SeqNMA: sequential network meta-analysis with monitoring boundaries — a methods and software demonstration on a synthetic network

Abstract

Background: Trial sequential analysis (TSA) applies group-sequential monitoring boundaries to a single cumulative pairwise meta-analysis, controlling type-I error as evidence accrues. Sequential monitoring has been extended to network meta-analysis (NMA) so that a network estimate can be declared conclusive once its cumulative z-score crosses an appended boundary (Nikolakopoulou et al., 2018). SeqNMA is an open, browser-based implementation of that idea, intended as a teaching and prototyping tool.

Methods: An in-browser engine rebuilds a contrast-based random-effects NMA at each chronological study addition. For every pairwise comparison it computes a required information size from the heterogeneity-adjusted diversity design effect D\xc2\xb2 (Wetterslev et al., 2009; the diversity, not the cluster, design effect), and monitors the cumulative network z-score against an O'Brien-Fleming alpha-spending efficacy boundary z_k = z_\xce\xb1 / \xe2\x88\x9a(information fraction), Bonferroni-corrected across all T(T\xe2\x88\x921)/2 comparisons, with a non-binding beta-spending futility wedge. The boundary mathematics were independently re-derived and numerically confirmed.

Results (illustrative, synthetic): On a deliberately synthetic, reproducibly seeded 12-study network of four treatments, the A-versus-C comparison crosses its monitoring boundary after study 10 (cumulative z = 3.34, log-odds-ratio 0.40), while A-versus-D remains inconclusive at 66% of its required information (z = 0.63). These numbers demonstrate the machinery; they are not an empirical finding about any treatment.

Conclusions: Browser-based sequential NMA makes accumulating-evidence monitoring transparent and reproducible. Two limitations are essential: the method ASSUMES but does not TEST network consistency (no design-by-treatment interaction or node-splitting), so every crossing is valid only if direct and indirect evidence agree; and the worked example is synthetic, a software demonstration rather than evidence about real interventions.