|
Abstract:
|
Delayed enrollment of subjects into a time-to-event study produces a sample in which larger event times are overrepresented. The corresponding covariate sample is similarly biased. Missing covariates often arise in epidemiologic studies when data is expensive or invasive to collect. Ignoring observations with missing values can produce biased and inefficient results. Several general missing data strategies have been proposed, including inverse probability weighting (IPW), augmented IPW (AIPW), and multiple imputation (MI). We evaluate these approaches when applied to estimating the hazard ratio with the Cox model through simulation studies under a variety of truncation scenarios. In particular, we highlight the effect of selection bias on modeling the missing covariate distribution in AIPW and MI. Additionally, we illustrate characteristics of the above approaches under left truncation by applying them to a study of Parkinson’s disease dementia. Careful consideration of method assumptions and the amount of truncation is necessary to obtain valid results in left truncated samples with missing covariates.
|
