You are on page 1of 40

The Effects of Sex Education on Teen Sexual Activity and Teen Pregnancy

Author(s): Gerald S. Oettinger


Source: Journal of Political Economy , Vol. 107, No. 3 (June 1999), pp. 606-644
Published by: The University of Chicago Press
Stable URL: https://www.jstor.org/stable/10.1086/250073

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access
to Journal of Political Economy

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
The Effects of Sex Education on Teen Sexual
Activity and Teen Pregnancy

Gerald S. Oettinger
University of Texas at Austin

This paper empirically examines the relationship between enroll-


ment in sex education and subsequent sexual behavior for U.S.
teenagers during the 1970s. The estimates indicate that enrollment
in sex education was associated with earlier sexual activity for fe-
males in this cohort. Sex education also was associated with earlier
pregnancy for some groups of females, but these effects are smaller
and not always statistically significant. For both types of transitions,
the effect of sex education appears to have been larger for women
with fewer alternative sources of sexual information. In contrast,
sex education had much less impact on male transitions into sexual
activity. Within-family analyses using sibling data reveal qualita-
tively similar patterns. Overall, the evidence suggests that sex edu-
cation in the 1970s had some causal impact on teen sexual behav-
ior, probably in significant part by providing information that
enabled teens to alter the risks of sexual activity.
Ignorance is not bliss. . . . Children who know more,
whose parents have taught them about these things,
or if they’ve learned it at school, are far less likely to
engage in early sexual activity. And if they do, they’re
far more likely to be responsible, to use appropriate
methods of contraception or control. It’s the children
who have not had this kind of education who get in
trouble. [Joycelyn Elders, former U.S. Surgeon
General; interview in Progressive (March 1995)]
I thank Steve Bronars, Melissa Famulari, Dan Hamermesh, Preston McAfee, two
anonymous referees, and seminar participants from Texas, Texas A&M, Rice, Hous-
ton, and the National Bureau of Economic Research 1995 Summer Institute for
helpful comments. All errors are my own.

[Journal of Political Economy, 1999, vol. 107, no. 3]


 1999 by The University of Chicago. All rights reserved. 0022-3808/99/0703-0006$02.50

606

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 607
Rarely has there been a more thoroughly discredited
idea than the notion that widespread availability of
contraceptives and knowledge of their use leads to a
diminished rate of teen pregnancy and abortion. To-
day, with much more sex education and availability of
birth control than there was 25 years ago, we also have
higher rates of teen pregnancy, soaring illegitimacy,
and a much higher abortion rate. [Ralph Reed, for-
mer executive director of the Christian Coalition;
Washington Times (February 14, 1995)]

I. Introduction
The last several decades have witnessed a dramatic rise in the rates of
sexual activity and pregnancy among unmarried teens in the United
States (Zelnik and Kantner 1980; Hofferth, Kahn, and Baldwin 1987;
Forrest and Singh 1990), despite a significant expansion of school-
based sex education (Kenney 1989; Kirby et al. 1994). At least in
part because of these trends, a contentious debate rages about the
effects of sex education on teenage sexual behavior. Critics of sex
education, in addition to their moral objections to these programs,
claim that sex education leads to higher rates of teen sexual activity
and teen pregnancy. In contrast, advocates of sex education argue
that such programs do not encourage sexual activity, but rather re-
duce the risks of illness and unwanted pregnancy among teens who
are already sexually active.
The present paper subjects these competing claims to empirical
scrutiny. I analyze data from the National Longitudinal Survey of
Youth (NLSY) on individuals who were in their teen years during
the 1970s and early 1980s to see how enrollment in sex education
affects hazard rates into sexual activity and pregnancy for teenagers.
Not surprisingly, the evidence suggests that the truth lies between
the more extreme views on either side of the debate.
In particular, I find that prior enrollment in sex education was
associated with a higher hazard rate into sexual activity for females
in this cohort. I also find that prior enrollment in sex education may
have been associated with a higher pregnancy hazard rate for certain
groups of females, but these effects are not always significant and
are always considerably smaller than the effects for transitions into
sexual activity. For both types of transitions, I find that the effects
of sex education were larger among women who appear to have had
fewer alternative sources of sexual information. In contrast, I find
little relationship between sex education and male transitions into

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
608 journal of political economy
sexual activity. I find similar patterns in within-family analyses that
use data on sibling pairs to control for unobserved family or commu-
nity effects that might be correlated with enrollment in sex educa-
tion. Overall, the evidence suggests that sex education in the 1970s
probably had some causal influence on teen sexual behavior and
that the provision of information that enabled teens to alter the risks
of sexual activity was probably an important source of this causal
influence. Finally, simulations suggest that universal sex education
in the midteen years probably would have had little effect on teen
sexual behavior but that universal sex education in the preteen years
might have led to significant increases in female sexual activity and
pregnancy in the early and middle teenage years.
This paper is not the first to investigate the effects of sex education
on teen sexual behavior.1 Several quasi-experimental evaluations of
specific sex education curricula have been undertaken recently, but
the results from these studies shed little light on the effects of sex
education as it has been implemented in the past and their general-
izability is unclear. Four studies have examined the effects of sex
education using broad-based individual survey data with informa-
tion on the timing of sex education relative to sexual activity and
pregnancy. Dawson (1986) and Marsiglio and Mott (1986), in analy-
ses of teen females, both find some evidence that prior enrollment
in sex education is associated with earlier transitions into sexual ac-
tivity but not into pregnancy. Ku, Sonenstein, and Pleck (1992,
1993), in analyses of a more recent cohort of teen males, find that
instruction on reproductive biology and contraceptive methods is
associated with earlier transitions into sexual activity whereas instruc-
tion on ‘‘resistance skills’’ is associated with later transitions, less
frequent sexual activity, and fewer sexual partners.
Although my results do not differ dramatically from those of Daw-
son (1986) and Marsiglio and Mott (1986), the current study makes
several new contributions. First, in contrast to all of the previous
research, I develop a theoretical model of a teen’s decision to be
sexually active and use it to derive the impacts of several different
types of sex education. The model also generates several simple pre-
dictions about how the magnitude of the effect of sex education
should vary with certain individual characteristics, and, as I explain
later, these predictions provide some leverage for distinguishing
true effects of sex education from spurious effects due to unobserved
heterogeneity or reporting bias. Second, as mentioned above, I ex-
ploit the presence of siblings in the NLSY data to estimate models

1
Kirby et al. (1994) exhaustively review the existing empirical evidence on the
effects of sex education.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 609
that control for unobserved fixed family or community determinants
of sexual activity or pregnancy that may be correlated with enroll-
ment in sex education. While previous authors have acknowledged
the potential importance of such unobserved heterogeneity, none
has dealt with it empirically. Finally, I characterize the effects of sex
education on teen sexual behavior more completely than previous
studies by estimating models for both males and females, by ex-
tending the analysis to the early teen years, and by conducting simu-
lations of various sex education policies to gauge the magnitudes of
the estimated effects of sex education.
The remainder of the paper is organized as follows. Section II
outlines a simple model of teen sexual behavior and uses it to analyze
the effects of different types of sex education. Section III develops
the econometric methodology used in the empirical analysis. Sec-
tion IV presents and interprets the empirical results. Finally, Section
V presents a conclusion.

II. Theory
Rational individuals become sexually active at the first age at which
the perceived benefits from sex exceed the perceived costs. Apart
from immediate gratification, sex might produce benefits by advanc-
ing an ongoing relationship, by enhancing social status within the
peer group, or by producing knowledge. On the other hand, sex
might impose costs by adversely affecting an ongoing relationship,
by prompting parental or community sanction, or by causing illness
or undesired pregnancy.2 Clearly, sex education will influence the
decision to be sexually active only insofar as it alters the expected
costs or benefits. Thus the absolute magnitude of the effect of sex
education on any given teen’s sexual behavior should rise with the
amount of new information provided by such instruction. Without
further assumptions about individuals’ prior information and the
exact content of sex education, however, theory is silent about the
sign of the effect of sex education.
To explore the possible effects of sex education, I consider a
model in which teens make a discrete choice between sexual activity
and abstinence at each age. If a teen chooses to be sexually active,
then utility (i.e., the sum of the current-period utility flow and ex-
pected discounted lifetime utility as of next period) is uncertain be-
cause the consequences of sexual activity—for example, whether a
pregnancy occurs—are also uncertain. In contrast, if a teen chooses

2
Obviously, the risk of pregnancy is a net benefit of sexual activity if a child is
desired.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
610 journal of political economy
abstinence, then the risks attendant with sexual activity are avoided
and utility is nonstochastic. I assume that decisions about sexual ac-
tivity maximize expected utility given the perceived probability distri-
bution over the outcomes of sexual activity.
Anticipating the empirical work, I assume that sexual activity intro-
duces uncertainty only through a risk of pregnancy. Thus there are
two possible states of the world in the next period if sexual activity
is chosen: ‘‘pregnancy’’ and ‘‘no pregnancy.’’3 For a sexually active
individual, let the sum of current-period utility and expected dis-
counted future utility as of next period be U 0 if a pregnancy occurs
and U 1 if a pregnancy does not occur. Let the sum of current-period
utility and expected discounted future utility as of next period be V
for an individual who abstains from sex. Finally, let p denote the
perceived probability that sexual activity results in pregnancy, where
0 ⱕ p ⱕ 1. A teen chooses to be sexually active if and only if pU 0 ⫹
(1 ⫺ p)U 1 ⬎ V. Equivalently, with Z ⬅ (U 1 ⫺ V )/(U 1 ⫺ U 0 ), teens
with U 1 ⬎ (⬍) U 0 choose to be sexually active if and only if p ⬍
(⬎) Z.
Before I proceed, several remarks are in order. First, heterogene-
ity in preferences and the evolution of preferences with age imply
that the indirect utilities U 0 , U 1 , and V vary both across individuals
of a given age and over time for a given individual. Likewise, hetero-
geneity in fecundity and reproductive knowledge and the evolution
of these variables with age imply that p varies both across individuals
of a given age and over time for a given individual. Thus the model
predicts cross-section and age-related variation in sexual behavior.
Second, the model abstracts from the decision of frequency of sexual
activity because data on the frequency of teenage sexual activity are
not available in the NLSY.4 I have analyzed a model that incorporates
a choice of frequency of sexual activity, however, and the predicted
effects of (different types of) sex education are quite similar to those
derived below for the simpler model. Third, I assume that p is exoge-
nous, but I have analyzed a model in which teens can alter p through
a choice of costly preventive effort and the predicted effects of (dif-
ferent types of) sex education are basically the same as those derived
below for the simpler model.5 Finally, I initially assume that each

3
In principle, the model could be extended easily to allow sexual activity to intro-
duce other risks (e.g., illness, termination of the relationship, and parental sanc-
tion). I ignore these other risks, however, because pregnancy is the only outcome
of sexual activity for which data are available in the NLSY.
4
The NLSY actually did ask questions about the current frequency of sexual activity
in the 1983–85 interviews, but most sample participants were well past their teen
years by this time.
5
In both extensions, one must model the impact of sex education on a teen’s
opportunity set a bit differently than in the simple model analyzed in the text. Details
of the extended models are available from the author on request.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 611
TABLE 1
Predicted Fractions of Teens Who Are Sexually Active and Who Become
Pregnant for Alternative Rankings of U 0, U 1, and V

Utility Predicted Fraction of Teens Predicted Fraction of Teens Who


Ordering Who Are Sexually Active Become Pregnant
U1 ⬎ V ⬎ U0 ∫0Z dF(p) ⫽ F(Z ) ∫0Z pdF(p) ⫽ ZF(Z) ⫺ ∫0Z F(p)dp
U1 ⬎ U0 ⬎ V ∫01 dF(p) ⫽ 1 ∫01 pdF(p) ⫽ 1 ⫺ ∫01 F(p)dp
V ⬎ U1 ⬎ U0 0 0
V ⬎ U0 ⬎ U1 0 0
U0 ⬎ U1 ⬎ V ∫01 dF(p) ⫽ 1 ∫01 pdF(p) ⫽ 1 ⫺ ∫01 F(p)dp
U0 ⬎ V ⬎ U1 ∫Z dF(p) ⫽ 1 ⫺ F(Z )
1
∫Z pdF(p) ⫽ 1 ⫺ ZF(Z ) ⫺ ∫Z1 F(p)dp
1

Note.—Z ⬅ (U 1 ⫺ V )/(U 1 ⫺ U 0).

teen correctly perceives her value of p, but I relax this assumption


later when considering the effects of sex education that reveals the
true value of p.
The fractions of teens who choose to be sexually active and who
become pregnant depend on the joint distribution of U 0 , U 1 , V, and
p in the teen population, but it is useful to begin by analyzing be-
havior conditional on U 0 , U 1 , and V. Thus let F(p) denote the (con-
ditional) cumulative distribution function for p, with F(0) ⫽ 0,
F(1) ⫽ 1, and F(p) weakly increasing for p ∈ (0, 1). There are six pos-
sible orderings of U 0 , U 1 , and V, as shown in table 1. Teens with
V ⫽ max(U 0 , U 1 , V ) prefer abstinence to sex regardless of the risk
of pregnancy. In contrast, teens with V ⫽ min(U 0 , U 1 , V ) prefer to
be sexually active for any risk of pregnancy. Finally, teens with U 1 ⬎
V ⬎ U 0 choose to be sexually active only if the pregnancy risk is
sufficiently small (i.e., p ⬍ Z ), whereas teens with U 0 ⬎ V ⬎ U 1
choose to be sexually active only if the risk of pregnancy is suffi-
ciently large (i.e., p ⬎ Z). Table 1 gives expressions for the fractions
of teens who are sexually active and who experience a pregnancy
for each possible ordering of U 0 , U 1 , and V. The unconditional frac-
tions of teens who are sexually active and who become pregnant are
simply weighted averages of the conditional proportions shown in
table 1 and would be obtained by integrating over the joint density
function for U 0 , U 1 , and V.
The fractions of teens who are sexually active and who become
pregnant, conditional on U 0 , U 1 , and V, also have a convenient
graphical representation. Figure 1 depicts a case in which U 1 ⬎ V ⬎
U 0 . The fraction of teens who are sexually active, ∫0Z dF(p) ⫽ F(Z ),
is given by the height of the distribution function at p ⫽ Z. The
fraction of teens who become pregnant, ∫0Z pdF(p) ⫽ ZF(Z ) ⫺
∫0Z F(p)dp, is given by area A to the left of the distribution function
between 0 and F(Z ). Analogous diagrams exist for the other rank-

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
612 journal of political economy

Fig. 1.—Fraction of teens who are sexually active and fraction of teens who be-
come pregnant.

ings of U 0 , U 1 , and V. Given an assumption about how sex education


alters the basic parameters (U 0 , U 1 , V, or p) describing a teen’s
choice problem, one can use these diagrams to illustrate how sex
education affects rates of teen sexual activity and teen pregnancy.
I analyze the effects of three distinct types of sex education, which
I refer to as ‘‘risk-altering,’’ ‘‘utility-altering,’’ and ‘‘risk-revealing’’
sex education. Risk-altering sex education provides information that
allows sexually active teens to alter the risk of pregnancy (p). Courses
that discuss the menstrual cycle or the proper usage and efficacy of
alternative contraceptive methods would fall in this category. Utility-
altering sex education changes the (expected) utilities (U 0 , U 1 , or
V ) that teens receive in different states of the world. Courses that
teach strategies for resisting sex (thereby increasing V ), that high-
light the costs of teen parenthood (presumably reducing U 0 ), or that
provide information about abortion (likely increasing U 0 ) are lead-
ing examples of this type of sex education. Finally, risk-revealing sex
education provides accurate information to teens who initially may
misjudge pregnancy risks. Instruction on contraceptive methods
would be purely risk-revealing if teens know the correct usage and
relative effectiveness of all types of contraception but are misin-
formed about absolute effectiveness. In practice, all three types of

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 613

Fig. 2.—Effects of risk-altering sex education

sex education may coexist, and any given sex education course might
combine all three archetypes.
Consider first the effects of risk-altering sex education on teen
sexual activity and teen pregnancy. Obviously, the provision of risk-
altering sex education has no effect on sexual activity or pregnancy
among teens who prefer abstinence to sex for any p (i.e., V ⬎ U 0 ⬎
U 1 or V ⬎ U 1 ⬎ U 0 ). These teens all abstain whether or not they
receive sex education.
For teens who prefer sexual activity for at least some values of p
but wish to avoid a pregnancy (i.e., U 1 ⬎ V ⬎ U 0 or U 1 ⬎ U 0 ⬎
V ), enrollment in risk-altering sex education should reduce (at least
weakly) the risk of pregnancy from any given level of sexual activity.
This reduction in risk is shown in figure 2 by the leftward shift in
the pregnancy risk distribution function from F(p) to G(p). For
teens with U 1 ⬎ V ⬎ U 0 , who are sexually active only if p ⬍ Z, the
reduction in risk brought about by sex education increases the frac-
tion who are sexually active from F(Z ) to G(Z ). The effect on the
pregnancy rate among these teens is ambiguous, however; the re-
duced risk of pregnancy among teens who would have been sexually
active without sex education lowers the fraction who become preg-
nant by area A, but the induced increase in the fraction of teens
who are sexually active raises the fraction who become pregnant by

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
614 journal of political economy
area B. The ambiguous effect of sex education on pregnancy among
teens with U 1 ⬎ V ⬎ U 0 is exactly analogous to the ambiguous effect
of automobile safety improvements on auto-related injuries dis-
cussed by Peltzman (1975). In contrast, teens with U 1 ⬎ U 0 ⬎ V are
sexually active for all values of p, and therefore risk-altering sex edu-
cation has no effect on the fraction of these teens who are sexually
active but unambiguously reduces the fraction who become preg-
nant by area A ⫹ C.
Finally, for teens who prefer sexual activity for at least some values
of p and who desire a pregnancy (i.e., U 0 ⬎ V ⬎ U 1 or U 0 ⬎ U 1 ⬎
V ), risk-altering sex education should increase (at least weakly) the
risk of pregnancy from any given level of sexual activity. This in-
creased risk can be seen in figure 2 as a rightward shift in the preg-
nancy risk distribution function from G(p) to F(p). Thus, for teens
with U 0 ⬎ V ⬎ U 1 , who are sexually active only if p ⬎ Z, risk-altering
sex education raises the fraction who choose to be sexually active
from 1 ⫺ G(Z ) to 1 ⫺ F(Z ) and raises the fraction who become
pregnant by area B ⫹ C. For teens with U 0 ⬎ U 1 ⬎ V, who are sexually
active for any value of p, risk-altering sex education cannot raise the
fraction who are sexually active, but it does raise the fraction who
become pregnant by area A ⫹ C.
The predicted effects of risk-altering sex education are summa-
rized in column 2 of table 2. Risk-altering sex education clearly is
expected to increase the overall fraction of teens who are sexually
active. However, both the sign of the effect on the fraction of teens
who become pregnant and the relative magnitude of the effect on
sexual activity versus the effect on pregnancy are theoretically ambig-
uous and depend on the joint distribution of U 0 , U 1 , and V in the
teen population.6
Next, consider the effects of utility-altering sex education on teen
sexual behavior. It is clear that an increase (decrease) in U 0 , an in-
crease (decrease) in U 1 , or a decrease (increase) in V all have qualita-
tively identical effects since each raises (lowers) the expected utility
of sexual activity relative to abstinence. Suppose first that utility-
altering sex education raises the relative expected utility from sexual
activity. This has no effect on sexual activity or pregnancy among
teens whose utility orderings before sex education are U 1 ⬎ U 0 ⬎ V
or U 0 ⬎ U 1 ⬎ V since these youths are all sexually active even in the

6
The fact that most individuals become sexually active during their teen years
implies that the orderings V ⬎ U 1 ⬎ U 0 and V ⬎ U 0 ⬎ U 1 are rare by the late teen
years and that the orderings U 1 ⬎ U 0 ⬎ V and U 0 ⬎ U 1 ⬎ V must be rare in the
early teen years. In addition, most sexually active teen females use contraception
and most pregnant teen females classify their pregnancies as unintended (Forrest
and Singh 1990), which implies that U 1 ⬎ U 0 for most sexually active teen females.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 615
TABLE 2
Predicted Effects of Different Types of Sex Education on Rates of Teen
Sexual Activity and Teen Pregnancy for Alternative Rankings of U 0, U 1,
and V and in Overall Teen Population

Type of Sex Education


Utility-Altering Risk-Revealing
Risk-
Utility Impact on Altering Case 1 Case 2 Case 1 Case 2
Ordering (1) (2) (3) (4) (5) (6)
U1 ⬎ V ⬎ U0 % sexually active ⫹ ⫹ ⫺ ⫺ ⫹
% had a pregnancy ? ⫹ ⫺ ⫺ ⫹
U1 ⬎ U0 ⬎ V % sexually active 0 0 0 or ⫺ 0 0
% had a pregnancy ⫺ 0 0 or ⫺ 0 0
V ⬎ U1 ⬎ U0 % sexually active 0 0 or ⫹ 0 0 0
% had a pregnancy 0 0 or ⫹ 0 0 0
V ⬎ U0 ⬎ U1 % sexually active 0 0 or ⫹ 0 0 0
% had a pregnancy 0 0 or ⫹ 0 0 0
U0 ⬎ U1 ⬎ V % sexually active 0 0 0 or ⫺ 0 0
% had a pregnancy ⫹ 0 0 or ⫺ 0 0
U0 ⬎ V ⬎ U1 % sexually active ⫹ ⫹ ⫺ ⫹ ⫺
% had a pregnancy ⫹ ⫹ ⫺ ⫹ ⫺
Overall % sexually active ⫹ ⫹ ⫺ ? ?
% had a pregnancy ? ⫹ ⫺ ? ?

Note.—Risk-altering sex education provides information that allows teens to alter the pregnancy risk associ-
ated with sexual activity. Utility-altering sex education provides information that alters the utilities that teens
receive in the different states: in case 1, the information raises the relative utility of sexual activity, and in
case 2 it reduces the relative utility of sexual activity. Risk-revealing sex education provides information that
alters the perceived pregnancy risk from sexual activity (because teens are assumed to be misinformed about
the true risks): in case 1, the information raises the perceived pregnancy risk, and in case 2 it reduces the
perceived pregnancy risk. When two effects are listed in the columns for utility-altering sex education, the
first gives the effect of a very small change in the utilities and the second gives the effect of a change in the
utilities that is large enough to change the individual’s utility ordering.

absence of sex education. In contrast, the fraction of sexually active


teens in the groups with U 1 ⬎ V ⬎ U 0 or U 0 ⬎ V ⬎ U 1 increases, and
since pregnancy risks are unchanged, the fraction of teens in these
groups who become pregnant increases as well. The effect of utility-
altering sex education on teens with U 1 ⬎ V ⬎ U 0 is shown graphi-
cally in figure 3, where participation in sex education raises Z ⬅ (U 1
⫺ V )/(U 1 ⫺ U 0 ) from Z 1 to Z 2 , thereby increasing the fraction of
sexually active teens from F(Z 1 ) to F(Z 2 ) and increasing the fraction
of teens who become pregnant by area A.7 Finally, sex education that
raises the relative utility of sexual activity increases sexual activity and
pregnancy among teens who initially abstain for any value of p (i.e.,
7
The graph in fig. 3 assumes that the change in indirect utilities induced by sex
education is small enough that the utility ordering remains U 1 ⬎ V ⬎ U 0 after the
provision of sex education. Figure 3 also illustrates the effect of utility-altering sex
education on teens with U 0 ⬎ V ⬎ U 1 . For these teens, sex education causes Z to
decrease from (say) Z 2 to Z 1 , causes the fraction of sexually active teens to increase
from 1 ⫺ F(Z 2 ) to 1 ⫺ F(Z 1 ), and causes the fraction of teens who become pregnant
to rise by area A.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
616 journal of political economy

Fig. 3.—Effects of utility-altering sex education

V ⬎ U 1 ⬎ U 0 or V ⬎ U 0 ⬎ U 1 ) if the change in indirect utilities is


large enough so that V ≠ max(U 0 , U 1 , V ), but otherwise it has no
effect.
The effects of utility-altering sex education that raises the relative
utility of sexual activity on the fractions of teens who are sexually
active and who become pregnant are summarized in column 3 of
table 2. Unlike risk-altering sex education, the fractions of teens in
the overall population who are sexually active and who experience
a pregnancy both increase unambiguously in response to this type
of sex education. The analysis for utility-altering sex education that
reduces the relative utility of sexual activity is entirely analogous and
is therefore not presented, but its effects are listed in column 4 of
table 2. Not surprisingly, utility-altering sex education that reduces
the relative utility of sexual activity lowers both sexual activity and
pregnancy in the overall teen population.
Finally, consider the effects of risk-revealing sex education on
teen sexual behavior. Learning that the actual pregnancy risk differs
from the initially perceived risk clearly has no effect on the fractions
who are sexually active or pregnant for teens who abstain from sex-
ual activity for any value of p (i.e., V ⬎ U 1 ⬎ U 0 or V ⬎ U 0 ⬎ U 1 ).
Likewise, risk-revealing sex education does not affect these fractions
for teens who are sexually active for any value of p (i.e., U 1 ⬎ U 0 ⬎

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 617

Fig. 4.—Effects of risk-revealing sex education

V or U 0 ⬎ U 1 ⬎ V ). Risk-revealing sex education does affect the


behavior of teens with U 1 ⬎ V ⬎ U 0 or U 0 ⬎ V ⬎ U 1 , however, who
are sexually active only for certain values of p. For simplicity, I exam-
ine its effects when all teens initially underestimate the risk of preg-
nancy and when all teens initially overestimate the risk of preg-
nancy.8 If all teens initially underestimate (overestimate) the risks
of pregnancy, then risk-revealing sex education will decrease (in-
crease) the fractions sexually active and pregnant for teens with U 1 ⬎
V ⬎ U 0 but will increase (decrease) the fractions sexually active and
pregnant for teens with U 0 ⬎ V ⬎ U 1 . Figure 4 illustrates the effects
of risk-revealing sex education on teens with U 1 ⬎ V ⬎ U 0 who under-
estimate the risk of pregnancy prior to sex education; the provision
of sex education changes the perceived risk distribution from F(p)
to G(p), reduces the fraction of sexually active youths from F(Z ) to
G(Z ), and reduces the fraction of youths who become pregnant by
area A.9
8
Clearly, neither set of beliefs is consistent with rational expectations. Many ob-
servers, however, believe that teens systematically underestimate the risks associated
with sexual activity. On the other hand, there is evidence that individuals tend to
overestimate the actual risks of unlikely events (e.g., Viscusi 1990).
9
The effects of risk-revealing sex education if teens initially overestimate pregnancy
risks can also be seen in fig. 4 if G(p) is interpreted as the initial perceived risk
distribution and F(p) is interpreted as the true risk distribution. The effects of risk-

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
618 journal of political economy
Columns 5 and 6 of table 2 summarize the effects of risk-revealing
sex education on teen sexual activity and teen pregnancy for the
case in which all teens underestimate pregnancy risks and the case
in which all teens overestimate pregnancy risks, respectively. In both
cases, the effects in the overall teen population depend on the rela-
tive sizes of the populations satisfying U 1 ⬎ V ⬎ U 0 and U 0 ⬎ V ⬎
U 1 . Both introspection and the evidence cited in note 6 suggest that
the former group is much larger, which implies that risk-revealing
sex education will reduce (increase) both sexual activity and preg-
nancy if all teens initially underestimate (overestimate) the risk of
pregnancy. Of course, a more plausible specification of initial risk
perceptions is that they are correct for the average teen but that
some teens overestimate risks whereas others underestimate risks.
In this case, there are no clear predictions about the signs of the
effects of risk-revealing sex education on teen sexual activity and
teen pregnancy. Even in this case, though, it seems likely that sex
education should have a similar effect on both outcomes.
In summary, risk-altering sex education should increase teen sex-
ual activity, but its effect on teen pregnancy is ambiguous. In con-
trast, utility-altering sex education that raises (lowers) the relative
utility of sexual activity should increase (decrease) both teen sexual
activity and teen pregnancy. Likewise, under the plausible assump-
tion that most sexually active teens prefer to avoid a pregnancy, risk-
revealing sex education that causes teens to revise the perceived risk
of pregnancy downward (upward) should increase (decrease) both
teen sexual activity and teen pregnancy. The predicted effects of the
different types of sex education are all essentially unchanged in the
extended model that allows teens to choose the frequency of sexual
activity.10 Of course, since in practice sex education probably com-

revealing sex education on teens satisfying U 0 ⬎ V ⬎ U 1 can be illustrated in fig. 4


in an analogous manner.
10
In particular, the predicted effects of sex education on the fractions of youths
who are sexually active in the extended model are identical to those shown in table
2. Moreover, one can show that the predicted effects of sex education on frequencies
of sexual activity in the various groups almost always reinforce the effects on the
fractions sexually active. (The only possible exception pertains to the presumably
small number of teens who desire a pregnancy. For them, risk-altering or risk-reveal-
ing sex education that raises the actual or perceived risk of pregnancy can only
increase the fraction who are sexually active but would reduce frequency for any
teens who initially were having sex more frequently than they found enjoyable solely
to raise the odds of a pregnancy.) Thus, since sex education is expected to move
both the fraction active and the frequency of activity among the already active in
the same direction, the effect of sex education that does not alter actual pregnancy
risks on the fraction of teens who become pregnant should have the same sign but
a larger magnitude than the effect on the fraction who are sexually active. In con-
trast, the effect of risk-altering sex education on the fraction of teens who become
pregnant remains ambiguous since the lower risk of pregnancy per episode of sexual

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 619
bines aspects of all these archetypes, theory does not place a priori
sign restrictions on the effects of sex education. Moreover, because
some of the different causal models yield fairly similar predictions,
determining the relative role played by the different types of sex
education appears difficult. Nevertheless, certain empirical find-
ings—for example, a smaller (i.e., less positive) effect of sex educa-
tion on pregnancy than on sexual activity—at least would suggest a
nontrivial role for risk-altering sex education.
In addition to the predictions summarized in table 2, the theory
also suggests that the magnitudes of these effects should vary system-
atically with certain individual characteristics. For example, the the-
ory suggests that sex education should have a larger effect on sexual
activity for teen females than for teen males. To see this, note from
table 2 that the teens whose sexual activity decisions are most respon-
sive to all types of sex education are those with U 1 ⬎ V ⬎ U 0 or U 0 ⬎
V ⬎ U 1 . These inequalities are more likely to hold if the net cost
(benefit) of an undesired (desired) pregnancy is large. Thus one
would expect sex education to have a larger impact on female sexual
activity than on male sexual activity as long as unintended teen preg-
nancies impose larger costs on females.
Recall as well that the theory implies that sex education should
affect sexual activity and pregnancy only to the extent that it presents
new information affecting the expected costs or benefits of sexual
activity. Thus sex education should have a greater impact on teens
who gain a lot of new information from a sex education course,
namely, teens without low-cost alternative sources of sexual informa-
tion. Older teens presumably have better alternative sources of sex-
ual information than younger teens, and therefore the effects of sex
education on sexual activity and pregnancy should be larger in the
early teen years than in the late teen years. Older siblings are another
low-cost source of sexual information, and therefore, other things
equal, the effects of sex education should be larger for teens without
older siblings than for teens with older siblings.
These additional predictions potentially can help distinguish
causal effects of sex education from spurious effects. Spurious effects
of sex education could arise from unobserved heterogeneity be-
tween recipients and nonrecipients of sex education. For example,
unobserved parental attitudes and community norms that indepen-
dently influence teen sexual behavior also might affect whether or
not a youth receives sex education instruction. Spurious effects also
could arise if sex education influences reported teen sexual activity or

intercourse tends to offset the increases in the fraction of sexually active teens and
the frequency of sexual activity.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
620 journal of political economy
if teen sexual activity influences reported enrollment in sex education.
Under any of these scenarios, however, there is no particular reason
why the magnitudes of the effects of sex education should differ
systematically between males and females, between the early teen
years and the late teen years, or, especially, between teens with older
siblings and teens without older siblings. In contrast, the theory de-
veloped above explicitly predicts such systematic differences. Thus
evidence in favor of these auxiliary predictions would lend support
to a causal interpretation of any measured effects of sex education.

III. Data and Empirical Methodology


The empirical analysis of teen sexual activity and teen pregnancy
uses data from the NLSY, a panel survey of American youths born
between 1957 and 1964 that began annual interviews in 1979 with
12,686 individuals. Because households were the primary sampling
unit in the initial survey, all youths born between 1957 and 1964 in
each sampled household were included in the panel. As a result, the
NLSY contains several thousand sibling pairs, and I exploit this as-
pect of the data in the empirical work. The NLSY originally consisted
of a sample representative of the U.S. youth population, a supple-
mental sample of youths from low-income or minority households,
and a small supplemental sample of young people in the military.
The military sample was discontinued after 1984, however, and the
present paper uses data from only the civilian samples.
Although it primarily focuses on labor market activity, the NLSY
asked about the age of first sexual intercourse at the 1983, 1984, and
1985 interviews; the age of enrollment in sex education at the 1984
interview; and the pregnancy history of each female respondent at
each even-year interview starting with 1982.11 I use the data on age
of first sexual activity and age of first pregnancy to estimate models
of the determinants of male transitions into sexual activity and fe-
male transitions into both sexual activity and pregnancy. Since I am
interested in teen sexual behavior, I limit attention to transitions
made at or before age 19. I focus on the main questions suggested
by the theory: whether sex education hastens or delays transitions
into sexual activity and pregnancy, whether the effects of sex educa-
tion are similar for both sexual activity and pregnancy, and whether

11
The NLSY also collected information on the frequency of sexual intercourse in
the previous 30 days in the 1983, 1984, and 1985 interviews and on current contra-
ceptive use in all the even-numbered interview years starting with 1982. Unfortu-
nately, these data provide information about current sexual behavior rather than
sexual behavior during the teenage years.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 621
the magnitudes of the effects of sex education vary with age, sex, or
presence of an older sibling.
I specify a hazard rate model to analyze the determinants of transi-
tion times.12 The hazard model framework easily handles right-cen-
sored spells (in the present problem, individuals who do not become
sexually active or pregnant at or before age 19) and time-varying
covariates (in the present problem, previous enrollment in sex edu-
cation). I adopt the familiar proportional hazards specification and
initially posit that the hazard rate into sexual activity (pregnancy)
for person i at age t, λ i (t), takes the form
λ i (t) ⫽ λ 0 (t)e X i (t )␤ , (1)
where λ 0 (t) is the baseline hazard for all individuals of age t, X i (t)
is a vector of observed covariates for individual i at age t, and ␤
is a parameter vector to be estimated. In the present problem, the
elements of ␤ that measure the effects of prior enrollment in sex
education are of particular interest. Under the assumption that X i (t)
can be treated as fixed within each year of age t, the likelihood func-
tion follows straightforwardly from (1).
I estimate the hazard models using the semiparametric method
developed in Meyer (1990), which does not assume any restrictive
form for the baseline hazard function. Instead, dummy variables are
defined for each possible time of transition, and the vector of esti-
mated coefficients for these dummies determines the shape of the
baseline hazard function.13 The baseline hazard reflects the risk of
transition at each age, given survival until that age, for an individual
with ‘‘baseline’’ characteristics. I also estimated the transition mod-
els using Cox partial likelihood (Cox and Oakes 1984) and using a
logit hazard specification (e.g., Hoynes and MaCurdy 1994) with a
nonparametric baseline hazard and obtained qualitatively identical
estimates of ␤.
The hazard rate specification in (1) assumes that there exists no
unobserved heterogeneity in the risk of teenage transitions into sex-
ual activity or pregnancy. However, unobservable family-specific or
community-specific characteristics, such as attitudes toward teen sex-
ual activity, probably influence teenage sexual behavior. If so, esti-
mates that ignore this unobserved heterogeneity will be biased. In

12
The earlier papers of Dawson (1986) and Marsiglio and Mott (1986) estimate
sequences of logit models, one for each age, for the probability of becoming sexually
active. That approach roughly approximates a hazard model but uses the data less
efficiently.
13
Since the NLSY measures age of first intercourse and age of first pregnancy only
in full years, the time periods in this analysis are one year in length, and separate
baseline transition risks are estimated for each year of age between 10 and 19.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
622 journal of political economy
the hazard model framework, unobserved heterogeneity typically is
assumed to be independent of the explanatory variables (i.e., a ‘‘ran-
dom effect’’) so that estimation simply requires specifying a mar-
ginal distribution for the unobserved component.14 In the current
problem, however, it is likely that the unobserved determinants of
teen sexual behavior are correlated with whether and when a teen-
ager is enrolled in sex education. Since the unobserved heterogene-
ity is not a random effect in this case, I instead consider a fixed-
effects estimation strategy below.
Unobserved determinants of teen sexual behavior might be corre-
lated with enrollment in sex education for at least two reasons. First,
decisions about sex education curricula are made largely at the
school district level (Kenney and Alexander 1980; Kenney 1989),
and local decisions regarding sex education are likely to be corre-
lated with unobserved community influences on the sexual behavior
of local teens. For example, communities that provide early sex edu-
cation might be those that are least disapproving of teen sexual activ-
ity. On the other hand, early sex education might be provided pre-
cisely in communities in which teen sexual activity is rare and seems
unlikely to be encouraged by such instruction. Second, enrollment
in sex education also might be correlated with unobserved family
determinants of teen sexual behavior because virtually all states re-
quire schools either to obtain explicit parental consent before pro-
viding sex education or to excuse students from such instruction on
parental request (Kenney and Alexander 1980; Kenney 1989).15
The discussion above suggests modifying (1) to allow the hazard
rate to depend on unobserved fixed family and community charac-
teristics. I do so by constructing separate samples of sister pairs and
brother pairs from the NLSY data and then specifying a model with
an unobserved heterogeneity term for each same-sex sibling pair.16
Thus the hazard rate takes the form
λ ij (t) ⫽ λ 0 (t)e X ij (t)␤⫹θ j , (2)
where j indexes same-sex sibling pairs, i ∈ {1, 2} indexes the two
14
The unobserved heterogeneity often is assumed to follow a gamma distribution,
which yields a closed form for the likelihood function. Heckman and Singer (1984)
develop an approach that allows for a nonparametric unobserved heterogeneity dis-
tribution.
15
The limited evidence suggests that, conditional on the sex education curriculum
in place at a school, children enrolled in that school are rarely withheld from sex
education courses by their parents. In particular, Kenney and Orr (1984) report that
a study of 12 California school districts found that only about 2 percent of parents
held their children out of sex education. Unobserved family effects could still mat-
ter, however, if a nontrivial number of families relocate or enroll their children in
private schools in response to the sex education curricula in local public schools.
16
In the few NLSY households that contain three or more sibling respondents of
the same sex, the two oldest siblings are used to form the sibling pair.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 623
siblings in each pair, θ j captures the unobserved heterogeneity, and
everything else is defined as in (1). Note that it is not assumed that
θ j and X ij (t) are uncorrelated, but it is assumed that θ j is identical
for sisters (brothers) from the same family.
With θ j and X ij (t) possibly correlated, consistent estimation of ␤
requires eliminating θ j from the likelihood function. Lancaster
(1990, pp. 268–69) shows that this can be accomplished by estimat-
ing a partial likelihood model and that, in the present problem with
exactly two spells (siblings) per family, partial likelihood yields the
fixed-effect conditional logit model (Chamberlain 1980). Intuitively,
the partial likelihood contribution of each sibling pair is just the
probability of the observed ordering of transition ages within the sib-
ling pair, conditional on the age of first transition in the sibling pair.
Thus, if the first sexual activity (pregnancy) in sibling pair j occurs
at age t, the probability that sibling i, and not sibling k, makes this
transition at age t is

λ ij (t) λ 0 (t)e X ij (t )␤⫹θj e [X ij(t)⫺X kj (t )]␤


⫽ ⫽ .
λ ij (t) ⫹ λ kj (t) λ 0 (t)e X ij (t )␤⫹θ j ⫹ λ 0 (t)e X kj (t )␤⫹θ j 1 ⫹ e [X ij (t )⫺X kj (t)]␤
(3)
Clearly, (3) has the logit form and is not a function of the unob-
served heterogeneity, θ j . Notice, however, that only the subvector of
␤ corresponding to covariates that vary within same-sex sibling pairs
can be identified. Moreover, only sibling pairs in which the age of
first sexual activity (pregnancy) differs between siblings and is less
than or equal to 19 for at least one sibling contribute to the likeli-
hood function because in the other sibling pairs behavior can be
‘‘explained’’ entirely by the unobserved heterogeneity.
These fixed-effect estimates of the effect of sex education provide
additional evidence on whether the measured effect of sex educa-
tion in the cross section represents a true causal effect. In particular,
finding an effect of sex education in the cross section but not within
same-sex sibling pairs would strongly suggest that the apparent ef-
fects of sex education are driven by unobserved heterogeneity,
whereas a robust effect of sex education both in the cross section
and within same-sex sibling pairs would make a causal interpretation
more credible. Nevertheless, the fixed-effects estimates might not
reflect the true causal effect of sex education on teen sexual behav-
ior for two reasons. First, the unobserved determinants of teen sex-
ual behavior are probably not identical for same-sex siblings. For
example, siblings of different ages have different peer groups, and
parental or community attitudes toward teen sexual activity may
change over time. These unobserved individual-specific influences
on teen sexual activity will generate biased estimates of the effects

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
624 journal of political economy
of sex education if they are correlated with enrollment in sex educa-
tion. Second, the data on sexual behavior and enrollment in sex
education both likely contain (possibly nonrandom) measurement
error.17 I consider these possible sources of bias in more detail after
presenting the estimation results. While it is clear that the empirical
evidence must be interpreted cautiously, the estimates reported here
represent a more careful analysis of the effects of sex education on
teen sexual behavior than has appeared in the previous empirical
research.

IV. Empirical Analysis


A. Descriptive Statistics
Table 3 presents the cumulative fractions of males and females who
have had sexual intercourse by each age between 10 and 19 and the
cumulative fraction of females who have had a pregnancy by each
age between 10 and 19, overall and by racial category. For each out-
come, the sample includes all observations in the full NLSY civilian
sample with valid data, and sample weights are applied to produce
statistics representative of the U.S. population born between 1957
and 1964. Most youths became sexually active during their teenage
years, but males report earlier sexual activity than females in all ra-
cial groups and blacks of both sexes report earlier sexual activity
than their white and Hispanic counterparts. These differences are
all strongly significant (p-value ⬍ .001) and agree with the findings
in previous studies. The earlier sexual activity of black males in the
NLSY is particularly striking but is entirely consistent with evidence
from other surveys (Zelnik and Kantner 1980; Ku et al. 1993). Col-
umns 9–12 document the empirical importance of teen pregnancy;
overall, 32.2 percent of the women in the 1957–64 birth cohort had
a pregnancy by age 19, and the fractions are considerably higher
among minorities.18 The evidence in table 3 also implies that the
17
On the basis of a comparison with adjusted vital statistics data from the National
Center for Health Statistics, Marsiglio and Mott (1986) conclude that abortions
among teenage women may be underreported in the NLSY by as much as 50 per-
cent. If one accepts this number and assumes that all live births are reported, the
total number of teen pregnancies might be undercounted in the NLSY by as much
as 20 percent. On the basis of comparisons with other broad surveys of teens, Mar-
siglio and Mott also report more tentatively that early teenage sexual activity among
females might be underreported in the NLSY.
18
This percentage falls to a still substantial 18.4 percent if the sample is limited
to women who remain unmarried through age 19. Restricting the sample to women
who remain unmarried through age 19 causes a smaller percentage reduction in
the fraction of black women who ever had a teen pregnancy, consistent with the
higher rate of out-of-wedlock births in the black population. Table 3 reports the
incidence of pregnancy among all teen women because unintended pregnancy
might be a significant cause of teen marriage.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
TABLE 3
Cumulative Percentages of Males and Females Who Have First Sexual Intercourse by Each Age and Cumulative Percentage
of Females Who Have First Pregnancy by Each Age, Separately by Race

Males: Percentage Had Sexual Females: Percentage Had Sexual Females: Percentage Had a
Intercourse Intercourse Pregnancy
Hispanic Black Other All Hispanic Black Other All Hispanic Black Other All
Age (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
10 2.2 7.7 1.0 2.0 ⬍.1 .3 .1 .2 ⬍.1 .1 ⬍.1 ⬍.1
11 3.2 11.9 1.3 2.8 .1 .5 .3 .3 .1 .1 ⬍.1 ⬍.1
12 6.1 20.2 3.3 5.7 .3 1.2 .5 .6 .2 .2 .1 .1
13 9.6 29.9 6.9 10.2 1.1 3.1 1.7 1.9 .6 1.6 .5 .6
14 18.3 44.2 12.5 17.1 4.0 9.6 5.0 5.6 2.1 5.6 1.4 2.1
15 33.0 62.0 23.2 29.1 11.6 20.6 11.9 13.1 6.6 13.0 4.1 5.5
16 51.1 79.5 43.1 48.5 24.7 40.6 26.7 28.5 14.9 23.3 8.5 11.0
17 69.5 88.3 60.6 64.9 41.0 61.3 43.9 46.2 23.0 32.8 14.3 17.4
18 81.9 94.0 75.6 78.5 58.7 79.6 63.5 65.5 33.5 41.1 21.8 25.3
19 87.5 95.7 82.2 84.4 70.3 86.8 74.1 75.7 42.6 49.9 28.3 32.2
N 910 1,392 3,077 5,379 942 1,438 3,271 5,651 912 1,418 3,186 5,516
Note.—The reported distributions were calculated using the NLSY sample weights to produce statistics representative of the U.S. population aged 14–21 on January 1, 1979. The
row labeled N shows the number of sample observations in each category.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
626 journal of political economy
TABLE 4
Percentage of Individuals Who Have a Sex Education Course by Each
Age, Separately by Gender and Race

Males Females
Age Hispanic Black Other All Hispanic Black Other All
10 .5 .8 1.2 1.1 1.6 1.9 3.5 3.1
11 1.3 1.8 3.4 3.1 3.3 3.9 7.2 6.5
12 3.7 3.0 6.7 6.0 7.7 6.5 12.7 11.6
13 7.1 7.1 11.9 10.9 11.1 12.0 18.6 17.2
14 10.9 12.8 18.2 17.0 16.7 19.5 26.6 25.0
15 19.1 19.0 25.4 24.2 23.7 28.3 35.2 33.6
16 26.7 27.0 33.4 32.1 33.8 38.9 44.5 43.1
17 36.3 35.9 43.6 42.1 41.7 47.3 53.1 51.6
18 44.0 45.0 50.0 49.0 47.0 54.4 59.8 58.3
19 48.0 50.5 53.6 52.8 50.4 57.2 62.0 60.6
N 883 1,346 2,955 5,184 903 1,373 3,174 5,450
Note.—The reported distributions were calculated using the NLSY sample weights to produce statistics
representative of the U.S. population aged 14–21 on January 1, 1979. The row labeled N shows the number
of sample observations in each category.

fraction of sexually active teen females who had been pregnant is


higher at each age for minorities than for whites, suggesting racial
differences in the frequency of sexual activity or the use of effective
contraception.
Table 4 shows the cumulative fraction of youths who had ‘‘ever
had a course relating to sex education’’ as of each age between 10
and 19, once again weighted to be representative of the youth popu-
lation. Only about half of all males and 60 percent of females in the
1957–64 birth cohort received formal sex education instruction by
age 19. In all racial groups, women were enrolled in sex education
more frequently and at younger ages than men, whereas among
both men and women, blacks and Hispanics received formal sex edu-
cation instruction less frequently and at later ages than whites.19
Given the large fraction of youths in this birth cohort who never had
a sex education course, it should not be surprising that nearly 60
percent of the women and over 75 percent of the men who became
sexually active by age 19 did so before taking a sex education course.
Thus, whatever the effect of sex education on teen sexual behavior,
it is doubtful that sex education was a primary cause of early teenage
sexual activity and pregnancy in this cohort. Finally, unreported logit
estimates indicate that the probability of having had a sex education
course by age 19 was much lower for youths born earlier in the 1957–
19
The finding that women were more likely to receive sex education instruction
and to receive it at younger ages is consistent with evidence from school district
administrative data (Sonenstein and Pittman 1984).

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 627
64 birth cohort, youths raised in the South, and youths with less
educated parents. The latter two effects suggest that there was sub-
stantial variation across communities in the prevalence of school-
based sex education during the 1970s and underscore the value of
estimation that allows for unobserved family fixed effects.

B. Hazard Model Estimates


Tables 5–7 present selected coefficient estimates from the models
for transitions into sexual activity and pregnancy; complete estima-
tion results are available from the author on request. For brevity, I
focus on the measured effects of sex education, but the other re-
ported coefficients indicate strong associations between teen sexual
behavior and several demographic and behavioral characteristics.20
Each table presents three sets of estimates. First, I present semipara-
metric estimates of (1), the proportional hazard model without un-
observed heterogeneity, using the full sample of females (or males)
with complete data. Second, I repeat the semiparametric estimates
of (1), this time using only the subsample of sister (or brother) pairs
with data for both siblings. Finally, I present partial likelihood esti-
mates of (2), the proportional hazard model that allows for fixed
unobserved family/community heterogeneity, again using the sub-
sample of sister (or brother) pairs with complete data for both sib-
lings.21 As described earlier, partial likelihood for the present prob-
lem is just the fixed-effects conditional logit estimator. For each
sample and estimation method, I report the results for two alterna-
tive specifications of the effect of sex education: in the first, I allow
the effect to differ between ages 10–15 and ages 16–19; in the sec-
ond, I let the effect vary between teens who have an older sibling
and those who do not.22
Before I turn to the results, several points about the data require
20
In addition, although I do not report the estimated birth year dummy coeffi-
cients, there is strong evidence that teens from more recent birth cohorts had higher
hazard rates into sexual activity and pregnancy. Similarly, while I do not report the
nonparametric baseline hazard estimates, the hazard rates into both sexual activity
and pregnancy show a clear age pattern. For sexual activity, the baseline hazard
function for both males and females is basically zero at the youngest ages (e.g., 10–
11), rises monotonically up to age 18, and then drops off noticeably at age 19. For
pregnancy, the baseline hazard is essentially zero through age 13 and then rises
monotonically up to age 19.
21
These samples are somewhat smaller than the sibling pair samples in the semi-
parametric hazard estimation, however, because only sibling pairs for which the age
of transition differs between siblings and is less than 19 for at least one sibling con-
tribute to the sample.
22
The main effect for age greater than 15 is (properly) excluded from all the
specifications. In the semiparametric hazard model, this main effect is collinear with
the covariates identifying the baseline hazard. In the conditional logit model, this
main effect is not included because both siblings are observed at the same age, and

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
628 journal of political economy
discussion. First, I track transitions into sexual activity (pregnancy)
occurring only between ages 10 and 19. Thus the very few individuals
who report initial sexual activity (pregnancy) before age 10 are
dropped from the sample, and first transitions that occur after age
19 are treated as censored spells at age 19.23 Second, the reported
estimates are based on samples that include all person-years up to
the age of transition (or age 19), regardless of the respondent’s mari-
tal status at that age; but all the results are qualitatively the same if
the sample includes only unmarried person-years. Third, because
the NLSY records age of sex education and age of first sexual activity
(pregnancy) only in full years, one cannot tell whether sex education
precedes or follows first sexual activity (pregnancy) for respondents
who report that both occurred at the same age. I report estimates
that assume that sex education followed first sexual activity (preg-
nancy) for individuals who report that both events took place at the
same age, but the results are qualitatively identical if one assumes
that sex education preceded first sexual activity (pregnancy) for
these individuals.
Table 5 presents the estimates for female transitions into sexual
activity. Sex education has a positive and significant effect on the
sexual activity hazard rate for all groups of teen females in all the
various samples and specifications.24 However, except in column 1,
the sex education interactions suggest that the effect of sex educa-
tion is much smaller in the later teen years and for teens with older
siblings, and in four of these specifications, the interaction term is
at least marginally significant at conventional significance levels.
Both the smaller effects of sex education for apparently better-
informed teens and the robustness of the pattern of estimated effects
of sex education to the inclusion of family fixed effects lend support
to a causal interpretation of the estimates. However, restricting the
sample to sibling pairs and controlling for fixed effects both cause
the estimated effects of sex education to increase markedly.25 I dis-
cuss possible biases in the fixed-effects estimates in subsection D.
therefore there is no within-sibling pair variation in the indicator for age greater
than 15.
23
Treating transitions that occur after age 19 as censored spells at age 19 is desir-
able if the effects of some covariates differ between the teen years and adulthood.
Nevertheless, the results for transitions into sexual activity are qualitatively un-
changed if spells of virginity are not censored at age 19.
24
Various hypothesis tests for the significance of the estimated sex education ef-
fects are reported at the bottom of the table. The estimated effects of sex education
in specifications that do not include any sex education interaction terms (not re-
ported in table 5) are .1808 for the full sample, .3651 for the sister pair subsample
without family fixed effects, and .6537 for the sister pair subsample with family fixed
effects. The p-value for each of these estimated effects of sex education is less than
.001.
25
The difference between the estimates in cols. 1 and 2 and the estimates in cols.
3 and 4 results entirely from changes in sample composition. In unreported esti-

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 629
Table 6 shows the estimates for female transitions into pregnancy.
The point estimates suggest that sex education generally has a posi-
tive effect on the pregnancy hazard, but the effects are uniformly
smaller and less significant than the estimated effects for transitions
into sexual activity.26 As in table 5, the effects of sex education on
the pregnancy hazard are much larger and more significant in the
early teen years and for women without older siblings.27 Also as in
table 5, allowing for family fixed effects does not change the pattern
of the estimated effects of sex education, but it does increase their
magnitudes. The smaller effect of sex education on the pregnancy
hazard as compared with the sexual activity hazard is inconsistent
with the predicted effects of utility-altering and risk-revealing sex
education, but not with the predicted effect of risk-altering sex edu-
cation. Thus a causal interpretation of the evidence would suggest
that the provision of risk-altering information must have been a
prominent component of sex education courses in the 1970s.
Finally, table 7 presents the estimates for male transitions into sex-
ual activity. Sex education never has a significant effect on the haz-
ard rate in the early teen years or for males without older siblings,
and the magnitudes of the estimated effects are often small. In con-
trast to the theory, sex education appears to have a larger positive
effect on the hazard rate for males older than 15.28 The effects of
sex education also appear to be slightly larger in the specifications
that allow for family fixed effects. Perhaps most interesting, a com-
parison with the results in table 5 indicates that sex education has

mates, I find that the effect of sex education is somewhat larger for more recent
birth cohorts. Consequently, the larger estimated effect of sex education in the sister
pair subsample can be explained partially by the fact that the sister pair subsample
contains disproportionately many teens from the later birth cohorts. The larger mea-
sured effect of sex education in more recent birth cohorts could be a true effect
resulting from changes in sex education curricula over time or an artificial conse-
quence of greater recall error in reported enrollment in sex education for individu-
als from earlier birth cohorts (who were longer out of high school when these data
were collected).
26
The estimated effects of sex education in specifications that do not include any
sex education interaction terms (not reported in table 6) are .0664 for the full sam-
ple, .1739 for the sister pair subsample without family fixed effects, and .3931 for
the sister pair subsample with family fixed effects, with p-values of .188, .119, and
.080, respectively. These effects are much smaller than the analogous effects for
transitions into sexual activity (see n. 24).
27
In four of the six specifications in table 6, the main effect of sex education is
significant and the sex education interaction term is marginally significant at con-
ventional significance levels.
28
The estimated effects of sex education in specifications that do not include any
sex education interaction terms (not reported in table 7) are .0722 for the full sam-
ple, .1210 for the brother pair subsample without family fixed effects, and .2939 for
the brother pair subsample with family fixed effects, with p-values of .041, .076, and
.206, respectively. These effects are much smaller than the effects for female transi-
tions into sexual activity (see n. 24).

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
TABLE 5
Selected Coefficient Estimates from Models of Transition into Sexual Activity for Females

All Females Sister Pairs Sister Pairs


(1) (2) (3) (4) (5) (6)

630
Have older sibling .0787 .0740 .1431 .0177 ⫺.1022 ⫺.3119
(.0503) (.0389) (.1206) (.0904) (.2370) (.1982)
Prior sex education .1887 .3398 .5947 .6655 1.1189 .9327
(.0691) (.0916) (.1585) (.1932) (.3648) (.4407)
Prior sex education ⫻ have older sibling ⫺.0103 ⫺.2858 ⫺.5711
(.0782) (.1754) (.3818)
Prior sex education ⫻ age ⬎ 15 ⫺.1841 ⫺.3441 ⫺.3424
(.0986) (.2070) (.4860)
Black .2473 .2484 .3817 .3849
(.0427) (.0427) (.0923) (.0922)
Hispanic ⫺.2681 ⫺.2677 ⫺.2591 ⫺.2578
(.0573) (.0573) (.1205) (.1206)
Parents’ highest grade ⬍ 12 .1122 .1125 .0485 .0496

This content downloaded from


(.0383) (.0383) (.0809) (.0809)
Parents’ highest grade ⱖ 16 ⫺.3304 ⫺.3311 ⫺.3084 ⫺.3203

186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00


All use subject to https://about.jstor.org/terms
(.0522) (.0522) (.1070) (.1068)
Live in nonintact family at age 14 .3739 .3735 .3585 .3547 .2062 .2178
(.0360) (.0360) (.0788) (.0788) (.3846) (.3843)
Never attend church .7552 .7547 .8936 .8944 .6260 .5927
(.0512) (.0512) (.1135) (.1135) (.3351) (.3343)
Attend church ⬍ 1 time/month .5784 .5783 .7413 .7339 .7640 .7267
(.0433) (.0432) (.0887) (.0887) (.2542) (.2529)
Attend church 1–3 times/month .3783 .3791 .4719 .4806 .4829 .4739
(.0457) (.0457) (.0946) (.0944) (.2325) (.2323)
Family-specific fixed effects? no no no no yes yes
Number of individuals 4,945 4,945 1,208 1,208 854 854
Number of person-years 41,126 41,126 10,309 10,309 854 854
Log-likelihood ⫺9,104.1 ⫺9,102.4 ⫺2,080.7 ⫺2,080.6 ⫺275.3 ⫺276.1
p-value for hypothesis that:
Sex education coefficient ⫽ 0 .006 ⬍.001 ⬍.001 .001 .002 .034
Sex education interaction coefficient ⫽ 0 .895 .062 .103 .096 .135 .481
Sex education coefficient ⫹ sex education
interaction coefficient ⫽ 0 ⬍.001 ⬍.001 ⬍.001 ⬍.001 .006 .004

631
Note.—Asymptotic standard errors are in parentheses. Cols. 1–4 report semiparametric estimates of proportional hazard models. Cols. 5 and 6 report estimates of fixed-effect
conditional logit models. The church attendance variables reflect behavior during the previous year at the time of the 1979 interview. Other explanatory variables in the models in
cols. 1–4, not reported in the table, are dummies for each year of age between 10 and 19 (which yield a nonparametric estimate of the baseline hazard), a dummy for urban residence
at age 14, a dummy for residence in the South at age 14, dummies for religious affiliation at age 14 (Catholic, non-Christian religion, and no religious affiliation), and dummies for
year of birth (1957–63). Other explanatory variables in the models in cols. 5 and 6, not reported in the table, are a dummy for urban residence at age 14, a dummy for residence in
the South at age 14, and dummies for year of birth (1957–63).

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
TABLE 6
Selected Coefficient Estimates from Models of Transition into Pregnancy for Females

All Females Sister Pairs Sister Pairs


(1) (2) (3) (4) (5) (6)
Have older sibling .0430 .0118 .0162 ⫺.1441 .0573 ⫺.2174
(.0721) (.0557) (.1823) (.1354) (.3050) (.2432)
Prior sex education .1245 .2617 .4666 .5976 .9297 .8180
(.0996) (.1303) (.2392) (.2838) (.4368) (.6245)
Prior sex education ⫻ have older sibling ⫺.0757 ⫺.3634 ⫺.6596
(.1121) (.2636) (.4569)
Prior sex education ⫻ age ⬎ 15 ⫺.2261 ⫺.4877 ⫺.4876
(.1404) (.3058) (.6631)
Black .3620 .3624 .7219 .7264
(.0593) (.0593) (.1373) (.1374)
Hispanic .2216 .2221 .1030 .1004
(.0787) (.0787) (.1800) (.1801)
Parents’ highest grade ⬍ 12 .3638 .3641 .4943 .4942
(.0518) (.0518) (.1162) (.1163)

632
Parents’ highest grade ⱖ 16 ⫺.7440 ⫺.7450 ⫺.7760 ⫺.7892
(.0963) (.0963) (.2166) (.2164)
Live in nonintact family at age 14 .4315 .4307 .3847 .3790 .1921 .2328
(.0491) (.0491) (.1125) (.1125) (.4422) (.4393)
Never attend church .6303 .6281 .7091 .6980 .1013 .0924
(.0717) (.0717) (.1639) (.1637) (.3880) (.3862)
Attend church ⬍ 1 time/month .3815 .3805 .4631 .4530 .0841 .0089
(.0630) (.0630) (.1361) (.1358) (.3004) (.2970)
Attend church 1–3 times/month .3139 .3143 .2523 .2530 .2508 .2295
(.0649) (.0649) (.1424) (.1424) (.2931) (.2922)
Family-specific fixed effects? no no no no yes yes
Number of individuals 4,945 4,945 1,208 1,208 554 554
Number of person-years 45,681 45,681 11,343 11,343 554 554
Log-likelihood ⫺6,317.6 ⫺6,316.6 ⫺1,335.8 ⫺1,335.5 ⫺185.9 ⫺186.6
p-value for hypothesis that:

This content downloaded from


Sex education coefficient ⫽ 0 .211 .045 .051 .035 .033 .190
Sex education interaction coefficient ⫽ 0 .499 .107 .168 .111 .149 .462

186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00


Sex education coefficient ⫹ sex education

All use subject to https://about.jstor.org/terms


interaction coefficient ⫽ 0 .391 .511 .403 .355 .259 .166
Note.—See note to table 5.
TABLE 7
Selected Coefficient Estimates from Models of Transition into Sexual Activity for Males

All Males Brother Pairs Brother Pairs

(1) (2) (3) (4) (5) (6)


Have older sibling .0740 .0815 .2618 .2329 .4021 .3999
(.0433) (.0380) (.0904) (.0792) (.1897) (.1720)
Prior sex education .0494 ⫺.0461 .2168 ⫺.0519 .2927 ⫺.0389
(.0786) (.0925) (.1603) (.1690) (.3720) (.3294)
Prior sex education ⫻ have older sibling .0306 ⫺.1201 .0015
(.0893) (.1780) (.3814)
Prior sex education ⫻ age ⬎ 15 .1483 .2190 .5490
(.1026) (.1879) (.4248)
Black .7500 .7486 .8247 .8233
(.0422) (.0422) (.0829) (.0829)
Hispanic .0241 .0227 .1337 .1273
(.0528) (.0527) (.0946) (.0946)
Parents’ highest grade ⬍ 12 .0087 .0085 ⫺.1889 ⫺.1870
(.0376) (.0376) (.0711) (.0710)

633
Parents’ highest grade ⱖ 16 ⫺.3124 ⫺.3126 ⫺.4432 ⫺.4441
(.0488) (.0488) (.0920) (.0920)
Live in nonintact family at age 14 .2943 .2944 .3082 .3097 .2229 .2154
(.0357) (.0357) (.0704) (.0704) (.3200) (.3204)
Never attend church .4457 .4464 .5135 .5141 .5764 .5760
(.0476) (.0476) (.0868) (.0868) (.2677) (.2679)
Attend church ⬍ 1 time/month .4311 .4317 .4050 .4037 .1602 .1698
(.0440) (.0440) (.0839) (.0839) (.2316) (.2311)
Attend church 1–3 times/month .2608 .2620 .3366 .3336 .1514 .1574
(.0465) (.0465) (.0846) (.0844) (.2147) (.2141)
Family-specific fixed effects? no no no no yes yes
Number of individuals 4,618 4,618 1,364 1,364 1,072 1,072
Number of person-years 33,074 33,074 9,730 9,730 1,072 1,072
Log-likelihood ⫺9,618.4 ⫺9,617.5 ⫺2,828.4 ⫺2,828.0 ⫺358.7 ⫺357.8
p-value for hypothesis that:
Sex education coefficient ⫽ 0 .530 .618 .176 .759 .431 .906

This content downloaded from


Sex education interaction coefficient ⫽ 0 .732 .148 .500 .244 .997 .196
Sex education coefficient ⫹ sex educa-

186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00


tion interaction coefficient ⫽ 0 .089 .026 .252 .050 .176 .057

All use subject to https://about.jstor.org/terms


Note.—See note to table 5.
634 journal of political economy
a smaller effect on the sexual activity of males than of females. This
finding is consistent with the theory if, as seems likely, unintended
pregnancies impose greater costs on women.

C. Simulating the Effects of Expanding Sex Education


To get a sense of the magnitudes of the effects of sex education on
teen sexual behavior, I use the hazard model estimates to predict
how the cumulative fraction of youths who become sexually active
(pregnant) by each age between 10 and 19 would have differed if
sex education had been more widespread.29 For brevity, I consider
two simple but dramatic expansions of school-based sex education:
mandatory sex education at age 15 for all youths who have not previ-
ously had a sex education course and mandatory sex education at
age 10 for all youths who have not previously had a sex education
course. Examining the actual sex education experience of the 1957–
64 birth cohort (shown in table 4) confirms that either policy would
have represented a huge expansion of sex education. The predicted
effects of these policies are shown in tables 8 and 9: table 8 is based
on the semiparametric hazard model estimates for the full teen sam-
ples (specifically, col. 2 of tables 5–7) and table 9 is based on the
analogous estimates for the sibling pair subsamples (specifically, col.
4 of tables 5–7).30 The next subsection considers whether the esti-
mates that underlie these simulations should be given a causal inter-
pretation.
The simulations in table 8 indicate that universal sex education
at age 15 would have had very small impacts on teen sexual behavior.
For example, the maximum predicted absolute increase in the per-
centage of women who are sexually active is 2.5 percentage points
at age 18. The maximum predicted percentage increase in the fraction
of women who are sexually active occurs at age 16 and is only about
5 percent. The predicted (absolute and percentage) increases in
pregnancy and male sexual activity are much smaller still. Universal
sex education at age 10 apparently would have had considerably
stronger impacts, at least for females, reflecting the larger estimated

29
I do this by first computing each individual’s predicted probability of having
made a transition by each age using the estimated hazard model coefficients, the
individual’s personal characteristics, and an assumed time path of the sex education
indicator. I then calculate the predicted cumulative fraction of teens who are sexu-
ally active (pregnant) by each age as a weighted average of these individual predicted
probabilities using the NLSY sample weights.
30
The estimates from cols. 2 and 4 are used because these specifications give
higher log-likelihood values than those in cols. 1 and 3. The fixed-effects estimates
are not used for this exercise because computing the survival probabilities requires
estimates of the baseline hazard function.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
TABLE 8
Predicted Percentage of Youths Who Have Become Sexually Active or Pregnant by Each Age, under Different Assumptions
about the Provision of Sex Education, Based on Full-Sample Hazard Model Estimates

Females: Percentage Sexually Females: Percentage Who Had Males: Percentage Sexually Active
Active under Pregnancy under under
Sex Sex Sex Sex Sex Sex
Education Education Education Education Education Education
Status Quo at 15 at 10 Status Quo at 15 at 10 Status Quo at 15 at 10
Age (1) (2) (3) (4) (5) (6) (7) (8) (9)
10 .1 .1 .1 .0 .0 .0 1.0 1.0 1.0
11 .2 .2 .3 .0 .0 .0 2.1 2.1 2.1
12 .6 .6 .7 .1 .1 .2 5.8 5.8 5.6
13 2.0 2.0 2.6 .6 .6 .7 10.8 10.8 10.4
14 5.8 5.8 7.6 2.3 2.3 2.8 18.6 18.6 17.9
15 13.7 13.7 17.3 6.2 6.2 7.6 31.0 31.0 30.0
16 29.1 30.6 33.4 12.3 12.4 13.6 49.6 50.8 50.2
17 46.8 49.2 51.1 18.9 19.2 20.2 65.4 67.0 66.6
18 65.6 68.1 69.2 26.4 26.7 27.7 78.3 79.8 79.6
19 75.6 77.8 78.5 33.4 33.8 34.6 83.7 85.0 84.9
Note.—Predictions are based on the hazard model estimates presented in col. 2 of tables 5–7. The predictions in cols. 1, 4, and 7 are computed using the actual data on sex
education enrollment in the samples. The predictions in cols. 2, 5, and 8 are computed under the assumption that all respondents are enrolled in sex education at age 15 if they have
not been enrolled in sex education previously. The predictions in cols. 3, 6, and 9 are computed under the assumption that all respondents are enrolled in sex education at age 10
if they have not been enrolled in sex education previously.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
TABLE 9
Predicted Percentage of Youths Who Have Become Sexually Active or Pregnant by Each Age, under Different Assumptions
about the Provision of Sex Education, Based on Sibling Pair Sample Hazard Model Estimates

Females: Percentage Sexually Females: Percentage Who Had Males: Percentage Sexually Active
Active under Pregnancy under under
Sex Sex Sex Sex Sex Sex
Education Education Education Education Education Education
Status Quo at 15 at 10 Status Quo at 15 at 10 Status Quo at 15 at 10
Age (1) (2) (3) (4) (5) (6) (7) (8) (9)
10 .0 .0 .0 .0 .0 .0 1.2 1.2 1.2
11 .0 .0 .0 .0 .0 .0 2.5 2.5 2.4
12 .1 .1 .2 .0 .0 .0 6.2 6.2 5.9
13 .6 .6 1.0 .5 .5 .8 10.9 10.9 10.5
14 4.1 4.1 6.8 1.5 1.5 2.4 19.0 19.0 18.3
15 11.5 11.5 17.8 4.9 4.9 7.4 30.3 30.3 29.2
16 25.6 28.4 33.3 9.5 9.9 12.2 47.1 48.9 48.2
17 43.0 47.7 51.0 15.1 15.7 17.8 64.2 66.8 66.4
18 62.4 67.6 69.4 21.2 22.1 24.0 78.1 80.5 80.3
19 72.9 77.5 78.7 27.5 28.6 30.2 82.9 85.1 84.9
Note.—Predictions are based on the hazard model estimates presented in col. 4 of tables 5–7. For further explanation, see note to table 8.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 637
effects of sex education on behavior in the early teen years. For ex-
ample, the predicted increase in the percentage of females who are
sexually active by age 15 is 3.6 percentage points, which represents
a substantial 26 percent increase. More generally, the simulations
imply that sex education at age 10 would have led to larger absolute
increases in the fractions of females who had become sexually active
or pregnant by each age than sex education at age 15 would have,
and the percentage increases would have been quite substantial in
the early and midteen years.
Table 9 reproduces the simulations using the hazard model esti-
mates from the sibling pair samples. The basic pattern of the pre-
dicted effects is the same as in table 8, but the magnitudes are some-
what larger. Overall, the simulations suggest that increases in sexual
activity and pregnancy among teen females in response to universal
sex education would have been modest if the course was given at
age 15 but would have been considerably larger if the course was
given at age 10.

D. Discussion of Findings
The evidence indicates that prior enrollment in sex education was
associated with earlier initial sexual activity for women from the
1957–64 birth cohort. Sex education also may have been associated
with earlier first pregnancies for these women, but this effect is
smaller and less significant than the effect on sexual activity. For
both outcomes, sex education generally seems to have had a larger
impact in the early teen years and for women without older siblings.
Finally, the association between sex education and transitions into
sexual activity was much smaller and weaker for males from this co-
hort than for females.
Are these results evidence of a causal influence of sex education
on teen sexual behavior, and, if so, what is the source of this influ-
ence? The robustness of the pattern of estimated effects in the within-
sibling pair analyses certainly makes a causal interpretation more
plausible. Moreover, only the risk-altering model of sex education
can easily explain the smaller measured effect of sex education on
pregnancy relative to sexual activity. However, several issues need to
be addressed before one accepts the conclusion that sex education
courses in the 1970s had a causal impact on teen sexual behavior,
in large part by providing information that allowed teens to reduce
the risks of sexual activity.
To start with, it is worth checking whether sex education courses
during the 1970s actually had a prominent risk-altering component.
The evidence from surveys of school staff (Orr 1982; Sonenstein and

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
638 journal of political economy
Pittman 1984) and students (Zelnik and Kim 1982; Dawson 1986;
Marsiglio and Mott 1986) strongly supports this view. In particular,
all these studies indicate that contraception, family planning, and
pregnancy risks were major topics in most sex education courses dur-
ing this era and contraception, when discussed, was likely to receive
extensive coverage (Orr 1982).
Second, although I have not analyzed the effect of sex education
on the frequency of teen sexual activity because of lack of data, ex-
tending the theory to incorporate a choice of frequency does not
substantively change the predicted effects of any of the types of sex
education on the fractions of teens who are sexually active or who
become pregnant.31 Thus, while the effects of sex education on the
sexual activity hazards reported above probably do not capture the
total impact of sex education on all dimensions of teen sexual activ-
ity, the evidence on the outcomes that I can observe still is potentially
useful for evaluating the theory.32
Third, and most important, the question remains whether the es-
timated effects of sex education accurately measure the causal ef-
fects of sex education on behavior. Cross-section estimates of the ef-
fects of sex education might be biased for three distinct reasons:
unobserved family or community heterogeneity correlated with
enrollment in sex education, unobserved individual heterogeneity
correlated with enrollment in sex education, or individual-specific
measurement error in reported sexual behavior or reported enroll-
ment in sex education. The within-sibling pair analyses eliminate
the first potential source of bias. If the other two potential sources
of bias were negligible, the rise in the estimated effects of sex educa-
tion in the within-sibling pair analyses would imply that, during the
1970s, sex education courses were more prevalent in communities
in which unobserved propensities toward teen sexual activity and
pregnancy were low. However, if the two individual-specific sources
of error are not negligible, the within-sibling pair estimates also
might be biased, perhaps more than the cross-section estimates.
Thus, in the remainder of this section, I consider the importance
of these two other potential sources of bias.
Enrollment in sex education might be correlated with unobserved
individual-specific differences in the probability of teen sexual activ-

31
See n. 10 for further discussion of this extension of the model.
32
Ku et al. (1992) find that sex education does influence the frequency of sexual
activity. In addition, Kirby et al. (1994) report that several evaluations of specific
sex education curricula find delayed initiation of sexual activity and increased con-
traceptive use, but not decreased incidence of pregnancy, among course participants.
These results can be reconciled easily if sex education affects the frequency of sexual
activity among already sexually active teens.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 639
ity for a few reasons. First, within any family, parents might selectively
decide whether and at what age to enroll each child in sex education
on the basis of their perception of that child’s propensity toward
teen sexual activity. If the teens most inclined toward early sexual
activity are more likely to be enrolled in sex education (or to be
enrolled at younger ages), then the within-sibling pair estimates of
the effect of sex education will be upward-biased. Second, unob-
served community influences on sexual behavior might vary between
siblings in a way that is correlated with enrollment in sex education.
For example, if communities adopt sex education curricula in re-
sponse to unexplained local increases in teen sexual activity and if
teen sexual activity continues to increase after the introduction of
sex education, the within-sibling pair estimates again will be upward-
biased.33
While the evidence is limited, unobserved individual heterogene-
ity does not seem to offer a compelling explanation for the empirical
findings. First, conditional on where they choose to enroll their chil-
dren, parents seem to withhold their children from sex education
courses extremely rarely (see n. 15). Similarly, it seems doubtful that
more than a handful of families enroll only some of their children
in private schools and make the decision on the basis of each child’s
unobserved propensity toward sexual activity. In addition, most
school districts had fairly rigid, grade-based sex education curricula
(Orr 1982; Sonenstein and Pittman 1984), which suggests that par-
ents had little ability to influence the age at which a teen received sex
education, conditional on the school in which the teen was enrolled.
Finally, the very small age difference in most of the sibling pairs
suggests that between-sibling differences in community environ-
ment were also probably very small.34
A more likely source of bias in the estimated effects of sex educa-
tion is individual-specific measurement error in either reported sex-
ual behavior or reported enrollment in sex education. As mentioned
in note 17, vital statistics data suggest that pregnancies are under-
reported in the NLSY (Marsiglio and Mott 1986). Age of first sexual
activity also is likely to be subject to reporting error, both because
of intentional misreporting and, to a lesser extent (given the salience
of the event), because of lapses in recall. Finally, the data on sibling

33
However, since I include birth year dummies in all the analyses, this type of
bias will arise only if the part of the local community time trend in environmental
influences on teen sexual activity that is orthogonal to the national time trend in
environmental influences on teen sexual activity is correlated with local adoption
of sex education curricula.
34
In the data, 66 percent (84 percent) of all sibling pairs are 2 (3) years or less
apart in age.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
640 journal of political economy
pairs also suggest the presence of some measurement error in re-
ported enrollment in sex education. In particular, in a nontrivial
minority of sibling pairs, either the older sibling is the only one who
reports enrollment in sex education or the older sibling reports en-
rollment in sex education at a younger age than the younger sibling
does. While this pattern need not always indicate reporting error,
the growth in school-based sex education over the time period of
the data suggests that this pattern should have been rare.
I have performed additional analyses that indicate that the esti-
mated effects of sex education cannot be explained away as arti-
facts of purely random measurement error. To begin with, random
misclassification error in a binary explanatory variable biases the
associated coefficient estimate toward zero (Aigner 1973). Thus, if
enrollment in sex education is subject to random misreporting,
the estimated effects of sex education understate the true effects.
Random misclassification error in a binary dependent variable also
causes conventional estimation methods to be biased, but Haus-
man, Scott-Morton, and Abrevaya (1996) develop techniques for ob-
taining consistent estimates under the assumption that misclassifica-
tion is uncorrelated with the explanatory variables. I implemented
the Hausman et al. approach on a sequence of logit models that
approximate the hazard model specification, and the estimated ef-
fects of sex education were robust to allowing for random misclassi-
fication error in reported age of first sexual activity (or pregnancy).35
Ruling out nonrandom reporting error as the explanation for the
observed effects of sex education is much more difficult since the
‘‘right’’ correlation between reporting error and observables can ex-
plain any set of findings. Nevertheless, for several plausible stories
of misreporting, one would expect to observe a spurious positive
correlation between enrollment in sex education and the probability
of transition into sexual activity (or pregnancy). This would occur,
for example, if prior enrollment in sex education made individuals
more willing to report early sexual activity or pregnancy or if individ-
uals who become sexually active or pregnant at an early age were
more likely to recall (or invent) previous instruction in sex educa-
tion.
Still, nonrandom reporting error seems unlikely to offer a con-
vincing explanation for the full set of findings for a few reasons.

35
In particular, the specifications that allow for misclassification error produce
point estimates and standard errors on the sex education variables very similar to
those of the analogous specifications that do not. Although the misclassification
probabilities themselves are estimated quite imprecisely, the point estimates suggest
that women do not underreport teen sexual activity but substantially underreport
teen pregnancy.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 641
First, in unreported analyses I find that the estimated effects of sex
education do not differ significantly depending on whether the re-
spondent answered the questions on sex education and sexual be-
havior in the presence of a friend or relative or in private. Such
differences would have been expected, however, if nonrandom re-
porting error were the main source of the effects of sex education
and if, as seems likely, the confidentiality of the interview setting has
a systematic impact on misreporting. Second, ages of initial sexual
activity and pregnancy were obtained in retrospective interviews con-
ducted when most sample members were in their early to middle
twenties. By these ages, it seems doubtful that misreporting of teen
sexual behavior would be highly correlated with whether the individ-
ual had a sex education course many years before.36 Finally, to ex-
plain the complete set of findings as an artifact of nonrandom re-
porting error, one must explain why the reporting error is larger
for individuals with older siblings. While there is no obvious explana-
tion why reporting error would exhibit this pattern, the causal mod-
els of sex education provide an intuitive explanation for why the
effects of sex education should be mitigated for teens with older
siblings.37 Thus, while reporting error may affect the magnitude of
the estimated effects of sex education, the overall pattern of the re-
sults suggests that sex education did have some causal impact on
teen sexual behavior in the 1970s.

V. Conclusion
This paper has examined whether and how sex education affected
teen sexual behavior in the 1970s. In contrast to previous studies,
the empirical analysis is guided by a rational choice model of sexual
activity. Moreover, the empirical work goes beyond previous studies
of the effects of sex education both by testing more predictions of
the theory and by controlling for unobserved family-specific (or
community-specific) heterogeneity. The main findings are that en-
rollment in sex education was associated with a significant increase
in the hazard rate into sexual activity for females and a smaller, and
often insignificant, increase in the pregnancy hazard rate. For both

36
Moreover, most of the data on teen sexual activity and teen pregnancy were
obtained in an earlier survey round than the data on enrollment in sex education,
so there is no concern that the questions about enrollment in sex education influ-
enced the responses to questions about sexual behavior, or vice versa.
37
Similarly, any pure unobserved heterogeneity explanation of the full set of mea-
sured effects of sex education must explain why the correlation between the unob-
served heterogeneity and enrollment in sex education differs systematically between
individuals who have older siblings and those who do not.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
642 journal of political economy
outcomes, the effect of sex education was larger for young women
who appeared to have fewer low-cost alternative sources of sexual
information. The evidence suggests that sex education had some
causal influence on teen sexual behavior, and it seems likely that an
important source of this causal link was the provision of information
that altered the risks of sexual activity. However, reporting error and
unobserved individual heterogeneity may be partially responsible for
the measured effects of sex education, and therefore the estimates
should be interpreted cautiously.
For two reasons, the empirical findings reported here probably
offer only limited insight into the effects of sex education on sexual
behavior for the current generation of teens. First, as already dis-
cussed at length, the current estimates might not capture the causal
effects that sex education had on teen sexual behavior in the 1970s
because of reporting errors or unobserved factors. Perhaps equally
important, the emergence of acquired immune deficiency syndrome
(AIDS) or other changes over the last two decades may have altered
the nature of sex education or the perceived costs and benefits of
teen sexual activity, thereby possibly changing the average effect of
sex education.38 Ascertaining the effects of sex education for the cur-
rent generation will require more recent data. At the same time,
obtaining richer data would enable more informative analyses. For
example, one could test the theory more fully and could gain a more
detailed understanding of the effects of sex education with data on
a wider set of sexual behaviors (e.g., frequency of sexual activity,
contraceptive use) and on the content and comprehensiveness of
sex education courses. Furthermore, measurement error problems
could be reduced with school transcript data on enrollment in sex
education. Nevertheless, the framework developed here should be
a useful foundation for future work.

References
Aigner, Dennis J. ‘‘Regression with a Binary Independent Variable Subject
to Errors of Observation.’’ J. Econometrics 1 (March 1973): 49–59.
Chamberlain, Gary. ‘‘Analysis of Covariance with Qualitative Data.’’ Rev.
Econ. Studies 47 ( January 1980): 225–38.
38
For example, recent surveys (Forrest and Silverman 1989; Kenney, Guardado,
and Brown 1989) suggest that the emergence of AIDS has led to more uniform and
comprehensive sex education curricula that place greater emphasis on developing
‘‘resistance skills’’ and encouraging abstinence. If these curricular changes are actu-
ally effective in raising the relative utility of abstinence, then one might expect cur-
rent sex education to discourage teen sexual activity. Perhaps consistent with this
view, Ku et al. (1993) find that teen males who receive instruction on ‘‘resistance
skills’’ make later transitions into sexual activity, whereas those who get instruction
on contraception and reproductive biology make earlier transitions.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
effects of sex education 643
Cox, D. R., and Oakes, D. Analysis of Survival Data. London: Chapman and
Hall, 1984.
Dawson, Deborah A. ‘‘The Effects of Sex Education on Adolescent Behav-
ior.’’ Family Planning Perspectives 18 ( July/August 1986): 162–70.
Forrest, Jacqueline D., and Silverman, Jane. ‘‘What Public School Teachers
Teach about Preventing Pregnancy, AIDS, and Sexually Transmitted Dis-
eases.’’ Family Planning Perspectives 21 (March/April 1989): 65–72.
Forrest, Jacqueline D., and Singh, Susheela. ‘‘The Sexual and Reproductive
Behavior of American Women, 1982–1988.’’ Family Planning Perspectives
22 (September/October 1990): 206–14.
Hausman, Jerry A.; Scott-Morton, Fiona M.; and Abrevaya, Jason. ‘‘Misclassi-
fication of the Dependent Variable in a Discrete-Response Setting.’’
Manuscript. Cambridge: Massachusetts Inst. Tech., Dept. Econ., 1996.
Heckman, James J., and Singer, Burton. ‘‘A Method for Minimizing the
Impact of Distributional Assumptions in Econometric Models for Dura-
tion Data.’’ Econometrica 52 (March 1984): 271–320.
Hofferth, Sandra L.; Kahn, Joan R.; and Baldwin, Wendy. ‘‘Premarital Sex-
ual Activity among U.S. Teenage Women over the Past Three Decades.’’
Family Planning Perspectives 19 (March/April 1987): 46–53.
Hoynes, Hilary, and MaCurdy, Thomas E. ‘‘Has the Decline in Benefits
Shortened Welfare Spells?’’ A.E.R. Papers and Proc. 84 (May 1994): 43–
48.
Kenney, Asta M. Sex Education and AIDS Education in the Schools: A Survey of
State Policies, Curricula and Program Activities. New York: Guttmacher Inst.,
1989.
Kenney, Asta M., and Alexander, Sharon J. ‘‘Sex/Family Life Education in
the Schools: An Analysis of State Policies.’’ Family Planning/Population
Reporter 9 ( June 1980): 44–52.
Kenney, Asta M.; Guardado, Sandra; and Brown, Lisanne. ‘‘Sex Education
and AIDS Education in the Schools: What States and Large School Dis-
tricts Are Doing.’’ Family Planning Perspectives 21 (March/April 1989): 56–
64.
Kenney, Asta M., and Orr, Margaret T. ‘‘Sex Education: An Overview of
Current Programs, Policies, and Research.’’ Phi Delta Kappan 65 (March
1984): 491–96.
Kirby, Douglas, et al. ‘‘School-Based Programs to Reduce Sexual Risk Be-
haviors: A Review of Effectiveness.’’ Public Health Reports 109 (May–June
1994): 339–60.
Ku, Leighton C.; Sonenstein, Freya L.; and Pleck, Joseph H. ‘‘The Associa-
tion of AIDS Education and Sex Education with Sexual Behavior and
Condom Use among Teenage Men.’’ Family Planning Perspectives 24
(May/June 1992): 100–106.
———. ‘‘Factors Influencing First Intercourse for Teenage Men.’’ Public
Health Reports 108 (November–December 1993): 680–94.
Lancaster, Tony. The Econometric Analysis of Transition Data. Cambridge:
Cambridge Univ. Press, 1990.
Marsiglio, William, and Mott, Frank L. ‘‘The Impact of Sex Education on
Sexual Activity, Contraceptive Use, and Premarital Pregnancy among
American Teenagers.’’ Family Planning Perspectives 18 ( July/August
1986): 151–62.
Meyer, Bruce D. ‘‘Unemployment Insurance and Unemployment Spells.’’
Econometrica 58 ( July 1990): 757–82.
Orr, Margaret T. ‘‘Sex Education and Contraceptive Education in U.S. Pub-

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms
644 journal of political economy
lic High Schools.’’ Family Planning Perspectives 14 (November/December
1982): 304–13.
Peltzman, Sam. ‘‘The Effects of Automobile Safety Regulation.’’ J.P.E. 83
(August 1975): 677–725.
Sonenstein, Freya L., and Pittman, Karen J. ‘‘The Availability of Sex Educa-
tion in Large City School Districts.’’ Family Planning Perspectives 16
( January/February 1984): 19–25.
Viscusi, W. Kip. ‘‘Do Smokers Underestimate Risks?’’ J.P.E. 98 (December
1990): 1253–69.
Zelnik, Melvin, and Kantner, John F. ‘‘Sexual Activity, Contraceptive Use
and Pregnancy among Metropolitan-Area Teenagers: 1971–1979.’’ Family
Planning Perspectives 12 (September/October 1980): 230–37.
Zelnik, Melvin, and Kim, Young J. ‘‘Sex Education and Its Association with
Teenage Sexual Activity, Pregnancy, and Contraceptive Use.’’ Family Plan-
ning Perspectives 14 (May/June 1982): 117–26.

This content downloaded from


186.96.209.78 on Mon, 22 May 2023 02:29:12 +00:00
All use subject to https://about.jstor.org/terms

You might also like