Impact of Selection Bias on Treatment Effect Size Estimates in Randomized Trials of Oral Health Interventions: A...

Emerging evidence suggests that design flaws of randomized controlled trials can result in over- or underestimation of the treatment effect size (ES). The objective of this study was to examine associations between treatment ES estimates and adequacy of sequence generation, allocation concealment, and baseline comparability among a sample of oral health randomized controlled trials. For our analysis, we selected all meta-analyses that included a minimum of 5 oral health randomized controlled trials and used continuous outcomes. We extracted data, in duplicate, related to items of selection bias (sequence generation, allocation concealment, and baseline comparability) in the Cochrane Risk of Bias tool. Using a 2-level meta-meta-analytic approach with a random effects model to allow for intra- and inter-meta-analysis heterogeneity, we quantified the impact of selection bias on the magnitude of ES estimates. We identified 64 meta-analyses, including 540 randomized controlled trials analyzing 137,957 patients. Sequence generation was judged to be adequate (at low risk of bias) in 32% ( n = 173) of trials, and baseline comparability was judged to be adequate in 77.8% of trials. Allocation concealment was unclear in the majority of trials ( n = 458, 84.8%). We identified significantly larger treatment ES estimates in trials that had inadequate/unknown sequence generation (difference in ES = 0.13; 95% CI: 0.01 to 0.25) and inadequate/unknown allocation concealment (difference in ES = 0.15; 95% CI: 0.02 to 0.27). In contrast, baseline imbalance (difference in ES = 0.01, 95% CI: -0.09 to 0.12) was not associated with inflated or underestimated ES. In conclusion, treatment ES estimates were 0.13 and 0.15 larger in trials with inadequate/unknown sequence generation and inadequate/unknown allocation concealment, respectively. Therefore, authors of systematic reviews using oral health randomized controlled trials should perform sensitivity analyses based on the adequacy of sequence generation and allocation concealment.