Evaluating the Applicability of Data-Driven Dietary Patterns to Independent Samples with a Focus on Measurement Tools for Pattern Similarity

Abstract

BackgroundDiet is a key modifiable risk for many chronic diseases, but it remains unclear whether dietary patterns from one study sample are generalizable to other independent populations. ObjectiveThe primary objective of this study was to assess whether data-driven dietary patterns from one study sample are applicable to other populations. The secondary objective was to assess the validity of two criteria of pattern similarity. MethodsSix dietary patterns—Western (n=3), Mediterranean, Prudent, and Healthy— from three published studies on breast cancer were reconstructed in a case-control study of 973 breast cancer patients and 973 controls. Three more internal patterns (Western, Prudent, and Mediterranean) were derived from this case-control study’s own data. Statistical analysisApplicability was assessed by comparing the six reconstructed patterns with the three internal dietary patterns, using the congruence coefficient (CC) between pattern loadings. In cases where any pair met either of two commonly used criteria for declaring patterns similar (CC ≥0.85 or a statistically significant [P<0.05] Pearson correlation), then the true similarity of those two dietary patterns was double-checked by comparing their associations to risk for breast cancer, to assess whether those two criteria of similarity are actually reliable. ResultsFive of the six reconstructed dietary patterns showed high congruence (CC >0.9) to their corresponding dietary pattern derived from the case-control study’s data. Similar associations with risk for breast cancer were found in all pairs of dietary patterns that had high CC but not in all pairs of dietary patterns with statistically significant correlations. ConclusionsSimilar dietary patterns can be found in independent samples. The P value of a correlation coefficient is less reliable than the CC as a criterion for declaring two dietary patterns similar. This study shows that diet scores based on a particular study are generalizable to other populations.