Comparisons of Propensity Score Methods for Time to Event Outcomes: Evaluation through Simulations and Oral Squamous Cell Carcinoma Case Study

Authors: Sophie Yunfei Ma1 , Badr Id Said2 , Ali Hosni2 , Wei Xu1,3, Sareh Keshavarzi3*
*Corresponding author: Sareh Keshavarzi

Affiliations:
1. Biostatistics Division, Dalla Lana School of Public Health, University of Toronto, Toronto, Ontario, Canada
2. Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Ontario, Canada.
3. Department of Biostatistics, Princess Margaret Cancer Center, Toronto, Ontario, Canada

Key words: Causal Inference, Propensity Score Methods, Survival Outcomes

Observational studies are sometimes preferred over randomized trials because of having greater timeliness, lower cost, and more generalizability. The issue with the observational studies is that the causal inference cannot be simply made due to potential confounding effects. Propensity scores (PS) analysis is a well-known approach that can be used to reduce the confounders' effect on the treatments, interventions, or exposure estimates in observational studies. Few guidelines are available regarding the choice of PS approaches or covariate adjustment for the best performance in a given set of data. In this study, we conducted extensive scenarios of Monte Carlo simulations comparing the performance of conventional covariate adjustment and eight other common PS approaches to estimate the average treatment effect for the overall population on time-to-event outcomes. We also implemented the aforementioned PS approaches to compare the effect of receiving postoperative radiation therapy and having an engraftable tumor on different time-to-event clinical outcomes in Oral Squamous Cell Carcinoma Cancer patients. In our simulations, 1:1 nearest neighbor with and without caliper matching provided less precise estimates especially with few outcome events and low treatment prevalence. Inverse probability of treatment weighting and covariate adjustment performed well in most cases and produced unbiased estimates with small uncertainty. In the case study, results showed consistency across stratification and covariate adjustment. In practice, care should be taken in selecting the appropriate PS methods over covariate adjustments.