EFFECT OF MEASUREMENT ERROR ON ENERGY-ADJUSTMENT MODELS IN NUTRITIONAL EPIDEMIOLOGY

The use and interpretation of energy-adjustment regression models in n utritional epidemiology has been vigorously debated recently. There ha s been little discussion, however, regarding the effect of dietary mea surement error on the performance of such models. Contrary to conventi onal assumptions invoked in the standard treatment of the effect of me asurement error in regression analysis, reporting errors in dietary st udies are usually biased, correlated with true nutrient intakes and wi th each other, heteroscedastic, and nonnormally distributed. Methods d eveloped in this paper allow for this more complex error structure and are therefore more appropriate for dietary data, For practical illust ration, these methods are applied to data from the Women's Health Tria l Vanguard Study. The results demonstrate considerable shrinkage in th e magnitude of the estimated main exposure effect in energy-adjustment models due to attenuation of the true effect and contamination from t he effect of an adjusting covariate. In most cases, this shrinkage cau ses a sharply reduced statistical power of the corresponding significa nce test in comparison with measurement without error. These results e mphasize the need to understand the measurement error properties of di etary instruments through validation/calibration studies and, where po ssible, to correct for the impact of measurement error when applying e nergy-adjustment models.