Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment

arXiv:2602.17041 · stat.ME · Submitted 2026-02-19 · Read on arXiv

Listen

Transcript

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

Tom: We're looking at the paper: "Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem".

Jane: Here's the gist of the paper. Detailed Summary of "Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment" This paper formalizes the theoretical underpinnings of Population-Adjusted Indirect Comparisons (PAICs) by reframing the core challenge not merely as an adjustment procedure, but as a problem of transportability—determining when an estimated treatment effect learned in one patient population can be validly applied to a different, decision-relevant target population.

Tom: This focus is crucial for Health Technology Assessment (HTA), where PAIC-derived relative effects are routinely used in cost-effectiveness models defined for varied target populations.

Jane: Core Concepts: Collapsibility vs.

Tom: Transportability The paper establishes a critical distinction between two related but distinct concepts: 1.

Title and authors: Tom: Let's talk about that title again, "Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment."

Jane: It’s a big shift because they are reframing PAICs not just as an adjustment technique, but as a problem of transportability.

Lu: They are essentially saying that the real challenge isn't just calculating the adjustment; it's determining when an effect learned in one patient group is actually valid to apply to a different, more relevant target population.

Meng: So instead of thinking about how to clean up the data, we need to think about whether that cleaned-up number can actually travel from study A to decision-making group B.

Tom: That’s the core idea, and it ties directly into health technology assessment where we need evidence that applies beyond just the specific trial population.

Jane: They are using this transportability concept to guide how we use these PAIC-derived relative effects in cost-effectiveness models defined for varied target populations.

Lu: It’s a methodological refinement, focusing on the underlying causal estimand and its relationship with the covariate distribution of different populations.

Meng: So, if you're building a model that needs to predict outcomes for patients outside the trial group, you need to know if your PAIC estimate is transportable or not.

Tom: And the authors show that this concept forces us to look closely at what kind of effect we are actually estimating—marginal versus conditional.

Jane: That distinction is crucial because it dictates whether the underlying mathematical properties support a simple adjustment or require more complex modeling, like those PAIC methods they review.

The paper's summary: Tom: So what’s the actual summary of this paper? It boils down to clarifying the relationship between collapsibility and transportability.

Jane: They summarize that collapsibility is a property that holds within a single population, while transportability is about transferring an estimand across different populations.

Lu: The key finding they highlight is that marginal effects are typically population-dependent because they involve averaging after a nonlinear transformation, unlike conditional effects which might be invariant in indirect comparisons.

Meng: So, if you're looking at a hazard ratio and it’s not collapsible, you should expect it to be population-dependent unless the effect modifiers are shared.

Tom: That leads into the central difficulty: unadjusted transport is often invalid because we don't account for these population differences in how those effects aggregate.

Jane: They show that direct transportability only holds under very narrow conditions, like when there’s no effect modification by covariates whose distributions differ between populations.

Lu: When those conditions aren't met, they suggest using methods like MAICs or STCs to model the conditional transportability instead of relying on simple unadjusted comparisons.

Meng: I need to know how much extra work this modeling adds before I can deploy one of these PAIC methods into a real production environment.

Tom: The paper points out that even when effect modifiers are shared, marginal effects are still generally population-dependent for common measures like odds ratios and hazard ratios.

Jane: So the practical implication is that we need to be transparent about the assumptions we make about those covariates before we claim our PAIC estimate applies broadly.

The paper's improvements: Tom: Now let's look at what improvements the paper suggests for applying these PAIC methods.

Lu: The authors are suggesting that we need to be more rigorous in checking the four criteria—collapsibility, estimand type, shared effect modifier assumption, and scale alignment—before we use any PAIC estimate.

Meng: So the improvement is essentially a checklist: if you don't meet those criteria, you shouldn't blindly apply the comparison.

Tom: And they are suggesting that for non-collapsible measures like hazard ratios or odds ratios, we should be upfront that they are generally population-dependent and not directly transportable.

Jane: They are advocating for a more transparent approach where we explicitly state when an estimate is just populationspecific rather than decision-relevant.

Lu: For the methods, they suggest selecting specific PAIC tools based on what you need to achieve—for instance, Mean Difference estimation might be better if you want direct transportability under SEMA and scale alignment.

Meng: That sounds like we’re moving toward automated model selection for PAIC methods based on the required transportability properties.

Tom: And they are also suggesting that when dealing with network meta-analysis or simulated treatment comparisons, we need to be more careful about how they handle baseline risk heterogeneity between those structures.

Jane: It’s a call for more robust modeling, especially when you're dealing with unanchored comparisons, because standard MAIC or STC estimates often can't be transported beyond the comparator population without further assumptions.

Conclusion: Tom: To wrap up, the main implication is that applying PAIC-derived treatment effects depends on four things: the type of effect modification, whether you target a marginal or conditional estimand, if you assume shared effect modifiers across treatments, and the alignment between scales.

Jane: So we’re ending up with a framework that tells us exactly when to trust an estimate versus when it should be interpreted as populationspecific rather than decision-relevant.

Lu: It emphasizes that for non-collapsible measures like odds ratios or hazard ratios, SEMA alone is not enough to guarantee direct transportability.

Meng: So the practical warning is that if you're relying on SEMA alone, you should flag the risk of introducing bias in Step two transport when moving to a target population <ref:2602.17041#pg1>.

Tom: And we’re suggesting that for non-collapsible measures, using methods like ML-UMR might be a better path forward than just sticking with the standard tools.

Jane: It’s a call for more principled use of indirect evidence in health technology assessment, making sure we aren't applying estimates where they don't belong.

Lu: This paper sets the foundation for how we need to think about these comparisons when they interact with real-world heterogeneity.

Thermo Fisher Scientific

stat.ME

Submitted: 2026-02-19

Updated: 2026-10-28

Comments: 26 pages (excluding supplement and references), 7 figures, 1 table

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

Key concepts

Collapsibility: This is a property of an effect measure that describes how conditional effects within a single, fixed patient population combine to form a marginal effect. It is an internal characteristic of the measure itself, focusing on how different covariate distributions aggregate within one specific study group.
Transportability: This addresses whether a causal effect estimated in one source population can be reliably applied to a different target population with different characteristics. It requires re-expressing or re-averaging the estimate under the new distribution, making it a cross-population property.
Marginal Effects vs. Conditional Effects: Marginal effects are generally population-dependent because averaging outcomes after transformation changes depending on the covariate distribution in each population. Conditional effects may be more robust and possess better transportability properties under certain modeling assumptions.

Terminology

Health Technology Assessment — 19× in this paper
Causal Inference — 9× in this paper · explained in 5 episodes
Generalizability — 5× in this paper · explained in 6 episodes
Core — 3× in this paper · explained in 2 episodes
Technical Support Document — 3× in this paper
National Institute for Health and Care Excellence — 2× in this paper
Robustness — 2× in this paper · explained in 6 episodes
Clarity — 1× in this paper · explained in 3 episodes
Commentary — 1× in this paper · explained in 2 episodes
Composition — 1× in this paper · explained in 2 episodes
Decision Support Unit — 1× in this paper
Elias — 1× in this paper · explained in 2 episodes
European Union — 1× in this paper
Exchangeability — 1× in this paper · explained in 2 episodes
Interpretability — 1× in this paper · explained in 15 episodes
Paper Review — 1× in this paper · explained in 3 episodes
Structural Properties — 1× in this paper · explained in 2 episodes

Summary

Detailed Summary of Reframing Population-Adjusted Indirect Comparisons as a Transportability Problem: An Estimand-Based Perspective and Implications for Health Technology Assessment

This paper formalizes the theoretical underpinnings of Population-Adjusted Indirect Comparisons (PAICs) by reframing the core challenge not merely as an adjustment procedure, but as a problem of transportability—determining when an estimated treatment effect learned in one patient population can be validly applied to a different, decision-relevant target population. This focus is crucial for Health Technology Assessment (HTA), where PAIC-derived relative effects are routinely used in cost-effectiveness models defined for varied target populations.

Core Concepts: Collapsibility vs. Transportability

The paper establishes a critical distinction between two related but distinct concepts:

Collapsibility: This is a within-population property of an effect measure. It concerns how covariate-specific (conditional) treatment effects aggregate to form marginal effects within a single, fixed population defined by a specific covariate distribution.
Transportability: This is a cross-population property of an estimand. It addresses the fundamental question: under what conditions can a causal effect estimated in one source population (P 1) be validly applied to another target population (P 2) that possesses a different covariate distribution? Transportability inherently requires re-expressing or re-averaging the estimand under the new covariate distribution.

The Challenge of Population Dependence in Estimands

The central difficulty arises because treatment effects themselves are often population-dependent. The paper distinguishes between:

Marginal Effects: These involve averaging outcomes after a typically nonlinear transformation. They are generally population-dependent, meaning Marg P 1 not equal to Marg P 2 when covariate distributions differ across populations, making unadjusted transport invalid.
Conditional Effects: These effects may be invariant in indirect comparisons under appropriate modeling assumptions, offering more favorable transportability properties.

The paper demonstrates that even when effect modifiers (covariates) are shared across treatments, marginal effects are generally population-dependent for commonly used non-collapsible measures like hazard ratios and odds ratios. Conversely, collapsible and conditional effects defined on the linear predictor scale exhibit more favorable transportability properties.

Conditions for Direct Transportability

The paper rigorously defines the conditions under which a treatment effect estimand is directly transportable from a source population (P 1) to a target population (P 2). Direct transportability holds if:

Marg P 1 to P 2 = Marg P 1 = Marg P 2

This equality is generally only met under restrictive conditions, such as the absence of effect modification by covariates whose distributions differ between populations, or when the relevant effect-modifying covariates share the same distribution in both P 1 and P 2. Differences in baseline characteristics (e.g., age, disease severity) frequently induce treatment effect heterogeneity, violating these conditions and invalidating unadjusted transport.

Formalizing Conditional Transportability via PAIC Methods

When direct transportability fails, the paper focuses on conditional transportability, which is achieved through specific PAIC methodologies designed to model treatment effect heterogeneity:

The paper reviews three primary methods for implementing PAICs, highlighting their different targets and mechanisms:

Matching-Adjusted Indirect Comparisons (MAICs):

Mechanism: Estimates balancing weights using the method of moments and utilizes Equation 12 to conditionally transport marginal effects from the index study population (P 1) to the comparator study population.
Target: Inherently targets marginal effects.

Simulated Treatment Comparisons (STCs):

Mechanism: Can target both marginal and conditional effects. When targeting marginal effects, STC conditionally transports estimates from P 1 to the comparator study population using g-computation, as defined in Equation 11.

Multilevel Network Meta-Regression (ML-NMR):

Mechanism: Extends g-computation (Equation 11) to accommodate multiple arms and studies (Equation 26). It is designed to be robust, allowing results to be conditionally transported to any target population under additional modeling assumptions.

Conclusion and Implications for HTA

The paper concludes that the validity of applying PAIC-derived treatment effects depends jointly on four factors:

The type of effect modification (collapsible vs. non-collapsible).
The estimand being targeted (marginal vs. conditional).
Whether the relevant effect modifiers are shared across treatments (Shared Effect Modifier Assumption - SEMA).
The alignment between the scale of effect modification and the chosen effect measure.

The findings clarify precisely when applying PAIC estimates to desired target populations is justified, when additional assumptions are necessary, and when PAIC estimates should instead be interpreted as populationspecific rather than decision-relevant. This research strongly supports a more transparent and principled use of indirect evidence in HTA and related decision-making contexts.

Improvements for AI systems

Improved evidence synthesis for health technology assessment (HTA) by explicitly distinguishing between conditional and marginal estimands in population-adjusted indirect comparisons (PAICs). The system will now determine if a PAIC-derived relative effect is directly transportable based on three joint criteria: the type of effect modification, the collapsibility of the measure, and alignment between the scale of effect modification and the effect measure.
Enhanced decision support for cost-effectiveness models by identifying when PAIC estimates are population-specific rather than decision-relevant. The system can flag instances where applying a PAIC estimate to a target population requires additional assumptions or where it should instead be interpreted as populationspecific rather than decision-relevant.
More transparent and principled application of indirect evidence in HTA by automating the identification of the two-step transport process in pairwise PAICs. The system will differentiate between Step 1 (conditional transport to the comparator population) and Step 2 (implicit direct transport to a decision-relevant population), highlighting where SEMA is sufficient versus where it is insufficient for direct transportability.
More robust estimation of treatment effects across diverse patient populations by providing specific guidance based on outcome type and effect measure. For non-collapsible measures like hazard ratios or odds ratios, the system will alert users that they are generally population-dependent and not directly transportable, even when SEMA holds.
Automated model selection for PAIC methods based on required transportability properties. The system will suggest specific methods like Mean Difference estimation if the goal is direct transportability under SEMA and scale alignment, contrasting them with methods that rely on assumptions that may fail due to non-collapsibility or scale misalignment.
Improved handling of baseline risk heterogeneity in network meta-analysis (NMA) and simulated treatment comparisons (STC). The system will recognize the distinction between STC's single intercept anchored to the source population and ML-NMR's ability to incorporate study specific intercepts, allowing it to better assess uncertainty when transportability relies on these structural differences.
Advanced identification of unanchored comparison challenges through extension methods. For unanchored PAICs, the system will recognize that standard MAIC/STC estimates typically cannot be transported beyond the comparator population unless ΔBC is directly transportable, prompting a switch to more complex frameworks like ML-UMR and its shared prognostic factor assumption (SPFA).
Risk assessment of model misspecification during transport steps. The system will assess the risk that Step 2 transport bias is introduced if SEMA fails, noting that SME alone is not sufficient to guarantee direct transportability of marginal treatment effects for non-collapsible measures like odds ratios and hazard ratios.

Abstract

Population-adjusted indirect comparisons (PAICs) are widely used to synthesize evidence when randomized controlled trials enroll different patient populations and head-to-head comparisons are unavailable. Although PAICs adjust for observed population differences across trials, adjustment alone does not ensure transportability of estimated effects to decision-relevant populations for health technology assessment (HTA). We examine and formalize transportability in PAICs from an estimand-based perspective. We distinguish conditional and marginal treatment effect estimands and show how transportability depends on effect modification, collapsibility, and alignment between the scale of effect modification and the effect measure. Using illustrative examples, we demonstrate that even when effect modifiers are shared across treatments, marginal effects are generally population-dependent for commonly used non-collapsible measures, including hazard ratios and odds ratios. Conversely, collapsible and conditional effects defined on the linear predictor scale exhibit more favorable transportability properties. We further show that pairwise PAIC approaches typically identify effects defined in the comparator population and that applying these estimates to other populations entails an additional, often implicit, transport step requiring further assumptions. This has direct implications for HTA, where PAIC-derived effects are routinely applied within cost-effectiveness and decision models defined for different target populations. Our results clarify when applying PAIC-derived treatment effects to desired target populations is justified, when doing so requires additional assumptions, and when results should instead be interpreted as population-specific rather than decision-relevant, supporting more transparent and principled use of indirect evidence in HTA and related decision-making contexts.