There exist remarkably large differences in body weights and obesity prevalence between black and white women in the United States; and crucially, these differences are a significant contributor to black-white inequalities in health. In this article, we investigate the most proximal explanations for the weight gap: namely, differences in diet and exercise. More specifically, we decompose black-white differences in body mass index and waist-to-height ratio into components reflecting black-white differences in energy intake and energy expenditure. The analysis indicates that overconsumption is much more important than a lack of exercise in explaining the weight gap, which suggests that diet interventions will have to play a fundamental role if the weight gap between black and white women is to decline.
The proportion of Americans who are overweight or obese has reached alarming levels; approximately 66% of U.S. adults are overweight, and 32% are obese (Ogden et al. 2006).1 This has significant consequences because the overweight and obese have much higher rates of cardiovascular disease, diabetes, hypertension, cancer, and premature death than normal-weight individuals (Flegal et al. 2005; Fontaine et al. 2003; Mokdad et al. 2003). Furthermore, the excess medical expenditures that result from the diagnosis and treatment (including hospitalization costs) of obesity-related diseases are enormous: obesity-related medical expenditures are estimated to cost the United States $75 billion each year (in 2003 dollars) and to be responsible for between 4.3% and 7% of total health care expenditures (Colditz and Wang 2008).
Worse still, the prevalence and consequences of obesity are not spread evenly across the population. Obesity rates are much higher among some racial and ethnic minorities, and these differences are particularly apparent for women: approximately 54% of African American (hereafter, “black”) women are obese, and 15% are extremely obese (BMI ≥ 40); comparable figures for non-Hispanic white (hereafter, “white”) women are 30% and 6%, respectively (Ogden et al. 2006). Moreover, the female black-white “weight gap” has increased by around five percentage points since the late 1970s despite the considerable narrowing of other racial gaps during this period and the considerable gains in the general health of Americans.
The female weight gap is a significant problem because it is causing black-white disparities in rates of illness. For example, Brancati et al. (2000) reported that black women are more than twice as likely to develop Type 2 diabetes than white women, and that the black-white differences in modifiable risk factors (particularly adiposity) account for almost 50% of the excess risk. In addition, the weight gap will be contributing to existing black-white disparities in socioeconomic status if obese persons are discriminated against in the labor market; see evidence in Baum and Ford (2004), Cawley (2004), and Morris (2006). With regard to the weight gap, Penn and Zalesne (2007:183) stated that “[w]hile we focus on many of the challenges of the black community such as improving education and creating new opportunities for young people, this clear and statistically significant problem remains essentially unaddressed, effectively swept under the rug, despite high human and social cost.”
The most often considered explanation for the weight gap concerns black-white differences in socioeconomic status (SES). The SES explanation has been tested, implicitly or explicitly, in a number of studies that regress body mass index (BMI) on measures of race and individual SES (e.g., Burke and Heiland 2008; Chou et al. 2004; Lakdawalla and Philipson 2002). These studies have found that blacks exhibit significantly higher BMI than whites, even after controlling for individual differences in SES. For example, Burke and Heiland (2008) found that controlling for educational attainment, income, occupation, location of residence, and marital status reduces the weight gap by only 0.71 BMI units: from 4.04 to 3.33. Further evidence against the SES explanation is that the income-obesity gap has narrowed quite dramatically since the 1970s, but the race-obesity gap has not.
Another explanation that has received some attention, predominantly in the medical and epidemiology literatures, is that there are important genetic differences between races. In particular, some studies have shown that the basal metabolic rate (energy expended in maintaining basic body functions at rest) is lower for blacks than whites (e.g., Carpenter et al. 1998; Sharp et al. 2002; Weyer et al. 1999). However, if a black-white difference in basal metabolic rate is the principal cause, we would also expect to observe a weight gap for men, which we do not: approximately 34% of black men are obese, compared with 31% of white men (Ogden et al. 2006).
The absence of conclusive SES and genetic explanations for the female weight gap highlights the need for further research. In this article, we investigate the most proximal explanations for the gap: namely, differences in diet (i.e., energy intake), exercise (i.e., energy expenditure), and age. Uncovering whether black-white differences in energy intake or energy expenditure are causing the weight gap is important because before effective policy can be enacted to reduce the gap, we must ascertain whether unhealthy eating practices or sedentary lifestyles are mostly responsible.
Our methodological approach involves decomposing black-white differences in adiposity (i.e., fat) distributions with the Blinder-Oaxaca and the DiNardo-Fortin-Lemieux decomposition procedures. An important issue in performing the decompositions is that adiposity, energy intake, and energy expenditure are often measured with considerable error. We respond to the issue of measurement error by using a number of adiposity measures, adjusted energy intake, and objectively measured energy expenditure. The main result from the analysis is that overconsumption is much more important than a lack of exercise in explaining the female weight gap. For example, using our preferred specification, the proportion of the gap explained by differences in energy intake is more than three times larger than the proportion explained by differences in energy expenditure for each measure of adiposity. The finding suggests that if the remarkably large weight gap between black women and white women is to decline, diet interventions will have to play a fundamental role.
Our aim is to determine the proportions of the female weight gap that can be attributed to black-white differences in diet and exercise. In addition, we determine the proportion of the gap that is due to differences in the age distributions of the black and white populations. Age differences are also examined because age affects adiposity independently of energy intake and expenditure levels. We do not control for other individual-level characteristics, such as socioeconomic or marital status, because they should affect adiposity only through their effect on intake and expenditure.
The main advantages of the BO decomposition procedure are that it is well known, is based on standard statistical methods (i.e., linear regression), and provides easily interpretable results. For these three reasons, we use the procedure in this article. An important limitation, however, is that the researcher must specify a parametric model of the relationship between outcomes and observable characteristics. The parametric nature of the BO decomposition is particularly problematic for our application because the relationship between adiposity and the variables representing energy intake and energy expenditure are highly nonlinear (Burke and Heiland 2007). The nonlinearity implies that empirical models assuming adiposity are linearly separable in intake and expenditure, such as that used in the BO decomposition, are misspecified.
To overcome this potential misspecification, we also use the DiNardo-Fortin-Lemieux (DFL) decomposition (DiNardo et al. 1996).2 The DFL procedure does not require a parametric model of the relationship between adiposity and the variables representing energy intake and energy expenditure. Another advantage of using the DFL procedure is that we can consider changes across the entire distribution. For example, we are able to decompose changes at multiple points of the distribution (e.g., 25th and 75th percentiles) and decompose changes in the proportion of people who are overweight and obese. This is an important advantage because differences in average BMI or obesity rates may not fully reflect important differences in adiposity, especially if the distributions have different shapes.
The DFL decomposition involves reweighting the black sample of women to have the same distributions of intake, expenditure, and age as the white sample.3 The opposite counterfactual of reweighting white women is not examined because there are few whites in our data with high energy intake levels: the maximum caloric intake for whites is 4,597, and the maximum caloric intake for blacks is 5,129; the maximum sugar intake for whites is 392 grams compared with the maximum sugar intake for blacks of 614 grams. Thus, it is not possible to speculate about the BMI-intake relationship at high energy intake levels without making a functional form assumption and extrapolating out of the observed energy intake range for whites (Barsky et al. 2002). The difference between the actual BMI distributions of whites and blacks and a series of counterfactual BMI distributions form the basis of the decompositions underlying our empirical results. More specific details of this procedure are outlined in the following subsection.
DiNardo-Fortin-Lemieux Decomposition Procedure
This equation expresses the marginal BMI distribution as the product of two conditional distributions: the distribution of BMI conditional on BMI determinants and race, and the distribution of BMI determinants conditional on race.
The first term in Eq. 3 denotes the conditional BMI distribution, given all BMI determinants and race. The second term is the conditional distribution of energy intake given energy expenditure, age, and race. The third term is the conditional distribution of energy expenditure, given age and race. Finally, the fourth term is the distribution of age conditional on race only.
Repeating this procedure, we can also define a counterfactual BMI distribution fC(b) that would result if white women retained their own conditional distributions of energy intake, energy expenditures, and age but had the same conditional BMI distribution of black women.
In Eq. 6, the first term on the right-hand side captures the effect of differences in energy intake on the differences in BMI, the second captures the effect of energy expenditures differences, and the third captures the effect of age differences. The last term captures the unexplained portion of differences in BMI.
Similarly, reweighting the observed BMI distribution for r = 2 with suitable constructed weights provides the counterfactual distributions fB(b) and fC(b), respectively.
In the literature applying the DFL decomposition, there is no clear consensus regarding how best to estimate the probabilities (or propensity scores) that are used to create the weights. Here, we follow Bell and Pitt (1998) and Cobb-Clark and Hildebrand (2006) in using a logit model with independent variables entered additively and linearly.5 In contrast, for example, DiNardo et al. (1996) and Butcher and DiNardo (2002) chose to include some interaction terms in their binary choice model. Recently, Millimet and Tchernis (2009) suggested that there might be some benefits to overspecifying the propensity score model by including higher-order or interaction terms. Based on Monte Carlo evidence, they found little penalty for overfitting propensity scores and, in fact, found numerous cases in which overspecifying the model proved beneficial. In our case, if we are to include interaction terms in our propensity score model, it would make sense to do so only within each category and not across categories. Doing the latter is problematic because it is not clear, for example, whether one should place an interaction term involving an intake and expenditure variable in the intake or expenditure category. When we experimented with using a logit model with all possible interaction terms, the results are qualitatively very similar. Hence, we chose to present results only for the simpler model in this article.
Note that our framework, which partitions BMI determinants into n = 3 components, allows a total of 2n – 1 = 7 counterfactual distributions and weights for black women to be computed. The counterfactual distributions fA(b), fB(b), and fC(b) discussed earlier in this article are not the only ways in which we can write out the counterfactual distributions; this is because in the black women’s BMI distribution, we can introduce up to n different distributions of white women’s characteristics at once. These seven alternative counterfactual distributions in turn allow a total of n! = 6 possible decomposition sequences, with Eq. 6 representing just one of the possible ways of expressing the decomposition. Given that we have no a priori preference for one weighted counterfactual distribution over another, we follow the approach used in Cobb-Clark and Hildebrand (2006): we perform each possible decomposition and then present the simple average across all possible decompositions.
Another important issue that arises when allowing several factors in the decomposition is the order in which these factors enter the decomposition: that is, intake first, expenditures second, and age third. Importantly, the proportion of the BMI growth that is attributable to each factor varies, depending on the sequence. Some authors choose a predetermined order and then use the reverse order as a robustness check (e.g., DiNardo et al. 1996). Our approach is based on using an order in which the most endogenous term will be represented by the first term in the conditional distribution, and the most exogenous term the last. In our context, the factors are considered in the primary sequence of (1) intake, (2) expenditures, and (3) age; a second candidate sequence is (1) expenditures, (2) intake, and (3) age. We focus on these two sequences because it is likely that either energy intake or energy expenditure play the leading role in explaining black-white BMI differences. Because employing either of these sequences makes little qualitative difference to our results, we focus on presenting results based on using the primary sequence.
Measurement Error in Intake, Expenditure, and Adiposity
An important issue that arises when investigating the role of energy intake is that it is often measured with considerable error. For example, energy intake is usually underreported, and the extent of underreporting is associated with individual-level characteristics, such as weight, obesity, and sex (Lichtman et al. 1993; Macdiarmid and Blundell 1998; Nielsen and Adair 2007; Schoeller 1995). In a decomposition framework, the measurement error will bias the estimated importance of a factor toward zero. We overcome the issue in two ways. First, we use high-quality data from the National Health and Nutrition Examination Surveys (NHANES), which contain detailed dietary interviews and objectively measured energy expenditure. Second, we provide decomposition results based on samples that are adjusted for the underreporting of energy intake.
To adjust the sample for underreporting, we follow the most common approach in the public health and medical literatures and classify an individual’s self-reported food intake values as implausible if the individual’s total energy intake (EI) is considerably smaller or larger than that individual’s estimated energy expenditure (EE) (see Goldberg et al. 1991). More specifically, we classify an observation as implausible if the EI:BMR ratio (called the “intake ratio”) is more than two standard deviations (SD) smaller or larger than the average EE:BMR ratio (referred to as the “physical activity level” or the “average daily metabolic rate”), where BMR denotes an individual’s estimated basal metabolic rate. We use estimated age- and gender-specific values of average physical activity levels that are based on a summary of 74 medical studies (Black 2000: Table 1).6 These values are considered conservative (lower bound) estimates because most of the studies oversampled white-collar workers relative to workers in manual occupations (Black 2000). We estimate the intake ratio by using the total EI values calculated by NHANES and by estimating BMR for each individual using the formula in Mifflin et al. (1990).7
Figure 1 presents a scatterplot of BMI versus self-reported EI, with plausible observations represented by a circle and implausible observations represented by an “x”. Also shown is the estimated linear relationship between BMI and EI using all observations and using plausible observations only. The figure demonstrates that most of the implausible observations are those with low reported EI values: average caloric intake of implausibly low observations equals just 1,357 kcal, and the average BMI of the women with these low observations is 31.24. To help put this figure in perspective, the average female in our analysis sample—45 years old, 1.64 m tall, weighing 76.8 kg (BMI = 28.6), and exercising little or not at all—would require approximately 1,800 calories per day to maintain her current weight.8 Figure 1 also demonstrates that the correlation between BMI and EI using all observations is approximately zero and is significantly positive when using eligible observations only.
Measurement error is also an issue when examining the role of EE. Measurement of physical activity is typically self reported and also centered on leisure-time physical activities, such as walking, cycling, or other sport-related activities. However, these standard self-reported measures of leisure-time activities do not cover many day-to-day domestic-based activities (e.g., household chores, gardening, or yard work) that involve substantial EEs. As a result, prevalence estimates of physical activity may be underestimated (Weller and Corey 1998). Even if measures of both leisure-time and non-leisure-time activities are available, they rely on perfect recall and comprehension by the survey participants. Recently, the National Institutes of Health (NIH) has supported improvements in data collection by including an objective measure of physical activity, which involves requiring survey respondents to wear physical activity monitors for several consecutive days. We employ such objective measures of physical activity in our article.
Finally, although adiposity has almost universally been measured in the social science literature using BMI, a limitation of using BMI as a measure of adiposity is that it cannot distinguish fat from muscle mass. As a result, health risks tend to be overstated in muscular persons and understated in older persons. Burkhauser and Cawley (2008) showed that the limitation leads to an overestimation of the black-white weight gap. Therefore, in addition to the decomposition of BMI, we decompose differences in the black and white distributions of the waist-to-height ratio (WHR). WHR is chosen as an ancillary measure because a number of studies find that WHR is an exceptional tool for quantifying adiposity. For example, Lee et al. (2008) conducted a meta-analysis to determine which of the prominent indices of overweight and obesity is the best discriminator, and found that WHR was the best discriminator for hypertension, diabetes, and elevated cholesterol in both sexes.
Data, Definitions, and Descriptive Statistics
The data we use come from the 2003–2004 and 2005–2006 National Health and Nutrition Examination Surveys (NHANES). NHANES is a nationally representative cross-sectional survey that has been used repeatedly in the analysis of U.S. obesity. The surveys consist of interview, examination, and laboratory components. They collect information on individual’s demographics (e.g., age and gender); socioeconomic status (e.g., labor income and government program participation); health; and, most importantly for this research, physical measurements, food consumption, and EE.9 The survey oversamples Mexican Americans, African Americans, adolescents aged 12–19, pregnant women, and persons aged 60 and older to assure adequate representation, and includes weights to make the sample nationally representative.
The unique feature of NHANES is its use of mobile examination centers (MEC) to obtain objective body measurements from all survey participants, in addition to medical examinations and laboratory tests. In the 2003–2004 and 2005–2006 versions, trained technicians obtained measurements of individuals: weight; height; leg and arm length; arm, waist, thigh, and calf circumference; and triceps and scapular skinfold. We use this body measurement information to compute our two measures of adiposity: namely, BMI and WHR.10 Bioelectrical impedance analysis (BIA) readings are also available in NHANES 2003–2004 and enable one to construct measures of fat-free mass and body fat using conversion equations (Burkhauser and Cawley 2008), but they are not available in NHANES 2005–2006. Using observations solely from NHANES 2003–2004 results in a sample size that is too small for our analysis to be reliably undertaken, and so we do not analyze the BIA readings.
For our main analysis, we focus exclusively on non-Hispanic white and black female NHANES participants aged 20–74 who had valid weight, height, and waist measurements, and who had valid information regarding EI and EE. The original sample size in NHANES 2003–2004 and NHANES 2005–2006 is a combined total of 20,470 individuals, of which 10,420 are female. When restricted to blacks and whites, the sample drops to 6,791. When we also impose the age restriction, the sample is reduced to 3,164. Further restricting it to those with valid physical activity recordings and energy input measurements leaves us with a sample size of 1,768 for our analysis.
Measuring Energy Intake and Energy Expenditure
Energy intake and energy expenditure are notoriously difficult to measure on either an individual or a population level. A benefit of using the NHANES survey, however, is that it has been at the forefront of using new survey methodologies and technology to obtain accurate estimates of these items. In 2002, NHANES began using a new automated dietary interview system developed by the U.S. Department of Agriculture (USDA). Dietary information is first collected from NHANES respondents in MECs, where a trained dietary interviewer records all food eaten by the respondent in the prior 24-hour period (midnight to midnight) with a computer-assisted multiple-pass dietary interview. The type and amount of foods consumed are recalled with help from aids, such as abstract food models, special charts, measuring cups, and rulers, which help in quantifying the amounts consumed. Special probes were also used to help recall commonly forgotten items, such as condiments, accompaniments, fast foods, and alcoholic beverages, among others. All participants are also asked to complete a second 24-hour dietary recall (Day 2) interview. The NHANES Day 2 dietary recalls are collected by telephone approximately 3 to 10 days after the MEC exam. The information obtained from both dietary interviews is converted using the Food and Nutrient Database for Dietary Studies into variables representing the weight (in grams) of each food type. These are the variables that we use to represent EI in the decomposition analysis. More specifically, we use eight variables representing the average number of grams consumed of protein, starch, sugar, fiber, saturated fat, monounsaturated fat, polyunsaturated fat, and alcohol.
Physical activity measurement is characterized by the synthesis of information on the type, frequency, intensity, and duration of activity over a specified period. The NIH has recently supported improvements in the methods for assessing physical activity by including an objective measure of physical activity in NHANES, beginning with NHANES 2003–2004. Energy expenditure is objectively measured using physical activity monitors—ActiGraph accelerometers (ACC)—that detect all locomotion-type activities, such as walking or jogging. The monitors were attached to an elastic belt and worn at the right hip by people older than 6 years who did not have impairments to walking or wearing the monitor. The benefit of the monitors is that they provide a means of capturing nonstructured activities that are often difficult for survey respondents to self-report. We use data from the monitors that, according to the NHANES data-quality procedures, are reliable.11 Furthermore, data are used only if the participant wore the monitor for at least four days and for at least 600 minutes per day; although participants were asked to wear the monitor for seven consecutive days, some participants did not wear the monitor for the whole week. Comparing objective and subjective measures of physical activity in NHANES 2003–2004, Troiano et al. (2008) found that accelerometer-measured activity is substantially lower than self-reported levels. They suggested that caution be taken when interpreting self-reported physical activity in guiding the design of intervention efforts.
We follow the approach used in Ness et al. (2007) and code one minute of ACC data as one minute of either moderate physical activity or vigorous physical activity, depending on the value of the ACC count within that minute. From these definitions, we then construct three variables to represent energy expenditure: mean duration (minutes) of moderate activity bouts, mean duration (minutes) of vigorous activity bouts, and mean intensity count per minute, where a bout is defined as a continual stretch of physical activity.12
Table 1 reports for white and black females summary statistics for our BMI and WHR adiposity measures. As expected, the statistics reveal that black women have significantly higher levels of adiposity than white women. The figures for black women and white women for each adiposity measure are 31.7 and 28.1 for mean BMI, 0.61 and 0.57 for mean WHR, and 0.55 and 0.32 for proportion obese (BMI > 30).13 To better illustrate the different locations and shapes of the white and black BMI and WHR distributions, we plot kernel density estimates of BMI and WHR in Figs. 2 and 3, respectively. The estimates clearly show that the black BMI and WHR distributions lie to the right of the white distributions, and also that the black distributions have substantially more density at higher values—values where health risks are particularly high. The black and white distributions are also differently shaped. For example, the skewness of the black BMI distribution equals 0.60, whereas the skewness of the white BMI distribution equals 1.29. It is the difference in skewness that is partially responsible for the dramatic differences in the proportion of the populations that are obese. The difference in shapes supports the use of the DFL decomposition, which is able to decompose differences at any point in the distribution, rather than only at mean values.
The EI and EE descriptive statistics in Table 1 also reveal some significant differences. Black women consume less fiber and alcohol but significantly more sugar than white women. Also, average duration of moderate physical activity is significantly less for black women. There are, however, no significant differences in protein, starch, and fat consumption, and there are no significant differences in mean duration of vigorous physical activity and mean physical activity per minute. Thus, descriptive statistics based on mean differences do not indicate which of EI and EE is most to blame for the weight gap. For this, we must rely on our decomposition analysis.
Finally, the sample averages suggest that black females in the NHANES sample are, on average, younger than white females. The significant age difference likely has a negative impact on the weight gap: adiposity tends to increase with age because of the strong influence age has on metabolism and body composition. In estimating the propensity score, we use a set of age dummy variables (five-year intervals) rather than continuous age to allow for nonlinear effects.
The female weight gap is decomposed using three measures of adiposity (average BMI, average WHR, and proportion obese), two estimation samples (total sample and sample with plausible intake values), and two decomposition procedures (BO and DFL), with each combination presented in Table 2. A comparison across combinations provides an indication of which factor is most important in explaining the gap. Unfortunately, it is not possible to confidently declare that one combination is superior to all others because there is no consensus in the literature on what are the best measures of adiposity and EI. However, our preferred combinations are those using the DFL decomposition, chiefly because it is semiparametric, and those using plausible energy intake values, because substantial measurement error in self-reported intake is well established.
Decompositions of average BMI are presented in the top panel of Table 2, decompositions of average WHR are presented in the middle panel, and decompositions of proportion obese are presented in the bottom panel. The first row in each panel shows the total difference in the adiposity measure between black and white women for each combination of estimation sample and decomposition procedure. These values differ across columns within each panel because the BO procedure decomposes mean BMI and WHR, and the DLF procedure decomposes median BMI and WHR. Total difference values will also differ because the DFL procedure restricts the sample to those that are in the region of common support.14 The additional rows in each panel show the portion of the total difference that can be attributed to differences in EI, EE, age, and unobserved factors. We emphasize that differences in adiposity between blacks and whites can be caused only by systematic differences in intake, expenditure, and biology/genetics. This relationship might be viewed as almost being an accounting identity. Of course, individual-level characteristics such as education, marital status, and region of residence can affect obesity, but these variables affect adiposity only indirectly through their effects on intake and expenditure decisions. Adding these variables to the decomposition would likely reduce the size of the unobservable category because of measurement error, but we deliberately choose not to do so in order to focus on the role that intake and expenditure play in explaining the obesity gap.
A comparison across rows in each panel shows that the most important observed factor is EI. In each of the 12 decompositions, EI is estimated to explain more of the weight gap than both EE and age. For example, using our preferred specification of plausible intake values and DFL decomposition procedure, differences in EI are estimated to explain approximately 48% of the gap in average BMI, 44% of the gap in average WHR, and 38% of the gap in proportion obese. In comparison, differences in energy expenditure are estimated to explain approximately 13% of the gap in average BMI, 16% of the gap in average WHR, and 11% of the gap in proportion obese. However, none of the EE effects are statistically different from zero. Age differences are estimated to explain little of the difference in adiposity, although what little it does explain is, in fact, negative. Negative estimates indicate that the younger average ages of black women work to make the observed weight gap smaller rather than larger. Figure 4 graphically displays these results.
Table 2 also shows that unobservables play an important role in explaining the weight gap. Large unobservable effects are not uncommon in decomposition analyses. For example, unobservables are the largest factor in the DFL analyses of Hyslop and Mare (2005), Daly and Valletta (2006), and Cobb-Clark and Hildebrand (2006). To a large extent, this is unavoidable if some factors have a negative impact on the gap and other factors have a positive impact on the gap, as we observe in our analysis; however, it is still important to consider possible explanations for the unobservable effects. One likely explanation is measurement error. The unobservables category captures any differences in adiposity, diet, and exercise not captured by observed differences. Therefore if, for example, BMI poorly represents adiposity, then the unobservable effect will be large. Another explanation is the genetic differences between white and black women. Genetic differences in basal metabolic rates could create weight differences even if no differences exist in EI and EE. In a review of 15 studies, Gannon et al. (2000) provided support for this explanation. They concluded that black women living in the United States may be particularly vulnerable to obesity because of relatively lower EE for their metabolic size. We lack measured data on BMRs, so we cannot examine this issue in detail; however, it is unlikely that genetic differences are the main cause of the female weight gap given that no gap exists for men.
A comparison across columns in each panel shows that the contributions of EI and EE to the weight gap are generally larger in the DFL procedure than in the BO procedure. For example, between columns 1 and 3, the proportions explained by differences in EI increase from 6.1% to 25.5% for average BMI, 13.3% to 26.4% for average WHR, and 9.8% to 25.5% for proportion obese. The equivalent numbers for differences in EE are 3.3% to 8.2%, 4.0% to 9.5%, and 3.0% to 7.8%. One simple explanation for the different results is that the BO procedure compares mean adiposity whereas the DFL procedure compares median adiposity. Given the dramatic differences in the shape of the black and white adiposity distributions (e.g., mean BMI gap equals 3.56, and median BMI gap equals 4.33), this could substantially impact the estimated effects. Another explanation is that the BO procedure relies on a simple linear approximation of the relationship between adiposity and intake, expenditure, and age. Although one can add higher-order terms to attempt to capture the curvature of the true function, the effects will be over- or underestimated in different segments of the adiposity distribution unless the chosen nonlinear specification is correct.15 In contrast, the semiparametric DFL approach avoids the issue of making functional form assumptions of inputs, expenditure, and age by the use of a reweighting scheme. A third difference between the procedures is that the DFL procedure imposes a common support requirement, ensuring that any combination of characteristics observed for blacks can also be observed among whites. This explains why the sample sizes differ for BO decomposition (N = 1,768) and DFL decomposition (N = 1,748). Finally, even with well-specified regression models and similarly shaped distributions, the BO and DFL decomposition procedures are not numerically equivalent (Dinardo 2002).
Another clear difference between columns is that the proportion of the weight gap explained by EI is much higher for the sample that omits implausible intake values, with most of the increase coming at the expense of unobserved factors. For example, the percentage of average BMI differences explained by EI increases from 25.5% to 47.9%, and the percentage explained by unobservables decreases from 71.3% to 38.7%. The pattern supports our hypothesis that the unobserved factors component largely reflects the existence of measurement error.
An advantage of the DFL procedure is that it allows decompositions at any point in the distributions, not only at the mean. In Table 3, we exploit this advantage and decompose the weight gap in the 25th and 75th percentiles of the BMI and WHR distributions.16 Results from these decompositions are consistent with the results in Table 2. In general, the effects of EI differences are again much larger than the effects of EE differences, although few of the differences are statistically significant because of smaller sample sizes as we move away from the median.
Interestingly, our results contrast starkly with those contained in Burke and Heiland (2008). They estimated regression models containing variables representing exercise, intake, and smoking history, and found that these variables account for 0.25 BMI units or 6.4% of the gap in mean BMI. In addition, they concluded from their regression results that “black women’s lower exercise levels figure most prominently among the behavioral factors that contribute to their having higher BMI” (p. 21). The differences in the proportion explained and the prominence of energy expenditure are likely a consequence of the different measures used for physical activity. Instead of using ActiGraph accelerometer data for measuring physical activity, Burke and Heiland (2008) constructed a discrete variable with three categories based on responses to yes-or-no questions concerning whether the individual participated in a given type of activity for at least 10 minutes during the previous month. In addition, it may also be because they used an OLS regression approach and implausible intake values. As shown in column 1 of the first panel, the observed factors explain little of the difference in average BMI; however, even in this specification, intake explains more than expenditure.
A number of factors may be contributing to black women’s higher EI. One candidate is that predominantly black neighborhoods are different from predominantly white neighborhoods. For example, Baker et al. (2006) found black-white differences in the availability of healthy food retailers: black areas, regardless of income, were less likely than predominantly white higher-income communities to have access to foods that enable individuals to make healthy choices. Cultural factors possibly also contribute to higher intake. For example, a number of studies have found black-white differences in conceptions of ideal body size, with black women tending to select larger images of ideal size (Fitzgibbon et al. 2000; Flynn and Fitzgibbon 1998; Lovejoy 2001). These studies have also found black-white differences in attitudes toward dieting and exercise, with black women tending to be more satisfied with their own size and less likely to be attempting to lose weight.
Dramatic differences exist in body weight and obesity prevalence between black and white women in the United States, and crucially, these differences are a significant contributor to black-white inequalities in health. In this article, we provide one of the few empirical investigations of the female weight gap by assessing the relative culpability of energy intake and energy expenditure. Uncovering whether EI or EE dominates is important because before effective policy can be enacted to reduce the weight gap, we must ascertain whether unhealthy eating practices or sedentary lifestyles are mostly responsible. Furthermore, understanding which factor dominates helps to highlight the social and environmental differences that are indirectly at fault.
Our empirical approach is based on decomposing black-white differences in BMI and WHR into four components: energy intake, energy expenditure, age, and unobservables. Importantly, we use objectively measured EE data based on physical activity monitors, and self-reported food intake adjusted for measurement error. The results from our decomposition analyses suggest that increased EI is the driving force behind the weight gap.
Few existing studies consider the roles of EI and EE in explaining black-white BMI differences; however, a large literature investigates their roles in explaining the considerable growth in BMI over time. The relative culpability of “gluttony versus sloth” is in fact the subject of some dispute, with a number of studies suggesting that physical inactivity is the major cause (Heini and Weinsier 1997; Philipson 2001; Prentice and Jebb 1995; Weinsier et al. 1998) and a number suggesting that overconsumption is largely to blame (Bleich et al. 2008; McCrory et al. 2002; Cutler et al. 2003; Nielsen and Popkin 2003). Although we don’t investigate changes in obesity rates over time, the results from this article provide some new evidence to this ongoing debate.
The main implication of our results is that policies aimed at helping black women to reduce their food intake may be more successful in reducing the weight gap than policies aimed at increasing physical activity, although improving physical activity levels must clearly be part of the solution. Common policy suggestions include improving access to healthy food options, using pricing strategies to promote the purchase of healthy foods, and increasing media promotion of healthy eating (Bleich et al. 2008; French et al. 2001). The specific targeting of black female eating practices, however, likely requires more specialized policies. For such policies to be developed, more research is needed to improve our understanding of the factors driving the black-white differences in energy intake.
We thank Richard Burkhauser and Deborah Cobb-Clark for valuable comments and discussions. We are also grateful to two anonymous referees for useful suggestions that helped improve the paper and also to participants at the 2009 Australian Health Economics Conference for comments. All errors are our own.
The Centers for Disease Control (CDC) and the World Health Organization (WHO) define “overweight” and “obesity” as a body mass index (BMI) value of more than 25 and more than 30, respectively, where BMI is the ratio of weight measured in kilograms, to squared height measured in meters.
The BO decomposition also involves a reweighting but of a very different nature. In a basic BO decomposition, group differences in the independent variables are weighted by regression coefficients in order to determine the part that is explained by the independent variables. When one uses coefficients from the pooled model over both groups as the reference coefficients, the pooled coefficients can be expressed as , where W is a weighting matrix given by W = diag(b – bB) diag(bW – bB)–1; and b, bB, and bW denote the coefficients from a pooled model, the white model, and the black model, respectively.
This presentation draws on Cobb-Clark and Hildebrand (2006).
To account for unequal probabilities of selection resulting from the complex survey design in NHANES, relevant sampling weights provided by NHANES were also used in our estimation.
For example, a female aged 30–39 would be considered to be reporting plausible caloric intake information if she reports consuming daily calories that are between 1.18 to 2.18 times her BMR.
The formula is .
This estimate is based on the calorie counter on the Mayo Clinic website (http://www.mayoclinic.com/health/calorie-calculator/NU00598) assuming an inactive lifestyle. It is essentially based on a BMR equation with an allowance for different levels of physical activity.
NHANES questionnaires and a comprehensive listing of laboratory and examination components are posted on the NHANES website (http://www.cdc.gov/nchs/nhanes.htm).
Because NHANES collects nationally representative data that use actual height and weight measurements, rather than self-reported data, it is often referred to as the “gold standard” for studies that focus on obesity in the United States.
SAS code is available from the NIH for analyzing ActiGraph 7164 Physical Activity Monitor (PAM) data from the 2003–2004 NHANES. These programs are written to import and analyze accelerometer data downloaded from the National Center for Health Statistics. For this article, we adapt this code to similarly analyze PAM data from the 2005–2006 NHANES.
The ActiGraph AM-7164 device was programmed to detect and record the magnitude of acceleration or “intensity” of movement; acceleration data in NHANES were stored in memory according to a one-minute interval. The intensity count is the intensity value recorded by the device, and each minute has an intensity value. The intensity files were reviewed for outliers and unreasonable values. The criteria used for reasonable ranges of activity count data were based on published literature and expert judgment.
The averages and propensities reported in Table 1 for the analysis sample with nonmissing values of EI and EE variables are very similar to those for the entire sample. This suggests that our analysis sample does not constitute an unusual group of respondents.
Common support is imposed in our analysis by dropping all observations whose propensity score is smaller than the minimum and larger than the maximum in the opposite group. Implementing the common support condition ensures that any combination of characteristics observed for blacks can also be observed among whites, and that we are not comparing the incomparable.
We also performed the BO procedure using an equation that included interaction terms in the intake category and expenditure category and found that the results were very similar to the results presented in Table 2.
It is common to also decompose differences at the extremes of the distribution—for example, at the 10th and 90th percentiles. We do not follow this approach here because of a low density of observations at the extremes of the distribution, which will result in estimates that are too imprecise.