### The 2006 and 2007 American Time Use Surveys

Data for the current investigation come from the 2006 and 2007 public-use files of American Time Use Surveys (ATUS) and have the advantage of providing valid, reliable measures of time spent in both energy intake and energy expenditure related activities over one 24-hour period [26, 27]. The extraordinary level of detail in the ATUS allows us to separate time spent eating into time spent eating where eating is the respondent's primary focus and secondary eating time (i.e., time when the respondent's primary activity was something other than eating, but when eating was still taking place).

ATUS respondents are drawn from households that had completed their final interview for the Current Population Survey in the preceding 2-5 months. Each respondent is randomly selected from among each household's members, age 15 and older. Half complete a diary for a weekday and half complete a diary for a weekend day.

Information from the ATUS interviews is linked to information from the 2006 and 2007 Eating and Health module interviews [28, 29] so that we also have data on the respondent's height and weight. BMI is calculated from self-reported weight in kilograms divided by self-reported height in meters squared. It should be noted that although self-reported BMI has been commonly used in past studies [30–34], some have found that it results in a modest under-estimation of overweight and/or obesity rates [35–37] while others have found it to be a valid and reliable way to measure BMI for nonelderly adults [38].

We restrict our ATUS sample to those respondents who are between the ages of 25 and 64, inclusive. Younger respondents are excluded so as to avoid the inclusion of individuals whose eating habits may be dictated by their parents. Respondents over age 64 are excluded because these individuals are more likely to have health conditions that may affect some aspects of their time use. We also restrict our sample to those respondents whose BMI ranges from 16.0 to 60.0, inclusive. These BMI restrictions lead to the elimination of 5 male respondents (1 with BMI > 60.0 and 4 with BMIs < 16.0) and 17 female respondents (5 with BMIs >60.0 and 12 with BMIs < 16.0). In addition, we eliminate 12 respondents who report spending more than 15 hours being physically active, 18 respondents who report spending more than 20 hours sleeping and 4 respondents who report spending more than 20 hours watching television. These restrictions are made to reduce the potential influence of leverage points and outliers. Finally, we exclude women who are pregnant as their reported BMIs are likely not reflective of their usual BMIs. These sample restrictions result in a sample of 8,856 women and 7,586 men in our study.

We focus on seven time-use categories that are potentially related to energy balance. The first category measures the amount of primary time the respondent spends eating and drinking (i.e., time where eating and drinking has her/his primary attention).^{d} Secondary eating time is captured by the amount of time the respondent reports eating as a secondary activity (i.e., time where something else has her/his primary attention). Secondary time spent drinking anything other than plain water is measured separately. Other food related activities are measured by the time spent in food preparation and clean-up excluding related travel time.

Physical activity cannot be adequately measured by simply summing the time respondents report spending in exercise and sports as we would end up omitting things like bicycling to work, chasing after a toddler, and doing physically demanding household chores. Thus, rather than use only time spent in the ATUS sports and exercise categories, we sum time spent in all activities in the ATUS activity lexicon that generate metabolic equivalents (METs) of 3.3 or more. We select these activities based on the work done by Tudor-Locke et al. [39] who have linked the ATUS time use lexicon to the Compendium of Physical Activities. We choose a threshold of 3.3 METs because this captures activities such as exterior house cleaning, lawn and garden work, caring for and helping household children, playing sports with household children, active transportation time (i.e., walking or biking), as well as most forms of sports, exercise, and recreation. It excludes such routine household activities such as interior housekeeping and playing with children in non-sports.^{e} The compendium also identifies time spent in certain occupations (i.e., building and grounds cleaning and maintenance, farming, construction and extraction) as generating a minimum of 3.3 METs. To control for occupational physical activity requirements, we include a dummy variable in the male equation that takes on a value of "1" if the respondent works in one of these occupational categories. Only a handful of female respondents report working in these fields and thus we exclude this dummy from the female regressions. We sum only spells of 10 minutes or more of physical activity time because prior work has established 10 minutes as the minimum duration necessary to impact an individual's energy balance [40].

Finally, we use two measures of inactivity: television/video viewing time and time spent sleeping. These two measures have been associated with BMI and/or obesity risk in previous studies that have related single categories of time use to BMI [8, 9, 11–14].

### Analysis Approach

To examine the relationship between time use and BMI, ideally one would have longitudinal data on time use in various activities. Unfortunately, longitudinal time diary data do not exist. While some surveys do gather information on typical time use, methodological research has shown such questions provide less valid and reliable measures when compared to diary data [26, 27, 41].

Conceptually, cross-sectional time diary data of the type available in the ATUS have two disadvantages. First, time spent in various activities on any given day may deviate from an individual's usual time use patterns. As such, there is measurement error in the independent time use variables that likely bias the coefficient estimates toward zero [42]. Second, any observed association between time use and BMI obtained using cross-sectional data may reflect reverse causality. For example, having a high BMI may lead one to spend less time being physically active. To address both data shortcomings, we adopt a model of time use where BMI and time use are simultaneously determined.

In our model, BMI is a function of time use, biological traits (e.g., age, gender, race/ethnicity, health status) and socio-demographic characteristics (e.g., marital status, number of children, employment status, and education). Decisions about how much time to spend in various activities is a function of household roles (e.g., self-identification as the primary meal preparer, self-identification as the primary grocery shopper), structural factors (e.g., number of children in the home, marital status, employment status, gender, race/ethnicity, age, weekend or weekday diary, season of the year, rural residence, region of residence), prices (e.g., the respondent's wage rate, grocery prices), and income.

Data on wage rates in the ATUS are limited to those individuals who report both hours of work and earnings. To avoid the possibility of selection bias that could be introduced by excluding those who are not employed, we elect to use predicted hourly opportunity costs of time generated from wage regressions estimated using the corresponding years of the March Supplement to the Current Population Survey (CPS). We use individuals age 25-64 in the March Supplement to estimate wage equations that correct for sample selection bias using the techniques developed by James Heckman [43]. Equations are estimated separately for women and men using the appropriate CPS weights. Coefficients from these equations are used to generate predicted hourly opportunity cost of time for each individual in our ATUS sample. A random error is added to each predicted wage based on a mean of zero and a variance that is equal to the variance of the estimating equation.^{f} Estimates of offered wage rates provide approximate opportunity cost estimates of the value of time for employed individuals and lower-bound estimates of the value of time for non-employed individuals [43].

The ATUS contains a categorical measure of annual household income. The categorical nature of this variable coupled with item-specific non-response made it less than ideal to use on our analyses. Consequently, we again turn to the March Supplement to the CPS. For individuals age 25-64, we estimate a regression using the appropriate CPS weights where total, annual nonwage income for the household is the dependent variable. Coefficients from this equation are then used to generate predicted nonwage income values for our sample of respondents in the ATUS. A random error is added to each predicted nonwage income value based on a mean of zero and a variance that is equal to the variance of the estimating equation.^{g}

Grocery price information comes from the Council for Community and Economic Research's (C2ER) state-based cost of living index for 2006 and 2007. C2ER provides expenditure weighted, quarterly metropolitan and micropolitan price information [44].^{h} The only detailed geographic information contained in the ATUS is the respondent's state of residence and residential urbanicity. Thus, our linkage of grocery price information is done based on information about the respondent's state of residence, urban/rural status, and the quarter in which the respondent was interviewed. In those rare cases where the respondent was located in a micro area within a state that had no micro grocery price index, we use the state-wide average. Initially, we also included an index measuring non-grocery prices but this was dropped from our analyses once it was determined that the simple correlation between the grocery price index and the non-grocery price index was .89.

We estimate three different sets of equations separately for men and women. In the first formulation, we estimate a model where our time use measures are treated as predetermined variables that affect BMI. We then estimate an instrumental variables model that recognizes that the time use and BMI causality may run in both directions when one is analyzing cross-sectional data of the sort used here. In the final formulation, we estimate reduced form models of BMI. In this formulation, BMI is estimated as a function of the biological and socio-demographic variables and the strictly exogenous factors that are posited to affect time use [45]. Essentially, these latter two estimation approaches both incorporate the hypothesis that time use and BMI are simultaneously determined.

Key to identifying the preferred model is undertaking tests for endogeneity and then, if endogeneity is confirmed, identifying "instruments" that are correlated to time use but unrelated to the error term in the BMI equation [45]. We test for endogeneity by estimating the Durbin-Wu-Hausman F-statistic [46]. Strength of the instruments is assessed by calculating a variation on the squared partial correlation between the instruments excluded from the second stage and the endogenous regressors [47]. Independence of the instruments from the error term in the BMI equation is assessed by calculating Hansen's J statistic [46].

The instrumental variables used to identify the system in our application are self-identification as the primary meal preparer, self-identification as the primary grocery shopper, whether the diary day was a weekend, whether the diary day was in the summer, whether the diary day came from 2007, the grocery price index, the hourly opportunity cost of time, and the household's annual nonwage income. The instrumental variables approach involves first estimating the time use equations and using the coefficients from these equations to generate predicted time use values for all respondents in the sample. These predicted values are then included as regressors in the BMI equations. If all of the necessary conditions are met, the estimated coefficients using this approach are purged of possible reverse causation. This approach has the added advantage of also addressing the typical time use measurement issue since the predicted values may be thought of as approximating usual time spent in the various activities.

Separate equations are estimated for women and men to allow for the possibility that there are biological factors related to gender that interact with time use and are associated with BMI. All analyses are weighted using the appropriate ATUS weights. The ATUS weights compensate for the survey's oversampling of certain demographic groups, the oversampling of weekend day diaries, and differential response rates across demographic groups [48]. Estimation is done using Stata 11.0 and SAS 9.2.