Comparative study of outcome measures and analysis methods for traumatic brain injury trials

Abstract

Batteries of functional and cognitive measures have been proposed as alternatives to the Extended Glasgow Outcome Scale (GOSE) as the primary outcome for traumatic brain injury (TBI) trials. We evaluated several approaches to analyzing GOSE and a battery of four functional and cognitive measures. Using data from a randomized trial, we created a "super" dataset of 16,550 subjects from patients with complete data (n=331) and then simulated multiple treatment effects across multiple outcome measures. Patients were sampled with replacement (bootstrapping) to generate 10,000 samples for each treatment effect (n=400 patients/group). The percentage of samples where the null hypothesis was rejected estimates the power. All analytic techniques had appropriate rates of type I error (≤5%). Accounting for baseline prognosis either by using sliding dichotomy for GOSE or using regression-based methods substantially increased the power over the corresponding analysis without accounting for prognosis. Analyzing GOSE using multivariate proportional odds regression or analyzing the four-outcome battery with regression-based adjustments had the highest power, assuming equal treatment effect across all components. Analyzing GOSE using a fixed dichotomy provided the lowest power for both unadjusted and regression-adjusted analyses. We assumed an equal treatment effect for all measures. This may not be true in an actual clinical trial. Accounting for baseline prognosis is critical to attaining high power in Phase III TBI trials. The choice of primary outcome for future trials should be guided by power, the domain of brain function that an intervention is likely to impact, and the feasibility of collecting outcome data.

Keywords: Glasgow Outcome Scale; outcome measures; research design; statistical data analysis; traumatic brain injury.

FIG. 1.

Flow diagram of the simulation design.

FIG. 2.

Power curves for different outcome measures and analysis methods (sample n=400 per treatment group). For average percentile method, only results from Wilcoxon rank sum test are displayed in the figure (similar results were noted with t-test). For GEE, we only show the results under the assumption of exchangeable correlation structure (similar results were noted under the assumption of unstructured correlation). GOSE, Extended Glasgow Outcome Scale; GEE, generalized estimating equations.

FIG. 3.

Power curves for adjusted analyses using regression-based methods (sample n=400 per treatment group). Analyses were adjusted for the baseline probability of unfavorable outcome (i.e., GOSE categories: 1–4), as calculated using the International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) core prognostic model. Power curve for sliding dichotomy method (10 prognostic groups) is shown for comparison (red curve). For average percentile method, only results from Wilcoxon rank sum test are displayed in the figure (similar results were noted with t-test). For GEE, we only show the results under the assumption of exchangeable correlation structure (similar results were noted under the assumption of unstructured correlation). GOSE, Extended Glasgow Outcome Scale; GEE, generalized estimating equations.