Matching-Adjusted Indirect Comparison (MAIC), Step by Step

Matching-adjusted indirect comparison (MAIC) is used when your own trial has individual patient data (IPD), but the trial you need to compare against only exists in published, aggregate form — a data-availability gap that a standard Bucher-method indirect comparison cannot fully resolve, because Bucher’s method requires aggregate summary data from both trials and provides no mechanism for correcting cross-trial differences in patient mix. MAIC reweights the IPD trial’s patients, individual by individual, so their weighted aggregate characteristics match the published aggregate characteristics of the comparator trial, then compares outcomes using that reweighted population. It is one of the population-adjustment methods described in NICE Decision Support Unit Technical Support Document 18 and is now routine in industry-sponsored health technology assessment (HTA) submissions.

The scenario MAIC solves: IPD on one side, aggregate data on the other

Industry-sponsored comparative-effectiveness work often runs into a specific, recurring problem. A manufacturer running its own trial has full individual patient data (IPD): every enrolled patient’s baseline characteristics, treatment arm, and outcome. But the competitor product it needs to compare against was evaluated in a separate trial run by someone else, and all that’s publicly available for that trial is the published paper — aggregate summary statistics (mean age, percentage with a given comorbidity, overall response rate) with no patient-level records.

This is a genuinely different data-availability scenario from the one the Bucher method assumes. Bucher’s calculation — Effect(B vs. C) = Effect(A vs. B) − Effect(A vs. C) — only needs the two trials’ summary effect estimates and variances; it has no way to adjust for the fact that the A-vs-B trial’s population might be systematically older, sicker, or otherwise different from the A-vs-C trial’s population. When both trials are aggregate-only, that mismatch has to be handled by judgment (the similarity/transitivity assumption) rather than by calculation. MAIC exists for the case where one side of the comparison has IPD available — because IPD is what makes it possible to actually reweight a population rather than just assume it’s comparable.

The core idea: reweight the IPD trial to look like the aggregate trial

MAIC does not pool the two trials’ patients into one dataset — it changes how much each IPD patient counts. Each individual patient in the IPD trial is assigned a weight so that, once weighted, that trial’s own aggregate characteristics (mean age, % female, % with prior treatment, disease severity distribution, and so on) match the aggregate characteristics reported in the comparator trial’s publication. A patient whose characteristics differ from what the comparator trial’s population looked like on average gets down-weighted; a patient who looks more like the comparator trial’s typical enrollee gets up-weighted. The result is a synthetic, reweighted version of the IPD trial’s population that resembles the comparator trial’s population on every characteristic the reweighting targeted — closer to a like-for-like comparison than comparing the two trials’ raw, unadjusted populations.

The method, step by step

1. Identify the characteristics to match

Select the baseline characteristics — prognostic factors and, especially, treatment-effect modifiers (variables that plausibly change the size of the treatment effect, not just the outcome’s baseline level) — that are reported in both the IPD trial and the comparator trial’s publication. A characteristic can only be matched if the comparator trial’s paper actually reports it in aggregate form (typically a mean, proportion, or standard deviation); anything the published trial didn’t report cannot be used, however clinically relevant it might be.

2. Estimate weights via method of moments

Each IPD patient’s weight is calculated using a method-of-moments approach: a set of coefficients is estimated (commonly by minimizing an objective function equivalent to logistic-regression weighting) such that, when applied to the IPD patients, the weighted sample’s means for every matched characteristic exactly equal the comparator trial’s published aggregate values. Patients are not dropped; every patient keeps some weight, but the weights are no longer equal across patients the way they would be in an unweighted analysis.

3. Check the effective sample size

Reweighting concentrates influence on the patients who most resemble the comparator trial’s population, which reduces statistical precision even though no patients are physically removed. The standard way to quantify this is the effective sample size (ESS) — a function of the weights that is always less than or equal to the original number of patients, and often substantially less when the two trial populations differed a great deal on the matched characteristics. A large drop in ESS relative to the original IPD sample size is a signal that the two populations barely overlapped, and that the resulting comparison rests on a small effective amount of information even though the nominal patient count looks large.

4. Compare outcomes using the reweighted population

With weights assigned, the outcome of interest is recalculated for the IPD trial using the reweighted patients, and that reweighted outcome is compared against the comparator trial’s published aggregate outcome. Two structures are used depending on what’s available:

• Anchored MAIC — used when both trials share a common comparator arm (for example, both trials have a placebo or standard-of-care arm alongside the treatment of interest). The comparison is made on the relative-effect scale within each trial first, then the two relative effects are compared indirectly — the same logic as the Bucher method, but with the IPD trial’s population reweighted first. This preserves randomization within each individual trial and requires a weaker assumption than the unanchored case.
• Unanchored MAIC — used when there is no shared comparator arm to anchor through, so outcomes are compared directly between the reweighted IPD arm and the comparator trial’s arm. This requires the much stronger assumption that all effect modifiers AND all prognostic factors influencing the outcome have been matched — not just effect modifiers — because there is no within-trial randomization left to control for anything unmatched. Unanchored MAIC is markedly more fragile and is generally used only when no connected network of common comparators exists at all.

The core limitation: you can only match what was published

MAIC’s reweighting step can only adjust for a characteristic if the comparator trial’s own publication reported it in aggregate form. If a genuine effect modifier exists but the comparator trial’s paper never published a mean or proportion for it, MAIC has no way to include it in the weighting — the imbalance on that characteristic remains completely unadjusted, silently, in the final result. This is a real and frequently raised limitation in HTA appraisals of MAIC submissions: reviewers at bodies like NICE routinely ask sponsors to justify why a particular set of effect modifiers was chosen and whether any plausible ones were omitted because the comparator publication simply didn’t report them. Unlike an underpowered trial, this isn’t a limitation that more data collection on the IPD side can fix — the ceiling is set entirely by what the other trial’s authors chose to publish.

Two further limitations compound this: the reduced effective sample size after weighting (step 3 above) means the standard errors on a MAIC result are typically wider than a naive comparison of the raw trial populations would suggest, and unanchored MAIC in particular carries an assumption (no unmatched prognostic factors, not just effect modifiers) that is rarely fully verifiable. For all of these reasons, HTA bodies generally treat MAIC results as a supporting, exploratory form of evidence rather than a substitute for a real head-to-head trial, and expect a full accounting of which characteristics were matched, which were considered but unavailable, and why.

MAIC vs. the Bucher method: which one fits your data

These two methods are not competing choices for the same situation — which one applies is determined by what data you actually have access to:

• Use the Bucher method when only published, aggregate-level results are available for both trials being compared. It is simpler, requires no patient-level access, and is the appropriate two-treatment special case of network meta-analysis when trial populations are judged similar enough on the similarity/transitivity assumption.
• Use MAIC when you hold individual patient data for one trial (typically your own sponsored trial) but only published aggregate data exists for the comparator trial, and you have specific concerns that the two trial populations differ on characteristics that would otherwise bias a simple Bucher-style comparison.

Both methods share the same underlying purpose — estimating a treatment comparison where no direct head-to-head trial exists — and both ultimately depend on an unverifiable assumption about unmeasured or unreported differences between trial populations. MAIC narrows that assumption to unreported characteristics specifically; Bucher’s method leaves it as a broader judgment call about overall trial similarity. Neither is a full substitute for a real head-to-head trial, and reviewers evaluating either should look for a clear justification of the method chosen and the characteristics used.

Frequently asked questions

Does MAIC require access to the comparator trial’s raw data?

No — that is precisely the scenario MAIC is designed for. It needs individual patient data for only one of the two trials being compared; the other trial’s contribution is its published aggregate results.

Is anchored or unanchored MAIC preferred?

Anchored MAIC is preferred whenever a shared comparator arm exists, because it relies on a weaker assumption (only effect modifiers need to be matched, and within-trial randomization is preserved). Unanchored MAIC’s stronger assumption — that all prognostic factors and effect modifiers are matched — makes its results more fragile and is used only when no connected comparator network exists.

Can MAIC fully replace a head-to-head trial in an HTA submission?

No. HTA bodies generally treat MAIC as supporting, exploratory evidence rather than a substitute for direct comparative evidence, precisely because it can only adjust for characteristics the comparator trial’s publication actually reported.

What happens to sample size after MAIC reweighting?

The effective sample size (ESS) — not the raw patient count — determines the comparison’s statistical precision after weighting, and ESS is always less than or equal to the original IPD sample size. A large reduction signals limited population overlap between the two trials.