We performed a pairwise meta-analysis of direct treatment effects comparing each treatment with the control group in the respective original study. Extracted data were combined using a random-effects model, with effect sizes expressed as incidence rate ratio (IRR) of developing HZ with corresponding 95% confidence interval (CI). Between-study heterogeneity in the meta-analysis was assessed using the I-squared metric, with values of < 25%, 25–50%, or > 50% indicating low, moderate, or high heterogeneity, respectively.
14,15
NMA was performed using a multivariate, random-effects meta-regression model with restricted maximum likelihood for variance estimation. Models were fitted using a frequentist approach in STATA (Version 14.0, StataCorp, College Station, TX, USA) through a network module based on the
‘mvmeta’ command for multiple treatment comparisons. Between-study variances were equalized, correlations were set to 0.5, and CIs were estimated based on asymptotic error variance and normal distribution. We used NMA to combine direct and indirect comparative data for the various therapeutic options across a network of studies into a single effect size and to rank the treatments. We first categorized the systemic anti-psoriatic interventions into eight groups, including corticosteroids (CS), conventional synthetic DMARDs (csDMARDs), tsDMARDs, biologics, phototherapy/acitretin, biologics + csDMARDs and/or CS, csDMRADs + CS, and control (
Figure 2(a)); summary results were presented as the IRRs of HZ for each group. To understand which treatments have a higher risk of HZ, we further stratified treatments into 12 groups, including CS, csDMARDs, JAKi (tofacitinib), PDE4 inhibitor, infliximab, anti-TNF-α (except infliximab), anti-IL-17, -23 or -12/23, phototherapy, acitretin, anti-TNF-α + csDMARDs and/or CS, csDMRADs + CS, and control (
Figure 2(b)). Since previous studies suggested that infliximab has a higher risk of serious infection than other anti-TNF-α drugs,
16,17 we isolated infliximab from the anti-TNF-α group in our analysis. Forest plots were used to represent quantitative results. To rank risk of HZ for all treatments, we calculated the surface under the cumulative ranking area (SUCRA) and quantified the ranks; larger SUCRA scores indicate a higher ranking, which suggests a higher probability of developing HZ.
18 Cumulative ranking probability plots were used to visually inspect ranking probabilities and uncertainty of various treatments.
18 Local inconsistency between direct and indirect estimates in the network was analyzed using a side-splitting model and global inconsistency was evaluated by comparing the fit of the consistency and inconsistency models using a design-by-treatment interaction model.
19,20 We used comparison-adjusted funnel plots in which treatments were ordered according to their point estimate and Egger’s tests to examine potential publication bias and the small study effect.
21