Baseline characteristics and statistical power in randomized controlled trials: selection, prognostic targeting, or covariate adjustment?

Abstract

Objective: Heterogeneity of patients is a common problem in randomized controlled trials (RCTs) in various fields of clinical research. We aimed to investigate the potential benefits of different approaches for dealing with heterogeneity in a case study on traumatic brain injury (TBI).

Design and setting: Statistical modeling studies in three surveys and six randomized controlled trials.

Patients: Individual patient data (n = 8033) from the IMPACT database.

Interventions: We investigated the statistical power and efficiency of randomized controlled trials (RCTs) in relation to (1) selection according to baseline characteristics, (2) prognostic targeting (i.e., excluding those with a relatively extreme prognosis), and (3) covariate-adjusted analysis. Statistical power was expressed as the required sample size for obtaining 80% power and efficiency as the relative change in study duration, reflecting both gains in power and adverse effects on recruitment. Uniform and targeted treatment effects were simulated for 6 month unfavorable outcome.

Results: For a uniform treatment effect, selection resulted ina sample size reduction of 33% in the surveys and 5% in the RCTs, but decreased recruitment by 65% and 41%, respectively. Hence, the relative study duration was prolonged (surveys: 95%; RCTs: 60%). Prognostic targeting resulted in sample size reductions of 28% and 17%, and increased relative study duration by 5% in surveys and 11% in the RCTs. Covariate adjustment reduced sample sizes by 30% and 16%, respectively, and did not affect recruitment. For a targeted treatment effect, the sample size reductions by selection (surveys: 47%; RCTs: 20%) and prognostic targeting (surveys: 49%; RCTs: 41%) were larger and adverse effects on recruitment smaller.

Conclusions: The benefits of selection and prognostic targeting in terms of statistical power are reversed by adverse effects on recruitment. Covariate adjusted analysis in a broadly selected group of patients is advisable if a uniform treatment effect is assumed, since there is no decrease in recruitment.