[Paper Review] Towards causally interpretable meta-analysis: transporting inferences from multiple studies to a target population

Issa J Dahabreh, Lucia C. Petito+3 more|arXiv (Cornell University)|Mar 27, 2019

Advanced Causal Inference Techniques50 references4 citations

TL;DR

This paper proposes methods to transport causal inferences from multiple randomized trials to a well-defined target population using individual participant data from trials and baseline covariates from a target sample. It establishes identification conditions for average treatment effects, introduces estimators based on outcome modeling and inverse probability weighting, and demonstrates through the HALT-C trial that transported estimates can be more relevant than conventional meta-analytic summaries when populations differ.

ABSTRACT

We take steps towards causally interpretable meta-analysis by describing methods for transporting causal inferences from a collection of randomized trials to a new target population, one-trial-at-a-time and pooling all trials. We discuss identifiability conditions for average treatment effects in the target population and provide identification results. We show that assuming inferences are transportable from all trials in the collection to the same target population has implications for the law underlying the observed data. We propose average treatment effect estimators that rely on different working models and provide code for their implementation in statistical software. We discuss how to use the data to examine whether transported inferences are homogeneous across the collection of trials, sketch approaches for sensitivity analysis to violations of the identifiability conditions, and describe extensions to address non-adherence in the trials. Last, we illustrate the proposed methods using data from the HALT-C multi-center trial.

Motivation & Objective

• Address the lack of causal interpretability in standard meta-analyses when trial populations differ from the target population.
• Enable transportability of average treatment effects from multiple randomized trials to a new, well-defined target population.
• Provide identification conditions and estimation methods that account for differences in effect modifier distributions across trials and the target.
• Facilitate evidence synthesis with clear causal interpretation by integrating individual-level data from trials and covariate data from the target population.
• Support sensitivity analysis and homogeneity assessment to evaluate robustness of transported inferences.

Proposed method

• Define average treatment effects in the target population under nonparametric identification conditions, assuming consistency and positivity.
• Propose two main estimators: one based on conditional outcome mean modeling (equation 6) and another using inverse odds weighting (equation 7) for single-trial transport.
• Extend the approach to pool inferences across all trials using a combined estimator (equation 8) and weighting method (equation 9) for the full collection.
• Use inverse probability weighting to adjust for differences in baseline covariate distributions between trial populations and the target population.
• Implement estimators in R with provided code, supporting both parametric working models and doubly robust estimation strategies.
• Apply sensitivity analysis techniques to assess the impact of violations in identifiability assumptions, such as unmeasured confounding or model misspecification.

Experimental results

Research questions

• RQ1 Under what conditions can average treatment effects from multiple randomized trials be causally interpreted in a new target population?
• RQ2 How can we estimate the average treatment effect in a target population when trial participants differ in baseline covariates from the target?
• RQ3 What are the implications for the data-generating process when assuming transportability from all trials to the same target population?
• RQ4 How can we assess whether transported treatment effects are homogeneous across trials or vary systematically?
• RQ5 What methods can be used to evaluate the robustness of transported inferences to violations of identification assumptions?

Key findings

• The proposed methods establish nonparametric identification conditions for average treatment effects in the target population under standard causal assumptions.
• In the HALT-C trial application, transportability analyses yielded more relevant estimates than unadjusted trial-specific results, especially for underrepresented groups.
• The treatment effect estimate transported from all nine centers to the target population (S=0) was -43.7 (95% CI: -52.2, -35.2) using outcome modeling, compared to -45.7 in the target trial itself.
• Weighting-based estimators produced similar results (e.g., -42.4, 95% CI: -52.6, -32.3), suggesting consistency across methods.
• Heterogeneity in transported effects across centers was detected, indicating that trial-level effects are not homogeneous and should not be pooled naively.
• Sensitivity analysis and model diagnostics revealed that assumptions about positivity and correct model specification significantly affect the validity of transported estimates.