3	0.1	212.1	1.3
Low vs. high mother’s education	3.3	3.0	0.2	0.1	212.2	1.0
Hispanic vs. white	3.6	6.8	−2.1	−1.0	221.7	−9.0
Low vs. high reading materials	2.7	2.5	0.2	0.1	212.3	0.9
Non-Catholic vs. Catholic	2.3	2.1	0.2	0.1	212.9	0.7
Other race/ethnicity vs. white	0.5	0.2	0.2	0.1	211.8	0.9

Note: Comparisons for AFQT, mother’s age at first birth, mother’s education, and reading materials involve the means of these variables ∓ 1.0 sd, while the comparisons for number of siblings involve the mean of this variable ± 1.0 sd.

Source: As in Table 1

How much does the probability of a premarital first birth decline with later onset? Simple calculations provide some answers, with the caution that these answers are rough and vary slightly with differences between groups in estimated median ages at onset and with the duration of exposure. For example, the last column of Table 5 shows that median age at onset is 7.5 months later for women from an intact family relative to those from a nonintact family. For 60 months of exposure, Table 5 shows that the estimated increase in the probability of a premarital first birth associated with earlier onset is 2.5 percentage points (1.7 + 0.8), thus yielding a 0.33 percentage point decrease in the probability of a premarital first birth per month of onset delay (2.5 percentage points / 7.5 months). For 90 months of exposure, there is a 0.39 percentage point decrease per month of onset delay (results not shown). A similar pattern holds across all covariates we examine, with a one-month delay in median age at onset generating a roughly 0.4 percentage point decline in the probability of a premarital first birth in the period following first sexual intercourse.

A naive extrapolation based on the above might suggest that, for example, a 10-month delay in onset could yield as much as a 4 percentage point decrease in the probability of a premarital first birth. Our results suggest two important cautions to such conclusions. First, our estimates suggest that to have major impacts, effective policies would likely need to delay onset by at least 10 months. But as Table 5 shows, the largest differences in age at median onset occur when comparing whites and Hispanics (9 months), those from nonintact and intact families at age 14 (7.5 months), those with low and high AFQT scores (6.0 months), and those with mothers who began childbearing at early vs. late ages (5.4 months). Thus, sizeable declines in premarital first births would likely require policies that could delay onset by more than the amounts observed between these groups. Second, our results clearly identify the critical and very large role of post-onset risks. Differences in post-onset risks are, in turn, likely to reflect differences in post-onset behaviors, such as the frequency of sexual activity following onset and contraceptive knowledge, use, and consistency. These issues thus point to potential limitations of policies that seek only to delay onset but that do not also seek to influence post-onset behavior.

Our data are observational with our regression analyses yielding associations only; hence, the results in Tables 4 and 5 should be understood as stylized decompositions involving regression-adjusted quantities. If so, how might our estimates of the direct and indirect components in Tables 4 and 5 change were there to be important factors that we do not observe? A statistical “folk” theorem is that direct effects typically dominate indirect effects except when the magnitudes of indirect effects are large (for a more formal statement, see, e.g., Duncan 1975); thus, our results may be unsurprising to many. Still, abstinence proponents might correctly argue that our decompositions rely on mere associations, which are, in turn, subject to the usual omitted variable biases. However, the structure of our sequential hazard model, together with this particular pattern of findings, implies a smaller class of relevant unobservables than is usually the case for the expectations of abstinence proponents.

For concreteness, let w be a critical omitted variable (for example, family wealth) that is correlated with a particular x (say, intact family structure). Then as usual, omitting w would, in general, bias estimates of x, including biases in decomposition results. But an omitted w such as wealth would typically have the same signed effect as family structure; hence, not controlling for w would usually produce too large estimates of both the direct and indirect effects of x; conversely, controlling for w would typically decrease the estimated magnitude of both the direct and indirect effects of x. If so, then controlling for w would not typically alter our particular finding—that the direct effect of x is far larger than the combined indirect effects of x—if the biases for the direct and indirect effects were similar in magnitude to one another.

This suggests that a less usual scenario is needed to reverse the empirical finding that direct effects dominate indirect effects. For example, consider a w with the property that omitting w produces large biases in estimates of direct effects but no biases in the indirect effects of x. If so, the “true” direct effect of family structure can be smaller than the indirect effects of family structure. Note, however, that this possibility also leaves estimates for the indirect effects of x unchanged, posing difficulties for abstinence proponents in that our empirical results suggest not only that direct effects dominate indirect effects, but that indirect effects are also small in absolute magnitude. Thus, what would be needed ideally is a w whose omission leads to upwardly biased direct effects but downwardly biased indirect effects. Typically, one would expect that the correlation between variables such as family structure and wealth would also generate similar signed effects on both the 0 → T₁ and T₁ → T₂ transitions. But if this does not hold—if, for example, wealth in our were to have a positive effect on the 0 → T₁ transition but a negative effect on the T₁ → T₂ transition—then omitting such a w could yield a too small estimate of the indirect effect of x and a too large estimate of the direct effect of x. Omitting such a w could then mask the pattern of effects posited by abstinence proponents—large indirect effects of x via onset delay but negligible direct effects of x on risks following onset. But as these examples make clear, it is more difficult to posit an unobservable w correlated with x in which the effects of w and x have the same signs for one transition but opposite signs for the other transition. Thus, omitted variables can lead, as always, to biased inferences, but the class of unobservables that, if controlled, would yield results consistent with the expectations of abstinence proponents is narrower than is usually the case.

We have simplified our analyses by censoring at first marriage, yet first marriage will be an important source of heterogeneity in the period following onset. Our analyses also did not distinguish between cohabiting and non-union first births, yet factors such as socioeconomic disadvantage are known to be associated with both early onset and increased cohabitation (see, e.g., Seltzer 2000; Raley 2001; Wu and Thomson 2001; Smock et al. 2005; McLanahan and Percheski 2008). Nevertheless, our main finding—that onset timing has at best a small influence on the probability of a premarital first birth—is unlikely to be affected by either cohabitation or first marriage even though both almost certainly will be important factors influencing post-onset birth risks. We rarely regard later events as causes of earlier events; similarly, neither cohabitation nor first marriage are likely to be jointly endogenous with onset, since most young women are not seeking to enter a cohabiting union or to marry when having sex for the first time. And as noted above, other omitted variables are unlikely to substantially bias our finding of the small influence, relative to post-onset factors, of onset timing on the probability of a premarital first birth.

DISCUSSION

In this study, we have heeded two central dictums in demography—to get exposures right and to trace the consequences of differential exposure to risk. Previous research has largely ignored differences in onset timing and thus says little about the consequences of differential exposure to risk implied by earlier or later ages at sexual onset. By contrast, our sequential hazard model of premarital first births stipulates that no woman is at risk of a birth until she becomes sexually active, thus explicitly acknowledging that all births, save those resulting from assisted fertility techniques such as in vitro fertilization, are the product of sexual intercourse. Our model thus focuses attention on two behavioral subprocesses—an onset process by which some never-married women become sexually active and a post-onset process by which some never-married but sexually active women proceed to a premarital first birth.

Our empirical results confirm previous findings that women from disadvantaged backgrounds initiate sexual activity earlier and have more premarital first births than those from more advantaged backgrounds. Our hazard regressions reveal strikingly different effects of covariate for the transitions to sexual activity and to a premarital first birth following onset. A standard set of sociodemographic variables have large and statistically significant associations with premarital first birth risks following onset. These same variables have smaller associations with onset risk, and only a subset are statistically significant. These findings provide empirical hints that some aspects of these two processes may be behaviorally distinct. Results from decompositions comparing median onset timing for disadvantaged and advantaged groups show that, net of controls, differences in exposure to risk have only a small influence on the probability of a premarital first birth, but that group differences in post-onset risks have a sizeable influence on the probability of a premarital first birth. This pattern—small effects of differential exposure to risk but sizeable differences in post-onset risks—holds for all the variables we examine.

Why might this be? Some insight can be gleaned by contrasting what is characteristic of sexual onset vs. a premarital first birth. A higher risk of an event usually implies more individuals experiencing the event, and our results show that this holds for premarital first births. Thus, we find that sizeable and statistically significant group differences in post-onset premarital first birth risks imply equally sizeable group differences in the probability of such births. Sexual onset is different. Premarital sexual activity was both close to universal and highly compressed in this birth cohort of women, with estimated group differences in median ages at onset of 0.7 to 9.0 months after conditioning on background controls. Taken together, near universality and compression mean that even sizeable and statistically significant group differences in onset risks imply only modest group differences in onset timing.

These results also speak, in part, to policies such as the abstinence provisions in the 1996 Personal Responsibility and Work Opportunity Reconciliation Act by providing answers to “what if” questions such as “What might be the expected consequence for premarital first births were a policy to delay onset by an amount equal to that for women from intact vs. nonintact families?” Our results suggest that for premarital first births, onset timing plays a quite minor role relative to post-onset risks. This negative finding—that onset timing plays only a minor role—is, we believe, likely to be robust against factors not observed in our data in that the class of unobservables needed to reverse this finding will be far smaller than usual.

Our results thus suggest that policies seeking to influence post-onset behaviors may be far more effective than those seeking to delay sexual onset. But our findings do not distinguish between a variety of post-onset factors, including the frequency of post-onset sexual activity; contraceptive use, knowledge, and consistency; whether pregnancies are planned or unplanned; abortion availability and utilization; and how such options are weighed when individuals confront different sets of circumstances. Our results thus speak to the question of when policies to reduce premarital first births might be most effective, but say little about which post-onset behaviors might be most effectively targeted.

Further caution is warranted in that the outcomes we analyze have been evolving rapidly. In the United States, there have been sharp declines in teen fertility and steady increases in nonmarital fertility, but only modest increases in women’s age at first sexual intercourse, with premarital sexual activity remaining nearly universal. These trends reinforce our belief that our central finding—that for the probability of a premarital first birth, onset timing plays a quite minor role relative to post-onset risks—will likely hold for more recent birth cohorts of women. What has undoubtedly changed are factors influencing post-onset risks—post-onset contraception has improved among sexually active teens and young adults, and there have been sharp increases in births to cohabiting couples. This study thus provides only a first step towards examining these and other proximate determinants, but our findings also suggest that future research on such factors may hold substantial promise for better understanding the behavioral subprocesses underlying nonmarital fertility.

Appendix Table 1.

Estimated coefficients and standard errors from Cox and splined piecewise Gompertz proportional hazard models for: (a) the unconditional transition to a premarital first birth; (b) the transition to first sexual intercourse; and (c) the transition to a premarital first birth conditional on onset.

	Conventional		Sequential
	0→T2		0 →T1		T1 →T2
	Cox	Gompertz	Cox	Gompertz	Cox	Gompertz
Race and ethnicity
Black	0.88*** (0.09)	0.88*** (0.09)	0.12* (0.05)	0.12* (0.05)	0.76*** (0.09)	0.76*** (0.09)
Hispanic	0.14 (0.16)	0.14 (0.16)	−0.38*** (0.08)	−0.39*** (0.08)	0.54** (0.16)	0.54** (0.16)
Other	0.07 (0.12)	0.06 (0.12)	0.04 (0.04)	0.04 (0.04)	0.01 (0.12)	0.02 (0.12)
Family background
Mother’s education	−0.05*** (0.01)	−0.05*** (0.01)	−0.01 (0.01)	−0.01 (0.01)	−0.05*** (0.01)	−0.05*** (0.01)
Number siblings	0.06*** (0.01)	0.06*** (0.01)	0.01 (0.01)	0.01 (0.01)	0.06*** (0.01)	0.06*** (0.01)
Catholic	−0.20* (0.09)	−0.20* (0.09)	0.03 (0.03)	0.03 (0.03)	−0.19* (0.09)	−0.19* (0.09)
Reading materials	−0.12** (0.04)	−0.12** (0.04)	−0.02 (0.02)	−0.02 (0.02)	−0.12** (0.04)	−0.12** (0.04)
Intact family at age 14	−0.62*** (0.07)	−0.62*** (0.07)	−0.34*** (0.03)	−0.34*** (0.03)	−0.34*** (0.07)	−0.34*** (0.07)
Mother’s age at first birth	−0.05*** (0.01)	−0.05*** (0.01)	−0.03*** (0.00)	−0.03*** (0.00)	−0.04*** (0.01)	−0.04*** (0.01)
Ability
AFQT	−0.48*** (0.04)	0.49*** (0.04)	−0.15*** (0.02)	−0.15*** (0.02)	−0.44*** (0.04)	−0.44*** (0.04)
Sexual maturation and onset
1 if has reached menarche by age t	1.15* (0.56)	1.12* (0.51)	1.03*** (0.12)	0.95*** (0.11)
Age (in months) at onset					−0.006* (0.003)	−0.009** (0.003)
Early onset (age < 15)					−0.01 (0.13)	0.01 (0.12)
Late onset (age ≥ 21)					0.34 (0.39)	0.39 (0.38)
Duration following onset
u ∈(0,7]					−0.91*** (0.15)	−0.91*** (0.15)
u ∈(7,14]					---	---
u ∈(14,36]					−0.15 (0.11)	−0.20(0.11)
u ∈(36, ∞]					−0.12 (0.16)	−0.30 (0.15)

Note: All models include dummy variables for missing values of: mother’s education, mother’s age at first birth, family structure at age 14, reading materials, AFQT, and calendar month of first sexual intercourse. See text for additional details.

^*p < .05

^**p < .005

^***p < .0005 (two-tailed tests)

Source: As in Table 1