Reading, Writing, and Sex: the Effect of Losing Virginity on Academic Performance
By Sabia, Joseph J
Controlling for a wide set of individual- and family-level observables available in the National Longitudinal Study of Adolescent Health, ordinary least squares (OLS) estimates show that sexually active adolescents have grade point averages that are approximately 0.2 points lower than virgins. However, when information on the timing of intercourse decisions is exploited and individual fixed effects are included, the negative effect of sexual intercourse disappears for females, but persists for males. Taken together, the results of this study suggest that while there may be adverse academic spillovers from engaging in intercourse for some adolescents, previous studies’ estimates are overstated due to unmeasured heterogeneity. (JEL I10, I21, I18) ABBREVIATIONS
GPA: Grade Point Average
IV: Instrumental Variable
NLSY97: National Longitudinal Survey of Youth 1997
OLS: Ordinary Least Squares
2SLS: Two-stage Least Squares
While much of the policy discussion surrounding efforts to induce teenagers to delay first intercourse tend to focus on the potential health benefits of abstinence, increasing attention has been paid to possible spillover effects. In particular, some proponents of abstinence claim that delaying intercourse can enhance self- control, encourage greater future orientedness, and facilitate human capital accumulation. For example, the Family Research Council, a conservative domestic policy advocacy organization, has argued that maintaining an abstinent lifestyle can facilitate the development of character traits that enhance human capital:
Abstinence-until-marriage builds character and self-control. Unlike slapping on a condom, self-control must be cultivated over time. It is not a technique to master but a deeply rooted prize to nurture. When properly developed, it will help teens become adults that are effective long-range planners …. Just as self-control in the sexual arena benefits other areas of life, likewise the practice of immediate gratification of sexual urges encourages impulsiveness in many areas of life. (Diggs 2002)1
While several studies have carefully examined the effect of teenage childbearing on schooling and labor market outcomes (see, e.g., Angrist and Evans 1996; Bronars and Grogger 1994; Huffman, Foster, and Furstenberg 1993; hotz, McEhOy, and Sanders 2005; hotz, Mullin, and Sanders 1997; Klepinger, Lundberg, and Plotnick 1999; Rosenzweig and Wolpin 1995), fewer have explored whether becoming sexually active adversely affects early human capital accumulation. Those studies that have examined whether there are negative educational spillovers of engaging in sex at an early age have not adequately controlled for unmeasured characteristics associated with both sex decisions and academic performance (BiUy et al. 1988; Brooke et al. 1994; Costa et al. 1995; Dorius, Heaton, and Steffen 1993; Jessor et al. 1983; Meilman 1993; Mott and Marsiglio 1985; Rector and Johnson 2005; Schvaneveldt et al. 2001; Upchurch and McCarthy 1990). The central contribution of this study will be to provide more credible estimates of the causal effect of becoming sexually active on adolescent academic performance by exploiting information on the timing of intercourse decisions, which will permit the estimation of individual fixed-effects models.
One might expect a negative relationship between losing virginity and academic performance for several reasons. Becoming sexually active might cause a decline in academic performance because adolescent sex may psychologically distress or emotionally distract teenagers, causing them to pay less attention to coursework. However, the direction of causality may run in the opposite direction. Poor academic performance may cause adolescents to become sexually active. Teenagers may become disillusioned or depressed due to receiving low grades and may psychologically compensate for their feelings of academic inadequacy by seeking fulfillment in sex. Or, it may be that there is no causal link between early teen sex and academic performance but rather an association due to unmeasured heterogeneity. If the least academically motivated or least able adolescents choose to engage in sexual intercourse, and this motivation level is unmeasured, ordinary least squares (OLS) estimates will be biased toward negative academic consequences of becoming sexually active.
Using data from the National Longitudinal Study of Adolescent Health (Add Health), this study carefully examines the relationship between becoming sexually active and academic performance. Controlling for a wide set of individual- and family-level observables, OLS estimates consistently show that nonvirgins have grade point averages (GPAs) that are approximately 0.2 points lower than virgins. For adolescent females, the negative relationship disappears after including individual fixed effects, suggesting little evidence of a causal link. But for adolescent males, this relationship persists after controlling for time-invariant unobservables. An instrumental variables (IV) identification strategy produces results that are generally consistent with fixed- effects findings, though the estimated effects are weaker, likely due to weak instruments. Taken together, these findings suggest that the negative relationship between early adolescent sex and academic achievement is quite sensitive to controls for unmeasured heterogeneity, and that previous studies’ estimates of negative spillovers are overstated.
II. THEORETICAL AND EMPIRICAL LITERATURE
A. Theoretical Literature
The economy, psychology, and sociology literatures each offer explanations for why we might expect a negative relationship between teenagers’ becoming sexually active and human capital accumulation. One psychological theory suggests that losing one’s virginity has adverse emotional effects on teenagers (see a discussion of this issue in Cutler et al. 2001; lessor and lessor 1975; Rector, Johnson, and Noyes 2003; Sabia 2006; Stiffman et al. 1987), which may cause them to be unable to devote sufficient mental energies to their studies. A related physiological theory suggests that teenagers may go through hormonal changes that make concentration on coursework more difficult. If becoming sexually active has important psychological or physiological effects on teenagers, then emotional instability, psychological distraction, or physiological changes could lead to diminished capacity in preparing for academic classes, resulting in a decline in grades.
Engaging in first intercourse may also serve as an information revelation mechanism for teenagers. The revelation of the true immediate benefits of sex may cause teenagers to change their short- run investment decisions. Thus, for example, if the realized benefits of sexual intercourse are higher than the ex-ante anticipated benefits, adolescents may substitute time and energy away from investments in human capital and toward investments in future obtainment of sex. While this theory does not provide an explanation for why teenagers become sexually active in the first place, it may explain change in human capital accumulation following exits from virginity.
Related to the information revelation hypothesis, problem behavior syndrome theory, advanced by psychologists and sociologists, suggests that immersion in problem behaviors, such as early sexual activity, causes a change in the fundamental outlook of adolescents, causing them to want to explore other antisocial behaviors (Alien, Leadbeater, and Aber 1994; Capaldi, Crosby, and Stoolmiller 1996; Costa et al. 1995; Donovan, Jessor, and Costa 1988; Elliott and Morse 1989; Parrel, Danish, and Howard 1992; Harvey and Spigner 1995; McLean and Flanigan 1993; Peterson, Moore, and Furstenberg 1991; Rosenbaum and Kandel 1990; Schvaneveldt et al. 2001; Whitbeck et al. 1993). Problem syndrome theory predicts that involvement in early sexual activity causes a change in an adolescent’s mindset such that he would want to devote more time to antisocial behaviors and less time investing in human capital. As with the information revelation hypothesis, an important limitation of this framework is that it does not offer an explanation for why adolescents begin engaging in “problem” behaviors.
Economic theory and social exchange theory provide an explanation for the types of adolescents who will select into early sex: those with lower opportunity costs of sex (see, e.g., Becker 1980; Nye 1979; Small, Silverberg, and Kerns 1993). Students with the lower levels of academic achievement may be those most likely to choose to engage in sexual intercourse because these adolescents have the least to lose from the potential consequences of sex given that they may have limited future job and college opportunities. Psychological distress, excitable distraction, pregnancy, or sexually transmitted diseases would be less costly to these adolescents, relative to students who anticipate greater future economic gains. Moreover, students with higher discount rates or who are less risk averse are more likely to select into sexual activity. Given that discount rates, degrees of risk aversion, and anticipated future prospects are difficult to measure, failing to adequately control for unobserved heterogeneity will likely result in estimates biased toward adverse academic effects. B. Empirical Literature
Several studies in the psychology and sociology literature have found a statistical link between early initiation into sexual intercourse and academic achievement. Using cross-sectional estimation techniques that do not account for the endogeneity of sex decisions, many studies have found that initiating sexual intercourse early, particularly earlier than age 15, is associated with significantly lower academic goals and achievement (see, e.g., Billy et al. 1988; Brooke et al. 1994; Costa et al. 1995; Jessor et al. 1983; Meilman 1993; Mott and Marsiglio 1985; Schvaneveldt et al. 2001). A few studies have shown that, relative to virgins, sexually active adolescents are more likely to drop out of school and are less likely to attend college (Dorius, Heaton, and Steffen 1993; Rector and Johnson 2005; Upchurch and McCarthy 1990). However, none of these studies has controlled for the endogeneity of intercourse. Schvaneveldt et al. (2001) recognize this problem and use longitudinal data to try to tease out the direction of causality. However, the authors do not use individual fixed effects or IV techniques. Rather, they measure sexual activity prior to GPA was measured, and conclude that the direction of causality can be established by this temporal ordering. But one might easily imagine that fixed unmeasured characteristics associated with sexual activity at time period t are also correlated with GPA at time period t + 1.
The health economics literature has seen substantial growth in the number of studies that have examined the relationships among adolescent delinquent behaviors, and have stressed the importance of controlling for the endogeneity of delinquent behaviors.2
A forthcoming study by Sabia (2007b) finds that early teen sexual activity is associated with diminished school attachment, but the relationship is quite sensitive to unmeasured heterogeneity. However, no studies in the literature have specifically examined the relationship between teen sexual activity and academic achievement. The outcome examined in this study, GPA, is of particular interest given that several recent studies have found that high school grades are an important determinant of future human capital accumulation and earnings (Betts and Morrell, 1999; Rose and Betts, 2004; Grogger and Eide, 1995; Cohn et al, 2004).
Similar studies in the labor economics literature have studied the relationship between out-of-wedlock childbearing by young girls and future earnings using family-fixed effects, IV, and twin births to control for the endogeneity of pregnancy (see, e.g., Angrist and Evans 1996; Bronars and Grogger 1994; Hoffman, Foster, and Furstenberg 1993; Klepinger, Lundberg, and Plotnick 1999; Rosenzweig and Wolpin 1995). Most of these studies have found significant adverse effects of teenage childbearing. However, more recent studies (hotz, McElroy, and Sanders 2005; hotz, Mullin, and Sanders 1997) that have used miscarriages to provide exogenous variation in pregnancy to identify the causal effect of nonmarital births have found no evidence of adverse effects. hotz and his colleagues conclude that the negative relationship between teenage childbearing and adverse labor market outcomes can largely be explained by selection.
The current study builds upon the previous literature by exploiting information on the timing of adolescent intercourse decisions to better isolate the causal effects of early teen sex on academic performance. This is an important contribution to the literature because virgins and nonvirgins may differ not only on observed characteristics that have been controlled for in previous research but also on unobserved characteristics that are correlated with academic achievement. By controlling for fixed individual- level unmeasured heterogeneity, this study will be better able to determine the appropriateness of interpreting the association between virginity and academic performance causally. Moreover, to examine the robustness of fixed-effects estimates, an IV strategy is employed to explicitly control for the potential endogeneity of sex decisions.
III. THEORETICAL FRAMEWORK
A rational adolescent is assumed to maximize utility, U(s, GPA, L)-where s is sexual intercourse, GPA a measure of academic performance,3 and L leisure-subject to a budget constraint, a time constraint, and an educational skill production function. From this maximization problem, the adolescent’s reduced form demand for sexual intercourse (SEX) and human capital production function (GPA) can be derived:
(1) GPA = f(m, pe, a, h, t, q, z)
(2) SEX = g(p, r, Y, m, a, pe, h, q, z)
where m is student motivation, pe is parental effort and involvement, a is student ability, A is mental and physical health, t is time spent studying, q is school quality, z are taste shifters, p is the shadow price of sex, r are the prices of substitutes for sex, and Y is income.
Becoming sexually active is expected to affect GPA principally through its effects on adolescent motivation (m), time spent studying (t), and mental and psychological health (captured in h). However, GPA may also affect the propensity to exit virginity through its effect on psychological well-being (captured in h). And finally, it may be that GPA and the propensity to lose one’s virginity are related by common observable or unobservable characteristics, such as motivation (m) or ability (a).
Much of the existing virginity-human capital literature has treated sex decisions as exogenous and presented OLS estimates of the production function in Equation (1):
(3) GPA^sub ij^ = alpha + betaSEX^sub i^ + X^sub i^delta + X^sub j^gamma + epsilon^sub ij^
where GPA^sub ij^ is the GPA of adolescent i in family j, SEX a dummy variable equal to 1 if the adolescent has ever had sexual intercourse and equal to O if the adolescent has not, X^sub i^ a vector of adolescent-specific characteristics that capture inputs in Equation (1), and X^sub j^ a vector of family-level characteristics that measure parental involvement and effort in education. However, the identification assumption in Equation (3), E(epsilon|SEX) = 0, is likely to be violated given that teen sex decisions are potentially endogenous.
A potentially more credible identification strategy not yet explored in the virginityhuman capital literature is an individual fixed-effects model. With two periods of data, an individual fixed- effects model of the following form may be estimated:
The above model will control for individualspecific time- invariant unmeasured determinants of GPA that are correlated with sexual behavior.4 However, the identification assumption underlying the fixed-effects strategy, E(epsilon^sub t+1^ – epsilon^sub t^ | SEX^sub t+1^ – SEX^sub t^) = 0, may be violated if time-varying unobservables are correlated with both the decision to become sexually active and with changes in academic performance. For example, adolescents may experience changes in hormones that affect both attention to school work and the probability of becoming sexually active. Moreover, changes in peer relationships or in the home environment may affect both outcomes. Time-varying unmeasured heterogeneity of this form could bias the estimate of o toward adverse academic effects of sexual activity.
An alternative identification strategy would be to explicitly model endogenous sex decisions via IV. This requires estimating the schooling production function as the second-stage of a two-stage least squares model, where instruments (Z) provide exogenous variation in sexual intercourse. An IV strategy could, in principle, expunge endogeneity bias if the instruments are sufficiently powerful predictors of intercourse and are uncorrelated with unmeasured determinants of academic performance, E(epsilon|Z) = 0. However, as discussed below, finding high-quality instruments in the Add Health data is challenging.
The Add Health provides a rich data source to analyze the relationship between adolescent sexual activity and educational performance. The Add Health data set is a school-based nationally representative longitudinal survey containing information from students, their parents, and school administrators in the mid- 1990s. In the first wave of data collection (April 1995 to December 1995), students from 7th to 12th grade were asked questions about their schooling, personality, family, romantic relationships, health behavior, peer groups, neighborhoods, and sexual activity. The Add Health survey was conducted by the Carolina Population Center at the University of North Carolina at Chapel Hill and contained detailed information on adolescent health behaviors and academic outcomes. Bearman, Jones, and Richard (1997) discussed sampling methods and interview strategies in detail. Adolescents were then reinterviewed in the subsequent (1995-96) academic year.
Information on sensitive topics such as sex choices, contraceptive use, and attitudes about sex was collected so as to minimize reporting error. Students were given private laptop computers, which allowed them to anonymously respond to questions, and respondents were assured that the interviewer would never see their responses, nor would anyone be able to link their answers with their name. Parents and school administrators were also interviewed. Parents, usually mothers, were asked about their relationships with their children, their families, and their backgrounds. School administrators were asked questions about how their schools were organized and what types of courses were offered to students. The Add Health data set also contains contextual variables, which provide information on the legal, socioeconomic, and demographic background of the region where the adolescent resides.
The key dependent variable used in the analysis is a constructed measure of self-reported GPA.5 Adolescents are asked separate questions about the grades they received in their most recent English/language arts, math, science, and social studies/history classes. The responses adolescents could offer were A, B, C, and D or lower. From these survey items, I assign a 4.0 for a reported grade of “A,” 3.0 for a reported grade of “B,” 2.0 for a reported grade of “C,” and 0.5 for a reported grade of “D or lower.” A cumulative GPA is then constructed, giving equal weight to each grade.6 One criticism of this measure of academic performance is that it is a self-reported measure. Thus, one might be concerned with inflated grade reports, which could be particularly problematic if such misreporting is correlated with virginity status. However, the mean GPAs measured in the Add Health data set do not appear to differ substantially from the National Longitudinal Survey of Youth 1997 (NLSY97) or the High School and Beyond data sets, each of which provide transcript data. Using data from the NLSY97, Rothstein (2007) reported the mean GPA for high school students of 2.5 for males and 2.8 for females. GPAs reported in the Add Health data set are slightly higher perhaps due to inflated reporting of grades or due to the fact that Add Health does not permit reports of plus or minus grades.7
The key independent variable of interest is a measure of whether the adolescent has ever engaged in sexual intercourse.8 As expected, the percentage of teens who had ever engaged in sexual intercourse rises with age and is higher for males than females. About 11.5 percent of 13- to 14-year-old females and 15.4% of 13- to 14-year- old males report that they have engaged in sexual intercourse at least once; by age 17-18, the percentage of nonvirgins is 54.6% for females and 58.5% for males.9 These estimates are generally similar to those reported in the 1995 National Survey of Family Growth and the 1995 National Survey of Adolescent Males. Given that GPA declines slightly with age, and rates of sexual intercourse increases with age, it will be important to estimate separate models by age, as well as to control for age effects in regression models to ensure that unobserved age trends are not driving the negative correlation between teen sex and academic achievement.
Table 1 presents weighted means of the dependent variable and key control variables by age and virginity status. Across ages, the mean GPA of nonvirgins is consistently lower than the mean GPA of virgins, and this difference is significant. The remaining control variables listed in Table 1 are used in the OLS models. The inclusion of these variables is designed to capture the input measures described in the educational skill production function (Equation ): student motivation, parental effort, student effort, school quality, and health.10 The measures used in this study improve upon much of the previous literature because the Add Health data include a wide set of observable characteristics that capture parental schooling sentiments.11
Given the longitudinal nature of the Add Health data, individual fixed-effects models of the form described in Equation (4) may be estimated, where GPA and sexual activity are measured in successive academic years.12 Control variables in the individual fixed-effects model are those that change over time in the data and are noted in footnote 1 of Table 1.
Identification of the IV model requires plausible exclusion restrictions that are strongly correlated with teen sex decisions but are uncorrelated with unmeasured characteristics that affect GPA. These variables include the number of county-level family planning service providers per 10,000 population, whether there is an abortion provider in the county, whether the adolescent’s school provides or refers students to family planning materials, whether school policy requires the transfer of pregnant students to alternate schools, the adolescent’s randomly selected schoolmates’ perceptions of sex, and parental attitudes about sex.13 Descriptions and means of these variables are listed in Table 1.
Because there are multiple instruments, overidentification tests can provide suggestive evidence on the credibility of the exogeneity assumption of the IV model. However, there is some reason to be concerned that some of these instruments may be correlated with unmeasured determinants of schooling. For example, while the GPA equation includes several measures of pro-schooling parental sentiment, ’4 parental attitudes toward their children having sex may be correlated with unmeasured schooling expectations. Similarly, schoolspecific policies on pregnancy policies may be correlated with grading standards. Finally, measures of the sexual attitudes of an adolescent’s randomly selected schoolmates capture peer effects that are likely to be associated with parents’ choice of their children’s learning environment (see, e.g., Evans, Oates, and Schwab 1992; Gaviria and Raphael 2001; Sacerdote 2001). Parents who place their children among schoolmates who have more permissive attitudes toward sex may be less likely to care about their children’s academic performance. Thus, the per capita number of county-level family planning service providers and the availability of an abortion provider in the county may be more plausibly exogenous instruments.15
Lewbel (2006) noted that the assumptions underlying his approach had also been exploited to identify correlated random-coefficients models (Heckman and Vytlacil 1998). Moreover, Rigobon (2002, 2003), Klein and VeUa (2003), King, Sen tana, and Wadhwani (1994), and Sentena and Fiorentini (2001) have exploited heteroskedasticity to identify models in a manner similar to that proposed by Lewbel (2006). Learner (1981) and Feenstra (1994) also exploited heteroskedasticity to aid in identification. Several recent papers have used similar approaches, using plausible restrictions on higher order moments rather than traditional instruments to aid in identification (Cragg 1997; Dagenais and Dagenais 1997; Sabia 2007a, b; Erickson and Whited 2002; Lewbel 1997; Rummery, Vella, and Verbeek 1999).
A. OLS Estimates
Estimation results are found in Tables 2-5.16 In Table 2, OLS estimates of the GPA production function are presented to replicate existing findings in the literature. Results are obtained using data from Add Health’s baseline wave for adolescents aged 15-16. The findings in Table 2 suggest robust evidence of a negative relationship between early adolescent sex and GPA across model specifications. Models 1 and 2 include clearly exogenous variables as controls. Model 1 includes race, sex, and age as covariates, while Model 2 adds controls for household income, mother’s education, household structure, and regional effects. These results reflect that sexually active teenagers have GPAs that are 0.34-0.39 points lower than those who are not sexually active.
The remaining specifications add further controls for inputs in the human capital production function that are arguably endogenous but capture important inputs that are likely to be correlated with adolescent sex decisions. Model 3 includes measures of parental involvement in adolescent education as well as measures of the harmoniousness of the parent-child relationship, while Models 4 and 5 control for physical health, mental health, and employment. While the coefficient on the sex parameter falls to -0.26, it remains strongly significant. Model 6 adds a control for romantic relationship status to separate the effects of being in a relationship from being sexually active; the results remain unchanged.17 The specification in column 7 controls for adolescent’s college aspirations and innate intelligence, measured by the Add Health Picture and Vocabulary Test Score. The coefficient on sexual activity becomes smaller (-0.20) but remains significant. And finally, Model 8 includes a control for alcohol consumption, which is expected to be positively correlated with sexual activity and negatively associated with GPA. As expected, the coefficient on intercourse falls slightly but remains highly significant.
Taken together, the findings in Table 2 suggest robust evidence of a significant negative relationship between sexual activity and GPA, with a magnitude around -0.20. These results are consistent with much of the previous psychological and sociological literature (Billy et al. 1988; Brooke et al. 1994; Costa et al. 1995; Jessor et al. 1983; Meilman 1993; Mott and Marsiglio 1985; Schvaneveldt et al. 200l).18 Moreover, the estimate in column 8 may be considered a lower bound because sexual activity may also affect many of these arguably endogenous variables that affect GPA. The remaining regressions include the full set of controls listed in Model 8.19
B. Fixed-Effects Estimates
Table 3 compares OLS estimates to schoolfixed effects and individual-fixed effects estimates. These models allow heterogeneous effects of sexual activity by age and gender. Each estimate in Table 3 comes from a separate regression model estimated on a sample that is restricted to those adolescents who have nonmissing information on cumulative GPA and virginity status in consecutive academic periods.20
OLS estimates on this sample, found in row 1, are generally consistent with the estimates shown in Table 2. Estimated coefficients are generally larger for males than females. For 17- to 18-year-olds, there is no significant relationship for either males or females, but this is driven by the smaller selected sample. Adolescents who have grades in consecutive academic periods are those that graduate high school at later ages. Hence, because of the selected sample as well as the reduced power of the design, it is not surprising to find insignificant relationships. The remaining discussion will, therefore, focus on adolescents aged 13-16.21
One form of heterogeneity that could bias OLS estimates of the relationship between virginity and academic performance toward adverse academic effects is school-level heterogeneity. If, for example, schools with the most permissive sexual attitudes are of the lowest unobserved academic quality, then students in these schools may have low grades because their schooling environment is not as committed to encouraging academic success. This type of school-level heterogeneity would tend to bias OLS estimates toward adverse academic effects of becoming sexually active. Models including school fixed effects are estimated in row 2. These results suggest that unobserved school quality is not an important source of bias in OLS estimates. However, individual-level unmeasured heterogeneity remains an important concern. Students of the highest unobserved discipline or academic ability may be those who are most likely to choose to delay intercourse. Individual fixed-effects models are presented in row 3. These models control for several time- varying observable characteristics: whether the adolescent is in a romantic relationship, alcohol consumption, employment, body mass index, self-perception of bad health, aspirations to attend college, parental sentiments toward college education, number of family dinners per week, attempted suicides, quality of the parent-child relationship, and age.22
For females, there is strong evidence that fixed individual- level unobserved heterogeneity biases OLS estimates toward adverse academic effects of becoming sexually active. After controlling for individual fixed effects, the relationship between virginity and academic performance becomes statistically insignificant, with the magnitude of the estimated parameter falling over tenfold.23
For males, however, the negative relationship between losing virginity and academic performance is robust to the inclusion of individual fixed effects, though the magnitude is smaller. Becoming sexually active is associated with a 0.18- to 0.19-point decline in GPA. For males aged 15-16, the magnitude of the relationship falls (from -0.34 to -0.18), suggesting some evidence of selection into sexual activity based on unobserved characteristics associated with lower academic performance.
In row 4, the robustness of the individual fixed effects results to propensity score matching is examined so as to assure common support on observable characteristics. The propensity score matching exercise is executed by estimating a probit model of the change in virginity status between Waves 1 and 2 (=0 if no change, = 1 if exit virginity) on the set of baseline individual- and family-level observables listed in Table I.24 Adolescents are matched within caliper of 0.10 and without replacement. Then, a simple first- differences estimate is obtained using the matched sample. The fixed- effects propensity scorematched estimate of the relationship between losing virginity and academic performance is generally consistent with the standard individual fixed-effects estimate. For females, there continues to be no evidence of significant relationship. For males aged 15-16, there is still a significant negative relationship, with the magnitude slightly higher than the individual fixed-effects estimate (-0.28 vs. -0.18). For males aged 13-14, the magnitude of the coefficient remains stable, but the standard error is inflated due to the more stringent common support requirement.25
Taken together, the findings in Table 3 suggest that, for females, the negative relationship between becoming sexually active and academic performance can be explained by individual-level unmeasured heterogeneity. Thus, a causal interpretation of results presented in previous studies in the literature is not appropriate. For adolescent males, however, the negative relationship is robust to the inclusion of individual fixed effects and deserves further exploration.
C. Robustness Tests
Table 4 examines the robustness of individual fixed-effects estimates to changes in model specification, sample selected, and definitions of the dependent variable. First, note that the estimated relationship between virginity and academic performance is never significant for females, across any specification. Thus, the evidence that unmeasured heterogeneity can explain the negative association between sexual activity and GPA remains fairly strong for females. The remaining discussion of Table 4 focuses on males.26
The chief concern with the individual fixed-effects identification strategy is that there may be unobserved time- varying individual-level characteristics that are associated with both exiting virginity and reduced grades. In Table 4, the robustness of the fixed-effects findings to important observables is examined. Rows 1 and 2 reflect that the fixed-effects results are not sensitive to the inclusion or exclusion of measured time- varying characteristics.
One potentially important time-varying characteristic that is difficult to measure is puberty, the omission of which may be especially problematic for younger teens (aged 13-14). If the loss of virginity is simply a proxy for the onset of puberty, then puberty-and not exiting virginity-may create hormonal changes that diminish cognitive ability. If this is the case, then policies designed to delay first intercourse will not significantly improve academic performance. In row 3, some observable measures of beginning puberty are included. Adolescent boys are asked about the degree of facial hair and underarm hair that that they have, and girls are asked about the curves on their body and the onset of menstruation. When these measures are included, the estimated relationship between becoming sexually active and academic performance remains unchanged.
Another concern is that exiting virginity simply serves as a crude proxy for teenage parenthood. That is, adolescents who have had sex might be parents, and it may be the responsibilities of parenthood rather than losing virginity that diminishes academic performance. Moreover, perhaps it is not losing one’s virginity that causes adverse academic outcomes, but rather, becoming sexually active and engaging in unsafe sex. This is a reasonable explanation if worries about pregnancy or STDs cause less attention to one’s studies. In row 4, individuals who engaged in sexual intercourse without contraception or have been/caused a pregnancy are excluded. Across models, fixed-effects estimates do not change.27
The identification assumption of the individual fixed-effects model requires common unobserved time trends between those whose virginity status does not change and those that exit virginity, but this may not be a reasonable assumption if the effects of virginity are cumulative. Thus, in row 5, the sample is restricted to those adolescents who are virgins at baseline. The findings continue to show a significant negative relationship between becoming sexually active and academic performance.28
Taken together, the results in Table 4 suggest that for males, the significant negative relationship between losing virginity and academic performance for males cannot be fully explained by fixed individual-level unobserved heterogeneity, the onset of puberty, the onset of pregnancy, or engaging in unsafe sex. This may suggest some evidence of negative academic spillovers for males, which may not be trivial in magnitude. If permanent, GPA declines of this magnitude could have an impact on the quality of college to which an adolescent may gain admittance (Manski and Wise 1983). The education literature suggests that college quality could have important effects on future earnings, particularly for private elite colleges (see Brewer, Eide, and Ehrenberg 1999).
However, an important caveat to the individuals’ fixed-effects results is that they may not provide unbiased estimates of the effects of early teen sex on GPA if there are timevarying unobservables correlated with the decision to become sexually active and with academic performance. For example, unmeasured changes in peer groups or family environment might influence both outcomes. Thus, to test the robustness of fixed-effects results, an IV identification strategy is undertaken to explicitly model the endogeneity of sex decisions.
D. IV Estimates
Two-stage least squares (2SLS) estimates of the relationship between losing virginity and academic performance for adolescent males are presented in Table 5, along with OLS estimates for comparison.29 To examine the credibility of the instrument exogeneity assumption, two suggestive tests are conducted: (1) instruments are included in the OLS GPA model and tests of their individual and joint significance are presented, and (2) overidentification tests are presented for the IV models. The instruments are never individually or jointly significant at the 5% level in any of the OLS models and the overidentification tests in the IV models suggest that the instruments are valid.
In Table 5, several IV models are presented to test the sensitivity of findings to choice of exclusion restrictions. Both the standard IV and heteroskedasticity-identified IV models continue to provide some evidence of a negative relationship between engaging in sexual intercourse and academic performance for males, though the effects are imprecisely estimated and are often small in magnitude. Because of the large estimated standard errors on 2SLS estimates, I cannot reject OLS estimates. One important reason for the imprecisely estimated parameters is the weakness of the instruments in predicting sexual intercourse. For example, county-level family planning services are generally only marginally significant predictors of intercourse for males. The strongest predictors of intercourse-parental and peer attitudes-are those measures that we are most concerned may be correlated with unmeasured determinants of school achievement. Thus, while the 2SLS estimates are generally consistent with fixed-effects estimates, caution should be taken in their interpretation. In summary, the evidence presented suggests that the relationship between early teen sex and academic performance is sensitive to unmeasured heterogeneity. For females, there is little evidence of a causal relationship after controlling for individual unobserved heterogeneity, while for males, the relationship is more robust, suggesting some evidence of modest educational spillovers.
Using data from the National Longitudinal Study of Adolescent Health, this study estimates the relationship between becoming sexually active and adolescent academic performance. While OLS and school-fixed effects estimates suggest that adolescents who remain virgins have GPAs that are 0.2 points higher than those who become sexually active, I show that relationship can be explained, in part, by unmeasured characteristics associated with selection into sexual activity. For adolescent females, the inclusion of individual fixed effects results in estimated academic effects becoming small and insignificant. For males, however, the result is robust to the inclusion individual fixed effects but becomes weaker after controlling for the endogeneity of sex decisions. These findings suggest that previous studies’ estimates of the negative effects of early adolescent sex are overstated.
One intriguing finding in this study is possible evidence of heterogeneous academic effects of sexual intercourse by gender. One explanation for this finding may be based in biological differences or differences in the revelation of new information. After having sex for the first time, boys may be more likely than girls to become single minded in pursuing sexual conquests. The experience of first sexual intercourse may reveal new information to males on the immediate benefits of sex, and this information may induce boys to choose immediate investments in sex over schooling. For example, teenage boys may realize a social status gain from losing their virginity and view additional sexual “conquests” as a means to achieve even greater social status. Females may not experience such status gains from pursuing sex over education. An important area for future research would involve empirical tests of the information revelation hypotheses of teen sex. To what extent do teens update their beliefs about the perceived benefits and costs of sex after their first sexual experience, and do these updates vary by sex, age, and race? Moreover, an empirical investigation of the impact of delaying intercourse on future human capital accumulation will be important in understanding whether there are long-run non-sex- related benefits of abstinence.
While the strength of this study relies on its exploitation of the timing of first intercourse to identify the effects of early adolescent sex, one of its important limitations is the lack of powerful instruments to explicitly address the endogeneity of virginity decisions. Future work should pay careful attention to modeling the endogeneity of sex decisions.
1. Similar sentiments have been articulated by other socially conservative organizations, including the Heritage Foundation (Rector and Johnson 2005) and Concerned Women for America (Wallace and Warner 2002).
2. For example, Sen (2002) and Rees, Argys, and Averett (2001) examined the relationship between alcohol consumption and teen sexual activity. Facing a similar selection bias problem described above, Sen and Rees et al. established the need for exogenous variation in drinking to identify a causal relationship between drinking and sex. Sen (2002) used beer taxes as an instrument variable since beer taxes are theoretically believed to influence drinking behavior, but not sexual intercourse, except through drinking. Rees, Argys, and Averett (2001) used state requirements that schools offer alcohol and drug prevention education, per capita local and state expenditures on police protection, the number of arrests per violent crime in the county of residence, and the number of total arrests per crime in the county of residence as instruments. After controlling for the endogeneity of drinking, they are better able to make informed statements about the appropriateness of inferring a causal relationship between drinking and sex.
3. This can be thought of as a proxy for future consumption.
4. To assure that there is sufficient common support on observable characteristics among those whose virginity status does change between period t and t + 1, adolescents by nearest propensity score. First, a probit model of the probability of changing virginity status is estimated: SEX^sub t+1^ – SEX^sub t^ = Phi(delta^sub n^X^sub in^ + gamma^sub m^X^sub jm^) where phi is the standard normal distribution. Then, those whose virginity status did change are matched to those whose virginity status did not change by nearest propensity score, where the difference between each treated and untreated adolescent’s predicted probability is no greater than 0.10 (a within caliper estimate). After adolescents are “matched,” a first-difference estimate is obtained. Thus, while the fixedeffects propensity score-matched estimator assures common support on observables among adolescents, as well as controls for fixed individual unobserved heterogeneity, this estimator may still be biased if there are time-varying unobservables associated with entrance into sexual intercourse and changes in academic performance.
5. The Add Health survey item corresponding to grades is, “At (the most recent grading period/last grading period in the spring), what was your grade in __?”
6. An alternative to creating a continuous numerical GPA variable as a measure of academic achievement would be to leave the grade measure as a categorical variable. This would imply multinomial probit or multinomial logit estimation. Such models produce results similar to what is presented here.
7. Moreover, when I examined transcript-reported grades on high school seniors using data from the NLSY97, I find that the correlation between GPA and virginity status is similar to that found in Add Health.
8. The survey item corresponding to this question is, “Have you ever had sexual intercourse? When we say sexual intercourse, we mean when a male inserts his penis into a female’s vagina. ” Note that this definition does not address the timing of most recent intercourse. Nonvirgins can be currently sexually active or not currently sexually active. Three different measures of “current sexual activity” were constructed to try to better isolate this timing issue: (1) sexually active in the past year, (2) sexually active in the past schoolyear, and (3) sexually active in the previous 3 months. None of the results using these definitions of sex was substantively different from the results presented in the paper.
9. See Appendix A for weighted means of GPA and independent variables by age and gender.
10. A measure of innate student ability is also included, the Add Health Picture Vocabulary Test score. This is an abridged version of the Peabody Picture Vocabulary Test that measures an adolescent’s receptive vocabulary, verbal ability, and scholastic aptitude. This test was administered at the in-home survey in Wave 1 of Add Health data collection.
11. One of the strongest correlates of adolescent academic achievement is parental schooling (Miller and Sneesby 1988; Schvaneveldt et al. 2001; Teachman 1987; Thornton and Camburn 1989). Other measured variables capture parental preferences about higher education, parental involvement in adolescents’ schooling, and parental relations with the adolescent. Measures of the adolescent’s physical and mental health include body mass index, physical health and mental health body mass index (see, e.g., Sabia 2007a).
12. Between-wave variations in these measures are presented in Appendix B.
13. Each of these measures is theoretically expected to influence teenage sex decisions but not academic performance. The availability of county-level family planning services and the presence of an abortion provider are each expected to reduce the costs of sex by providing low-cost contraception information and services. School policies that raise the costs of pregnancy by requiring pregnant teenage girls to attend alternate schools are expected to reduce the likelihood of teen sex. Students attending schools with schoolmates that have more permissive attitudes toward sex are more likely to have lower search costs for a sexual partner than a student attending a school with schoolmates that have more conservative attitudes toward sex. And students who have parents with more permissive attitudes toward sex are more likely to engage in sexual activity because stigma costs are low.
14. These variables include (1) whether the parent moved to the neighborhood for the quality of the schools, (2) whether the parent is a member of the Parent-Teacher Association, (3) whether the parent prioritizes scholastic brilliance by their children, (4) whether the mother has graduated from college, (5) whether the parent talks with the adolescent about school work, and (6) whether the parent strongly disapproves if the child does not attend college.
15. In addition to the instruments described above, I also attempt several others. First, following Eisenberg (2004) and Argys and Rees (2006), I create a variable measuring the interaction of the young adolescent’s grade level with the school structure to capture whether the adolescent attended a school with older peers or with younger peers. One might expect that younger adolescents attending schools with older peers might be more likely to engage in sexual activity than younger adolescents attending schools with younger peers. However, this instrument was never a significant predictor of intercourse. Moreover, when I included state-level parental consent laws and mandatory waiting periods as exclusion restrictions, these instruments were never individually or jointly significant. 16. OLS, school-fixed effects, and individual-fixed effects models are estimated using Add Health’s design effect and are weighted (Chantala 2003).
17. In specifications not presented here, I estimate models separately for those who report being in a romantic relationship and those who do not. The coefficient estimate is significant in each specification but is larger in the sample for those not in a romantic or romantic-like relationship. While this may reflect heterogeneous effects of sex by relationship status, but it does not persist in later fixed-effects estimates.
18. The OLS estimates in Tables 2 are robust to different definitions of the dependent variable by academic subject, across age categories, and across race groups. Alternative specifications are available upon request. Appendix C presents coefficient estimates on control variables for OLS models run by age and gender.
19. For example, the Add Health Picture and Vocabulary Test score may also capture a measure of academic achievement, which may be affected by becoming sexually active. Thus, the effect of teen sex on academic performance could, in principle, be biased downward. Estimates of the relationship between key inputs-student effort, ability, health, and parental involvement-and academic performance are consistent with theoretical expectations and the previous literature. Students with higher innate ability, measured by the Add Health Vocabulary Test Score, have significantly higher grades, as do students who aspire to a college education. Students in bad health or with higher body mass indexes have significantly lower grades, consistent with Sabia (2007a). Moreover, mental health shocks through a friend’s attempted suicide are associated with a significantly lower mean GPA. Greater parental involvement and effort in their children’s education and greater parental tastes for education are associated with significantly higher academic performance, consistent with much of the literature (Miller and Sneesby 1988; Schvaneveldtetal. 2001;Teachman 1987; Thorn ton and Camburn 1989). Greater absences from school are associated with significantly lower grades.
20. A small percentage of observations (<4%) reported not being virgins in Wave I, but in Wave II reported that they were virgins. These observations are dropped from the analysis. Receding these individuals as nonvirgins in both waves does not change the results presented. Estimation results include dummy variables for missing information on control variables. However, restricting the data to only observations on which there are nonmissing observations for each of the control variables in both waves of data does not change the findings.
21. When I do not restrict the sample of 17- to 18-year-olds to including nonmissing grade data in subsequent academic years (thus allowing for individual fixed-effects estimates), I find that school fixed-effects estimates of the relationship between exiting virginity and academic performance are similar to those presented in Table 2.
22. Simple first-difference models that included no additional covariates were also estimated. The results produced similar results as those that included each of these variables. Estimates of coefficients on time-varying covariates are available upon request.
23. This finding is robust when the sample includes only those whose romantic relationship status has changed.
24. These probit models indicate that for 13- to 14-year-olds, adolescents who aspire to attend college and have mothers who are college graduates are less likely to exit virginity. Those that are in a romantic relationship, are in a single-parent household, or have experienced a recent suicide attempt by a friend or family member are more likely to exit virginity. For 15- to 16-year-olds, recent suicides are positively associated with exits from virginity, and greater parental involvement is associated with a lower likelihood of exiting.
25. If I relax the caliper requirement to 0.25, the coefficient is statistically significant at the 10% level.
26. Appendix B shows the means and variation in GPA and losing virginity between waves of data. Note that gender-specific differences in findings cannot be explained by greater within- person variation in intercourse or grades for males relative to females.
27. However, for 13- to 14-year-old males, the coefficient is no longer significant due to the larger standard error caused, in part, by the sample size reduction. These restrictions reduce the sample size by 14% from 1,255 to 1,082.
28. Several other robustness checks, not presented here, included controls for perceived popularity, unexcused absences from school, and degree of attentiveness in class. The inclusion of any of these observed measures did not change the individual fixed-effects estimates. Moreover, one might be concerned that the timing of sexual intercourse could be important. Those who are no longer virgins but are not currently sexually active might not see much of an effect on grades. Similarly, those who became sexually active between waves of the survey may have become sexually active at the end of the academic year, resulting in little effect on grades. Because the Add Health data do contain some information on the month- specific timing of intercourse, this issue was explored. The individual fixed-effects findings were robust to controls for the timing of intercourse.
29. F tests of the joint significance of the instruments and the added explanatory power (partial R^sup 2^) of the instruments are presented. Judged by traditional relevance standards suggested by Staiger and Stock (1997), weak instruments do not appear to be an especially important problem. Moreover, for the Lewbel models, p values for the Breusch-Pagan test for heteroskedasticity in the first-stage intercourse equation are presented. As noted above, first-stage heteroskedasticity is required for identification using the Lewbel approach.
Allen, J. P., B. J. Leadbeater, and J. L. Aber. “The Development of Problem Behavior Syndromes in At-Risk Adolescents.” Development and Psychopathology, 6, 1994, 323-42.
Angrist, J., and W. Evans. “Children and their Parent’s Labor Supply: Evidence from Exogenous Variation in Family Size.” American Economic Review [or National Bureau of Economic Research Working Paper No. 5778], 1996.
Argys, L. M., and D. I. Rees. “Searching for Peer Group Effects: A Test of the Contagion Hypothesis.” Working Paper, University of Colorado-Denver, 2006.
Bearman, P. S., J. Jones, and U. J. Richard. “The National Longitudinal Study of Adolescent Health: Research Design.” http:// www.cpc.unc.edu/projects/addhealth/ design.html, 1997. [accessed Jan 8, 2007]
Becker, G. A Treatise on the Family. Cambridge, MA: Harvard University Press, 1980.
Belts, J., and D. Morrell. “The Determinants of Undergraduate Grade Point Average,” The Journal of Human Resources, 34, 1999, 268- 293.
Billy, J. O. G., N. S. Landale, W. R. Grady, and D. D. Zimmerle. “Effects of Sexual Activity on Adolescent Social and Psychological Development.” Social Science Quarterly, 51, 1988, 190-212.
Brewer, D. J., E. Eide, and R. G. Ehrenberg. “Does It Pay to Attend an Elite Private College? Evidence on the Effects of College Type on Earnings.” Journal of Human Resources, 34, 1999, 104-23.
Bronars, S., and J. Grogger. “The Economic Consequences of Unwed Motherhood: Using Twin Births as a Natural Experiment.” American Economic Review, 84, 1994, 1141-56.
Brooke, J. S., E. B. Balka, T. Abernathy, and B. A. Hamburg. “Sequence of Sexual Behavior and its Relationship to Other Problem Behaviors in African American and Puerto Rican Adolescents.” Journal of Genetic Psychology, 1551, 1994, 107-14.
Capaldi, D. M., L. Crosby, and M. Stoolmiller. “Predicting the Timing of First Sexual Intercourse for AtRisk Adolescent Males.” Child Development, 67, 1996, 344-59.
Chantala, K. Introduction to Analyzing Add Health Data. Carolina Population Center, UNC-Chapel HiU. http:// www.cpc.unc.edu/projects/ addhealth/files/analyze. pdf, 2003. [accessed Jan 8, 2007]
Cohn, E., S. Cohn, D. Balch, and J. Bradley. “Determinants of Undergraduate GPAs: SAT Scores, High School GPA, and High School Rank.” Economics of Education Review, 23, 2004, 577-86.
Costa, F. M., R. Jessor, J. E. Donovan, and J. D. Fortenberry. “Early Initiation of Sexual Intercourse: The Influence of Psychological Unconventionality.” Journal of Research on Adolescence, 5, 1995, 93-121.
Cragg, J. “Using Higher Moments to Estimate the Simple Errors-in- Variables Model.” Rand Journal of Economics, 28, 1997, S71-91.
Cutler, D. M., E. L. Glaeser, and K. E. Norberg. “Explaining the Rise in Youth Suicide.” National Bureau of Economic Research Working Paper No. 7713, 2001.
Dagenais, M. G., and D. L. Dagenais. “Higher Moment Estimators for Linear Regression Models with Errors in the Variables.” Journal of Econometrics, 76, 1997, 193-222.
Diggs, J. R. “Why Congress Should Reauthorize Title V.” Family Research Council, http://www.frc.org/get. cfm?i=PD02D3ampv=PRINT, 2002. [accessed Jan 8, 2007]
Donovan, J. E., R. lessor, and F. M. Costa. “Syndrome of Problem Behavior in Adolescence: A Replication.” Journal of Consulting and Clinical Psychology, 56, 1988, 762-5.
Dorius, G. L., T. B. Heaton, and P. Steffen. “Adolescent Life Events and Their Association with the Onset of Sexual Intercourse.” Youth and Society, 25, 1993, 3-23.
Eisenberg, D. “Peer Effects for Adolescent Substance Use: Do They Really Exist?” Working Paper, University of California-Berkeley, 2004.
Elliott, D., and B. J. Morse. “Delinquency and Drug Use as Risk Factors in Teenage Sexual Activity.” Youth and Society, 21, 1989, 32- 60.
Erickson, T., and T. Whited. “Two-Step GMM Estimation of the Errors-in-Variables Model Using Higher Order Moments.” Econometric Theory, 18, 2002, 776-9. Evans, W. N., W. E. Dates, and R. M. Schwab. “Measuring Peer Group Effects: A Study of Teenage Behavior.” Journal of Political Economy, 100(5), 1992, 966-91.
Farrel, A. D., S. J. Danish, and C. W. Howard. “Relationship between Drug Use and Other Problem Behaviors in Urban Adolescence.” Journal of Consulting and Clinical Psychology, 60, 1992, 705-12.
Feenstra, B. “New Product Varieties and the Measurement of International Prices Measurement Error Models.” American Economic Review, 84, 1994, 157-77.
Gaviria, A., and S. Raphael. “School-Based Peer Effects and Juvenile Behavior.” The Review of Economics and Statistics, 83, 2001, 257-68.
Grogger, J., and E. Eide. “Changes in College Skills and the Rise of the College Wage Premium.” The Journal of Human Resources, 30, 1995, 280-310.
Harvey, S. M., and C. Spigner. “Factors Associated with Sexual Behavior among Adolescents: A Multivariate Analysis.” Adolescence, 30, 1995, 253-64.
Heckman, J., and E. Vytlacil. “Instrumental Variables Methods for the Correlated Random Coefficient Model.” Journal of Human Resources, 33, 1998, 974-87.
Hoffman, S. D., E. M. Foster, and F. F. Furstenberg. “Re- Evaluating the Costs of Teenage Childbearing.” Demography, 30, 1993, 1-13.
Hotz, V. J., S. McElroy, and S. Sanders. “Teenage Childbearing and Its Life Cycle Consequences: Exploiting a Natural Experiment.” Journal of Human Resources, 40, 2005, 683-715.
Hotz, V. J., C. Mullin, and S. Sanders. “Bounding Causal Effects Using Data from a Contaminated Natural Experiment: Analyzing the Effects of Teenage Childbearing.” Review of Economic Studies, 64, 1997, 576-603.
Jessor, R., F. Costa, L. Jessor, and J. E. Donovan. “Time of First Intercourse: A Prospective Study.” Journal of Personality and Social Psychology, 44, 1983, 608-28.
Jessor, S., and R. Jessor. “Transition from Virginity to Non- Virginity among Youth: A Social Psychological Study over Time.” Developmental Psychology, 11, 1975, 473-84.
King, M., E. Sentana, and S. Wadhwani. “Volatility and Links between National Stock Markets.” Econometrica, 62, 1994, 901-23.
Klein, R., and F. Vella. “Identification and Estimation of the Triangular Simultaneous Equations Model in the Absence of Exclusion Restrictions Through the Presence of Heteroskedasticity.” Working paper at Rutgers University, NJ, 2003.
Klepinger, D., S. Lundberg, and R. Plotnick. “How Does Adolescent Fertility Affect the Human Capital and Wages of Young Women?” Journal of Human Resources, 34, 1999, 421-48.
Learner, E. “Is it a Demand Curve or is it a Supply Curve? Partial Identification Through Inequality Constraints.” Review of Economics and Statistics, 63, 1981, 319-27.
Lewbel, A. “Constructing Instruments for Regressions with Measurement Error When No Additional Data are Available, with an Application to Patents and R&D.” Econometrica, 65, 1997, 1201-13.
_____. “Using Heteroskedasticity to Identify and Estimate Mismeasured and Endogenous Regressor Models.” Boston College Working Papers in Economics 587, Boston College, 2006.
Manski, C. F., and D. A. Wise. College Choice in America. Cambridge, MA: Harvard University Press, 1983.
McLean, A. L., and B. J. Flanigan. “Transition Marking Behaviors of Adolescent Males at First Intercourse.” Adolescence, 28, 1993, 579-95.
Meilman, P. W. “Alcohol Induced Sexual Behavior on Campus.” Journal of American College Health, 42, 1993, 27-31.
Miller, B. C., and K. R. Sneesby. “Educational Correlates of Adolescents’ Sexual Attitudes and Behavior.” Journal of Youth and Adolescence, 17, 1988, 521-30.
Mott, F. L., and W. Marsiglio. “Early Childbearing and Completion of High School.” Family Planning Perspectives, 17, 1985, 234-7.
Nye, F. J. “Choice, Exchange and the Family, ” in Contemporary Theories about the Family, Vol. 2, edited by W. R. Burr, R. Hill, I. F. Nye, and I. L. Reiss. New York: The Free Press, 1979, 1-41.
Peterson, J. L., K. A. Moore, and F. F. Furstenberg. “Television Viewing and Early Initiation of Sexual Intercourse: Is there a Link?” Journal of Homosexuality, 21, 1991, 93-118.
Rector, R. E., and K. A. Johnson. Teenage Sexual Abstinence and Academic Achievement. A Report of the Heritage Center for Data Analysis, Washington, DC: Heritage Foundation, 2005. http:// www.heritage. org/Research/Whitepaper10272005-1.cfm [accessed Jan 8, 2007]
Rector, R. E., K. A. Johnson, and L. R. Noyes. 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, 2003. http://www.neritage.org/Research/ Abstinence/cda0304.cfm [accessed Jan 8, 2007]
Rees, D. I., L. M. Argys, and S. L. Averett. “New Evidence on the Relationship between Substance Use and Adolescent Sexual Behavior.” Journal of Health Economics, 20, 2001, 835-45.
Rigobon, R. “The Curse of Non-Investment Grade Countries.” Journal of Development Economics, 69, 2002, 423-49.
_____. “Identification Through Heteroskedasticity.” Review of Economics and Statistics, 85, 2003, 777-92.
Rose, H., and J. Betts. “The Effect of High School Courses on Earnings.” The Review of Economics and Statistics, 86, 2004, 497- 513.
Rosenbaum, E., and D. B. Kandel. “Early Onset of Adolescent Sexual Behavior and Drug Involvement.” Journal of Marriage and the Family, 52, 1990, 783-98.
Rosenzweig, M., and K. Wolpin. “Sisters, Siblings and Mothers: The Effects of Teen-Age Childbearing on Birth Outcomes in a Dynamic Context.” Econometrica, 63, 1995, 303-26.
Rothstein, D. S. “High School Employment and Youths’ Academic Achievement. Journal of Human Resources, 42(1), 2007, 194-213.
Rummery, S., F. Vella, and M. Verbeek. “Estimating the Returns to Education for Australian Youth via Rank-Order Instrumental Variables,” Labour Economics, 6, 1999, 491-507.
Sabia, J. J. “Does Early Adolescent Sex Cause Depressive Sym