Multiple Event Analysis of Injuries Using Adaptations to the Cox ProportionalHazards Model

摘要

Understanding the epidemiology of injuries is of great importance to the UnitedStates Military. However, there is presently limited information regarding statistical methodology, as it pertains to a setting where an individual can experience multiple injuries. This thesis explores the use of three statistical models, each with distinctive underlying assumptions, that are commonly used in recurrent failure time settings. Each was applied to the same data set of United States Army Airborne soldiers (n=1214) . The outcome of interest was lower extremity or low back injury, and only the first and second injury events were examined. The methods employed were two Cox Proportional Hazards Models, each representing a separate injury event; the Andersen-Gill (AG) Multiplicative Hazards Model, which employs a counting process formulation; and the Prentice, Williams and Petersen (PWP) Model, where the multiple events are modelled via stratification. The final results for the Cox Model to first injury, and the first strata of the PWP Model are equivalent. The final AG Model, yielded coinciding covariates to the first injury event in the other models, with minimal differences between the parameter estimates. Similarly, the final Cox Model for the second injury event and the second strata of the PWP Model are equivalent; however, they produce different risk factors than the Cox Model for first event and the first strata of the PWP Model. The comparison of the different methodologies demonstrate that the PWP Model is best suited for the multiple injury setting. The facts that both the baseline hazard and the parameter estimates alter by event, and that it allows for easy comparison between strata (injury events), justifies this claim.