← allmeta

NMA Global Inconsistency — LR test

Fit the NMA under the consistency model (one effect per basic parameter) and the full inconsistency model (one effect per design). The residual-Q difference is distributed as χ² and provides an omnibus test of consistency across the network.

Input

Contrasts — treatmentA, treatmentB, effect, SE, design(optional)

Model comparison

Model	Q	df	τ²	I²

Likelihood-ratio test

Per-design effect on the observed contrast

Design	Observed effect	Consistency-model prediction	Residual

Residual-by-design plot

Method: fit WLS consistency model (X' W X)⁻¹ X' W y with pooled τ². Fit "design model": effect = basic-parameter sum + design-specific deviation. Q_cons - Q_full has Chi-squared distribution on (n_designs - n_basic) df. R-parity: consistency / full-design / ΔQ match netmeta(common=TRUE) decomp.design() (Total / Within designs / Between designs) and the LR p uses an exact χ² CDF — verified to 1e-4 (hub/shared/tests/nma-global-inconsistency-parity.spec.mjs).