Developing Pseudo-Observation and Multiple Imputation Approaches for Analysis of Dependently Censored Survival and Quality-Adjusted Survival Data.

Abstract

Dependent censoring is a common issue in survival and quality-adjusted survival analysis. This thesis develops pseudo-observation and multiple imputation approaches for analysis of these types of data. Our motivating survival analysis example takes place in the lung allocation setting, where more urgent patients are removed for transplant. Our motivating quality-adjusted survival analysis example takes place when a quality-of-life function is applied to follow-up time in breast cancer patients, inducing dependent censoring on the quality-of-life timescale.

In Chapters II and III we consider transplant urgency, defined as the days of life lived in one year without transplant. Waitlist candidates desire urgency estimates when being advised of their prognosis. This urgency measure is also used in organ allocation as part of the lung allocation score (LAS). Priority for transplant changes on a daily basis as lung candidates deteriorate, with more urgent patients leaving the waitlist as they receive lung offers. Leaving the waitlist before observing time-to-death without transplant induces bias when summarizing unadjusted outcomes.

In Chapter II, we develop an inverse-weighted pseudo-observation approach for estimating restricted means subject to dependent censoring. In Chapter III, a multiple imputation approach for completing dependently censored datasets is developed that incorporates time-varying factors predicting removal from the waitlist. The advantage of the latter approach is that many different types of analyses may be conducted using traditional software once the completed datasets are constructed. Simulations show that both the pseudo-observation approach to estimating restricted means and restricted mean analysis on multiply imputed datasets give good inference. We compare estimates of transplant urgency, transplant benefit and lung allocation scores when using the two approaches.

Many end-stage lung patients enter the waitlist to improve their quality of life rather than extend their life. Quality-of-life data are not collected by the U.S. Organ Procurement and Transplant Network since it is not included in lung prioritization. However, in the event this changes, statistical methodology must be available to estimate quality-adjusted urgency and transplant benefit. We describe how to multiply impute censored quality-adjusted survival outcomes in Chapter IV, applying our method to the International Breast Cancer Study Group Ludwig Trial V.