Approximate-Bayes NMA (WLS + MVN posterior)

Input

Contrasts — treatmentA, treatmentB, effect, SE Effect on any scale (log-OR, SMD, MD). Negative = A better if the outcome is bad. Reference treatment Used as the comparator in the output table. Preferred direction

Relative effects vs reference

vs ref	Estimate	SE	95% CrI	P(better than ref)

Treatment rankings (SUCRA)

Network diagram

Caveat: this is a consistency-only frequentist model with a Bayesian-style posterior approximation. It does not formally test loop inconsistency (use node-splitting or design-by-treatment in dedicated NMA software). Heterogeneity is captured via a pooled τ² across all contrasts; the DL estimator here is a univariate approximation to the NMA-specific version in netmeta. R-parity: on a consistent network (τ²=0) the WLS point estimates and SE match netmeta() to 1e-4 and the MVN-posterior SUCRA tracks netrank() P-score within Monte-Carlo error (hub/shared/tests/bayesian-nma-parity.spec.mjs).