Pseudo-observation process regression for survival outcomes

Abstract: Pseudo-observation methods have become useful tools for regression with censored survival data, especially when the quantity of interest is not a standard hazard-based parameter. Most current applications focus on estimating outcomes at a fixed and limited number of time points, such as survival probability, cumulative incidence, or restricted mean survival. In many applications, however, the main goal is to describe how the entire outcome curve evolves over time for patients with different covariate profiles. We propose a general framework that extends pseudo-observation regression from a small set of time points to the full time course of a survival outcome. The idea is to construct pseudo-observations over time and combine them with a process-regression approach, so that covariate effects can be studied continuously rather than only at selected times. This yields an estimated covariate-specific survival curve and includes existing pseudo-observation regression methods as a special case.  The proposed framework is flexible and can be used for several survival-type outcomes, including overall survival, cumulative incidence in competing risks, and related functionals. It also provides a bridge between familiar pseudo-observation methods and broader regression approaches for time-varying outcomes. We discuss the main methodological ideas, practical computation, and extensions to settings with covariate-dependent censoring. The approach is motivated by the need for interpretable regression methods that target clinically meaningful quantities directly, while making fuller use of follow-up information over time. This work aims to broaden the scope of pseudo-observation methods from pointwise analysis to process-level estimation and prediction.   Joint work with Omer Moyal