Evaluation of amyloid change as a surrogate for cognitive decline: demonstration in individual‐level data from the A4 study of solanezumab

Abstract

Background

Differential amyloid change has served as a surrogate outcome in Alzheimer's disease trials, allowing accelerated approval of aducanumab and lecanemab. Individual‐level data from the A4 study support the demonstration of novel methods to evaluate amyloid's validity as a surrogate for cognitive decline.

Methods

In 812 participants, cognitive change was measured using the Clinical Dementia Rating Sum of Boxes (CDR‐SB) score. Instrumental‐variable analysis estimated the effect of amyloid change; mediation analysis quantified proportion of cognitive effects mediated by amyloid change.

Results

Each 10‐Centiloid reduction in amyloid due to randomization to treatment was associated with a 0.026 higher CDR‐SB score (95% confidence interval [CI]: −0.013, 0.065). Amyloid reduction mediated 14.6% of solanezumab's effect on cognition (95% CI: −122%, 208%).

Discussion

Near‐zero effect estimates for amyloid change on cognitive decline in the A4 study suggest minimal impact of limited amyloid change reduction in populations with little disease progression. The broader question of validity of amyloid as a surrogate outcome cannot be conclusively answered in data from the A4 study due to study‐intrinsic limitations. Replication in anti‐amyloid trials with larger treatment effects will evaluate whether amyloid is an appropriate surrogate outcome.

Highlights

• This study evaluated amyloid change's validity as a surrogate for cognitive and functional decline in AD drug trials.

• Newly available individual‐level trial data from the A4 study enabled the application of epidemiologic and econometric methods to assess amyloid's impact on cognition.

• IV and causal mediation analyses estimated the effect of amyloid change on cognitive outcomes.

• Amyloid change mediated 15% of solanezumab's cognitive effect, though estimates were imprecise due to limited disease progression in the sample and solanezumab's minimal amyloid removal.

• Applying the same methods to data from trials of more effective anti‐amyloid drugs could validate amyloid as a surrogate outcome, guide related regulatory decisions, and influence treatment strategy.

1. BACKGROUND

Under the US Food and Drug Administration's Accelerated Approval Program, change in amyloid has served as a surrogate marker for cognitive and functional outcomes in Alzheimer's disease (AD) trials. Aducanumab ¹ and lecanemab were approved based on amyloid reduction. ² Subsequently, lecanemab and donanemab received full approval by demonstrating cognitive and functional benefits. ³ Rigorous evaluation of amyloid as a surrogate endpoint for cognitive decline using clinical trial data has not been reported.

While some argue that results from recent agents clearly demonstrate a causal role of amyloid in cognitive decline, ⁴ , ⁵ others speculate about possible non‐amyloid mechanisms. Skepticism about the safety and efficacy of amyloid‐targeting drugs persists. ⁶ , ⁷ With better tools to analyze and integrate existing trial results, there is potential to move beyond opinion‐based debates toward evidence‐based conclusions. To this end, several earlier meta‐analyses used aggregated, trial‐arm‐level data from anti‐amyloid drug trials to evaluate the effect of amyloid change on cognitive decline and obtained conflicting results; the most recent of those analyses reported small overall benefits. ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹²

Existing literature that attempts to formally evaluate the effect of amyloid reduction or reduction in accumulation on cognitive decline is limited by the lack of availability of individual‐level data. Specifically, existing analyses assume drug effects are fully mediated by changes in amyloid. This assumption can be tested in individual‐level but not in group‐level data. In addition, group‐level averages may mask important features of amyloid‐cognition relationships, such as widely hypothesized non‐linear associations. ¹³ , ¹⁴ Finally, additional statistical assumptions may be required when working with aggregated data. ⁸

Drawing on methods from epidemiology and econometrics, ¹⁵ , ¹⁶ we propose the following sequential analyses for evaluating surrogate outcomes in individual‐level clinical trial data: (1) using causal mediation analysis, quantify the fraction of the drug effect on the outcome mediated by the putative surrogate ¹⁷ and (2) using instrumental‐variable (IV) analysis, estimate the ratio of the effect of randomization on the outcome to its effect on amyloid reduction, interpreted as the treatment effect per unit change in the surrogate. If in analysis 1 the drug effect is largely through the surrogate, then the effect of treatment on the outcome per change in the surrogate approximates the causal effect of surrogate on the outcome. ¹⁵

Taken together, if these analyses demonstrate that the surrogate mediates a substantial portion of the treatment effect ¹⁷ and that changes in the surrogate yield clinically meaningful improvements in the outcome, they provide strong evidence for the surrogate's validity. Typical approaches to evaluate surrogacy examine within‐arm correlations between the putative surrogate and outcome. ¹⁸ , ¹⁹ In contrast, our approach leverages the benefits of randomization in IV analyses to evaluate whether amyloid is an appropriate surrogate for clinical outcomes, and it has the major advantage of allowing for within‐arm adjustment of confounding in mediation analyses.

The full implementation of these proposed methods requires individual‐level data on change in amyloid and change in cognition or function over time, from individuals randomized to varying doses of anti‐amyloid drugs (including placebo). Their application in this context has been hindered by a lack of familiarity with these methods for the evaluation of surrogate outcomes and by restricted access to individual‐level data from anti‐amyloid drug trials.

In this manuscript, we first illustrate our approach via simulated data. We then demonstrate the appropriate analyses in the newly available individual‐level Anti‐Amyloid Treatment in Asymptomatic Alzheimer's Disease (A4) study data, a large study of solanezumab for the prevention of cognitive decline in preclinical dementia. We do not expect to resolve the question of surrogacy in data from the A4 trial alone, as this trial did not demonstrate treatment effects. As additional trial data become available (e.g., from the TRAILBLAZER‐ALZ‐2 trial of donanemab ²⁰ ), we recommend that the approaches modeled here be implemented to rigorously assess amyloid as a surrogate outcome for cognitive decline.

2. METHODS

2.1. A4 study

The A4 study was a Phase 3 randomized‐controlled clinical trial of solanezumab, an anti‐amyloid monoclonal antibody, for the reduction of cognitive decline in individuals with elevated amyloid assessed via amyloid positron emission tomography (PET) but no clinical symptoms. ²¹ Final results reported in 2023 indicated that solanezumab did not slow cognitive decline. ²² Additional details are reported elsewhere. ²¹ , ²³ , ²⁴ , ²⁵

We perform secondary data analysis using deidentified individual‐level data from the A4 study. All participants provided informed consent at the time of enrollment. Institutional Review Board (IRB) approval was obtained at each of the trial sites. Data were obtained from the A4 and Longitudinal Evaluation of Amyloid Risk and Neurodegeneration (LEARN) Study Data Package.

2.2. Amyloid‐PET imaging

All participants underwent amyloid‐PET imaging at first and last study visits. ²³ Standardized uptake value ratios (SUVRs) were converted to Centiloids using the following formula: Centiloids = 183.07×(Florbetapir SUVR) − 177.26.

2.3. Cognitive and functional outcomes

Full details on cognitive testing are described in Table S1 and elsewhere. ²⁶ The A4 study's primary endpoint was change in the Preclinical Alzheimer Cognitive Composite (PACC) at 4.5 years. For interoperability, we examined PACC subcomponents: Mini‐Mental State Examination (MMSE) and Digit Symbol Substitution Test (DSST). The study also assessed Clinical Dementia Rating Sum of Boxes (CDR‐SB) and Alzheimer's Disease Cooperative Study–Activities of Daily Living (ADCS‐ADL). We emphasize CDR‐SB, the primary endpoint in recent successful monoclonal antibody trials and the only measure where higher scores indicate worse function. We refer to all cognitive and functional outcomes evaluated as “cognitive outcomes” for clarity and parsimony, with some loss of linguistic precision.

RESEARCH IN CONTEXT

1. Systematic review: Existing literature estimating the effects of amyloid change on cognitive change uses only aggregated, rather than individual‐level, data and furnishes conflicting results. Newly available A4 trial data enable novel, potentially more powerful methods to assess amyloid's impact on cognitive decline.

2. Interpretation: Using individual‐level data from A4 trial participants, we applied epidemiologic and econometric methods, causal mediation, and IV analysis to evaluate amyloid change as a surrogate. Our results suggest amyloid change mediated 15% of solanezumab's small effect, though confidence intervals were wide due to limited amyloid removal and limited disease progression in this sample.

3. Future directions: Use of amyloid as a surrogate outcome for regulatory decisions for Alzheimer's treatments should be based on evaluations of the link between amyloid change and cognitive change. Future research should apply methods used here to trials of more effective anti‐amyloid therapies to evaluate whether amyloid is an appropriate surrogate outcome.

3. ANALYSIS

3.1. Estimates of change and covariates

Change in amyloid for each participant was determined by taking the difference between Centiloids collected at baseline and the final study visit at approximately 240 weeks. We estimated cognitive changes using an annual trajectory by taking the slope of a linear model fit with respect to time, giving cognitive change per year for each individual.

Analyses were adjusted for sex, APOE ε4 gene dose, baseline age, baseline cognition, and baseline amyloid. Since participants were randomly assigned to treatment, adjustment for baseline cognition would not be expected to induce bias. ²⁷

3.2. Methods for evaluation of a surrogate

Figure 1 provides conceptual models (directed acyclic graphs) of mediation and IV analyses for examining the effect of amyloid change. ²⁸

FIGURE 1.

Conceptual models of instrumental‐variable and mediation analyses for analyses examining the effect of amyloid change. (A) Effect of anti‐amyloid therapy on cognition is partially mediated by change in amyloid, but confounders C (e.g., sex, APOE ε4, and baseline age, cognition, and amyloid) influence both change in amyloid and change in the outcome. Mediation analyses estimate the fraction of the effect of randomization on cognition that operates via changes in amyloid but requires that we know all common causes (confounders) C of change in amyloid and change in cognition and assumes no exposure‐induced mediator‐outcome confounding. (B) The instrumental‐variable (IV) estimate provides the ratio of the effect of randomization to treatment on cognition to the effect of randomization on change in amyloid; under the stronger assumption that the drug's effect operates entirely via amyloid, this provides an estimate of the effect of amyloid change on cognitive change. Because this assumption may not be met, we interpret the IV estimate as the effect of randomization to treatment on cognition, scaled by the change in amyloid.

3.3. Simulation of illustrative data

To show how mediation and IV estimates depended on relationships between change in cognition and change in amyloid, we simulated data under three hypothetical drug/trial scenarios (X, Y, Z). Each scenario varied assumptions about drug effects on amyloid, drug effects on cognition, and observability of disease progression in the trial sample. Simulation details are given in Table S2.

Scenario X shows a large, donanemab‐like amyloid reduction in amyloid ²⁰ with the arrest of cognitive decline in the treatment arm, with ongoing decline in placebo. Scenario Y shows a similar amyloid reduction but no cognitive benefit, with comparable cognitive decline observed in both trial arms. Scenario Z shows a small, solanezumab‐like effect that slows amyloid accumulation but does not remove it (consistent with the drug's affinity for monomers). Cognitive decline is observed in neither treatment nor control arms in Scenario Z's trial sample, precluding inference about the effect of the drug on cognition through amyloid.

3.4. Implementation of analyses

For mediation, assumptions include no unmeasured confounding of the amyloid–cognition relationship and no exposure‐induced mediator‐outcome confounding. ¹⁶ For IV, we did not require the exclusion restriction assumption but instead interpreted the IV results descriptively as the effect of treatment on cognition scaled by change in amyloid. ²⁹ Further discussion of mediation and IV can be found in the Supplemental methods.

3.4.1. Mediation analyses

We specified a series of linear regression models to evaluate the relationships between treatment assignment, amyloid change, and cognitive outcomes. The first model, the mediator model, regressed amyloid change on treatment group assignment to estimate the effect of solanezumab versus placebo on amyloid change. The second model, the outcome model, regressed cognitive change on both amyloid change and treatment group, as well as confounders for adjusted models. The proportion of the total effect mediated through amyloid change is calculated as the ratio of the mediated effect (average causal mediation effect) to the total effect (average treatment effect) using the mediate function from the R mediation package (version 4.5.0) with the two models as inputs. Confidence intervals were obtained using quasi‐Bayesian Monte Carlo with 1000 simulations. ³⁰ , ³¹

3.4.2. IV analyses

We estimated the effect of randomization on cognitive change per change in amyloid using the IVreg function from the R AER package (version 1.2.14). In IV analyses, we adjusted for the same confounders as in the mediation analyses in both regressions, but this is for the purpose of improving precision and not for confounding adjustment.

4. RESULTS

For 812 study participants with complete before/after amyloid‐PET measures, complete CDR‐SB measures, and at least two observations for all other cognitive measures evaluated, the mean age was 71.47 (SD 4.45) years, and 59.2% of participants were women. The between‐group difference (treatment minus placebo) in average amyloid change was −8.33 (95% CI: −11.25, −5.40) Centiloids. Average between‐group difference in annual change in CDR‐SB was 0.02 (95% CI: −0.01, 0.06) points per year. The correlation between treatment arm and amyloid change was −0.19 (95% CI: −0.26, −0.12), p < 0.01. Among the solanezumab treated (N = 396), the correlation between dose escalation day relative to baseline and amyloid change was −0.07 (95% CI: −0.17, 0.02), p = 0.17. Summary statistics for additional quantities are given in Table 1.

TABLE 1.

Characteristics of treated and untreated groups and overall. Mean and standard deviation are shown for all covariates used in the analysis, along with mean changes in function, cognition, and amyloid.

	Placebo (N = 416)	Solanezumab (N = 396)	Overall (N = 812)
Age	71.48 (4.46)	71.46 (4.43)	71.47 (4.45)
Sex
Female	247 (59.4%)	234 (59.1%)	481 (59.2%)
Male	169 (40.6%)	162 (40.9%)	331 (40.8%)
APOE ε4 allele copies
0	161 (38.7%)	150 (37.9%)	311 (38.3%)
1	222 (53.4%)	210 (53.0%)	432 (53.2%)
2	33 (7.9%)	36 (9.1%)	69 (8.5%)
Baseline amyloid burden on PET, Centiloids	65.65 (31.06)	65.43 (31.47)	65.54 (31.24)
Before/after change in amyloid PET, Centiloids	21.13 (20.64)	12.80 (21.75)	17.07 (21.58)
Baseline CDR‐SB score	0.05 (0.17)	0.06 (0.16)	0.05 (0.17)
Change in CDR‐SB score per year	0.10 (0.22)	0.13 (0.27)	0.11 (0.25)
Baseline MMSE score	28.79 (1.24)	28.86 (1.21)	28.83 (1.22)
Change in MMSE score per year	−0.09 (0.36)	−0.10 (0.41)	−0.09 (0.39)
Baseline DSST score	49.36 (9.61)	49.05 (10.31)	49.21 (9.95)
Change in DSST score per year	−0.41 (1.31)	−0.41 (1.38)	−0.41 (1.34)
Baseline ADL partner score	43.48 (2.66)	43.56 (2.38)	43.52 (2.53)
Change in ADL partner score per year	−0.30 (0.97)	−0.40 (1.12)	−0.35 (1.04)
Baseline PACC score	0.11 (2.59)	0.21 (2.71)	0.16 (2.65)
Change in PACC score per year	−0.22 (0.77)	−0.25 (0.91)	−0.23 (0.84)

Abbreviations: ADCS–ADL, Alzheimer's Disease Cooperative Study–Activities of Daily Living; APOE ε4, apolipoprotein E with ε4 allele; CDR‐SB, Clinical Dementia Rating Sum of Boxes; DSST, Digit Symbol Substitution Test (Wechsler Adult Intelligence Scale); MMSE, Mini‐Mental State Examination; PACC, preclinical Alzheimer's cognitive composite.

Figure 2 is intended to facilitate the interpretation of the mediation analysis results for this and future studies. If amyloid mediates 100% of the effect of solanezumab treatment on cognition, the IV estimate corresponds to the causal effect of amyloid on cognition. Otherwise, the IV estimate gives the effect of randomization to treatment on cognition, scaled per 10 Centiloids of amyloid change; this estimate may over‐ or underestimate the effect of amyloid change on cognitive change, depending on the percentage mediated.

FIGURE 2.

Guide to evaluation of amyloid change as surrogate outcome using mediation and instrumental‐variable (IV) analyses. Idealized scenarios in which change in amyloid has no effect on cognition (A) or in which randomization affects cognition only via change in amyloid (B). More realistic scenarios (C1, C2, and C3) assume that randomization influences cognition via multiple mechanisms, including amyloid reduction. In this setting, the interpretation of the IV estimate is guided by the mediation estimate. Here, we assume an IV estimate that is positive, reflecting the putative assumption that a reduction in amyloid accumulation reduces cognitive decline. If a cognitive measure such as the Clinical Dementia Rating Sum of Boxes, where increases reflect cognitive worsening, is used, or if the surrogate causes harm, the same reasoning applies to the sign‐reversed quantity. Confounders of the relationship between amyloid reduction and cognitive decline are omitted for parsimony but must be adjusted for to obtain valid mediation results; adjustment for such confounders is not essential for IV estimates but may improve precision.

Figures 3 and 4 are intended to facilitate interpretation of the A4 findings via comparison to simulation results. Figure 3 shows the distribution of changes in the clinical outcome (cognitive function) and potential surrogate (amyloid) for simulated hypothetical trials X, Y, and Z and for individuals in the A4 study. In the A4 trial, placebo and treatment groups are similar, with very little change in amyloid or cognition. Figures S1 and S2 are intended to illustrate that cognitive decline was minimal across cognitive outcomes in A4. Figure S1 shows group mean and median scores for all visits for each cognitive measure used. Figure S2 shows changes in these scores by modeled trajectory (primary analysis) or first‐to‐last visit differences (sensitivity analyses). Figure 4 shows individual‐level changes in amyloid and cognition, a visual representation of IV model fits, and mediation and IV estimates for hypothetical trials and the A4 study.

FIGURE 3.

Change in amyloid and change in cognition for simulation Scenarios X, Y, and Z and as observed in the Anti‐Amyloid Treatment in Asymptomatic Alzheimer's Disease (A4) trial. Simulated scenarios X and Y reflect large drug‐induced changes in amyloid, akin to those seen in recent trials like TRAILBLAZER‐ALZ 2. In Scenario X, treatment confers significant cognitive benefit, while in scenario Y treatment confers no cognitive benefit. In Scenario Z, there is only a small difference between amyloid change in treated and placebo groups and no average disease progression in either group. In A4, there is a small difference in amyloid change between the placebo and solanezumab groups, and neither function nor cognition changed for most participants in either the treatment or placebo arms.

FIGURE 4.

Change in cognition versus change in amyloid for simulation scenarios and the Anti‐Amyloid Treatment in Asymptomatic Alzheimer's Disease (A4) trial. Rate of change in cognition is plotted against change in amyloid. Unadjusted mediation and instrumental‐variable (IV) estimates are presented in simulation scenarios and the A4 trial. Crosses represent ± 1 standard deviation centered at group mean values, and the slope of regression lines connecting the intersection of crosses is equivalent to the IV estimate. Estimates of the percent mediated are not expected to provide precise results in scenarios where there is no overall effect of randomization to treatment on cognitive outcome, as in Scenarios Y and Z and in the A4 trial. Plotted points were jittered to improve visibility. IV estimates are presented as per 10‐Centiloid (CL) reductions (in further accumulation) and, thus, are sign reversed compared to the slopes of fitted lines.

Scenario X: A simulated trial in which the study population experiences change in cognitive outcomes. Drug X effectively removes cerebral amyloid and benefits cognition.

Scenario Y: A simulated trial in which the study population experiences change in cognitive outcomes. Drug Y effectively removes amyloid but has no impact (positive or negative) on cognition.

Scenario Z: A simulated trial in which the study population does not experience meaningful change in the cognitive outcomes. Drug Z does not remove amyloid and does not impact cognition.

Mediation analysis suggests that amyloid change mediates 14.6% of solanezumab's effect on cognitive change (95% CI: −122%, 208%). The IV‐estimated effect of treatment on annual change in CDR‐SB change per 10‐Centiloid reduction was 0.026 (95% CI: −0.013, 0.065). The remaining IV and mediation results are qualitatively similar and shown in Table 2. There was no evidence of non‐linear effects of amyloid on CDR‐SB by treatment group (p = 0.27) (Figure S3). The sensitivity analyses performed using change in cognitive outcomes from the first in‐study assessment visit and the end‐of‐study visit are shown in Tables S3 and S4 and are consistent with the main results. Table S5 gives percentages of individuals in each treatment arm with declines in amyloid.

TABLE 2.

Mediation and instrumental‐variable results for effects of reductions in amyloid accumulation on cognition in A4 trial.

Method	Measure	Reported value	Unadjusted (95% CI)	Adj. Set 1 (95% CI)	Adj. Set 2 (95% CI)
Mediation	CDR‐SB	Proportion mediated	0.359 (−2.874, 3.706)	0.327 (−3.605, 3.853)	0.146 (−1.216, 2.084)
Mediation	MMSE	Proportion mediated	0.391 (−9.558, 9.035)	0.337 (−10.773, 7.504)	0.240 (−6.116, 7.625)
Mediation	DSST	Proportion mediated	−0.074 (−4.929, 8.377)	−0.022 (−4.884, 8.516)	−0.014 (−4.117, 2.734)
Mediation	ACDS‐ADL partner score	Proportion mediated	0.374 (−2.869, 4.246)	0.342 (−3.425, 3.771)	0.235 (−2.742, 2.186)
Mediation	PACC	Proportion mediated	0.535 (−6.677, 11.915)	0.499 (−7.010, 9.179)	0.338 (−6.253, 6.403)
IV	CDR‐SB	Effect of randomization on annual rate of cognitive change, scaled per 10 CL reduction	0.027 (−0.014, 0.068)	0.026 (−0.014, 0.066)	0.026 (−0.013, 0.065)
IV	MMSE	Effect of randomization on annual rate of cognitive change, scaled per 10 CL reduction	−0.012 (−0.076, 0.052)	−0.010 (−0.073, 0.052)	−0.009 (−0.070, 0.051)
IV	DSST	Effect of randomization on annual rate of cognitive change, scaled per 10 CL reduction	0.010 (−0.213, 0.232)	0.014 (−0.204, 0.231)	0.009 (−0.203, 0.220)
IV	ACDS‐ADL partner score	Effect of randomization on annual rate of cognitive change, scaled per 10 CL reduction	−0.125 (−0.298, 0.048)	−0.121 (−0.291, 0.049)	−0.119 (−0.283, 0.045)
IV	PACC	Effect of randomization on annual rate of cognitive change, scaled per 10 CL reduction	−0.040 (−0.178, 0.099)	−0.035 (−0.168, 0.098)	−0.046 (−0.171, 0.079)

Note: Mediation results give the proportion of total effect on cognition mediated by amyloid. Instrumental‐variable results give the effect of randomization to treatment on annual cognitive change scaled per 10‐CL reduction in amyloid accumulation. Cognitive outcomes include the CDR‐SB, MMSE, DSST, ADCS‐ADL partner, and overall PACC scores. Adjustment set 1 includes baseline age, sex, and APOE ε4 carrier status. Adjustment set 2 additionally includes baseline amyloid and baseline cognitive measure.

Abbreviations: ADCS‐ADL: Alzheimer's Disease Cooperative Study–Activities of Daily Living; APOE ε4, apolipoprotein E with ε4 allele; CDR‐SB: Clinical Dementia Rating Sum of Boxes; CI, confidence interval; CL, Centiloid; DSST: Digit Symbol Substitution Test (Wechsler Adult Intelligence Scale); MMSE: Mini‐Mental State Examination; PACC: preclinical Alzheimer's cognitive composite.

5. DISCUSSION

We reanalyzed individual‐level A4 data to evaluate whether amyloid reduction was a valid surrogate for cognitive and functional decline. We could not precisely estimate the percentage of the effect of solanezumab treatment mediated by amyloid reduction, with both 0% and 100% mediation falling within the 95% CI. The IV estimates, which give the effect of treatment per 10‐Centiloid reduction, are close to zero. This is consistent with a previously published intent‐to‐treat estimate ²³ and more precise than previous IV estimates in aggregated data. ⁸ Regardless of the proportion mediated, amyloid reduction is unlikely to impact cognition over 5 years in populations with minimal cognitive decline.

Our goal was not to relitigate the primary results of A4 but to pilot how causal inference methods could be applied in practice. Although these methods require assumptions and are not without limitations, they provide a more rigorous framework than correlation analyses alone for assessing amyloid's validity as a surrogate endpoint. Our results are consistent with those of the original analysis and highlight what these approaches can and cannot reveal when applied to a trial with very little disease progression. Specifically, the importance of this analysis is its demonstration of an approach to evaluating amyloid reduction as a surrogate outcome using clinical trials data – an approach that is immediately feasible and actionable with individual‐level data already collected in prior trials. Causal mediation and IV methods are not yet widely used in AD trials, largely due to limited access to individual‐level data and unfamiliarity with these approaches among trialists. Broader data sharing and methodological training could facilitate wider adoption and strengthen surrogate validation efforts.

From the mediation and IV results using the A4 trial alone we cannot conclude amyloid change is a reliable surrogate for cognitive change. This is because (1) we cannot conclude amyloid largely or entirely mediates the effect of treatment on cognition and function and (2) the estimated effect of a 10‐Centiloid reduction on cognitive change is precise and close to zero.

However, we also cannot rule out amyloid change as a reliable surrogate for change in cognition using the A4 trial data. The mediation results were inconclusive because most participants did not cognitively decline (even if mean decline was non‐zero), randomization to treatment did not appreciably modify this, and the effect of solanezumab on amyloid was small – inadequate to fully prevent treated individuals from continuing to accumulate amyloid. This contrasts with trials of the newly approved drugs lecanemab and donanemab. These drugs not only prevent amyloid accumulation but also dramatically reduce baseline amyloid levels in populations with significant cognitive decline. ²⁰ , ³² The IV estimates in A4 participants were close to zero since there was no appreciable difference in cognitive change between groups. Reproductions in subsequent trials will not suffer from this limitation.

Rigorous methodological approaches are needed to evaluate efficacy and safety, resolve ongoing debates, ⁴ , ⁵ , ⁶ , ⁷ and inform future drug approval processes. Evaluating the relationship between amyloid reduction and cognitive changes using these tools would provide clear evidence for or against amyloid as a surrogate. If treatment‐induced changes in amyloid contribute to cognitive benefits, this provides modest support for amyloid as a surrogate marker for treatment efficacy. ¹⁷ However, if amyloid is not the primary determinant of cognitive and functional outcomes, granting accelerated approval solely on the basis of amyloid reduction would not be justified. We note that the challenges raised here in the context of amyloid‐targeting drugs echo long‐standing concerns in other therapeutic domains, particularly in oncology: reliance on surrogate markers such as tumor response or progression‐free survival has often failed to predict survival or quality‐of‐life improvements.

IV methods preserve the benefits of randomization and avoid biases inherent in within‐arm correlations between surrogate and outcome. While some current approaches examine within‐treatment‐arm correlations, these will be biased if unmeasured factors affect both surrogate and outcome. For example, APOE ε4 carriers may start with greater amyloid, potentially leading to greater removal with lecanemab or donanemab treatment; yet APOE ε4 carriers might be expected to have the most cognitive decline due to other, preexisting biomarker changes downstream in the cascade. In such cases, greater amyloid removal or reductions in accumulation within a treatment arm could misleadingly appear linked to worse cognitive outcomes. In addition, noise in cognitive measures likely largely obfuscates any expected associations or inverse associations between amyloid and cognition. Thus, if we rely solely on methods that do not model differences between treatment arms, we may underestimate the utility of a potential surrogate. IV has the additional benefit of producing estimates free of regression dilution bias in contexts with measurement error in the mediator. Centiloids are measured with error, ³³ and thus the IV estimate produces an un‐attenuated estimate of the effect of amyloid on cognition. In contrast, within‐arm correlations that do not preserve the benefits of randomization would be attenuated even in the absence of confounding by APOE ε4 or other biasing factors.

The use of individual‐level data confers additional advantages. First, while IV analysis is possible with aggregated data, mediation analysis is not. As shown in Figure 2, mediation aids in IV interpretation. Without it, we cannot assess whether a surrogate fully mediates the treatment–outcome relationship. ¹⁷ Estimating the percentage mediated helps determine whether the IV estimate reflects, overstates, or understates the true causal effect. To interpret the IV estimate as the surrogate's causal effect on the outcome, ⁸ , ⁹ , ¹⁰ one must assume full mediation – an assumption that cannot be supported with aggregate data. Second, individual‐level data allow testing for non‐linearity in the effect of the putative surrogate on the outcome, which has important implications for treatment optimization. Third, individual‐level data allow for the assessment of effect heterogeneity in the effect of the putative surrogate on the outcome. Fourth, the IV estimator used with aggregated data is unbiased but loses precision when the change in surrogate overlaps considerably within groups. ⁸ More precise IV results are possible with individual‐level data. Finally, individual‐level adjustment for covariates can also improve IV precision. For drugs that more effectively remove amyloid, this gain may increase the chances of correctly identifying a valid surrogate.

While IV analysis is not common in AD research, the core ideas are a century old and have been both formally ⁹ , ¹⁰ , ¹⁵ and informally adopted in the context of amyloid‐targeting drugs. At the 2022 Alzheimer's Association International Conference, for example, during a well‐attended and controversial discussion on newly approved amyloid‐targeting drugs, Yaning Wang presented a plot of mean cognitive change versus mean amyloid change across multiple trials. ¹⁰ , ¹⁵ , ³⁴ Examining this relationship by treatment arm or trial is both an intuitive way to visualize cross‐trial patterns and a fundamental component of IV analyses.

Yet there are scientific risks associated with using IV in an informal manner, including overlooking key assumptions (e.g., the exclusion restriction assumption and homogeneity in effect per amyloid change across different drugs and populations), failing to quantify uncertainty, and lacking tools to perform formal evaluation of mechanistic effects of interest. For example, Wang argued that his visualization showed a non‐linear relationship – greater amyloid removal or reductions in accumulation yielding proportionally larger cognitive benefits. While plausible, this was not statistically tested and had no measures of uncertainty. A similar plot published by a Biogen research team, for example, did not visually suggest the same non‐linear trend. ¹⁰ Furthermore, arguments via aggregated data visualization obscure non‐linear surrogate‐outcome relationships by showing only study means. Formal statistical estimators can test such claims of non‐linearity: there is no need to “eyeball” patterns. Lastly, trial effects vary for reasons independent of amyloid change. In A4, solanezumab did not show benefit because the placebo group did not clinically progress – results for cognitive outcomes likely would have been null even with a drug that effectively removed amyloid. Given the strong financial incentives to interpret ambiguous results in a particular direction and entrenched scientific camps, rigorous statistical tools, rather than informal interpretations, are needed.

Proposed analyses are not comprehensive since other frameworks evaluate surrogacy ³⁵ , ³⁶ and in a wider range of settings (e.g., for time‐to‐event outcomes ¹⁷ and in meta‐research ³⁷ ). Another group of authors introduced a framework for evaluating amyloid surrogate outcomes, including in individual‐level data of gantenerumab. ¹² However, as the authors note, their methods are a tool for subgroup analysis that does not preserve the benefits of randomization for causal inference. Evaluations of safety and overall efficacy, examining within‐treatment‐arm correlations, and other approaches ¹⁸ , ¹⁹ are still valuable in this context; these approaches should be viewed as complementary to ours because the various approaches require different assumptions. Nonetheless, mediation ¹⁹ and IV analyses are unique in that they can adjust for within‐arm confounding and preserve the benefits of randomization and they are valuable as part of a framework to evaluate amyloid as a surrogate outcome for cognition. Finally, our mediation analyses considered amyloid change as a single mediator of the treatment‐cognition relationship. We did not fit multiple mediator models, and thus we cannot rule out the possibility that the observed associations reflect alternative or additional pathways correlated with amyloid change.

In conclusion, using mediation and IV methods to analyze individual‐level data from trials of amyloid‐targeting drugs can improve our understanding of amyloid as a surrogate outcome for cognition. Their application to the A4 data produced inconclusive results due to study‐specific limitations. A logical next step is to perform these analyses using data that already exist – individual‐level data from recent trials in populations with clinical progression – and using interventions that dramatically change amyloid.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

CONSENT STATEMENT

This work represents a secondary data analysis using deidentified individual‐level data from the A4 study. All study participants provided informed consent at the time of enrollment in the study. IRB approval was obtained at each of the trial sites. The use of deidentified data in the present analysis was determined to be exempt from additional IRB review via consultation with the Brown University IRB staff. These data are publicly available to qualified researchers with a short application at a4studydata.org.

Supporting information

Supporting Information

Supporting Information

Ackley SF, Flanders M, Murchland A, et al. Evaluation of amyloid change as a surrogate for cognitive decline: demonstration in individual‐level data from the A4 study of solanezumab. Alzheimer's Dement. 2026;12:e70205. 10.1002/trc2.70205