|
Abstract:
|
Exposure data with repeated measures from environmental studies are commonly right-skewed and left-censored. The use of linear mixed effects model incorporating maximum likelihood method for repeated measures data with non-detects has been discussed to model log-normal exposure outcomes. However, this modeling has a disadvantage that assumes a correctly specified distribution for the random effect, which is practically unknown, and the estimation methods can result in bias and imprecision in finite-sample data even when distributional assumptions are met. Marginal modeling provides an alternative to analyze data with repeated measurements, in which the parameter interpretations are with respect to population-averaged means. We outline the theories of three marginal models, i.e., generalized estimating equations (GEEs), quadratic inference functions (QIF), and generalized method of moments. With these approaches, we propose to incorporate fill-in methods. In a simulation study and application examples, we demonstrate that the GEE method works well in terms of estimating the regression parameters, particularly in small sample sizes, while the QIF outperforms in large-sample settings.
|
