点击显示 收起
1 From the Jean Mayer US Department of Agriculture Human Nutrition Research Center on Aging at Tufts University, Boston (PKN and KLT), and the National Institute on Aging, National Institutes of Health, Baltimore (DM, JH, and RA)
2 Supported in part by the US Department of Agriculture, Agricultural Research Service, contract number 53-3K06-01; the National Institutes of Health, National Institute on Aging Intramural Program; and the General Mills Bell Institute of Health and Nutrition. 3 Reprints not available. Address correspondence to PK Newby, Jean Mayer USDA Human Nutrition Research Center on Aging at Tufts University, 711 Washington Street, 9th Floor, Boston, MA 02111. E-mail: pknewby{at}post.harvard.edu.
ABSTRACT
Background: Sixty-five percent of US adults are overweight, and 31% of these adults are obese. Obesity results from weight gains over time; however, dietary determinants of weight gain remain controversial.
Objective: Our objective was to examine whether food patterns derived from exploratory factor analysis are related to anthropometric changes. We hypothesized that we would derive a healthy food pattern and that it would predict smaller changes in body mass index (BMI; in kg/m2) and waist circumference (in cm) than would other food patterns in models adjusted for baseline anthropometric measures.
Design: The subjects were 459 healthy men and women participating in the Baltimore Longitudinal Study of Aging. Diet was assessed by using 7-d dietary records, from which 40 food groups were formed and entered into a factor analysis.
Results: Six food patterns were derived. Factor 1 (reduced-fat dairy products, fruit, and fiber) was most strongly associated with fiber (r = 0.39) and loaded heavily on reduced-fat dairy products, cereal, and fruit and loaded moderately on fruit juice, nonwhite bread, nuts and seeds, whole grains, and beans and legumes. In a multivariate-adjusted model in which the highest and lowest quintiles were compared, factor 1 was inversely associated with annual change in BMI (ß = –0.51; 95% CI: –0.82, –0.20; P < 0.05; P for trend < 0.01) in women and inversely associated with annual change in waist circumference (ß = –1.06 cm; 95% CI: –1.88, –0.24 cm; P < 0.05; P for trend = 0.04) in both sexes.
Conclusion: Our results suggest that a pattern rich in reduced-fat dairy products and high-fiber foods may lead to smaller gains in BMI in women and smaller gains in waist circumference in both women and men.
Key Words: Food patterns • dietary patterns • factor analysis • principal components • obesity • body composition • diet assessment • body mass index • BMI • waist circumference
INTRODUCTION
Sixty-five percent of US adults are currently overweight [body mass index (BMI; in kg/m2) 25], and 31% of these adults are obese (BMI 30) (1). Obesity results from weight gains over time; however, dietary determinants of weight gain remain controversial (2, 3). Inconsistent findings may be explained in part by limitations in the single-nutrient approach, which is commonly used in nutritional epidemiologic research (4, 5). Cluster and factor analyses have emerged as methods to empirically derive dietary (or food) patterns and possibly explain nutrient-disease relations.
In our previous study, we found that patterns derived by using cluster analysis were significantly related to longitudinal changes in anthropometric measures (6). Specifically, the healthy dietary pattern—which was high in fruit, reduced-fat dairy products, high-fiber cereal, and whole grains and low in red and processed meats, fast food, and soda—showed smaller gains in both BMI and waist circumference than did the other dietary patterns (white bread, sweets, alcohol, and meat and potatoes).
In the current study, we continued to examine prospectively whether eating patterns are related to anthropometric changes in women and men participating in the Baltimore Longitudinal Study of Aging (BLSA) by using a distinctly different patterning method—factor analysis. Specifically, we used principal components to extract factors, in which a given set of variables is transformed into a reduced set of correlated variables (7). Factor analysis is statistically quite different from cluster analysis. Whereas cluster analysis separates persons into mutually exclusive groups based on differences in mean food intakes (clusters), factor analysis separates foods into groups based on correlations between foods (factors), and persons receive a factor score for each of the derived factors. With the use of the factor analysis procedure, a person’s dietary pattern would be best represented by looking at his or her factor scores for each of the derived vectors. Because factor analysis studies focus on individual food vectors when examining relations rather than the total diet for a given person, we refer in this article to the derived factors as food patterns and not as dietary patterns.
Other studies that have used factor or principal components analysis have empirically derived a "healthy" food pattern that is high in fruit, vegetables, seafood, fiber, and other "healthy" foods (8-11). None of these studies, nor other studies using factor analysis (12-14), examined whether food patterns derived by using factor analysis are related prospectively to anthropometric changes. Our objective was to examine whether food patterns derived from exploratory factor analysis, specifically, principal components analysis, are related to changes in BMI and waist circumference. We hypothesized that, using factor analysis, we would derive a healthy food pattern and that this pattern would lead to smaller changes in BMI and waist circumference than would other dietary patterns.
SUBJECTS AND METHODS
Study population
The BLSA was initiated in 1958 to study the physical, mental, and emotional effects of aging in healthy, active people; the original study design and data collection were described in detail elsewhere (15). Briefly, initial study participants were white, male, community-dwelling volunteers aged 22–88 y and living in the Baltimore area. The study protocol expanded in 1978 to include women and minorities. Enrollment in the BLSA is open and participants return approximately every 24 mo (mean time interval: 25 mo) for repeated measurements of height, weight, body composition, diet, and a variety of other physiologic, psychological, and behavioral measures. The Institutional Review Boards of the Johns Hopkins Bayview Medical Research Center and the Gerontology Center approved the BLSA protocol, and all subjects gave written informed consent for their participation.
This analysis uses the same data set as described in our previous paper (6) to facilitate comparison of the factor and cluster analyses in deriving food and dietary patterns, respectively. In summary, our study population was limited to those entering the study on or after 1980 (n = 921) to avoid bias from changing dietary trends over time, and persons aged <30 y and >80 y were excluded (n = 138). Of the remaining 783 subjects, persons who had not completed 4 d of dietary records were omitted (n = 9), as were subjects whose food group intake appeared implausible (>6 SDs from the mean for each food group) (n = 63). An additional 134 subjects were omitted because they did not have 2 measures of height or weight (at the time the dietary record was collected and at follow-up). Finally, all subjects who received a diagnosis of cancer, diabetes, stroke, or heart disease either before or at baseline were omitted (n = 190) to create a disease-free cohort. After all exclusions, 459 subjects were available for the analysis. An additional 10 subjects were omitted from the waist circumference analysis because they did not have baseline or follow-up information on waist circumference; thus, data for 449 subjects were used in this analysis.
Dietary assessment
Dietary intake was assessed with the use of 7-d dietary records; reports detailing dietary collection methods and dietary intake in the BLSA population were published previously (16-18). In summary, study participants were instructed by trained dietitians to complete 7-d food records. Food records were completed at home by the participants and were sent back to the study center. Before 1992, the subjects were given food models and a booklet of food pictures to help them assess portion size. In 1994, the subjects were given a portable scale to weigh food portions. The participants were contacted by telephone with any questions about the diet records.
Dietary records from 1984 to 1991 were originally coded and entered into a nutrient database maintained by the BLSA, whereas diet records completed since 1994 were coded and entered into the Minnesota Nutrition Data System (Nutrition Coordinating Center, University of Minnesota, Minneapolis) at Tufts University; no dietary data were collected in 1992 and 1993. Dietary data from 1984 to 1991 were then reentered into the Minnesota Nutrition Data System, and nutrient intakes were back-adjusted to correct for changes in the food supply (eg, changes in nutrient content resulting from cereal fortification) by using data from the US Department of Agriculture to correspond to appropriate time intervals (19, 20).
Food pattern derivation
To perform a food pattern analysis with the use of principal components analysis, individual foods and food ingredients from the dietary records were first aggregated into groups. We formed 40 food groups, mainly according to macronutrient composition (eg, fat or fiber content) and culinary use; several foods (eg, pizza and eggs) composed their own groups (Appendix A). Where possible, foods were separated into full- and reduced-fat groups (eg, high-fat dairy and reduced-fat dairy products).
View this table:
APPENDIX A. The 40 food groups used in the factor analysis
Food groups may be entered into the principal components analysis as absolute weight in grams, number of servings, or the percentage of energy intake from each food group (21). As before (6), we chose to consider the food-group variables in terms of the percentage of energy contributed by each of the 40 groups to derive factors that were in proportion to total energy intake. Daily food intakes were calculated for each subject at baseline from the average intake of each food (group) over the 7-d dietary record. The percentage contribution from each food group for each subject was then entered into the principal components analysis.
The PROC FACTOR procedure in SAS (version 8.2; SAS Institute, Cary, NC) was used to perform the analysis. The procedure uses principal components analysis and orthogonal rotation (varimax option in SAS) to derive noncorrelated factors. As with other empirically derived pattern procedures, principal components analysis requires preselection of the number of factors. To decide which number of factors to retain, we ran factor solutions ranging from 2 to 15, including food-group components with an eigenvalue > 1 and examined both the scree plots and the factors themselves to see which set of factors most meaningfully described distinct food patterns. From these analyses, a 6-factor solution was selected. Factor loadings were calculated for each food group across the 6 factors (food patterns). A factor score was then calculated for each subject for each of the 6 factors, in which the standardized intakes of each of the 40 food groups were weighted by their factor loadings and summed (7); the sums were then standardized (
Anthropometric and covariate assessment
Anthropometric measurements were made following standardized procedures (23) and were fully described elsewhere (24). In summary, weight and height were measured for each subject at each visit, from which the BMI was calculated. Waist circumference was measured at each visit with an inelastic tape at the narrowest part of the torso at the end of expiration (24).
Demographic data were collected from each study participant at the first visit and were used to adjust for potential confounding in regression analyses. Race-ethnicity, physical activity, smoking habit, education, and vitamin supplement use were determined by questionnaire at the time that dietary records were collected. Physical activity was measured by an adapted version of the Harvard Alumni questionnaire, which asked participants about all daily activities (eg, activities at home, work, and during recreation or sports). The amount of time spent for each activity was summed across all activities to determine the daily energy output per body weight (kJ/kg) and was described previously (16, 25).
Statistical analyses
Factors were divided into quintiles, and sample means and frequencies were calculated at the baseline visit at which time dietary data were collected. The percentage of subjects who were overweight (BMI = 25–29.99) or obese (BMI 30) was calculated by using recommended international cutoffs (26). Partial correlation coefficients (adjusted for age, sex, and energy intake) were calculated between each food pattern vector and selected macronutrients to further describe the patterns. Mean nutrient and food group intakes were calculated for each quintile of factor 1 (reduced-fat dairy products, fruit, and fiber) because it was the dominant food pattern vector.
Separate regression analyses were performed for each factor to test whether food patterns predicted changes in BMI and waist circumference. Factors were treated categorically (quintiles), in which the lowest quintile was the reference group; a test for trend was also performed by using a categorical variable that ranked the median factor score for each quintile. Annual change in BMI was calculated by subtracting BMI at visit 1 from BMI at visit 2, dividing by the time interval between visits (months), and then multiplying by 12 mo. Annual change in waist circumference was calculated by using the same algorithm. We tested the a priori hypotheses that the healthy food pattern we derived would predict smaller changes in BMI and waist circumference than would other food patterns in models adjusted for baseline anthropometric measures (baseline BMI and baseline waist circumference), age, sex, and sociodemographic covariates. Because total energy intake is the mechanism through which nutrients and foods may affect BMI, it is arguable whether energy should be included in the model; we fit a final model for each analysis in which we added total energy to the multivariate model to see whether adjustment for baseline energy would change our estimates. We also added terms to adjust for development of disease (stroke, cancer, diabetes, and heart disease) during the follow-up period. Because age may not be linearly related to BMI, we fit regression models using quadratic terms for age. We tested for interactions between food pattern and sex by using a cross-product term.
In secondary analyses, we performed 2 tests to compare the factor solution discussed in this study with the cluster solution reported in our previous study (6). In the first test, we calculated the mean factor scores for each cluster by using the same 40 food groups used in the current study. In the second test, we used Pitman’s test to compare the SD of the residuals from the each of the regression analyses to examine which patterning method better predicted change in BMI and waist circumference. All analyses were performed by using the Statistical Analysis System (SAS) for WINDOWS (version 8.2; SAS Institute, Cary, NC).
RESULTS
Factor loadings for the 6 food patterns and the name we assigned to each pattern are presented in Table 1. Individual factor loadings are interpreted similarly to correlation coefficients, in which the most positive values contribute most to the factor score, and the most negative values contribute least to the factor score. We derived a healthy, fiber-rich food pattern (factor 1), which loaded heavily on reduced-fat dairy products, fruit, and fiber and loaded moderately on fruit juice, nonwhite bread, nuts and seeds, whole grains, and beans and legumes. Factor 1 was the most dominant food pattern in the population and explained 5.8% of the variance in intake, whereas each of the remaining 5 factors explained between 4.8% (factor 2) and 3.9% (factor 6) of the variance. Together, the 6 factors explained 27% of the variance in intake (data not shown). For simplicity, we do not present results for factors 4–6 for the remainder of the analyses because these patterns were not dominant and are not as interpretable as are the first 3 patterns.
View this table:
TABLE 1. Factor loadings for 6 food patterns at baseline for 459 adults participating in the Baltimore Longitudinal Study of Aging1
Sample characteristics at baseline are presented for the lowest and highest quintiles of each food pattern (Table 2). Subjects in the highest quintile of factor 1 had the lowest BMI (23.9 ± 0.4) and the smallest waist circumference (81.1 ± 0.9 cm), whereas subjects in the lowest quintile of factor 1 had the highest BMI (25.9 ± 0.4) and the largest waist circumference (86.3 ± 0.9 cm). The highest percentage of women (59.8%) and vitamin users (58.7%) were seen in the highest quintile of factor 1.
View this table:
TABLE 2. Sample characteristics for the lowest (Q1) and highest (Q5) quintiles of 3 food patterns identified at baseline for 459 adults participating in the Baltimore Longitudinal Study of Aging1
Factor 1 and factor 3 (sweets) were both strongly correlated with carbohydrate (r = 0.57 and r = 0.46, respectively); however, factor 1 was more strongly associated with dietary fiber (r = 0.39), whereas factor 3 was more strongly associated with sucrose (r = 0.40) (Table 3). Factor 2 (protein and alcohol) was moderately correlated with protein (r = 0.43) and alcohol (r = 0.36). Factor 5 (fatty meats) was directly associated with fat (r = 0.32) and saturated fat (r = 0.35) (data not shown for factor 5). An analysis of factor 1 by quintiles indicated that the percentage of energy intake from carbohydrate was 44.6% in quintile 1 and 55.2% in quintile 5, whereas the percentage of energy intake from fat was 37.2% in quintile 1 and 30.5% in quintile 5 (Table 4). Fiber intake increased >10 g from quintile 1 to quintile 5. Food groups with high positive factor loadings were highest in quintile 5 (eg, reduced-fat dairy products, cereal, and fruit), whereas food groups with high negative factor loadings were highest in quintile 1 (eg, white bread and refined grains, meat, and processed meat).
View this table:
TABLE 3. Partial Pearson correlation coefficients (r) between each of 3 food patterns and daily energy and nutrient intakes at baseline for 459 adults participating in the Baltimore Longitudinal Study of Aging1
View this table:
TABLE 4. Daily energy, nutrient, and food group intakes across quintiles (Q1–Q5) of factor 1 at baseline for 459 adults participating in the Baltimore Longitudinal Study of Aging1
For all subjects, the mean annual change in BMI was 0.11, and the mean annual change in waist circumference was 0.84 cm. The smallest mean annual change in BMI was seen in the highest quintile of factor 1 for both women (–0.12 ± 0.09; P < 0.05) and men (0.02 ± 0.08; P > 0.05), where P > 0.05 denotes that changes in BMI that were not significantly different from 0. Mean annual change in waist circumference was smallest in the highest quintile of factor 1 (0.18 ± 0.28 cm; P > 0.05) for both sexes (data not shown).
Regression results associating change in BMI across quintiles for all 6 factors are presented in Table 5. A test for interaction between food pattern and sex indicated that the effect of food pattern on change in BMI was significantly modified by sex for factor 1 (P = 0.008), but not for other factors (P > 0.05); therefore, the results are presented separately for women and men for factor 1 only. For women, the highest quintile of factor 1 was inversely associated with annual change in BMI in a comparison with the lowest quintile (ß = –0.51; 95% CI: –0.82, –0.20; P < 0.05) and the test for trend was significant (P < 0.01) in a multivariate-adjusted model. Factor 1 was inversely but not significantly related to annual change in BMI in men in a comparison of the fifth with the first quintile (ß = –0.14; 95% CI: –0.36, 0.08; P > 0.05; P for trend = 0.47). Factor 2 was directly associated with change in BMI in a comparison of the fifth with the first quintile (ß = 0.20; 95% CI: 0.04, 0.36; P < 0.05), and the test for trend was marginally significant (P = 0.05). Factors 4–6 were not significantly related to annual change in BMI (data not shown).
View this table:
TABLE 5. Linear regression analysis ß coefficients (and SEs) for quintiles 1–5 (Q1–Q5) of 3 food patterns for predicting annual change in BMI in 459 adults participating in the Baltimore Longitudinal Study of Aging1
The results of the regression analysis for waist circumference are shown in Table 6; no significant interactions were seen between food pattern and sex. Factor 1 was inversely associated with annual change in waist circumference in a comparison of the highest with the lowest quintile (ß = –1.06 cm; 95% CI: –1.88, –0.24; P < 0.05), and the test for trend was significant (P = 0.04) in a multivariate-adjusted model. Quintile 5 of factor 3 was directly and significantly related to annual change in waist circumference compared with quintile 1 (P for trend = 0.04). Quintile 4 of factor 4 (vegetable fats and vegetables) and of factor 6 (eggs, bread, and soup) were directly and significantly related to annual change in waist circumference compared with the lowest quintile (data not shown for factors 4–6). Estimates for both the BMI and waist circumference analyses were similar when total energy was included in the model and when the models were adjusted for the development of disease during the follow-up period. Inclusion of quadratic terms for age in the model did not improve model fit.
View this table:
TABLE 6. Linear regression analysis ß coefficients (and SEs) for quintiles (Q1–Q5) of 3 food patterns for predicting annual change in waist circumference in 459 adults participating in the Baltimore Longitudinal Study of Aging1
When we compared factor scores across the 5 clusters, we found that the highest mean score for factor 1 was seen for those subjects in the healthy cluster (
DISCUSSION
We derived 6 food patterns using principal components analysis in a sample of healthy adults. Our hypothesized "healthy pattern," or factor 1 (reduced-fat dairy products, fruit, and fiber), explained the greatest amount of variance and had high positive loadings for reduced fat dairy products, cereal, fruit, fruit juice, nonwhite bread, nuts and seeds, whole grains, and beans and legumes and high negative loadings for white bread and refined grains, processed meat, potatoes, meat, and non-diet soda. Factor 1 was inversely related to annual change in BMI in women, but not in men, and was inversely associated with waist circumference in both sexes. Our findings were strengthened by the use of a prospective design, the use of dietary records (which are the "gold standard"), and the adjustment for many potential confounders in our regression models.
The healthy pattern we observed is most similar to the cold-foods pattern derived in the study by Maskarinec et al (14), which had high factor loadings for cold breakfast cereals, fruit juice, and fruit. It also shares common elements with the cereals patterns derived in several other studies (21, 27, 28). Factor 1 is similar to the "healthy" cluster derived in our previous study, whereas the protein and alcohol and sweets factors are less similar to the alcohol and sweets clusters, respectively. Although these similarities point to some comparability in the methods, the extent to which factors and clusters yield different solutions is attributable to methodologic differences.
Our healthy pattern is less similar to the "prudent" pattern described in other studies (10, 29, 30), which is high in vegetables, fruit, legumes, and fish. Also, we did not observe a separate "Western" pattern, which is high in processed meat, refined grains, sweets and desserts, and French fries (10, 29). Rather, we derived a pattern, factor 1, that had high positive factor loadings for healthy foods and high negative factor loadings for less healthy foods. In our study, then, the "Western" pattern previously described (10, 29) appears embedded in factor 1, ie, persons with higher scores for factor 1 were consuming a healthy food pattern, whereas those with lower scores for factor 1 were consuming a Western food pattern.
Different effects of food patterns on BMI by sex were previously reported (10, 14, 27, 29, 31, 32), and an interaction between food patterns and sex was also observed in relation to coronary heart disease (33). In this study, of those in the top quintile of factor 1, 59% were vitamin users and 87% were never smokers, which suggests that a healthy diet is part of a healthy lifestyle (34). Although we adjusted for these factors in our analysis, there may be other potential confounders associated with dietary intake and change in BMI that may be different in men and women, for which we were not able to adjust. Our finding of an interaction between a fiber-rich food pattern and sex may reflect a true association or may be due to lack of power, lack of variation in factor 1 scores for men, different overlap in factor scores between men and women, uncontrolled confounding, measurement error, or some other reason. More research is needed to elucidate whether inconsistent findings between women and men are due to methodologic limitations or whether the biological effects of diet are truly modified by sex.
We are not aware of any studies that have examined the relation between food patterns derived by using factor analysis and waist circumference. The inverse relation seen here between factor 1 and anthropometric changes may be explained by the roles of fiber (35), glycemic index (36), and energy density (37-39) on food intake and appetite, as previously discussed (6); the roles of dairy products and calcium; or some othre reason. For example, the dietary fiber intake was >10 g greater in the highest quintile of factor 1 than in the lowest quintile of factor 1, with similar intakes in quintiles 2–4, which possibly explained why only the highest quintile was associated with changes in waist circumference and BMI. Understanding the reasons for the associations between the other food patterns we derived and anthropometric changes, however, is more difficult. Factor 2 was most strongly correlated with alcohol, and factor 3 was most strongly correlated with sucrose; these nutrients may help to explain the effects we observed. Because a person’s total dietary pattern is represented by his or her scores on all of the derived factors, an examination of individual factor relations may obfuscate diet-disease relations within a person. Perhaps, because of this limitation, investigators have created mutually exclusive groups of persons based on their factor scores across several food patterns [eg, the healthiest pattern combined those in the highest quintile of the prudent pattern and the lowest quintile of the Western pattern (29). This approach, however, is not practical for a 6-factor solution. We therefore focused our analysis and discussion on the hypothesized healthy food pattern.
Our derived food patterns explained almost 30% of the variance in dietary intake, which is less than that observed in some studies (10, 29, 30) but similar in magnitude to that observed in other studies (22, 27). The amount of variance explained by individual factors may be expected to vary across studies, because it is a function of both the number of variables included in the analysis (7) and the correlation matrix itself (28). When we compared the factor solution from this study with the cluster solution from our previous study (6), no significant difference between solutions was observed in predicting change in BMI, but the factor solution presented here better predicted change in waist circumference. More research is needed to further examine whether significant differences in patterning solutions are consistently observed in explaining health outcomes.
Our study has several limitations. Our sample has limited variability in ethnicity, and education and food patterns differed according to these variables (4, 40). However, it is not clear whether the patterns we derived would have a different biologic effect in persons with different educational levels or ethnicities. Food patterns may also differ by sex (4, 22, 41, 42). Women tend to have healthier diets than do men (43, 44), which may explain why 60% of those in the top quintile of factor 1, our healthy pattern in this study, were women. In addition, we were unable to adjust for the role of genetics in our models, which may modify the relation between diet and adiposity (45).
The principal components method itself also has limitations that stem from several subjective decisions that investigators must make. First, one must decide how to treat the dietary input variables and whether these variables should be adjusted for total energy intake, because these decisions will affect the correlations between the variables and, hence, the factor solution. In this study, we chose to consider the dietary variables as a percentage of energy to obtain factors proportional to total energy intake. However, some reports have adjusted for intakes by using the nutrient residual approach (46) and others have used total grams (47) or servings (8, 10) unadjusted for total energy; other studies have transformed food group variables logarithmically to improve normality (14). Research is needed to examine how the treatment of dietary variables affects the factor solution, which may lead to insight on what is the best method for handling variables in nutritional studies using pattern analysis.
Another limitation of factor analysis is that the investigator is forced to prespecify the number of factors. Although eigenvalues, scree plots, and interpretability (7) are used to guide the investigator in determining the best factor solution, ultimately such a decision is subjective. The number of factors extracted by other studies has varied, ranging from 2 (48, 49) to 10 (50). The aforementioned reasons for the increasing complexity of analyzing >2 factors may explain why investigators using factor analysis are inclined to derive or report less, rather than more, food patterns. Our choice to retain 6 factors was based on eigenvalues, scree plots, interpretability, and similarity in range to other studies, in addition to our a priori belief that there are likely >2 meaningful food patterns in nature.
In summary, we derived a healthy food pattern and observed prospective associations with this pattern and BMI in women and with waist circumference in both sexes. Food patterns differ between populations, including by age (50), religion (51), and ethnicity (49), but the prudent, Western, and cereal-based patterns as well as the sweets and alcohol patterns seem to be reasonably reproducible among populations. Our results suggest that a fiber-rich food pattern that is high in reduced-fat dairy products, cereal, fruit, fruit juice, nonwhite bread, nuts and seed, whole grains, and beans and legumes may lead to smaller gains in BMI in women and smaller gains in waist circumference in both women and men.
ACKNOWLEDGMENTS
We gratefully acknowledge Peter Bakun and Janice Maras for assistance with data processing and preparation of dietary variables and Ning Qiao for statistical programming assistance.
RA and JH contributed to the original design and to data collection for the BLSA. DM prepared the nondietary data for this analysis. PKN was responsible for the design and analysis of the study and drafted the manuscript. KLT contributed to the analysis and design of the study and oversaw the Tufts collaboration with the BLSA. All authors made critical comments during the preparation of the manuscript and fully accept responsibility for the work. No author had a financial interest or professional or personal affiliations that compromised the scientific integrity of this work.
REFERENCES