A posteriori dietary patterns: how many patterns to retain?

Principal component analysis (PCA) and cluster analysis are used frequently to derive dietary patterns. Decisions on how many patterns to extract are primarily based on subjective criteria, whereas different solutions vary in their food-group composition and perhaps association with disease outcome. Literature on reliability of dietary patterns is scarce, and previous studies validated only 1 preselected solution. Therefore, we assessed reliability of different pattern solutions ranging from 2 to 6 patterns, derived from the aforementioned methods. A validated food frequency questionnaire was administered at baseline (1993-1997) to 39,678 participants in the European Prospective Investigation into Cancer and Nutrition-The Netherlands (EPIC-NL) cohort. Food items were grouped into 31 food groups for dietary pattern analysis. The cohort was randomly divided into 2 halves, and dietary pattern solutions derived in 1 sample through PCA were replicated through confirmatory factor analysis in sample 2. For cluster analysis, cluster stability and split-half reproducibility were assessed for various solutions. With PCA, we found the 3-component solution to be best replicated, although all solutions contained ≥1 poorly confirmed component. No quantitative criterion was in agreement with the results. Associations with disease outcome (coronary heart disease) differed between the component solutions. For all cluster solutions, stability was excellent and deviations between samples was negligible, indicating good reproducibility. All quantitative criteria identified the 2-cluster solution as optimal. Associations with disease outcome were comparable for different cluster solutions. In conclusion, reliability of obtained dietary patterns differed considerably for different solutions using PCA, whereas cluster analysis derived generally stable, reproducible clusters across different solutions. Quantitative criteria for determining the number of patterns to retain were valuable for cluster analysis but not for PCA. Associations with disease risk were influenced by the number of patterns that are retained, especially when using PCA. Therefore, studies on associations between dietary patterns and disease risk should report reasons to choose the number of retained patterns.

Dietary patterns within a population are more reproducible than those of individuals

Insight into the stability of dietary behavior over time is important, because only a single measurement of diet is often available to study the association between eating behavior and the occurrence of chronic diseases many years after baseline data collection. Little is known about changes in dietary patterns over time. The current study examined the (internal) stability and reproducibility of dietary patterns and the transition of individuals between patterns over time from 3 surveys within one study population by using cluster analysis. The dietary intake of participants in the Doetinchem Cohort Study in 6113, 4916, and 4520 adults in 1993-1997, 1998-2002, and 2003-2007, respectively, was measured using a validated food-frequency questionnaire. Stability and reproducibility of dietary patterns were studied by examining the optimal number of clusters per survey by comparing the contribution of food groups to total energy intake within the clusters over time and by studying transitions of individuals between clusters over time. A low-fiber bread pattern and a high-fiber bread pattern were identified in all 3 surveys. Over time, dietary patterns were comparable in terms of foods contributing most to total energy intake, suggesting good reproducibility. Nevertheless, only 41.8% of the participants were consistently assigned to the same dietary pattern for all 3 surveys. This implies that, over time, similar dietary patterns were found at the group level, but that ignoring individual transitions between dietary patterns during follow-up may lead to misclassification of a large proportion of the study population.

Dietary patterns derived from principal component- and k-means cluster analysis: long-term association with coronary heart disease and stroke

BACKGROUND AND AIMS

Studies comparing dietary patterns derived from different a posteriori methods in view of predicting disease risk are scarce. We aimed to explore differences between dietary patterns derived from principal component- (PCA) and k-means cluster analysis (KCA) in relation to their food group composition and ability to predict CHD and stroke risk.

METHODS AND RESULTS

The study was conducted in the EPIC-NL cohort that consists of 40,011 men and women. Baseline dietary intake was measured using a validated food-frequency questionnaire. Food items were consolidated into 31 food groups. Occurrence of CHD and stroke was assessed through linkage with registries. After 13 years of follow-up, 1,843 CHD and 588 stroke cases were documented. Both PCA and KCA extracted a prudent pattern (high intakes of fish, high-fiber products, raw vegetables, wine) and a western pattern (high consumption of French fries, fast food, low-fiber products, other alcoholic drinks, soft drinks with sugar) with small variation between components and clusters. The prudent component was associated with a reduced risk of CHD (HR for extreme quartiles: 0.87; 95%-CI: 0.75-1.00) and stroke (0.68; 0.53-0.88). The western component was not related to any outcome. The prudent cluster was related with a lower risk of CHD (0.91; 0.82-1.00) and stroke (0.79; 0.67-0.94) compared to the western cluster.

CONCLUSION

PCA and KCA found similar underlying patterns with comparable associations with CHD and stroke risk. A prudent pattern reduced the risk of CHD and stroke.