Evaluating Various Estimating Equation Approaches to Handle Right-Censored Covariates in Regression

While right-censored 'time to event' outcomes in regression have been studied for decades, handling these 'time to event' variables as covariates in regression is of growing interest. For a biomedical example where censored covariates are present, we consider Huntington disease. Knowledge of how an individual's symptoms progress in Huntington disease is invaluable for identifying helpful interventions. While modeling this progression is not easy, a promising solution is to piece together short-term clinical data from patients and align the pieces along their time of clinical diagnosis. However, time to clinical diagnosis is not observed for everyone, usually due to drop-out or limited resources. This variable is right-censored, which creates a unique statistical problem where symptoms are modeled as a function of a right-censored covariate. In a regression framework, it is desirable to produce consistent, efficient, and robust estimators. When the data are fully observed, there are standard estimation procedures to aid in achieving this goal, but estimation is less straightforward when covariates are censored. In Chapter 2, we clarify when the complete case estimator is guaranteed to be consistent and draw connections between various censoring mechanisms that lead to consistency. In Chapter 3, we evaluate the robustness of several different proposed methods under informative censoring including inverse probability weighting, maximum likelihood, and augmented inverse probability weighting. All methods will be applied to Huntington disease observational study data to analyze various symptoms in these patients as a function of time to clinical diagnosis. In Chapter 4, we present an R Package, EECovCensRd, that in addition to providing user-friendly access to all previously proposed methods, implements an ‘augmented’ complete case estimator.