Alcohol and Body Weight in United States Adults DAVID F. WILLIAMSON, MS, PHD, MICHELE R. FORMAN, PHD, NANCY J. BINKIN, MD, MPH, EILEEN M. GENTRY, PATRICK L. REMINGTON, MD, MPH, AND FREDERICK L. TROWBRIDGE, MD, MPH Abstract: Alcohol contributes more than 10 per cent of the total caloric intake of adult drinkers in the United States. However, the effect of alcohol on body weight has not been adequately studied in the general population. The association between weight and frequency of alcohol consumption was examined in two national crosssectional surveys: the Second National Health and Nutrition Examination Survey (HANESII; n = 10,929) and the Behavioral Risk Factor Surveys (BRFS; n = 18,388). Linear multiple regression was used to estimate the independent effect of alcohol on weight, adjusting for smoking, age, diet practices, physical activity, race, education, and height. Among men, alcohol had only a slight effect on weight in either survey. However, among women, alcohol was

Introduction It is estimated that alcohol accounts for over 10 per cent of the total caloric intake of adult drinkers in the United States.' Although there has been considerable examination of the relationship between smoking and body weight,2" the association between alcohol and body weight in the general population has been less closely studied. Studies of alcoholics and of healthy volunteers indicate that high daily intakes of alcohol are associated with loss of body weight."2 Population-based studies that have examined this association have been equivocal, although some have suggested that drinkers have lower body weights than nondrinkers.3"- 17 Smokers are more likely to drink alcohol than nonsmokers,'3'18 9 hence alcohol may confound the association between smoking and lower body weight. It is also possible that alcohol may modify the weight-lowering effect of smoking. However, neither confounding, nor effect modification by alcohol has been closely studied in reference to the effects of smoking on body weight. In this paper, we examine the association between alcohol and body weight using two independent sets of nationally representative cross-sectional data, the Second National Health and Nutrition Examination Survey and the Behavioral Risk Factor Surveys.

Methods Second National Health and Nutrition Examination Survey (HANESII) The HANESII was a cross-sectional household survey

conducted between February 1976 and February 1980, with a total sample of 27,801 persons residing in the United States. Because the sample was selected to be representative of the civilian, noninstitutionalized population ages six months through 74 years, the obtained estimates are in general very reliable and valid. These data were also weighted to ensure that they represented the 1978 US population. Details of the survey sample design are presented elsewhere. From the Division of Nutrition, Centers for Disease Control (all authors) and the Department of International Health, Johns Hopkins School of Hygiene and Public Health (Dr. Forman). Address reprint requests to David F. Williamson, PhD, 1600 Clifton Road, NE, Bldg. 3-SB45, Centers for Disease Control, DHHS, USPHS, Atlanta, GA 30333. This paper, submitted to the Journal March 6, 1987, was revised and accepted for publication May 20, 1987.

© 1987 American Journal of Public Health 0090-0036/87$1.50 1 324

associated with a substantial reduction in weight, which was as large as the effect of smoking. Compared with nondrinkers, women who consumed alcohol 7-13 times per week had the greatest reduction in weight: -3.6 kg (95% confidence limits [CL] = -5.6, -1.5 kg) in HANESII and -3.2 kg (95% CL = -4.9, -1.5 kg) in BRFS. Alcohol confounded the association between smoking and weight, and among women it accounted for nearly 45 per cent of the weight-lowering effect of smoking. Alcohol also diminished the weight-lowering effect of smoking in men, while in women the smoking effect was slightly enhanced. Further studies are needed to understand the causal mechanisms by which alcohol is associated with body weight. (Am J Public Health 1987; 77:1324-1330.)

We examined HANESII data for the subsample of the 12,520 adults whose age at examination was 18-74 years (5,922 men and 6,598 women) and who received a dietary, medical history, and health history interview, as well as a physical examination. Pregnant women (N = 113) and persons whose records had unknown values for any analytic covariates (N = 1,478) were excluded. The remaining sample included 5,265 men and 5,664 women. Behavioral Risk Factor Surveys (BRFS)

Between April 1981 and October 1983, 29 states carried out cross-sectional, random-digit telephone surveys among adults 18 years of age and older. These surveys used a core questionnaire and sampling methodology developed in cooperation with the Centers for Disease Control. A supplemental survey was conducted in the remaining states (except Hawaii) in order to obtain a sample representative of civilian, noninstitutionalized adults living on the US mainland. The combined data from the state surveys and supplemental survey were weighted to represent the 1982 US adult population. Because of greater sampling variability, national estimates from BRFS are less reliable than comparable estimates from HANESII; however, this survey was included in the analysis in order to provide independent confirmation of results. Details of the survey design and the statistical methodology used in aggregating the individual surveys have been previously described.23-26 In the BRFS, the total sample was 22,236 persons (9,495 men and 12,741 women). For our analysis we excluded records from those states in which information on race (N = 1,483) and physical activity was not collected (N = 1,527). Records of pregnant women (N = 320), as well as records of those with extreme outlying weights and heights (2 300 per cent of median weight-for-height) or with unknown values for any analytic covariates were also excluded (N = 518). The remaining sample had 8,110 men and 10,278 women. Variables Used in the Analysis

Body weight (kg) is the dependent variable in this study. In HANESII, body weight in light examination clothing was measured to the nearest 0.25 pounds and converted to the nearest 0.01 kg. In BRFS, weight in light clothing without shoes was self-reported to the nearest pound and later converted to kilograms. Information on current smoking practices was collected in both surveys using three similar questions. Current smokAJPH October 1987, Vol. 77, No. 10

ALCOHOL AND BODY WEIGHT ers were defined as those reporting that they had smoked at least 100 cigarettes during their lifetime and continued to smoke. Former smokers were defined as those who reported smoking 100 or more cigarettes but who no longer smoked. Never-smokers were defined as those who reported smoking fewer than 100 cigarettes in their lifetime. Current smokers were also asked the average number of cigarettes smoked per day. In our analyses, smoking is categorized as: none, 1-15, 16-25, and 26 or more cigarettes per day. Information on current drinking practices was recorded in a different manner in the two surveys. The HANESII respondents were asked to report the number of times they drank and to indicate whether this occurred less than once a week. They were also asked how often they drank on a weekly or daily basis for beer, wine, and liquor separately. Nondrinkers were defined as those who reported they never consumed alcoholic beverages. For those reporting consumption of more than one type of alcoholic beverage, the drinking frequencies were summed across the three alcohol types. In order to sum across different frequency categories (e.g., less than once a week for wine + daily for beer + once a week for liquor) the categories were also expressed on a daily basis. A frequency of less than once a week was arbitrarily assumed to be once a month and assigned a value of 0.03 times per day (1/30); frequencies reported on a weekly basis were divided by seven. According to the average number of times alcohol was normally consumed, HANESII respondents were classified into: none, less than once per week, 1-6 times per week, 7-13 times per week, and 14 or more times per week. BRFS "nondrinkers" were defined as those who reported that on average they consumed alcohol less than once a month during the previous 12 months, 69 per cent of whom had not drunk alcohol within the previous year. Current drinkers were asked to estimate, on either a weekly or monthly basis, the average number of times they drank alcoholic beverages, and the average number of drinks they consumed when they drank. According to the average number of drinks usually consumed, BRFS respondents were classified into one of five categories: none, less than I drink per week, 1-6 drinks per week, 7-13 drinks per week, and 14 or more drinks per week. Information on consumption of beer, wine, and liquor was not recorded separately. In HANESII, two questions were asked about physical activity. The first question asked for a self-reported estimate of the usual amount of physical exercise from recreational activities as indicated by: much, moderate, or little to none. The second question asked the respondent for a daily estimate of nonrecreational physical activity as indicated by: very active, moderately active, or quite inactive. In the BRFS, five questions were asked about physical activity: frequency of vigorous exercise, its duration, primary form of exercise, level of physical effort on the job, and frequency of light recreational activities (e.g., bowling, gardening). A value of 0 corresponded to someone who reported having a sedentary occupation, less than one hour per month of vigorous physical activity, and less than three times per week of light recreational activity. The highest value (8) corresponded to someone reporting a physically active occupation, four or more hours per week of vigorous exercise, and more than four times per week of light recreational activity. This ordinal variable was inversely related to

Results

In HANESII and BRFS analyses, the variables height, age, race, education, and weight-loss diet status were ex-

Among men in HANESII, 40.7 per cent (SEM = 0.9%) smoked and 75.8 per cent (SEM = 1.1%) drank; smokers were 1.1 times more likely to drink than never-smokers (82%

body weight.

AJPH October 1987, Vol. 77, No. 10

pressed identically. Total daily caloric intake (KCAL), which was not available in BRFS, was included as a covariate in the HANESII analyses and was estimated from a single 24-hour dietary recall. This estimate of total caloric intake included calories from alcohol. Data Analysis To estimate the independent effects of alcohol and smoking on body weight, we used analysis of covariance in a sex-specific linear multiple regression framework. The regression coefficients for each level of drinking or smoking represent the mean difference in body weight between those reporting a level of drinking or smoking and the referent group (never-smokers who do not drink). The joint effects of alcohol and smoking on body weight were estimated by including 12 smoking-by-drinking interaction terms (3 levels of smoking x 4 levels of drinking), together with their main effects, in the regression model. The multivariate analyses included a set of covariates, notably age, height, race, education, weight-loss diet status, physical activity, and total caloric intake (HANESII only). These variables were chosen a priori because they were recognized as potential confounding factors in the association of body weight with the exposures, alcohol and smoking. The categories used for each of these covariates are listed in the Appendix. Among the continuous variables, height and dietary energy intake were nearly linear with body weight, but age had a curvilinear relationship; therefore age was included together with its quadratic term (age2). As a result of the complex sample design in HANESII and BRFS, statistical analysis based on assumptions of simple random sampling may lead to an underestimation of the variance of the estimated effects. Therefore, all reported multiple regression analyses used SURREGR, a specialized regression program for complex sample survey data. All prevalence estimates and estimates of univariate means were analyzed using SESUDAAN, which also computes variances by taking into account the sampling design and sampling weights 27.28 The distribution of the dependent variable, body weight, was slightly skewed to the right. Therefore, all analyses reported here was replicated using the log transformation of body weight. Analyses were also replicated using body mass index (weight/height2) as the dependent variable. In both cases, the findings were unchanged. In this paper the results are presented in terms of body weight to aid in the interpretation of findings. In both surveys the data that were excluded before creation of the analytic samples were examined for evidence of potential bias. For those excluded cases in which weight, height, and age were known, the crude associations with smoking and drinking status were very similar to those reported for the analytic sample. It is doubtful that any systematic bias occurred because of these exclusions. Because of the generally large sample sizes reported in this study, the importance of the estimated associations are based on the magnitude and direction of effects, and their consistency between surveys. Standard errors and 95 per cent confidence limits (CL) are given in order to judge the precision of the reported estimates.

1 325

WILLIAMSON, ET AL. TABLE 1-Sex-specific Crude Estimates of Mean Body Weight, Height, and Age by Smoking and Drinking Status, in the HANESII and BRFS Surveys Men

HANESII Current smokers Former smokers Never smokers Drinkers Nondrinkers BRFS Current smokers Former smokers Never-smokers Drinkers Nondrinkers

Women

N

Weight (kg)

Height (cm)

2038 1721 1506 3699 1566

76.3 [0.3] 79.8 [0.4] 78.6 [0.4] 77.9 [0.2] 78.4 [0.5]

175.6 [0.2] 175.0 [0.2] 175.7 [0.3] 175.8 [0.2] 174.3 [0.2]

2790 2209 3111 5573 2537

78.5 [0.5] 81.8 [0.7] * 78.6 [0.5] 80.0 [0.4] 78.4 [0.5]

-

-

177.9 177.8 177.5 178.5 176.3

[0.2] [0.3] [0.3] [0.2] [0.4]

Age

N

Weight (kg)

[0.3] [0.5] [0.6] [0.3] [0.6]

1723 809 3132 2838 2826

64.1 [0.4] 67.2 [0.7] 66.1 [0.3] 63.8 [0.3] 67.9 [0.4]

162.3 162.4 161.1 162.3 161.0

[0.2] [0.3] [0.2] [0.2] [0.2]

38.5 [0.4] 44.5 [0.6] 43.1 [0.5] 39.0 [0.4] 45.6 [0.5]

39.6 [0.71 47.7 [0.8] 37.9 [0.7] 38.2 [0.5] 47.8 [0.7]

2949 1634 5695 4904 5374

61.8 [0.4] 63.3 [0.6] 63.6 [0.3] 61.4 [0.3] 64.5 [0.4]

163.8 163.7 162.8 163.9 162.5

[0.3] [0.3] [0.2] [0.2] [0.2]

39.5 [0.6] 46.2 [0.9] 45.6 [0.5] 39.2 [0.5] 48.5 [0.6]

38.9 47.5 37.2 39.1 46.5

Height (cm)

Age

*Standard error of mean (SEM) in brackets.

versus 71%). In the BRFS, 33.7 per cent (SEM = 1.1%) smoked and 68.4 per cent (SEM = 1.1%) drank; smokers were 1.2 times more likely to drink than never-smokers (75% versus 63%). Among women in HANESII, 33.5 per cent (SEM = 0.7%) smoked, and 56.9 per cent (SEM = 1.5%) drank; smokers were 1.4 times more likely to drink than neversmokers (70% versus 50%). In the BRFS, 29.7 per cent (SEM = 0.9%) smoked and 50.2 per cent (SEM = 1.0%) drank; smokers were 1.5 times more likely to drink than neversmokers (62% versus 40%). Table 1 gives the sample sizes and weighted estimates of mean body weight, height, and age of men and women according to their self-reported drinking and smoking status in the HANESII and BRFS surveys. The mean height of HANESII respondents was consistently below that of those in BRFS regardless of sex, and the measured body weight of women in HANESII was above the self-reported weight of women in BRFS. These differences may reflect biases known to be associated with self-reported height and weight.29 Among men in both the HANESII and BRFS surveys, former smokers were heavier than current smokers and never-smokers. In HANESII, the mean body weight of current smokers was below that of never-smokers, while in BRFS the mean weight of current smokers was similar to that of never-smokers. In HANESII, men who drank weighed less than nondrinkers; in BRFS the converse was true. Among women, the surveys showed a more consistent pattern. In both surveys, current smokers weighed less than never- and former smokers, while drinkers weighed less than nondrinkers. Nondrinkers had the heaviest mean body weights in both surveys. Height did not appear to be accounting for the observed differences in body weight between nondrinkers and the other groups; however, in both surveys female nondrinkers were older. The alcohol and smoking categories in Table 1 are not mutually exclusive. To determine if alcohol was confounding the association between smoking and body weight, and to estimate the independent effects of alcohol and smoking on body weight, two regression models were analyzed. The first model included only the smoking variables, together with the covariates listed in the Appendix. (In order that only nondrinkers who never smoked were in the referent group, the first model also included a variable which indicated whether or not the never-smokers drank. This variable was deleted from the second model.) The second regression 1 326

model included the drinking variables in addition to the variables in the first model. By comparing the size of the regression coefficients for the smoking variables between the two models, we estimated the degree to which adjustment for alcohol changes the association between body weight and smoking. At the same time the second model gave an estimate of the independent effects of smoking and alcohol on body weight, adjusted for each other, and for the remaining covariates. These analyses were limited to current and never-smokers. Figure 1 shows the sex-specific associations between body weight and smoking, before and after adjusting for alcohol consumption, in the HANESII and BRFS surveys. In HANESII men, adjustment for alcohol reduced the estimated effect of smoking on body weight by 24 per cent for those smoking 1-15 cigarettes per day, 18 per cent for those smoking 16-25 cigarettes per day, and 31 per cent for those smoking more than 25 cigarettes per day. The weighted average of the adjustment effects across the three smoking levels among men in HANESII was 23 per cent. In the BRFS, adjustment for alcohol did not change the estimated mean effect of smoking on body weight in men. In women the result of adjustment for alcohol was stronger and more consistent than in men. Adjustment for alcohol reduced the estimated effect of smoking on body weight by 52 per cent, 38 per cent, and 43 per cent for the three smoking levels in HANESII, and by 43 per cent, 32 per cent and 59 per cent in BRFS. The weighted average of the three adjustment effects in women was 45 per cent and 42 per cent in the HANESII and BRFS surveys, respectively. Table 2 gives the estimated independent effects on body weight of alcohol and smoking from the second regression model. In both surveys there was only a slight independent effect of alcohol on body weight in men. The analysis for men in HANESII was repeated without the total caloric intake variable and the results remained nearly unchanged. This indicates that there is little collinearity between total caloric intake and drinking frequency in men in this study. Among women, both the HANESII and BRFS results indicated that there was a substantial independent association between alcohol and lower body weight, and the size of the alcohol effect was comparable to that of smoking. In both surveys there was a trend among women to increasingly lower body weight with increasing consumption of alcohol for the first three levels of drinking, with an attenuated effect of alcohol at the highest level of drinking. This was especially AJPH October 1987, Vol. 77, No. 10

ALCOHOL AND BODY WEIGHT MEN

TABLE 2-Sex-specific Linear Multiple Regression Estimates of the Independent Effects of Smoking and Alcohol on Body Weight* Men

0

ev-

HANESII

-

1-15 16-25 26+ Drinking (times)