Causally-interpretable meta-analysis using aggregate data

arXiv:2605.27272 · stat.ME · Submitted 2026-05-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Causally-interpretable meta-analysis using aggregate data".

Jane: Here's the gist of the paper. Evidence syntheses and meta-analyses are used to inform clinical practice guidelines and health economic evaluations by estimating summary results that integrate treatment effect estimates from multiple randomized trials The method proposes a new version of CIMA that only requires aggregate data from trials, addressing the limitations of traditional meta-analysis methods while relying only on aggregate data The proposed method leverages trial-reported estimates to build moment equations for identifying and estimating a parametric conditional average treatment effect (CATE) function The identification of the CATE function is achieved by leveraging information from both marginal and one-at-a-time subgroup treatment effects from trials, and descriptive summary statistics of baseline covariates The method involves constructing moment equations to estimate the parameters of a parametric CATE function via the generalized method of moments (GMM) The average treatment effect in a new target population is obtained by marginalizing the CATE function over the target population covariate distribution Identification and Estimation Framework The framework establishes a causal link between trial results and a target-population estimand by learning a parametric approximation to the conditional average treatment effect (CATE) function from aggregate data The method relies on assumptions A1 and A3 to express the target population average treatment effect as E

g(X)S = 0: , where g(x) is the CATE function The true parameter θ0 is identified as the unique element of Θ satisfying Z α(x; η0) g(x; θ) dQ(x) − τ0 = 0J Performance and Application in Simulation A simulation study was conducted to examine the finite-sample behavior of CIMAgD under stylized data-generating mechanisms, comparing it against standard meta-analysis and individual participant data (IPD) benchmarks The method demonstrated good finite-sample performance with coverage ranging from approximately 94% to 97% in the multi-trial setting Clinical Relevance and Decision Making The application of the method to SGLT2 inhibitors in patients with heart failure demonstrated that the proposed method can quantify treatment-effect heterogeneity across clinically relevant subgroups in a new target population The results suggest treatment-effect heterogeneity across clinically relevant subgroups in the target population and show how the proposed method can be used to quantify such heterogeneity for decision-making The approach avoids test-based decisions and directly targets a well-defined estimand for decision-making The proposed method extends causally interpretable evidence synthesis to settings where individual-level trial data are unavailable while retaining a more explicit causal identification and inferential framework The resulting estimator remains grounded in a clearly defined target population and supports standard large-sample inference under the stated assumptions This framework may be particularly useful for clinical decision-making, guideline development, and health technology assessment The method relies on subgroup treatment effects reported by the trials to identify the CATE function In conclusion, we developed a method for estimating average treatment effects in a user-specified target population using aggregate data from randomized trials Appendix Details on Estimator Structure The asymptotic linear representation of the CATE estimator explicitly shows the first-order propagation of uncertainty from multiple data sources The variance estimation involves a plug-in estimator for the sampling variance that accounts for dependencies across different data sources The asymptotic linear representation shows that the first-order uncertainty arises from three components: (1) estimation of the stacked moments using the base distribution Q; (2) estimation of the treatment effects within each trial; and (3) estimation of the exponential-tilting parameters using information from both the trials and the base distribution Appendix Details on Relative Treatment Effects The method can be adapted to estimate indirect treatment comparisons in a target population by using relative treatment effects The indirect comparison estimand is defined as ψ12 ≡ E

Y (2) − Y (1) S = 0: Issa J Dahabreh et al., Toward causally interpretable meta-analysis: transporting inferences from multiple randomized trials to a new target population.

Tom: Epidemiology, 31(3):334–344, 2020a Issa J Dahabreh et al., Study designs for extending causal inferences from a randomized trial to a target population.

Jane: American journal of epidemiology, 190(8):1632–1642, 2021 Issa J Dahabreh et al., Efficient and robust methods for causally interpretable meta-analysis: Transporting inferences from multiple randomized trials to a target population.

Tom: Biometrics, 79(2):1057–1072, 2023 Muthiah Vaduganathan et al., Sglt2 inhibitors in patients with heart failure: a comprehensive meta-analysis of five randomised controlled trials.

Paper summary: Jane: So, to recap the paper "Causally-interpretable meta-analysis using aggregate data", they argue that conventional meta-analyses struggle when there’s a lot of treatment effect heterogeneity because their random-effect assumptions don't hold up in reality >

Tom: Right. The core thesis is that they provide a rigorous framework, CIMA, for specifying and identifying the causal effect in a new target population using only aggregate data from randomized trials >

Lu: They establish how to link the trial results to that target population estimand by learning a parametric approximation of the CATE function from those aggregate statistics >

Meng: The key claim seems to be that you can get causally-interpretable indirect treatment comparisons in the target population, which is useful for guiding clinical decisions >

Lalam: They also establish the asymptotic properties of this estimation procedure, which means they’ve looked at what happens as the number of trials gets really big >

Tom: So it's moving beyond just estimating an average effect across all trials and trying to account for how that average changes depending on where you are in terms of patient characteristics >

Jane: It matters because it offers a way to quantify treatment effect heterogeneity in a new target population without relying on assumptions about the distribution of those effects across trials >

Conclusion: Tom: So, looking at this paper "Causally-interpretable meta-analysis using aggregate data", it’s really about taking evidence from multiple randomized trials and making it causal in a specific group of people >

Jane: The authors are trying to show that this approach keeps the explicit causal identification and inferential framework even when you don't have individual patient trial data available >

Lu: It’s useful because it supports standard large-sample inference under the assumptions they set up, making it applicable for guideline development and health technology assessment >

Meng: From an engineering standpoint, they’ve created a new R package, CIMAgD that you can actually use to implement this methodology >

Lalam: So the paper offers a method for estimating average treatment effects in a user-specified target population using just aggregate data from randomized trials >

University of Groningen · University Medical Center Utrecht · Groningen Research Institute of Pharmacy, University of Groningen · Dutch National Institute of Public Health and the Environment (RIVM) · Harvard T.H. Chan School of Public Health

stat.ME

Submitted: 2026-05-26

Updated: 2026-11-16

Comments: Updated version

Code: https://github.com/qingyshi/CIMAgD

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: Evidence syntheses and meta-analyses are used to inform clinical practice guidelines and health economic evaluations by estimating summary results that integrate treatment effect estimates from…

Key concepts

Conditional Average Treatment Effect (CATE)
The CATE is a function that describes how the average treatment effect changes depending on an individual's characteristics. This method tries to learn this function using aggregate trial data, which helps in understanding treatment heterogeneity across different patient groups.
Generalized Method of Moments (GMM)
GMM is a statistical technique used here to estimate the parameters of the CATE function. It works by constructing moment equations based on the observed data and then finding the parameter values that make these moments match, helping to accurately model how treatment effects vary.
Target Population Estimand
This refers to a specific average treatment effect that researchers want to estimate for a defined group of patients, such as heart failure patients. The method aims to connect the results from many different trials back to this specific target group's outcome.
Causal Interpretation
The framework provides a way to draw causal inferences even when individual patient data from trials are missing. It establishes a clear link between trial results and the target population by using assumptions that ensure the resulting estimate is causally interpretable.

Terminology

Summary

Evidence syntheses and meta-analyses are used to inform clinical practice guidelines and health economic evaluations by estimating summary results that integrate treatment effect estimates from multiple randomized trials The method proposes a new version of CIMA that only requires aggregate data from trials, addressing the limitations of traditional meta-analysis methods while relying only on aggregate data

Methodology for Aggregate Data Utilization

The proposed method leverages trial-reported estimates to build moment equations for identifying and estimating a parametric conditional average treatment effect (CATE) function The identification of the CATE function is achieved by leveraging information from both marginal and one-at-a-time subgroup treatment effects from trials, and descriptive summary statistics of baseline covariates The method involves constructing moment equations to estimate the parameters of a parametric CATE function via the generalized method of moments (GMM) The average treatment effect in a new target population is obtained by marginalizing the CATE function over the target population covariate distribution

Identification and Estimation Framework

The framework establishes a causal link between trial results and a target-population estimand by learning a parametric approximation to the conditional average treatment effect (CATE) function from aggregate data The method relies on assumptions A1 and A3 to express the target population average treatment effect as E[g(X)S = 0], where g(x) is the CATE function The true parameter θ0 is identified as the unique element of Θ satisfying Z α(x; η0) g(x; θ) dQ(x) − τ0 = 0J

Performance and Application in Simulation

A simulation study was conducted to examine the finite-sample behavior of CIMAgD under stylized data-generating mechanisms, comparing it against standard meta-analysis and individual participant data (IPD) benchmarks The method demonstrated good finite-sample performance with coverage ranging from approximately 94% to 97% in the multi-trial setting

Clinical Relevance and Decision Making

The application of the method to SGLT2 inhibitors in patients with heart failure demonstrated that the proposed method can quantify treatment-effect heterogeneity across clinically relevant subgroups in a new target population The results suggest treatment-effect heterogeneity across clinically relevant subgroups in the target population and show how the proposed method can be used to quantify such heterogeneity for decision-making The approach avoids test-based decisions and directly targets a well-defined estimand for decision-making

Conclusion on Causal Interpretation

The proposed method extends causally interpretable evidence synthesis to settings where individual-level trial data are unavailable while retaining a more explicit causal identification and inferential framework The resulting estimator remains grounded in a clearly defined target population and supports standard large-sample inference under the stated assumptions This framework may be particularly useful for clinical decision-making, guideline development, and health technology assessment The method relies on subgroup treatment effects reported by the trials to identify the CATE function In conclusion, we developed a method for estimating average treatment effects in a user-specified target population using aggregate data from randomized trials

Appendix Details on Estimator Structure

The asymptotic linear representation of the CATE estimator explicitly shows the first-order propagation of uncertainty from multiple data sources The variance estimation involves a plug-in estimator for the sampling variance that accounts for dependencies across different data sources The asymptotic linear representation shows that the first-order uncertainty arises from three components: (1) estimation of the stacked moments using the base distribution Q; (2) estimation of the treatment effects within each trial; and (3) estimation of the exponential-tilting parameters using information from both the trials and the base distribution

Appendix Details on Relative Treatment Effects

The method can be adapted to estimate indirect treatment comparisons in a target population by using relative treatment effects The indirect comparison estimand is defined as ψ12 ≡ E[Y (2) − Y (1) S = 0] Issa J Dahabreh et al., Toward causally interpretable meta-analysis: transporting inferences from multiple randomized trials to a new target population. p. 1 Epidemiology, 31(3):334–344, 2020a Issa J Dahabreh et al. p. 0, Study designs for extending causal inferences from a randomized trial to a target population. p. 2 American journal of epidemiology, 190(8):1632–1642, 2021 Issa J Dahabreh et al. p. 0, Efficient and robust methods for causally interpretable meta-analysis: Transporting inferences from multiple randomized trials to a target population. p. 1 Biometrics, 79(2):1057–1072, 2023 Muthiah Vaduganathan et al. p. 2, Sglt2 inhibitors in patients with heart failure: a comprehensive meta-analysis of five randomised controlled trials. p. 1 The Lancet, 400(10354):757–767, 2022 Thomas M. p. 5 Cover and Joy A. p. 5 Thomas. Maximum Entropy, chapter 12, pages 409–426. John Wiley & Sons, Hoboken, New Jersey, 2nd edition, 2006 Lars Peter Hansen. Large sample properties of generalized method of moments estimators. p. 7 Econometrica: Journal of the econometric society, pages 1029–1054, 1982 Issa J Dahabreh and Miguel A Hernán. p. 0 Extending inferences from a randomized trial to a target population. p. 1 European journal of epidemiology, 34(8):719–722, 2019 Issa J Dahabreh et al. p. 0, Extending inferences from a randomized trial to a new target population. p. 1 p. 1 Statistics in medicine, 39(14):1999–2014, 2020b Tim P Morris, Ian R White, and Michael J Crowther. Using simulation studies to evaluate statistical methods. p. 1 Statistics in medicine, 38(11):2074–2102, 2019 Leana Bellanca et al. p. 8, Incidence and prevalence of heart failure in england: a descriptive analysis of linked primary and secondary care data - the pulse study. p. 2 BMC Cardiovascular Disorders, 23(1):374, 2023 James E Signorovitch et al. p. 1, Matching-adjusted indirect comparisons: a new tool for timely comparative effectiveness research. p. 1 Value in Health, 15(6):940–947, 2012 Antonio Remiro-Azócar et al. p. 0, Parametric g-computation for compatible indirect treatment comparisons with limited individual patient data. p. 1 Research synthesis methods, 13(6):716–744, 2022 David M Phillippo et al. p. 0, Multilevel network meta-regression for population-adjusted treatment comparisons. p. 1 Journal of the Royal Statistical Society Series A: Statistics in Society, 183(3):1189–1210, 2020 Kollin W Rott et al., Causally interpretable meta-analysis combining aggregate and individual participant data. p. 0 American journal of epidemiology, 194(7): 2060–2068, 2025 Tat-Thang Vo et al. p. 0, Integration of aggregated data in causally interpretable meta-analysis by inverse weighting. p. 0 arXiv preprint arXiv: 2503. p. 005634, 2025 Issa J Dahabreh and Miguel A Hernán. p. 0 Extending inferences from a randomized trial to a new target population. European journal of epidemiology, 39(9): 957, 2024 Whitney K Newey and Daniel McFadden. p. 0 Large sample estimation and hypothesis testing. p. 8 Handbook of econometrics, 4:2111–2245, 1994

Simulation Study Details

The simulation study explored the method's performance in both multi-trial and single-trial settings using stylized data-generating mechanisms that mimic real-world heterogeneity The simulation results showed that CIMAgD performed well, with coverage ranging from approximately 94% to 97% in both settings

Application Results

The application of the method to SGLT2 inhibitors in patients with heart failure showed that the estimated treatment effect ranged from a reduction of 58 events per 1000 patients with LVEF ≤ 40% to a reduction of 28 events per 1000 patients with LVEF > 40% 40%, no history of HHF, and no diabetes> This suggests treatment-effect heterogeneity across clinically relevant subgroups in the target population The proposed method produced estimates that were close to the target-population estimand and performed similarly to the IPD g-formula benchmark

Limitations of Aggregate Data Reliance

A limitation noted is that the method relies on subgroup treatment effects reported by the trials to identify the CATE function In practice, trials may omit subgroup results, which can lead to identification failure or unstable estimates if the available subgroup information is sparse or poorly aligned with the chosen CATE model The method requires large-sample behavior within trials rather than a large number of trials for its validity The method extends causally interpretable evidence synthesis to settings where individual-level trial data are unavailable while retaining a more explicit causal identification and inferential framework The authors thank Veerle Coupé, Paul Gustafson, Spyros Balafas, and Steef Konings for their helpful comments and feedback on an early stage of this work The code for the simulation study is publicly accessible at https://github.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Transport causal inferences from multiple randomized trials to a new target population by learning a parametric approximation to a conditional average treatment effect function from aggregate data. This allows for estimating the average treatment effect in a clearly defined target population even when individual patient data is unavailable.

  2. Identify and estimate the canonical CATE function using aggregate data by leveraging information from both marginal and one-at-a-time subgroup treatment effects reported by trials, along with descriptive summary statistics of baseline covariates. This method builds moment equations for identifying and estimating a parametric conditional average treatment effect function via the generalized method of moments.

  3. Obtain causally interpretable indirect treatment comparisons in a target population by extending the method to estimate E[Y (2) − Y (1) S = 0] using estimated parametric CATE functions from different trials. This provides a direct comparison of two interventions in the target population without requiring direct head-to-head trials.

  4. Perform transportability analyses for single trials by estimating time stratified causal effects, allowing for the estimation of corresponding average treatment effects at specific follow-up timepoints in the target population. This is achieved by stratifying the CATE function on discrete followup timepoints.

  5. Estimate the variance of causal effect estimators using a plug-in estimator derived from the asymptotic linear representation, which accounts for uncertainty from multiple data sources and trial-level estimates. This provides robust standard errors and confidence intervals for both the estimated CATE parameters and the target population average treatment effect.

Sources

Related papers