Reporting guide

Which graphics each method needs, and which diagnostics apply where

No single graph can validate a meta-analysis or indirect comparison. The central rule is simple and frequently broken: pair every result plot with a plot of the assumption or data structure that makes the result credible. A forest plot without a heterogeneity diagnostic is incomplete. An NMA ranking without geometry, inconsistency checks, and effect estimates is incomplete. A MAIC estimate without its weights, effective sample size, and covariate support is incomplete. An ML-NMR result without effect-modifier and integration diagnostics is incomplete. An unanchored ML-UMR result without explicit sensitivity analyses is hard to defend.

Minimum graphical sets

These sets synthesize what the methodological literature and current software make possible. They are recommendations, not requirements of any single reporting guideline.

Method	What should normally be shown
Pairwise MA	Forest plot with prediction interval; a heterogeneity or influence diagnostic (Baujat, leave-one-out); a funnel plot only when there are enough studies; moderator plots when heterogeneity is explained
NMA	Network graph; relative effects as a forest, league table, or interval plot; an inconsistency diagnostic (node splitting, net heat); contribution or evidence flow when provenance matters
Bucher ITC	Anchored comparison diagram; transitivity plot of effect modifiers
MAIC	Balance before and after weighting; weight distribution; effective sample size; covariate support; the adjusted effect
STC and G-computation	Covariate support; outcome model functional form and residual diagnostics; marginal effects in the target population; extrapolation sensitivity
ML-NMR	Network and transitivity plots; effect-modifier curves; marginal effects in a stated target population; the integration error diagnostic; standard Bayesian diagnostics
NMI	Subgroup data availability; the interpolation relationship; the adjusted-effect forest
ML-UMR	Population support and prognostic-factor diagnostics; absolute outcome and marginal-effect plots; posterior and model checks; explicit sensitivity plots for unanchored assumptions

Which diagnostic applies to which population-adjustment method

Population-adjusted comparisons need a stricter graphical standard than conventional meta-analysis because validity depends on which covariates were adjusted, whether populations overlap, whether the model extrapolates, and whether an outcome or weighting model is credible. The table separates weighting, outcome-regression, integrated-likelihood, and subgroup-interpolation approaches.

Diagnostic	MAIC	STC / G-comp	ML-NMR	NMI	ML-UMR
Love / SMD balance	Core	Useful	Useful	Useful	Useful
Weight distribution	Core	No	No	No	No
Effective sample size	Core	No	No	No	No
Covariate distributions and support	Core	Core	Core	Core	Core
Selection-score overlap	If IPD allow	Occasionally	Rare	Rare	Rare
Residual and calibration plots	No	Core	Core	Limited	Core
Effect-modifier curves	Limited	Core	Core	Core	Core
Numerical integration error	No	No	Core	No	Implementation dependent
Subgroup interpolation	No	No	No	Core	No
Absolute outcome curves	Optional	Core	Strongly useful	Possible	Core
Target-population marginal effects	Limited	Core	Core	Core	Core
Strong-assumption sensitivity	Core if unanchored	Core if unanchored	Important	Important	Essential

Software landscape

R has by far the deepest purpose-built ecosystem: meta and metafor for pairwise meta-analysis, netmeta for frequentist NMA, multinma for Bayesian NMA and ML-NMR, orchaRd and metaviz for newer displays, maicplus for MAIC, and mlumr for ML-UMR. Stata is strong for standard meta-analysis and NMA through its official meta suite and the network and network graphs packages (network map, network sidesplit, ifplot, netfunnel, intervalplot, netleague, sucra, mdsrank, clusterank). Python is much less specialized: statsmodels covers basic meta-analysis and forest plots, and most other graphics must be built with matplotlib, PyMC, and ArviZ. Generic plotting code should not be presented as if it were validated method-specific software.