The rise of precision medicine has shifted statistical analysis from population-averaged treatment effects to subgroup-specific effects, highlighting the importance of tailoring treatments to distinct patient subgroups. Longitudinal data, in particular, provides an opportunity to study how treatment effects evolve over time. However, methods for analyzing subgroup-specific, time-varying treatment effects in longitudinal settings remain limited. Most current approaches rely on two-step procedures with restrictive assumptions, such as a fixed number of subgroups or prespecified correlation structures. In this paper, we propose a novel semiparametric subgroup-varying-coefficient model that simultaneously estimates all unknowns, including the number of subgroups and subgroup-specific treatment effects, directly from longitudinal data without requiring restrictive assumptions. We develop a penalized Expectation-Maximization (EM) algorithm that simultaneously identifies the subpopulations and estimates the treatment effects in different subgroups. We establish the theoretical properties of the proposed estimators, including consistency and asymptotic normality. The efficacy and robustness of the method are demonstrated through extensive simulations and applications to two real-world datasets: the Standard and New Antiepileptic Drugs study and the Mayo Clinic trial on primary biliary cirrhosis. In both cases, our approach successfully identifies meaningful patient subgroups and reveals subgroup-specific, time-varying treatment sensitivities.
Zhang et al. (Fri,) studied this question.