Estimating heterogeneous treatment effects for latent subgroups in observational studies

Area of Science:

• Biostatistics
• Medical Research Methodology

Background:

• Identifying patient subgroups with differential treatment responses is crucial in medical research.
• Subgroup analyses carry a risk of misleading results, necessitating robust exploratory methods.
• Latent subgroup structures, not directly observed, pose challenges for standard statistical models.

Purpose of the Study:

• To propose a semiparametric Bayesian latent model for exploratory subgroup effect analysis.
• To assess the model's utility in identifying latent subgroups associated with observed covariates.
• To enhance model robustness using propensity score matching for covariate balancing.

Main Methods:

• Developed a semiparametric Bayesian latent model flexible in capturing data-driven latent structures.
• Employed propensity score matching to balance covariate distributions and improve model robustness.
• Conducted simulation studies to evaluate the model's performance in identifying subgroup structures and estimating effects.

Main Results:

• The proposed model successfully identified latent subgrouping structures in simulation studies.
• Propensity score matching adjustment yielded robust estimates even with a misspecified outcome model.
• Real-data analysis revealed significant subgroup effects on steroid avoidance in pediatric kidney transplant patients, missed by standard regression.

Conclusions:

• The semiparametric Bayesian latent model provides a valuable tool for exploratory subgroup effect analysis.
• Propensity score matching enhances the reliability of subgroup effect estimates, particularly in complex clinical settings.
• This approach can uncover clinically relevant subgroup effects missed by conventional methods, improving personalized medicine.