Introduction In the context of the growing obesity epidemic [1], it has been suggested that increases in caloric availability and thus energy intake [2], irrespective of changes in physical activity, are enough to explain the observed increases in weight at the population level [3]. Theories about the causes of change in energy intake are numerous, but tend to focus on one of three areas: increases in the frequency of eating/drinking occasions (EOs) [4]–[6], especially snacking [7]; increases in the typical portion sizes (PSs) of foods and beverages [8]–[11]; or changes in the energy density (ED) of the foods consumed (termed “volumetrics” by Rolls and colleagues) [12]–[14]. Much of this research focuses on the effect of the ED or PSs of individual foods, or preload conditions [15],[16], on energy intake at a given meal The limited epidemiological work has been cross-sectional [17]. At least one study [18] also examined whether increasing PS has an effect beyond a single EO, reporting that increased PSs over 2 d resulted in increased energy intake and that the increase on day one was not compensated for on the second day [18]. Several studies have also confirmed that when individuals consume meals that are lower in ED, their daily energy intake is also lower [19]–[22]. Taken together these findings suggest that it is the total meal (combination of foods and beverages consumed at a given EO), not just individual foods consumed, that is important in determining total energy (TE) intake and should be the focus of research. A small body of research has examined the combined effect of changes in both the ED and PS of foods with respect to energy intake [23]–[25], but to our knowledge similar research does not exist for the other possible combinations of ED, EO frequency, and PS, nor have these factors been examined all together in either large-scale epidemiological studies or clinical trials. Further, the research on PSs has focused mainly on separate foods and beverages (e.g., sugar-sweetened beverages or cheeseburgers), ignoring both their potential effect on each meal or snack occasion and the relationship of overall PSs of all other meals and snacks to daily totals. To address this knowledge gap, in the present study we examine the relative contribution of changes in the frequency of EOs, PS for each EO, and ED for each EO to changes in TE intake using nationally representative samples of US adults between 1977 and 2006. Methods Study Population Cross-sectional nationally representative dietary intake data of adults 19 y and older were taken from four US food surveys. United States Department of Agriculture (USDA) data came from respondents of the Nationwide Food Consumption Survey (NFCS) of 1977–78 (n = 17,464) and the Continuing Survey of Food Intakes of Individuals (CSFII) 1989–91 (n = 8,340) and 1994–96, 1998 (CSFII 1994–98, n = 9,460). We also combined two consecutive National Health and Nutrition Examination Survey (NHANES) surveys, 2003–04 and 2005–06, into a single analytic sample (NHANES 2003–06, n = 9,490). The USDA and NHANES surveys are based on stratified area probability samples of non-institutionalized US households in the 48 contiguous [26] or all 50 states [27]. Detailed information about each survey and its sampling design has been previously published [26]–[29], and a comparison of the sampling and 24-h recall intake methodologies can be found in Table S1. The study was approved by the Institutional Review Board at the University of North Carolina at Chapel Hill. Dietary Data The NFCS 1977–78 and CSFII 1989–91 surveys collected dietary intake data over three consecutive days using a single-interviewer-administered multiple-pass 24-h dietary recall followed by a self-administered 2-d diet record using methods developed by the USDA. Dietary data from these surveys consisted of all foods consumed at and away from home (24-h recall) and a comprehensive record of all foods eaten on the day of the interview and the following day (2-d record). This USDA dietary methodology was later integrated into the CSFII 1994–98 and NHANES 2003–06 surveys, which utilized two nonconsecutive days of interviewer-administered 24-h dietary recalls (3–10 d apart). In order to maintain consistency across studies we utilized the first day of available 24-h recall dietary data. We excluded all reported instances of water being consumed as a separate food item from all surveys, as this information was not collected in the same manner across exams. In the later NHANES exams, water (as a beverage) was probed for specifically, which resulted in a dramatic increase in the reported instances of water consumption. Food and Beverage Definitions Foods and beverages were defined and grouped according the UNC-CH food-grouping system [30]. Briefly, foods and beverages were grouped into 101 nutrient-based food groups (including 16 beverage groups) according to fat and fiber content. Because “dish” identifiers are not available in NHANES, it is not possible to accurately and confidently link foods consumed separately but which might constitute a single dish, e.g., milk and cereal consumed at the same EO. In instances such as this, cereal is identified as a food, while milk is identified as a beverage. Although it seems possible to make educated guesses about foods like milk and cereal, there are many more assumptions required to assign a single “food” status to something like milk consumed with macaroni and cheese. Therefore, in all cases where any beverage was consumed in the same EO as a food, the beverage is considered independent of the food. The UNC-CH food-grouping system has been used previously in studies of beverage [2],[31] and dietary intake [32],[33] specifically, as well as studies examining snacking [7],[34] and overall eating behaviors [35],[36]. Defining Eating Occasions EOs, either meals or snacks, were self-defined by respondents in both the USDA and NHANES surveys. Meals were defined by the respondent as breakfast/brunch, lunch, and dinner/supper, while the snack category included those EOs defined by the respondent as “snack,” plus related snacking occasions (i.e., food and/or coffee/beverage breaks). All occasions that were identified as snacks but were consumed within 15 min of each other were combined into a single snacking event. Also, some people defined foods eaten at the same time as both a snack and a meal. As an example, suppose an individual reported consuming a sandwich and a bag of chips (eaten at the same time). This individual identified the sandwich as lunch and the chips as a snack. In instances where this occurred, both items were considered eaten as part of a single EO (lunch), rather than as two separate EOs (lunch and a snack). Beverages consumed alone, and not identified as a meal, were considered snacks. The number of meals and snacks was then summed for each individual for a total number of EOs. All foods were assigned to a specific EO in 1977–78 and in 2003–06, while in the other two surveys 0.15% of food items were not assigned because they had neither an EO name nor a time associated with the EO. This method of assigning meals and snacks has been previously employed to study overall eating [37] and snacking behavior [7] in a sample of US adults. Total Energy, Portion Size, and Energy Density We calculated per EO measures for energy intake, PS, and ED. For each individual, the daily total gram weight (PS) and total daily energy of all foods consumed were summed over a 24-h period and divided by the total number of EOs as a measure of per EO PS and per EO energy. ED was then generated by dividing energy (per EO) by PS (per EO). This was done for foods and beverages separately. Decomposition Algorithm Mathematical decomposition has been applied to many measures of changes in health and behavior (e.g., mortality and fertility rates [38]–[40]). We define total daily energy intake (TE) (kcal/d) as the number of daily EOs multiplied by the average PS (grams) per EO multiplied by the average ED (kcal/g) of each EO, as in the following equation: (1) Using this equation, we then estimate the proportionate contribution (a partial derivative) of changes in each of these components to overall changes in total daily energy intake by taking the derivative of changes in TE with respect to changes in PS, ED, and the number of EOs, holding the other two factors constant at their mean [38]. Briefly, for each component of total daily energy, the change between two time points (e.g., 1977–78 and 1989–91) is multiplied by the average of the other two components. This derivative is calculated for each component (PS, ED, and EOs) and the values summed to generate the full derivative for change in TE intake, as shown in the following equation: (2) To annualize change, these values were then divided by the number of years between each wave of data collection (i.e., results comparing 1977–78 to 1989–91 were divided by 12.5 [mean year points: 1990–1977.5 = 12.5]). The resulting output is interpreted as the annual change in energy (kcal/d/y) that is attributed to changes in PS, ED, and EOs, with sign indicating the direction of change. Statistical Analysis All analyses were conducted using Stata 11 (StataCorp). We used survey commands to account for survey design: weighting and clustering. All values were adjusted to the 1977–78 age–gender–race/ethnicity sample distribution and are reported as mean (or percent) and standard error. Values were then annualized to account for the unequal spacing within and between exam years. To test for statistical differences in sample characteristics (not PS, ED, EO, or TE) comparing all years to each other, we used independent two-sided t tests with p≤0.05 set for statistical significance using the Bonferroni correction for multiple comparisons. Results Overall The sample population in 1977–78 was significantly younger and had a higher percentage of non-Hispanic white males with 12 or fewer years of education compared to the later exam years. The population in 1977–78 also had a lower percentage of individuals of Hispanic and “non-Hispanic other” race/ethnicity and a lower percentage of persons living at or above the 350% poverty income ratio (Table 1). 10.1371/journal.pmed.1001050.t001 Table 1 Characteristics of study populations across exam years. Sample Characteristic Subcategory Exam Periods 1977–78 1989–91 1994–98 2003–06 Sample size 17,228 10,501 9,338 9,018 Age (y) 44±0.26 45±0.46 45±0.36 46±0.48a Female (%) 41±0.52 53±0.66a 52±0.63a 52±0.44a Race/ethnicity (%) Non-Hispanic white 83±1.33 80±1.16 76±1.84a 72±2.15a , b Non-Hispanic black 11±1.06 11±0.62 11±1.05 11±1.31 Hispanic 5±0.66 7±1.07 9±1.48a 11±1.34a , b Non-Hispanic other 1±0.15 2±0.59a 4±0.44a , b 5±0.43a , b Education (%) <12 y 32±1.06 19±0.68a 16±0.82a , b 17±0.94a 12 y/GED 32±0.68 35±0.58a 35±1.07a 26±0.72a , b , c 13–15 y (
