A Comparison of Statistical Methods for Variable Selection Under Competing Risks: Modelling Death Possibly Related to Total Pancreatectomy with Islet Auto Transplantation (TPIAT)

Presented by Losha Ndemeno-Tegomoh
Masters Candidate in Biostatistics
Plan B Adviser: Dr. Anne Eaton

Competing risks can arise in survival analysis when the occurrence of one event prevents the occurrence of another. Competing risks should be taken into consideration because, if ignored, the presence of a competing risk can bias the results of statistical analysis. The Fine-Gray proportional sub distribution hazards model is a commonly-used method for regression analysis that accounts for the presence of competing risks. Several methods have been proposed for variable selection in Fine-Gray regression. This study used data from 555 participants that have undergone total pancreatectomy with islet auto transplantation (TPIAT) at the University of Minnesota 5 – 20 years ago to investigate three such variable selection methods: stepwise selection based on the BICcr, a criteria tailored to Fine-Gray regression, random survival forests, and LASSO. The event of interest was death possibly related to TPIAT, with death due to other causes treated as a competing risk. By 15 years post-TPIAT, 9.2% (95% CI: 0.01%, 29.5%) of participants experienced death possibly related to TPIAT. Age and pancreatitis duration were significant predictors for the incidence of TPIAT-related death. All variable selection methods identified these variables as among the most important. Multiple variables that were correlated with one another and may be related to pancreas health might also be important predictors for the incidence of TPIAT-related death. Different variable selection methods selected different variables from within this subset. Fibrosis severity level may not correlate with the risk of death related to TPIAT in a linear fashion, which may explain why it was not identified as important by some variable selection methods.