📊 statistics

On the magnitude, sign and ranking of recanting-twin path-specific effects

This paper systematically evaluates the reliability of recanting-twin path-specific effects as approximations for natural path-specific effects by deriving non-parametric bounds on their magnitude differences and demonstrating through simulations that sign and ranking disagreements are frequent and heavily influenced by unobservable correlations in intermediate confounders, particularly in non-monotonic settings.

Original authors: Tran Trong Khoi Le, Pham Hien Trang Tu, Nhat Long Ngo, Tat-Thang Vo

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tran Trong Khoi Le, Pham Hien Trang Tu, Nhat Long Ngo, Tat-Thang Vo

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

Imagine you are trying to figure out exactly how a new video game console (the "treatment") makes you play better (the "outcome"). You know the console helps, but how? Does it work because the controller feels better (Mediator A)? Or because the graphics look sharper (Mediator B)? Or maybe the better graphics make the controller feel even more responsive, creating a chain reaction? This is the heart of causal mediation analysis: a branch of statistics that tries to untangle the specific pathways of cause and effect.

Usually, scientists can map these paths easily. But sometimes, things get messy. Imagine that the better graphics (Mediator A) actually change how the controller works (Mediator B). Now, the two mediators are tangled up. In the world of statistics, this is called "intermediate confounding," and it creates a logical paradox known as the "recanting witness." It's like asking a witness to testify about two different timelines at the same time, which breaks the rules of standard math. To fix this, researchers invented a clever workaround called recanting-twin effects. Instead of asking the witness to be two people at once, they create a "twin" version of the witness—a random copy drawn from the population—to help calculate the answer. This new method is mathematically safe and can be calculated from real data, but a big question remained: Does this "twin" version tell the truth about the real, messy world, or is it just a convenient lie?

This paper is a deep dive into that question. The authors, a team of statisticians from France, the Netherlands, and Belgium, set out to see if the "recanting-twin" method gives the same answers as the "natural" method (the ideal, perfect answer we can't quite calculate) when the rules of the game get complicated. They didn't just guess; they built a massive digital playground with 320 different scenarios to test how these two methods compare.

Here is what they found. First, they discovered that if the world is simple and linear (like a straight line), the twin and the natural methods agree perfectly. They are identical twins in that case. But life is rarely a straight line. When the relationships are curved, messy, or change direction for different people (non-monotonic), the twins start to drift apart.

The researchers ran their simulations and found that in these messy, non-linear worlds, the recanting-twin method often disagrees with the natural truth. Specifically, when they looked at the ranking of the paths (which path is the most important?), the twins got the order wrong between 15% and 56% of the time. When they looked at the sign (is the effect positive or negative?), the twins disagreed with the truth between 5% and 43% of the time.

The paper suggests that the main culprit for these mix-ups is a hidden, invisible factor: the correlation between what would have happened to the "intermediate confounder" under different conditions. It's like trying to guess how a character would have acted in a different movie, but you can't see the script. Because this correlation is unobservable, we can't always know when the twin method is leading us astray.

However, the paper doesn't say the twin method is useless. The authors derived some mathematical "safety nets" (upper and lower bounds) that can tell us if the difference between the twin and the truth is huge or tiny, especially when the outcome is something simple like "yes" or "no." If the difference is small, the twin is a good approximation. But if the difference is large, we have to be very careful.

In short, the recanting-twin method is a brilliant tool that lets us solve problems we couldn't solve before, but it's not a magic wand. In complex, real-world situations where cause and effect twist and turn, the twin might tell a slightly different story than the truth. The authors suggest that while we can use this method to get a general idea, we must be cautious about using it to decide which path is the "most important" or to determine if an effect is positive or negative, unless we are sure the world behaves in a simple, predictable way. The paper leaves us with a clear warning: in the tangled web of cause and effect, the twin is a helpful guide, but it might occasionally point you in the wrong direction.

Technical Summary: On the Magnitude, Sign, and Ranking of Recanting-Twin Path-Specific Effects

Problem Statement
Causal mediation analysis aims to decompose the total effect of an exposure on an outcome into specific causal pathways. However, when intermediate confounders exist (variables affected by the exposure that also confound the mediator-outcome relationship), standard natural path-specific effects (PSEs) become non-identifiable. This issue, known as the "recanting witness" problem, arises because defining certain natural PSEs requires a counterfactual variable to simultaneously take values under different exposure conditions, a contradiction that prevents identification from observed data.

To address this, Díaz (2023) proposed "recanting-twin" PSEs, a class of randomized interventional effects that are identifiable from observed data by replacing the recanting witness with a random draw from its distribution. While these effects satisfy the path-specific sharp null criteria and align with natural PSEs in direction under monotonicity, it remains unclear how well they approximate natural PSEs in terms of magnitude, sign, and relative ranking when monotonicity is violated (i.e., in non-linear, non-monotonic settings). Specifically, it is unknown whether the recanting-twin framework reliably preserves the direction of effects or the ranking of pathway importance compared to the unidentifiable natural counterparts.

Methodology
The authors employ a combination of analytical derivation and simulation studies to characterize the behavior of recanting-twin PSEs relative to natural PSEs.

Analytical Derivation:
- Linear Settings: The authors first derive closed-form expressions for both natural and recanting-twin PSEs under linear outcome models. They demonstrate that the two estimands coincide exactly if there is no interaction between the mediator and the intermediate confounder in the outcome model, or under specific strong assumptions regarding the data-generating process (e.g., constant individual-level effects).
- Non-parametric Bounds: For general settings with binary outcomes, the paper derives non-parametric, identifiable upper and lower bounds for the absolute difference between recanting-twin and natural PSEs. These bounds rely on the fact that the total indirect effect through the intermediate confounder is identifiable in both frameworks. The lower bound is derived via the triangle inequality, while the upper bound utilizes the constraints of binary outcomes and structural relationships between potential outcomes.
Simulation Study:
- A comprehensive simulation study was conducted across four scenarios involving binary or continuous mediators ( $Z, M$ ) and outcomes ( $Y$ ).
- The study evaluated 320 numerical settings, varying key parameters: the direct effect of exposure on the intermediate confounder, the effect of the confounder on the mediator, the effect of the confounder on the outcome, and crucially, the correlation ( $\rho$ ) between the counterfactual values of the intermediate confounder ( $Z(0)$ and $Z(1)$ ).
- The authors assessed the frequency of disagreements in sign (whether the effect is positive or negative) and ranking (the relative magnitude ordering of the four causal paths) between the two estimand types.

Key Contributions and Results

Conditions for Equivalence: The paper establishes that natural and recanting-twin PSEs are identical in linear models without mediator-intermediate confounder interactions. In the presence of such interactions, they diverge, with the discrepancy depending on the strength of the interaction, the treatment effect on the confounder, and the correlation between counterfactual confounder values.
Identifiable Bounds: The derivation of non-parametric upper and lower bounds for the absolute difference between the two effect types provides a practical tool for applied researchers. These bounds allow for an assessment of whether the magnitude discrepancy between recanting-twin and natural PSEs is substantial, even without identifying the unobserved counterfactual correlations.
Simulation Findings on Disagreement:
- Ranking: In non-linear, non-monotonic settings, the ranking of path-specific effects frequently disagrees between the two frameworks. Disagreement rates ranged from 15% to 56% across the four scenarios.
- Sign: Disagreement in the sign of the effects occurred in 5% to 43% of settings.
- Drivers of Inconsistency: The frequency of disagreement in both sign and ranking is strongly influenced by the unobservable correlation ( $\rho$ ) between the counterfactual values of the intermediate confounder ( $Z(0)$ and $Z(1)$ ). Specifically, as $\rho$ becomes more negative, ranking discrepancies become more frequent.
- Predictive Value of Intermediate Confounding: The parameter $\psi_{P2 \lor P3}$ , which identifies the presence of intermediate confounding, was found not to be associated with the extent of ranking or sign disagreement. Thus, detecting intermediate confounding does not predict the degree of inconsistency between the two PSE types.
- Magnitude of Disagreement: While sign disagreements occurred frequently in certain scenarios (e.g., binary $Z$ , continuous $M$ , binary $Y$ ), the absolute difference in magnitude during these disagreements was often small (typically below 0.05).

Significance and Limitations
The paper concludes that while recanting-twin PSEs offer a viable, identifiable alternative to natural PSEs in the presence of intermediate confounding, they should be interpreted with caution in non-monotonic settings. The findings suggest that relying solely on recanting-twin PSEs to determine the direction or relative importance of causal pathways may lead to incorrect conclusions, particularly when the correlation between counterfactual confounder values is negative.

The authors emphasize that because the parameters driving these inconsistencies (specifically the correlation between counterfactuals) are generally unidentifiable in practice, researchers cannot easily determine when the recanting-twin approximation is sufficient. The paper calls for future work to identify specific conditions under which the two frameworks yield consistent conclusions.

Limitations acknowledged by the authors include:

The simulation parameters were not tied to specific clinical or policy contexts, limiting the assessment of substantive impacts.
The proposed non-parametric bounds are not smooth or pathwise differentiable, making the establishment of asymptotic properties for standard non-parametric estimation challenging, though bootstrap procedures remain viable for parametric approaches.

Technical Summary: On the Magnitude, Sign, and Ranking of Recanting-Twin Path-Specific Effects

More like this