The method landscape

What each approach identifies, where it breaks, and how to choose

The catalog is organized by problem. This page is organized by method: which approach answers which question, what each one assumes, and where each one fails. It exists so an entry in the catalog can be read against the method it constrains.

The distinction that organizes everything

The decisive division is anchored versus unanchored evidence, not matching versus regression versus multilevel modeling.

With an anchor, randomization still identifies the within-trial relative effect, and population adjustment is about transporting effect modification. Without one, the analyst is reconstructing a causal contrast across studies: every prognostic factor and effect modifier that creates a cross-study outcome difference becomes relevant, and residual confounding is generally not testable.

A richer model does not move a problem across that line.

Choosing a method

flowchart LR
    A[Decision problem] --> B{Connected randomised network?}
    B -- Yes --> C{Effect modifiers plausibly balanced?}
    C -- Largely yes --> D[Standard ITC / NMA]
    C -- No, IPD available in subset --> E[Anchored PAIC]
    E --> F[ML-NMR preferred when feasible]
    E --> G[MAIC or STC only with explicit diagnostics]
    B -- No --> H{Single-arm or disconnected evidence?}
    H -- Yes --> I[Unanchored comparison]
    I --> J{Conditional constancy plausible?}
    J -- Weak --> K[Do not make strong causal claims]
    J -- Strong but uncertain --> L[Unanchored MAIC / STC / ML-UMR plus aggressive sensitivity analysis]
    H --> M[Alternative linking: ALM, baseline RE, dose-response MBNMA, class effects]

The branch that matters most is the one ending in “do not make strong causal claims.” That is a legitimate destination, not a failure of the analysis.

Where the bias enters

flowchart TB
    X[Measured baseline covariates] --> S[Study membership]
    X --> Y[Outcome]
    Z[Effect modifiers] --> T[Treatment effect]
    Z --> S
    T --> Y
    U[Unmeasured design / care differences] --> S
    U --> Y
    O[Overlap quality] --> W[Weight stability / extrapolation]
    W --> Y

Read this for what it says about \(U\). Weighting and outcome regression both act on \(X\) and \(Z\). If the determinants of the cross-study outcome difference sit in \(U\) instead, no amount of balancing on \(X\) repairs the comparison; at best it disguises the bias more elegantly. That observation is why quantitative bias analysis is the largest category in this catalog.

What each method identifies

Method	Evidence structure	Nominal estimand	Main identifying assumptions	Principal unresolved weakness
Bucher adjusted ITC	Two connected RCTs, common comparator	Relative effect in an implicitly common population	Transitivity; compatible effect scale; no material modifier imbalance	No explicit target population; fragile to modifier imbalance and design differences
Aggregate-data NMA	Connected RCT network	Network relative effects	Homogeneity, transitivity, consistency	Population mismatch and effect modification often only weakly assessed
IPD-NMA	IPD from all or many trials	Conditional or standardized effects	Correct individual-level model; exchangeability; IPD availability mechanism	IPD availability is selective; high-dimensional interactions remain difficult
Anchored MAIC	IPD index trial, AgD comparator, common comparator	Usually a marginal effect in the AgD trial population	All relevant modifiers balanced; overlap; correct weighting moments	ESS collapse, incomplete moments, target ambiguity, non-collapsibility, second-stage transport
Anchored STC	IPD index trial, AgD comparator, common comparator	Often a conditional effect at target means unless standardized	Correct outcome model; all relevant modifiers included; valid extrapolation	Model misspecification; conditional and marginal incompatibility; target joint distribution unavailable
Parametric g-computation	IPD source plus target summaries	Compatible marginal effect in a specified target	Correct outcome model and target-distribution model	Sensitive to misspecification; target-summary uncertainty usually ignored
ML-NMR	Mixed IPD and AgD in a connected network	Population-specific effects from one hierarchical model	Conditional constancy; individual-level model; integration distribution	Computational burden; partial identifiability; within and between interaction conflation
Unanchored MAIC or STC	Disconnected trials or single-arm studies	Cross-study contrast in a chosen target	All prognostic factors and modifiers that differ across studies measured and correctly adjusted	No randomization anchor; extreme vulnerability to omitted variables and design heterogeneity
Baseline synthesis, reference prediction	Disconnected AgD, possibly single-arm	Relative effects linked through baseline-outcome structure	Exchangeable or modelled baseline effects; correct arm-level model	Baseline risk is not a universal bridge; ecological and design confounding remain
Dose-response MBNMA	Treatments linked through dose or exposure	Effects along a shared dose-response function	Common pharmacologic curve form and modifiers	A wrong curve or class assumption produces precise but artificial links
Class-effect NMA	Treatments grouped into classes	Shrunk treatment effects and class contrasts	Within-class exchangeability; correct class variance	Class definitions are contestable and may hide genuine heterogeneity
Additive CNMA	Multicomponent treatments sharing components	Component and reconstructed treatment effects	Additivity; stable component effects across subnetworks	Cross-gap component constancy; synergy and treatment-version differences may invalidate the bridge
Component ML-NMR	Component network with mixed IPD and AgD	Component-adjusted contrasts at target covariate values	CNMA assumptions plus transportability and stable component-by-modifier interactions	Strong, partly untestable structural assumptions; marginal standardization incomplete
Doubly robust transport	Source IPD plus target IPD or adequate summaries	Target-population marginal effect	Conditional exchangeability plus either a correct weight model or a correct outcome model	No protection against unmeasured modifiers, positivity failure, or both models being wrong
Partial identification	Any structure with uncertain assumptions	Bounds or a bias-adjusted set	Explicit bounds rather than point identification	Requires defensible sensitivity parameters; less convenient for decision models

This table is not a ladder. A richer model helps only when its added structure is scientifically defensible and adequately informed by data. A faster sampler does not cure non-identification, a doubly robust estimator does not cure positivity failure, perfect covariate balance does not cure unmeasured effect modification, and a narrow interval does not establish that the target population was the right one.

How to evaluate a method

Simulation studies in this field routinely report bias without naming the population-specific estimand the bias is relative to, which makes the number uninterpretable. A usable evaluation covers six families at minimum.

Metric family	Minimum metrics
Estimand recovery	Bias, relative bias, RMSE, mean absolute error, sign error, target-standardized bias
Uncertainty calibration	95% coverage, empirical to nominal standard-error ratio, posterior calibration, type I error, power
Decision quality	Incremental net-benefit decision error, wrong-treatment probability, rank-reversal frequency
Overlap and weighting	ESS, ESS over nominal \(n\), coefficient of variation of weights, maximum weight, entropy, proportion trimmed
Model adequacy	Calibration slope, posterior predictive checks, leave-one-study-out predictive error, divergence counts, \(\hat{R}\), effective posterior draws
Computation and reproducibility	Runtime, failure rate, seed sensitivity, sensitivity to starting values, container reproducibility

The decision-quality row is the one most often omitted. A method that is slightly more biased but better calibrated for a reimbursement decision may be preferable to one that looks efficient but is anti-conservative when overlap is weak.

flowchart LR
    A[Define estimand] --> B[Data-generating mechanism]
    B --> C[Connected / disconnected evidence]
    C --> D[Manipulate: overlap, modifiers, misspecification]
    D --> E[Fit competing methods]
    E --> F[Evaluate against the declared estimand]
    F --> G[Publish code, seeds, manifests]

Software

Verified against CRAN and GitHub during the audit; versions as at 25 July 2026.

Package	Status	Strengths	Cautions
`multinma`	CRAN 0.9.1	The most mature open Bayesian NMA, NMR, and ML-NMR platform; mixed IPD and AgD; target integration; class effects	Runtime and integration burden; disconnected-network support sits in an open, unmerged pull request
`maicplus`	CRAN 0.1.2	Maintained MAIC workflow; anchored and unanchored binary and survival analyses	Open needs include continuous and count outcomes, trimming, weight-estimation-aware variance, and non-proportional-hazards measures
`netmeta`	CRAN 3.6-1	Mature frequentist NMA; component NMA through `discomb`; rare-event methods	Not a population-transport platform
`outstandR`	Active	Compatible marginal effects by model-based standardization	Pairwise focus; sensitive to the target distribution and outcome model
`drMAIC`	CRAN 0.1.0	Early doubly robust MAIC with diagnostics	New; requires independent audit and adversarial simulation
`gemtc`, `MBNMAdose`, `crossnma`	CRAN	Bayesian NMA, dose-response linking, cross-design synthesis	None has a native population-transport layer

Two packages that appear in older workflows are no longer maintained. maic was archived from CRAN on 22 July 2025 because the maintainer’s email had become undeliverable, and Roche/MAIC states in its own README that it “is no longer developed”, recommending maicplus instead. Legacy installs belong in reproducibility archaeology, not in a new workflow.

No package unifies target-population definition, connected and disconnected evidence, marginal and conditional estimands, doubly robust transport, component structure, complex outcomes, missing data, sensitivity analysis, and regulatory reporting. That gap is itself catalogued, under software and benchmarks.