Re-Evaluating the Impact of Including Patients with Bilateral Conditions in Orthopaedic Clinical Research Studies: When 1 + 1 Does Not Equal 2.

Abstract

BACKGROUND Orthopaedic studies frequently include subjects with bilateral conditions. Failure to account for bilateral conditions can lead to spurious associations. The performance of different methods for addressing this issue, especially in populations that include subjects with unilateral and bilateral conditions, has not been rigorously evaluated. The purpose of the present study was to test 3 different methods for analyzing bilateral data: (1) analyzing all limbs as independent subjects (naïve), (2) randomly selecting 1 limb per subject (random), and (3) accounting for correlation between limbs with use of a linear mixed model (LMM). METHODS We simulated a hypothetical randomized controlled trial in which Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) scores were collected at a baseline and a 2-year visit. We simulated 2 scenarios: Scenario 1 (in which there was truly no difference between groups [mean difference = 0]) and Scenario 2 (in which there was truly a difference between groups [mean difference = 10]). We varied the prevalence of bilateral involvement from 10% to 100% within each scenario. We evaluated method performance on the basis of bias (difference from the simulated true effect), power (1 - type-II error), type-1 error rate, and 95% confidence interval (CI) coverage. RESULTS Bias (difference from simulated true effect) was similar across all methods. In Scenario 2 (true difference between groups), CI coverage was lowest with use of the naïve method (median, 87.8%; range, 85.3% to 93.5%) relative to the random method (median, 95.1%; range, 94.5% to 95.6%) and the LMM method (median, 95.1%; range, 94.5% to 95.5%). In Scenario 1 (no difference between groups), the type-1 error rate was highest for the naïve method (median, 11.3%; range, 6.7% to 14.7%) relative to the LMM method (median, 4.9%; range, 4.5% to 5.3%) and the random method (median, 5.0%; range, 4.5% to 5.2%). CONCLUSIONS Failure to account for bilateral conditions led to biased CIs and an increased type-1 error rate. Due to the fact that bias was similar across the methods, decreased model performance using the naïve method was likely attributable to underestimation of the standard error. Orthopaedic studies involving subjects with bilateral conditions warrant special considerations that can be addressed using simple (random) or more complex (LMM) methods. CLINICAL RELEVANCE Adherence to robust methodological practices is an essential but underappreciated component of the translation of evidence into clinical practice. Our work is meant to be educational, providing clinical researchers with the knowledge and skills to address a common challenge within the field.