Linear regression calibration: theoretical framework and empirical results in EPIC, Germany.
Annals of nutrition & metabolism 2002 ; 46: 2-8.
DOI : 10.1159/000046746
PubMed ID : 11914509
Large scale dietary assessment instruments are usually based on the food frequency technique and have therefore to be tailored to the involved populations with respect to mode of application and inquired food items. In multicenter studies with different populations, the direct comparability of dietary data is therefore a challenge because each local dietary assessment tool might have its specific measurement error. Thus, for risk analysis the direct use of dietary measurements across centers requires a common reference. For example, in the European prospective cohort study EPIC (European Prospective Investigation into Cancer and Nutrition) a 24-hour recall was chosen to serve as such a reference instrument which was based on a highly standardized computer-assisted interview (EPIC-SOFT). The 24-hour recall was applied to a representative subset of EPIC participants in all centers. The theoretical framework of combining multicenter dietary information was previously published in several papers and is called linear regression calibration. It is based on a linear regression of the food frequency questionnaire to the reference. The regression coefficients describe the absolute and proportional scaling bias of the questionnaire with the 24-hour recall taken as reference. This article describes the statistical basis of the calibration approach and presents first empirical results of its application to fruit, cereals and meat consumption in EPIC Germany represented by the two EPIC centers, Heidelberg and Potsdam. It was found that fruit could be measured well by the questionnaire in both centers (lambdacirc; = 0.98 (males) and lambdacirc; = 0.95 (females) in Heidelberg, and lambdacirc; = 0.86 (males) and lambdacirc; = 0.7 (females) in Potsdam), cereals less (lambdacirc; = 0.53 (males) and lambdacirc; = 0.4 (females) in Heidelberg, and lambdacirc; = 0.53 (males) and lambdacirc; = 0.44 (females) in Potsdam), and that the assessment of meat (lambdacirc; = 0.72 (males) and lambdacirc; = 0.65 (females) in Heidelberg, and lambdacirc; = 0.49 (males) and lambdacirc; = 0.42 (females) in Potsdam) has a center-specific bias. The application of the calibration approach to the questionnaire data will change the ranking of the two centers following the data of the reference instrument, and not well-measured food items will exhibit considerably less variation compared to the original data. We conclude that calibration is a necessary step in multicenter studies. However, this exercise shows that the current statistical framework is not yet sufficiently developed for a broad application.