|
Abstract:
|
Censoring due to detection limits is common in biomedical and environmental data sets. When the interest lies in inference for an outcome modeled via a generalized linear model, we propose several multiple imputation strategies. We develop a fully conditional specification approach that takes into account the informative nature of missingness and two joint modeling approaches for directly modeling the marginal distributions of the covariates subject to detection limits. We compare these strategies, as well as several previously proposed approaches, through extensive simulation studies that consider an array of underlying data generation models.
|
