Journal of Labor Research

, Volume 33, Issue 2, pp 225–250

The Effect of Exercise on Earnings: Evidence from the NLSY

Authors

Article

DOI: 10.1007/s12122-011-9129-2

Cite this article as:
Kosteas, V.D. J Labor Res (2012) 33: 225. doi:10.1007/s12122-011-9129-2

Abstract

This paper investigates whether engaging in regular exercise leads to higher earnings in the labor market. While there has been a recent surge of interest by economists on the issue of obesity, relatively little attention has been given to the economic effects of regular physical activity apart from its impact on body composition. I find that engaging in regular exercise yields a 6 to 10% wage increase. The results also show that while even moderate exercise yields a positive earnings effect, frequent exercise generates an even larger impact. These findings are fairly robust to a variety of estimation techniques, including propensity score matching.

Keywords

ExerciseEarningsObesityPropensity score matching

JEL Codes

I19J21

Introduction

It is widely acknowledged that regular exercise has a beneficial impact on well-being. In addition to the positive impacts on heart health, weight and a variety of other medical issues, studies in the psychology and biology/medicine literatures show that exercise leads to improved mental function (Etnier et al. 1997; Tomporowski 2003; Hillman et al. 2008), psychological condition (Folkins and Sime 1981; Spence et al. 2005) and higher energy levels (Puetz 2006). Exercise has also been found to have an indirect effect on job satisfaction by directly impacting enthusiasm at work (Thogerson-Ntoumani 2005). All three of these traits can translate into higher earnings by raising productivity. As outlined in Lechner (2009), in addition to the direct effect, exercise can have indirect positive impacts on labor market outcomes either by serving as a signal to potential employers that the individual is dedicated and disciplined or through social networking effects. The findings in these studies suggest that individuals who regularly engage in physical activity may have a lower probability of being unemployed and higher wages relative to non-exercisers. While the existing literature suggests a potential connection between regular physical activity and labor market outcomes, identification of actual links between the two is not well established. Determining whether exercise can lead to higher wages or better occupational choices will contribute to a more complete understanding of the full benefits of exercise and healthy lifestyle choices.

As evidenced by the President’s Council on Fitness, Sports and Nutrition (www.fitness.gov) and myriad community level programs, government officials have made attempts to raise the level of physical activity for the average American. To justify the resources spent on these campaigns and activities, government officials tout the many benefits of exercise, including the potential to reduce health care expenditures. In spite of these efforts, nearly 30% of American adults remain physically inactive (Ruhm 2005). Given the many health benefits associated with exercise and the easy availability of this knowledge, it is clear that many individuals view engaging in exercise as providing significant disutility. If the highly touted health benefits are not enough to overcome the immediate disutility associated with exercise, perhaps evidence of pecuniary benefits combined with the established health benefits can encourage more individuals to participate in regular physical activity. The identification of a link between exercise and earnings also has implications for estimating the returns to employee wellness programs which have become popular with many firms by supporting the hypothesis that exercise leads to higher productivity.

The present paper contributes to the literature by employing propensity score matching (PSM) to US data in order to estimate the treatment effect of exercise on weekly earnings. To provide a baseline for comparison, each model is fitted via ordinary least squares and fixed effects estimation. In these models, only frequent exercise (at least three times per week) is consistently associated with higher wages. The PSM estimates show that frequent exercise is associated with a nearly 7% increase in wages for men and an even larger increase for women. Furthermore, men might begin to accrue pecuniary benefits by exercising at least once per week while it appears women’s wages only show a positive wage association for frequent exercise (three plus times per week).

Literature Review

Concern over the nation’s increasing obesity rate and stubbornly low rates of regular exercise has resulted in significant research effort into the economic causes and consequences of obesity and healthy lifestyle choices. One strand of literature investigates the own-wage and spillover effects of obesity. Consistent with widely held beliefs, Chou et al. (2004) find that obesity and BMI are negatively correlated with income in the United States. Other studies have investigated the relationship between obesity and wages, treating income as the dependent variable. Using fixed effects estimation and sibling differences to control for unobserved fixed effects, Baum and Ford (2004) find a wage penalty associated with obesity. Using both fixed effects and instrumental variables estimation, Cawley (2004) finds a negative correlation between obesity and wages for white women, but not for other demographic groups. Finally, Averett and Korenman (1996) find that obese women have significantly lower family incomes. While the authors find some evidence of labor market discrimination, the majority of the earnings difference is due to marriage probabilities and spousal earnings. Using English data Morris (2006) shows differing impacts of body composition on occupational attainment for men and women, with a negative impact for the latter group. Han et al. (2009a, b) show that the wage penalty associated with obesity varies with age, increasing with age after workers reach their mid-twenties. The authors also find that the wage penalty is larger in occupations that require interpersonal skills. Han et al. (2009a) find that teenage obesity can indirectly affect wages through educational attainment for both genders in addition to a direct effect for women. Research also has shown that the well documented wage penalty associated with obesity may in fact be due to employers’ passing on the higher costs of health care in the form of lower wages (Bhattacharya and Bundorf 2009). Gregory and Ruhm (2011) also take a deeper look at the relationship between wages and obesity finding a non-linear relationship with wages peaking well below the cutoff for being overweight. Mocan and Tekin (2011) find evidence that in addition to a direct effect on wages, obesity can impact wages indirectly through a negative impact on self-esteem.

In addition to the studies on obesity, there is a growing literature examining the effects of non-traditional measures of human capital on labor market outcomes. One branch of this research has looked into the labor market effects of health related choices such as drinking alcohol (Auld 2005; Hamilton and Hamilton 1997; Jones and Richmond 2006; MacDonald and Shields 2001; Renna 2008; van Ours 2004; Ziebarth and Grabka 2009) and smoking (Auld 2005; van Ours 2004). While the studies on alcohol consumption yield mixed results, several show that moderate amounts of use are associated with higher earnings (Auld 2005; Hamilton and Hamilton 1997; MacDonald and Shields 2001), a result that may stem from social networking effects akin to those posited by Lechner (2009). Jones and Richmond (2006) apply propensity score matching techniques to the National Longitudinal Surveys of Youth (NLSY79) dataset to show a negative effect of alcoholism on earnings while Balsa and French (2009) use similar techniques to evaluate whether there is a causal relationship between alcohol consumption and other labor market outcomes such as absenteeism in Uruguay.

There is also a related literature investigating the earnings effect associated with beauty. In an early paper, Hamermesh and Biddle (1994) find that plain looking people earn less than average looking people, with a penalty greater than the wage gain for good looking people. These results hold true for both men and women and persist in spite of sorting by better looking people into occupations where looks are more important. Johnston (2010) shows that blondes earn more than non-blondes and have spouses with higher average earnings. In addition to an earnings penalty, Mocan and Tekin (2010) find less attractive individuals are more likely to commit crimes, possibly as a result of their lowered economic prospects. Biddle and Hamermesh (1998) find that more attractive lawyers earn more and are less likely to enter the public sector. While these studies examine U.S. labor markets, Harper (2000) finds a significant beauty premium for British workers as well. Using experimental data on Argentinean university students, Mobius and Rosenblat (2006) also find a beauty premium and further show that this premium is due to higher confidence levels, the perception that more attractive employees are also more able and possess superior oral skills. Examining data on Chinese females, Hamermesh et al. (2002) find that expenditures on cosmetics and clothing lead to higher beauty ratings, which in turn leads to higher wages. Thus, the beauty premium uncovered in these papers is independent of the other channels through which body composition might affect labor market outcomes. Finally, studies also provide evidence of a link between height and earnings (Persico et al. 2004 and Case and Paxson 2008). A recent paper by Cipriani and Zago (2011) finds the wage premium associated with beauty may in fact reflect productivity differences rather than pure discrimination. These studies are relevant to the obesity and earnings literature since very overweight individuals are generally considered less attractive in our society. The studies focusing on the effects of facial beauty generally control for weight or BMI so that their findings are independent of BMI. Still, exercise can have an indirect effect on earnings through the beauty-wage connection by affecting perceived beauty. Two individuals with similar BMIs may have very different levels of physical fitness and attractiveness (facial beauty aside).

Another growing body of work in the medicine and business literatures examines the returns to company sponsored wellness programs. Through these programs, employers take steps to encourage healthy lifestyle choices and may include subsidized or free health club memberships. A primary motivation for providing these benefits is the potential to bring down health insurance costs. Studies attempting to estimate the return on investment (ROI) for these programs have focused on this category of savings. For example, Chung et al. (2009) estimate DaimlerChrysler Canada’s ‘Tune up Your Heart’ program reports potential savings through the reduction in several health risk factors observed when workers adhered to the program. By moving workers into lower risk categories, the program has the potential to lower the cost of life insurance, disability and prescription drug benefit expenditures. Berry et al. (2010) conducted field visits to ten employers to analyze the potential benefits of their wellness programs. In one case, the company reported a return on investment ratio of six to one based on the reduction on health care related expenditures. The study also highlights the importance of potential productivity effects associated with wellness programs, but does not provide any hard evidence on this topic. Evidence of these direct effects resulting from exercise would present another source of returns to these company sponsored programs.

Despite all of the recent attention paid to the importance of exercise and healthy lifestyle choices, relatively little attention has been given to the earnings effects of regular physical activity apart from its impact on body composition. Ruhm (2005 and 2000) investigates the effect of business cycle fluctuations on exercise, finding that exercise increases and obesity decreases during economic downturns. In particular, the increase in activity appears to stem from a decline in work hours, not from the decline in incomes.1 While these studies show that economic factors affect exercise, they do not indicate whether the reverse is true. The only other published study (that the author is aware of) to investigate the direct labor market effects of regular physical activity finds a positive earnings effect when applying semi-parametric techniques to German data (Lechner 2009). Lechner finds that participating in sports at least monthly has a positive effect on both monthly earnings and hourly wages for men and women, but does not have any significant impact on employment. The current paper differs from Lechner’s in a couple significant ways. First, the analysis uses US, not German data. Second, the current paper presents results from multiple estimators, not just a matching estimator. Third, the present paper considers the possibility that different levels of physical activity might have differing impacts on earnings. Lechner’s study only looks at the effects of participating in sports at least monthly.

Methodology

This paper tests the hypothesis that regular exercise may yield labor market benefits in the form of higher earnings. Simple linear regression techniques may not be adequate to separate causality from simple correlation. There are three potential explanations for the correlation between exercise and earnings: 1) exercise leads to higher wages; 2) wage changes affect individuals’ level of physical activity; 3) other, unobserved factors cause differences in both exercise frequency and earnings.

Exercise can lead to higher wages by raising a worker’s productivity. Regular physical activity has been linked to improved mental function, psychological wellbeing and energy levels, all of which can result in increased productivity and translate into higher earnings. The issue at hand is how to disentangle this explanation from the other two.

Explanations in the second category suggest the causal link runs in the opposite direction. Wage changes can affect exercise frequency through the labor supply decision. As is well known, if the substitution effect dominates the income effect, a wage increase will induce individuals to work more hours, taking less leisure time and leaving less time for exercise. If the income effect dominates, the opposite will occur. Thus, higher wages can lead to either more or less exercise. Given the relatively small estimates for labor supply elasticity (French 2004), I expect this mechanism to be secondary in magnitude, at best. Additionally, the empirical model controls for hours of work per week, which should capture most of this effect.2 Finally, I exploit the (short) panel structure of the data and estimate the impact of past and future exercise on current earnings as a test of reverse causality.

Explanations in the third category all point to an endogeneity problem when attempting to identify a causal relationship between exercise and labor market outcomes. Unobserved factors that may affect both commitment to exercise and earnings include, but are not limited to differences in discount rates and discipline. People with lower discount rates and greater discipline are less likely to put off exercising on a regular basis; they are also more likely to undertake investments in their human capital and work more diligently, leading to higher potential earnings. The empirical models include a proxy for the discount rate while an extended model includes a measure of post-schooling training as an attempt to account for this effect. However, there may be other sources of omitted variables bias.

While omitted variables are likely to cause least squares estimates to be biased upwards, there is the potential for an attenuation bias due to measurement error. In this case, both intentional and unintentional misreporting of exercise frequency creates measurement error in the exercise variable. In part, this problem will arise since the National Longitudinal Surveys of Youth (NLSY) did not ask about exercise intensity or duration. Consider two individuals who enjoy riding bicycles. The first individual may ride his bicycle at a leisurely pace for 15 min a day 3 days per week while the second goes out for thirty plus mile rides at a twenty mile-per-hour pace for 3 days a week. Both might report exercising three times per week. While this report is accurate for the second individual, the first person’s level of activity does not meet either the duration or intensity criteria for a session of ‘vigorous’ exercise. The use of indicator variables to measure exercise might eliminate some of the measurement error, but cannot eliminate all of it.

In the absence of adequate controls for discipline and discount rates, fixed effects estimation may be employed to eliminate the endogeneity bias. This approach assumes that these unobservable characteristics are time invariant, which over a short period of time is a plausible assumption for unobservable characteristics such as time preference and discipline. However, other unobservable factors such as a change in motivation to exercise (such as a health scare) may also lead to changes in other behaviors that also affect earnings. Even if we accept the time invariance of discipline and discount rates and have eliminated reverse causality as an explanation, fixed effects estimates may exacerbate the attenuation that results from measurement error in the key explanatory variable when applied to very short panels.3 Thus, it is possible for fixed effects to underestimate the causal effect of exercise on earnings. Furthermore, fixed effects estimation relies on within-unit (in the case the individual) changes in the key variables to identify the model coefficients. This can eliminate a large percentage of the variation in the dependant variable and many of the explanatory variables. In the sample employed in this study, 60% of the individuals with two usable observations report a change in exercise frequency. However, the indicator variable for frequent exercise changes value for only 20% of these respondents. Thus, identification of the coefficient on the exercise variables relies on information for only a subset of the sample group (particularly when looking at frequent exercise).

In light of these econometric issues, I employ propensity score matching routines to estimate the average treatment effect of participation in regular exercise. In order to estimate the average treatment effect, propensity score matching takes place in three stages. In the first stage, a probit (or logit) model is estimated for the probability of belonging to the treatment group and the propensity score estimated.4 Next, observations from the treatment group (those who exercise frequently) are matched to those not in the treatment group (those who do not exercise frequently) based on their propensity scores and the sample is tested to see if the individuals in the treated group share the same characteristics (as measured by the covariates) as their matches from the untreated group. This is referred to as the balancing criterion. In the stratification matching procedure employed in this paper, the observations are placed into blocks according to the value of the propensity score obtained in the first stage. Then, the samples are checked to see if the treated and untreated groups differ in their underlying characteristics by testing for differences in the mean values of these characteristics between the two groups within each block. If the balancing requirement is satisfied we move on to step three. In the final step, observations undergoing treatment are matched with observations that did not undergo the treatment and the effect of treatment is obtained by comparing the mean difference of the dependent variable across the two groups.

While PSM estimation provides an alternative to regression analysis, there are potential drawbacks to this approach. In order to produce unbiased estimates of the treatment effect, PSM requires large sample sizes (not a problem in the present study), substantial overlap between the treatment and comparison groups (Figs. 1 and 2 show this is the case for the present samples) and a rich set of covariates to estimate the propensity score. PSM rests on the assumption that assignment to treatment and control groups is random after conditioning on observable characteristics. Omitting variables which affect both assignment to the treatment (engagement in regular exercise) and the outcome variable (weekly earnings) from the first stage can lead to biased estimates (Heckman et al. 1997). Thus, estimates obtained via PSM may eliminate some, but not all of the bias present when estimating treatment effects via least squares estimation. In particular, the two unobservable characteristics of greatest concern in the present study are discipline and discount rates. The dataset does provide a variable which can proxy for the discount rate (albeit imperfectly) but does not have any direct measures for discipline. This leads to the question of how discipline affects earnings. Discipline can have an indirect effect on earnings through its effect on education, investment in post-schooling training, obesity (through diet and exercise) and a more direct effect manifested through effort at work. We can control for the indirect effects, but do not observe individuals’ effort on the job. Thus, our best bet would be to include proxies for an individual’s discipline. Lack of an adequate proxy means that PSM may not eliminate all of the bias found when using least squares estimation. One potential proxy is post-schooling education, since time spent in training is a signal of the individual’s commitment to her career. As a robustness check, results are presented controlling for cumulative time spent in training between 1985 and 2006.
https://static-content.springer.com/image/art%3A10.1007%2Fs12122-011-9129-2/MediaObjects/12122_2011_9129_Fig1_HTML.gif
Fig. 1

P-score distribution for male sample. Shows the propensity score distribution for the treatment group (those who exercise frequently) and the reference group (those who do not exercise frequently) for the male sample

https://static-content.springer.com/image/art%3A10.1007%2Fs12122-011-9129-2/MediaObjects/12122_2011_9129_Fig2_HTML.gif
Fig. 2

P-score distribution for female sample. Shows the propensity score distribution for the treatment group (those who exercise frequently) and the reference group (those who do not exercise frequently) for the female sample

The possibility of not eliminating all sources of bias may lead some to take a ‘kitchen sink’ approach, including as many variables as possible in the first stage estimation to obtain the propensity score. However, this approach runs the risk of understating the treatment effect if some of these variables are themselves affected by the treatment (Smith and Todd 2005). For example, exercise leads to greater fitness and lower BMI scores. Thus, controlling for body composition in the first stage probit model will lead to an attenuated estimate in the second stage. However, body composition may also capture information about otherwise unobservable, but important characteristics, such as discipline. Taking a more conservative approach argues in favor of including these variables in the first stage.

It should also be noted that the results obtained by PSM estimation are not directly comparable to those obtained via ordinary least squares, fixed effects or instrumental variables estimation. The coefficients obtained from a regression model are interpreted as the average treatment effect (ATE), while those obtained via PSM are interpreted as the average treatment effect for the treated (ATT). If the treatment effect does not vary across the population, then the two effects are the same. Otherwise, they may differ and we should be careful to interpret the estimates appropriately.

Data and Empirical Models

This paper employs the 1998 and 2000 waves of the National Longitudinal Surveys of Youth 1979 dataset (NLSY79). The NLSY79, which conducted surveys every year starting in 1979 through 1994, then in even numbered years, began with an initial sample of 12,686 individuals. The initial sample contained oversamples of poor white individuals and members of the armed forces. The military and poor white oversamples were dropped in 1985 and 1991. The present study is restricted to the 1998 and 2000 waves because the exercise variable is only available in those years. During those years, survey participants ranged in age from 33 to 41, thus we are capturing workers during their prime working years and at a time in the life cycle when weight gain starts to set in. Observations were excluded from the final sample if the individual reported working fewer than 500 h or more than 3,500 per year or a weekly income less than one-hundred dollars. These restrictions are employed to remove observations where the individual has a weak labor market attachment or where there may be significant reporting error in the hours worked. All monetary values have been converted to 2006 dollars using the consumer price index.

The survey contains several questions about individuals’ activities. Respondents were asked to answer the following question: “How often do you participate in vigorous physical exercise or sports—such as aerobics, running, swimming, or bicycling?” Responses were placed in the following categories: never, less than once a month, one to three times each month, once or twice a week, three times or more each week. Dummy variables for each level of activity are constructed from this primary variable, excluding the no exercise category. The categorical variables are labeled rarely exercise (less than once a month), infrequent exercise (1–3 times each month), moderate exercise (1–2 times per week) and frequent exercise (3 or more times per week).

These exercise variables are not perfect measures of physical activity. There are two major drawbacks in addition to the potential measurement error discussed earlier. First, they do not capture variations in amount of time spent exercising. An individual who goes for a 2 h bicycle ride twice a week spends more time on physical activity than an individual who jogs 3 days a week for 30 min each time. Second, the data do not capture variations in exercise intensity, which may be an important factor determining the benefits of exercise. In spite of these limitations, the exercise variables do contain important information on exercise habits and provide a useful instrument for determining whether there is a link between physical activity and labor market outcomes.

Body composition variables are constructed using information on each individual’s height and weight. Each individual’s body mass index (BMI) is calculated using the standard formula: BMI=Weight in pounds*703/Height in inches squared. Individuals can be grouped into the following categories given their BMI: underweight if BMI = <18.5; normal weight if18.5<= BMI<25; overweight if 25<=BMI<30; obese if BMI > 30. I then construct indicator variables for whether an individual is overweight or obese (so that underweight and normal weight individuals serve as the reference group).

Five of the explanatory variables included in each model are measured at a single point in time: the armed forces qualifying test (AFQT) score percentile, information on athletics participation in high school (dummy variable equal to one if the individual participated in athletics in high school and zero otherwise), whether the individual watches/attends sporting events as part of his leisure activities (as an indicator variable), the individual’s height and a proxy for the discount rate. The AFQT score was measured in 1980 while respondents were asked to retrospectively report on their high school athletics participation in 1985 (by which time the youngest respondents were 20 years old) and the individual’s adult height was also reported in 1985. The sports attendance/viewing variable was collected in 2004 along with other information about leisure activities. In addition to these variables, the discount rate proxy (which I refer to as the impatience variable from here on) is measured in 2006. This variable is the response to the question “How much additional money would you need to postpone a $1,000 payoff 1 year?” Since many individuals’ response is zero, the variable is constructed as the log of one plus the actual dollar amount reported. While not a perfect measure of the individual’s discount rate, it should provide useful information. From here on, this variable is referred to as ‘impatience.’ Since each variable is measured at a single point in time, they are dropped during fixed effects estimation.

Empirical Models

In addition to the variables described above, the Mincerian wage models contain standard controls including highest grade completed, age, log tenure with the current employer in years, log tenure in years squared, indicator variables for whether the individual is female, black or Hispanic and an indicator variable for whether the individual is covered by a union contract/collective bargaining agreement and the number of hours worked per week. These estimates provide a base to compare the results from the PSM routines. Additionally, the coefficient estimates for the human capital variables provide a means through which to put the treatment effect of exercise into context.

The literature on obesity and earnings indicates a properly specified empirical model also needs to control for body composition. Additionally, Lakdawalla and Philipson (2009) find that BMI is negatively related to job strenuousness. Occupation indicators should capture most of this effect, but will not capture differences in job strenuousness and other factors that vary within occupational group. It is for this reason that each model includes indicator variables for whether the individual is overweight or obese. Thus, the estimated impact of exercise on earnings is independent of its effect on a person’s weight and body composition. Additionally, Ettner (1996) finds increases in income lead to increases in both physical and mental health. This raises the question of causality and also indicates the need to control for health measures to make sure that we are not picking up the effects of changes in health in the coefficient for our exercise variable. To account for other health issues, the models include controls for whether the individual reports a health condition that limits either the type or amount of work that she can perform.5 In both cases, an indicator variable is constructed taking a value of one if the individual reports having that limitation and zero otherwise.

Ruhm’s (2005) finding that a reduction in hours of work is a likely cause of the negative correlation between exercise and macroeconomic conditions indicates it is also critical to control for hours of work in order to account for the effect of labor market conditions on exercise. Including information on local unemployment rates might capture some of the change in labor market conditions but it will not account for differences across occupations. It would also lead to a significant drop in sample size. Therefore, each model also includes a set of occupation indicator variables.

Additionally, each model contains indicator variables for whether the individual participated in athletics during high school and whether viewing/attending sporting events is one of the respondent’s leisure time activities. Controlling for participation in high school athletics ensures that the results are not confounded by the wage effects of participating in high school athletics established in several papers (Barron et al. 2000; Eide and Ronan 2001; Stevenson 2010).6 Watching sports may provide networking opportunities for individuals, thus it is important to control for this factor as well.

It has also been suggested that employer variables such as firm size may affect both earnings and the offering of workplace sponsored wellness programs. The NLSY does contain a measure of firm size (number of employees at the respondent’s location); however, including the variable leads to a substantial reduction in sample size. Estimating the models including this variable does not have a significant impact in the coefficient estimates for the exercise variables. Furthermore, the first stage probits for the PSM routine fail to uncover a significant relationship between firm size and the probability of exercising frequently.

The second part of the empirical analysis employs PSM routines to estimate a causal relationship between exercise and earnings. The PSM model estimates the effect of frequent exercise (at least three times per week). The first stage probit model includes most of the same explanatory variables included in the augmented Mincer equations plus four additional variables. The probit models drop the health limits the amount of work variable in order to satisfy the balancing condition. The health limits amount of work variable was not statistically significant in the first stage probit models in any case; excluding it from the model does not reduce the model’s predictive power or have a significant impact on the estimated propensity score. The respondent’s marital status is included since it might affect motivation to exercise (to the extent that individuals exercise in order to attract a mate). The number of children in the household and an indicator variable for whether there is a child under 7 year old are included in the model since the time demands associated with raising children leaves less time for exercise, especially for the primary care giver (thus we expect the number of children to have a larger impact on exercise for women). In light of studies showing an earnings effect of height, adult height is included in case taller individuals are more likely to participate in sports or exercise, particularly team sports such as basketball and volleyball. All empirical models are estimated separately for men and women.

Summary Statistics

There are some notable differences in exercise frequencies reported in the NLSY79 and other data sets. For the sample used in the present paper, nearly 32% of the respondents never exercise, 16% exercise less than once per month, 14% exercise one to three times each month, 19% exercise once or twice per week and another 19% exercise three or more times per week. The statistic for no exercise is similar to that reported in Ruhm (2005), who analyzed the Behavioral Risk Factor Surveillance System (BRFSS) survey. However, in Ruhm’s (2005) sample, 42% of respondents exercise regularly.7 In the NLSY79 sample, we see that 19% exercise three times or more per week and less than 39% exercise at least once a week. A big reason for the difference between the two samples may lie in the definition of exercise in the BRFSS. The BRFSS includes walking and gardening in their definition of exercise, where the present study focuses on more vigorous physical activity. One might postulate the difference in the two sets of summary statistics arises from the present paper’s focus on employed individuals. However, there is little change in the exercise frequency statistics for the full sample (employed and non-employed individuals) compared to the sample of employed individuals. Additionally, the lack of any duration attached to the exercise frequency question may lead to a difference in how individuals report their exercise activity. For example, a respondent may not report a 20 min jog in the NLSY sample believing it was too short to qualify, while this episode would be captured in the BRFSS.

Table 1 provides summary statistics for the key variables for the full sample and by gender. Given the discussion of exercise above, I will focus on some of the other key variables. In the full sample (column 1) we see that 36.7% of the sample is overweight while 28.3% is classified as obese (so that 65% of the sample is at least overweight). Nearly four percent report their health limits the kind of work they can perform and 3% report being limited in the amount of work. Average weekly income is $782.45 (in 2006 dollars). A comparison of the summary statistics by gender shows that men earn significantly more than women. Men are also more likely to be categorized as being overweight although the incidence of obesity is roughly comparable between the two groups. While men average more than five additional hours of work per week compared to women, they are also more likely to exercise regularly. This discrepancy may be due to a disparity in time devoted to home production, leaving women with less free time (and energy) for exercise.
Table 1

Summary statistics

 

All

Men

Women

Weekly income

782.45 (665.75)

925.55 (766.37)

631.56 (497.04)

Rarely exercise

0.159 (.366)

0.152(.359)

0.167 (.373)

Infrequent exercise

0.143 (.35)

0.155 (.362)

0.131 (.337)

Moderate exercise

0.192 (.394)

0.226 (0.418)

0.156 (0.363)

Frequent exercise

0.196 (.397)

0.235 (0.424)

0.155 (0.362)

Impatience

0.644 (.770)

0.615 (.756)

0.675 (.784)

Body mass index

27.73 (5.84)

27.92 (4.819)

27.54 (6.75)

Overweight

0.367 (.482)

0.454 (.498)

0.275 (.447)

Obese

0.283 (.45)

0.273 (.446)

0.293 (.455)

Health limits kind of work

0.039 (.192)

0.031 (.173)

0.047 (.211)

Health limits amount of work

0.03 (.171)

0.021 (.144)

0.039 (.195)

Hours worked per week

42.8 (9.28)

45.48 (8.59)

39.98 (9.15)

Highest grade completed

13.18 (2.34)

13.05 (2.41)

13.31 (2.25)

AFQT score percentile

0.404 (.282)

0.407 (.296)

0.402 (.267)

Age

38 (2.47)

37.85 (2.49)

38.17 (2.44)

Job tenure in years

5.78 (5.5)

5.87 (5.60)

5.69 (5.4)

Covered by union contract

0.05 (.218)

0.04 (.197)

0.06 (.238)

Female

0.487 (.5)

0 (.0)

1 (.0)

Black

0.283 (.45)

0.272 (.445)

0.293 (.455)

Hispanic

0.18 (.384)

0.184 (.388)

0.176 (.381)

Height in inches

67.25 (4.1)

70.13 (2.933)

64.22 (2.74)

Athletics in high school

0.403 (0.49)

0.448 (0.497)

0.356(0.479)

Watch or attend sporting events

0.66 (0.474)

0.788 (0.409)

0.525 (0.499)

Observations

6,190

3,177

3,013

Table provides means with standard errors in parentheses

Exercise frequencies defined as follows: Rare=less than once a month, Infrequent=1–3 times per month, Moderate=1–2 times per week, Frequent=3+ times per week

Table 2 provides summary statistics for some of the key variables by level of physical activity. The table highlights the differences between exercisers and non-exercisers. Frequent exercisers earn roughly three-hundred and sixty-two dollars more per week on average compared to their non-exercising counterparts. They are much less likely to be obese; however they have a higher incidence of overweight. In fact, we see the incidence of overweight increase with exercise frequency (except for infrequent and moderate exercisers who have roughly the same incidence). This might be explained by the fact that individuals who regularly engage in rigorous weight resistance training have greater muscle mass, which raises the BMI. The inaccuracy of the BMI as a measure of body composition for these individuals is well established, particularly with respect to men (Prentice and Jebb 2001). A closer look at the data shows that this observation only holds true for the men in the sample; women who exercise regularly (at least once per week) have lower rates of both overweight (25 versus 28.5%) and obesity (18.8 versus 34.9%) relative to non-regular exercisers. Conversely, men who exercise regularly have a higher incidence of overweight (49.2 compared to 42.1%) and a lower rate of obesity (22.2 versus 31.9%). Frequent exercisers also average nearly 1.35 additional years of schooling completed and have a higher average AFQT score percentile. These comparisons underscore the importance of dealing with the potential endogeneity of the exercise variable since body composition, educational attainment and exercise frequency may all be codetermined by unobservable factors such as self-control and ability to delay gratification.
Table 2

Summary statistics for select variables

Exercise frequency:

Never

Rare

Infrequent

Moderate

Frequent

Weekly income

621.66 (462.4)

717.46 (549.6)

807.67 (645.17)

781.21 (743.73)

983.61 (860.52)

Highest grade completed

12.57 (2.167)

12.95 (2.144)

13.25 (2.19)

13.53 (2.411)

13.92 (2.5)

AFQT score percentile

0.321 (0.253)

0.405 (0.274)

0.446 (0.279)

0.439 (0.29)

0.471 (0.295)

Overweight

0.326 (0.469)

0.349 (0.477)

0.388 (0.487)

0.384 (0.486)

0.415 (0.493)

Obese

0.364 (0.481)

0.321 (0.467)

0.263 (0.441)

0.238 (0.426)

0.18 (0.385)

Athletics in high school

0.290 (0.454)

0.390 (0.488)

0.432 (0.496)

0.458 (0.498)

0.517 (0.5)

Watch or attend sporting events

0.554 (0.497)

0.634 (0.482)

0.699 (0.459)

0.731 (0.444)

0.751 (0.432)

Number of observations

1,915

987

885

1,189

1,214

Table provides means with standard errors in parentheses

Exercise frequencies defined as follows: Rare=less than once a month, Infrequent=1–3 times per month, Moderate=1–2 times per week, Frequent=3+ times per week

Results

Regression Based Results

Table 3 presents the results when the model is estimated via OLS and fixed effects. All models control for year and occupation fixed effects. Since no exercise is the excluded exercise frequency category, the coefficients on each of the exercise variables are interpreted as the wage premium relative to individuals who never exercise. The OLS results for men (column 1) show a fairly uniform earnings premium through weekly exercise, with a larger premium associated with frequent exercise. T-tests (reported at the bottom of the table) for the equality of the coefficient on frequent exercise with each of the other exercise coefficients show that these differences are significant at least at the 10% level. While impatience does show a negative correlation with earnings, the coefficient is not statistically significant. Being overweight or obese does not have a statistically significant relationship with earnings for men. However, men who report that their health limits the kind of work they can perform do earn substantially lower wages, indicating that physical health is important in the types of jobs many men perform. The remaining coefficients are of the expected signs: education, ability (as measured by the AFQT score), participation in high school athletics, consuming sporting events and job tenure are all associated with higher wages. The fixed effects estimates do not show a statistically significant correlation between earnings and any level of exercise. According to the t-tests, the coefficient on frequent exercise is only statistically significantly different from that on the infrequent exercise indicator variable (and even then only at the 10% level).
Table 3

Exercise frequency and earnings

 

Men

Women

OLS

FE

OLS

FE

Rarely exercise

0.052a (0.027)

0.039 (0.026)

0.018 (0.026)

−0.013 (0.032)

Infrequent exercise

0.05a (0.028)

−0.016 (0.024)

0.014 (0.028)

0.0095 (0.037)

Moderate exercise

0.061* (0.025)

0.034 (0.025)

0.055* (0.027)

−0.033 (0.037)

Frequent exercise

0.105** (0.027)

0.034 (0.027)

0.129** (0.028)

0.079a (0.044)

Impatience

−0.019 (0.012)

 

−0.015 (0.011)

 

Overweight

0.005 (0.021)

0.0073 (0.034)

−0.027 (0.021)

0.159** (0.065)

Obese

−0.014 (0.023)

−0.0084 (0.058)

−0.079** (0.022)

0.211** (0.102)

Health limits kind of work

−0.221** (0.071)

0.115 (0.077)

−0.024 (0.053)

0.075 (0.095)

Health limits amount of work

0.021 (0.08)

−0.062 (0.097)

−0.09 (0.06)

−0.124 (0.10)

Hours worked per week

0.016** (0.0013)

0.0056** (0.0016)

0.022** (0.0012)

0.0096** (0.002)

Highest grade completed

0.051** (0.0054)

0.024 (0.043)

0.066** (0.0053)

0.056 (0.038)

AFQT score percentile

0.368** (0.044)

 

0.432** (0.048)

 

Athletic clubs in high school

0.043* (0.019)

 

0.013 (0.02)

 

Watch or attend sporting events

0.059** (0.022)

 

0.07** (0.018)

 

Age

0.0064 (0.004)

0.043 (0.027)

−0.0074a (0.0042)

−0.008 (0.042)

Log of tenure in years

0.069** (0.0075)

0.029** (0.01)

0.055** (0.0085)

0.018 (0.016)

Log of tenure in years squared

0.0095** (0.0033)

−0.0028 (0.01)

0.021** (0.0036)

0.0024 (0.0082)

Covered by union contract

0.053 (0.04)

0.025 (0.034)

−0.021 (0.04)

0.057 (0.059)

Female

Black

−0.091** (0.024)

 

0.092** (0.024)

 

Hispanic

0.016 (0.024)

 

0.147** (0.028)

 

T-test rare=frequent

3.60a

0.03

11.73**

3.86*

T-test infrequent=frequent

3.89*

3.21a

11.47**

2.08

T-test moderate=frequent

2.94a

0

4.99*

6.76**

R-squared

0.4596

0.1661

0.4374

0.13578

Observations

3,177

3,177

3,013

3,013

Log weekly earnings is the dependent variable

Robust standard errors are reported in parentheses

All models control for year and occupation fixed effects

T-tests are for the equality of coefficients between the frequent exercise indicator and the other exercise frequency indicator variables

a, *, ** denote significance at the 10%, 5% and 1% levels, respectively

By contrast, the results for women exhibit the same pattern as for the full sample. The coefficient on the frequent exercise indicator variable is still statistically significant in the fixed effects model. Comparing other estimates across genders, we see that body composition does not have a significant correlation with earnings for men. Consistent with prior research, the OLS results indicate that attractiveness plays a larger role in labor market success for women. It is also possible that differences in labor force participation rates between men and women could affect these results.

The female only OLS estimates show a positive and increasing association between exercise and earnings (column 3). The correlation becomes significant with moderate exercise (at least once per week) and increases substantially with frequent exercise (three plus times per week). Consistent with the existing literature, obesity has a strong, negative correlation with earnings. In contrast to the male sample, a health issue limiting the type of work she can perform is not associated with lower wages, indicating that men and women sort into different jobs (even after controlling for broad occupation categorical variables as covariates). This is consistent with a large body of literature showing that men and women sort into different types of jobs.

The models estimated via OLS show that being black or Hispanic is associated with higher wages for females while being black is associated with lower wages for men. Further restricting the sample to women who worked at least 1,500 h over the past year significantly reduced the coefficient on the black indicator (b = 0.024) and cuts the indicator on Hispanic by one-third.8 Thus, these results may reflect differences in selection into the workforce between black and non-black, non-Hispanic females. These results are consistent with a pattern where black, female part-time workers are more likely to come from higher up on the ability distribution relative to white, female part-time workers. Furthermore, removal of the AFQT score from the empirical model results in a sign reversal for being black and substantial reduction of the coefficient on the Hispanic variable in the female sample. This is by no means the first paper to find that minority women earn more than white women, controlling for a host of other factors. In a recent paper Miller (2011) analyzes the effects of motherhood timing on career outcomes using the 1979–2000 waves of the NLSY79 and also finds higher wages for Hispanic and black women. She also attributes this (seemingly) unusual result on the inclusion of the AFQT score and sample selection.

The fixed effects estimates continue to show a positive earnings premium associated with frequent exercise (now significant at the 10% level). Moderate exercise no longer exhibits a significant correlation with earnings. The FE results for women also show a highly unusual result in that being overweight or obese is associated with substantially higher earnings. However, these results are a result of the very short nature of the panel (there are an average of 1.3 observations per female). The positive correlations between overweight and obese and earnings disappear when a more basic wage equation (excluding the exercise variables amongst others which are not available in other years of the survey) is fitted using data for the 1989–2000 waves of the NLSY via fixed effects estimation. There are also relatively few women moving into or out of the obese category between the two survey years, raising the possibility that a few outliers are driving these results.

The OLS results indicate that both exercise and body composition has a stronger impact on wages for women relative to men. These results are consistent with the hypothesis that beauty and fitness matter more for women, while the results for the overweight and obese indicator variables are consistent with previous findings in the obesity literature. While these results are consistent with the conjecture that exercise has significantly greater health benefits for women, I am not aware of strong support for this notion from the medical literature. Given Bhattacharya and Bundorf’s (2009) findings, the larger exercise-wage premium for women may be due to the greater health care savings for fit and healthy women relative to men (remember that the women in this sample are in their 30s, still in their child bearing years and a time when they are more likely to have young children at home). Investigating this potential explanation is beyond the scope of the present paper. Finally, it is possible that the larger wage-exercise premium for women reflects the difference in rates of physical activity between the sexes. If the pecuniary returns to exercise are not uniform, assuming individuals with the highest returns are more likely to engage in regular physical activity, and the distribution of these returns is the same for men and women, then we would expect to obtain larger correlations between exercise and earnings for women based on the fact that a significantly smaller fraction of women report frequent or moderate exercise frequency.

Overall, the results by exercise frequency indicate the earnings effect is not uniform across levels of activity. Frequent exercise shows the most robust, positive association with earnings. Therefore, the PSM estimates focus on frequent exercise. These results contrast with Lechner’s (2009) finding that exercising at least once a month is enough to generate wage benefits for German workers.

Propensity Score Matching Estimates

Table 4 presents the results from the first stage for the propensity score matching estimates. The balancing condition is satisfied for each of the explanatory variables across each stratum. Columns 1–2 present the results for men and women, respectively. There is a small change in the sample sizes compared to those reported in Table 3. A small number of observations are dropped during the first stage probit because some of the occupation indicators perfectly predict failure of the dependent variable (i.e. none of the individuals working in that occupation exercise frequently). These reductions are very modest (nine observations lost for the sample of men and fourteen for the women) and should not have any significant impact on the estimates. For men, the number of children does not show a significant association with frequent exercise, while married men are less likely to exercise frequently. Men with young children in the household are also less likely to exercise frequently. The balancing condition is satisfied with the full set of explanatory variables for both samples. Unlike the male sample, having a health condition that limits the kind of work a woman can perform does not affect exercise. Having more children is associated with a lower probability of exercising frequently; however having young children does not affect exercise. These results are consistent with the observation that women still tend to perform the majority of the child raising duties, taking away from time available for other activities. Figures 1 and 2 illustrate the propensity score distributions for the male and female samples, respectively. They show significant overlap between the treated and untreated observations, with only the upper tails showing a region without overlap. In particular, the upper end of the distribution for the women only sample ends with a propensity score below 0.5 for those who do not exercise frequently while the propensity score distribution for those who do exercise frequently continues up to estimates above 0.6. Overall, there is enough overlap along propensity score estimates between the treated and untreated observations to support the PSM routine.
Table 4

First stage results from propensity score matching

Variable

Men

Women

Impatience

−0.065a (0.036)

−0.062 (0.039)

Height

0.0068 (0.0092)

0.011 (0.011)

Athletic clubs in high school

0.13* (0.056)

0.187** (0.062)

Watch or attend sporting events

0.145* (0.067)

0.144* (0.061)

Overweight

0.044 (0.061)

−0.17* (0.07)

Obese

−0.278** (0.072)

−0.445** (0.077)

Health limits kind of work

−0.398* (0.167)

0.039 (0.141)

Hours worked per week

−0.0055a (0.003)

0.00006 (0.0033)

Highest grade completed

0.07** (0.016)

0.058** (0.017)

AFQT score percentile

0.041 (0.133)

0.067 (0.152)

Age

−0.0045 (0.012)

−0.0047 (0.014)

Number of children

0.013 (0.025)

−0.062* (0.028)

Children under seven

−0.135* (0.068)

0.0047 (0.077)

Married

−0.165** (0.064)

0.039 (0.064)

Black

0.105 (0.07)

−0.163* (0.086)

Hispanic

0.099 (0.074)

0.04 (0.087)

Observations

3,168

2,999

Log likelihood

−1,632.54

−1,203.82

Table presents the coefficient estimates from the probit models used to estimate the propensity score

An indicator variable for frequent exercise is the dependent variable

Both models include year and occupation indicators

a, *, ** denote significance at the 10%, 5% and 1% levels, respectively

Table 5 presents the estimates for the average treatment effect on the treated (those who exercise frequently) for PSM using the stratification matching method.9 The results for both genders show a positive and statistically significant earnings premium associated with exercise, with a larger impact for women. However, given the size of the standard errors, we cannot say that this difference is statistically significant. The estimates for both genders fall between the OLS and fixed effects estimates. This pattern is consistent with OLS overestimating the true effect due to omitted variables bias and fixed effects underestimating the true effect because of an attenuation bias due to measurement error. However, this interpretation must be viewed with caution since the PSM estimates the ATT as opposed to the ATE, which may be different. Overall, the results show that engaging in frequent exercise is associated with a significant earnings premium. Women (men) who exercise frequently earn 11.9 (6.7) percent more on average than women (men) who do not exercise frequently (this includes all other levels of physical activity and inactivity). Using the earnings effect of schooling obtained via OLS estimation (presented in Table 3), where an additional year of schooling results in a 6.6 (5.1) percent earnings gain for women (men), the results indicate the engaging in frequent exercise-earnings premium is equal to nearly one and one-third additional years of schooling completed for men and 1.8 additional years of schooling for women.
Table 5

Propensity score matching estimates

 

Men

Women

Average treatment effect

0.067* (0.028)

0.119** (0.033)

Observations in treatment

748

458

Observations in control

2,420

2,541

Log weekly earnings is the dependent variable

Table presents the estimated average treatment effect for the treated of engaging in frequent exercise using the stratification matching method

Results are interpreted as the wage premium associated with frequent exercise compared to not exercising frequently

*, ** denote significance at the 10%, 5% and 1% levels, respectively

Robust standard errors are in parentheses

While the stratification method is widely used to estimate models via propensity score matching, the second stage estimates were obtained using two additional methods: nearest neighbor and kernel estimation. These results, presented in appendix Table 9, show a fairly robust, positive relationship between frequent exercise and weekly earnings. The nearest neighbor (kernel) estimates are smaller (larger) than those obtained using the stratification method. The nearest neighbor method, applied to the sample of men is the only one that does not yield a statistically significant coefficient. This may be due to the significant reduction in sample size as the nearest neighbor method only compares each treated observation to one nearest neighbor. The statistics show that there are fewer non-treated neighbors than treated observations, meaning that some neighboring observations are used multiple times.

Robustness Check: Sample Restriction

I restrict the sample by excluding part-time workers (keeping only those who worked more than 30 h per week), the severely underweight (BMI < 16), the morbidly obese (BMI > 40) and workers with an hourly wage at or below the minimum wage in 1998 and 2000 (which stood at $5.15). The results using the OLS and FE estimators are presented in Table 6. While all models contain the full set of covariates, only the results for the exercise frequency and body composition variables are presented for the sake of brevity. The results for the exercise variables are highly similar to those presented in Table 3. The coefficients on the frequent exercise variable maintain the same level of significance, except where the FE estimator is applied to the sample of females only. While the corresponding result in Table 3 indicates a nearly eight percent increase in wages associated with frequent exercise, the results from the restricted sample show a nearly 10% wage increase with a coefficient that is now significant at the 5% level. Another important change is the substantial decrease in the magnitude of the coefficients for overweight and obese in the fixed effects specification for the female sample. The coefficient on the obese indicator is no longer significant while that on the overweight indicator is only significant at the 10% level. It appears that much of the positive correlation between becoming overweight or obese was due to individuals who reported hourly wages below the minimum wage.10
Table 6

Regression-based results for the restricted sample

 

Men

Women

OLS

FE

OLS

FE

Rarely exercise

0.049a (0.027)

0.024 (0.024)

−0.004 (0.028)

−0.001 (0.027)

Infrequent exercise

0.046a (0.028)

−0.024 (0.024)

0.0041 (0.031)

0.037 (0.038)

Moderate exercise

0.06* (0.026)

0.032 (0.025)

0.037 (0.029)

0.0059 (0.038)

Frequent exercise

0.094** (0.027)

0.032 (0.028)

0.119** (0.029)

0.097* (0.046)

Overweight

0.0006 (0.021)

0.017 (0.031)

−0.036a (0.022)

0.094a (0.05)

Obese

−0.017 (0.023)

0.03 (0.052)

−0.069** (0.024)

0.112 (0.076)

T-test rare=frequent

2.62

0.07

11.51**

4.38*

T-test infrequent=frequent

2.94a

3.77a

9.58**

1.35

T-test moderate=frequent

1.76

0

4.71*

3.67a

R-squared

0.4562

0.123

0.4128

0.1546

Observations

3,009

3,009

2,413

2,413

All models contain the full set of covariates, including year and industry indicators

Log weekly earnings is the dependent variable

T-tests are for the equality of coefficients between the frequent exercise indicator and the other exercise frequency indicator variables

The samples exclude part-time workers, those reporting an hourly wage below the federal minimum wage, the morbidly obese and the severely underweight

a, *, ** denote significance at the 10%, 5% and 1% levels, respectively

Robust standard errors are in parentheses

The results for the PSM routine on the same restricted samples (Table 7) also show similar results to those obtained for the more inclusive samples. In the first stage of the PSM routine, the height variable was excluded from the full and female only samples in order to meet the matching criterion. This exclusion should not pose any serious problems for the matching estimator since the height variable did not yield a significant coefficient in the first stage for any of the three samples. The results indicate that men who exercise frequently earn 6.6% more than those who do not (compared to 6.7% in the broader sample) while women who exercise frequently earn 11.2% more (compared to 11.9%). These results support the primary findings reported in Table 5.
Table 7

Propensity score matching estimates for the restricted sample

 

Men

Women

Average treatment effect

0.066* (0.028)

0.112** (0.033)

Observations in treatment

711

382

Observations in control

2,289

2,018

Table presents the estimated average treatment effect for the treated of engaging in frequent exercise using the stratification matching method

The samples exclude part-time workers, those reporting an hourly wage below the federal minimum wage, the morbidly obese and the severely underweight

Results are interpreted as the wage premium associated with frequent exercise compared to not exercising frequently

*, ** denote significance at the 10%, 5% and 1% levels, respectively

Robustness Check: Post-Schooling Training

As an additional robustness check, each model was estimated including a measure of post-schooling training (using the broader sample). The variable measures the total number of weeks spent in training (both on-the-job and outside of work) between 1985 and 1996. Since this variable is only measured once, it is dropped from the fixed effects model. This variable was not included in the primary results because it leads to a substantial reduction in total sample size from 6,190 to 3,563 combined observations for men and women. The results are presented in appendix Table 10. These results support those presented in Tables 4 and 6. In fact, both the OLS and PSM estimates including the training variable are larger than those obtained in the models excluding the training variable. Only the fixed effects estimates are smaller. Estimating the model using this smaller sample, but excluding the training variable yields similar results to those presented in Table 10. Thus, it appears that the reduction in the coefficient is due entirely to the change in the estimation sample.

Robustness Check: Past and Future Exercise

As an additional robustness check, I exploit the (short) panel nature of the data. Appendix Table 11 presents estimates for the full sample replacing the frequent exercise indicator variable with an indicator variable for future frequent exercise (panel A) and past frequent exercise (panel B). The models are estimated via OLS and PSM. The model estimated via OLS contains the same explanatory variables as the models presented in Table 3, while the PSM estimates use the same first stage covariates as the models presented in Table 4. The results show that future exercise does not have a statistically significant correlation with current earnings for either gender using either estimator. By contrast, the results in panel B show that past exercise does have a statistically significant correlation with current earnings for men. While the results for the female sample are not statistically significant, much of this is due to the increase in the standard errors resulting from the loss of sample size. According to these results, it does not appear that reverse causality is driving the results presented in Tables 3 and 5.

Robustness Check: Regular Exercise and Earnings

Finally, Table 8 presents results from a model estimated where the set of exercise variables are replaced with a single indicator variable for regular exercise (at least once per week).11 Thus, the results presented in this table show the wage premium for those who exercise regularly relative to those who do not. The results show a positive association between exercise and earnings for both men and women and continue to show a larger earnings premium for women. The exception comes from the FE estimates. However, we should bear in mind the issues associated with applying FE estimation to short panels. The estimated average treatment effect for men is similar to that presented in Table 5. This is consistent with the FE estimates which showed similar exercise-earnings premium for men with moderate and frequent exercise. By contrast, the average treatment effect for women is smaller than the one presented in Table 5; consistent with the increasing exercise-wage premium for women presented in Table 4.
Table 8

Relationship between regular exercise and earnings

 

OLS

FE

PSM

Panel A: Men

Regular exercise

0.052** (0.017)

0.03* (0.017)

0.068** (0.023)

Panel B: Women

Regular exercise

0.084** (0.02)

0.011 (0.03)

0.086** (0.026)

Table presents the results of the empirical model estimated with an indicator for regular exercise replacing the other categorical exercise variables

Regular exercise is defined as exercising at least weekly

The OLS and FE models include the full set of covariates

In order to meet the balancing criterion, the occupation indicators were dropped from the first stage of the PSM routine

The PSM estimates report the average treatment effect for the treated of regular exercise compared to not engaging in regular exercise using the stratification method

*, ** denote significance at the 10%, 5% and 1% levels, respectively

Robust standard errors are in parentheses

Conclusions

This paper finds a positive and significant treatment effect of engaging in frequent exercise on labor market earnings. The results are fairly robust to a variety of estimation techniques, including propensity score matching. As a next step, data with more detailed information on exercise duration, type and frequency should be used to examine whether the relationships found in this study are robust to various types of exercise and whether they depend on exercise intensity or if an intensity threshold exists. Furthermore, more work is needed to investigate other potential labor market benefits of exercise such as lower unemployment rates or changes in the likelihood of promotion. Increasing public awareness of the labor market benefits of exercise may provide another tool in motivating people to adopt more active lifestyles.

From an employer’s perspective, the present findings combined with earlier findings that exercise can positively affect job satisfaction (Thogersen et al. 2005) and the connection between job satisfaction and multiple workplace factors such as absenteeism (Clegg 1983) and productivity (Mangione and Quinn 1975) suggest employer sponsored exercise programs and gym memberships may have a positive impact on firms’ financial health beyond their impact on attracting workers and lowering health insurance premiums. More work is needed to assess the effectiveness of these programs in raising productivity and profitability.

Footnotes
1

While the life-cycle hypothesis predicts that individuals will substitute towards leisure when wages are low (as they are during an economic downturn) it is important to bear in mind that hours worked is not a decision taken unilaterally by the worker. A significant portion of the decline in average work hours during an economic downturn is likely dictated by employers.

 
2

While hours of work may be fairly unresponsive to wage changes, individuals may respond to wage increases by decreasing the amount of time they spend on household work (by eating out more often or hiring a cleaning service) and spending more time on leisure activities or exercising. If exercise services are a normal good, then the income effect of a wage or salary increase may lead to greater consumption of these services.

 
3

The sample examined in this paper contains a maximum of two observations per individual.

 
4

Thus, I can no longer employ the categorical exercise variable. Instead, the primary PSM estimation focuses on the effects of frequent exercise.

 
5

Robustness checks address the issue of reverse causality. These tests are described later in the paper.

 
6

Only 17% of individuals who did not participate in athletics in high school report frequent exercise in the current sample compared to 27% for those who did participate in high school athletics. The percentages for exercising at least once per week are 32.7 and 47.4 for non-high school athletes and high school athletes, respectively.

 
7

Ruhm defines regular exercise as exercising three or more times a week for at least 20 min. The NLSY79 did not collect information on the duration or intensity of exercise, only the frequency.

 
8

The coefficient on black is no longer statistically significant.

 
9

Estimates using kernel and nearest neighbor matching are presented in appendix table 9 as a robustness check. The results are qualitatively similar with those obtained using the stratification matching method. In most cases, the estimates obtained using the other matching methods are even larger than the ones presented in table five.

 
10

Fitting the model using each category of exclusions separately did not yield substantially reduced coefficients on the overweight and obese variables for the part-time and extreme BMI restrictions. The large drop in the estimated coefficients occurs when low-wage workers are excluded from the sample.

 
11

In order to meet the balancing criterion, the occupation and year indicators were dropped from the first stage of the estimation routine for both samples. Neither the occupation nor the year indicators exhibited a significant correlation with the exercise variable in the first stage indicating that it is safe to exclude them.

 

Copyright information

© Springer Science+Business Media, LLC 2011