Early adolescent sex and diminished school attachment: selection or spillovers?

1. Introduction

The public policy debate over abstinence-only versus comprehensive sex education has generally focused on the effects of such programs on adolescent sexual behaviors and health outcomes. (1) However, less attention has been paid to whether the effective promotion of abstinence yields human capital-related benefits. This study explores the relationship between early adolescent sexual activity and four measures of school attachment--suspension from school, unexcused absences, affinity for school, and aspirations to attend college.

There are several theoretical reasons to expect a negative relationship between early adolescent sex and school attachment. First, sexual intercourse may have adverse psychological, hormonal, or physiological effects on teens that impede human capital accumulation. Moreover, unprotected sex may lead to pregnancies or sexually transmitted infections, which may adversely affect school attachment. Second, it may be that causality runs in the opposite direction. Poor academic achievement may cause teens to psychologically compensate for feelings of loss by engaging in sexual activity. Finally, it may be that there is no causal link between teen sex and school attachment, but rather an association driven by unobserved heterogeneity. Those with the lowest unobserved propensity for human capital accumulation may choose to engage in sexual intercourse at early ages. Carefully addressing the endogeneity of sexual activity will be a central empirical challenge of this study.

Using the National Longitudinal Study of Adolescent Health, I estimate the relationship between teen sexual activity and school attachment. Controlling for a wide set of individual-level and family-level observables, ordinary least squares (OLS) and school fixed effects (SFE) estimates reflect that sexual activity is associated with diminished school attachment for both adolescent males and females. However, after controlling for individual unmeasured heterogeneity via individual fixed effects (IFE) and instrumental variables (IV), the evidence of a causal link is weaker, with modest adverse effects of adolescent sex observed for younger teenagers.

For adolescents aged 13-15, IV estimates suggest that early sexual activity is associated with a modestly higher likelihood of reporting at least one unexcused absence from school and a diminished aspiration to attend college. Taken together, these findings suggest that any positive educational spillovers associated with delaying first intercourse tend to be modest and concentrated among the youngest teenagers.

2. Literature Review

There are several theories in the economic, psychological, and sociological literature that explain a relationship between teen sexual activity and school-related outcomes. Standard rational choice theory suggests that an adolescent will choose to engage in sexual intercourse if the expected gains from engaging in sexual intercourse are greater than the expected costs (see, for example, Nye 1979). The economic theory of fertility, outlined by Becker (1980), implies that those with the lowest shadow prices of sex will choose to engage in it. This theory predicts that those who plan to accumulate the least amount of human capital and have the dimmest future economic prospects will be most likely to exit virginity because they have the least to lose from the consequences of engaging in sexual activity at early ages. Hence, those who are suspended from school, skip classes, feel less a part of their school, and have the least aspirations to attend college are most likely to choose to exit virginity. (2) Thus, when estimating the effect of exiting virginity on school-related outcomes, carefully addressing unmeasured heterogeneity is critical in discerning the nature of the sex-schooling relationship.

While there are strong theoretical reasons to expect that adverse schooling outcomes will impact teens' sexual decisions, there are several plausible theoretical reasons to expect that exiting virginity will cause adverse schooling outcomes. First, losing one's virginity may lead to teen pregnancies or sexually transmitted diseases, which impede schooling. Second, independent of pregnancy effects, there may be physiological or psychological effects of early teen sex that will impact schooling attachment. Entry into sexual activity could lead to concentration problems for teens, with mental resources substituted away from schooling and toward prurient interests. (3) At the same time, early teen sex could cause teens to lose interest in school, altering their taste for education.

Moreover, the experience of sexual intercourse may serve as an information revelation mechanism. Once an adolescent has engaged in sexual intercourse, new information may be revealed to him that changes his future investment decisions. For example, if the ex post benefits of sex are higher than the ex ante anticipated benefits, adolescents may choose to substitute away from investments in schooling and toward short-run investments in sexual conquests. In this sense, the experience of sexual intercourse may serve as an information revelation mechanism that causes teenagers to update their beliefs about the benefits and costs of future investment decisions.

Sociologists and psychologists have posited other theories to explain how sex may affect schooling decisions. Some of these theories are more informative than others. For example, problem behavior syndrome theory posits that immersion in problem behaviors, such as early sexual intercourse, causes a change in an adolescent's mindset such that they will want to explore other nonconformist, antisocial behavior (Donovan, Jessor, and Costa 1988; Elliot and Morse 1989; Rosenbaum and Kandel 1990; Peterson, Moore, and Furstenberg 1991; Farrel, Danish, and Howard 1992; McLean and Flanigan 1993; Whitbeck et al. 1993; Allen, Leadbeater, and Aber 1994; Costa et al. 1995; Harvey and Spigner 1995; Capaldi, Crosby, and Stoolmiller 1996; Schvaneveldt et al. 2001). This theory predicts that the change in mindset or peers caused by sexual experiences could cause teenagers to explore activities such as cutting school, reducing focus in class, and engaging in actions that would result in school suspensions. However, immersion theory does not glean any insight into why adolescents choose to immerse themselves in these behaviors in the first place.

Another psychological theory used to explain changes in behavior caused by the experience of early first intercourse involves depression. A few studies (see, for example, Rector, Johnson, and Noyes 2003) have found that those who are sexually active are more likely to be depressed and to contemplate suicide. If having sex causes teens to become more depressed or to diminish self-esteem, they may find it more difficult to psychologically cope with schooling responsibilities and become less attached to school. However, no recent studies have convincingly shown evidence of a causal relationship between teen sex and depression. (4)

While there is an extensive empirical literature exploring the impact of teenage childbearing on future human capital accumulation (see, for example, Geronimus and Korenman 1992; Hoffman, Foster, and Furstenberg 1993; Bronars and Grogger 1994; Rosenzweig and Wolpin 1995; Angrist and Evans 1996; Hotz, Mullin, and Sanders 1997; Klepinger, Lundberg, and Plotnick 1999; Hotz, McElroy, and Sanders 2005), there is far less empirical work examining the effect of losing virginity on schooling outcomes.

The sociology and psychology literature suggests a strong statistical link between teenage entrance into sexual activity and diminished academic achievement, even after controlling for a wide set of individual-level and family-level observables. Several studies have found that teenagers who have sexual intercourse are more likely to have lower academic achievement and lower academic goals (Jessor et al. 1983; Mott and Marsiglio 1985; Billy et al. 1988; Meilman 1993; Brooke et al. 1994; Costa et al. 1995; Schvaneveldt et al. 2001; Rector and Johnson 2005; Sabia, forthcoming). However, in most of these studies, cross-section OLS models are employed, making it difficult to determine whether there is a causal relationship or simply a correlation due to unobserved heterogeneity. One exception is Sabia (forthcoming), who uses fixed effects and IV methods to show that early teen sex may adversely affect junior high school grade point average (GPA) for adolescent males, but not for females.

This study contributes to the teen sex-human capital literature by carefully examining the sensitivity of the relationship between early sexual activity and school attachment to unobserved heterogeneity. Unlike previous studies in the literature that have relied on cross-section identification strategies, this study presents fixed effects and instrumental variables estimates, which rely on more credible identification assumptions. These methods will permit more appropriate interpretations of the nature of the relationship between adolescent sexual activity and school-related outcomes.

3. Data

The dataset used for this analysis is the National Longitudinal Study of Adolescent Health (Add Health), a school-based nationally representative survey with information from adolescents, parents, and school administrators. (5) Adolescents were interviewed in consecutive academic years, in 1994-1995 (wave 1) and again in 1995-1996 (wave 2). In each wave, information on sexual activity, schooling, family, personality, peers, health behaviors, attitudes, and neighborhoods was collected. To reduce the likelihood of reporting error on sensitive topics such as sexual experience, adolescents were assured by interviewers that their answers could not be matched with their names by anyone. Moreover, not even the interviewer knew of the responses since questions were answered via a computer given to the adolescent. This is important inasmuch as adolescents may be shy in reporting truthful information on personal subjects to interviewers.

In addition to adolescent interviews, parents were interviewed in wave 1 and school administrators were interviewed in waves 1 and 2. Mothers provided information about the nature of their relationships with their adolescent children, their families, backgrounds, involvement in school activities, and motivations for decision making. School administrators provided information on school policies, procedures, academic requirements, and general school organization.

This analysis focuses on four key measures of school attachment: out-of-school suspensions, unexcused absences from school, affinity for school, and aspirations to attend college. While these variables are not perfect measures of school attachment, they are expected to be good proxies. For example, while the number of times an adolescent skips a full day of school without excuse is a useful measure of school attendance, a variable capturing how often he cuts classes and how many classes he misses would be desirable as well. However, missing full days of school is expected to be highly correlated with frequent cutting of individual classes and is interpreted as such. Similarly, rather than suspensions, a measure of the number of detentions for the adolescent might be preferable since detention is a more common school punishment for failure to complete school work, skipping class, and general insubordination. Suspensions are generally thought to be associated with more severe law-breaking activities such as underage drinking and drug use. However, this is not an entirely accurate assessment of suspension policies. While suspension is a common punishment for illegal activities, it is also a tool used to punish adolescents who have had chronic detentions, have verbally abused teachers, and have cheated on school work. (6)

The weighted means and standard deviations for the key dependent variables and independent variables from wave 1 are found in Table 1, by gender and age. The first dependent variable is a dichotomous measure of suspensions, SUSPEND. In each wave, adolescents are asked, "During the [current] school year, have you received an out-of-school suspension?" (7) In response, 5.6% of males aged 13-15 and 11.1% of 16- to 18-year-olds reported an out-of-school suspension. The percentages were lower for females, at 3.4% of 13- to 15-year-olds and 5.6% of 16- to 18-year-olds.

The second dependent variable measures the number of unexcused absences from school, SKIP. In response, 17.3% of 13- to 15-year-old males and 16.3% of 13- to 15-year-old females reported at least one unexcused absence from class in the current academic year. For 16- to 18-year-olds, the numbers were higher, at 32.6% of males and 27.8% of females. (8)

The third variable measures an index of affinity the adolescent has for his school, PARSCH. Adolescents are shown the statement, "You feel like you are a part of your school" and are asked to answer whether they agreed or disagreed with the assertion. The variable is coded to one if adolescents report that they "strongly agree" with the statement, to two if they "agree" with the statement, to three if they report they "neither agree nor disagree," to four if they "disagree," and to five if they "strongly disagree." Across adolescent males and females, most students report that they agree with the statement, with means ranging from 2.0 to 2.2. (9)

The final dependent variable measures an index of the adolescent's aspirations to attend college, COLLEGE. Adolescents are asked, "On a scale of 1 to 5, where 1 is low and 5 is high, how much do you want to go to college?" The means response to this variable was around 4.5 across adolescents, indicating a generally strong preference to attend college. (10)

The key variable of interest is whether the adolescent has ever engaged in sexual intercourse, INTERCOURSE. The survey asks students, "Have you ever had sexual intercourse? When we say sexual intercourse, we mean when a male inserts his penis into a female's vagina." As expected, the proportion of teenagers who have had sexual intercourse rises with age, with 16% to 17% of 13- to 15-year-olds and 43% to 44% of 16- to 18-year-olds reporting that they had engaged in sexual intercourse at least once. Reports of sexual activity in Add Health compare favorably with other large nationally representative datasets. For example, the 1995 National Survey of Family Growth and the 1995 National Survey of Adolescent Males find that 24% of 15-year-old females and 27% of 15-year-old males were sexually active. This compares with 25% of 15-year-old females and 31% of 15-year-old males in my Add Health sample. This suggests little evidence of underreporting of sexual activity in this sample.

It is important to distinguish the effects of sexual intercourse on school attachment from the effects of entering into a romantic relationship. Thus, a key control variable is the adolescent's report of whether he or she is in a romantic or romantic-like relationship, ROMANTIC. Moreover, controls for other risky health behaviors that are correlated with sexual activity--alcohol consumption (BINGEYR) and drug use (DRUG)--are also included.

Other key family-level and individual-level socioeconomic characteristics associated with schooling outcomes are included in Table 1. They include measures of intelligence, race, age, family structure, physical and mental health, religiosity, puberty onset, and a series of measures capturing parental attitudes toward education. (11)

4. Methods

OLS and Fixed Effects

Most studies estimating the relationship between teen activity and school attachment have used cross-sectional models of the following form:

[A.sub.i] = [alpha] + [[beta].sub.1] [INT.sub.i] + [n.summation over (j=2)] [[beta].sub.j] [X.sub.ij] + [[epsilon].sub.i], (1)

where [A.sub.i] is a measure of school attachment for individual i, [INT.sub.i] measures whether the adolescent has engaged in sexual intercourse, and X is a set of individual- and family-level observable characteristics. However, the key identification assumption, E([epsilon] | [INT) = 0, may be violated in the presence of unmeasured heterogeneity or reverse causality.

To address the potential problem of fixed individual-level unobserved heterogeneity, IFE models are estimated as

[A.sub.it + 1] - [A.sub.i] = [[beta].sub.1]([INT.sub.it + 1] - [INT.sub.it]) + [n.summation over (j = 2)] [[beta].sub.j]([X.sub.ijt + 1] - [X.sub.ijt]) + ([[epsilon].sub.it + 1] - [[epsilon].sub.it]). (12) (2)

The identification assumption here requires common unobserved time trends, E([[epsilon].sub.t+l] - [[epsilon].sub.t] | [INT.sub.t+1] - [INT.sub.t]) = 0. However, this assumption will be violated if there are time-varying unobservables correlated with selection into sexual intercourse and with changes in school attachment. For example, if an adolescent's personal discount rate rises or the perceived opportunity costs of sex decline, then we might observe a spurious correlation between exiting virginity and decreased school attachment. Moreover, reverse causality may lead to estimates biased toward adverse schooling effects. Adolescents who become less attached to school may choose to substitute energies away from school and toward sexual relationships.

Instrumental Variables

To address the endogeneity of sexual intercourse, instrumental variable models are estimated via two-stage least squares (2SLS). The first stage virginity equation is given by

[INT.sub.i] = [alpha] + [summation over (k = 1]) [kappa][Z.sub.i] + [n.summation over (j = 2)] [[lambda].sub.j][X.sub.ijt] + [bar.[omega].sub.ij], (13) (3)

where [Z.sub.i] is the set of exclusion restrictions. In order for instruments to be valid, they must be uncorrelated with unobserved determinants of school attachment, E([epsilon] | Z) = 0.

Seven exclusion restrictions are used in the analysis. The means of these variables are shown in Appendix B. First, two school policies that are expected to influence adolescent sexual activity, but not directly influence school attachment, are included in the vector Z: the presence of day care services provided in the school or by the district (DA YCARE) and school sex education that has a family planning component (SEXEDFAM). (14)

Providing day care services to enrolled students is expected to reduce the costs of sexual activity by reducing the direct costs of out-of-wedlock births. Family planning focused sex education is also expected to decrease the likelihood of virginity by providing low-cost contraceptive services and, perhaps, by destigmatizing teen sex. (15)

Two measures of the adolescent's schoolmates' sexual attitudes are also included as exclusion restrictions: the mean percentage of schoolmates who associate entrance into sexual activity with feelings of strong guilt (SCHGUILT) and the mean percentage of schoolmates who believe that their attractiveness will be enhanced by losing virginity (SCHATT). Note that these are not measures of attitudes of the adolescents' chosen peers, but rather of the sample of adolescents randomly selected in the school by administrators of the Add Health study. Schoolmates' attitudes toward sex are expected to impact the cost of exiting virginity because it affects the ease with which an adolescent can find a sexual partner. An adolescent attending a school with schoolmates who have more conservative attitudes toward sex faces a higher cost of finding a sexual partner. One would not expect that schoolmates' attitudes toward sex would be directly correlated with school attachment except through their effect on engaging in sexual activity.

Finally, three family-level or individual-level characteristics are included in Z: parental disapproval of teen sex, parent-child discussions of sex, and intensity of religious views. Adolescents who perceive that their parents strongly disapprove of them having sexual intercourse (PARSEX) will have a higher price of sex relative to those who do not perceive such strong parental attitudes. One would expect that adolescents with higher prices of sex will be more likely to remain virgins. Second, I include a measure of whether the parent has spoken to the adolescent about sex (TALKSEX), which may be an important informational path through which adolescent sexual decisions are made. And finally, a measure of the adolescent's religiosity is included: whether the adolescent believes that the Bible is the literal word of God and considers him or herself to be a born-again Christian (BBORN). Adolescents with strong religious convictions are more likely to remain virgins.

While each of the above exclusion restrictions is theoretically expected to be a strong predictor of adolescent sexual activity, an important critique of these instruments is that one or more measures may be correlated with unobservable characteristics that are correlated with school attachment. For example, an adolescent's schoolmates' attitudes are not necessary exogenous inasmuch as parents choose the schools that their children attend. Parents who place their children among schoolmates who have conservative attitudes toward sex are more likely to care about their children's school attachment. These parents are likely to put in more effort toward ensuring that their children are focused on school. (16) Moreover, families that encourage long-run goal setting and delayed gratification, both of which are likely to be correlated with greater school attachment, are also likely to disapprove of sex and to encourage virginity.

One response to concerns about the validity of the instruments is that the richness of the Add Health data allows controls for several observable measures of family-level proschooling sentiment, many of which are noted in Table 1: whether the parent is a member of the Parent Teacher Association (PTA), whether the parent moved to the neighborhood because of the local school quality (NEIGHBOR), whether the parent strongly values scholastic brilliance (BRILLIANT), whether the parent has helped the adolescent with recent school work (PARHELP), parental educational attainment (COLLEGRAD), and parental attitudes about the adolescent's college attendance (PARDISCOL). If these observable family-level schooling sentiment measures capture important schooling attitudes, then the instruments should be valid. Furthermore, because of the use of several exclusion restrictions in Z, formal overidentification tests are presented to examine the validity of the instruments.

However, given concerns about the identification assumptions underlying the standard IV model described above, an alternative approach offered by Lewbel (2006) is also used. The Lewbel IV method allows the inclusion of the above exclusion restrictions in both the intercourse and school attachment equations and identifies the 2SLS model off of heteroskedasticity in the intercourse equation. In the presence of heteroskedastic disturbances in Equation 3, Lewbel (2006) shows that [[beta].sub.1] can be consistently estimated using (Z - [bar.Z])[??] as exclusion restrictions, where [??] are the estimated residuals from Equation 3, and Z is a vector of observed exogenous variables that can be a subset of X or can equal X. In the Lewbel IV models presented in this paper, the variables that serve as Z are the above exclusion restrictions and urbanicity of the community. Thus, in the Lewbel formulation, the variables in Z are included in both Equations 1 and 3, with the identification of the parameters coming from heteroskedastic disturbances in 3. The identification assumption simply requires heteroskedasticity in Equation 3 (and Equation 4 to permit reverse causality) and cov(Z, [epsilon][omega]) = 0. These assumptions are shown by Lewbel (2006) to be features that are quite common in models in which the correlation of errors ([epsilon] and [omega]) are due to an unobserved common characteristic.

Lewbel (2006) notes that the assumptions underlying his approach have also been exploited to identify correlated random coefficients models (Heckman and Vytlacil 1998). Moreover, King, Sentana, and Wadhwani (1994); Sentana and Fiorentini (2001); Rigobon (2002, 2003); and Klein and Vella (2003) have exploited heteroskedasticity to identify models in a manner similar to that proposed by Lewbel (2006). Leamer (1981) and Feenstra (1994) also exploit heteroskedasticity to aid in identification. As Lewbel (2006) notes, several recent papers have proposed restrictions on higher order moments rather than traditional instruments as an alternative method of identification (Cragg 1997; Dagenais and Dagenais 1997; Lewbel 1997; Vella and Verbeek 1997; Erickson and Whited 2002). Thus, in the absence of standard instruments that fail the relevance or exogeneity tests, the Lewbel approach provides a potentially more credible strategy for identifying the causal effects of early teen sex. Of course, technical identification of this form lacks the intuitive appeal of a cleverly conceived natural experiment.

5. Results

Estimation results are presented in Tables 2-9. Separate models are estimated by gender and age to allow for heterogeneous effects of first intercourse. Taken together, the results reflect consistent evidence of a negative association between teen sexual intercourse and school attachment. However, after controlling for individual unmeasured heterogeneity with IFE and IV, evidence of a causal link is somewhat weaker. These findings suggest that unobserved individual traits can explain much of the relationship, and that any negative spillovers tend to be modest and concentrated among younger teens aged 13-15.

OLS and Fixed Effects Estimates

Table 2 presents OLS, school fixed effects, and IFE estimates of the relationship between sexual intercourse and school attachment. OLS and school fixed effects estimates are obtained using baseline data (wave 1), and IFE estimates are obtained using data from both wave 1 and wave 2. IFE estimates include time-varying observable characteristics noted in Table 1. (17) Each estimate presented in Table 3 comes from a separate regression model. (18)

Columns 1-6 of the first row of Table 2 present OLS and fixed effects estimates of the relationship between sexual activity and the probability of suspensions for adolescent males; this information is presented in columns 7-12 for females. Consistent with much of the previous schooling-teen sex literature, linear probability model estimates show robust evidence of a significant positive relationship between engaging in sexual activity and the probability of suspension. Moreover, this significant relationship persists after controlling for school fixed effects and individual fixed effects for some adolescents.

While the estimates in the first row could suggest the presence of negative educational spillovers of losing virginity, they do not necessarily imply positive spillovers of abstinence. Adolescents who are no longer virgins could be (i) currently practicing unsafe sex, (ii) currently practicing safe sex, or (iii) not currently sexually active. It may be that unprotected sex or current sexual activity is driving the negative relationship observed in the first row of estimates. This may be the case if, for example, fear of contracting sexually transmitted diseases or teen pregnancy causes diminished school attachment. To conclude that abstinence is positively associated with school attachment, one would need to show that the outcomes of virgins are better than the outcomes of those who are either engaging in safe sex or not currently sexually active. Next, I explore two different "control" groups for virgins.

The estimates presented in the second row of Table 2 compare the outcomes of those who are virgins with the outcomes of those who (i) have engaged in protected sex at most recent intercourse or (ii) are not virgins, but have not been sexually active during the academic year. The estimates in the second row continue to show a significant positive relationship between losing one's virginity--even to protected sex--and the probability of being suspended.

One concern with the estimates presented in the first two rows is that they do not control for oral or anal sex by virgins or nonvirgins. It may be that adolescents have engaged in other unmeasured forms of sex. While waves 1 and 2 of the Add Health data do not have explicit information on oral or anal sex, there is information on whether the adolescent has engaged in any form of genital contact with their most recent romantic partner. In the final row of Table 2, virgins who reported genital contact with a recent romantic partner are excluded from the sample. Thus, these regressions compare the outcomes of virgins who have had no genital contact with romantic partners ("super" virgins) to those who have had protected sexual intercourse at most recent intercourse. The evidence continues to show a positive relationship between sexual intercourse and the probability of suspension, even after controlling for fixed individual-level unobserved heterogeneity. However, given the inability to control for oral or anal sex among sexually active adolescents, it could be that the negative relationship between intercourse and school attachment is explained by sexual activity other than intercourse.

Tables 3-5 present OLS and fixed effects results for the three other measures of school attachment: probability of an unexcused school-day absence, strength of school affinity, and intensity of college aspirations. Across all measures of school attachment and across all definitions of the "control" group, OLS and school fixed effects estimates continue to show a negative relationship between sexual intercourse and school attachment. However, for these outcomes, the inclusion of individual fixed effects often eliminates this relationship suggesting that time-invariant unmeasured heterogeneity can explain much of the negative relationship. Still, IFE estimates continue to show that sexual intercourse diminishes school affinity for males and reduces the college aspiration for young females.

However, caution should be taken in interpreting IFE estimates causally and concluding that there are positive educational spillovers associated with delaying first intercourse. Unmeasured time-varying characteristics may be correlated with both entrance into sexual intercourse and with measured school attachment. For example, while the IFE models do control for changes in self-reported mental health and for the onset of puberty, unmeasured changes in self-esteem or peer pressure may be associated with both entrance into sexual intercourse and with diminished school attachment. Moreover, reverse causality may play an important role. Teenagers who become less attached to school may be more likely to substitute energies away from schooling and toward developing sexual relationships. Thus, in the remaining tables, the endogeneity of sexual intercourse is more carefully addressed with an instrumental variables strategy.

Instrumental Variables Estimates

In Tables 6-9, standard IV and Lewbel IV estimates are presented. (19) OLS estimates using the IV sample are also presented to allow direct comparison with IV estimates. F statistics on the joint significance of the instruments, [R.sup.2] values for the first-stage regressions, p values from Sargan overidentification tests, and p values from the first-stage Breusch-Pagan heteroskedasticity tests are presented in the tables. Coefficient estimates on the exclusion restrictions for the standard IV models are found in Appendix B.

Consistent with theoretical expectations, school-based day care services, family planning services, schoolmate linking of sex with attractiveness, and parental discussions of sex with adolescents are each generally positively associated with the probability of nonvirginity. Strong parental disapproval with sex, born-again Christianity, and schoolmate linking of sex with guilt are each generally negatively associated with the probability of nonvirginity.

Table 6 presents results on suspensions. Consistent with previous findings, OLS estimates on the IV sample show consistent evidence of a positive relationship between sexual activity and the probability of suspension across each comparison of nonvirgins versus virgins (see rows 1 to 3).

After addressing endogeneity bias with standard instrumental variables techniques, IV estimates continue to show a significant positive relationship between adolescent sexual activity and the probability of suspension for both males and females across ages. In each model, F statistics on the joint significance of the instruments are generally greater than I0, suggesting strong instrument relevance. Sargan overidentification tests fail to reject the null hypothesis that the instruments are uncorrelated with estimated unobserved determinants of schooling.

While overidentification tests may be somewhat reassuring, the magnitudes of the standard IV estimates are of concern. They are uniformly greater in magnitude than OLS estimates, on the order of five to seven times as large. Hausman tests reflect that these differences are statistically significant. One explanation for this result could be that OLS estimates are actually biased downward. This could be the case if unobserved characteristics positively associated with school attachment are also positively associated with losing virginity. One such unmeasured characteristic could be popularity. Popular students may be more attached to school--often because of involvement in school activities and sports--and also more likely to find sexual partners. However, in unreported specifications that include controls for involvement in sports and self-perceived popularity, the magnitudes of the IV estimates are not substantially reduced.

An alternative explanation for why standard IV estimates are significantly larger than OLS estimates is the presence of weak or invalid instruments. Stock and Yogo (2005) suggest that one of the consequences of weak instruments may be large size distortions in parameter estimates. Moreover, overidentification tests may not be sufficiently powerful to detect whether the instruments are valid. Thus, caution should be taken in interpreting standard IV results as evidence of a causal relationship between sexual activity and school attachment.

An alternate IV approach that allows the above exclusion restrictions to be included in the school attachment equation is the Lewbel strategy. Results from these models are also presented in Table 6. The identification assumption of the Lewbel IV models requires heteroskedasticity in the intercourse and schooling equations, both of which are found in these data, and cov(Z, [epsilon][omega]) = 0. For 13- to 15-year-old males and females, there is strong evidence of heteroskedasticity, leading to stronger instruments. Note that the F tests on the joint significance of the exclusion restrictions are higher for the Lewbel IV models than for the standard IV models, and that the [R.sup.2] on the first-stage regressions suggests a stronger goodness of fit. However, for 16- to 18-year-olds, Breusch-Pagan tests reveal less strong evidence of heteroskedasticity, leading to weaker instruments. Still, one might prefer Lewbel IV estimates to standard IV estimates for 16- to 18-year-olds even in the presence of weaker instruments if one is concerned that the exclusion restrictions in the standard IV model (Z) are correlated with unobserved determinants of schooling.

Across specifications, Lewbel IV estimates show little evidence of a significant relationship between adolescent sexual activity and the probability of an out-of-school suspension. This finding suggests that OLS, fixed effects, and standard IV estimates may be biased toward negative schooling outcomes and casts some doubt on the hypothesis that sexual intercourse has important educational spillovers.

Results in Tables 7-9 present IV estimates for the remaining measures of school attachment. In Table 7, Lewbel IV estimates reflect some evidence that sexual intercourse increases the probability of having an unexcused absence from school for the youngest teens (aged 13-15). Hausman tests show that Lewbel IV estimates are statistically equivalent to OLS estimates for these adolescents. Tobit estimates, presented in Appendix C, suggest that early adolescent sex is associated with a three- to five-day higher number of unexcused absences per year, which may indicate a modest decline in school attachment. (20) These findings reflect that delaying first intercourse among the youngest teenagers could have modest positive spillover effects on this dimension of school attachment.

The findings on school affinity (Table 9) reflect that after controlling for the endogeneity of school affinity via the Lewbel approach, there is little evidence of a causal link between postponing sex and greater affinity for one's school. However, for college aspirations, the evidence is mixed. For adolescent males, while those who have engaged in sexual intercourse are less likely to aspire to attend college, there is little evidence of a causal link, as indicated by insignificant IV estimates. But, for females, there is some evidence of a significant negative relationship between sexual activity and the intensity of preference to attend college. However, the magnitude of the estimate is rather modest, representing about one-quarter standard deviation difference between virgins and nonvirgins. Moreover, this relationship becomes smaller when comparing virgins who have had no genital contact to those who are having protected sex.

Taken together, the results in Tables 2-9 suggest that while OLS estimates of the relationship between virginity and schooling outcomes are generally consistent with much of the existing empirical literature (see, for example, Jessor et al. 1983; Mott and Marsiglio 1985; Billy et al. 1988; Meilman 1993; Brooke et al. 1994; Costa et al. 1995; Schvaneveldt et al. 2001; Sabia, forthcoming), the estimated association is weaker after controlling for unmeasured heterogeneity, suggesting that adolescents with the lowest unobserved propensity for academic achievement choose to engage in sexual activity. These findings are generally consistent with Sabia (2006c), who found that unmeasured heterogeneity can explain the negative relationship between adolescent sex and grade point average for some teens. (21)

The negative academic spillovers associated with early teen sex appear to be modest and concentrated among younger teenagers. Lewbel IV estimates reflect that adolescent males aged 13-15 who engage in sexual activity are 15 to 17 percentage points more likely to have positive unexcused absences from school, and adolescent females aged 13-15 are 10 to 12 percentage points more likely to have unexcused absences. For adolescent females, there is also some evidence that early adolescent sex may diminish the intensity of aspirations to attend college, though the magnitude of the relationship is small. (22) Thus, while selection can explain much of the negative relationship between early adolescent sex and school attachment, there may be modest, positive spillovers associated with delaying intercourse for young teens.

6. Conclusions

While the public policy debate over abstinence-only sex education often involves a discussion of the health benefits of abstinence, this study presents evidence on whether there are nonsexual, human capital--related spillovers to exiting virginity. Consistent with much of the existing empirical literature, this study finds evidence of a negative relationship between adolescent sexual activity and school attachment. Early teen sex is associated with a higher probability of suspension, a higher probability of unexcused absences from school, a lessened affinity for school, and reduced aspirations to attend college. However, after controlling for individual unmeasured heterogeneity with individual fixed effects and instrumental variables, the estimated relationship is somewhat weaker, reflecting the importance of controlling for unobservables. Negative schooling spillovers that persist appear to be fairly modest and concentrated among the youngest teens.

One should be cautious in interpreting the findings of this study as support for abstinence-only sex education. The estimated magnitudes of the relationship between sexual activity and school attachment are rather small in magnitude. For example, among adolescent girls, the mean preference for college aspiration among nonvirgins is less than one-quarter standard deviation lower than the mean preference among virgins. Thus, future research is necessary on whether delaying intercourse has important effects on school attainment.

Finally, even if delaying the time until first intercourse reduces some adverse school-related outcomes and these human capital improvements are economically meaningful, a general equilibrium analysis might find that adolescents will find other counterproductive outlets for their sexual energies, such as reckless driving, or fighting in and outside of school. Thus, while cost-benefit analyses of sex education programs designed to delay the time until first intercourse should consider both health and schooling-related benefits of abstinence, the findings of this study should not necessarily be interpreted to imply that promoting abstinence will enhance social welfare. A more complete general equilibrium analysis is necessary to draw such a policy conclusion.

Appendix A
OLS Estimates of Relationship between Any
Sexual Intercourse and School Suspension

                                 Males

                    Age 13-15            Age 16-18

INT              0.042 ** (0.017)     0.089 *** (0.020)
DRUG             0.036 ** (0.014)     0.075 *** (0.022)
DRINK           -0.016 (0.019)        0.017 (0.022)
PREFCOL         -0.026 * (0.014)     -0.049 *** (0.017)
ROMANTIC         0.006 (0.010)        0.000 (0.017)
AFDC             0.000 (0.010)       -0.025 (0.024)
BMI             -0.002 * (0.001)     -0.000 (0.002)
WORK            -0.011 (0.012)        0.006 (0.016)
INCOME          -0.017 (0.011)       -0.001 (0.011)
SINGLPAR        -0.020 (0.017)       -0.024 (0.028)
BLACK            0.049 ** (0.023)     0.088 *** (0.028)
HISPANIC         0.008 (0.021)       -0.010 (0.023)
AHPVT           -0.000 (0.003)       -0.001 * (0.001)
DEPRESS          0.028 (0.042)        0.052 (0.039)
BIRTHWGT         0.030 (0.029)        0.027 (0.060)
BADHEALTH        0.015 (0.039)        0.096 * -0.056
FIGHT            0.001 (0.012)        0.040 * (0.022)
NOTGETAL         0.107 (0.114)       -0.061 (0.069)
OLDERSIB         0.010 (0.011)       -0.027 * -0.015
FRIENDSUICIDE    0.002 (0.015)       -0.006 (0.002)
FAMSUICIDE      -0.001 (0.029)       -0.100 *** (0.038)
NEIGHBOR        -0.020 * (0.011)     -0.005 (0.020)
PTA             -0.015 (0.011)       -0.002 -0.018
WORKOUT         -0.025 (0.018)       -0.007 (0.021)
BRILLIANT       -0.003 (0.012)        0.015 (0.016)
PARTALK         -0.016 (0.015)        0.063 * (0.036)
WEEKENDTM        0.029 * (0.016)     -0.019 (0.015)
DINNERWK         0.000 (0.004)       -0.005 * (0.003)
PARDISCOL        0.014 (0.014)       -0.028 (0.023)
AGE13           -0.072 ** (0.029)    --
AGE14            0.009 (0.014)       --
AGE17                   --            0.005 (0.025)
AGE18                   --            0.045 (0.039)
RELIGWK          0.005 (0.017)       -0.036 (0.031)
RELIGMO         -0.006 (0.018)        0.007 (0.032)
RELIGYR          0.011 (0.017)        0.001 (0.030)
COLLGRAD        -0.016 (0.011)       -0.011 (0.017)
NOTCARE         -0.034 (0.039)        0.083 (0.079)
MIDWEST         -0.007 (0.016)        0.041 * (0.022)
WEST             0.006 (0.019)        0.006 (0.029)
SOUTH           -0.002 (0.016)       -0.012 (0.020)
SUSVERB          0.010 (0.013)        0.020 (0.019)
RURAL           -0.022 (0.020)        0.062 ** (0.026)
SUBURB          -0.005 (0.020)        0.016 (0.018)
PUBLIC          0.006 (0.014)         0.059 ** (0.024)
N                      2530                 2301

                                Females

                    Age 13-15            Age 16-18

INT              0.028 * (0.016)      0.041 *** (0.010)
DRUG             0.002 (0.011)        0.041 ** (0.019)
DRINK            0.018 (0.015)        0.011 (0.014)
PREFCOL         -0.004 (0.012)       -0.056 *** (0.021)
ROMANTIC        -0.004 (0.012)        0.005 (0.011)
AFDC             0.026 (0.022)       -0.013 (0.024)
BMI             -0.002 * (0.001)      0.000 (0.002)
WORK            -0.008 (0.010)        0.014 (0.015)
INCOME           0.005 (0.006)       -0.015 * (0.009)
SINGLPAR        -0.005 (0.013)        0.027 (0.019)
BLACK            0.053 *** (0.019)    0.081 *** (0.020)
HISPANIC         0.026 (0.017)       -0.025 (0.018)
AHPVT           -0.0005 * (0.0003)   -0.0006 * (0.0004)
DEPRESS         -0.037 (0.035)        0.033 * (0.019)
BIRTHWGT         0.007 (0.035)       -0.110 *** (0.039)
BADHEALTH        0.008 (0.027)       -0.013 (0.024)
FIGHT            0.012 (0.012)        0.018 (0.013)
NOTGETAL        -0.035 *** (0.013)    0.048 (0.101)
OLDERSIB         0.006 (0.008)       -0.007 (0.013)
FRIENDSUICIDE    0.006 (0.012)        0.010 (0.017)
FAMSUICIDE      -0.024 (0.019)       -0.054 ** (0.025)
NEIGHBOR        -0.014 (0.010)        0.005 (0.012)
PTA              0.011 (0.011)        0.006 (0.020)
WORKOUT         -0.007 (0.012)       -0.005 (0.019)
BRILLIANT        0.009 (0.008)        0.018 (0.012)
PARTALK          0.003 (0.013)       -0.001 (-0.021)
WEEKENDTM        0.005 (0.011)        0.025 (0.018)
DINNERWK        -0.001 (0.002)       -0.003 (0.002)
PARDISCOL        0.014 (0.010)       -0.010 (0.013)
AGE13           -0.006 (0.017)       --
AGE14            0.018 *** (0.006)   --
AGE17           -0.006 (0.017)       -0.038 ** (0.015)
AGE18            0.018 *** (0.006)    0.002 (0.032)
RELIGWK         -0.001 (0.016)        0.022 (0.018)
RELIGMO         -0.014 (0.018)        0.001 (0.021)
RELIGYR         -0.002 (0.020)        0.016 (0.023)
COLLGRAD        -0.016 (0.011)       -0.005 (0.017)
NOTCARE         -0.012 (0.020)        0.040 (0.040)
MIDWEST          0.002 (0.015)        0.016 (0.020)
WEST            -0.014 (0.013)       -0.031 (0.016)
SOUTH           -0.011 (0.014)       -0.014 (0.015)
SUSVERB         -0.031 ** (0.013)    -0.036 ** (0.018)
RURAL           -0.010 (0.010)        0.057 ** (0.025)
SUBURB          -0.015 (0.010)       -0.008 (0.012)
PUBLIC           0.031 *** (0.012)    0.014 (0.016)
N                      3008                 2304

Source: Computed by the author.

Standard errors of the estimates appear in parentheses.
Each model also includes dummy variables for religious
affiliation, number of excused absences from school,
grade in school, and extent of puberty onset, measured
by facial and underarm hair for males and breast size
and curves for females.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Appendix B
Means of Exclusion Restrictions Used in Standard IV Models

                                                           Males

                                                  Age 13-15   Age 16-18

DAYCARE    School provides some form of day         0.189       0.187
             care for adolescents who have         (0.391)     (0.390)
             children
SEXEDFAM   Sex education mandate with family        0.208       0.215
             planning services provided or         (0.406)     (0.411)
             referred by school
BBORN      Adolescent believes Bible is literal     0.251       0.214
             word of God and considers himself     (0.434)     (0.410)
             to be born-again Christian
SCHGUILT   Proportion of schoolmates who            0.082       0.138
             associate strong guilt with sexual    (0.075)     (0.061)
             activity
SCHATT     Proportion of schoolmates who            0.052       0.093
             associate increased attractiveness    (0.043)     (0.042)
             with sexual activity
PARSEX     Parent strongly disapproves of           0.628       0.361
             adolescent engaging in sexual         (0.483)     (0.480)
             activity
TALKSEX    Parent has spoken with adolescent        0.304       0.386
             about sexual activity                 (0.460)     (0.487)
N                                                   3127        3958

                    Females

           Age 13-15   Age 16-18

DAYCARE      0.206       0.189
            (0.405)     (0.392)
SEXEDFAM     0.210       0.216
            (0.405)     (0.411)
BBORN        0.272       0.233
            (0.445)     (0.423)
SCHGUILT     0.085       0.137
            (0.077)     (0.060)
SCHATT       0.054       0.093
            (0.043)     (0.041)
PARSEX       0.812       0.555
            (0.391)     (0.497)
TALKSEX      0.418       0.456
            (0.493)     (0.498)
N            3488        3825

Source: Computed by the author.

Standard deviations of means appear in parentheses.

Appendix C
First-Stage Estimates on Excluded
Instruments from Standard IV Models

                                Males
Exclusion
Restrictions       Age 13-15            Age 16-18

DAYCARE         0.015 (0.017)        0.056 *** (0.018)
SEXEDFAM       -0.011 (0.017)       -0.000 (0.020)
BBORN          -0.043 ** (0.018)     0.000 (0.020)
SCHATT          0.482 *** (0.209)    0.248 (0.189)
SCHGUILT       -0.094 (0.121)       -0.476 *** (0.132)
PARSEX         -0.136 *** (0.015)   -0.108 *** (0.015)
TALKSEX         0.030 ** (0.015)     0.076 *** (0.014)
N                     3127                 3958

                               Females
Exclusion
Restrictions       Age 13-15            Age 16-18

DAYCARE         0.051 *** (0.016)   -0.006 (0.018)
SEXEDFAM        0.027 * (0.016)      0.045 ** (0.019)
BBORN          -0.027 * (0.016)     -0.039 ** (0.019)
SCHATT          0.430 ** (0.191)     0.641 *** (0.185)
SCHGUILT       -0.168 (0.107)       -0.194 * (0.129)
PARSEX         -0.146 *** (0.016)   -0.143 *** (0.015)
TALKSEX         0.057 *** (0.013)    0.080 *** (0.014)
N                     3488                 3825

Source: Computed by the author.

These estimates are from virgin vs. nonvirgin regressions.
The dependent variable in these models is INT.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Appendix D
Tobit Estimates of Unexcused Absences and Robustness
of Female Specifications to Pregnancy

                                   Males                   Females

                          Age 13-15   Age 16-18   Age 13-15   Age 16-18

SKIP (Tobit)               3.56 ***    5.35 ***    5.38 ***    3.39 ***
                          (0.805)     (0.835)     (0.883)     (0.895)
N                           2901        3083        3351        2996
SKIP (OLS for SKIP > 0)    2.13        3.62 ***    3.29 *      1.33
                          (1.34)      (1.17)      (1.75)      (1.05)
N                            545        1245         604        1006
SKIP (preg control)        0.082 ***   0.085 ***   0.108 ***   0.066 **
                          (0.029)     (0.027)     (0.030)     (0.027)
N                           2906        3088        3352        2998
COLLEGE (preg control)    -0.111      -0.156 **   -0.109 *    -0.121 **
                          (0.075)     (0.063)     (0.068)     (0.061)
N                           2905        3084        3352        2996

Source: Computed by the author.
They key independent variable in all above
models is INT. All regressions are weighted.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Thanks to Julie Hotchkiss and two anonymous referees for useful comments and suggestions. I am also grateful to Don Kenkel for helpful comments. This research uses data from Add Health, a program project designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris, and funded by grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due to Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Persons interested in obtaining data files from Add Health should contact Add Health, Carolina Population Center, 123 W Franklin Street, Chapel Hill, NC 27516-2524 (http://www.cpc.unc.edu/addhealth/contract.html).

Received January 2006; accepted August 2006.

References

Allen, J. P., B. J. Leadbeater, and J. L. Aber. 1994. The development of problem behavior syndromes in at-risk adolescents. Development and Psychopathology 6(2):323-42.

Angrist, Joshua, and William Evans. 1996. Children and their parent's labor supply: Evidence from exogenous variation in family size. American Economic Review 88:450-77 (also see NBER Working Paper 5778, September 1996).

Becker, Gary. 1980. A treatise on the family. Cambridge, MA: Harvard University Press.

Billy, J. O. G., N. S. Landale, W. R. Grady, and D. D. Zimmerle. 1988. Effects of sexual activity on adolescent social and psychological development. Social Science Quarterly 51:190-212.

Bronars, S., and J. Grogger. 1994. The economic consequences of unwed motherhood: Using twin births as a natural experiment. American Economic Review 84(5): 1141-56.

Brooke, J. S., E. B. Balka, T. Abernathy, and B. A. Hamburg. 1994. Sequence of sexual behavior and its relationship to other problem behaviors in African American and Puerto Rican adolescents. Journal of Genetic Psychology 1551(1):107-14.

Capaldi, D. M., L. Crosby, and M. Stoolmiller. 1996. Predicting the timing of first sexual intercourse for at-risk adolescent males. Child Development 67(2):344-59.

Chantala, K. 2003. "Introduction to Analyzing Add Health Data." Accessed 2 April 2006. Available http://www.cpc.unc.edu/projects/addhealth/files/analyze.pdf

Costa, F. M., R. Jessor, J. E. Donovan, and J. D. Fortenberry. 1995. Early initiation of sexual intercourse: The influence of psychological unconventionality. Journal of Research on Adolescence 5(1):93-121.

Cragg, J. 1997. Using higher moments to estimate the simple errors-in-variables model. Rand Journal of Economics 28:S71-S91.

Dagenais, M. G., and D. L. Dagenais. 1997. Higher moment estimators for linear regression models with errors in the variables. Journal of Econometrics 76:193-222.

Dawson, Deborah H. 1986. The effects of sex education on adolescent behavior. Family Planning Perspectives 18:162-70.

Donovan, J. E., R. Jessor, and F. M. Costa. 1988. Syndrome of problem behavior in adolescence: A replication. Journal of Consulting and Clinical Psychology 56:762-5.

Elliott, D., and B. J. Morse. 1989. Delinquency and drug use as risk factors in teenage sexual activity. Youth and Society 21:32-60.

Erickson, T., and T. Whited. 2002. Two-step GMM estimation of the errors-in-variables model using higher order moments. Econometric Theory 18:776-99.

Evans, W. N., W. E. Oates, and R. M. Schwab. 1992. Measuring peer group effects: A study of teenage behavior. Journal of Political Economy 100:966-91.

Farrel, A. D., S. J. Danish, and C. W. Howard. 1992. Relationship between drug use and other problem behaviors in urban adolescence. Journal of Consulting and Clinical Psychology 60:705-12.

Feenstra, B. 1994. New product varieties and the measurement of international prices measurement error models. American Economic Review 84:157-77.

Geronimus, A. T., and S. Korenman. 1992. The socioeconomic consequences of teen childbearing reconsidered. Quarterly Journal of Economics 107:1187-1214.

Harvey, S. M., and C. Spigner. 1995. Factors associated with sexual behavior among adolescents: A multivariate analysis. Adolescence 30(118):253-64.

Heckman, J., and E. Vytlacil. 1998. Instrumental variables methods for the correlated random coefficient model. Journal of Human Resources 33:974-87.

Hoffman, S. D., E. M. Foster, and F. F. Furstenberg. 1993. Re-evaluating the costs of teenage childbearing. Demography 30(1):1-13.

Hotz, V. J., S. McElroy, and S. Sanders. 2005. Teenage childbearing and its life cycle consequences: Exploiting a natural experiment. Journal of Human Resources 40(3):683-715.

Hotz, V. J., C. Mullin, and S. Sanders. 1997. Bounding causal effects using data from a contaminated natural experiment: Analyzing the effects of teenage childbearing. Review of Economic Studies 64:576-603.

Jessor, R., F. Costa, L. Jessor, and J. E. Donovan. 1983. Time of first intercourse: A prospective study. Journal of Personality and Social Psychology 44:608-28.

King, M., E. Sentana, and S. Wadhwani. 1994. Volatility and links between national stock markets. Econometrica 62:901-23.

Kirby, Douglas. 2001. Emerging answers: Research findings on programs to reduce teen pregnancy. In The national campaign to prevent teen pregnancy. Washington, DC: The National Campaign to Prevent Teen Pregnancy, pp. 1-19.

Klein, R., and F. Vella. 2003. Identification and estimation of the triangular simultaneous equations model in the absence of exclusion restrictions through the presence of heteroskedasticity. Unpublished paper.

Klepinger, Daniel, Shelley Lundberg, and Robert Plotnick. 1999. How does adolescent fertility affect the human capital and wages of young women? Journal of Human Resources 34:421-48.

Ku, Leighton C., Freya L. Sonenstein, and Joseph H. Pleck. 1992. The association of AIDS education and sex education with sexual behavior and condom use among teenage men. Family Planning Perspectives 24:100-6.

Ku, Leighton C., Freya L. Sonenstein, and Joseph H. Pleck. 1993. Factors influencing first intercourse for teenage men. Public Health Reports 108:680-94.

Leamer, E. 1981. Is it a demand curve or is it a supply curve? Partial identification through inequality constraints. Review of Economics and Statistics 63:319-27.

Lewbel, Arthur. 1997. Constructing instruments for regressions with measurement error when no additional data are available, with an application to patents and R&D. Econometrica 65:1201-13.

Lewbel, Arthur. 2006. Using heteroskedasticity to identify and estimate mismeasured and endogenous regressor models. Boston College Working Papers in Economics 587, Boston College Department of Economics.

Manski, Charles F., and David A. Wise. 1983. College choice in America. Cambridge, MA: Harvard University Press.

Marsiglio, William, and Frank L. Mott. 1986. The impact of sex education on sexual activity, contraceptive use, and premarital pregnancy among American teenagers. Family Planning Perspectives 18:151-62.

McLean, A. L., and B. J. Flanigan. 1993. Transition marking behaviors of adolescent males at first intercourse. Adolescence 28:579-95.

Meilman, P. W. 1993. Alcohol induced sexual behavior on campus. Journal of American College Health 42:27-31.

Miller, B. C., and K. R. Sneesby. 1988. Educational correlates of adolescents' sexual attitudes and behavior. Journal of Youth and Adolescence 17:521-30.

Mott, F. L., and W. Marsiglio. 1985. Early childbearing and completion of high school. Family Planning Perspectives 17:234-7.

Nye, F. J. 1979. Choice, exchange and the family. In Contemporary theories about the family 2, edited by W. R. Burr, R. Hill, I. F. Nye, and I. L. Reiss. New York: The Free Press, pp. 1-41.

Oettinger, Gerald S. 1999. The effects of sex education on teen sexual activity and teen pregnancy. Journal of Political Economy 107(3):606-44.

Peterson, J. L., K. A. Moore, and F. F. Furstenberg. 1991. Television viewing and early initiation of sexual intercourse: Is there a link? Journal of Homosexuality 21:93-118.

Rector, Robert E., and Kirk A. Johnson. 2005. Teenage sexual abstinence and academic achievement. Conference paper. Accessed 17 April 2006. Available http://www.heritage.org/Research/Welfare/whitepaper10272005-1.cfm

Rector, Robert E., Kirk A. Johnson, and Lauren R. Noyes. 2003. Sexually active teenagers are more likely to be depressed and to attempt suicide. A report of the Heritage Center for Data Analysis. Washington, DC: Heritage Foundation.

Rigobon, R. 2002. The curse of non-investment grade countries. Journal of Development Economics 69:423-49.

Rigobon, R. 2003. Identification through heteroskedasticity. Review of Economics and Statistics 85:777-92.

Rosenbaum, E., and D. B. Kandel. 1990. Early onset of adolescent sexual behavior and drug involvement. Journal of Marriage and the Family 52:783-98.

Rosenzweig, M., and K. Wolpin. 1995. Sisters, siblings and mothers: The effects of teen-age childbearing on birth outcomes in a dynamic context. Econometrica 63(2):303-26.

Sabia, Joseph J. 2006a. Does early adolescent sex affect depressive symptoms? Journal of Policy Analysis and Management 25(4):803-25.

Sabia, Joseph J. 2006b. Does sex education affect adolescent sexual behaviors and health? Journal of Policy Analysis and Management 25(4):783-802.

Sabia, Joseph J. Forthcoming. Reading, writing, and sex: The effect of losing virginity on adolescent academic performance. Economic Inquiry.

Schvaneveldt, Paul L., Brent C. Miller, E. Helen Berry, and Thomas R. Lee. 2001. Academic goals, achievement, and age at first sexual intercourse: Longitudinal, bidirectional influences. Adolescence 36(144):767-87.

Sentana, E., and G. Fiorentini. 2001. Identification, estimation, and testing of conditional heteroskedastic factor models. Journal of Econometrics 102:143-64.

Stock, J. H., and M. Yogo. 2005. Testing for weak instruments in IV regression. In Identification and inference for econometric models: A festschrift in honor of Thomas Rothenberg. Cambridge, MA: Harvard University Press.

Teachman, J. 1987. Family background, educational resources, and educational attainment. American Sociological Review 52:548-57.

Thornton, A., and D. Camburn. 1989. Religious participation and adolescent sexual behavior and attitudes. Journal of Marriage and the Family 51:641-53.

Vella, F., and M. Verbeek. 1997. Order as an instrumental variable: An application to the return to schooling. Unpublished paper.

Whitbeck, L. B., R. D. Conger, R. L. Simons, and M. Kao. 1993. Minor deviant behaviors and adolescent sexual activity. Youth and Society 25:24-37.

(1) There is a fair amount of empirical literature on this question, focusing on the impact of different types of sex education programs on abstinence, unprotected sex, pregnancy, and sexually transmitted diseases (see, for example, Dawson 1986; Marsiglio and Mott 1986; Ku, Sonenstein, and Pleck 1992, 1993; Oettinger 1999; Kirby 2001; Sabia 2006b).

(2) This theory also assumes that there is positive assortative mating among adolescents. That is, males and females with the lowest opportunity costs of sex or that have similar lifestyle values will choose to mate together.

(3) Engaging in sexual intercourse could increase competition among sexually active adolescents, causing greater conflicts in school and less devotion and concentration on schooling pursuits.

(4) In fact, Sabia (2006a) shows that the findings by Rector, Johnson, and Noyes (2003) can be explained by fixed individual-level unobserved heterogeneity. However, the mental health measures used in each study are self-reported and it may be that early teen sex may adversely affect unobserved measures of mental health.

(5) Note that by definition, the sample used for this paper includes only those who are enrolled in school. Thus, this is a selected sample of nondropouts. Thus, I cannot estimate the relationship between sexual intercourse and the schoolrelated behaviors of those who drop out of school.

(6) Out-of-school suspension is also a common punishment (67% of schools in the Add Health survey) for fighting with another student. As mentioned above, physical confrontation--particularly among males might be an outcome associated with exiting virginity. If teen sex creates physical confrontations among adolescent males who are competing with other males for sexual conquests, the suspension measure may capture this. Hence, while imperfect, the measures of school attachment I use are likely to adequately capture the outcomes I endeavor to measure.

(7) In wave 1, the adolescent was asked whether he or she has ever received an out-of-school suspension. Using information on the grade in which the student is enrolled and information on when the student received his or her last suspension, I am able to construct a variable on whether the adolescent was suspended in the current school year. In wave 2, the question is directly asked about suspensions in the current school year, so such construction is not necessary.

(8) Across the entire sample, 70% reported no unexcused absences, 21% reported one to five unexcused absences, and 9% reported more than five unexcused absences.

(9) Of the adolescents asked, 26% reported they strongly agreed, 48% stated they agreed, 15% said they neither agreed nor disagreed, 9% said they disagreed, and 4% said they strongly disagreed.

(10) Of the adolescents asked, 3% reported "l," 3% reported "2," 9% reported "3," 13% reported "4," and 72% reported "5."

(11) Several studies have shown that parental education level is a strong predictor of their children's academic achievement (Teachman 1987; Miller and Sneesby 1988; Thornton and Camburn 1989; Schvaneveldt et al. 2001; Sabia, forthcoming). Thus, it may also be the case that parental schooling is associated with their children's attachment to school. I create an indicator variable named COLGRAD, set equal to zero if the mother has not graduated college and set equal to one if she has. Approximately 25% of mothers in the sample reported they had graduated from college. It also may be the case that other proschooling parental attitudes and actions are associated with their children's attachment to school. Over 50% of parents reported that their family moved to their current neighborhood because of the local school system (NEIGHBOR). One might expect the children of parents who made the decision to move to a neighborhood because of school quality to have greater attachment to their schools. Table 1 also presents a measure of whether the parent was a member of the PTA, since this reflects greater household-level attachment to school.

(12) This equation can be interpreted as a production function for academic performance.

(13) This equation can be interpreted as the adolescent's reduced form demand for sexual activity.

(14) In the Add Health school administrator questionnaire, administrators are asked whether their schools provide or refer students for family planning services and whether there are day care services for students provided in the district.

(15) Additional state- or school-level exclusion restrictions were attempted, including sexually transmitted disease prevalence, state abortion laws, and the presence of family planning clinics in the geographic area, but none were strong predictors of virginity.

(16) The endogeneity of school-level peer effects has been discussed at length in the literature (see, for example, Evans, Oates, and Schwab 1992).

(17) Time-invariant controls from the OLS estimates are excluded from the IFE models. Some of these time-invariant controls may change over time but are not available in the data. In particular, survey items asked of parents are not repeated in wave 2. In addition to the noted time-varying covariates, IFE models also include variables capturing school changes and the onset of puberty. Regressions presented in Tables 2-6 are weighted using Add Health's clustered survey design effect (Chantala 2003).

(18) Estimated parameters on control variables of the suspension equation are found in Appendix A. Estimates on control variables for all models are available upon request of the author. Linear probability models produce results similar to unreported probit and logit models.

(19) Robustness tests are conducted to examine whether Lewbel IV estimates are sensitive to the choice of variables included in Z. Parameter estimates were not especially sensitive to choice of variables included in Z, but when, for example, the community characteristics were omitted, estimates were less precise. Moreover, given that the Lewbel IV approach requires heteroskedasticity in the intercourse equation for identification, all remaining models presented are unweighted.

(20) Also presented in Appendix D are OLS estimates of the relationship between sexual activity and the number of unexcused absences among those adolescents who reported positive unexcused absences. IV Tobit models are not presented, but produce similar results. However, the likelihood function for Lewbel IV Tobit models did not converge.

(21) However, Sabia (forthcoming) did find that for adolescent males, entrance into sexual intercourse was associated with a 0.15 point reduction in GPA. If this effect were permanent, a point reduction of this magnitude could marginally affect the quality of college to which a high school student could gain admission (Manski and Wise 1983).

(22) Note that teenage pregnancy cannot explain these results. Appendix D presents estimates on unexcused absences and college aspirations that control for a reported pregnancy. Estimates do not change.

Joseph J. Sabia, Department of Housing and Consumer Economics, 204 Consumer Research Center, House C, University of Georgia, Athens, GA 30602, USA; E-mail [e-mail].

Table 1. Definitions and Weighted Means of Variables

Dependent Variables
  SUSPEND (a)           Received an out-of-school suspension in
                          current school year
  SKIP (a)              Had at least one unexcused absence from school
  PARSCH (a)            Adolescent feels part of school (= 1 strongly
                          agree ... = 5 strongly disagree)
  COLLEGE (a)           Adolescent aspires to attend college (= 1 low
                          preference ... = 5 high preference)
Independent Variables
  INT (a)               = 1 if ever had intercourse; = 0 if never had
                          intercourse
  RINT (a,b)            = 1 if had protected sex recently or nonrecent
                          sex; = 0 if never had intercourse
  DRUG (a)              Consumed any illegal drugs in last 12 months
  DRINK (a)             Had five or more drinks in one sitting at
                          least once last year
  PREFCOL (a,c)         Strongest preference for attending college
  ROMANTIC (a)          Adolescent in a romantic relationship
  AFDC                  Family received Aid to Families with Dependent
                          Children/Temporary Assistance for Needy
                          Families benefits
  BMI (a)               Body mass index in m/[kg.sup.2]
  WORK (a)              Adolescent worked during academic year
  INCOME                Household income (log)
  SINGLPAR              Single parent household
  BLACK                 Adolescent is Black
  HISPANIC              Adolescent is Hispanic
  AHPVT                 Add Health Picture and Vocabulary Test Score
  DEPRESS (a)           Adolescent reports depression
  BIRTHWGT              Adolescent birth weight (in 000s grams)
  BADHEALTH (a)         Adolescent reports being in bad health
  FIGHT                 Parent reports fighting between spouses in
                          the household
  NOTGETAL              Parent reports severe relationship problems
                          with adolescent
  OLDERSIB              Adolescent has older sibling
  FRIENDSUICIDE (a)     Adolescent reports friend made a suicide
                          attempt in last year
  FAMSUICIDE (a)        Adolescent reports family member attempted
                          suicide in last year
  NEIGHBOR              Parent moved to neighborhood because of school
                          quality
  PTA                   Parent is member of Parent Teacher Association
  WORKOUT               Mother works outside of the home
  BRILLIANT             Parent views scholastic brilliance is
                          important quality for child
  PARHELP (a)           Parent recently helped child with recent
                          schoolwork
  WEEKENDTM             Parent sets strict curfew time on weekends
  DINNERWK (a)          Number of times per week adolescent eats
                          dinner with family
  PARDISCOL (a)         Parent strongly disapproved of child not
                          attending college
  AGE13                 Age 13
  AGE14                 Age 14
  AGE17                 Age 17
  AGE18                 Age 18
  RELIGWK (a)           Attends religious services at least once per
                          week
  RELIGMO (a)           Attends religious services once per month
  RELIGYR (a)           Attends religious services once per year
  COLLGRAD              Parent is college graduate
  NOTCARE               Adolescent believes parent does not care about
                          him/her
  MIDWEST               Lives in midwest region
  WEST                  Lives in western region
  SOUTH                 Lives in southern region
  SUSVERB               School has suspension policy if student
                          verbally attacks teachers
  RURAL                 Lives in rural region
  SUBURB                Lives in suburban region
  PUBLIC (a)            Adolescent attends public school
  N

                                      Males

                          Age 13-15        Age 16 18

Dependent Variables
  SUSPEND (a)           0.056 (0.229)    0.111 (0.313)
  SKIP (a)              0.173 (0.379)    0.326 (0.469)
  PARSCH (a)             2.06 (0.964)      2.16 (1.00)

  COLLEGE (a)            4.53 (0.941)      4.39 (1.05)

Independent Variables
  INT (a)               0.165 (0.371)    0.428 (0.495)
  RINT (a,b)            0.102 (0.303)    0.322 (0.467)

  DRUG (a)              0.214 (0.410)    0.364 (0.481)
  DRINK (a)             0.155 (0.362)    0.379 (0.485)
  PREFCOL (a,c)         0.740 (0.439)    0.666 (0.472)
  ROMANTIC (a)          0.422 (0.494)    0.589 (0.492)
  AFDC                  0.112 (0.315)    0.100 (0.301)

  BMI (a)                 21.9 (4.45)      23.4 (4.41)
  WORK (a)              0.465 (0.499)    0.614 (0.487)
  INCOME                 3.21 (0.870)     3.61 (0.891)
  SINGLPAR              0.313 (0.464)    0.343 (0.475)
  BLACK                 0.121 (0.327)     0.16 (0.366)
  HISPANIC              0.104 (0.306)    0.114 (0.317)
  AHPVT                  103.2 (14.0)     103.4 (13.9)
  DEPRESS (a)           0.036 (0.185)    0.066 (0.248)
  BIRTHWGT               4.77 (0.208)     4.78 (0.193)
  BADHEALTH (a)         0.054 (0.226)    0.049 (0.217)
  FIGHT                 0.223 (0.416)     0.21 (0.407)
  NOTGETAL              0.007 (0.083)    0.009 (0.096)
  OLDERSIB              0.412 (0.492)    0.359 (0.480)
  FRIENDSUICIDE (a)     0.121 (0.327)    0.133 (0.340)
  FAMSUICIDE (a)        0.039 (0.193)    0.037 (0.188)

  NEIGHBOR              0.525 (0.499)    0.530 (0.499)
  PTA                   0.356 (0.479)    0.351 (0.477)
  WORKOUT               0.737 (0.440)    0.747 (0.435)
  BRILLIANT             0.657 (0.475)    0.647 (0.478)

  PARHELP (a)           0.108 (0.310)    0.057 (0.233)
  WEEKENDTM             0.288 (0.453)    0.418 (0.493)
  DINNERWK (a)            5.44 (2.12)      4.38 (2.47)
  PARDISCOL (a)         0.662 (0.473)    0.653 (0.476)
  AGE13                 0.305 (0.460)               --
  AGE14                 0.354 (0.478)               --
  AGE17                            --    0.374 (0.484)
  AGE18                            --    0.114 (0.318)
  RELIGWK (a)           0.400 (0.490)    0.332 (0.411)
  RELIGMO (a)           0.209 (0.407)    0.215 (0.406)
  RELIGYR (a)           0.155 (0.362)    0.209 (0.421)
  COLLGRAD              0.219 (0.413)    0.230 (0.421)
  NOTCARE               0.012 (0.108)    0.012 (0.107)
  MIDWEST               0.324 (0.468)    0.338 (0.473)
  WEST                  0.154 (0.361)    0.155 (0.362)
  SOUTH                 0.379 (0.485)    0.352 (0.478)
  SUSVERB               0.787 (0.409)    0.798 (0.401)

  RURAL                 0.147 (0.354)    0.172 (0.378)
  SUBURB                0.604 (0.489)    0.600 (0.491)
  PUBLIC (a)            0.924 (0.265)    0.918 (0.274)
  N                          2530             2301

                                      Females

                          Age 13-15        Age 16 18

Dependent Variables
  SUSPEND (a)           0.034 (0.182)    0.056 (0.229)
  SKIP (a)              0.163 (0.370)    0.278 (0.483)
  PARSCH (a)             2.01 (0.949)     2.20 (0.998)

  COLLEGE (a)             4.6 (0.848)     4.59 (0.876)

Independent Variables
  INT (a)               0.156 (0.363)    0.438 (0.496)
  RINT (a,b)            0.092 (0.289)    0.323 (0.460)

  DRUG (a)              0.211 (0.408)    0.316 (0.465)
  DRINK (a)             0.166 (0.372)    0.305 (0.461)
  PREFCOL (a,c)         0.765 (0.424)    0.776 (0.417)
  ROMANTIC (a)          0.457 (0.498)    0.663 (0.473)
  AFDC                  0.144 (0.351)    0.111 (0.315)

  BMI (a)                 21.7 (4.29)      22.4 (4.37)
  WORK (a)              0.447 (0.497)    0.578 (0.494)
  INCOME                 3.52 (0.868)     3.57 (0.880)
  SINGLPAR              0.336 (0.472)    0.352 (0.478)
  BLACK                 0.143 (0.350)     0.17 (0.375)
  HISPANIC              0.113 (0.316)    0.106 (0.308)
  AHPVT                  101.4 (14.0)     101.6 (14.4)
  DEPRESS (a)           0.089 (0.285)    0.105 (0.307)
  BIRTHWGT               4.73 (0.191)     4.74 (0.205)
  BADHEALTH (a)         0.066 (0.249)    0.077 (0.267)
  FIGHT                 0.198 (0.399)    0.191 (0.393)
  NOTGETAL              0.009 (0.094)    0.011 (0.105)
  OLDERSIB               0.44 (0.497)    0.386 (0.487)
  FRIENDSUICIDE (a)     0.256 (0.437)    0.207 (0.406)
  FAMSUICIDE (a)        0.052 (0.222)    0.045 (0.208)

  NEIGHBOR              0.506 (0.500)    0.535 (0.499)
  PTA                   0.358 (0.479)    0.341 (0.474)
  WORKOUT               0.742 (0.438)    0.739 (0.439)
  BRILLIANT             0.689 (0.463)    0.646 (0.478)

  PARHELP (a)           0.103 (0.305)    0.074 (0.262)
  WEEKENDTM             0.201 (0.401)    0.316 (0.465)
  DINNERWK (a)            5.30 (2.23)      4.16 (2.51)
  PARDISCOL (a)         0.666 (0.472)    0.695 (0.461)
  AGE13                 0.306 (0.461)               --
  AGE14                 0.341 (0.474)               --
  AGE17                            --    0.353 (0.478)
  AGE18                            --     0.09 (0.286)
  RELIGWK (a)           0.454 (0.498)    0.391 (0.488)
  RELIGMO (a)           0.192 (0.394)    0.188 (0.391)
  RELIGYR (a)           0.155 (0.362)    0.203 (0.403)
  COLLGRAD              0.206 (0.404)    0.205 (0.404)
  NOTCARE               0.026 (0.159)    0.033 (0.180)
  MIDWEST               0.330 (0.470)    0.348 (0.477)
  WEST                  0.164 (0.370)    0.159 (0.366)
  SOUTH                 0.379 (0.485)    0.361 (0.481)
  SUSVERB               0.787 (0.409)    0.818 (0.386)

  RURAL                 0.166 (0.372)    0.181 (0.385)
  SUBURB                0.583 (0.493)    0.573 (0.495)
  PUBLIC (a)            0.933 (0.250)    0.934 (0.249)
  N                          3008             2304

Source: Computed by the author.
Standard deviations of variables appear in parentheses.

(a) Time-varying variables available in both waves of data.

(b) The construction of this variable implies a smaller sample size
since those who have engaged in unprotected recent sex are excluded.
Sample sizes for this variable are noted in Table 2.

(c) This variable is not included in college aspirations models since
it is derived from it.

Table 2. Fixed Effects Estimates of Relationship between
Adolescent Sexual Activity and the Probability of Suspension

                                    Males

                                  Age 13-15

                     OLS (1)       SFE (2)       IFE (3)

Nonvirgin vs.       0.042 **      0.039 **      0.177 ***
  Virgin           (0.018)       (0.018)       (0.036)
  N                   2530          2530          2530
Safe Current vs.    0.030 *       0.031         0.172 ***
  Virgin           (0.019)       (0.022)       (0.048)
  N                   2246          2246          2246
Safe vs. "Super"    0.038 *       0.041         0.204 ***
  Virgin           (0.024)       (0.027)       (0.069)
  N                   1886          1886          1886

                                    Males

                                  Age 16-18

                     OLS (4)       SFE (5)       IFE (6)

Nonvirgin vs.       0.088 ***     0.094 ***     0.056 *
  Virgin           (0.020)       (0.020)       (0.033)
  N                   2301          2301          2301
Safe Current vs.    0.077 ***     0.083 ***     0.044
  Virgin           (0.024)       (0.025)       (0.034)
  N                   1797          1797          1797
Safe vs. "Super"    0.047 *       0.049         0.008
  Virgin           (0.028)       (0.030)       (0.045)
  N                   1452          1452          1452

                                   Females

                                  Age 13-15

                     OLS (1)       SFE (2)       IFE (3)

Nonvirgin vs.       0.031 *       0.032 *       0.025
  Virgin           (0.016)       (0.017)       (0.026)
  N                   3008          3008          3008
Safe Current vs.    0.057 **      0.061 **      0.070 **
  Virgin           (0.024)       (0.024)       (0.031)
  N                   2609          2609          2609
Safe vs. "Super"    0.048 **      0.049 **      0.068 **
  Virgin           (0.023)       (0.023)       (0.033)
  N                   2137          2137          2137

                                   Females

                                  Age 16-18

                     OLS (4)       SFE (5)       IFE (6)

Nonvirgin vs.       0.044 ***     0.042 ***     0.068 ***
  Virgin           (0.011)       (0.011)       (0.025)
  N                   2304          2304          2304
Safe Current vs.    0.043 **      0.041 **      0.086 ***
  Virgin           (0.017)       (0.016)       (0.027)
  N                   1788          1788          1788
Safe vs. "Super"    0.036 *       0.024         0.110 ***
  Virgin           (0.020)       (0.020)       (0.039)
  N                   1445          1445          1445

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and SEE models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. IFE models include only observable characteristics
that change between data waves, including whether the adolescent
changed schools. These variables are denoted in Table 1.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 3. Fixed Effects Estimates of Relationship between Adolescent
Sexual Activity and the Probability of Unexcused Absences

                               Males

                             Age 13-15

                 OLS (7)      SFE (8)      IFE (9)

Nonvirgin vs.    0.128 ***    0.112 ***    0.037
  Virgin        (0.028)      (0.028)      (0.038)
  N                2583         2583         2583
Safe Current     0.131 ***    0.114 ***    0.064
  vs. Virgin    (0.041)      (0.040)      (0.045)
  N                2227         2227         2227
Safe vs.         0.093 **     0.071 *      0.057
  "Super"       (0.040)      (0.041)      (0.055)
  Virgin
  N                1914         1914         1914

                               Males

                             Age 16-18

                 OLS (10)     SFE (11)     IFE (12)

Nonvirgin vs.    0.099 ***    0.083 **     0.010
  Virgin        (0.033)      (0.033)      (0.042)
  N                2601         2601         2601
Safe Current     0.064 *      0.039       -0.024
  vs. Virgin    (0.035)      (0.036)      (0.047)
  N                1989         1989         1989
Safe vs.         0.112 ***    0.096 **     0.037
  "Super"       (0.041)      (0.043)      (0.080)
  Virgin
  N                1620         1620         1620

                              Females

                             Age 13-15

                 OILS (7)     SFE (8)      IFE (9)

Nonvirgin vs.    0.107 ***    0.104 ***    0.008
  Virgin        (0.025)      (0.026)      (0.036)
  N                3067         3067         3067
Safe Current     0.089 **     0.074 **     0.001
  vs. Virgin    (0.038)      (0.038)      (0.040)
  N                2638         2638         2638
Safe vs.         0.052        0.021       -0.023
  "Super"       (0.037)      (0.036)      (0.056)
  Virgin
  N                2158         2158         2158

                              Females

                             Age 16-18

                 OLS (10)     SFE (11)     IFE (12)

Nonvirgin vs.    0.080 ***    0.059 *     -0.018
  Virgin        (0.030)      (0.031)      (0.044)
  N                2555         2555         2555
Safe Current     0.094 ***    0.075 **     0.033
  vs. Virgin    (0.029)      (0.030)      (0.037)
  N                1933         1933         1933
Safe vs.         0.079 **     0.078 **     0.097 *
  "Super"       (0.034)      (0.033)      (0.053)
  Virgin
  N                1570         1570         1570

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and SFE models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. IFE models include only observable characteristics
that change between data waves, including whether the adolescent
changed schools. These variables are denoted in Table 1.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 4. Fixed Effects Estimates of Relationship between
Adolescent Sexual Activity and School Affinity

                                  Males

                                Age 13-15

                    OLS (1)      SFE (2)      IFE (3)

Nonvirgin vs.       0.141 *      0.144 *      0.039
  Virgin           (0.086)      (0.089)      (0.083)
  N                   2529         2529         2529
Safe Current vs.    0.161 *      0.152 *      0.230 **
  Virgin           (0.091)      (0.094)      (0.090)
  N                   2246         2246         2246
Safe vs.            0.113        0.089        0.081
  "Super"          (0.081)      (0.088)      (0.087)
  Virgin
  N                   1886         1886         1886

                                  Males

                                Age 16-18

                    OLS (4)      SFE (5)      IFE (6)

Nonvirgin vs.       0.093        0.121        0.247 ***
  Virgin           (0.071)      (0.075)      (0.094)
  N                   2299         2299         2299
Safe Current vs.    0.060        0.084        0.276 **
  Virgin           (0.084)      (0.089)      (0.112)
  N                   1796         1796         1796
Safe vs.           -0.041       -0.016        0.156
  "Super"          (0.089)      (0.094)      (0.121)
  Virgin
  N                   1452         1452         1452

                                 Females

                                Age 13-15

                    OLS (I)      SFE (2)      IFE (3)

Nonvirgin vs.       0.098        0.158 **     0.020
  Virgin           (0.066)      (0.072)      (0.078)
  N                   3007         3007         3007
Safe Current vs.    0.025        0.073        0.049
  Virgin           (0.081)      (0.085)      (0.078)
  N                   2609         2609         2609
Safe vs.           -0.006        0.060        0.003
  "Super"          (0.083)      (0.094)      (0.097)
  Virgin
  N                   2137         2137         2137

                                 Females

                                Age 16-18

                    OLS (4)      SFE (5)      IFE (6)

Nonvirgin vs.       0.015        0.012       -0.054
  Virgin           (0.076)      (0.080)      (0.096)
  N                   2305         2305         2305
Safe Current vs.   -0.069       -0.107       -0.136
  Virgin           (0.079)      (0.082)      (0.096)
  N                   1788         1788         1788
Safe vs.           -0.023       -0.081       -0.089
  "Super"          (0.088)      (0.095)      (0.114)
  Virgin
  N                   1445         1445         1445

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and SEE models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. IFE models include only observable characteristics
that change between data waves, including whether the adolescent
changed schools. These variables are denoted in Table 1.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 5. Fixed Effects Estimates of Relationship between
Adolescent Sexual Activity and College Aspirations

                              Males

                            Age 13-15

                OLS (7)      SFE (8)      IFE (9)

Nonvirgin      -0.130 *     -0.136 *     -0.100
  vs. Virgin   (0.078)      (0.080)      (0.077)
  N               2581         2581         2581
Safe Current   -0.145       -0.163 *     -0.078
  vs. Virgin   (0.101)      (0.100)      (0.091)
  N               2275         2275         2275
Safe vs.       -0.182 *     -0.185 *     -0.094
  "Super"      (0.100)      (0.103)      (0.106)
  Virgin
  N               1978         1978         1978

                              Males

                            Age 16-18

                OLS (10)     SFE (11)     IFE (12)

Nonvirgin      -0.101 *     -0.083       -0.112
  vs. Virgin   (0.057)      (0.061)      (0.079)
  N               2636         2636         2636
Safe Current   -0.045       -0.032       -0.122
  vs. Virgin   (0.065)      (0.070)      (0.078)
  N               1894         1894         1894
Safe vs.        0.017        0.011       -0.166 *
  "Super"      (0.060)      (0.063)      (0.088)
  Virgin
  N               1546         1546         1546

                             Females

                            Age 13-15

                OLS (7)      SEE (8)      IFE (9)

Nonvirgin      -0.178 ***   -0.173 ***   -0.228 ***
  vs. Virgin   (0.066)      (0.066)      (0.064)
  N               3064         3064         3064
Safe Current   -0.104       -0.098       -0.291 ***
  vs. Virgin   (0.102)      (0.105)      (0.074)
  N               2635         2635         2635
Safe vs.       -0.151       -0.136       -0.321 ***
  "Super"      (0.100)      (0.096)      (0.103)
  Virgin
  N               2155         2155         2155

                             Females

                            Age 16-18

                OLS (10)     SEE (11)     IFE (12)

Nonvirgin      -0.066       -0.048        0.008
  vs. Virgin   (0.079)      (0.080)      (0.062)
  N               2379         2379         2379
Safe Current   -0.004       -0.004        0.019
  vs. Virgin   (0.084)      (0.090)      (0.037)
  N               1801         1801         1801
Safe vs.        0.079 **     0.078 **     0.007
  "Super"      (0.034)      (0.033)      (0.096)
  Virgin
  N               1471         1471         1471

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and SFE models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. IFE models include only observable characteristics
that change between data waves, including whether the
adolescent changed schools. These variables are denoted in Table 1.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 6. IV Estimates of Relationship between Adolescent
Sexual Activity and the Probability of Suspension

                                        Males

                                      Age 13-15

                                     Standard IV   Lewbel IV
                         OLS (1)         (2)          (3)

Nonvirgin vs.            0.036 ***    0.144 **      -0.005
  Virgin                 (0.011)       (0.062)      (0.033)
F stat on                   --          15.2         70.0
  instruments
First-stage [R.sup.2]       --          0.34         0.42
Sargan Overid p-            --          0.63         0.70
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          3127         3217         3127
Safe Current vs.         0.024 **     0.140 ***     -0.010
  Virgin                 (0.012)       (0.069)      (0.024)
F stat on                   --          13.3         118.9
  instruments
First-stage [R.sup.2]       --          0.31         0.49
Sargan Overid p-            --          0.57         0.93
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          2917         2917         2917
Safe vs. "Super"         0.021 *      0.147 **       0.007
  Virgin                 (0.013)       (0.067)      (0.026)
F stat on                   --          14.4         105.5
  instruments
First-stage [R.sup.2]       --          0.36         0.52
Sargan Overid               --          0.62         0.88
  p-value
Males
Breusch-Pagan               --           --          0.00
  p-value
N                          2624         2624         2624

                                        Males

                                      Age 16-18

                                     Standard IV   Lewbel IV
                         OLS (4)         (5)          (6)

Nonvirgin vs.            0.095 ***    0.256 ***     -0.009
  Virgin                 (0.012)       (0.076)      (0.222)
F stat on                   --          15.6          1.8
  instruments
First-stage [R.sup.2]       --          0.33         0.33
Sargan Overid p-            --          0.22         0.78
  value
Breusch-Pagan p-            --           --          0.71
  value
N                          3958         3958         3958
Safe Current vs.        0.089 ***     0.262 ***     0.220 *
  Virgin                 (0.012)       (0.074)      (0.132)
F stat on                   --          14.7          4.4
  instruments
First-stage [R.sup.2]       --          0.32         0.32
Sargan Overid p-            --          0.21         0.79
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          3359         3359         3359
Safe vs. "Super"        0.097 ***     0.369 ***      0.172
  Virgin                 (0.014)       (0.080)      (0.231)
F stat on                   --          14.8          1.4
  instruments
First-stage [R.sup.2]       --          0.40         0.40
Sargan Overid               --          0.35         0.62
  p-value
Breusch-Pagan               --           --          0.02
  p-value
N                          2974         2974         2974

                                       Females

                                      Age 13-15

                                     Standard IV   Lewbel IV
                         OLS (1)         (2)          (3)

Nonvirgin vs.            0.020 **      0.105 **     -0.005
  Virgin                 (0.008)       (0.044)      (0.027)
F stat on                   --          18.7         46.0
  instruments
First-stage [R.sup.2]       --          0.31         0.38
Sargan Overid p-            --          0.06         0.89
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          3488         3488         3488
Safe Current vs.          0.014       0.092 **      -0.017
  Virgin                 (0.009)       (0.046)      (0.023)
F stat on                   --          20.1         80.7
  instruments
First-stage [R.sup.2]       --          0.27         0.40
Sargan Overid p-            --          0.25         0.67
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          3221         3221         3221
Safe vs. "Super"          0.006        0.083 *      -0.005
  Virgin                 (0.010)       (0.044)      (0.027)
F stat on                   --          21.7         74.8
  instruments
First-stage [R.sup.2]       --          0.31         0.43
Sargan Overid               --          0.14         0.35
  p-value
Breusch-Pagan               --           --          0.00
  p-value
N                          2911         2911         2911

                                       Females

                                      Age 16-18

                                     Standard IV   Lewbel IV
                         OLS (4)         (5)          (6)

Nonvirgin vs.             0.017 *     0.123 ***      0.001
  Virgin                 (0.009)       (0.047)      (0.134)
F stat on                   --          22.9          2.4
  instruments
First-stage [R.sup.2]       --          0.39         0.39
Sargan Overid p-            --          0.55         0.29
  value
Breusch-Pagan p-            --           --          0.89
  value
N                          3825         3825         3825
Safe Current vs.          0.012       0.092 **       0.046
  Virgin                 (0.010)       (0.043)      (0.094)
F stat on                   --          23.1          4.2
  instruments
First-stage [R.sup.2]       --          0.39         0.40
Sargan Overid p-            --          0.63         0.49
  value
Breusch-Pagan p-            --           --          0.00
  value
N                          3167         3167         3167
Safe vs. "Super"          0.008       0.099 **      0.263 *
  Virgin                 (0.011)       (0.047)      (0.152)
F stat on                   --          22.2          2.1
  instruments
First-stage [R.sup.2]       --          0.46         0.46
Sargan Overid               --          0.58         0.22
  p-value
Breusch-Pagan               --           --          0.00
  p-value
N                          2800         2800         2800

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and IV models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. For standard IV models, exclusion restrictions
include parental disapproval of adolescent sex, school-level
measures of strong guilt and attractiveness associated with
sex, state sex education mandates and family planning services,
parental discussions with adolescent about sex, day care services
provided in the public school, and whether the adolescent believes
the Bible is the literal word of God and considers himself to be a
born-again Christian.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 7. IV Estimates of Relationship between Adolescent
Sexual Activity and the Probability of Unexcused Absences

                                        Males

                                      Age 13-15

                                      Standard     Lewbel
                           OLS (7)     IV (8)      IV (9)

Nonvirgin vs. Virgin      0.105 ***   0.435 ***   0.245 ***
                           (0.018)     (0.102)     (0.052)
F stat on instruments        --         15.2        70.0
First-stage [R.sup.2]        --         0.34        0.42
Sargan Overid p-value        --         0.63        0.72
Breusch-Pagan p-value        --          --         0.00
N                           3127        3217        3127
Safe Current vs. Virgin   0.112 ***   0.470 ***   0.167 ***
                           (0.020)     (0.110)     (0.040)

F stat on instruments        --         13.3        118.9
First-stage [R.sup.2]        --         0.31        0.49
Sargan Overid p-value        --         0.25        0.34
Breusch-Pagan p-value        --          --         0.00
N                           2917        2917        2917
Safe vs. "Super" Virgin   0.112 ***   0.462 ***   0.153 ***
                           (0.021)     (0.111)     (0.018)

F stat on instruments        --         14.4        105.5
First-stage [R.sup.2]        --         0.36        0.52
Sargan Overid p-value        --         0.21        0.11
Breusch-Pagan p-value        --          --         0.00
N                           2622        2622        2622

                                        Males

                                      Age 16-18

                                      Standard     Lewbel
                          OLS (10)     IV (11)     IV (12)

Nonvirgin vs. Virgin      0.114 ***   0.313 ***    0.475 *
                           (0.017)     (0.102)     (0.310)
F stat on instruments        --         15.6         1.8
First-stage [R.sup.2]        --         0.33        0.33
Sargan Overid p-value        --         0.03        0.05
Breusch-Pagan p-value        --          --         0.71
N                           3958        3958        3958
Safe Current vs. Virgin   0.109 ***   0.313 ***     0.142
                           (0.018)     (0.105)     (0.187)

F stat on instruments        --         14.7         4.4
First-stage [R.sup.2]        --         0.32        0.32
Sargan Overid p-value        --         0.00        0.06
Breusch-Pagan p-value        --           -         0.00
N                           3359        3359        3359
Safe vs. "Super" Virgin   0.112 ***   0.261 ***     0.067
                           (0.020)     (0.109)     (0.327)

F stat on instruments        --         14.8         1.4
First-stage [R.sup.2]        --         0.40        0.40
Sargan Overid p-value        --         0.00        0.01
Breusch-Pagan p-value        --          --         0.02
N                           2974        2974        2974

                                       Females

                                      Age 13-15

                                      Standard     Lewbel
                           OLS (7)     IV (8)      IV (9)

Nonvirgin vs. Virgin      0.122 ***   0.397 ***    0.107 *
                           (0.017)     (0.093)     (0.056)
F stat on instruments        --         18.7        46.0
First-stage [R.sup.2]        --         0.31        0.38
Sargan Overid p-value        --         0.16        0.02
Breusch-Pagan p-value        --          --         0.00
N                           3488        3488        3488
Safe Current vs. Virgin   0.064 ***   0.461 ***   0.102 **
                           (0.020)     (0.100)     (0.048)

F stat on instruments        --         20.1        80.7
First-stage [R.sup.2]        --         0.27        0.40
Sargan Overid p-value        --         0.14        0.03
Breusch-Pagan p-value        --          --         0.00
N                           3221        3221        3221
Safe vs. "Super" Virgin   0.091 ***   0.405 ***   0.124 **
                           (0.020)     (0.093)     (0.049)

F stat on instruments        --         21.7        74.8
First-stage [R.sup.2]        --         0.31        0.43
Sargan Overid p-value        --         0.20        0.35
Breusch-Pagan p-value        --          --         0.14
N                           2911        2911        2911

                                       Females

                                      Age 16-18

                                      Standard     Lewbel
                          OLS (10)     IV (11)     IV (12)

Nonvirgin vs. Virgin      0.104 ***     0.089       0.264
                           (0.017)     (0.086)     (0.248)
F stat on instruments        --         22.9         2.4
First-stage [R.sup.2]        --         0.39        0.39
Sargan Overid p-value        --         0.00        0.06
Breusch-Pagan p-value        --          --         0.89
N                           3825        3825        3825
Safe Current vs. Virgin   0.092 ***     0.084       0.095
                           (0.019)     (0.084)     (0.184)

F stat on instruments        --         23.1         4.2
First-stage [R.sup.2]        --         0.39        0.40
Sargan Overid p-value        --         0.02        0.05
Breusch-Pagan p-value        --          --         0.00
N                           3167        3167        3167
Safe vs. "Super" Virgin   0.089 ***     0.054       0.051
                           (0.021)     (0.090)     (0.270)

F stat on instruments        --         22.2         2.1
First-stage [R.sup.2]        --         0.46        0.46
Sargan Overid p-value        --         0.02        0.14
Breusch-Pagan p-value        --          --         0.00
N                           2800        2800        2800

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and IV models include each of the individual-level, family-level,
and school-level observable characteristics described in Table 1.
For standard IV models, exclusion restrictions include parental
disapproval of adolescent sex, school-level measures of strong
guilt and attractiveness associated with sex, state sex education
mandates and family planning services, parental discussions with
adolescent about sex, day care services provided in the public
school, and whether the adolescent believes the Bible is the
literal word of God and considers himself to be a born-again
Christian.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 8. IV Estimates of Relationship between
Adolescent Sexual Activity and School Affinity

                                         Males

                                       Age 13-15

                                        Standard      Lewbel
                           OLS (1)       IV (2)       IV (3)

Nonvirgin vs. Virgin      -0.142 ***   0.817 ***      0.132
                           (0.047)      (0.261)      (0.135)
F stat on instruments         --          15.2         70.0
First-stage [R.sup.2]         --          0.34         0.42
Sargan Overid p-value         --          0.20         0.48
Breusch-Pagan p-value         --           --          0.00
N                            3127         3127         3127

Safe Current vs. Virgin    0.107 **    0.915 ***      0.033
                           (0.051)      (0.297)      (0.103)
F stat on instruments         --          13.3        119.6
First-stage [R.sup.2]         --          0.31         0.49
Sargan Overid p-value         --          0.28         0.90
Breusch-Pagan p-value         --           --          0.00
N                            2917         2917         2917

Safe vs. "Super" Virgin    0.087 *      0.698 **      -0.004
                           (0.054)      (0.285)      (0.015)
F stat on instruments         --          14.4        105.5
First-stage [R.sup.2]         --          0.36         0.52
Sargan Overid p-value         --          0.07         0.80
Breusch-Pagan p-value         --           --          0.00
N                            2624         2624         2624

                                         Males

                                       Age 16-18

                                        Standard      Lewbel
                           OLS (4)       IV (5)       IV (6)

Nonvirgin vs. Virgin       0.062 *       0.150        0.243
                           (0.037)      (0.225)      (0.669)
F stat on instruments         --          15.6         1.8
First-stage [R.sup.2]         --          0.33         0.33
Sargan Overid p-value         --          0.11         0.78
Breusch-Pagan p-value         --           --          0.36
N                            3958         3958         3958

Safe Current vs. Virgin     0.059        -0.046       0.182
                           (0.040)      (0.229)      (0.417)
F stat on instruments         --          14.7         4.4
First-stage [R.sup.2]         --          0.32         0.32
Sargan Overid p-value         --          0.01         0.13
Breusch-Pagan p-value         --           --          0.00
N                            3359         3359         3359

Safe vs. "Super" Virgin     -0.003       0.098        -0.209
                           (0.045)      (0.241)      (0.732)
F stat on instruments         --          14.8         1.4
First-stage [R.sup.2]         --          0.40         0.40
Sargan Overid p-value         --          0.05         0.62
Breusch-Pagan p-value         --           --          0.02
N                            2974         2974         2974

                                        Females

                                       Age 13-15

                                        Standard      Lewbel
                           OLS (1)       IV (2)       IV (3)

Nonvirgin vs. Virgin       0.082 *       0.021        0.029
                           (0.045)      (0.236)      (0.147)
F stat on instruments         --          18.7         46.0
First-stage [R.sup.2]         --          0.31         0.38
Sargan Overid p-value         --          0.04         0.01
Breusch-Pagan p-value         --           --          0.00
N                            3488         3488         3488

Safe Current vs. Virgin     0.048        0.196        0.047
                           (0.052)      (0.252)      (0.128)
F stat on instruments         --          20.1         80.7
First-stage [R.sup.2]         --          0.27         0.40
Sargan Overid p-value         --          0.01         0.16
Breusch-Pagan p-value         --           --          0.00
N                            3221         3221         3221

Safe vs. "Super" Virgin     0.029        0.152        0.085
                           (0.055)      (0.247)      (0.135)
F stat on instruments         --          21.7         74.8
First-stage [R.sup.2]         --          0.31         0.43
Sargan Overid p-value         --          0.03         0.10
Breusch-Pagan p-value         --           --          0.00
N                            2911         2911         2911

                                        Females

                                       Age 16-18

                                        Standard      Lewbel
                           OLS (4)       IV (5)       IV (6)

Nonvirgin vs. Virgin        0.027        0.072        -0.614
                           (0.040)      (0.195)      (0.576)
F stat on instruments         --          22.9         2.4
First-stage [R.sup.2]         --          0.39         0.39
Sargan Overid p-value         --          0.01         0.07
Breusch-Pagan p-value         --           --          0.89
N                            3825         3825         3825

Safe Current vs. Virgin     0.020        0.241        -0.787
                           (0.043)      (0.191)      (0.458)
F stat on instruments         --          23.1         4.2
First-stage [R.sup.2]         --          0.39         0.40
Sargan Overid p-value         --          0.01         0.30
Breusch-Pagan p-value         --           --          0.00
N                            3167         3167         3167

Safe vs. "Super" Virgin     0.016        0.713       -1.32 *
                           (0.048)      (0.205)      (0.690)
F stat on instruments         --          22.2         2.1
First-stage [R.sup.2]         --          0.46         0.46
Sargan Overid p-value         --          0.02         0.61
Breusch-Pagan p-value         --           --          0.00
N                            2800         2800         2800

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and IV models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. For standard IV models, exclusion restrictions
include parental disapproval of adolescent sex, school-level
measures of strong guilt and attractiveness associated with
sex, state sex education mandates and family planning services,
parental discussions with adolescent about sex, day care services
provided in the public school, and whether the adolescent believes
the Bible is the literal word of God and considers himself to be a
born-again Christian.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 9. IV Estimates of Relationship between Adolescent
Sexual Activity and College Aspirations

                                         Males

                                       Age 13-15

                                        Standard      Lewbel
                           OLS (7)       IV (8)       IV (9)

Nonvirgin vs. Virgin      -0.084 ***     0.020         0.09
                           (0.026)      (0.139)      (0.128)
F stat on instruments         --          15.2         70.0
First-stage [R.sup.2]         --          0.34         0.42
Sargan Overid p-value         --          0.46         0.13
Breusch-Pagan p-value         --           --          0.00
N                            3125         3125         3125

Safe Current vs. Virgin   -0.060 **      0.040        0.013
                           (0.027)      (0.150)      (0.054)
F stat on instruments         --          13.3        119.5
First-stage [R.sup.2]         --          0.31         0.49
Sargan Overid p-value         --          0.59         0.22
Breusch-Pagan p-value         --           --          0.00
N                            2915         2915         2915

Safe vs. "Super" Virgin   -0.067 **      0.019        -0.04
                           (0.029)      (0.146)      (0.058)
F stat on instruments         --          14.4        105.4
First-stage [R.sup.2]         --          0.36         0.52
Sargan Overid p-value         --          0.70         0.29
Breusch-Pagan p-value         --           --          0.00
N                            2622         2622         2622

                                         Males

                                       Age 16-18

                                        Standard      Lewbel
                           OLS (10)     IV (11)      IV (12)

Nonvirgin vs. Virgin      -0.066 ***     0.088        0.350
                           (0.024)      (0.146)      (0.453)
F stat on instruments         --          15.6         1.8
First-stage [R.sup.2]         --          0.33         0.33
Sargan Overid p-value         --          0.87         0.23
Breusch-Pagan p-value         --           --          0.71
N                            3958         3958         3958

Safe Current vs. Virgin   -0.060 **      0.014        0.095
                           (0.025)      (0.144)      (0.265)
F stat on instruments         --          14.6         4.4
First-stage [R.sup.2]         --          0.32         0.32
Sargan Overid p-value         --          0.76         0.23
Breusch-Pagan p-value         --           --          0.00
N                            3359         3355         3355

Safe vs. "Super" Virgin     -0.03        0.023        0.545
                           (0.027)      (0.154)      (0.503)
F stat on instruments         --          14.8         1.4
First-stage [R.sup.2]         --          0.40         0.40
Sargan Overid p-value         --          0.00         0.68
Breusch-Pagan p-value         --           --          0.02
N                            2970         2970         2970

                                        Females

                                       Age 13-15

                                        Standard      Lewbel
                           OLS (7)       IV (8)       IV (9)

Nonvirgin vs. Virgin      -0.111 ***     -0.100     -0.274 **
                           (0.039)      (0.200)      (0.125)
F stat on instruments         --          18.7         46.2
First-stage [R.sup.2]         --          0.31         0.38
Sargan Overid p-value         --          0.88         0.29
Breusch-Pagan p-value         --           --          0.00
N                            3487         3487         3487

Safe Current vs. Virgin    -0.082 *      -0.014       -0.144
                           (0.044)      (0.212)      (0.108)
F stat on instruments         --          20.0         80.8
First-stage [R.sup.2]         --          0.27         0.40
Sargan Overid p-value         --          0.87         0.30
Breusch-Pagan p-value         --           --          0.00
N                            3220         3220         3220

Safe vs. "Super" Virgin   -0.053 **      0.020        0.016
                           (0.025)      (0.112)      (0.061)
F stat on instruments         --          21.7         75.0
First-stage [R.sup.2]         --          0.31         0.44
Sargan Overid p-value         --          0.66         0.23
Breusch-Pagan p-value         --           --          0.00
N                            2910         2910         2910

                                        Females

                                       Age 16-18

                                        Standard      Lewbel
                           OLS (10)     IV (11)      IV (12)

Nonvirgin vs. Virgin      -0.134 ***    -0.334 *      -0.176
                           (0.036)      (0.176)      (0.505)
F stat on instruments         --          23.1         2.4
First-stage [R.sup.2]         --          0.39         0.39
Sargan Overid p-value         --          0.33         0.26
Breusch-Pagan p-value         --           -           0.89
N                            3824         3824         3824

Safe Current vs. Virgin   -0.104 ***    -0.317 *     -0.639 *
                           (0.037)      (0.166)      (0.372)
F stat on instruments         --          23.5         4.2
First-stage [R.sup.2]         --          0.39         0.40
Sargan Overid p-value         --          0.28         0.41
Breusch-Pagan p-value         --           --          0.00
N                            3166         3166         3166

Safe vs. "Super" Virgin     -0.047       -0.231       -0.02
                           (0.042)      (0.179)      (0.539)
F stat on instruments         --          22.2         2.1
First-stage [R.sup.2]         --          0.46         0.47
Sargan Overid p-value         --          0.20         0.45
Breusch-Pagan p-value         --           --          0.00
N                            2799         2799         2799

Source: Computed by the author.

Standard errors of the estimates appear in parentheses. Each
coefficient estimate is from a separate regression model. OLS
and IV models include each of the individual-level, family-
level, and school-level observable characteristics described
in Table 1. For standard IV models, exclusion restrictions
include parental disapproval of adolescent sex, school-level
measures of strong guilt and attractiveness associated with
sex, state sex education mandates and family planning services,
parental discussions with adolescent about sex, day care services
provided in the public school, and whether the adolescent believes
the Bible is the literal word of God and considers himself to be a
born-again Christian.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.