In epidemiology as well as in clinical studies our main interest lies in the estimation of the effect of exposure on disease. Most exposure data are subject to measurement error, and when ignored, they will bias our estimates of exposure-disease associati ons. It is important to understand the size of the measurement error in the exposure variables, the impact of measurement error on the associations, and, if possible, correct for it. This project will focus on correction for measurement error through esti mation of the distribution of the exposure variable. Estimating the density when we have observations measured with error is called density deconvolution. In clinical and epidemiological studies the estimation of the exposure distribution is of interest i n itself. But in most applications it is, of course, more relevant to use such an exposure estimate to relate it to some outcome variables of interest. Once we have the true exposure distribution one main issue is to construct corrected risk estimates tha t take the true exposure distribution into consideration. We will use the Generalized Additive Model (GAM) approach to relate the deconvolution estimate to the outcome variable. In our setting we will in this way obtain risk estimates of disease, correcte d for measurement error.
Our methodological projects will clarify
1. Methods for estimation of multivariate exposure distributions in measurement error models
2. Within the framework of the Generalized Additive Models (GAM) relate true exposure to outc ome.
With our collaborators we will study the following applications
1. Through the large, population-wide cohort "The Norwegian Women and Cancer Study" (NOWAC) we have excellent opportunities to study the impact of measurement errors in dietary data on risk estimates
2. Can findings from case-control studies that are often not confirmed in the larger cohort studies be explained by differential misclassification?