Causally Invalid Methods Make Misleading Arguments for Surrogate Outcomes

Michael D. Flanders, Michelle Caunca, Renaud La Joie, Lon S. Schneider, Sarah F. Ackley

---

Background

Individual participant data (IPD) from randomized controlled trials provide a rare opportunity to rigorously evaluate candidate surrogate outcomes.

A recently reported analytic approach—merge-and-aggregate, also called quantile aggregation—was applied to IPD from TRAILBLAZER-ALZ 2 and found a strong relationship between post-treatment amyloid and cognitive change. The analysis was interpreted as supporting amyloid as a surrogate endpoint.

However, IPD from randomized trials can only provide reliable evidence about surrogate outcomes when analyses preserve the causal information created by randomization.

---

Objective

To characterize and test the recently reported merge-and-aggregate method:

• Does it provide reliable evidence in support of amyloid or other candidate surrogate outcomes?

• Does it produce interpretable estimates of the amyloid–cognition relationship?

---

Approach

Reconstructing the Merge-and-Aggregate Procedure

1. Pool treatment and control arms.

2. Rank all participants by post-treatment amyloid.

3. Partition participants into quantiles.

4. Summarize each quantile using:

   • Median amyloid

   • Mean cognitive change

5. Regress the quantile-level summaries to estimate the amyloid–cognition association, reported as R².

Testing the Method

We evaluated the approach in two settings:

1. Simulated trial data

• 1,600 participants

• Calibrated to recent anti-amyloid trials

• Included only a weak amyloid–cognition relationship by design

• Ground truth was known

2. Randomized clinical trial data

• Individual-level data from the A4 trial of solanezumab

• Solanezumab showed no cognitive benefit in A4

---

Why the Method Is Causally Problematic

• Pooling trial arms discards treatment-allocation information, transforming a causally interpretable randomized trial into an observational analysis of unclear meaning.

• Grouping participants by a post-treatment variable—amyloid—creates an ecological analysis vulnerable to confounding.

• The reported R² values do not measure the strength of the amyloid–cognition relationship in the underlying individual-level data.

• Apparent variance in cognition “explained” by amyloid can arise as a statistical artifact of the aggregation procedure itself.

---

Findings

1. Simulated Trial Data

Even when the true amyloid–cognition relationship was intentionally weak, merge-and-aggregate produced a strong apparent association.

Interpretation: The method can generate persuasive-looking relationships that do not reflect the known underlying causal structure.

2. Real Anti-Amyloid Trial Data

The method produced a similar apparent amyloid–cognition relationship in:

• TRAILBLAZER-ALZ 2, a successful donanemab trial

• A4, a solanezumab trial with no demonstrated cognitive benefit

Interpretation: Because the method generates similar findings from both a successful and an unsuccessful drug, it cannot reliably show that drug effects on cognition are explained by effects on amyloid.

---

Conclusions

• Merge-and-aggregate does not preserve the causal information provided by randomization.

• A causally interpretable randomized trial can become an observational, ecologically aggregated analysis after participants are grouped by post-treatment amyloid.

• Strong R² values from the aggregated analysis can be artifacts of the method rather than evidence that amyloid is a valid surrogate outcome.

• Reliable surrogate evaluation requires individual-level analyses that preserve randomization and use established causal methods.

Individual participant data from randomized trials remain among the strongest sources of evidence for identifying clinically reliable surrogate outcomes—but only when analyzed with causally valid methods.

---

Scan the QR code for our recent JAMA Neurology publication, including additional details and contact information.

---

References

1. Lu M, et al. JAMA Neurology. 2025;82(12):1251–1256. doi:10.1001/jamaneurol.2025.3869

2. Ackley SF, et al. Alzheimer’s & Dementia: Translational Research & Clinical Interventions. 2026;12(1):e70205. Published February 14, 2026. doi:10.1002/trc2.70205

3. Flanders MD, et al. JAMA Neurology. Published online May 18, 2026. doi:10.1001/jamaneurol.2026.1240

---

Funding and Data Source

Funding: NIH/NIA R00AG073454.

Data Source: A4 Study and A4/LEARN Data Package. Additional information and data are available at a4studydata.org.

A4/LEARN was funded by the National Institute on Aging/NIH, Eli Lilly and Co., the Alzheimer’s Association, GHR Foundation, and additional philanthropic donors.