Abstract:
|
In analyses of drug safety studies using large administrative datasets, confounding by indication is an important source of bias. The propensity score is often used for confounding control. When large numbers of covariates are available for analysis, methods such as the high-dimensional propensity score (hdPS) are often used; this method has been shown to have reasonable properties under a range of circumstances. However,we demonstrate that they may suffer from over-fitting when the number of exposed subjects is small relative to the number of covariates considered, in particular when the propensity score is used for inverse weighting. Standard propensity score methods may also introduce bias when relations between covariates and treatment are non-linear and the model is mis-specified. We compare the standard propensity score to a variety of machine-learning prediction algorithms to estimate the propensity score for use in inverse weighted estimation of a marginal structural Cox model; machine learning tools demonstrate small bias and mean square error.
|