In this section, we empirically investigate the relationship between BMI and hours of work. First, we use OLS to establish an association and then we use sibling fixed effects and IV regressions to explore whether the relationship is causal. We find consistent evidence indicating that BMI increases hours of work for White (both single and married) women. Furthermore, it seems that the relationship is causal. For other groups, either we do not find any relationship or the evidence is inconsistent. After that, we go through auxiliary analyses to check whether the marriage market is the driver behind the relationship between BMI and hours of work. Our analysis suggests that the marriage market is indeed the primary driver. Finally, we present evidence from a series of robustness checks to show that our results are robust to reasonable changes in sample selection criteria, estimation methods, and functional form assumptions.
Baseline analysis: single and married respondents
We start by reporting baseline results for married (Table 2) and single (Table 3) men and women. These tables report the coefficients of BMI in hours of work regressions for women (panels A, B, and C) and men (panels D, E, F) using three techniques: OLS, IV (the instrumental variable being same-sex sibling BMI), and sibling FE (fixed effects). In the case of the IV regressions, the first-stage F test for the instrument range from 16.4 (for married Hispanic men in Table 3 (E)) to 107.2 (for single White women in Table 4 (A)), suggesting that the instrument is strong by conventional standards.Footnote 9 The first two columns are for Whites, the next two for Blacks, and the last two for Hispanics. Columns 1, 3, and 5 do not include hourly wage rate as a control, while columns 2, 4, and 6 do. Our discussion focuses on the results controlling for own wage, for these results are more likely to indicate causal effects related to marriage markets. To keep the estimates comparable, we use the same samples for regressions using different methods. Hours of work are in logarithmic form.
Table 2 BMI coefficients in log hours of work regressions for married respondents: OLS, IV, and FE estimates
Table 3 BMI coefficients in log hours of work for single respondents: OLS, IV, and FE estimates
Table 4 BMI and hours work in men and women: does marital status matter?
The married
The OLS results from column 2 in Table 2 (panel A) suggest that among White women, a one-unit increase in BMI is associated with a statistically significant 0.66% increase in hours worked.Footnote 10 The estimate in column 1, not controlling for wage, is similar, suggesting that if there is an effect of BMI on hours worked, not much of it is channeled via effects of BMI on wage.
When applying the sibling FE method to our sample of married White women, we obtain results similar in size and significance to those obtained with OLS (comparing panels A and C Table 3). Furthermore, when we use the IV method (panel B) and control for wage, we continue to get results significant at the 5% level, but the estimated coefficient is larger than the OLS one: a one-unit increase in BMI leads to an almost 2% increase in White married women’s hours of work (comparing panels A and B, column 2).
Therefore, for White married women, the finding of a positive relation between body weight and hours of work is robust to the method of estimation. When own weight is instrumented with sibling’s weight, we get larger effects. One possible explanation for this pattern may be the following: suppose women with higher BMI value leisure more (i.e., work less). Since we do not observe preference for leisure, this is going to create a negative correlation between hours worked and BMI, thereby creating a downward bias in the OLS estimate. In IV regressions, we are using the sibling’s BMI to predict the BMI of a woman, and therefore, IV estimates do not suffer from this omitted-variable induced endogeneity bias.
Columns 3 to 6 in Table 3 (panel A) report OLS estimates for Black and Hispanic married women. Samples here are smaller, and the only significant coefficient is that for Black women (when we use IV and control for wage). Panels D–F of Table 3 report results for married men. Here, none of the BMI coefficients are statistically significant, regardless of whether men are White, Black, or Hispanic and regardless of method of estimation.
In sum, we find that White married women with higher BMI work more hours in the labor market using all three methods of estimation. We do not find such effects of BMI for minority married women or for married men. The finding for White married women is consistent with the framework presented in Section 2: thinner women get a higher price in marriage and thus obtain more access to consumption or more in-marriage financial transfers. The ethnic differential is consistent with the possibility that relative to White women, minority women are less likely to obtain in-marriage financial transfers or access to consumption.Footnote 11 The gender differential may be a function of the higher involvement of married women in household production and their lower earnings relative to that of married men and to the ensuing higher likelihood that they are on the receiving side of intra-marriage financial transfers.
The singles
Table 3 reports results for single women (panels A–C) and for single men (panels D–F). The OLS results from column 2 (panel A) suggest that among White single women, a one-unit increase in BMI is associated with a statistically significant 0.58% increase in hours worked (0.45% if hourly wage is not added as a control variable; column 1, Table 3, panel A). Comparing panel A with panel B indicates that a one-unit increase in BMI has a larger effect when the IV method is used: a one-unit increase in BMI leads to a 1.38% (column 2, panel B) increase in hours worked. The estimate is significant at a 5% level of significance. Thus, as in the case of White married women, both OLS and IV estimates suggest that heavier White single women work more hours. However, we do not find any significant results when the sibling FE method is used. One potential reason may be that the correlation between own BMI and sibling BMI is stronger (as suggested by larger first-stage F-stats for single women), and therefore, differencing may magnify the measurement error (Deaton 1995), rendering the estimates insignificant. OLS results also suggest that Black single women with higher BMI work more hours. This result is also robust to inclusion of wage.
OLS results in Table 3 (panels D and E) suggest that BMI is also positively associated with hours of work of White and Black single men. The result for White single men carries over when the IV method is applied (columns 1 and 2), but that is not the case for Black single men (columns 3 and 4). As for Hispanics, we find no association between body weight and hours of work for either single men or single women.
Comparing results with and without control for wage
Our results suggest that controlling for wage does not significantly change the association between BMI and hours worked, suggesting that the lower wage of high-BMI White women cannot explain the effect of BMI on hours of work. In addition, if anything, column 2 estimates (when we control for wage) are always bigger than column 1 estimates (without wage as a control). This is consistent with our discussion that BMI reduces wage in White women (Cawley 2004 among others) and women’s labor supply elasticity is positive (Blundell and MaCurdy 1999 among others).
Comparing results for married and single respondents
In the case of White women, the coefficients of BMI are similar for single women and married women. This holds using both the OLS and IV methods. For instance, using the IV method and controlling for wage, a one-unit increase in BMI leads to a 1.38% increase in hours worked by single women and a 2% increase in hours worked by married women. Both findings are significant at the 5% level of significance. This suggests that single women anticipate possible future benefits from being married, with heavier single women anticipating lower financial in-marriage transfers. To test whether the effect of BMI on hours of work is different in single and married respondents, we pooled the single and married samples, separately for each ethnic group, and included married status and an interaction between marital status and BMI in OLS regressions. We use OLS (and therefore ignore the endogeneity of BMI as well as marriage) as the purpose of these exercises is to check whether the level of association between BMI and hours of work varies by marital status. Table 4 reports the results for women (panel A) and men (panel B). We only report the coefficients of BMI, marital status, and the interaction between BMI and marital status. We list all other control variables below the table. Table 4 (panel A) shows that high-BMI women (White or Black) work more in the combined sample. Being married does not make a significant difference in women’s hours worked, suggesting that marital status itself does not change hours of work. Most importantly, the coefficient of the interaction term is never significant, suggesting that the level of association between BMI and hours of work does not vary with marital status.
Results for men (panel B) are somewhat different from the results for women. Tables 2 and 3 suggest this as well. Married men work more, and the coefficient of the interaction between BMI and marital status is significant for White men when we do not control for wage (column 1). However, once we control for wage (column 2), it is not significant anymore. One possible explanation is that high-BMI single men need more resources to attract potential mates, but once they are married, they do not face the need for the extra resources. This is consistent with the findings in Mukhopadhyay (2008).
It thus appears that in the case of Black and White men, as well as in the case of Black women, there is positive association between BMI and hours of work at the single stage, but not at the married stage. The only case where a positive association between BMI and hours of work is also observed at the married stage (and possibly larger for the marrieds than for the singles) is the case of White women. White women may be the exception for they may be more likely to receive in-marriage transfers from husbands compared to their Black or Hispanic counterparts. Therefore, there may be more of an in-marriage financial reward (or a reward in the form of extra access to consumption) for being thin in the case of married White women than in the case of married minority women. The result that the association between BMI and hours of work disappears for Black married women at the married state is consistent with Black women often being the principal earner in their couple (see Mincy et al. 2005). This may also reflect that married women’s access to their husband’s financial resources may vary across races (see Grossbard 2005).
In the rest of the paper, we follow multiple strategies to disentangle whether the relation between BMI and hours of work is driven by factors related to marriage markets or by labor markets. However, from now on we focus on women (especially White women), since the results so far suggest a relationship between BMI and hours of work mostly for White women.
Are marriage markets driving the relation between BMI and hours of work in women?
Married women
In the conceptual framework section (Section 2), we discussed that marriage markets can affect the relationship between BMI and hours of work through two potential channels. First, high BMI may mean lower spousal income. Second, conditional on spousal income, high BMI may mean lower bargaining power within marriage, i.e., less access to spousal income.Footnote 12 Therefore, our next step is to use data on married women and to estimate the association between women’s BMI and their spouse’s annual income (in $10,000). We also estimate whether the association between BMI and hours of work in married women is sensitive to inclusion (or lack thereof) of spousal characteristics (age, education, and annual income) in OLS regressions of hours of work.
Table 5 shows the results. Panel A presents the association between BMI and spouse’s annual income, and panel B presents the association between BMI and hours of work, with and without spousal characteristics. All regressions include all women’s characteristics including her wage. Columns 1, 3, and 5 (of panel A) only include women’s characteristics, whereas columns 2, 4, and 6 (of panel A) include the husband’s age and education as additional controls. Results in panel A show that for married White women, the coefficient of BMI is −0.112 (column 1), or in other words, a one-unit increase in BMI is associated with $1120 reduction in annual spousal income. Controlling for spouse’s age and education has little effect on the estimate. The coefficient of BMI is not significant for married Black women regardless of inclusion of age and education, and it is not significant in married Hispanic women when we include the husband’s age and education. This suggests that there is more of a material payoff for being thin in the case of White married women than in the case of Black or Hispanic married women.
Table 5 BMI, spousal annual income, and log hours of work for married women
Columns 2, 4, and 6 (of panel B) reproduce the OLS results reported in the corresponding columns of Table 2 (panel A). The following characteristics of the husband had been included: age, education, and annual income. Columns 1, 3, and 5 present results for the same regressions excluding the characteristics of the husband. Results show that the coefficients of BMI are unaffected by the addition of spousal characteristics. This suggests that high-BMI White women work more hours not because they are married to men earning less, but that they have less bargaining power in marriage and can access less of their spouse’s income. This is consistent with results reported by OQD.
Single women
Our results so far suggest that the association between BMI and hours of work in single women stems from how BMI affects expectations about marriage. For single women, we cannot pursue the direct strategy employed in Table 5 above. Instead, we investigate the rational expectations argument presented above by performing two more tests: (1) we use a unique question about marriage expectations and (2) we investigate how the association between BMI and single women’s hours of work changes when we control for predicted cumulative spousal income.
First, we use a question that was asked to the single NLSY97 respondents in some of the waves (2000, 2001, 2009, 2010, and 2011). The respondents were asked “Now think about five years from now, you will be [{AGE IN 5YRS}]. What is the percent chance that you will be married?” This question was not asked in the NLSY79, and therefore, this analysis can only be performed on NLSY97 women. As we discussed earlier, our hypothesis is that single women with high BMI expect smaller future income transfer from husbands, and therefore, they work more in the labor market even when they are single. It follows that the effect of BMI will be stronger in single women with higher self-assessed probability of getting married in the next 5 years, compared to their counterparts who have a lower self-assessed probability of getting married in the next 5 years. In a different but relevant context, Kureishi and Wakabayashi (2013) showed that marriage expectations could affect single women’s savings decisions: they found that single women who expect to get married save significantly less than single women who do not expect to get married in the next 3 years.
Accordingly, we re-estimate the OLS regressions for single women, controlling for probability of marriage. We add an interaction term between the self-assessed probability of marriage and BMI. This regression has the same set of control variables (including log wage) as in Table 3. However, sample sizes are smaller given that we only used the NLSY97 and that only selected waves of NLSY97 included the question about marriage expectations.
Figure 1 shows the marginal effect of BMI for single women and how it varies with the self-assessed probability of marriage, along with the 95% confidence interval. Panel A suggests that for single White women, the marginal effect of BMI on hours of work is insignificant when the expected probability of marriage is zero. The association between BMI and hours of work remains statistically insignificant as long the chance of getting married remains below 40%. However, the marginal effect is positive and significant for single White women when the chance of getting married in the next 5 years is at 50% or above. Panel B (panel C) shows the marginal effect of BMI for Black (Hispanic) single women. Results here are never statistically significant.
Next, we investigate how including predicted cumulative spousal income affects the association between BMI and single women’s hours of work. To do this, first, we compute the cumulative spousal income of all NLSY79 women at the time of the last survey, when all women were 50 or older. Cumulative spousal income depends on the annual income of the spouse and number of years the woman was married. If a woman was never married by the time of the last interview, then her cumulative spousal income is zero. We regress cumulative spousal income on the characteristics of women including their BMIs. Results (not shown) suggest that for White women, a one-unit increase in BMI is associated with $11,610 less in cumulative spousal income,Footnote 13 but we do not find any association for Black women. We then use the estimated coefficients to obtain the predicted cumulative spousal income of single women. Then we run regressions for single women in the entire NLSY data sets, with and without including this predicted cumulative spousal income.
Table 6 presents OLS regressions for single women. All regressions include all the controls used in Table 3, including their own wage. Thus, columns 1, 3, and 5 in this table reproduce the results presented in columns 2, 4, and 6 of Table 3 (panel A). In the regression results reported in columns 2, 4, and 6 of Table 6, we have added predicted cumulative spousal income as an additional control. Results show that for White single women, the coefficient of BMI goes down from 0.00581 to 0.00373 when we add predicted cumulative spousal income as an additional control. This suggests that one of the mechanisms by which BMI influences single White women’s hours of work is via BMI’s effect on predicted spousal income. However, this is not the case for Black single women: here, the effect of BMI does not seem to act via its effect on predicted spousal income: the coefficients of BMI in columns 3 and 4 of Table 6 are very similar. In the case of Hispanic women, we find no association in both columns 5 and 6.
Table 6 BMI and log hours of work for single women: with and without predicted cumulative spousal income
Adding up the results reported in Fig. 1 and Table 6, it appears that one reason why higher-BMI White single women work more in the labor market relative to their lower-BMI counterparts is that they expect lower future in-marriage income transfers. In turn, these expected future in-marriage transfers are lower because either heavier women are likely to get married to men with fewer resources or heavier women have less access to spousal resourcesFootnote 14 (i.e., a lower bargaining power). It seems that in the case of White women, both of these channels operate.
Can labor market mechanisms explain the relation between BMI and hours of work in single women?
In our Section 2, we discussed two additional labor market-based mechanisms that can explain the positive association between BMI and hours of work in single women. First, is it possible that high-BMI women are less healthy, and unhealthy women expect to spend fewer years in the labor force at older ages and therefore they work more hours at an early age to compensate for a shorter expected working span. Since the single sample is substantially younger than the married sample, we are most likely to observe this in the single sample. A second possibility is that high-BMI singles are more likely to work full time in order to qualify for health insurance benefits.Footnote 15
Next, we check whether these mechanisms can explain the relation between BMI and hours of work in women. First, we exclude individuals who report fair or poor health or have any kind of work limitations from our samples of single women (about 4% of the sample). Then we estimate the same OLS regressions to check whether the association between BMI and hours of work differs from the results we reported in Table 3. Results presented in Table 7 suggest that when we exclude women with health problems, our results still hold. Thus, it is unlikely that poor health explains the association between hours of work and BMI.
Table 7 BMI and log hours of work for single women: excluding women with health limitations
Next, we test whether our results for single women are driven by considerations related to access to health insurance. We estimate the association between BMI and hours of work separately for two groups: those with employers who offer health insurance as a benefit and those with employers who do not offer health insurance.Footnote 16 Table 8 reports coefficients of BMI in OLS regressions of hours of work for single women. All regressions include the standard set of controls including wage. Table 8 shows that the positive association of BMI and hours of work holds in both groups of single White women and single Black women. The smaller coefficient of BMI for White women with employer-provided insurance can be explained by the fact that most of these women already work full time (on an average 2008 vs. 1235 h/year for women do work for employers that do not offer health insurance), often a requirement to eligibility for such insurance. Therefore, there is not much room for additional hours of work. The group of women who do not have employer-provided health insurance is illuminating. These women’s hours of work are unlikely to be motivated by insurance (since the employer do not offer health insurance in the first place), but the positive BMI per hours of work coefficient is nevertheless found for them too.
Table 8 BMI and log hours of work for single women: by availability of employer-provided insurance
Robustness checks
We perform a number of robustness checks. First, all results reported so far were estimated for samples excluding respondents without siblings. Therefore, we re-estimate OLS regressions in Tables 2 and 3 for full samples, including those without any siblings. The results (presented in Table 10 in the Appendix) are qualitatively similar to those reported in panel A of Tables 2 and 3.
Second, we experiment with alternative assumptions about extreme values of BMI. As we discussed in Section 3, Winsorizing BMI (at 1%) involves replacing all BMI below 17.2 by 17.2 and replacing all BMI above 45.5 by 45.5. Here, we follow two alternative strategies: first, we ignore all BMIs that are below 17.2 or above 45.5, and second, we ignore all BMIs that are below 18.5 or above 40.0 (a strategy followed by OQD). Table 11 in the Appendix shows results using our three methods of estimation (OLS, IV, and sibling FE) when using these two alternative strategies, along with our baseline results. The results show that our results are robust to these alternative approaches.
Third, we check whether the association between BMI and hours of work varies across the hours of work distribution by using the UQR method described in Section 3. To the extent unobserved characteristics are responsible for different women choosing different hours of work, UQR can be a check on whether the OLS results are driven by this type of unobservable factors. An added benefit of using UQR is that it can help us infer whether health insurance motive is driving the association between BMI and hours of work. If the health insurance motive is the driving factor behind this association, then we are likely to see the effect in people who are right below the cutoff of health insurance eligibility (which is about 30 h/week or about 1500 h/year). The 1500 h/year is between the 40th and 50th percentile, depending on the race/marital status group. In contrast, if the marriage market is the driving force behind the association between BMI and hours of work, then we are likely to see an effect at all levels of hours of work. Figure 2 shows the coefficient of BMI and how it varies across the unconditional log (BMI) distribution for White women. The left panel is for single women, and the right panel is for married women. Figure 2 shows that the effect of BMI is positive and significant over most of the distribution for both single and married White women. The coefficient of BMI seems to be highest well before the 40th percentile for both single and married White women. This suggests that the health insurance motive is unlikely to explain the pattern represented in Fig. 2. Figures 3 and 4 present the corresponding results for Black and Hispanic women. Consistent with our OLS results, we find that the coefficient of BMI is positive and significant for single Black women over most of the distribution, while it is never significant for married Black women or for Hispanic women.
Fourth, in our analyses so far, we assumed that the relation between BMI and hours worked is linear. This assumption may not be accurate, and if the true relationship is non-linear, this may introduce bias in our estimates. To address this issue, we estimate a partially linear model where BMI enters the “hours worked” equation non-parametrically. We use Robinson’s (1988) double residual estimator.Footnote 17 Figure 5 presents the results from semiparametric regressions in which we treat BMI as an exogenous variable. The left panel shows the results for single women and the right panel for married women. To show the difference in the semiparametric fit across races, we do not include the 95% confidence interval in this figure. The left panel in Fig. 5 indicates that the relationship is broadly linear (and monotonic) until a BMI of 40 in the case of both White and Black single women. Beyond that, an increase in BMI is associated with a decline in hours worked, most likely because the health effects dominate other considerations. The semiparametric estimates for married White women (right panel of Fig. 5) suggest more evidence of non-linearity. Even in this case, the estimated increase in hours of work when BMI doubles from 20 to 40 is not statistically different from the OLS estimate.
Fifth, following a common practice in this literature, we created dummies for weight categories, to replace the linear BMI variable. We followed the literature and created four weight categories: underweight (BMI < 18.5), healthy weight (18.5 ≤ BMI < 25), overweight (25 ≤ BMI < 30), and obese (BMI ≥ 30). Then we estimated OLS regressions with weight categories as our independent variables of interest, with healthy weight as the omitted category and controlling for wage. The results from these regressions are presented in Table 12 in the Appendix. Results suggest White and Black women who are either overweight or obese work more hours than their healthy weight counterparts. This holds for single and married women. As for men, the overweight single men work more hours, but the obese do not.Footnote 18
Finally, so far we have focused on hours of work as our outcome variable and ignored the extensive margin or labor supply. It is also conceivable that BMI affects the labor-force participation decision. The estimates reported in Table 13 in the Appendix are from OLS and IV regressions. We tried to estimate the corresponding probit and IV-probit models, but the IV-probit likelihood does not converge in two out of six cases (single Black women and married Hispanic women). Since Angrist and Pischke (2009) argue that “IV methods capture local average treatment effects regardless of whether the dependent variable is binary, non-negative, or continuously distributed on the real line,” we report the estimates from linear IV models. Results shown in Table 13 in the Appendix suggest that BMI increases probability of employment in single White women (using OLS) and in married White women (using IV). We do not find any consistent pattern of relationship between BMI and employment probability.