Unit censored regression models for recurrent event with covariate adjustment
Abstract
Traditional survival analysis methods often fail to properly account for the substantial proportion of individuals who remain event-free during a fixed follow-up period, leading to biased or less interpretable estimates of restricted mean survival times. The existing models also rely on restrictive distributional assumptions and do not clearly separate the probability of staying event-free from the timing of events among those who experience them. This study addresses these gaps by introducing a flexible unit regression modeling (URM) framework that improves estimation and interpretability, even under complex censoring scenarios. A novel URM approach is proposed for analyzing τ-restricted censored time-to-event and recurrent event data. By explicitly modeling min(τ,T), where T is the event time, the URM framework captures both the point mass at τ and the distribution of event times within the follow-up period. The model decomposes min(τ,T) into τ[I(T≥τ)+(T/τ)I(T<τ)] and estimates its mean through a joint logistic and unit regression structure. Flexible distributions, including the simplex and unit gamma, are incorporated to better reflect diverse event-time patterns. The model estimation is performed using an expectation-maximization algorithm and a multiple imputation approach. The results show that the simplex distribution is more flexible than the unit gamma distribution for URM.