INTRODUCTION: Owing to a scarcity of head-to-head trials of biologics for Crohn’s disease (CD), network meta-analysis (NMA) has been used to assess the comparative efficacy of biologics by combining evidence from randomized controlled trials (RCTs). Broadly, there are 2 NMA modeling approaches: the frequentist, which assumes that model error follows a standard distribution and the Bayesian, which estimates the exact distribution from the available data. Here, we describe frequentist and Bayesian models used to evaluate biologics for inducing clinical remission and response in biologic-naïve patients with moderate to severe CD.
METHODS: Studies included in the NMAs (12 RCTs of 9 interventions) were from a subset of studies identified by a systematic literature review. Most RCTs defined clinical remission as Crohn’s Disease Activity Index (CDAI) < 150 and response as ≥ 100-point decrease in CDAI after 4–12 weeks of treatment. In RCTs with results from multiple time points, data from the time point representative of the end of the induction period (as per the Food and Drug Administration label for the biologic) were used. For each endpoint, 4 random-effects NMAs were conducted; 3 used frequentist models and 1 used a Bayesian model. Treatment effects were compared using odd ratios (ORs); 95% confidence intervals (CIs) and credible intervals (CrIs) were estimated for frequentist and Bayesian NMAs, respectively. Sensitivity analyses were performed with fixed-effects models.
RESULTS: For clinical remission and response, the CIs for ORs (biologic vs placebo) from the frequentist models differed from each other and from the CrIs generated by the Bayesian model (Figure 1). In the Bayesian NMA, pairwise comparisons favoring infliximab-based interventions over placebo for remission reached statistical significance (p < 0.05) but the ORs were characterized by the widest CrIs in the network. No other pairwise comparison reached statistical significance for either clinical remission or response. Results from the sensitivity analyses were broadly consistent with those of the primary analyses.
DISCUSSION: The differences in the CIs and CrIs from the frequentist models and the Bayesian model suggest that the distributional assumptions needed by the frequentist approach may be inappropriate for the network of trial evidence. The Bayesian NMA did not reveal superiority of any biologic over another for induction of clinical remission or response in biologic-naïve patients with moderate to severe CD.
