Objective: Estimating the overall successfulness of a treatment can be difficult when success is defined by freedom from multiple endpoints that are each subject to competing risks. We describe a method for modeling short-term competing outcomes.
Methods: We used polytomous categorical variable modeling to describe the 30-day onset of renal failure, neurologic deficit, stroke or death (events) following repair of 841 thoracoabdominal aortic aneurysms. This was to determine whether common risk factors had a multivariate association with these outcomes, and whether predictor variables might be positively associated with some outcomes and negatively associated with others. The goal was to determine whether a single aggregate-endpoint logistic model could accurately predict the probability of good outcome 30 days following surgery.
Results: When more than one event occurred in a single patient, the first (or most severe simultaneous) event was used for censoring. Five hundred and ninety-three out of 841 (70.5%) patients had no postoperative events. The most common event was renal failure. We detected five predictors that were significant for at least one of the four outcomes. These were age, poor preoperative renal function (RENAL), acute dissection, extent II aneurysm, and use of cerebrospinal fluid drainage and distal aortic perfusion (ADJUNCT). Only RENAL was significant for all outcomes. ADJUNCT was highly significant only for neurologic deficit in the polytomous analysis and dropped out of the aggregate-endpoint multiple logistic model.
Conclusion: Polytomous-outcome multivariate categorical modeling can detect effects missed by aggregate models, and is a valuable and statistically powerful method for evaluating risk factor effects on multiple competing endpoints.