点击显示 收起
1 From the Food and Nutrition Database Research Center, Department of Food Science and Human Nutrition, Michigan State University, East Lansing, MI (JMK, EJY, and WOS), and the Flint Public Schools, Flint, MI (LB).
2 Supported in part by unrestricted research grants from the Kellogg Company, the Egg Nutrition Center, and the Michigan Agricultural Experimental Station. 3 Address reprint requests to WO Song, 129 GM Trout FSHN Building, Food and Nutrition Database Research Center, Department of Food Science and Human Nutrition, Michigan State University, East Lansing, MI 48824. E-mail: song{at}msu.edu.
ABSTRACT
Background: Certain nutrients are well established as dietary risk factors for cardiovascular disease (CVD), but dietary patterns may be a better predictor of CVD risk.
Objective: This study tested the hypothesis that the complex dietary behaviors of US adults can be grouped into major dietary patterns that are related to risk factors for CVD.
Design: With the use of food-frequency questionnaire data from the third National Health and Nutrition Examination Survey, dietary patterns of healthy US adults (=" BORDER="0"> 20 y old; n = 13 130) were identified by factor analysis. Log-transformed biomarker data were associated with major dietary patterns after control for confounding variables in regression analyses. All statistical analyses accounted for the survey design and sample weights.
Results: Of 6 dietary patterns identified, 2 patterns emerged as the most predominant: the Western pattern was characterized by high intakes of processed meats, eggs, red meats, and high-fat dairy products, and the American-healthy pattern was characterized by high intakes of green, leafy vegetables; salad dressings; tomatoes; other vegetables (eg, peppers, green beans, corn, and peas); cruciferous vegetables; and tea. The Western pattern was associated (P < 0.05) positively with serum C-peptide, serum insulin, and glycated hemoglobin and inversely with red blood cell folate concentrations after adjustment for confounding variables. The American-healthy pattern had no linear relation with any of the biomarkers examined.
Conclusions: The identification of common dietary patterns among free-living persons is promising for characterizing high-risk groups at the US population level. The dietary patterns identified here are similar to those reported in other nonrepresentative samples and are associated with biomarkers of CVD risk, which confirms that dietary pattern analysis can be a valuable method for assessing dietary intakes when predicting CVD risk.
Key Words: Dietary patterns • biomarker • cardiovascular disease • NHANES III • food-frequency questionnaire • factor analysis
INTRODUCTION
Cardiovascular disease (CVD) is a major public health problem in the United States, and dietary intake is thought to exert a great influence on the risk of CVD. The known risk factors for CVD, however, including specific dietary components (eg, excess saturated fat), explain only a portion of the morbidity and mortality due to CVD. Because dietary intake is such a complex exposure variable, it is now recognized that we must develop and refine methods of assessing dietary intake that focus on the total diet and not just on individual dietary components (eg, nutrients). As a result, the study of patterns of intakes of nutrients, foods, and food groups has begun to emerge in nutrition research.
Assessment of the total diet takes into account all nutrient interactions and allows us to capture diet-disease or dietbiomarker relations without knowing the specific nutrient or food component involved (1). Because many nutrients are highly correlated within foods, it is difficult to examine their effects separately. Moreover, nutrient intakes may be confounded by entire dietary patterns (2). For example, ß-carotene supplementation failed to prevent chronic disease in prospective trials conducted after it was observed that the consumption of fruit and vegetables (ie, ß-carotene–containing foods) was associated with a lower incidence of disease (3–5). This association has been partially attributed to a combination of the other beneficial components of diets high in fruit and vegetables, such as folate, fiber, magnesium, potassium, flavonoids, and plant sterols (6).
Two large prospective studies reported that dietary patterns defined by factor analysis of dietary data collected by using a food-frequency questionnaire (FFQ) are associated with various sociodemographic characteristics, coronary heart disease (CHD) risk (nonfatal myocardial infarction and fatal CHD), and biomarkers of CVD risk (7–9). Findings from these studies, however, are limited to the men and women who participated in the Health Professionals Follow-up Study and the Nurses’ Health Study. Subjects were mostly white, > 40 y old, and of higher socioeconomic status, and thus we do not know if and how dietary patterns differ in the overall US adult population inclusive of all sociodemographic strata. Furthermore, even if dietary patterns of the US adult population were identifiable, we would not know whether they were associated with CVD risk, because the risk of CVD differs by age, sex, and race.
Whereas dietary patterns have been identified and associated with CVD risk in specific subpopulations, they should also be identified in nationally representative populations for better understanding of the effect of diet on CVD risk at the population level. Therefore, the purpose of this study was to test the hypotheses that 1) dietary intake data for the healthy US population can be systematically classified into distinct dietary patterns and 2) risk factors for CVD are associated with specific dietary patterns.
SUBJECTS AND METHODS
Data set
Subjects in this study were participants in the third National Health and Nutrition Examination Survey (NHANES III), conducted in 1988–1994 (10). The National Center for Health Statistics conducted the survey to obtain nationally representative information on the health and nutritional status of the US population. The NHANES III sample represents the total noninstitutionalized civilian US population aged =" BORDER="0"> 2 mo. In NHANES III, 39 695 persons were originally sampled over 6 y. Of those, 33 994 (86% of sampled subjects) were interviewed in their homes, and they completed the Household Adult Questionnaire and the Dietary Food Frequency Questionnaire (11). All interviewed persons were invited to the mobile examination center, where blood and urine specimens were obtained, and a number of tests and measurements, including body measurements and blood pressure testing, were performed (12).
Analytic sample
All adults aged =" BORDER="0"> 20 y (n = 18 125) were eligible for inclusion in this study. From this eligible sample, we excluded subjects who were pregnant (n = 288) or lactating (n = 95), who were taking drugs for hyperlipidemia or unspecified heart disease (n = 1251), who were told by a physician that they have diabetes (n = 1498), and who reported changing their diet in the past year for any reason (n = 3227). The final analytic sample in this study consisted of 13 130 persons aged =" BORDER="0"> 20 y who completed both the home questionnaire and the medical examination (11).
Dietary assessment methods
Within the framework of the FFQ, subjects were asked how often in the past month they had eaten specific food items (11). The foods were listed in groups that targeted those high in vitamins A and C and calcium but also represented all major food groups consumed by the US population (see Table 1 for food groups). Interviewers also asked about the consumption of other food and beverage items, which were included in the present analyses. It is important to note that portion sizes were not defined, and responses represent the number of times a food was consumed, as reported by the respondent: ie, the number of times consumed per day, per week, per month, or never. All frequency-of-consumption variables were standardized as "times per month" by using the conversion factors 4.3 wk/mo and 30.4 d/mo rounded to the nearest whole number (11). Usual food intake to establish dietary patterns was based on the FFQ (11). The food groups from the FFQ were collapsed to 35 defined food groups (Table 1) to closely approximate food groups that have been used previously in the literature. This was done to limit subjectivity in defining food groups and to allow us to replicate or refute studies already reported in the literature.
View this table:
TABLE 1. Food groupings used in the dietary pattern analysis1
Laboratory methods
Laboratory data in NHANES III were available from whole blood and sera. A questionnaire was administered before the phlebotomy (by venipuncture) to determine an examinee’s eligibility for all phlebotomy procedures: questions were asked to ascertain whether performance of the venipuncture would be safe, to determine and document fasting compliance, and to aid in analyzing the results of the laboratory tests performed. Examinees were instructed to fast for 10–16 h before the morning examination or for 6 h before the afternoon or evening examination (13).
Detailed specimen collection and processing instructions are discussed in the Manual for Medical Technicians, and the analytic methods used by each of the participating laboratories are described in the Laboratory Procedures Used for NHANES III; both of these manuals are part of a larger publication from the Centers for Disease Control and Prevention (13).
Data for all adults were available for serum total cholesterol, HDL-cholesterol, and triacylglycerol concentrations (14). LDL cholesterol was calculated for sample persons who reported fasting for =" BORDER="0"> 9 h and who had triacylglycerol concentrations 400 mg/dL (4.52 mmol/L) according to the equation developed by Friedewald et al (15). In the present study, we included triacylglycerol data only for subjects who reported fasting for =" BORDER="0"> 9 h. This is consistent with the guidelines recommended by the third National Cholesterol Education Program Adult Treatment Panel for lipoprotein analysis (16). Concentrations of red blood cell folate, serum C-peptide, serum insulin, and serum C-reactive protein, as well as glycated hemoglobin (%) and blood pressure, were available for all adults (14). Serum homocysteine data were available only in phase 2 of the survey (1991–1994), and plasma fibrinogen data were available only for subjects aged =" BORDER="0"> 40 y (14).
Statistical analysis
Statistical software
Data preparation was performed by using SAS software (version 8.1; 17). Because NHANES III was conducted in a stratified, multistage probability design, traditional methods of statistical analysis based on the assumption of a simple random sample were not applicable. As recommended by the National Center for Health Statistics, SUDAAN software (version 8.0; 18) was used to estimate descriptive and inferential statistics of interest and the associated variances. Sample weighting was used in NHANES III to account for the unequal probability of selection, noncoverage, and nonresponse bias. Older persons (aged > 60 y), African Americans, and Mexican Americans were oversampled to allow for more precise estimates of the health and nutritional characteristics of those specific population subgroups (12). Appropriate sample weights were applied in all statistical analyses to produce estimates of means and percentiles that can be generalized to the healthy adult US population.
Statistical methods
Because the distribution of the dietary data were extremely nonnormal, data values were truncated at 4 SDs above the mean and then log-transformed. Factor analysis (principal component) was used to derive food patterns based on the frequency of consumption of each of the 35 food groups obtained by collapsing the groups in the FFQ. The analysis was conducted by using FACTOR PROCEDURE in SAS. To account for the complex survey design of NHANES III, a correlation matrix was created from the weighted data on the 35 food groups after the variance was pooled within the sample strata by using PROC GLM in SAS. Next, a data step was performed to read the correlation matrix directly into FACTOR PROCEDURE in SAS. The factors were orthogonally transformed by using varimax rotation to achieve a structure with independent (nonoverlapping) factors.
In determining the number of factors to retain, a stepwise process was utilized. Eigenvalues > 1.25 and the interpretability of the factors were used as the initial cutoffs for reporting dietary patterns, which resulted in 6 dietary patterns. Next, we considered the Scree test, which clearly identified 2 major dietary patterns that then were used in further analyses with CVD risk factors. The factors were labeled on the basis of both interpretation of the data and previously published methods (19). A factor score created for each individual was based on the monthly intake frequencies of the 35 food groups and the standardized scoring coefficient of each food group for each factor. Thus, each person has a factor score for each factor that emerged from the data.
Mean (± SE) percentages were calculated by the linearization (Taylor series) variance estimation method for population indexes. Categorical variables were assessed by using a chi-square test. Ratio scale variables were assessed by using Wald F tests for determination of significance between means of biomarkers by quintile of dietary pattern scores, by using the first quintile as the reference group. Linear regression analyses were conducted between biomarkers of CVD risk and dietary pattern scores after control for the confounding variables of age, sex (male or female), ethnicity (non-Hispanic white, non-Hispanic black, or Mexican American), smoking status (yes or no), alcohol intake (nondrinker: 0 drinks/d; light drinker: > 0–0.5 drink/d; moderate drinker: 0.5 to < 2 drinks/d; heavy drinker: =" BORDER="0"> 2 drinks/d), vitamin or mineral supplement use (yes or no), BMI, physical activity (summation of the frequency of multiple leisure-time activities multiplied by the respective estimated oxygen consumption of each activity), and income (poverty income ratio calculated as the ratio of family income to a Census Bureau–determined poverty threshold). Univariate statistics of all dependent variables were assessed before regression analyses and log transformed where extremely nonnormal to more closely approximate a normal distribution.
RESULTS
The complete factor-loading matrix for the 2 major and 4 minor dietary patterns, subjectively labeled Western, American-healthy, Californian, Breakfast, Southwestern, and Convenience, according to the foods with high factor loadings within each factor, is shown in Table 2. These 6 dietary patterns represent 37% of the variance explained. Further statistical analyses are presented only for the 2 major dietary patterns, which together accounted for 20% of the variance explained. Subjects in the highest quintile of the Western dietary pattern (Table 3) were more likely to be younger, male, non-Hispanic black or Mexican American, less educated, of lower income, smokers, heavy drinkers, less physically active, and less likely to take vitamin or mineral supplements, whereas those in the highest quintile of the American-healthy dietary pattern were more likely to be older, female, white, more educated, of higher income, moderate drinkers, more physically active, and more likely to take vitamin or mineral supplements.
View this table:
TABLE 2. Factor-loading matrix for major and minor dietary patterns in healthy US adults1
View this table:
TABLE 3. Percentage of the study population in sociodemographic and lifestyle categories by quintiles (Q) of dietary pattern scores in healthy US adults1
Because the dietary pattern scores were related to sociodemographic and lifestyle characteristics, further statistical analyses controlled for the effect of sex, ethnicity, smoking status, alcohol intake, vitamin or mineral supplement use, BMI, physical activity, and income. Significant differences in BMI and concentrations of serum LDL and HDL cholesterol, red blood cell folate, glycated hemoglobin (%), and serum C-peptide and insulin across quintiles of Western dietary pattern scores even after control for confounding variables, are shown in Table 4. To test for linear relations, factor scores were modeled as continuous variables, and positive associations were found between the Western dietary pattern and glycated hemoglobin (P < 0.0001), serum C-peptide (P < 0.0001), and serum insulin (P < 0.001) concentrations. An inverse association was seen between the Western dietary pattern and red blood cell folate concentrations (P = 0.0001). Significant differences in BMI and concentrations of serum total and HDL cholesterol and glycated hemoglobin (%) across quintiles of the American-healthy dietary pattern are shown in Table 5, but none of these differences showed a significant linear relation (data not shown). In addition, systolic blood pressure, serum lipoprotein(a), plasma fibrinogen, and serum C-reactive protein were not found to be associated with the dietary patterns (data not shown).
View this table:
TABLE 4. Multivariate-adjusted biomarker values by quintiles (Q) of Western dietary pattern scores in healthy US adults1
View this table:
TABLE 5. Multivariate-adjusted biomarker values by quintiles (Q) of American-healthy dietary pattern scores in healthy US adults1
DISCUSSION
The underlying assumption of statistical data reduction with regard to food intake is that foods eaten together (ie, within the total diet) can be characterized as part of a dietary pattern that is more epidemiologically meaningful than are its individual components (20). The use of factor analysis to define dietary patterns has been criticized, however, for its subjective nature, and there is concern that results cannot be replicated across populations or even within the same population (21). Similar findings across studies would support the use of factor analysis in nutrition epidemiology. Therefore, this research attempted to replicate dietary patterns reported from other epidemiologic studies by using steps similar to theirs in the subjective decision-making process while also using data representative of the healthy US adult population. Results remarkably similar to those reported elsewhere (7, 8, 20, 22) were found, seemingly in spite of the facts that the NHANES III sample is extremely diverse in sociodemographic and lifestyle characteristics and that previously reported data were from age-, sex-, and race-specific samples. These results support the description of Tseng et al (23) of 2 fundamental US dietary patterns that they posit are a result of British culinary heritage (the Western pattern) and of efforts by nutrition science, industry, and government to promote diets that prevent illness (the American-healthy pattern).
The relation of the dietary patterns to sociodemographic and lifestyle characteristics supports the theory that healthy food choices are a part of a larger pattern of health-related characteristics and behaviors (24). Our results indicated relations between a Western dietary pattern and subject characteristics of being nonwhite, male, and less educated; having lower income; smoking; and having low physical activity and low vitamin or mineral supplementation. Dietary patterns may be independent from or may interact with other known risk factors for CVD, although many sociodemographic and lifestyle characteristics were statistically controlled for in the multivariate analyses. In addition, describing dietary patterns by using food intake only, not additional dietary behaviors such as meal and snack patterns, has been criticized (23), and further research may be indicated.
It is not surprising that the Western dietary pattern is associated with lower red blood cell folate concentrations, because of that pattern’s lower consumption of folate-containing foods. It is somewhat surprising, although not inconsistent with prior reports, that the Western dietary pattern is not associated with serum lipid concentrations. An analysis of dietary intake in free-living persons with the use of NHANES III data showed a positive relation between dietary fat intake and serum total cholesterol in men but not in women (25). Results similar to those reported in the current study were found in the Health Professionals Follow-up Study, in which a Western dietary pattern characterized by high intakes of red meat, processed meat, high-fat dairy products, and refined grains was associated with CHD mortality but not with serum lipid concentrations (7, 9). In addition, the Lyon Diet Heart Study revealed a strong protective effect of a Mediterranean dietary pattern in the prevention of CHD recurrence, although there were no differences in concentrations of total, HDL, and LDL cholesterol; triacylglycerol; or lipoprotein(a) between the control and experimental groups at the beginning or end of the study (26). This finding indicates that there are important risk factors other than serum lipids that mediate the relation between diet and CHD. Perhaps the association shown here between markers of glucose metabolism and the Western dietary pattern is indicative of the importance of the relation between dietary patterns and metabolic syndrome (which is itself related to CVD risk) (27).
The American-healthy dietary pattern identified in this research is somewhat similar to the prudent dietary pattern reported in both the Health Professionals Follow-up Study and the Nurses’ Health Study; however, it does not include high intakes of potentially beneficial food groups such as fish, legumes, and low-fat dairy products, as the prudent dietary pattern does. This led us to label the pattern American-healthy, which implies an effort toward a healthy dietary pattern, even if not exactly a prudent pattern. Because the dietary patterns described here are based on actual food intakes of free-living individuals and do not represent ideal dietary patterns, strong associations with specific biomarkers were not expected. Moreover, the 2 major dietary patterns described in this research represent only 20% of the between-person variance in dietary patterns, and thus further analyses of minor dietary patterns may show additional associations between dietary patterns and biomarkers of CVD risk. Also described here are 4 minor dietary patterns, and future research should explore associations between minor dietary patterns and chronic disease risk to more fully evaluate the diet-disease risk relation at the population level.
Dietary pattern analysis is likely to be a more robust and longer-term—and more accurate—measure of dietary habits than is the estimation of the intake of specific nutrients and their interactions. As with all nutrition research, however, dietary pattern analysis can be only as good as the dietary assessment method upon which it is based. Measurement errors inherent in the use of FFQs for dietary assessment include possible underreporting or overreporting of general food intake, selective underreporting or overreporting of the intakes of certain foods, or both. Accordingly, the limitations of FFQs also apply to dietary pattern analyses that are based on dietary information collected from an FFQ. However, because FFQs are designed to assess usual intake, and because they have lower cost and greater relative ease of administration, most large epidemiologic studies have used FFQs.
As previously mentioned, subjectivity was limited by intentionally choosing to include in the factor analysis variables (food groups) that have been used previously, but the exact re-creation of previously reported food groups was not possible. For example, because of the food grouping system used in NHANES III (eg, oatmeal and other hot cereals), our best attempt at replicating food groupings in other studies led us to include all hot cereals in the food group we called "whole grains." The inclusion of unrelated variables in a factor analysis can have the effect of redefining factors because of shared extraneous variance, whereas the exclusion of variables to simplify the factorial structure can lead to erroneous conclusions. Fine-tuning the food groups entered into a factor analysis may improve associations between dietary patterns and markers for disease risk.
A strength of this study is the statistical technique used to account for the complex survey design of NHANES III, ie, pooling the variance within the strata before creating a correlation matrix based on weighted data that could then be used in a factor analysis procedure with SAS software. Because the available software programs that will account for the sample design of NHANES III in producing correct variance estimates (eg, SUDAAN) cannot be used to conduct factor analyses, it was imperative to conduct an appropriate factor analysis to obtain estimates that could then be associated with biomarkers of CVD. To our knowledge, the use of this technique has not been reported elsewhere.
The use of multivariate data-reduction techniques (eg, factor analysis) to define dietary patterns is emerging in nutritional epidemiology as an important method of presenting a snapshot of the entire diet that can then be associated with risks of chronic disease. Dietary pattern analysis is an extension of the many years of research involving specific dietary components, and it remains critically important to continue investigating single dietary components for many reasons, including the correct interpretation of dietary pattern research. In the present study, no attempt was made to surmise cause-and-effect relations, both because of the cross-sectional design of NHANES III and because certain dietary patterns may be part of a larger pattern of healthy or unhealthy behaviors and thus may simply be a surrogate measure for other variables. However, even after we statistically controlled for many potentially confounding variables, dietary patterns alone were significant predictors of biomarkers of CVD risk. The dietary patterns identified here are similar to those reported in other nonrepresentative samples and are associated with biomarkers of CVD risk; both of these facts support the use of dietary patterns in guiding public health recommendations for dietary prevention of chronic disease.
ACKNOWLEDGMENTS
The authors acknowledge the contributions of Lorraine Weatherspoon, Norman Hord, Wanda Chenoweth, and Howard Teitelbaum, who provided substantial review and comments.
This study was designed collaboratively by JMK, EJY, LB, and WOS with the use of the national survey data accessible to the public. Extensive data analyses were carried out by JMK, EJY, and LB. The manuscript was written mainly by JMK under the guidance of WOS.
REFERENCES