
Study or design influence on a joint network fit
network_influence.RdRefit after omitting whole studies or all studies with a common treatment design, and summarize the resulting change in treatment estimates.
Arguments
- object
A fitted joint network model; for the summary method, the summary data frame returned by
network_influence().- unit
studyomits one independent study at a time;designomits all studies sharing the same observed set of treatments.- x
The summary data frame returned by
network_influence().- ...
Additional arguments passed to
print.data.frame().- digits
Number of decimal places, from 1 to 22. P-values below 0.0001 print as
<0.0001; other p-values use at least four decimal places.- detailed
Include the stored
estimablecolumn; defaultFALSE.- row.names
Display row names; default
FALSE.
Value
A list containing:
- summary
One row per omission with
omitted,unit,estimable,reason, andmax_effect_change.estimablemeans refitting and comparison with all original prediction targets succeeded.- fits
Named list of successfully fitted omission models. A fit can be present even if some original prediction targets could not be evaluated.
- unit
The requested omission unit.
The summary inherits from nma_influence_summary and data.frame. Its
default printout hides estimable while retaining all omission rows and their
reasons. The flag remains accessible as $summary$estimable;
print(result$summary, detailed=TRUE) displays it. Printing changes no data.
No multiple-testing adjustment or automatic study selection is performed.
Details
Omissions remove complete study covariance blocks, preserving alignment of the remaining observations, strata and grouping IDs. The same mean/covariance model is refitted. Heterogeneity is re-estimated unless originally fixed. Predictions are evaluated at every original reported treatment-pair/mean-stratum row, with moderators at the original summary values (normally zero).
For omission \(g\), report
\(\max_j|\hat d_j^{(-g)}-\hat d_j|\), where j ranges over the original
reported treatment comparisons and strata. This is an unstandardized maximum
absolute change on the model's effect scale. It does not summarize changes in
SEs, intervals or heterogeneity; inspect the returned fits to examine those.
It is not Cook's distance, a p-value, or a threshold for study removal. Maximum
changes across differently scaled outcomes need separate interpretation.
Large changes can arise from an informative study, sparse connectivity or a scientifically different design, without indicating an error. Evaluate influence with study quality, clinical relevance and uncertainty. Omissions that disconnect a network, remove a required reference/stratum, make covariance unidentified or fail optimization are flagged. Such failures do not establish incoherence.
Examples
d <- data.frame(study=1:9,a=rep(c("A","A","B"),3),
b=rep(c("B","C","C"),3),effect=c(.1,.5,.2,.4,.8,.1,.2,.4,.4),variance=.04)
fit <- network_meta(d,"study","a","b","effect","variance")
network_influence(fit)$summary
#> omitted unit reason max_effect_change
#> 1 study 0.0476190
#> 2 study 0.0095238
#> 3 study 0.0190476
#> 4 study 0.0380952
#> 5 study 0.0761905
#> 6 study 0.0476190
#> 7 study 0.0190476
#> 8 study 0.0380952
#> 9 study 0.0380952
network_influence(fit,unit="design")$summary
#> omitted unit reason max_effect_change
#> A | B design 0.0666667
#> A | C design 0.0666667
#> B | C design 0.0666667