Abstract:
|
Prediction performance methods for Cox models have traditionally included the concordance statistic. However, when the Cox model involves both time-varying covariates and competing risk, there is paucity of methods for model performance assessment. Our work involves creating a framework to support development of a concordance statistic for competing risk models with time-varying covariates. We first compare existing concordance methods for the less complex cases: Cox proportional hazards model with time-varying covariates (but not competing cause) and models with competing causes only. We then build a theoretical framework for developing a concordance statistic for the complex case. We compare these methods using simulation results, and an application to a large cohort of Veterans (n=115k) with type 2 diabetes to assess whether treatment with second-generation sulfonylurea medication for diabetes was associated with three events: hospital admission for a major adverse cardiac event (MACE), hospitalization for hypoglycemia, or death. We will also address the computational challenges for running these models on such a very large administrative healthcare dataset.
|