📊 statistics

Anchors Away: Navigating Unanchored Indirect Comparisons with Multilevel Unanchored Meta-Regression (ML-UMR)

This paper proposes Multilevel Unanchored Meta-Regression (ML-UMR), a Bayesian framework that extends multilevel network meta-regression to synthesize disconnected individual and aggregate data for estimating treatment effects across multiple populations while explicitly modeling the assumptions required for identification and transportability.

Original authors: Conor Chandler, Jack Ishak

Published 2026-06-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Conor Chandler, Jack Ishak

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Comparing Apples to Oranges (Without a Common Basket)

Imagine you are a health official trying to decide which of two new medicines is better. You have two pieces of evidence, but they are completely disconnected:

Study A tested Medicine A on a group of patients.
Study B tested Medicine B on a different group of patients.

There is no "Study C" where both medicines were tested side-by-side (no common anchor). Furthermore, the patients in Study A and Study B are different. Maybe Study A had older patients, while Study B had younger ones. Maybe Study A had sicker patients.

If you simply compare the results, you don't know if Medicine A is better because the drug works, or just because the patients in Study A were healthier to begin with. It's like trying to decide which car is faster by watching a race between a Ferrari on a smooth track and a minivan on a muddy road. You can't tell if the car is the problem or the road.

The Old Way: "Matching" and "Simulating"

Scientists have tried to fix this using methods called MAIC (Matching-Adjusted Indirect Comparison) and STC (Simulated Treatment Comparison).

The Analogy: Imagine you have a photo of the Ferrari drivers (Study A) and a photo of the Minivan drivers (Study B). To compare them, you try to "match" the drivers. You take the Ferrari drivers and say, "Okay, we will pretend these specific drivers are the same age and weight as the Minivan drivers."
The Flaw: These old methods usually only work for comparing two things at a time (pairwise). More importantly, they often calculate the result based on the original group of drivers (the Minivan group). But the health officials (decision-makers) care about a third group: the general public or a specific country's population. The old methods struggle to say, "Here is how these drugs would perform for our specific people," rather than just the people in the original study.

The New Solution: ML-UMR (The "Universal Translator")

The authors propose a new method called Multilevel Unanchored Meta-Regression (ML-UMR). Think of this as a "Universal Translator" for medical data.

Instead of just matching people, ML-UMR builds a detailed map of how patient characteristics (like age, weight, or disease severity) affect the outcome of the treatment.

The Map (The Model): It uses a mathematical model to draw a line between "Patient Features" and "Health Results."
The Bridge (Marginalization): It takes the data from Study B (which only has summary numbers, not individual details) and uses the map to "reconstruct" what the individual patients might have looked like.
The Destination (Transport): Once the map is built, it can project the results onto any population. It can answer: "If we gave these drugs to the specific people in our country, what would happen?"

Key Feature: It handles multiple treatments and multiple studies at once, creating a unified picture rather than just comparing two things in isolation.

The Catch: The "Shared Prognostic Factor" Assumption

The paper is very honest about a major limitation. To make this map work, you have to make a big guess called the Shared Prognostic Factor Assumption (SPFA).

The Analogy: Imagine you are trying to predict how fast a car goes based on the driver's height. You assume that height affects speed the same way for both the Ferrari and the Minivan.
The Reality: If height actually makes the Ferrari go faster but makes the Minivan go slower, your map is wrong, and your prediction will be biased.
The Paper's Claim: The authors state that ML-UMR does not fix this guess. It doesn't magically make the assumption true. Instead, it makes the guess explicit. It forces the scientists to say, "We are assuming height affects both drugs the same way." If that assumption is wrong, the results will be wrong.

However, the paper shows that if you have extra data (like breaking the groups into subgroups), you can relax this guess and get more accurate results.

What the Simulations Showed

The authors ran computer simulations to test their new method against the old ones. Here is what they found:

The "Home Field" Advantage: All methods (old and new) were good at predicting results for the original study groups. If you just wanted to know how the drugs worked for the people in the original studies, they all worked well.
The "Travel" Problem: When they tried to predict results for a different target population (the decision-makers' population), the old methods (MAIC/STC) often failed, especially if the drugs worked differently for different types of people (effect modification). They produced biased results.
The ML-UMR Win: The new method (ML-UMR) was much better at "transporting" the results to the new population, but only if the big guess (SPFA) was correct or if they had extra subgroup data to fix the guess.
Sample Size Trap: The paper warns that having more data (larger studies) doesn't fix the problem if the underlying assumptions are wrong. In fact, with more data, the method becomes very confident in its wrong answer, leading to a false sense of security.

The Bottom Line

This paper introduces a sophisticated new tool (ML-UMR) for comparing medical treatments when you don't have a head-to-head trial.

It's better than the old tools because it can handle complex situations with multiple drugs and can project results onto the specific population that matters for health policy decisions.
It's not magic. It still relies on strong assumptions about how patient characteristics affect outcomes.
Its main value is transparency. It forces researchers to clearly state their assumptions and test how sensitive their results are if those assumptions are slightly wrong.

In short, it's a more powerful and honest way to navigate the "unanchored" waters of medical evidence, helping decision-makers avoid getting lost when comparing treatments that haven't been tested against each other directly.

Technical Summary: Anchors Away: Navigating Unanchored Indirect Comparisons with Multilevel Unanchored Meta-Regression (ML-UMR)

1. Problem Statement

In Health Technology Assessment (HTA), decision-makers frequently encounter evidence networks that are disconnected, meaning treatments are evaluated in separate studies (e.g., single-arm trials) without a common comparator (such as a placebo or standard of care). In these "unanchored" settings, standard Network Meta-Analysis (NMA) is inapplicable because there is no shared anchor to link the evidence.

Existing population-adjusted indirect comparison (PAIC) methods, primarily Matching-Adjusted Indirect Comparison (MAIC) and Simulated Treatment Comparison (STC), are widely used to address this. However, the authors identify three critical limitations:

Pairwise Restriction: They are generally limited to pairwise comparisons and struggle to synthesize multiple treatments or studies simultaneously.
Target Population Misalignment: They typically estimate effects in the comparator study population. In HTA, the decision-relevant population is often the index study population or a specific jurisdiction. Transporting effects from the comparator to the index population often relies on implicit, untested assumptions.
Lack of Unified Framework: Current methods do not explicitly separate the identification of treatment effects (distinguishing treatment effects from population differences) from the transportability of those effects to a target population. This limits coherent inference across multiple treatments and decision contexts.

2. Methodology: Multilevel Unanchored Meta-Regression (ML-UMR)

The authors propose ML-UMR, a unified Bayesian outcome-regression framework designed to synthesize Individual Patient Data (IPD) and Aggregate Data (AgD) from fully disconnected evidence. It extends Multilevel Network Meta-Regression (ML-NMR) to unanchored settings.

Core Modeling Structure

ML-UMR utilizes a hierarchical outcome regression model where the relationship between outcomes, treatments, and baseline covariates (prognostic factors) is modeled at the individual level, even when only aggregate data is available for some studies.

Outcome Regression: For treatment $k$ , the conditional mean outcome is modeled as:
$g(\theta_k(x)) = \alpha_k + x^\top \beta_k$
where $g(\cdot)$ is a link function, $\alpha_k$ is a treatment-specific intercept, and $\beta_k$ represents covariate effects.
Joint Likelihood: The framework constructs a joint likelihood combining:
1. Individual-Level Component (IPD): Standard likelihood for studies with IPD.
2. Aggregate-Level Component (AgD): For studies with only AgD, the model integrates the conditional outcome model over the reconstructed covariate distribution of that study ( $f_B(x)$ ) to derive the marginal mean. This involves numerical integration (e.g., quasi-Monte Carlo) to avoid aggregation bias, particularly for non-linear link functions.
Identification Assumptions: Valid inference relies on conditional exchangeability across studies and the Shared Prognostic Factor Assumption (SPFA). Under SPFA, covariate effects ( $\beta$ $β$ ) are assumed to be identical across treatments ( $\beta_A = \beta_B$ $β_{A} = β_{B}$ ), implying no treatment-covariate interaction (effect modification).
- Relaxing SPFA: The framework allows for relaxing SPFA if subgroup-level AgD is available, enabling the estimation of treatment-specific covariate effects. Alternatively, structured sensitivity analyses (e.g., tipping point analyses) can be performed to assess the impact of violating SPFA.

Transportability

Once treatment effects are identified, ML-UMR transports them to any decision-relevant target population ( $\tilde{P}$ ) via population-specific standardization (Bayesian g-computation). This involves integrating the estimated outcome model over the covariate distribution of the target population ( $f_{\tilde{P}}(x)$ ).

Conditional Effects: Under SPFA and linear predictors, conditional effects are invariant to the target population.
Marginal Effects: Marginal effects (e.g., risk ratios, odds ratios) generally depend on the target population's covariate distribution and must be re-standardized for each specific population of interest.

Extensions

The paper outlines extensions for:

Multiple Treatments: Generalizing the pairwise formulation to $K$ treatments.
Multiple Studies per Treatment: Incorporating between-study heterogeneity via fixed or random study intercepts.
Survival Outcomes: Adapting the framework for time-to-event data by reconstructing pseudo-IPD from Kaplan-Meier curves and marginalizing over covariate distributions for the aggregate likelihood.

3. Simulation Study Results

The authors conducted a Monte Carlo simulation study comparing ML-UMR (under SPFA and with SPFA relaxed) against MAIC and STC using disconnected single-arm evidence with binary outcomes.

Comparator Population: All methods (ML-UMR, MAIC, STC) successfully recovered treatment effects in the comparator population with low bias and nominal coverage, even when SPFA was violated.
Target (Index) Population Transport:
- MAIC and STC: Exhibited substantial bias and poor coverage when transporting marginal effects (e.g., log risk ratios) to the index population, particularly under population imbalance or effect modification. This highlights the failure of implicit transport assumptions in these methods.
- ML-UMR (SPFA): Performed well when SPFA held but showed increasing bias as SPFA violations (effect modification) and population imbalance increased.
- ML-UMR (Relaxed SPFA): When subgroup data was used to relax SPFA, ML-UMR restored near-unbiased estimation and nominal coverage across all scenarios.
Robustness: Results were robust to moderate misspecification of covariate dependence structures.
Sample Size: Increasing sample size reduced variance but did not reduce structural bias; in fact, it led to tighter credible intervals around biased estimates, exacerbating under-coverage in scenarios with violated assumptions.

4. Key Contributions

Formalization of Assumptions: The paper explicitly distinguishes between identification assumptions (required to separate treatment effects from population differences in disconnected data) and transportability assumptions (required to apply those effects to a specific target population).
Unified Framework (ML-UMR): It introduces a generalizable Bayesian framework that synthesizes IPD and AgD from disconnected evidence, supporting multiple treatments, multiple studies, and multiple target populations simultaneously.
Demonstration of Transportability Risks: Through simulation, the authors demonstrate that while PAIC methods may appear robust in the comparator population, they can yield biased results in decision-relevant populations if transportability assumptions (specifically regarding effect modification and collapsibility) are not explicitly modeled and tested.

5. Significance and Claims

The authors claim that ML-UMR does not eliminate the inherent limitations of unanchored comparisons (which rely on strong, often unverifiable assumptions like SPFA). Instead, it makes these assumptions explicit within a transparent modeling framework.

The significance of this work lies in its alignment with HTA evidentiary requirements:

It allows analysts to estimate treatment effects directly in the decision-relevant target population rather than relying on implicit transport from a comparator population.
It facilitates structured sensitivity analyses to quantify how violations of identifying assumptions (e.g., effect modification) impact cost-effectiveness outcomes and reimbursement decisions.
It provides a principled alternative to current pairwise methods, offering a coherent approach for synthesizing disconnected evidence that is better suited for complex health economic evaluations.

The paper concludes by noting that future work should focus on quantitative bias analysis (QBA) for unanchored comparisons and extending these methods to partially connected networks where some evidence is anchored and some is not.