Reproducibility and Validity of A Posteriori Dietary Patterns: A Systematic Review

Abstract

The effective use of dietary patterns (DPs) remains limited. There is a need to assess their consistency over multiple administrations of the same dietary source, different dietary sources, or across different studies. Similarly, their generalizability should be based on a previous assessment of DP construct validity. However, to date, no systematic reviews of reproducibility and validity of a posteriori DPs have been carried out. In addition, several methodological questions related to their identification are still open and prevent a fair comparison of epidemiological results on DPs and disease. A systematic review of the literature on the PubMed database was conducted. We identified 218 articles, 64 of which met the inclusion criteria. Of these, the 38 articles dealing with reproducibility and relative and construct validity of DPs were included. These articles (published in 1999– 2017, 53% from 2010 onwards) were based on observational studies conducted worldwide. The 14 articles that assessed DP reproducibility across different statistical solutions examined different research questions. Included were: the number of food groups or subjects; input variable format (as well as adjustment for energy intake); algorithms and the number of DPs to retain in cluster analysis; rotation method; and score calculation in factor analysis. However, we identified at most 3 articles per research question on DP reproducibility across statistical solutions. From another 15 articles, reproducibility of DPs over shorter (≤1 y) time periods was generally good and higher than DP relative validity (as measured across different dietary sources). Confirmatory factor analysis was used in 15 of the included articles. It provided reassuring results in identifying valid dietary constructs characterizing the populations under consideration. Based on the available evidence, only suggestive conclusions can be derived on reproducibility across different statistical solutions. Nevertheless, most identified DPs showed good reproducibility, fair relative validity, and good construct validity