Analysis of spatially clustered survival data with unobserved covariates using SBART

Oregon Convention Center 

Popular regression methods for clustered survival data become inadequate when functional forms of covariates and interactions are unknown. Spatially correlated random cluster effects and unknown cluster-level covariates further complicate the issue. In this context, we propose a nonparametric approach, SBART (soft Bayesian additive regression trees) within the Bayesian ensemble learning paradigm (Basak et al., 2022). Our method addresses challenges like numerous clusters, variable cluster sizes, and data information for proper statistical augmentation of the unobserved covariate sourced from a data registry different from the survival study. We illustrate the practical implementation of our method and its advantages over existing methods by assessing the impacts of intervention in some county level and patient-level covariates to eliminate existing racial disparity in breast cancer survival in different Florida counties. The sources of clustered survival data with patient-level covariates and the data information for an unobserved county-level covariate are different. We also compare our method with existing analysis methods to demonstrate our advantages through simulation studies.

Soft Bayesian Additive Regression Trees (SBART)

Data Augmentation

Disparity

Markov Chain Monte Carlo (MCMC)

Metropolis Hastings (MH) Algorithm.

Unobserved covariate 

Section on Nonparametric Statistics