Does the individual patient data actually inform this contrast?

Population adjustment only helps if the adjusted edges actually carry the contrast you care about. They need not. In a component bridge the estimate of a contrast m' beta is a weighted combination of the observed edges,

Usage

edge_influence(object, treatment, comparator = NULL, tol = 1e-08, ...)

Arguments

object: A cpaic_bridge, cpaic_maic or cpaic_stc object.
treatment, comparator: The contrast of interest. comparator defaults to the network reference.
tol: Influence weights below this (relative to the largest) are treated as zero.
...: Unused.

Value

A data frame with one row per edge: studlab, treat1, treat2, has_ipd, and influence (the weight w_j). Edges are ordered by absolute influence. A warning is issued if any IPD edge has no influence on the requested contrast.

Details

$$m'\hat\beta = \underbrace{m' (X'WX)^{+} X'W}_{w} \, d ,$$

so edge j influences the answer only through its weight w_j. An IPD edge with w_j of zero contributes nothing to that contrast, and adjusting it changes nothing.

The weight uses a diagonal W of inverse edge variances. The fit itself is produced by netmeta::discomb(), which accounts for the within-study covariance of a multi-arm trial, so in a network containing multi-arm studies the weight reported here is a close approximation to the fitted estimator's influence rather than its exact value. It is intended as a screening diagnostic, to flag an IPD edge that carries little or no weight on the contrast; read a weight near zero as "this edge barely matters here", not as an exact sensitivity.

This matters because the usual diagnostic cannot detect the problem. In simulation, putting the IPD on an edge that does not bridge the gap left cMAIC numerically identical to the unadjusted analysis (bias +0.374, coverage 0.676) while effective_sample_size() happily reported an ESS of 999 out of 1000. A healthy ESS says the weights are well behaved; it says nothing about whether the reweighted edge is relevant to your estimand.

Examples

net <- cpaic_network(cpaic_bin_agd, sm = "OR", inactive = "Placebo")
br <- cnma_bridge(net)
edge_influence(br, treatment = "A+B+C")
#>   studlab treat1  treat2 has_ipd influence
#> 1      S1      A Placebo   FALSE 1.0000000
#> 2      S2      B Placebo   FALSE 1.0000000
#> 3      S3  A+B+C     A+B   FALSE 0.7065349
#> 4      S5  A+B+C   A+B+D   FALSE 0.2934651
#> 5      S4  A+B+D     A+B   FALSE 0.2934651