Node-splitting forest plot

Side-split plot, direct versus indirect forest, NMA composite forest

NMA
Direct, indirect, and network estimates for every comparison informed by both kinds of evidence, stacked as forest-plot rows.
NMAEstablished

Node-splitting forest plot example

Node-splitting (SIDDE) forest plot for the diabetes network: for each comparison with both direct and indirect evidence, the direct estimate, the indirect estimate, and the network estimate, with the number of direct trials, the direct evidence proportion, and I². Data: netmeta::Senn2013.
Family
Inconsistency
Purpose
Check local consistency by comparing direct and indirect evidence comparison by comparison.
Inputs
A fitted network meta-analysis.
Software
R netmeta::netsplit() + forest(), multinma (consistency = "nodesplit"), gemtc::mtc.nodesplit(); Stata network sidesplit

What it shows

Node splitting (Dias and colleagues) separates, for one comparison at a time, the direct evidence from the indirect evidence supplied by the rest of the network. The forest plot stacks the direct estimate, the indirect estimate, and the combined network estimate for each comparison. Disagreement between the first two is local inconsistency. The same layout shows at a glance where each network estimate comes from, which is why it is sometimes called an NMA composite forest.

How to read it

  • Groups: comparisons that have both direct and indirect evidence.
  • Rows within a group: direct estimate, indirect estimate, and network estimate (diamond).
  • Columns: number of direct trials, direct evidence proportion, and within-comparison \(I^2\).
  • Look for: direct and indirect intervals that do not overlap.

Interpretation

Most pairs agree closely. The largest discrepancies are metformin versus sulfonylurea (direct −0.37 against indirect −0.99) and rosiglitazone versus sulfonylurea, both informed by a single direct trial, with p-values of 0.19 and 0.16 for the difference. With intervals this wide, the absence of significant differences is weak reassurance.

Pitfalls

  • A non-significant difference does not establish consistency; direct and indirect estimates are often imprecise.
  • Testing every comparison creates multiplicity.
  • With multi-arm trials, the split can be defined in several ways (for example, SIDDE in netmeta versus the Dias approach), which changes results.
  • Loops share evidence, so node-split tests are not independent.

Code

library(netmeta)

data(Senn2013)

net <- netmeta(
  TE, seTE, treat1.long, treat2.long, studlab,
  data = Senn2013, sm = "MD",
  common = FALSE, reference.group = "Placebo"
)

# Separate direct and indirect evidence for every comparison informed by both
ns <- netsplit(net)
forest(ns, show = "both", nchar.trts = 6,
       col.square = "#1d4e89", col.diamond = "#1d4e89",
       col.square.lines = "#1d4e89")
network meta consistency
network sidesplit all, tau

References

  • Dias S, Welton NJ, Caldwell DM, Ades AE. Checking consistency in mixed treatment comparison meta-analysis. Stat Med. 2010;29:932-944. doi:10.1002/sim.3767
  • König J, Krahn U, Binder H. Visualizing the flow of evidence in network meta-analysis and characterizing mixed treatment comparisons. Stat Med. 2013;32:5414-5429. doi:10.1002/sim.6001