Matching-Adjusted Indirect Comparison (MAIC)

• MAIC is a population-adjusted indirect comparison method that reweights individual patient data to match aggregate covariate profiles from a comparator study.
• It employs exponential tilting to balance baseline covariates across studies, yielding a marginal treatment effect estimate in the target population.
• Key diagnostics such as effective sample size and geometric balance checks are essential to assess overlap and ensure stable inference in MAIC analyses.

Searching arXiv for MAIC and population-adjusted indirect comparison papers. Matching-Adjusted Indirect Comparison (MAIC) is a population-adjusted indirect comparison method used when two treatments have not been evaluated in a head-to-head trial, individual patient data (IPD) are available for one study, and only aggregate-level data (AgD) are available for the comparator study. In the canonical anchored setting, one trial compares $A$ versus a common comparator $C$ with IPD available, another compares $B$ versus $C$ with only AgD available, and MAIC reweights the IPD trial so that the weighted distribution of selected baseline covariates matches the comparator trial. The resulting weighted analysis estimates a marginal treatment effect for the target population defined by the AgD study, which can then be combined with the comparator study effect through an anchored indirect comparison (Remiro-Azócar, 2022, Remiro-Azócar, 2021). Within recent health technology assessment terminology, MAIC is typically treated as a member of the broader class of population-adjusted indirect comparisons (PAICs), alongside simulated treatment comparison (STC) (Baschet et al., 11 May 2026).

1. Definition, setting, and role in indirect comparisons

MAIC is designed for indirect treatment comparison when direct head-to-head evidence is unavailable and trial populations differ in effect modifiers. The standard use case is an anchored two-trial configuration: an $AC$ trial with IPD and a $BC$ trial with AgD. The aim is to estimate the effect of $A$ versus $B$ in the population of the $BC$ trial by first estimating the effect of $A$ versus $C$0 after reweighting the $C$1 IPD to resemble the $C$2 population, then combining this with the $C$3 trial’s effect estimate via an anchored contrast (Remiro-Azócar, 2022, Remiro-Azócar et al., 2020).

The method is positioned as a second-line option within a broader hierarchy of indirect comparison methods. A recent methodological guide for French health technology assessment places network meta-analysis first when a connected network exists, then subgroup-restricted network meta-analysis, then multilevel network meta-regression (ML-NMR), then external control arm approaches in non-randomized settings, followed by anchored PAICs such as MAIC and STC, with unanchored PAICs regarded as lower-evidence approaches (Baschet et al., 11 May 2026). This suggests that MAIC is generally used when conventional network synthesis is infeasible because only one trial provides IPD or because the network is too sparse for more integrated modeling.

MAIC has also been described as the weighting analogue of propensity-score weighting in a single observational study. In this analogy, the treatment-assignment model of causal inference is replaced by a trial-assignment model: patients in the IPD study are reweighted so that their covariate distribution mimics that of the target AgD study (Campbell et al., 2023). This characterization is technically important because it clarifies both the strengths of MAIC and its limitations under poor overlap, model misspecification, and non-collapsible effect measures.

2. Weighting construction and target estimands

The core MAIC construction assigns a non-negative weight to each IPD subject so that weighted covariate moments in the IPD trial match the reported covariate moments in the comparator trial. A standard formulation is

$C$4

with $C$5 chosen so that the weighted means of selected covariates equal the AgD means of the comparator study (Green et al., 12 Jan 2026). Equivalent formulations appear throughout the literature, including method-of-moments and entropy-balancing interpretations (Remiro-Azócar, 2022, Glimm et al., 2021).

In one common specification, if $C$6 denotes the covariate means reported in the comparator trial, the weights satisfy moment constraints of the form

$C$7

or, equivalently, after centering on $C$8,

$C$9

(Green et al., 12 Jan 2026). In the original anchored MAIC formulation summarized in later methodological work, the log-weight is written as the log-odds of belonging to the target trial rather than the index trial conditional on selected effect modifiers, and the parameters are obtained by minimizing a convex objective (Remiro-Azócar, 2022).

The resulting weighted analysis is typically a univariable weighted regression of outcome on treatment in the IPD trial, with no covariates in the outcome model. For binary outcomes, this is often a weighted logistic regression; for time-to-event outcomes, a weighted Cox model; for continuous outcomes, a weighted linear regression (Remiro-Azócar, 2022, Green et al., 12 Jan 2026). The treatment coefficient from that weighted regression is then interpreted as the treatment effect of $B$0 versus $B$1 in the reweighted target population.

A central issue in the MAIC literature is the precise estimand. Multiple papers emphasize that standard MAIC, as usually implemented, targets a marginal treatment effect in the target population rather than a conditional effect. In a logistic-outcome setting, the weighted univariable regression coefficient is a marginal log-odds ratio in the reweighted population (Remiro-Azócar et al., 2020). This matters because indirect comparisons frequently use non-collapsible effect measures such as log-odds ratios and log-hazard ratios, for which marginal and conditional effects differ even under perfect randomization (Remiro-Azócar et al., 2020, Remiro-Azócar, 2021).

A further limitation is that standard pairwise MAIC targets the comparator-study population. A discussion of estimands for population-adjusted indirect comparisons argues that current pairwise methods such as MAIC are restricted to marginal estimands specific to the comparator study sample, which may not coincide with the decision-relevant population for health technology assessment (Remiro-Azócar, 2021). This suggests that MAIC’s practical target population is often determined more by data availability than by policy relevance.

3. Assumptions, identification, and overlap

Anchored MAIC relies on the conditional constancy of relative treatment effects across trials given the observed effect modifiers. In the terminology of recent HTA guidance, randomized trial-based indirect comparisons require justification of similarity, transitivity, and exchangeability, and failures of homogeneity or consistency often arise from imbalances in treatment-effect modifiers across studies (Baschet et al., 11 May 2026). Applied to MAIC, the method attempts to enforce similarity on observed effect modifiers, but exchangeability still depends on all relevant effect modifiers being measured and correctly represented in the weighting model.

The required covariates are often divided into prognostic factors and effect modifiers. In anchored settings, several commentaries emphasize that MAIC is necessary when there are cross-trial imbalances in effect modifiers; if no effect modifiers are imbalanced, the standard anchored indirect comparison is typically more precise and may be more accurate than MAIC because weighting only reduces effective sample size without removing bias (Remiro-Azócar et al., 2020). By contrast, for marginal estimands under non-collapsible measures, other work argues that unbalanced prognostic variables must also be adjusted for, because marginal odds ratios and hazard ratios depend on the covariate distribution even when those covariates do not modify treatment effects (Campbell et al., 2023). This reflects a genuine estimand distinction in the literature: one line of work prioritizes the traditional anchored assumption of effect-modifier adjustment, whereas another focuses on the identification of marginal non-collapsible measures and therefore emphasizes broader covariate adjustment.

In unanchored settings, the assumptions become substantially stronger. French HTA guidance states that unanchored MAICs are highly discouraged because it is very difficult to verify the assumption of conditional constancy of absolute effects (Baschet et al., 11 May 2026). A later methodological paper on unanchored meta-regression similarly groups MAIC with other unanchored PAIC methods and states that valid inference requires strong and often unverifiable assumptions including conditional exchangeability, correct outcome-model specification, and adequate overlap (Chandler et al., 18 Jun 2026). This suggests that unanchored MAIC should usually be regarded as supportive rather than definitive evidence.

Positivity or overlap is a separate requirement. MAIC can only reweight to the target population if the target covariate vector is representable by the IPD population. A geometric analysis formalizes this by showing that the aggregate target vector must lie in the convex hull of the IPD covariate vectors; otherwise the MAIC moment-matching equations have no solution (Glimm et al., 2021). Even when a solution exists, if the target lies near the boundary of the convex hull, the method may assign extreme weights to a few patients, producing a very small effective sample size and unstable inference (Glimm et al., 2021, Remiro-Azócar, 2022).

4. Effective sample size, feasibility, and diagnostics

Effective sample size (ESS) is one of the most prominent MAIC diagnostics. In anchored PAICs, recent HTA guidance states that presentation and interpretation of ESS are critical, that method choice should not solely rely on ESS, and that a substantial decrease signals underlying issues with population comparability (Baschet et al., 11 May 2026). In the MAIC context, ESS is conventionally defined as

$B$2

which measures how much information remains after weighting (Baschet et al., 11 May 2026, Remiro-Azócar, 2022).

A large reduction in ESS indicates that only a small subset of IPD subjects effectively matches the comparator population, implying limited covariate overlap and loss of precision. Several papers connect low ESS to wider confidence intervals, greater sensitivity to individual observations, and possibly undercoverage when sandwich-based variance estimators are used (Remiro-Azócar, 2022, Remiro-Azócar et al., 2020). French HTA guidance accordingly recommends that ESS be reported and discussed, rather than treated as a technical detail, and interpreted as a signal of limited overlap and fragile inference (Baschet et al., 11 May 2026).

Beyond ESS, numerical feasibility itself can be assessed geometrically. A dedicated study on feasibility recommends checking whether the AgD covariate vector lies within the convex hull of the IPD covariate cloud using linear programming. If not, no set of MAIC weights can achieve exact balance, and software may nevertheless return a spurious numerical solution with poor balance and extremely concentrated weights (Glimm et al., 2021). The same paper proposes principal component analysis as a visualization tool for locating the AgD point relative to the IPD cloud, and Mahalanobis-distance or Hotelling-type diagnostics to assess whether matching is even needed because the two samples may already be compatible (Glimm et al., 2021).

The shape of the weight function also matters. Because standard MAIC uses exponential tilting, weights increase monotonically along a direction of steepest ascent in covariate space, which tends to assign the largest weights to IPD subjects at the edge of the data cloud rather than to those nearest the target mean. This was proposed as one contributing factor to contradictory MAIC results obtained by two different sponsors each using their own IPD and the other sponsor’s AgD (Glimm et al., 2021). A plausible implication is that MAIC can define subtly different effective populations depending on which trial provides IPD, even when each analysis satisfies its own balancing constraints.

5. Anchored and unanchored use, performance, and variants

Simulation studies consistently distinguish anchored from unanchored MAIC. In a large simulation study for survival outcomes with the log hazard ratio as effect measure, MAIC was found to yield unbiased treatment-effect estimates under no failures of assumptions, whereas standard unadjusted indirect comparisons were systematically biased under stronger effect-modifier imbalance and STC was biased because it typically targets a conditional rather than marginal effect (Remiro-Azócar et al., 2020). The same study reported that robust sandwich variance estimators for MAIC can underestimate variability when ESS is small (Remiro-Azócar et al., 2020).

A commentary on effect modification in anchored indirect comparisons emphasizes that anchored MAIC is needed when there are cross-trial imbalances in effect modifiers, but that standard indirect comparisons have greater precision and accuracy if there are no effect modifiers in imbalance (Remiro-Azócar et al., 2020). This is an important corrective to the common misconception that covariate adjustment is always beneficial: in an anchored randomized setting, weighting without genuine effect-modifier imbalance merely trades precision for no reduction in bias.

Other simulation work focuses on the interaction between positivity violations and covariate distribution shape. A 2025 Monte Carlo study compared MAIC matching first moments only (MAIC-1), MAIC matching first and second moments (MAIC-2), and a full-IPD propensity score weighting benchmark across normal, lognormal, and bimodal covariate scenarios. It reported that MAIC-1 remained unbiased in the presence of moderate positivity violations and non-normal covariates, whereas MAIC-2 and the full-IPD weighting benchmark appeared more sensitive to positivity violations (Serret-Larmande et al., 16 Jul 2025). The same study found that omission of key confounders produced substantial bias in all methods, reinforcing that correct covariate specification dominates the choice among weighting estimators (Serret-Larmande et al., 16 Jul 2025).

Several methodological extensions have been proposed. A two-stage extension, 2SMAIC, augments standard MAIC with a treatment-assignment model within the IPD trial so that the final weights balance covariates both across studies and between treatment arms within the IPD study. In simulation studies across two randomized trials, 2SMAIC improved precision and efficiency relative to standard MAIC in all scenarios while maintaining similarly low bias; gains were largest when the IPD trial was small, because the second stage corrected chance imbalances in prognostic covariates between study arms (Remiro-Azócar, 2022). However, when overlap between trial populations was poor and weights were extreme, the benefits were attenuated and truncation introduced a substantial bias–variance trade-off (Remiro-Azócar, 2022).

Another extension is augmented or doubly robust MAIC for externally controlled single-arm trials and unanchored indirect comparisons. A 2025 paper develops an estimator that combines entropy-balancing MAIC weights with a conditional outcome model, yielding a doubly robust estimator that is consistent if either the weighting model or the outcome model is correctly specified and that shows higher precision than non-augmented weighting estimators in binary-outcome simulations (Campbell et al., 30 Apr 2025). This paper characterizes standard MAIC as singly robust when the outcome model is non-linear and argues that augmentation improves both protection against misspecification and efficiency (Campbell et al., 30 Apr 2025).

6. Methodological debates, alternatives, and target populations

Several debates structure the current MAIC literature. The first concerns marginal versus conditional estimands. A 2020 commentary argues that MAIC targets a marginal effect and that simulation studies evaluating MAIC against a conditional truth in logistic or Cox models can falsely label it biased because of estimand mismatch due to non-collapsibility (Remiro-Azócar et al., 2020). A closely related commentary on anchored survival-data simulations makes the same point, noting that MAIC targets a marginal or population-average treatment effect, while some simulation studies define the truth as a conditional treatment coefficient from a Cox model (Remiro-Azócar et al., 2020). These arguments suggest that MAIC’s performance can only be judged relative to the estimand it is designed to estimate.

A second debate concerns the selection of covariates to balance. One line of work emphasizes that subject-matter knowledge is indispensable because effect-modifier identification is a low-dimensional but high-stakes problem: omitting a true effect modifier biases the indirect comparison, whereas including unnecessary covariates mainly costs precision (Remiro-Azócar et al., 2020). Another line, focused on marginal non-collapsible estimands, argues that purely prognostic factors must also be adjusted for in anchored settings if the goal is a marginal odds ratio or hazard ratio (Campbell et al., 2023). This suggests that covariate selection depends both on the underlying clinical science and on the target estimand.

A third debate concerns weighting versus outcome modeling. Outcome-modeling approaches such as STC or more general G-computation can be more efficient than MAIC when correctly specified, especially under poor overlap, because they extrapolate rather than relying entirely on weighted common support (Remiro-Azócar et al., 2023, Campbell et al., 2023). However, several methodological discussions stress that model-based extrapolation is assumption-heavy and difficult to validate in regions with limited overlap (Remiro-Azócar et al., 2023). Weighting approaches such as MAIC are viewed by some authors as more bias-robust because the balancing target is explicit and directly checkable (Remiro-Azócar et al., 2023), whereas others advocate doubly robust hybrids that combine weighting and outcome modeling (Campbell et al., 30 Apr 2025, Remiro-Azócar et al., 2023).

A fourth debate concerns the target population. A paper on target estimands for population-adjusted indirect comparisons argues that HTA agencies make reimbursement decisions at the population level and therefore require marginal estimands in explicitly relevant target populations. It contends that pairwise methods such as MAIC are restricted to comparator-study-specific marginal estimands, which may not coincide with the decision-relevant real-world population (Remiro-Azócar, 2021). In contrast, ML-NMR can standardize conditional estimates to any population of interest provided its covariate distribution is available (Remiro-Azócar, 2021).

This target-population issue is central to the so-called “MAIC paradox.” A 2025 paper on arbitrated indirect treatment comparisons describes the situation in which two sponsors each perform MAIC using their own IPD and the other sponsor’s AgD, thereby targeting different populations and potentially reaching opposite conclusions about which treatment is superior (Fang et al., 20 Oct 2025). To address this inconsistency, the paper proposes overlap-population targeting through trial-membership overlap weights, so that both analyses are conducted in a common overlap population rather than in each sponsor’s preferred comparator-trial population (Fang et al., 20 Oct 2025). This suggests that contradictory MAIC conclusions may arise not only from poor overlap or weight instability but also from inconsistent target estimands.

More recent work extends these concerns to unanchored settings. A Bayesian framework for multilevel unanchored meta-regression treats MAIC and STC as pragmatic pairwise tools that usually estimate marginal effects in the comparator study population and then often assume, implicitly, that these effects apply to a different decision-relevant population. Simulation studies in that paper show that MAIC can recover comparator-population effects well yet still be biased when those effects are naively transported to another population because marginal risk ratios and odds ratios are population-specific (Chandler et al., 18 Jun 2026). This suggests that even when MAIC works as intended, careless post hoc transportability assumptions can induce additional bias.

7. Reporting, health technology assessment, and software

Recent HTA guidance recommends early planning of indirect treatment comparisons, ideally prospectively in the statistical analysis plan, especially where HTA submission is foreseeable (Baschet et al., 11 May 2026). In the MAIC context, this means collecting baseline covariates that are likely to be reported across competitor trials, pre-specifying potential effect modifiers and prognostic factors, and anticipating ESS loss in sample-size planning (Baschet et al., 11 May 2026). The same guidance emphasizes transparent explanation of heterogeneity, similarity, exchangeability, overlap, and ESS, and notes that French HTA reviewers frequently criticize submissions for lack of homogeneity (Baschet et al., 11 May 2026).

Reporting recommendations derived from that guidance include explicitly labeling analyses as anchored or unanchored, stating assumptions about constancy of relative or absolute effects, reporting unweighted sample size, ESS, and the impact of weighting on uncertainty, and presenting balance diagnostics before and after weighting (Baschet et al., 11 May 2026). When unanchored MAIC is used, it should be positioned as supportive evidence with extensive caveats and sensitivity analyses (Baschet et al., 11 May 2026).

Software support has broadened. The 2026 outstandR package in R implements MAIC, STC, maximum-likelihood and Bayesian G-computation, and multiple imputation marginalization in a unified framework for IPD–AgD population-adjusted indirect comparison (Green et al., 12 Jan 2026). In the package, MAIC is implemented through exponential tilting to match AgD covariate means, followed by weighted outcome analysis and indirect contrast construction on the selected effect scale (Green et al., 12 Jan 2026). The package also exposes MAIC weights and ESS, facilitating standard diagnostics. This suggests a movement from fragmented MAIC implementations toward standardized PAIC workflows, although the underlying methodological concerns remain unchanged.

8. Summary

MAIC is a weighting-based population-adjusted indirect comparison method that uses IPD from one trial and AgD from another to estimate a marginal treatment effect in the comparator-study population. Its core operation is to assign subject-specific weights so that weighted baseline covariate moments in the IPD trial match those reported for the target AgD trial, after which a weighted treatment-only analysis yields a marginal effect estimate compatible with anchored indirect comparison (Remiro-Azócar, 2022, Remiro-Azócar et al., 2020).

The method is most defensible in anchored randomized settings with credible measurement of all relevant effect modifiers, adequate overlap, and explicit recognition that the estimand pertains to the comparator population unless a broader transportability argument is made (Baschet et al., 11 May 2026, Remiro-Azócar, 2021). ESS, balance diagnostics, and geometric feasibility checks are central to interpretation because extreme or unstable weights signal limited overlap and fragile inference (Glimm et al., 2021, Baschet et al., 11 May 2026). In unanchored settings, the required assumptions are much stronger and recent HTA guidance explicitly discourages unanchored MAIC except as limited supportive evidence (Baschet et al., 11 May 2026).

Methodological work since 2020 has sharpened three points. First, MAIC should be evaluated against the marginal estimands it targets, not against incompatible conditional truths under non-collapsible measures (Remiro-Azócar et al., 2020, Remiro-Azócar et al., 2020). Second, MAIC is only one member of a broader class of PAIC methods, and more integrated approaches such as ML-NMR or newer unanchored regression frameworks can target effects in arbitrary populations and larger treatment networks (Remiro-Azócar, 2021, Chandler et al., 18 Jun 2026). Third, MAIC’s practical performance is governed less by the elegance of its weighting equations than by overlap, covariate specification, and target-population choice; recent variants such as 2SMAIC and augmented MAIC seek to improve efficiency and robustness, but they do not remove the underlying dependence on exchangeability and positivity (Remiro-Azócar, 2022, Campbell et al., 30 Apr 2025).

In that sense, MAIC remains a technically important and widely used HTA method, but one whose validity depends on transparent estimand definition, careful covariate selection, explicit overlap diagnostics, and disciplined interpretation of what population the analysis actually represents.