Accounting for Withdrawal of Life-Sustaining Treatment in the Analysis of Traumatic Brain Injury Studies.

Abstract

Studies that aim to evaluate outcomes after severe traumatic brain injury (TBI) must account for patients who die after withdrawal of life-sustaining treatment (WLST). If we are willing to assume that some of the patients who die of WLST might have had a good outcome at 6 months, the choice of analytic approach may impact the results. In this study, 6-month clinical outcomes for patients with TBI were simulated under six different scenarios related to WLST. Each scenario represents different assumptions related to the decision to choose WLST and how that decision relates to the 6-month clinical outcome. For each simulated dataset and scenario, three analytic approaches were used to estimate the probability of a good outcome at 6 months: complete case analysis, worst-case imputation, and inverse probability weighted analysis. The bias of the estimate from each of the approaches was used to compare the performance of the analysis approaches. When the probability of WLST was equal for all patients (i.e., covariates were not factored into the WLST decision), both the complete case analysis and the inverse probability weighted analysis were unbiased. When only patients who would have a poor outcome at 6 months were eligible to have WLST, only the worst-case imputation analysis was unbiased. When the probability of WLST was a function of observed patient characteristics that were also related to 6-month outcome (e.g., age, injury severity), only the inverse probability weighted analysis was unbiased. Finally, when the probability of missingness was related to an unobserved patient characteristic, none of the approaches were unbiased. If some patients who die of WLST might have had a good outcome, inverse probability weighting could be considered to decrease bias associated with censoring or imputing poor outcomes for participants with WLST.