Statistics and Data Science Seminar: "Provider Profiling for Cox's Proportional Hazards Model by Fusion Penalty"

Abstract: Provider profiling is the evaluation of a medical practitioner’s performance on selected clinical or administrative outcomes. In this paper we are interested in profiling health care providers with respect to survival outcomes. Traditionally provider profiling is conducted by either fixed effects or random effects models. In the random effects models, the provider effects are often described by a random intercept with a specific distribution (usually normal), which could lead to over-simplified model and potential biases. On the other hand, in the fixed effects models, the degree of freedom for the provider effects is too large, which could result in a loss of efficiency. A new fused effects model with fusion penalty is proposed for estimation and inference, which achieve a proper balance in terms of efficiency and bias. It can classify providers into different groups, without knowing grouping information in advance. An efficient alternating direction method of multipliers algorithm is developed to implement the proposed method. We evaluate and compare the performance of our method to the fixed- and random-effects models by simulation studies. Our method is applied to profile kidney transplant centers using the national kidney transplant registry data.

Host: Likai Chen and Debashis Mondal