Regression Analysis for Multistate Models Using Time Discretization with Applications to Patients’ Health Status

Abstract

This paper addresses the estimation of multistate models in discrete time, which are widely used to describe complex event histories involving transitions between multiple health states. Accurate estimation of transition intensities and probabilities is essential for understanding disease progression and evaluating the impact of covariates. However, conventional estimators such as the Nelson–Aalen estimator often produce rough estimates, especially in sparse data settings. To improve estimation, we apply kernel smoothing to Nelson–Aalen estimators of transition intensities. Transition probabilities are then derived via product-integrals of the smoothed intensities. Covariate effects on transition intensities are modeled using the Cox proportional hazards model. Rather than modeling covariate effects on transition probabilities indirectly through their influence on transition intensities, we model them directly using pseudo-values of state occupation probabilities obtained through a jackknife procedure. These pseudo-values are treated as outcome variables in a Generalized Estimating Equation (GEE) framework. The proposed methodology is applied to patient visit data from a clinic in West Java, Indonesia, where it successfully captures both the progression dynamics across health states and the influence of key covariates.