Keisuke Ejima

Ph.D.

Indiana University

Post-doctoral fellow

IDATA (Learn more about this study)

IDATA-17

Oct 31, 2017

Does Exclusion of Extreme Reporters of Energy Intake Make Results Less Biased in Nutrition Studies?

Energy intake (EI) is a key factor of energy balance, as is energy expenditure, and known to be associated with obesity and is plausibly related to other health outcomes. Therefore, study of EI is of vital interest in nutrition/obesity studies. Self-report energy intake (SREI) derived from food diaries and other methods is frequently used in those study fields, however, it is well known that SREI is severely error-prone and usually underestimates (sometimes overestimates) actual energy intake. As an accurate measurement of EI, doubly labeled water (DLW) method is widely accepted.
To address the issue of SREI, in many nutrition/obesity studies, outliers (reporting extremely high or low EI) are excluded from analyses based on ‘Goldberg’s rule’ (Goldberg et al., EJCN 1991), which is believed to make results from analyses less biased. However, whether use of this procedure actually makes results less biased has not been tested in empirical studies and we know of no formal theoretical foundation (i.e., mathematical proof) for claiming that it does so.
In this study, we examine whether applying Goldberg’s rule to a dataset make results less biased in nutrition/obesity studies through a test case study using the NCI iData study dataset. In this study, three linear regression analyses will be conducted for each of several health measurements (e.g., weight, blood pressure and blood cholesterol level) as outcomes. The first and second analyses will both use SREI as a main predictor. However, implausible SREI measurements will be excluded from the second analysis using Goldberg’s rule. The third analysis will use EI estimated from DLW, which is a ‘gold standard’, as a main predictor. The estimated coefficients from each analysis are compared to evaluate whether applying Goldberg’s rule to SREI makes the estimated association between EI and the health outcomes less biased. We will also address the research question through a complementary set of computer simulation studies in which data are generated to emulate the distributions of the real observed data.