Skip to contents

Fit statistics for the additive component model

Source: R/diagnostics.R
additivity_test.Rd

Returns the Cochran Q statistics from netmeta::discomb().

Usage

additivity_test(object, ...)

Arguments

object

A cpaic_bridge or cpaic_fit object.

...

Unused.

Value

A one-row data frame with Q, df, pval (total additive-model lack of fit) and Q.diff, df.diff, pval.diff (the nested additivity test, NA on a disconnected network).

Details

Two different statistics are reported, and only one of them tests additivity:

  • Q (Q.additive) is the total lack of fit of the additive component model, pooling ordinary heterogeneity/inconsistency with any failure of additivity. It is not a test of additivity.

  • Q.diff = Q.additive - Q.standard is the nested test of the additivity restrictions themselves, and is the statistic to read. It exists only when a standard (non-additive) NMA is also estimable, i.e. on a connected network; on a disconnected network it is NA.

Neither statistic can test the assumption that actually bridges a disconnected network, namely that component effects (and, under population adjustment, component x effect-modifier interactions) are constant across sub-networks. There is by construction no cross-gap evidence against which to test it. A large p-value here is therefore not a licence to bridge; the assumption must be defended on clinical grounds (Veroniki et al. 2026).