Recognition has been growing that occupational cohort studies need to assess the impact of confounders and effect modifiers on exposure-response models. However, the statistical assumptions and methods to detect and adjust for confounding and effect modification have had weaknesses that need to be addressed, so this proposal will develop new methods to do so. Current methods for detecting effect modifiers in cohort studies rely on specific parametric model assumptions that may be inappropriate or arbitrary. We will develop a fully nonparametric method. Methods to adjust for confounding and effect modification also tend to be unstable and inefficient; we will develop an empirical Bayesian method that increases efficiency. Random measurement error effects can be amplified by effect modifiers, a problem that has not been studied; we will address the issue and develop methods to appropriately correct for this. Collinearity among covariates is common in occupational cohort studies (e.g., correlation between cumulative exposure and duration of employment), but methods have not been developed to address this for cohort studies. We will develop methods for both collinearity diagnostics and a ridge regression approach to minimize its effects. A related problem of the correlated variables being both confounders and intermediate variables will also be addressed. Simulations will be used to evaluate all these methods for statistical power and size, and the methods will be applied to real occupational cohort data. The proposed methods for modeling confounders and effect modifiers will have widespread application to a wide range of occupational or environmental cohort studies with exposure-response data.
