Person-Period survival models with time-dependent covariates: a comparative simulation and pharmacovigilance case study

Abstract

Person-period data are widely used in time-to-event analyses with time-dependent covariates when covariate histories are summarized over discrete intervals. In pharmacovigilance databases, cumulative drug exposure often requires reconstruction from irregular dosing records, requiring choices about interval width, covariate construction, and model interpretation. We compared the time-dependent Cox model, the Poisson log-linear model with an interval-length offset, and the discrete-time logit model. The Cox and Poisson models estimate closely related hazard- and rate-ratio effects under a piecewise exponential representation, whereas the logit model estimates conditional interval log-odds ratios. In simulations under proportional hazards models with Weibull baseline hazards, we varied sample size, follow-up time, interval width, Weibull parameters, and the time-dependent covariate trajectory. Across 1,000 valid replications per setting, the time-invariant coefficient was relatively stable, whereas the time-dependent coefficient was more sensitive to interval width and covariate evaluation time, particularly for logarithmic trajectories. Poisson and Cox estimates were generally similar on fine grids, whereas logit estimates differed more when interval event probabilities were not small. Additional scenarios generated from an accelerated failure time model with time-dependent covariates were examined as sensitivity analyses under model misspecification. A case study using the Japanese Adverse Drug Event Report database illustrates implementation and interpretation.