The effect of body weight on adolescent academic performance.

Sabia, Joseph J.

1. Introduction

A recent study by Cawley (2004) found evidence of a negative relationship between body weight and wages for white females, even after controlling for the endogeneity of body weight. If obesity causes white females' wages to be lower, this may reflect the presence of workplace discrimination against obese women or lower productivity levels for these workers, while the results presented in Cawley (2004) suggest that obesity may have an important negative economic effect. Our current understanding of the adverse economic impact of obesity may be understated if obesity also negatively affects early human capital accumulation. If increased body weight reduces the academic performance of adolescents or young adults, then the obesity-specific wage gap estimated by Cawley may reflect only part of the economic harm of obesity.

Exploring the effect of adolescent obesity on human capital accumulation is also important in the context of the current public policy environment. There are increased efforts by policy makers to fight childhood obesity by improving the nutritional quality of foods provided in public schools. For example, in September 2005, California governor Arnold Schwarzenegger signed legislation creating the nation's most rigorous nutrition standards in state public schools. Effective in July 2007, the new California law will limit the sale of many "junk foods," by regulating fat content, sugar content, and portion size. While the promotion of such school policies highlights the potential public health benefits of combating obesity, there may also be positive educational spillovers associated with improving adolescent body weight. In the context of the current policy environment for preventing and reversing childhood obesity, and building on the work of Cawley (2004), this study examines whether adolescent obesity adversely affects early human capital accumulation.

There are several reasons to expect a negative relationship between body weight and academic performance. First, it may be that poor academic performance causes higher body weight. This may be the case if, for example, adolescents choose to eat excessively to psychologically compensate for doing poorly in school. Second, obesity could cause a decline in academic performance. This could occur if teachers discriminate against overweight students by giving them poorer grades or if obesity has adverse psychological and physiological effects that impede productive studying. Finally, it may be that there is no causal link between body weight and academic performance, but rather an association that is explained by unobserved individual-level characteristics. For example, it may be that those with the least personal discipline expend the least amount of effort exercising and the least amount of effort studying.

Alternatively, there may be a positive relationship between body weight and academic performance. Poor academic performance may cause psychological stress, which reduces one's appetite and resultant body weight. Or, there may be a positive relationship between body weight and academic performance due to an unobserved heterogeneity. For example, if individuals must allocate their time between efforts to improve (or maintain) their physical well-being and efforts to enhance academic performance, then individuals with the least to gain from physical health investments (or the most to gain from investments in academic pursuits) may choose to devote more time and efforts toward academic endeavors and less toward monitoring and maintaining their weight. (1)

This paper examines the sensitivity of the association between adolescent body weight and academic performance to potential biases caused by unmeasured heterogeneity. Using the National Longitudinal Study of Adolescent Health, I estimate the relationship between several measures of adolescent body weight and grade point average (GPA). Ordinary least squares (OLS), instrumental variables (IV), and individual fixed effects (FE) estimates produce consistent evidence of a negative relationship between body weight and academic performance for white females aged 14-17. Conservative estimates reflect a difference in weight of 50 to 60 pounds (approximately two standard deviations) is associated with an 8 to 10 percentile difference in standing in the GPA distribution. For nonwhite females and white males, I find little evidence of a significant relationship between body weight and academic performance after controlling for unobserved heterogeneity. For nonwhite males, however, there is some evidence of a nonlinear relationship between body mass index (BMI) and GPA. Taken together, these findings indicate that adolescent obesity may have adverse academic consequences for white females. Thus, in principle, targeting obesity-reduction policies to adolescents may not only improve health outcomes, but may also have a positive impact on human capital accumulation.

2. Empirical Literature

Several empirical studies have found a negative association between adolescent obesity and academic achievement. Sargent and Blanchflower (1994) find that females who were obese at age 16 had lower reading and math test scores later in life than those who were not obese at the same age. Crosnoe and Muller (2004) show that adolescents in the 85th or higher percentile of the BMI distribution for their age-gender group have lower mean GPAs than those in the lower 85th percent of the distribution. They find that GPAs were even lower for obese adolescents in schools with higher rates of romantic relationships and lower average body size among students. Their results suggest that self-appraisal of weight, relative to one's peers, may have an independent effect on academic achievement.

Other related work has examined the relationship between obesity and educational attainment. Gortmaker et al. (1993) find that relative to women who were not overweight in 1981, women who were overweight in that same year had fewer years of education accumulated five years later. Sargent and Blanchflower (1994) also find that females who were obese at age 16 accumulated fewer years of schooling later in life than those who were not obese at age 16.

Each of these findings suggests evidence of a negative relationship between obesity and human capital accumulation, but in each of these studies, estimates could be biased by unmeasured characteristics associated with both obesity and educational attainment. With regard to heterogeneity bias, if the least disciplined individuals are most likely to become obese and to achieve less in school, and this level of personal discipline is unobservable, cross-section estimates of the effect of obesity on academic achievement will be biased upward. On the other hand, if unobserved time and effort must be allocated between monitoring or regulating one's physical health and investing in productive study time, and the most academically motivated individuals choose to devote more time to studying and less time to personal health care, then OLS estimates may be biased downward. Moreover, as noted above, it is not difficult to imagine reverse causality, whereby schooling outcomes could affect body weight.

While not specifically examining the relationship between body weight and academic achievement, related work has shown that the relationship between obesity and wages is quite sensitive to assumptions about unobservables (see, for example, Pagan and Davila 1997; Behrman and Rosenzweig 2001; Baum and Ford 2004; Norton and Han 2006). Using a sample of twins to control for unobserved family-level characteristics, Behrman and Rosenzweig (2001) find no significant relationship between obesity and wages. However, this may be due to the limited power of the design implied by small sample sizes. Pagan and Davila (1997) find a negative relationship between obesity and wages using instrumental variables. However, Cawley (2004) notes that their choice of instruments--health limitations and family poverty--may not be credible because they may be directly related to wages. Norton and Han (2006) use genetic information from specific genes linked to obesity as instruments in identifying the causal effects of obesity on female labor supply and wages. They find a small positive effect of obesity on employment probabilities and no effect on wages. While the instruments are credible, the relatively small sample size available for the wage equations (around 500) suggests that obesity effects may be imprecisely estimated.

Cawley (2004) provides convincing evidence of a significant negative relationship between body weight and wages using a large, nationally representative sample of workers from the National Longitudinal Survey of Youth (NLSY79). Cross-section estimates show a consistent negative relationship between obesity and wages for white women, Hispanic women, and black women. However, after including individual FE to control for fixed individual-level unobserved heterogeneity, he finds that the relationship between obesity and wages is only significant for white women, reflecting that selection on unobservables likely explains the strong association for black women and Hispanic women. Cawley finds similar results when he attempts to control for potential endogeneity bias with instrumental variables models, using a biological sibling's BMI as the key exclusion restriction.

Cawley (2004) offers one plausible explanation for race- and gender-specific differences in the effect of obesity on wages. He cites the sociological literature, which suggests that obesity may have a more adverse psychological effect on white women than on black and Hispanic women. In fact, for black and Hispanic females, Stearns (1997) finds that heavier weight is associated with greater self-perceived stability and power. Averett and Korenman (1996) find that obesity is correlated with lower self-esteem for white females (but not for nonwhite females or males), but that controlling for self-esteem does not fully explain race-specific differences in the association between obesity and wages.

This paper builds on the work by Cawley (2004) by examining whether adolescent obesity affects academic achievement. The current study is the first to explore the sensitivity of the association between body weight and academic performance to unmeasured heterogeneity. Using an IV strategy to control for endogeneity bias and individual FE models to control for time-invariant unobserved heterogeneity bias, this study presents evidence on the appropriateness of inferring a causal link between adolescent body weight and academic performance.

3. Methodology

OLS Model

The most common estimator used in the literature to identify the effect of obesity on education attainment is the OLS estimator, given by [[gamma].sub.1] in the following equation: (2)

[A.sub.1] = X'[[beta].sub.1] + [BW.sub.2][[gamma].sub.1] + [[epsilon].sub.1], (1)

where A is a measure of academic achievement, X is a vector of individual-level, family-level, and community-level observables, and BW is a measure of the adolescent's body weight. The estimate of [[gamma].sub.1] will only be an unbiased estimate of the effect of obesity on academic performance if there are no unobservable individual-, school-, or family-level characteristics correlated with both obesity and academic achievement; that is E([epsilon]\BW) = 0. If this identification assumption is violated, then the OLS estimator will be biased. This will be the case in the presence of endogeneity or heterogeneity bias.

IV Models

A common method of addressing endogeneity bias is through the use of instrumental variables. If the identification assumptions underlying the IV model are satisfied, then this estimate will control for any reverse causality whereby academic performance may cause changes in obesity. For example, poor grades may cause adolescents to psychologically compensate for their academic shortcomings by consuming more food. On the other hand, poor academic performance may cause an increase in stress, which may suppress appetite and reduce bodyweight. (3)

IV estimation requires finding observable characteristics that provide exogenous variation in adolescent body weight that are uncorrelated with academic achievement except through body weight. The two-stage least squares (2SLS) model jointly estimates the academic performance in Equation 1 and a body weight equation:

[BW.sub.2] = X'[[beta].sub.2] + [[epsilon].sub.2]. (2)

The classic IV identification assumption requires setting one or more elements in [[beta].sub.1] = 0. This implies that subset of X will serve as exclusion restrictions (Z) to identify the model.

Two exclusion restrictions are chosen for identification of the standard IV model: (i) the parent's report that the adolescent's biological mother suffers from an obesity problem, and (ii) the parent's report that the adolescent's biological father suffers from an obesity problem. (4) The identification assumption of the standard IV model first requires parental self-reports of obesity to be strongly correlated to/with adolescent obesity. This is theoretically expected to be the case because recent studies have found that approximately half of the variation in weight can be explained by genetics (Comuzzie and Allison 1998).

Identification of the model also requires that parental obesity not be correlated with unmeasured determinants of adolescent academic performance. This assumption may be problematic if parental obesity serves as a proxy for unobserved family-level environmental characteristics that are associated with schooling outcomes. For example, if parental obesity is correlated with a lack of motivation by parents and this motivation is both unmeasured and correlated with less monitoring of adolescents' school work, then parental obesity may have an independent effect on academic achievement, resulting in upwardly biased IV estimates.

However, there is some empirical evidence to suggest that a biologically related individual's BMI may serve as a credible instrument. Using the BMI of a sibling as his key exclusion restriction, Cawley (2004) provides a compelling case to suggest little empirical evidence of an effect of common household environment on body weight. In particular, he focuses on adoption studies that show that (i) the association in BMI between children and their biological parents is the same for children who live with their birth parents and those who live with adoptive parents. (Stunkard et al. 1986; Vogler et al. 1995), and (ii) the correlation in weight between biologically unrelated adopted children is statistically equal to zero (Grilo and Pogue-Geile 1991). This scientific evidence suggests that genetics rather than household environment is the most prominent influence on body weight.

Given the concern that parental obesity may be correlated with family-level schooling sentiment that could affect adolescents' academic performance, I control for several measures of family-level schooling sentiment. The Add Health dataset provides a rich set of family-level observable characteristics that capture schooling attitudes. These variables include: whether the parent moved to the neighborhood because of the quality of the local schools, whether the parent is a member of the Parent-Teacher Association, whether the parent prioritizes scholastic brilliance by their children, whether the mother has graduated from college, whether the parent talks with the adolescent about schoolwork, and the degree to which the parent monitors their child's curfew and friends.

While the scientific literature and the included schooling sentiment controls may enhance the credibility of the instrument exogeneity assumption, there remains a concern that parental obesity may capture unobserved genetic characteristics that contribute to both adolescent obesity and to adolescent academic performance. While I do control for a measure of innate academic intelligence using the student's Add Health Picture and Vocabulary Test (AHPVT) score, an unobserved genetic trait correlated with a genetic predisposition to obesity and a genetic predisposition to intelligence could lead to a violation of the identification assumption of the IV model. Because there are two exclusion restrictions for one potentially endogenous variable, overidentification tests are conducted to examine whether it is appropriate to reject the null hypothesis that the instruments are uncorrelated with the error term of the academic achievement equation.

Given concerns about the validity of the instruments described above, a second set of IV models are estimated that do not rely on the assumption that parental obesity is uncorrelated with adolescent academic performance. In the presence of heteroskedastic disturbances in Equation 2, Lewbel (2006) shows that [[gamma].sub.1] can be consistently estimated using (Z - [bar.Z])[[??].sub.2] as exclusion restrictions, where [[??].sub.2] are the estimated residuals from Equation 2, and Z is a vector of observed exogenous variables that can be a subset of X or can equal X. Thus, parental obesity is included in both Equations 1 and 2, with the identification of the parameters coming from heteroskedastic disturbances in Equation 2. Thus, the identification assumption requires cov (Z, [[epsilon].sup.2.sub.j]) = 0 (for j = 2 in the presence of a single unobserved common factor, and for both j = 1 and j = 2 in the presence of reverse causality) and cov (Z, [[epsilon].sub.1][[epsilon].sub.2]) = 0. Unlike the standard 2SLS case, no further restrictions are placed on Z. These assumptions required for the identification of this model are shown by Lewbel (2006) to be features that are quite common in models in which the correlation of errors ([[epsilon].sub.1] and [[epsilon[.sub.2]) is due to an unobserved common characteristic.

In models similar to Lewbel (2006), several researchers have exploited heteroskedasticity for identification (see, for example, King, Sentana, and Wadhwani 1994; Sentana and Fiorentini 2001; Rigobon 2002, 2003; Klein and Vella 2003). Learner (1981) and Feenstra (1994) also exploit heteroskedasticity to aid in identification. The assumptions underlying the Lewbel IV approach have also been exploited to identify correlated random coefficients models (Heckman and Vytlacil 1998). Moreover, as Lewbel (2006) notes, several recent papers have proposed restrictions on higher order moments rather than traditional instruments as an alternative method of identification (Vella and Verbeek 1997; Dagenais and Dagenais 1997; Lewbel 1997; Cragg 1997; Erickson and Whited 2002). The Lewbel IV approach will provide an additional identification method to test the robustness of standard IV results.

FE Model

Finally, while not explicitly controlling for reverse causality, an individual FE model will control for time-invariant individual-level unmeasured heterogeneity. The data used to estimate this model, described in the next paragraph, are from two waves of data collected in successive academic years. Thus, the individual FE model is a first differences model that uses individual-specific changes in weight and academic achievement to identify the effect of body weight on academic outcomes.

One limitation of estimating this model is that the inclusion of individual FEs may reduce the precision of estimates. For the individual FE design to be powerful enough to detect significant effects, there must be sufficient individual-specific variation in grades and body weight. Furthermore, in a first differences model utilizing two years of data, it is assumed that the effects of body weight on academic performance will occur contemporaneously. This may understate the true full effects of obesity on academic performance if there are important lagged effects.

The key advantage of a FE model is that it will control for individual-level characteristics that are unobservable to the researcher--such as unobserved discipline, psychological makeup, motivation, or genetic attractiveness--that may affect both body weight and academic performance. Given that the types of unobserved heterogeneity discussed to this point have been fixed individual-level unobserved characteristics, these FE estimates may produce unbiased estimates of the effect of body weight on academic achievement. However, if there are time-varying unobservables correlated with changes both in obesity and academic achievement, then FE estimates may be biased. (5)

While not achieving the level of internal validity that would be established in a randomized social experiment, the estimation of OLS, IV, and individual FE models will allow an examination of the sensitivity of the association between body weight and academic achievement to unobserved heterogeneity. These findings will inform the appropriateness of inferring a causal link between body weight and academic performance.

4. Data

This study utilizes data from the National Longitudinal Study of Adolescent Health (Add Health) to examine the relationship between adolescent obesity and academic achievement. The Add Health dataset is a school-based nationally representative longitudinal study that surveys adolescents enrolled in seventh to twelfth grade, their parents, and school administrators beginning in the 1994-1995 academic year. Wave 1 was collected during or just after the 1994-1995 academic school year, and Wave 2 was collected during the latter half or just after the 1995-1996 academic school year. Adolescents were asked questions on their education, health, family, romantic relationships, peer groups, neighborhoods, and sexual activity. Parents, mostly biological mothers, were also interviewed. Mothers were asked about their relationships with their children, their families, their backgrounds, their health status, and the health status of the adolescent's biological father.

The sample is restricted to 14-17-year-old adolescents, and the sample of females is further limited to those who reported not being pregnant, so as not to confound the effect of body weight on academic achievement with the impact of pregnancy. OLS, IV, and school FE models are estimated on a sample of 5129 adolescents from Wave 1 of the Add Health data. Individual FE models are estimated on a sample of 4218 adolescents with observations in both Wave 1 and Wave 2.

Dependent Variable

The dependent variable used to measure academic achievement is the combined math and English/language arts GPA. Adolescents are asked, "At (the most recent grading period/last grading period in the spring), what was your grade in --?" Questions are asked separately for math and English/language arts classes. (6) Adolescents could respond to this question with "A," "B," "C," or "D or lower." From these survey items, I created a measure of class-specific GPA, assigning a 4.0 for a reported grade of A, 3.0 for a reported grade of B, 2.0 for a reported grade of C, and 0.5 for a reported grade of D or lower. Then, giving each grade equal weight, I created an average math-English GPA for each adolescent.

In Table 1, I present weighted means and standard deviations of the dependent variable and all independent variables used in the OLS models, collected at baseline (Wave 1). Given the possible heterogeneous effects of body weight on academic achievement by sex and race, I conducted analyses separately for each of four subgroups. Thus, in Table 1, I present means for white (7) females, white males, nonwhite females, and nonwhite males. The mean math-English GPA (MEGPA) is highest for white females (2.93) and lowest for nonwhite males (2.44).

One limitation of this measure of the dependent variable is that it is self-reported. Thus, reported grade point averages may be upwardly biased if students provide inflated reports of their true grades. However, the average GPAs measured in the Add Health dataset do not appear to differ substantially from the National Longitudinal Survey of Youth 1979 or the High School and Beyond datasets, each of which provide transcript data. Oettinger (1999) reports the mean GPA for juniors in 1979-1983 to be 2.48. In the mid-1990s, the Add Health sample shows a mean MEGPA for juniors of 2.66 and a mean cumulative GPA of 2.79. The slightly higher GPAs reported in the Add Health dataset may be due to inflated reporting of grades or due to increased grade inflation over time.

Key Independent Variables

Following Cawley (2004), I use several measures of weight to estimate the effect of obesity on academic achievement. BMI is the standard measure of body weight in the epidemiological and medical literature, and was the key measure of body weight used by Cawley (2004). BMI is calculated as the body weight of an individual in kilograms divided by the height in meters squared. The mean BMI ranges from 21.6 for white females to 23.0 for nonwhite females and males.

The Centers for Disease Control and Prevention (CDC) provides age-sex specific measures of obesity for children aged 2-20. If an individual's BMI falls in the 5th percentile or lower in the age-sex specific BMI distribution, then the individual is clinically classified as underweight. If the individual's BMI falls in the 5th to 85th percentile, the individual is classified as having a normal body weight. An individual in the 85th to 95th percentile is classified as at-risk of being overweight (ATRISK). An individual in the 95th percentile or higher is classified as overweight (OBESE). In the sample, 1.7% of white females were underweight, 12.1% were at-risk of being overweight, and 4.9% were classified as overweight. A higher proportion of nonwhite females (28.3%) were at-risk of being overweight or overweight than white females (17%).

Another measure of obesity employed in this analysis is body weight in pounds, controlling for height in inches. The mean weight of white females was 128.2 pounds and for nonwhite females 132.2 pounds. The mean male weight was 154.4 pounds for whites and 149.8 for nonwhites. (8) Mean heights ranged from 5'4" for nonwhite women to 5'9" for white males.

Finally, the Add Health dataset asks adolescents about their perceptions of being overweight. If there are psychological effects resulting from obesity that cause adverse schooling outcomes, then self-perception of weight may be an important measure of obesity. Crosnoe and Muller (2004) suggest that the relative obesity of peers may influence the association between obesity and adverse schooling outcomes, thus implying a psychological component of the effect of obesity. Of all white females in the sample, 37.6% perceive themselves as overweight, compared with 14.5% of white females who are clinically defined as being at-risk for obesity or obese by the BMI classification. The correlation between these two measures for white females is just 0.48. Among those white females who believed they were overweight, only 36.6% were at risk of being overweight or overweight as classified by the BMI scale. Among those white females who were classified as overweight or overweight by the BMI scale, 94.3% believed they were overweight.

For other sex-race categories, 41.1% of nonwhite females perceived themselves as overweight compared with the 28.3% of nonwhite females who are classified as at-risk of being overweight or overweight (correlation = 0.54). The correlation between the two measures was higher for white males (0.59) than for nonwhite males (0.51).

Exclusion Restrictions

The two key instruments used in the IV models are (i) perceived obesity of the biological mother and (ii) perceived obesity of the biological father. The parent of the adolescent, usually the biological mother, is interviewed and asked, "Does the adolescent's biological mother now have [the health problem] of obesity?" The same question is asked of the parent with regard to the adolescent's biological father. Of the respondents with white female children, 19.6% reported that the biological mother had a current problem with obesity; 11.9% reported that the biological father had a problem with obesity.

One limitation of these instruments is that they are parental self-reports of perceived obesity rather than a direct measure of parental BMI. This could be problematic if there are unobserved sentiments correlated with a parental assessment of obesity that are also correlated with adolescent schooling outcomes. For example, parents who are more likely to report obesity problems might have other unobserved psychological problems that could negatively affect adolescents' academic performance. While direct measures of parental BMI would be preferred, given limitations of the data, such measures cannot be utilized. This data limitation, and its implication for the validity of the instruments, should be kept in mind when interpreting these IV estimates.

5. Empirical Results

Tables 2-5 present the main findings of this paper. (9) Only parameter estimates on the key body weight measures are presented in the tables, though a wide set of individual-level and household-level observables are included to control for characteristics that could influence both body weight and academic achievement. (10) Results presented are robust to choice of individual-and family-level covariates. Parameter estimates for non-obesity-related covariates are available in Appendix A.

OLS Estimates

Table 2 presents OLS estimates of the relationship between body weight and academic achievement for adolescents aged 14-17. Each column presents results from four separate regression models. These four regressions, which differ by the measure of body weight used, are estimated for each sex-race category. The first regression uses a continuous measure of BMI as the body weight measure (Row 1). The second model uses weight in pounds, controlling for inches (Row 2). The third specification uses the three CDC classifications of BMI: underweight, at-risk of being overweight, and overweight (Rows 3-5). The fourth regression model uses the adolescent's self-perception of being overweight (Row 6).

OLS findings for females are presented in Columns 1 and 2 of Table 2. In Column 1, I present results for white females. Under the assumption of a linear relationship between BMI and academic performance, there is a statistically significant negative relationship between BMI and grade point average for white females. For the average white female in the sample, a 50% increase in BMI would be associated with a 6.6% decline in GPA (approximately 0.2 GPA points). When obesity is measured as body weight in pounds, controlling for height in inches, I again find a significant negative relationship between weight and GPA for white females. Holding height constant, a 50-pound gain is associated with a 0.17 point decline in GPA.

In Rows 3-5, I allow the relationship between BMI on academic achievement to be nonlinear. Dummy variables are included for being underweight, at-risk of being overweight, and overweight, with the omitted category being those in the age-gender specific healthy BMI range. I find that, relative to healthy white females, overweight white females have a 0.182 point lower mean GPA. This finding reflects that much of the identifying variation in the relationship between BMI and GPA observed in Row 1 may be explained by variation in BMI at the high end of the BMI distribution. The findings in Models 1-3 confirm results in previous obesity-academic achievement studies (Sargent and Blanchflower 1994; Crosnoe and Muller 2004) and are consistent with recent obesity-wage studies (Pagan and Davila 1997; Cawley 2004).

In Row 6, I estimate the relationship between a perception of being overweight and GPA. Consistent with the other specifications, I find evidence of a significant negative relationship between the perception of being overweight and GPA. Controlling for observables, white females who perceive themselves to be overweight have a mean GPA that is 0.153 points lower than those who do not perceive themselves to be overweight.

While the relationship between body weight and academic performance is statistically significant for white females, the magnitude of the association appears to be quite small. It would take a weight difference of approximately 150 pounds (holding height constant) for there to be a one-half letter grade difference in GPA. Thus, my findings may appear to suggest that body weight is a significant, but practically unimportant, determinant of academic performance.

However, it is important to note that one need not observe GPA differences as large as one-half letter grade for there to be important economic consequences. For example, a weight difference of 55 pounds (approximately two standard deviations) is associated with a GPA difference of approximately 0.18 points. This represents an approximately 8 percentile lower ranking in the GPA distribution. (11) A GPA difference of this magnitude could have important effects on the quality of colleges to which a student could gain admission. Manski and Wise (1983) show that GPA differences approaching a decile could have nontrivial impacts on the probability of admissions to a high-quality college. To put the magnitude of this finding in context, Cawley finds that a 65-pound difference in weight (approximately two standard deviations) for adult white women is associated with a difference in wages of approximately 9%.

For nonwhite females (Column 2), the results are similar, with the magnitudes of the coefficients even larger than for whites. Whether body weight is measured by BMI or weight, controlling for height, there is a significant negative relationship between weight and academic achievement. As with white females, this relationship is driven by overweight nonwhite women. Relative to nonwhite females with a healthy BMI level, overweight nonwhite females have a mean GPA that is 0.273 points lower. However, unlike the finding for white females, I find that for black females, there is no significant correlation between the perception of being overweight and academic achievement.

Columns 3 and 4 present findings for males. In contrast to the findings for females, I find less consistent evidence of a significant relationship between weight and GPA for males. For white males (Column 3), after controlling for observables, I do not observe a significant negative relationship between obesity and academic achievement. This is consistent with findings by Sargent and Blanchflower (1994), who find no significant effect of obesity on reading and math scores for 16-year-old adolescent males. It is also consistent with Cawley (2004), who finds no relationship between obesity and wages for white males. For nonwhite males (Column 4), there is some evidence that weight may have nonlinear effects on academic achievement. Relative to nonwhite males with healthy BMI levels, obese nonwhite males have mean GPAs that are 0.180 points lower. However, nonwhite males who are underweight also have mean GPAs that are lower (0.472 points) than those who have healthy BMIs. This latter finding is consistent with Cawley (2004), who finds that underweight black males have lower mean wages than black males with healthy BMIs.

One must be cautious in interpreting OLS estimates causally because of the possibility of endogeneity bias or unobserved heterogeneity bias. The OLS estimates presented in Table 2 will only be unbiased estimates of the effect of obesity on academic achievement in the absence of reverse causality, whereby academic performance affects obesity, and if there are no unmeasured characteristics associated with both body weight and measured GPA. Thus, I present instrumental variables estimates and FE estimates to test the sensitivity of the OLS results to control for unmeasured heterogeneity.

Standard IV Estimates

To address the possibility of reverse causality in the relationship between obesity and academic performance, I estimate standard IV models, using the two measures of parental obesity as exclusion restrictions. These estimates appear in Table 3. OLS estimates are presented as well, so as to allow direct comparison. F-tests of the joint significance of the instruments are presented in the table along with p-values from the Sargan overidentification test and the [R.sup.2] from the first-stage models. Joint significance tests are consistently larger than 10 in most models, suggesting little evidence of weak instruments (Staiger and Stock, 1997). Moreover, p-values for Sargan overidentification tests are generally greater than 0.10, suggesting a failure to reject the null hypothesis that the exclusion restrictions are invalid.

After accounting for the endogeneity of body weight, I find consistent evidence of a negative relationship between body weight and GPA for white females (Column 2). Whether body weight is defined as BMI, weight in pounds (controlling for height), self-perception of being overweight, or obesity, this negative relationship persists. The findings are consistent with those in Cawley (2004), who finds negative effects of obesity on wages for white women, even after accounting for the endogeneity of weight via IVs. The IV estimates in Table 3 reflect little evidence that the endogeneity of body weight upwardly biases OLS estimates; in fact, Hausman tests reveal that the magnitudes of the coefficients are larger for standard IV models than for OLS models.

One interpretation of the differences in OLS and standard IV estimates for white females is that unobserved heterogeneity biases OLS estimates downward. For example, it may be that poor academic performance causes a reduction in bodyweight among white females. This might occur if, for example, poor grades cause psychological stress among white females that causes them to lose weight. However, it may be that the exclusion restrictions are correlated with other unobserved characteristics that could also impact adolescent academic achievement. While there is reason to believe that household-level environmental characteristics associated with parental obesity do not have an independent effect on body weight, as evidenced by much of the scientific literature (see, for example, Stunkard et al. 1986; Grilo and Pogue-Geile 1991; Vogler et al. 1995), parental self-reports of obesity may still be correlated with unobserved genetic determinants of adolescent academic performance. Moreover, parental self-reports of obesity may be correlated with unobserved household-level psychological effects that may adversely affect academic performance.

While the exclusion restrictions are strong predictors of body weight for white females (F-statistics between 17 and 29) and also satisfy the Sargan overidentification test, these statistical tests do not guarantee that the identification assumptions are not violated. The overidentification test may, in fact, be insufficiently powerful to detect weak instruments. In fact, the first-stage [R.sup.2] values are quite low, reflecting that the instruments may, indeed, be weak, resulting in large size distortions (Stock and Yogo 2004). Hence, the IV estimates in Column 2 should be cautiously interpreted.

For nonwhite females (Columns 4 and 5), OLS estimates of the relationship between body weight and academic performance are significant, but IV estimates are not. One interpretation of these findings is that unobserved heterogeneity explains the OLS estimates and that there is no causal link between obesity and GPA after controlling for the endogeneity of body weight. This interpretation is consistent with evidence from Cawley (2004), who finds that after controlling for the endogeneity of body weight, there is no significant association between obesity and wages. However, it is important to note that, unlike the findings for white females, the magnitudes of the OLS and IV estimates are quite similar for three of four specifications. IV estimates are insignificant due to inflated standard errors.

For white males (Columns 7 and 8), IV estimates are consistent with OLS estimates, indicating no significant relationship between obesity and academic achievement for white males. These findings are also consistent with the obesity-wage findings in Cawley (2004). For nonwhite males (Columns 10 and 11), however, I find evidence of a negative effect of weight on academic achievement. Whether body weight is measured via BMI, weight in pounds controlling for height, obesity, or perception of weight, there is a significant negative relationship between obesity and academic achievement for nonwhite males. However, I note that the instruments appear to be quite weak predictors of adolescent body weight in two of the four models.

Taken together, the IV estimates suggest evidence of a significant negative relationship between body weight and academic performance for white females and nonwhite males, but not for nonwhite females or white males. However, these estimates should be cautiously interpreted given the data available for the exclusion restrictions.

Lewbel IV Estimates

Because parental obesity may be directly correlated with adolescent academic performance, the Lewbel IV approach provides a method by which these measures can also be included in the academic performance equation. In these models, the elements in Z are included in X (i.e., the parental obesity measures), with heteroskedasticity in the body-weight equation used to identify the model. Across all specifications, Breusch-Pagan heteroskedasticity tests reveal the presence of heteroskedasticity in the first-stage body weight equation, as well as in the second stage.

Lewbel IV estimates are generally consistent with OLS estimates and are much smaller in magnitude than standard IV estimates. For white females (Column 3), the Lewbel IV estimates are closer in magnitude to OLS estimates than standard IV estimates in most cases. However, the standard errors are larger, leading to findings that are not consistently statistically significant. For nonwhite females (Column 6), there is inconsistent evidence of a negative relationship, but the p-values on the Sargan overidentification test suggest that the Lewbel instruments may not be uncorrelated with the error of the academic performance equation. The Lewbel results for white males and nonwhite males are generally consistent with OLS estimates. For white males, there is no evidence of a significant relationship between body weight and academic performance, while for nonwhite males, parameters on the body weight measures are negative and significant. (12,13)

Taken together, the Lewbel IV estimates are smaller in magnitude than standard IV estimates. The [R.sup.2] values on the first-stage equations are generally higher than those in the standard IV models, suggesting stronger instruments in some cases. The magnitudes of the Lewbel estimates appear more plausible and, in some specifications, suggest that reverse causality may not be a sufficient explanation for the negative relationship between body weight and academic performance for white females and for nonwhite males.

Individual FE

While not explicitly controlling for reverse causality, individual FE models control for individual-specific time-invariant unobserved heterogeneity. This is the form of unobserved heterogeneity--in the form of motivation, discipline, and self-esteem--that is theoretically posited to be the most likely source of heterogeneity bias in OLS estimates. These estimates are obtained by exploiting the longitudinal nature of the Add Health data, which have repeated observations on the same individuals in successive school years. Data from Waves 1 and 2 are used to estimate first difference models.

Identification of the effects of body weight on academic performance requires sufficient within-person variation in weight and GPA. (14) The variation in GPA across Waves 1 and 2, as well as the variation in key independent variables of interest are presented in Appendix B. While mean values have not changed much across successive waves, this masks some important person-specific variation across waves. The first difference models control for several time-varying observable characteristics theoretically believed to affect changes in GPA and body weight. (15)

Table 4 presents FE results along with OLS estimates for comparison. The FE sample is slightly smaller than the IV sample because the FE sample requires nonmissing observations on academic performance and obesity in both waves of data. OLS estimates on the FE sample are generally similar to those obtained with the previous, larger sample. For white females (Columns 1 and 2), I find that individual FE estimates are generally consistent with OLS estimates. Consistently, I find a significant negative relationship between body weight and academic achievement. Whether body weight is defined as BMI weight in pounds, controlling for height, or the perception of being overweight, FE models consistently show that an increase in body weight is associated with a decline in academic achievement. The magnitudes of the coefficients are generally statistically equivalent to OLS estimates. The one area where there is some difference between OLS and individual FE estimates is in the model that assumes a nonlinear relationship between body weight and academic achievement. While the parameter estimates on ATRISK and OBESE are negative, as expected, neither is significant. This likely occurs because there is insufficient within-person variation in OBESE to identify the model, as shown in Appendix B. When the definition of obesity is expanded to include those white females with BMIs in the 85th percentile or higher (ATRISK or OBESE) so as to allow additional identifying variation, I find that being at risk of being overweight or being overweight is associated with a statistically significant 0.18 point lower GPA. (16) One important caveat to the first differences model is that it assumes that any effects of bodyweight on academic performance occur contemporaneously.

Thus, if there are lagged adverse effects of obesity on academic performance, then individual FE estimates may be understated. (17)

In summary, for white females, OLS, IV, and individual FE estimates each suggest strong evidence of a negative relationship between body weight and academic performance. (18) OLS, Lewbel IV, and individual FE estimates reflect that a difference in body weight of 50 pounds (approximately two standard deviations) is associated with a 0.15 to 0.2 point difference in GPA. These findings on academic performance are consistent with Cawley (2004), who found that the negative relationship between obesity and wages for white women was robust to controls for unmeasured heterogeneity.

For nonwhite females, however, the evidence from individual FE models suggests that unobserved heterogeneity is the likely explanation for the negative association between obesity and academic achievement observed in the cross section. FE results for nonwhite females are presented in Columns 3 and 4. While OLS estimates consistently show a significant negative relationship, FE estimates show no significant association. This suggests that time-invariant unobservables rather than a causal link likely explains the positive association observed by OLS estimates. These findings are generally consistent with IV estimates, suggesting little evidence that obesity causes lower academic achievement for nonwhite females. These results are also consistent with Cawley (2004), who finds little evidence of a negative relationship between weight and wages for black females or Hispanic females.

For white males, there is little consistent evidence of a negative relationship between obesity and academic achievement (see Columns 5-6) in either OLS or individual FE estimates. For nonwhite males, I find no evidence of a significant negative relationship between obesity and academic achievement after controlling for individual-level time-invariant unobservables. In fact, there is some evidence that being underweight, relative to having a healthy BMI, is associated with lower academic achievement. This curious result is consistent with Cawley's (2004) findings that being underweight was associated with lower intelligence levels and educational attainment for black males. This finding is consistent with the hypothesis that, for nonwhite males, being underweight may actually result in a social stigma that could adversely affect academic performance.

Taken together, the estimation results presented in Tables 2-4 reflect that the negative relationship between obesity and academic achievement for white females is robust to controls for endogeneity bias and heterogeneity bias. I find no evidence that higher body weight causes lower grade point averages among nonwhite females. For males, the evidence is less clear. While there is some evidence of a nonlinear relationship between body weight and academic achievement, this relationship needs further empirical exploration given the inconsistency of IV and FE estimates.

Robustness of Findings

Given the strong evidence of a negative relationship between body weight and academic performance for white females, understanding the mechanisms by which body weight may cause a contemporaneous decline in grades is important. There are two hypotheses in the literature that may explain such a causal mechanism. First, there is a possibility of teacher-specific discrimination against overweight white females. Data limitations in Add Health preclude estimation of teacher FE models, which could shed some light on this question. If the inclusion of teacher effects reduced the magnitude and significance of the estimated relationship between body weight and GPA, this would suggest some evidence in support of the discrimination hypothesis.

While not permitting teacher FE models, the Add Health data do have information from the student on whether she gets along with her teachers. One might expect that in the presence of teacher discrimination, students might express displeasure at being treated unfairly. If the inclusion of this measure of the student's relationship with her teachers reduces the magnitude and significance of the relationship between body weight and wages, this may suggest some support for the discrimination hypothesis. However, when this variable is included in the individual FE model, the magnitude and significance of the relationship does not change, as shown in the first and second rows of Table 5. While this finding does not rule out the possibility of teacher-specific discrimination, it does cast some doubt on this hypothesis.

Another mechanism by which body weight may influence academic performance is suggested in the sociology literature. This literature suggests that the mental health or self-esteem of white women may be adversely affected by obesity, which, in turn, may affect academic performance. The Add Health data allows me to test the robustness of the relationship between body weight and academic performance to control for observed measures of mental health. Rows 3-5 of Table 5 present results of alternate individual FE specifications that control for several measures of adolescent mental health: self-assessed depression, frequent reports of "having the blues," loneliness, frequent difficulty with paying attention in class, and thoughts of suicide. Across all models, the relationship between BMI (or weight in pounds, controlling for height) and grade point average remains negative and significant after controlling for observed measures of mental health and physical health.

An alternative path through which obesity may reduce academic performance is through physical impairment. For example, the medical and public health literatures suggest a strong link between obesity and sleep apnea in adults (see, for example, Young et al. 1993). There is some evidence that sleep disorders of this sort may have significant adverse effects across several measures of cognition (Kales et al. 1985; Bedard et al. 1993; Adams et al. 2001; Naismith et al. 2004; El-Ad and Lavie 2005). While the Add Health data do not allow a direct test of whether obesity affects sleep disorders, adolescents are asked questions about whether they perceive themselves as being in good health, whether they are tired frequently (almost every day or every day), and whether they frequently wake up tired. When these variables are included in the model (see Rows 6 and 7 of Table 5), the magnitude of the coefficients on the key body weight measures remains unchanged. (19) The results in Table 5 suggest the importance of future research to understand the psychosocial and physical mechanisms by which obesity adversely affects the academic performance of white females.

6. Conclusions

Building on the work of Cawley (2004), this paper examines the relationship between adolescent body weight and academic performance. OLS, IV, and individual FE estimates suggest robust evidence of a negative relationship between body weight and academic achievement among white females. There is little consistent evidence of a robust negative relationship between body weight and academic performance among males or nonwhite women after controlling for various forms of unmeasured heterogeneity. These findings are consistent with the obesity wage findings in Cawley (2004), and can be interpreted in several ways. First, the results may suggest that the obesity-specific wage differential observed for white females can be partially explained by differences in human capital accumulation. Second, it may be that body weight impacts a common unobserved factor that affects both academic achievement and wages, such as self-esteem. Finally, the results may suggest that there is school- and market-level discrimination against obese white women.

The interpretation of results in this study requires a few important caveats. While statistically significant, the magnitude of the effect of changes in body weight on academic performance is likely to be rather small for the average white female. A difference of 50 to 60 pounds is associated with a 0.2 point difference in GPA. A body weight difference of this magnitude is quite large (approximately two standard deviations), but the impact of such a weight difference on GPA standing may not be trivial. A difference in GPA of 0.2 points is associated with an approximately 10 percentile difference in a student's position in the GPA distribution. If this effect is not transitory, but rather represents a permanent GPA difference, this could significantly affect the quality of college to which an adolescent could gain admittance (Manski and Wise 1983).

Future research is important in understanding the mechanisms by which obesity reduces academic performance among white adolescent females. Data that allow the inclusion of teacher FE would allow tests of whether teacher-specific fixed unobservables reduces the magnitude and significance of the relationship between body weight and academic performance. If the relationship were weakened after accounting for teacher-level unobserved heterogeneity, this would strengthen the hypothesis that teacher-specific discrimination explains the relationship.

However, in the absence of support for a discrimination hypothesis, further research is needed on the psychological or physiological mechanisms by which obesity may affect the human capital accumulation of white females. While I found little evidence that controlling for various measures of depression and self-worth reduces the magnitude or significance of the relationship between body weight and academic performance, increases in body weight may have important contemporaneous effects on other psychological traits--such as unmeasured self-esteem or stress--that affect academic performance of white females. Future work in this area would be an important contribution to the literature.

Regardless of the mechanism through which obesity affects adolescent human capital accumulation, the evidence presented here suggests that targeting anti-obesity efforts toward children could, in principle, enhance the collective human capital of the United States. However, in the absence of market failures, adolescent and parental choices over the production of health and human capital may result in outcomes that are socially efficient. If, however, the presence of imperfect health information or schooling externalities precludes the achievement of socially optimal private decision making, school policies aimed at improving nutrition or enhancing physical education could, in principle, enhance efficiency if the social benefits of such policies exceeded the social costs. Thus, in future research on this question, addressing the social welfare implications of policy changes will be important.

Appendix A
OLS Estimates of Relationship between BMI and Academic Performance

 White Females (1) Nonwhite Females (2)

BMI -0.018 *** (0.007) -0.015 ** (0.007)
SPORT 0.033 (0.065) 0.047 (0.062)
EXERCISE 0.100 * (0.055) 0.082 (0.056)
ASPIRE 0.391 *** (0.076) 0.455 *** (0.082)
AHPVT 0.018 *** (0.002) 0.007 *** (0.002)
PUBLIC -0.065 (0.083) -0.268 ** (0.110)
RURAL 0.141 (0.089) -0.051 (0.107)
SUBURBAN 0.045 (0.075) 0.001 (0.067)
SOUTH 0.062 (0.077) 0.070 (0.093)
WEST 0.021 (0.086) 0.012 (0.101)
MIDWEST -0.029 (0.077) 0.145 (0.112)
PARDISCOL -0.085 (0.059) -0.073 (0.074)
PARTEACH 0.097 * (0.059) 0.089 (0.064)
NGHBRHD 0.121 ** (0.057) 0.026 (0.060)
BRILLIANT 0.118 ** (0.054) -0.047 (0.062)
PARPROJECT -0.074 (0.119) 0.167 * (0.095)
PARTALK 0.033 (0.058) 0.141 ** (0.072)
SINGLEPAR -0.305 *** (0.108) -0.043 (0.080)
COLGRAD 0.185 *** (0.062) 0.205 *** (0.072)
PARWORK 0.002 (0.063) -0.132 * (0.073)
CURFEW 0.150 *** (0.057) -0.097 (0.063)
DINNERWK 0.021 * (0.013) 0.035 *** (0.012)
MEAT 0.108 (0.125) -0.082 (0.056)
RELIGIONWK 0.065 (0.075) -0.000 (0.088)
RELIGIONMO 0.044 (0.090) -0.011 (0.097)
RELIGIONYR -0.091 (0.089) 0.069 (0.105)
ROMANTIC 0.022 (0.068) -0.031 (0.064)
INTERCOURSE -0.080 (0.075) -0.180 ** (0.073)
OLDERSIB -0.114 * (0.056) -0.067 (0.060)
HHINC 0.007 (0.040) 0.004 (0.041)
DRINK -0.139 ** (0.060) -0.116 * (0.064)
AGE15 -0.017 (0.073) -0.082 (0.070)
AGE16 0.029 (0.081) 0.082 (0.089)
AGE17 -0.085 (0.095) 0.114 (0.097)
BLACK -- -0.326 *** (0.089)
HISPANIC -- -0.402 *** (0.087)
N 1472 1059

 White Males (3) Nonwhite Males (4)

BMI 0.001 (0.006) -0.020 *** (0.006)
SPORT 0.143 * (0.080) 0.211 * (0.107)
EXERCISE 0.072 (0.060) -0.025 (0.065)
ASPIRE 0.379 *** (0.069) 0.248 *** (0.076)
AHPVT 0.011 *** (0.003) 0.012 *** (0.002)
PUBLIC -0.113 (0.093) -0.145 (0.128)
RURAL 0.046 (0.088) 0.212 * (0.124)
SUBURBAN 0.036 (0.076) 0.118 (0.073)
SOUTH 0.021 (0.081) -0.099 (0.105)
WEST 0.061 (0.097) -0.287 ** (0.100)
MIDWEST 0.174 *** (0.082) -0.173 (0.118)
PARDISCOL 0.020 (0.060) 0.053 (0.068)
PARTEACH 0.111 * (0.062) -0.029 (0.070)
NGHBRHD 0.147 ** (0.059) 0.046 (0.061)
BRILLIANT -0.012 (0.060) 0.050 (0.071)
PARPROJECT 0.114 (0.097) 0.085 (0.114)
PARTALK 0.138 ** (0.063) -0.007 (0.069)
SINGLEPAR -0.033 (0.106) -0.213 ** (0.095)
COLGRAD 0.082 (0.066) 0.067 (0.076)
PARWORK -0.049 (0.066) -0.123 * (0.076)
CURFEW 0.024 (0.060) 0.133 ** (0.068)
DINNERWK 0.028 (0.016) 0.025 * (0.013)
MEAT -0.046 (0.015) -0.193 ** (0.080)
RELIGIONWK 0.038 (0.076) 0.038 (0.095)
RELIGIONMO 0.081 (0.086) -0.058 (0.106)
RELIGIONYR -0.032 (0.092) -0.000 (0.115)
ROMANTIC -0.012 (0.060) 0.044 (0.066)
INTERCOURSE -0.283 *** (0.077) -0.265 *** (0.071)
OLDERSIB -0.016 (0.060) -0.020 (0.065)
HHINC 0.122 *** (0.042) -0.014 (0.039)
DRINK -0.124 ** (0.062) -0.232 *** (0.068)
AGE15 0.060 (0.083) -0.050 (0.109)
AGE16 0.055 (0.083) 0.011 (0.107)
AGE17 0.181 ** (0.089) -0.011 (0.120)
BLACK -- -0.050 (0.122)
HISPANIC -- -0.250 ** (0.116)
N 1561 1055

All models also include a set of dummy variables for parental
monitoring. Nonwhite models include dummy variables for black
and Hispanic identified youth.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Appendix B1
Coefficient Estimates on Instruments in First-Stage of 2SLS
Models for Females

 White

 (1) (3) (5) (7)
 BMI POUNDS OBESE PEROVER

OBESEMOM 2.164 *** 12.65 *** 0.078 *** 0.190 ***
 (0.255) (1.52) (0.012) (0.033)
OBESEDAD 1.163 *** 6.91 *** 0.067 *** 0.081 **
 (0.309) (1.830) (0.015) (0.040)
N 1472 1472 1472 1472

 Nonwhite

 (9) (11) (13) (15)
 BMI POUNDS OBESE PEROVER

OBESEMOM 2.903 *** 17.47 *** 0.172 *** 0.211 ***
 (0.507) (2.970) (0.041) (0.059)
OBESEDAD 1.714 ** 10.51 *** 0.084 * 0.126 *
 (0.759) (4.370) (0.053) (0.079)
N 1059 1059 1059 1059

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Appendix B2
Coefficient Estimates on Instruments in First-Stage of 2SLS
Models for Males

 White

 (1) (3) (5) (7)
 BMI POUNDS OBESE PEROVER

OBESEMOM 2.564 *** 17.61 *** 0.140 *** 0.186 ***
 (0.285) (1.980) (0.016) (0.028)
OBESEDAD 2.095 *** 14.11 *** 0.107 *** 0.107 **
 (0.342) (2.37) (0.019) (0.033)
N 1561 1561 1561 1561

 Nonwhite

 (1) (3) (5) (7)
 BMI POUNDS OBESE PEROVER

OBESEMOM 2.465 *** 17.14 *** 0.150 *** 0.092 *
 (0.592) (4.13) (0.048) (0.048)
OBESEDAD 1.163 * 9.12 * 0.136 ** 0.129 **
 (0.709) (4.94) (0.063) (0.064)
N 1055 1055 1055 1055

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Appendix C
Means and Variation in Key Dependent and Independent Variables
Used in Individual FE Model

 White Females

 Wave 1 Wave 2

MEGPA 2.97 (0.874) 2.98 (0.835)
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 0.5] 64.7
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 24.6
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.5] 7.5
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 2.3

BMI 21.51 (3.74) 21.80 (4.07)
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 38.6
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 13.9
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 3.0] 6.2
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 4.0] 3.3

POUNDS 128.05 (24.50) 131.76 (26.55)
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 10] 28.0
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 20] 6.0
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 30] 2.5
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 40] 1.2

INCHES 64.7 (2.80) 65.0 (2.75)
UNDER 0.010 (0.090) 0.028 (0.166)
ATRISK 0.124 (0.329) 0.100 (0.300)
OBESE 0.047 (0.212) 0.058 (0.235)
% who [DELTA] BMI Category 12.7
% who [DELTA] in/out of ATRISK 8.8
% who [DELTA] in/out of OBESE 2.5
PEROVER 0.373 (0.484) 0.349 (0.479)
% who [DELTA] PEROVER 17.1
N 1209 1209

 Nonwhite Females

 Wave 1 Wave 2

MEGPA 2.77 (0.894) 2.74 (0.844)
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 0.5] 69.9
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 28.8
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.5] 7.4
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 1.8

BMI 22.71 (4.48) 22.93 (4.78)
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 43.9
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 19.3
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 3.0] 9.9
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 4.0] 5.0

POUNDS 131.31 (27.64) 135.46 (30.52)
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 10] 29.0
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 20] 5.8
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 30] 2.6
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 40] 0.9
INCHES 63.7 (2.91) 64.2 (2.83)
UNDER 0.019 (0.136) 0.036 (0.186)
ATRISK 0.169 (0.375) 0.150 (0.358)
OBESE 0.098 (0.298) 0.083 (0.277)
% who [DELTA] BMI Category 16.5
% who [DELTA] in/out of ATRISK 12.5
% who [DELTA] in/out of OBESE 3.6
PEROVER 0.399 (0.490) 0.395 (0.489)
% who [DELTA] PEROVER 15.9
N 823 823

 White Males

 Wave 1 Wave 2

MEGPA 2.78 (0.940) 2.77 (0.909)
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 0.5] 70.5
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 32.2
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.5] 12.8
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 5.4

BMI 22.45 (4.43) 23.05 (4.60)
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 49.7
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 22.3
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 3.0] 9.4
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 4.0] 4.9

POUNDS 152.92 (36.60) 162.00 (36.71)
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 10] 47.1
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 20] 16.7
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 30] 5.3
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 40] 1.6
INCHES 69.0 (3.62) 70.0 (3.18)
UNDER 0.023 (0.150) 0.022 (0.148)
ATRISK 0.149 (0.356) 0.149 (0.356)
OBESE 0.137 (0.344) 0.132 (0.339)
% who [DELTA] BMI Category 16.7
% who [DELTA] in/out of ATRISK 13.8
% who [DELTA] in/out of OBESE 5.2
PEROVER 0.225 (0.418) 0.224 (0.412)
% who [DELTA] PEROVER 11.2
N 1339 1339

 Nonwhite Males

 Wave 1 Wave 2

MEGPA 2.51 (0.935) 2.43 (0.923)
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 0.5] 67.7
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 29.3
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 1.5] 12.1
% with GPA [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 3.9

BMI 22.67 (4.33) 23.17 (4.40)
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 1.0] 46.1
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 2.0] 19.6
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 3.0] 8.9
% with BMI [DELTA]
 [greater than or equal to]
 [absolute value of 4.0] 4.1

POUNDS 149.69 (33.38) 157.44 (33.94)
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 10] 47.6
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 20] 14.7
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 30] 4.9
% with POUNDS [DELTA]
 [greater than or equal to]
 [absolute value of 40] 1.9

INCHES 68.0 (3.43) 68.9 (3.160
UNDER 0.014 (0.117) 0.020 (0.140)
ATRISK 0.177 (0.382) 0.142 (0.350)
OBESE 0.133 (0.340) 0.132 (0.338)
% who [DELTA] BMI Category 15.7
% who [DELTA] in/out of ATRISK 14.1
% who [DELTA] in/out of OBESE 4.7
PEROVER 0.200 (0.400) 0.203 (0.402)
% who [DELTA] PEROVER 12.7
N 822 822

The author wishes to thank two anonymous referees, Julie Hotchkiss, and participants at the March 2006 meeting of the Georgia Policy Leadership for Active Youth (PLAY) for helpful comments and suggestions. Thanks also to Nikki Williams for excellent editorial assistance. This research uses data from Add Health, a program project designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullah Harris, and funded by a grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Persons interested in obtaining data files from Add Health should contact Add Health, Carolina Population Center, 123 W. Franklin Street, Chapel Hill, NC 27516-2524 (http://www.cpc.unc.edu/addhealth/contract.html).

Received November 2005; accepted June 2006.

References

Adams, Nancy, Milton Strauss, Mark Schluchter, and Susan Redline. 2001. Relation of measures of sleep-disordered breathing to neuro-psychological functioning. American Journal of Respiratory and Critical Care Medicine 163:1626-31.

Aschenfelter, O., and A. Krueger. 1994. Estimates of the economic return to schooling from a new sample of twins. American Economic Review 84:1157-74.

Averett, Susan, and Sanders Korenman. 1996. The economic reality of the beauty myth. Journal of Human Resources 31:304-30.

Baum, C. L., and W. F. Ford. 2004. The wage effects of obesity: A longitudinal study. Health Economics 13:885-99.

Bedard, M. A., J. Montplaisir, J. Malo, F. Richer, and I. Rouleau. 1993. Persistent neuropsychological deficits and vigilance impairments in sleep apnea syndrome after treatment with continuous positive airway pressure. Journal of Clinical and Experimental Neuropsychology 15:330-41.

Behrman, Jere R., and Mark R. Rosenzweig. 2001. The returns to increasing body weight. Unpublished paper, University of Pennsylvania: Philadelphia, PA.

Bound, John, Charles Brown, and Nancy Mathiowetz. 2002. Measurement error in survey data. In Handbook of econometrics 5, edited by James Heckman and Ed Learner. New York: Springer-Verlag, pp. 3705-843.

Cawley, John. 2004. The impacts of obesity on wages. Journal of Human Resources 39:451-74.

Comuzzie, Anthony G., and David B. Allison. 1998. The search for human obesity genes. Science 280:1374-77.

Cragg, J. 1997. Using higher moments to estimate the simple errors-in-variables model. Rand Journal of Economics 28:S71-91.

Crosnoe, Robert, and Chandra Muller. 2004. Body mass index, adolescent achievement, and school context: Examining the educational experiences of adolescents at risk of obesity. Journal of Health and Social Behavior 45:393-407.

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

El-Ad, Baruch, and Peretz Lavie. 2005. Effect of sleep apnea on cognition and mood. International Review of Psychiatry 17:277-82.

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

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

Gortmaker, Steven L., Aviva Must, James M. Perrin, Arthur M. Sobol, and William H. Deitz. 1993. Social and economic consequences of overweight in adolescence and young adulthood. New England Journal of Medicine 329:1008-12.

Grilo, Carlos M., and Michael F. Pogue-Geile. 1991. The nature of environmental influences on weight and obesity: A behavioral genetic analysis. Psychological Bulletin 110:520-37.

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

Kales, A., A. B. Caldwell, R. J. Cadieux, A. Vela-Bueno, L. G. Ruch, and S. D. Mayes. 1985. Severe obstructive sleep apnea II: Associated psychopathology and psychosocial consequences. Journal of Chronic Diseases 38:427-34.

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

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

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

Lee, Lung-fei, and Jungsywan H. Sepanski. 1995. Estimation of linear and nonlinear errors-in-variables models using validation data. Journal of the American Statistical Association 90:130-40.

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

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

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

Naismith, S., V. Winter, H. Gotsopoulos, I. Hickie, and P. Cistulli. 2004. Neurobehavioral functioning in obstructive sleep apnea: Differential effects of sleep quality, hypoxemia and subjective sleepiness. Journal of Clinical and Experimental Neuropsychology 26:3-54.

Norton, E., and E. Han. 2006. Using genetic information to identify causal effects of obesity on female labor market outcomes. Unpublished manuscript, University of North Carolina-Chapel Hill.

Oettinger, Gerald S. 1999. Does high school employment affect high school academic performance? Industrial and Labor Relations Review 53:136-51.

Pagan, Jose A., and Alberto Davila. 1997. Obesity, occupational attainment, and earnings. Social Science Quarterly 8:756-70.

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

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

Sargent, J. D., and D. G. Blanchflower. 1994. Obesity and stature in adolescence and earnings in young adulthood: Analysis of a British birth cohort. Archives of Pediatrics and Adolescent Medicine 148:681-87.

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

Staiger, Douglas, and James H. Stock. 1997. Instrumental variables regression with weak instruments. Econometrica 65:557-86.

Stearns, Peter N. 1997. Fat history: Bodies and beauty in the modern West. New York: New York University Press.

Stock, J. H., and M. Yogo. 2004. Testing for weak instruments in IV regression. In Identification and inference for econometric models: A festschrift in honor of Thomas Rothenberg. Cambridge, MA: University Press.

Stunkard, A. J., T. I. A. Sorensen, C. Hanis, T. W. Teasdale, R. Chakraborty, W. J. Schull, and F. Schulsinger. 1986. An adoption study of human obesity. New England Journal of Medicine 314:193-96.

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

Vogler, G. P., T. I. A. Sorensen, A. J. Stunkard, M. R. Srinivasan, and D. C. Rao. 1995. Influences of genes and shared family environment on adult body mass index assessed in an adoption study by a comprehensive path model. International Journal of Obesity 19:40-5.

Young, T., M. Palta, J. Dempsey, J. Skatrud, S. Weber, and S. Badr. 1993. The occurrence of sleep-disordered breathing among middle-aged adults. New England Journal of Medicine 328:1230-35.

(1) For example, consider a student lacking innate athletic ability. Because such a student has a very low propensity to excel athletically or obtain an athletic scholarship, she may choose to devote more time to studying and less to participating in athletic activities, which could, in turn, raise her body weight. High schools and colleges clearly recognize the trade-off students must make between time spent on athletic endeavors and time spent studying. Athletes are often required to meet some minimum grade standards to remain eligible for athletic pursuits.

(2) An adolescent (or parent) is assumed to maximize the adolescent's utility, which is a function of human capital, health, and other consumption goods. Equation 1 may be interpreted as the production function for human capital, measured by academic performance, and Equation 2 to follow may be interpreted as the production function for health, measured by body weight.

(3) Alternatively, it may be that poor academic performance causes adolescents to reprioritize time away from scholastic endeavors and toward more productive athletic pursuits, which may reduce body weight.

(4) The parent that is most often interviewed is the biological mother.

(5) Thus, for example, if there are unobserved within-person changes in personal discipline or motivation, these changes could lead to biased estimates.

(6) Questions were also asked for history and science classes, but I chose to focus on grades in math and English courses for two reasons. First, math and English are the core courses taken by most adolescents in both waves of data collection. Including history and science classes in the GPA calculation in each wave (necessary for the FE estimation) would reduce the sample size substantially (24%), thereby diminishing the power of the design, which is of particular concern in this study because I endeavor to produce estimates separately by race and sex. Second, among those who have taken all four courses, the math/English/language arts GPA captures 78% of the variation in overall GPA. Thus, the gain in power is likely worth the reduction in precision in GPA measurement. However, the central findings in this study are generally robust to choice of dependent variable, with FE estimates most affected because of the dramatic reduction in sample size caused by restricting the sample to those students who took all four courses in consecutive academic years.

(7) These are non-Hispanic, non-mixed-race whites.

(8) An important limitation to this study is that height and weight data are self-reported. If there is sufficient measurement error, estimates of the effect of BMI on academic achievement may be biased toward zero and, as Aschenfelter and Krueger (1994) have shown, FE estimates can exacerbate this bias. However, the results presented in this study are consistent with those of Cawley (2004), who corrected for measurement error in the National Longitudinal Survey of Youth using information on true and reported height and weight from the Third National Health and Nutrition Examination Survey (see, for example, Lee and Sepanski 1995; Bound, Brown, and Mathiowetz 2002). The consistency of findings across datasets in our studies suggests that measurement error bias alone cannot explain the results of this study.

(9) All models are weighted with robust standard errors presented in parentheses.

(10) All cross-section models include dummy variables for age and control for race, intelligence (via AHPVT Score), whether living in a single parent household, whether the mother works outside the home, household income, whether the household receives Aid to Families with Dependent Children, the presence of older siblings in the household, public or private school, age of mother at birth of child, whether the parent attended college, whether the parent moved to the neighborhood because of the local schools, whether adolescent aspires to attend college, religious attendance, parental monitoring of school and friends, parental disapproval of failure to attend college, adolescent's alcohol consumption, whether the adolescent is in a romantic relationship, whether the adolescent is sexually active, frequency of family dinners, and share of people in the census tract with a high school diploma.

(11) This finding is consistent with other studies. For example, using NLSY79 data, Oettinger (1999) finds that a 0.2 point GPA reduction is associated with a 10 percentile reduction in standing in the GPA distribution.

(12) The choice of variables to include in Z to identify the Lewbel IV model requires Coy(Z, [[epsilon].sub.1][[epsilon].sub.2]) = 0 and Cov (Z, [[epsilon].sup.j.sub.2]) [not equal to] 0. Thus, unlike standard IV models, the model does not require that any of the Zs be uncorrelated with unmeasured determinants of academic performance. While not presented here, the robustness of the Lewbel IV results to choice of Z was examined. The results of these models suggest that the choice of variables to include in Z did marginally affect results. For example, the inclusion of the OBESEMOM ([Z.sub.1]) as a variable in Z did diminish the size and precision of the parameter estimate for white girls. In an auxiliary academic performance regression that included both OBESEMOM and ([Z.sub.1] - [Z.sub.1]) [[??].sub.2] as covariates, the coefficient on ([Z.sub.1] - [Z.sub.1]) [[??].sub.2] was found to be significantly associated with GPA. Thus, it is not clear that ([Z.sub.1] - [Z.sub.1]) [[??].sub.2] is an appropriate exclusion restriction to identify the Lewbel model. Thus, a separate set of Lewbel IV models were estimated, where the choices of variables in Z were restricted to a clearly more exogenous set of Zs: regional dummies and urbanicity measures. Auxiliary academic performance regressions did not find that the parameter estimates on any of the ([Z.sub.1] - [Z.sub.1]) [[??].sub.2] variables were significant. The results of these models were generally consistent in sign and magnitude with those presented in Table 3, but the parameters were imprecisely estimated.

(13) For nonwhite males, Lewbel IV estimates are consistently larger than OLS estimates, though they are not statistically different. One explanation for this finding may be that overweight nonwhite males have greater self-perceived unobserved stability and power (Stearns 1997).

(14) Another FE model considered was a school FE model, which can be estimated with the Add Health data. While not presented here, school FE models produced results that are statistically equivalent to OLS estimates.

(15) Time-varying covariates in the models include aspirations to attend college, parental involvement in student's schoolwork, parent's labor force participation, parental setting of weekend curfew, weekly religious attendance, whether in a romantic relationship, whether had sexual intercourse, and whether changed schools between interviews.

(16) Results of these alternate specifications are available upon request of the author.

(17) However, when I compare the OLS estimate of the relationship between BMI in Wave 1 on GPA in Wave 2 and the OLS estimate of the relationship between BMI in Wave 1 on GPA in Wave 1, I do not find that the estimates are statistically different.

(18) In unreported results, I also estimate Lewbel IV models on the fixed-effects sample. In these models, Lewbel IV estimates are generally similar to OLS estimates, as in Table 3.

(19) The findings are similar for OLS models.

Joseph J. Sabia, University of Georgia, Department of Housing & Consumer Economics, Athens, GA 30602, USA; E-mail jsabia@fcs.uga.edu.

Table 1. Weighted Means and Standard Deviations of Variables (a)

Variable Name Definition White Females

MEGPA Math and English GPA 2.93 (0.899)
MATH Math GPA 2.77 (1.15)
ENGLISH English GPA 3.09 (0.992)
BMI Body Mass Index (weight in
 kilograms/height in meters
 squared) 21.56 (3.82)
POUNDS Weight in pounds 128.21 (24.99)
INCHES Height in inches 64.6 (2.79)
UNDER BMI < 5th percentile for age-sex
 category 0.017 (0.128)
ATRISK BMI between 85th and 95th percentile
 for age-sex category 0.121 (0.326)
OBESE (b) BMI greater than 95th percentile for
 age-sex category 0.049 (0.216)
PEROVER Perceive oneself to be overweight 0.375 (0.484)
OBESEMOM Parent reports biological mother has
 obesity problem 0.196 (0.397)
OBESEDAD Parent reports biological father has
 obesity problem 0.119 (0.324)
SPORT Adolescent plays school sport 0.688 (0.464)
EXERCISE Adolescent exercises regularly 0.538 (0.499)
MEAT Meat consumption for breakfast 0.042 (0.202)
ASPIRE Adolescent aspires to attend college 0.792 (0.406)
AHPVT Add Health Picture-Vocabulary Test
 Score 105.8 (11.9)
PUBLIC Adolescent attends public school 0.929 (0.256)
RURAL Adolescent's school in rural area 0.213 (0.409)
SUBURBAN Adolescent's school in suburban area 0.598 (0.490)
SOUTH Adolescent lives in the southern
 region of United States 0.308 (0.462)
WEST Adolescent lives in western region
 of United States 0.130 (0.337)
MIDWEST (c) Adolescent lives in midwestern
 region of United States 0.419 (0.494)
PARDISCOL Strong parental disapproval if
 adolescent does not attend college 0.630 (0.483)
PARTEACH Parent a member of PTA 0.395 (0.489)
NGHBRHD Parent moved to neighborhood because
 of school system 0.558 (0.497)
BRILLIANT Parent believes adolescent being
 brilliant is top priority 0.631 (0.483)
PARPROJECT Parent recently aided adolescent
 with school project 0.077 (0.267)
PARTALK Parent recently spoke with
 adolescent about grades 0.674 (0.469)
SINGLEPAR Single-parent household 0.096 (0.295)
COLGRAD Parent graduated from college 0.285 (0.451)
PARWORK Parent is employed outside the home 0.758 (0.428)
CURFEW Parent has strict weekend curfew for
 adolescent 0.253 (0.435)
DINNERWK Number of days per week adolescent
 has dinner with family 5.09 (2.28)
RELIGIONWK Family attends religious services
 at least once per week 0.404 (0.491)
RELIGIONMO Family attends religious services
 about once per month 0.179 (0.383)
RELIGIONYR (d) Family attends religious services
 about once per year 0.192 (0.394)
NOMONITOR Parent does not monitor friends of
 adolescent 0.158 (0.365)
ROMANTIC Adolescent in romantic or
 romantic-like relationship 0.612 (0.487)
INTERCOURSE Adolescent engaged in sexual
 intercourse 0.274 (0.446)
OLDERSIB Older sibling in household 0.423 (0.494)
HHINC Log of household income (in 000s) 3.82 (0.706)
DRINK Adolescent consumed alcoholic
 beverages in absence of parents
 during previous month 0.537 (0.499)
AGE 15 Adolescent aged 15 0.318 (0.466)
AGE16 Adolescent aged 16 0.258 (0.438)
AGE 17 (e) Adolescent aged 17 0.157 (0.364)
N 1472

 Nonwhite
Variable Name Definition Females

MEGPA Math and English GPA 2.72 (0.903)
MATH Math GPA 2.56 (1.15)
ENGLISH English GPA 2.89 (l.04)
BMI Body Mass Index (weight in
 kilograms/height in meters
 squared) 22.69 (4.56)
POUNDS Weight in pounds 131.26 (30.81)
INCHES Height in inches 63.7 (2.99)
UNDER BMI < 5th percentile for age-sex
 category 0.017 (0.129)
ATRISK BMI between 85th and 95th percentile
 for age-sex category 0.160 (0.367)
OBESE (b) BMI greater than 95th percentile for
 age-sex category 0.101 (0.302)
PEROVER Perceive oneself to be overweight 0.397 (0.490)
OBESEMOM Parent reports biological mother has
 obesity problem 0.172 (0.378)
OBESEDAD Parent reports biological father has
 obesity problem 0.084 (0.277)
SPORT Adolescent plays school sport 0.602 (0.490)
EXERCISE Adolescent exercises regularly 0.574 (0.495)
MEAT Meat consumption for breakfast 0.143 (0.351)
ASPIRE Adolescent aspires to attend college 0.826 (0.379)
AHPVT Add Health Picture-Vocabulary Test
 Score 97.1 (14.5)
PUBLIC Adolescent attends public school 0.911 (0.285)
RURAL Adolescent's school in rural area 0.089 (0.284)
SUBURBAN Adolescent's school in suburban area 0.465 (0.499)
SOUTH Adolescent lives in the southern
 region of United States 0.462 (0.499)
WEST Adolescent lives in western region
 of United States 0.263 (0.441)
MIDWEST (c) Adolescent lives in midwestern
 region of United States 0.152 (0.359)
PARDISCOL Strong parental disapproval if
 adolescent does not attend college 0.736 (0.441)
PARTEACH Parent a member of PTA 0.327 (0.469)
NGHBRHD Parent moved to neighborhood because
 of school system 0.443 (0.497)
BRILLIANT Parent believes adolescent being
 brilliant is top priority 0.780 (0.414)
PARPROJECT Parent recently aided adolescent
 with school project 0.068 (0.252)
PARTALK Parent recently spoke with
 adolescent about grades 0.744 (0.436)
SINGLEPAR Single-parent household 0.234 (0.423)
COLGRAD Parent graduated from college 0.217 (0.412)
PARWORK Parent is employed outside the home 0.741 (0.438)
CURFEW Parent has strict weekend curfew for
 adolescent 0.229 (0.420)
DINNERWK Number of days per week adolescent
 has dinner with family 4.41 (2.57)
RELIGIONWK Family attends religious services
 at least once per week 0.493 (0.500)
RELIGIONMO Family attends religious services
 about once per month 0.228 (0.420)
RELIGIONYR (d) Family attends religious services
 about once per year 0.157 (0.364)
NOMONITOR Parent does not monitor friends of
 adolescent 0.307 (0.462)
ROMANTIC Adolescent in romantic or
 romantic-like relationship 0.495 (0.500)
INTERCOURSE Adolescent engaged in sexual
 intercourse 0.306 (0.461)
OLDERSIB Older sibling in household 0.432 (0.496)
HHINC Log of household income (in 000s) 3.36 (0.864)
DRINK Adolescent consumed alcoholic
 beverages in absence of parents
 during previous month 0.430 (0.495)
AGE 15 Adolescent aged 15 0.340 (0.474)
AGE16 Adolescent aged 16 0.247 (0.432)
AGE 17 (e) Adolescent aged 17 0.147 (0.354)
N 1059

Variable Name Definition White Males

MEGPA Math and English GPA 2.73 (0.955)
MATH Math GPA 2.71 (1.18)
ENGLISH English GPA 2.74 (1.10)
BMI Body Mass Index (weight in
 kilograms/height in meters
 squared) 22.55 (4.50)
POUNDS Weight in pounds 154.41 (36.86)
INCHES Height in inches 69.2 (3.59)
UNDER BMI < 5th percentile for age-sex
 category 0.021 (0.142)
ATRISK BMI between 85th and 95th percentile
 for age-sex category 0.152 (0.329)
OBESE (b) BMI greater than 95th percentile for
 age-sex category 0.133 (0.340)
PEROVER Perceive oneself to be overweight 0.226 (0.419)
OBESEMOM Parent reports biological mother has
 obesity problem 0.198 (0.399)
OBESEDAD Parent reports biological father has
 obesity problem 0.126 (0.332)
SPORT Adolescent plays school sport 0.836 (0.371)
EXERCISE Adolescent exercises regularly 0.528 (0.499)
MEAT Meat consumption for breakfast 0.078 (0.269)
ASPIRE Adolescent aspires to attend college 0.727 (0.446)
AHPVT Add Health Picture-Vocabulary Test
 Score 106.9 (0.116)
PUBLIC Adolescent attends public school 0.918 (0.274)
RURAL Adolescent's school in rural area 0.201 (0.401)
SUBURBAN Adolescent's school in suburban area 0.632 (0.482)
SOUTH Adolescent lives in the southern
 region of United States 0.343 (0.475)
WEST Adolescent lives in western region
 of United States 0.128 (0.335)
MIDWEST (c) Adolescent lives in midwestern
 region of United States 0.381 (0.486)
PARDISCOL Strong parental disapproval if
 adolescent does not attend college 0.612 (0.487)
PARTEACH Parent a member of PTA 0.396 (0.489)
NGHBRHD Parent moved to neighborhood because
 of school system 0.573 (0.495)
BRILLIANT Parent believes adolescent being
 brilliant is top priority 0.655 (0.476)
PARPROJECT Parent recently aided adolescent
 with school project 0.074 (0.262)
PARTALK Parent recently spoke with
 adolescent about grades 0.616 (0.486)
SINGLEPAR Single-parent household 0.085 (0.279)
COLGRAD Parent graduated from college 0.269 (0.444)
PARWORK Parent is employed outside the home 0.765 (0.424)
CURFEW Parent has strict weekend curfew for
 adolescent 0.337 (0.473)
DINNERWK Number of days per week adolescent
 has dinner with family 5.32 (2.05)
RELIGIONWK Family attends religious services
 at least once per week 0.367 (0.473)
RELIGIONMO Family attends religious services
 about once per month 0.193 (0.395)
RELIGIONYR (d) Family attends religious services
 about once per year 0.176 (0.381)
NOMONITOR Parent does not monitor friends of
 adolescent 0.164 (0.370)
ROMANTIC Adolescent in romantic or
 romantic-like relationship 0.528 (0.499)
INTERCOURSE Adolescent engaged in sexual
 intercourse 0.278 (0.448)
OLDERSIB Older sibling in household 0.395 (0.489)
HHINC Log of household income (in 000s) 3.80 (0.730)
DRINK Adolescent consumed alcoholic
 beverages in absence of parents
 during previous month 0.501 (0.500)
AGE 15 Adolescent aged 15 0.274 (0.446)
AGE16 Adolescent aged 16 0.246 (0.431)
AGE 17 (e) Adolescent aged 17 0.192 (0.394)
N 1561

Variable Name Definition Nonwhite Males

MEGPA Math and English GPA 2.46 (0.927)
MATH Math GPA 2.39 (1.21)
ENGLISH English GPA 2.54 (1.07)
BMI Body Mass Index (weight in
 kilograms/height in meters
 squared) 22.63 (4.48)
POUNDS Weight in pounds 149.00 (33.96)
INCHES Height in inches 67.9 (3.52)
UNDER BMI < 5th percentile for age-sex
 category 0.016 (0.124)
ATRISK BMI between 85th and 95th percentile
 for age-sex category 0.156 (0.363)
OBESE (b) BMI greater than 95th percentile for
 age-sex category 0.139 (0.348)
PEROVER Perceive oneself to be overweight 0.206 (0.405)
OBESEMOM Parent reports biological mother has
 obesity problem 0.163 (0.370)
OBESEDAD Parent reports biological father has
 obesity problem 0.086 (0.281)
SPORT Adolescent plays school sport 0.863 (0.344)
EXERCISE Adolescent exercises regularly 0.547 (0.498)
MEAT Meat consumption for breakfast 0.211 (0.408)
ASPIRE Adolescent aspires to attend college 0.737 (0.441)
AHPVT Add Health Picture-Vocabulary Test
 Score 97.8 (14.3)
PUBLIC Adolescent attends public school 0.915 (0.279)
RURAL Adolescent's school in rural area 0.101 (0.301)
SUBURBAN Adolescent's school in suburban area 0.491 (0.500)
SOUTH Adolescent lives in the southern
 region of United States 0.440 (0.497)
WEST Adolescent lives in western region
 of United States 0.281 (0.450)
MIDWEST (c) Adolescent lives in midwestern
 region of United States 0.168 (0.374)
PARDISCOL Strong parental disapproval if
 adolescent does not attend college 0.738 (0.440)
PARTEACH Parent a member of PTA 0.332 (0.471)
NGHBRHD Parent moved to neighborhood because
 of school system 0.433 (0.496)
BRILLIANT Parent believes adolescent being
 brilliant is top priority 0.721 (0.449)
PARPROJECT Parent recently aided adolescent
 with school project 0.089 (0.285)
PARTALK Parent recently spoke with
 adolescent about grades 0.687 (0.464)
SINGLEPAR Single-parent household 0.235 (0.424)
COLGRAD Parent graduated from college 0.262 (0.440)
PARWORK Parent is employed outside the home 0.749 (0.434)
CURFEW Parent has strict weekend curfew for
 adolescent 0.352 (0.478)
DINNERWK Number of days per week adolescent
 has dinner with family 4.42 (2.59)
RELIGIONWK Family attends religious services
 at least once per week 0.462 (0.499)
RELIGIONMO Family attends religious services
 about once per month 0.195 (0.397)
RELIGIONYR (d) Family attends religious services
 about once per year 0.148 (0.355)
NOMONITOR Parent does not monitor friends of
 adolescent 0.278 (0.448)
ROMANTIC Adolescent in romantic or
 romantic-like relationship 0.529 (0.499)
INTERCOURSE Adolescent engaged in sexual
 intercourse 0.426 (0.495)
OLDERSIB Older sibling in household 0.431 (0.495)
HHINC Log of household income (in 000s) 3.38 (0.894)
DRINK Adolescent consumed alcoholic
 beverages in absence of parents
 during previous month 0.378 (0.485)
AGE 15 Adolescent aged 15 0.301 (0.459)
AGE16 Adolescent aged 16 0.272 (0.445)
AGE 17 (e) Adolescent aged 17 0.176 (0.381)
N 1055

(a) Sample restricted to students enrolled in English/language
arts and math courses and had nonmissing observations for all
right-hand side variables in the OLS regression analysis.

(b) Omitted category is BMI in 5th to 85th percentile of BMI
distribution.

(c) Omitted category is northeast region of country.

(d) Omitted category is never attending religious services.

(e) Omitted category is age 14.

Table 2. OLS Estimates of Relationship between Obesity and Academic
Achievement for Adolescents Aged 14-17 (a)

 Females Males

 White (1) Nonwhite (2) White (3) Nonwhite (4)

BMI -0.018 *** -0.015 ** 0.003 -0.020 ***
 (0.007) (0.007) (0.006) (0.006)
POUNDS (b) -0.003 *** -0.003 ** 0.0002 -0.003 ***
 (0.001) (0.001) (0.0008) (0.001)
UNDER (c) 0.060 0.153 -0.013 -0.362 **
 (0.165) (0.157) (0.061) (0.152)
ATRISK (c) -0.034 0.069 0.066 -0.120
 (0.066) (0.081) (0.065) (0.087)
OBESE (c) -0.182 * -0.270 *** -0.049 -0.278 ***
 (0.099) (0.103) (0.069) (0.088)
PEROVER -0.153 *** 0.006 -0.052 -0.123
 (0.055) (0.060) (0.065) (0.080)
N 1472 1059 1561 1055

Standard errors presented in parentheses. All models weighted.

(a) All models include controls for age, intelligence (AHPVT Score),
race, whether in romantic relationship, whether sexually active,
whether household receives AFDC, age of mother at adolescent's birth,
whether single parent household, whether mother works outside home,
household income, whether public school, rural/urban/suburban,
region of country, presence of older siblings, alcohol consumption,
parental strictness, religious attendance, frequency of family
dinners, meat consumption at breakfast, whether adolescent aspires
to attend college, educational attainment of parent, whether parent
moved to neighborhood because of school system, parental monitoring
of school and friends, parental disapproval if adolescent does not
attend college, whether plays a sport, and whether exercises
regularly.

(b) Controlling for height in inches.

(c) Omitted category includes adolescents with BMI in 5th to
85th percentile for their age-sex category.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 3. IV Estimates of Relationship between Obesity and Academic
Achievement for Adolescents Aged 14-17 (a)

 White Female

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

BMI -0.018 *** -0.096 *** -0.012
 (0.007) (0.030) (0.012)
 F-stat on instruments F = 29.4 F = 12.8
 Sargan
 overidentification
 p-value p = 0.48 p = 0.31
 First-stage [R.sup.2] [R.sup.2] = 0.16 [R.sup.2] = 0.35
 Breusch-Pagan p-value -- p = 0.00

POUNDS (b) -0.003 *** -0.016 *** -0.003 *
 (0.001) (0.005) (0.002)
 F-stat on instruments F = 26.1 F = 12.6
 Sargan
 overidentification
 p-value p = 0.53 p = 0.38
 First-stage [R.sup.2] [R.sup.2] = 0.31 [R.sup.2] = 0.48
 Breusch-Pagan p-value -- p = 0.00

PEROVER -0.153 *** -1.011 *** -1.14 ***
 (0.055) (0.363) (0.377)
 F-stat on instruments F = 26.1 F = 0.6
 Sargan
 overidentification
 p-value p = 0.53 p = 0.67
 First-stage [R.sup.2] [R.sup.2] = 0.09 [R.sup.2] = 0.08
 Breusch-Pagan p-value -- p = 0.00

OBESE -0.177 * -1.85 *** -0.076
 (0.099) (0.257) (0.125)
 F-stat on instruments F = 17.0 F = 58.9
 Sargan
 overidentification
 p-value p = 0.64 p = 0.46
 First-stage [R.sup.2] [R.sup.2] = 0.10 [R.sup.2] = 0.67
 Breusch-Pagan p-value -- p = 0.00

N 1472 1472 1472

 Nonwhite Female

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

BMI -0.015 ** -0.015 -0.016
 (0.007) (0.021) (0.010)
 F-stat on instruments F = 22.3 F = 10.0
 Sargan
 overidentification
 p-value p = 0.68 p = 0.04
 First-stage [R.sup.2] [R.sup.2] = 0.17 [R.sup.2] = 0.40
 Breusch-Pagan p-value p = 0.00

POUNDS (b) -0.003 *** -0.002 -0.003 *
 (0.001) (0.003) (0.002)
 F-stat on instruments F = 23.1 F = 15.3
 Sargan
 overidentification
 p-value p = 0.68 p = 0.03
 First-stage [R.sup.2] [R.sup.2] = 0.31 [R.sup.2] = 0.50
 Breusch-Pagan p-value p = 0.00

PEROVER 0.006 -0.208 -0.353
 (0.060) (0.301) (0.354)
 F-stat on instruments F = 13.7 F = 0.57
 Sargan
 overidentification
 p-value p = 0.68 p = 0.07
 First-stage [R.sup.2] [R.sup.2] = 0.09 [R.sup.2] = 0.11
 Breusch-Pagan p-value p = 0.06

OBESE -0.287 *** -0.266 -0.277 **
 (0.102) (0.370) (0.130)
 F-stat on instruments F = 14.5 F = 19.3
 Sargan
 overidentification
 p-value p = 0.71 p = 0.11
 First-stage [R.sup.2] [R.sup.2] = 0.13 [R.sup.2] = 0.51
 Breusch-Pagan p-value p = 0.00

N 1059 1059 1059

 White Male

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

BMI 0.001 -0.029 0.016
 (0.006) (0.018) (0.010)
 F-stat on instruments F = 28.0 F = 15.1
 Sargan
 overidentification
 p-value p = 0.20 p = 0.31
 First-stage [R.sup.2] [R.sup.2] = 0.20 [R.sup.2] = 0.40
 Breusch-Pagan p-value p = 0.00

POUNDS (b) 0.0002 -0.004 0.002
 (0.001) (0.003) (0.001)
 F-stat on instruments F = 27.2 F = 30.7
 Sargan
 overidentification
 p-value p = 0.20 p = 0.37
 First-stage [R.sup.2] [R.sup.2] = 0.43 [R.sup.2] = 0.58
 Breusch-Pagan p-value p = 0.00

PEROVER -0.052 -0.478 0.011
 (0.065) (0.298) (0.119)
 F-stat on instruments F = 21.2 F = 9.6
 Sargan
 overidentification
 p-value p = 0.28 p = 0.64
 First-stage [R.sup.2] [R.sup.2] = 0.12 [R.sup.2] = 0.28
 Breusch-Pagan p-value p = 0.00

OBESE -0.061 -0.437 -0.047
 (0.068) (0.288) (0.103)
 F-stat on instruments F = 25.8 F = 28.6
 Sargan
 overidentification
 p-value p = 0.22 p = 0.41
 First-stage [R.sup.2] [R.sup.2] = 0.15 [R.sup.2] = 0.50
 Breusch-Pagan p-value p = 0.00

N 1561 1561 1561

 Nonwhite Male

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

BMI -0.020 *** -0.078 *** -0.031 ***
 (0.006) (0.030) (0.010)
 F-stat on instruments F = 14.5 F = 13.8
 Sargan
 overidentification
 p-value p = 0.28 p = 0.17
 First-stage [R.sup.2] [R.sup.2] = 0.10 [R.sup.2] = 0.42
 Breusch-Pagan p-value p = 0.00

POUNDS (b) -0.003 *** -0.011 *** 0.005 ***
 (0.001) (0.004) (0.002)
 F-stat on instruments F = 15.5 F = 13.5
 Sargan
 overidentification
 p-value p = 0.25 p = 0.16
 First-stage [R.sup.2] [R.sup.2] = 0.31 [R.sup.2] = 0.55
 Breusch-Pagan p-value p = 0.00

PEROVER -0.123 -1.14 * -0.099
 (0.080) (0.604) (0.102)
 F-stat on instruments F = 5.4 F = 18.6
 Sargan
 overidentification
 p-value p = 0.10 p = 0.77
 First-stage [R.sup.2] [R.sup.2] = 0.07 [R.sup.2] = 0.58
 Breusch-Pagan p-value p = 0.00

OBESE -0.249 *** -0.966 *** -0.270 **
 (0.087) (0.423) (0.124)
 F-stat on instruments F = 10.1 F = 15.0
 Sargan
 overidentification
 p-value p = 0.14 p = 0.25
 First-stage [R.sup.2] [R.sup.2] = 0.06 [R.sup.2] = 0.41
 Breusch-Pagan p-value p = 0.00

N 1055 1055 1055

Robust standard errors presented in parentheses for OLS and IV
models. Lewbel IV models not corrected for heteroskedasticity
because the heteroskedasticity in the first state is used to
identify the models.

(a) All models include controls for age, intelligence, race,
whether in romantic relationship, whether sexually active, whether
household receives AFDC, age of mother at adolescent's birth,
whether single parent household, whether mother works outside home,
household income, whether public school, rural/urban/suburban
region, region of country, presence of older siblings, alcohol
consumption, parental strictness, religious attendance, frequency
of family dinners, share in census tract w/ high school diploma,
whether adolescent aspires to attend college, educational attainment
of parent, whether parent moved to neighborhood because of school
system, parental monitoring of school and friends, and parental
disapproval if adolescent does not attend college.

(b) Controlling for height in inches.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 4. OLS and Individual FE Estimates of Impact of Obesity on
Academic Achievement for Adolescents Aged 14-17 (a)

 White Females Nonwhite Female

 (1) OLS (2) FE (3) OLS (4) FE

BMI -0.022 *** -0.031 ** -0.020 ** 0.001
 (0.008) (0.016) (0.008) (0.018)

POUNDS (b) -0.004 ** -0.005 * -0.004 *** -0.000
 (0.001) (0.003) (0.001) (0.004)

UNDER (c) -0.193 0.245 0.413 * 0.139
 (0.249) (0.267) (0.229) (0.137)

ATRISK (c) -0.059 -0.012 0.032 -0.048
 (0.087) (0.100) (0.093) (0.088)

OBESE (c) -0.310 ** -0.253 -0.312 *** -0.120
 (0.135) (0.190) (0.121) (0.196)

PEROVER -0.154 *** -0.111 * -0.054 0.205 **
 (0.057) (0.069) (0.068) (0.089)

N 1209 1209 823 823

 White Males Nonwhite Males

 (5) OLS (6) Fixed (7) OLS (8) Fixed

BMI 0.001 0.014 -0.019 ** 0.013
 (0.007) (0.019) (0.008) (0.025)

POUNDS (b) 0.0002 0.003 -0.003 *** 0.002
 (0.001) (0.003) (0.001) (0.005)

UNDER (c) -0.203 -0.134 -0.493 *** -0.354
 (0.182) (0.154) (0.190) (0.574)

ATRISK (c) 0.072 0.008 -0.233 ** -0.112
 (0.079) (0.096) (0.096) (0.129)

OBESE (c) -0.076 -0.205 -0.243 ** -0.256
 (0.083) (0.172) (0.104) (0.287)

PEROVER -0.020 -0.049 -0.117 0.226 **
 (0.067) (0.098) (0.093) (0.104)

N 1339 1339 822 822

Standard errors presented in parentheses. All models weighted.

(a) All OLS models include controls for age, whether in romantic
relationship, whether sexually active, whether household receives
AFDC, age of mother at adolescent's birth, whether single parent
household, whether mother works outside home, household income,
whether public school, rural/urban/suburban region, region of
country, presence of older siblings, alcohol consumption, parental
strictness, religious attendance, frequency of family dinners,
share in census tract w/high school diploma, whether adolescent
aspires to attend college, educational attainment of parent,
whether parent moved to neighborhood because of school system,
and parental monitoring of school and friends. FE models include
controls for the following time-varying covariates: aspirations
to attend college, whether had sexual intercourse, whether in
romantic relationship, parental involvement in adolescent's school
work, parent's labor force participation, adolescent's alcohol
consumption, religious attendance, athletic activity, parental
setting of weekend time limits, and whether the adolescent changes
schools.

(b) Controlling for height in inches.

(c) Omitted category includes adolescents with BMI in 5th to
85th percentile for their age/sex category.

* Significant at 10% level.

** Significant at 5% level.

*** Significant at 1% level.

Table 5. Robustness of Individual FE Estimates of Relationship
between BMI and Academic Performance for White Females (a)

Additional Controls BMI POUNDS (b)

No additional controls -0.031 ** -0.005 *
 (0.015) (0.003)
Model includes difficulty getting along with -0.033 ** -0.005 **
 teachers (0.015) (0.002)
Model includes self-assessed depression -0.033 ** -0.005 **
 (0.015) (0.003)
Model includes frequent loneliness, having -0.033 ** -0.005 **
 the blues, and suicidal thoughts (0.015) (0.003)
Model includes measure of frequent difficulty -0.034 ** -0.005 **
 paying attention in class (0.015) (0.003)
Model includes self-assessment of general bad -0.032 ** -0.005 **
 health (0.015) (0.002)
Model inclues measures of whether tired a lot -0.031 ** -0.005 **
 or wake up tired frequently (0.014) (0.003)
Model includes absences from school -0.033 ** -0.005 **
 (0.014) (0.003)
Model includes all above controls -0.035 ** -0.006 **
 (0.014) (0.003)
N 1566 1566

(a) All models control for the following time-varying covariates:
aspirations to attend college, whether had sexual intercourse, whether
in romantic relationship, parental involvement in adolescent's school
work, parent's labor force participation, adolescent's alcohol
consumption, religious attendance, athletic activity, and parental
setting of weekend time limits.

(b) Weight in pounds, controlling for height in inches as a
right-hand variable.

* Significant at 10% level.

** Significant at 5% level.