Diluting the signal: Class-level pooling and the ecological fallacy in meta-analyses of anti-amyloid monoclonal antibodies

The recently published Cochrane systematic review by Nonino et al. concludes that amyloid-β-targeting monoclonal antibodies produce no clinically meaningful benefit in mild cognitive impairment or mild dementia due to Alzheimer's disease, and advises that future research should abandon this mechanism of action. ¹ The conclusion has generated wide media coverage and has been invoked in early policy discussions concerning access to approved therapies. A careful reading of the review reveals that its principal inferential claim rests on a methodological decision with serious implications: the pooling of nine pharmacologically distinct agents (aducanumab, bapineuzumab, crenezumab, donanemab, gantenerumab, lecanemab, ponezumab, remternetug, and solanezumab) into a single class-level estimate. Only three of these (aducanumab, lecanemab, and donanemab) received FDA approval, and only two (lecanemab and donanemab) remain in active clinical use. Treating the pooled average as an informative verdict on the therapeutic class commits, in epidemiological terms, an ecological fallacy at the drug level.

The ecological fallacy is classically defined as the inference of individual-level relationships from aggregate-level data. The Cochrane Handbook itself explicitly cautions against this bias in the context of meta-analyses that aggregate study-level information, noting that associations observed at the group level may not reflect relationships at the individual level (section 9.6.5.5). ² The same logic applies with equal force when the aggregation is across drugs rather than across individuals within a drug. Pooling non-exchangeable interventions under a shared mechanistic label conceals effect modification and produces an average estimate that describes no clinically relevant comparator.

The authors of the Cochrane review justify class-level pooling on two grounds: shared mechanism of action, and common use of amyloid-PET as a measure of target engagement. ¹ Both premises conflate mechanistic kinship with clinical equivalence. The agents included differ substantially in target epitope, in the amyloid species preferentially bound (soluble monomers, oligomers, protofibrils, fibrils, or deposited plaques), in the degree and kinetics of in vivo plaque clearance achieved, and in the populations enrolled for their pivotal trials. Solanezumab, for instance, targets soluble monomeric amyloid and fails to reduce plaque burden. Lecanemab preferentially binds soluble protofibrils and achieves robust plaque clearance on amyloid-PET. ³ Donanemab targets pyroglutamate-modified plaque and produces near-complete clearance in many treated patients, with a clinically adjudicated slowing of decline in the TRAILBLAZER-ALZ 2 trial. ⁴ Collapsing these distinct pharmacological profiles into a single pooled estimate is not merely a conservative analytical choice: it is an active statistical operation that blurs the distinctions between agents that have failed clinically and agents that have modified the trajectory of symptomatic disease.

This problem is not unique to the Cochrane review. The 2024 meta-analysis by Ebell and colleagues similarly pooled eight anti-amyloid antibodies and concluded that none achieved the minimal clinically important difference, a conclusion inseparable from the choice of pool. ⁵ In contrast, two recent analyses demonstrate that a more defensible approach is feasible. Avgerinos et al. (2024) performed subgroup analyses by individual agent and by binding affinity, finding that donanemab and lecanemab produced the largest benefits and that antibodies without affinity for monomeric amyloid had the most favorable effects. ⁶ Wu et al. (2023) restricted their meta-analysis to FDA-approved antibodies, thereby preserving the regulatory signal. ⁷ These examples establish that stratified analysis is both feasible and informative, and that the pooling decision is methodological rather than inevitable.

Three concrete recommendations follow. First, pre-specified subgroup analyses stratified by regulatory approval status and antibody generation should be the default, not the exception, in syntheses of pharmacologically heterogeneous drug classes. Second, meta-regression exploring amyloid clearance as an effect modifier of cognitive outcome would directly test the biological hypothesis underlying any such review, and would disaggregate agents that cleared plaque from those that did not. Third, individual participant data meta-analysis, where feasible, allows effect modification to be examined at the appropriate level of inference, and should be pursued whenever regulatory-grade data are accessible.

The consequences of neglecting these distinctions extend beyond academic dispute. Health technology assessment bodies (NICE, EMA, the Brazilian CONITEC, among others) increasingly rely on meta-analytic syntheses to inform reimbursement and access decisions. When a pooled estimate is dominated by failed agents, the certainty of evidence assigned through GRADE to the aggregated outcome cannot be validly transferred to the individual approved agents. Policies built on such estimates risk denying access to therapies with demonstrated, if modest, clinical benefit, on the basis of evidence drawn largely from molecules that never reached clinical practice. Similarly, claims that future research should abandon amyloid-targeting mechanisms, when drawn from an analysis that blurs failed from successful agents, misdirect scientific priorities at a moment when precision matters.

The amyloid hypothesis remains contested, and the clinical gains of approved anti-amyloid antibodies are modest. Rigorous scrutiny is warranted. However, the standards of that scrutiny must themselves be rigorous. Meta-analyses that average heterogeneous agents under a shared mechanistic label, without stratification or justification of statistical homogeneity, do not advance understanding. And when such analyses shape public perception, regulatory reassessment, or funding allocation, the cost is paid by patients.