Quantifying Inconsistent Mediation: The Mediated Magnitude Share and the Degree of Suppression
summary
The gist
The paper introduces two complementary effect-scale measures, the Mediated Magnitude Share (MMS) and the degree of suppression…
In short
The paper introduces two measures, Mediated Magnitude Share (MMS) and Degree of Suppression, to better describe mediation when direct and indirect effects oppose each other. MMS quantifies the relative indirect effect magnitude compared to the sum of direct and indirect effects, while suppression measures how much total effect is lost due to opposing directions.
Key concepts
- Inconsistent Mediation
- This occurs when the direct effect (DE) and indirect effect (IE) have opposite signs. This leads to a total effect (|TE|) that is smaller than the sum of their individual magnitudes (|TE| < |DE| + |IE|), making standard proportion mediated measures unreliable.
- Mediated Magnitude Share (MMS)
- MMS calculates the relative size of the indirect pathway compared to the combined absolute magnitudes of both direct and indirect pathways. It is defined as MMS = |IE| / (|DE| + |IE|) and is useful because it remains bounded between 0 and 1, unlike proportion mediated (PM) in inconsistent cases.
- Degree of Suppression
- This measure quantifies the fraction of the combined magnitude that is cancelled out by opposing pathway directions. It is calculated as σ = 1 − |TE| / (|DE| + |IE|) and shows how much total effect is lost when DE and IE move in opposite directions.
Terminology used across episodes
This episode discusses
- Quantifying Inconsistent Mediation: The Mediated Magnitude Share and the Degree of Suppression · Paper Radio
The paper
Quantifying Inconsistent Mediation: The Mediated Magnitude Share and the Degree of Suppression · Read on arXiv
University of Texas at El Paso · International Centre for Diarrhoeal Disease Research, Bangladesh
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: "Quantifying Inconsistent Mediation".
Jane: Here's the gist of the paper. The paper introduces two complementary effect-scale measures, the Mediated Magnitude Share (MMS) and the degree of suppression, to provide bounded and interpretable characterizations of mediation systems where pathway directions are not uniform The authors establish these measures to address the limitations of the proportion mediated (PM) when direct and indirect effects operate in opposite directions Defining Inconsistent Mediation and its Problem Inconsistent mediation occurs when the direct effect (DE) and indirect effect (IE) have opposite signs, leading to a total effect (TE) that is smaller than the sum of their individual magnitudes (TE < DE + IE) This situation causes the proportion mediated (PM = IE/TE) to fall outside the
0, 1: range or become unstable as TE approaches zero For example, in the ATLAS program, PM ≈ −0.30 because IE = 0.042 and DE = −0.181 The authors argue that a measure of mediated contribution should use a reference denominator that remains well defined under both consistent and inconsistent mediation while ensuring a value bounded within the unit interval Mediated Magnitude Share (MMS) The Mediated Magnitude Share (MMS) quantifies the relative magnitude of the indirect pathway to the combined absolute magnitudes of the direct and indirect pathways Formally, MMS is defined as MMS = IE / (DE + IE) = m/L This measure is bounded between 0 and 1, with MMS = 1 if and only if DE = 0 Under consistent mediation, MMS reduces exactly to PM because TE has the same sign as DE and IE Degree of Suppression The degree of suppression quantifies the fraction of the combined magnitude that is cancelled by opposing directions, defined as σ = 1 − TE / (DE + IE) = 1 − TE / L Under consistent mediation, σ = 0, while under inconsistent mediation, it quantifies the proportion of the combined magnitude that is lost through opposing directions In the ATLAS example, MMS was 0.188 and the degree of suppression was σ = 0.376 Extension to Multiple Pathways and Binary Outcomes The framework is extended to multiple indirect pathways where TE = DE + ∑k i=1 IEi, allowing for suppression among various subsets of pathways For binary outcomes on the ratio scale, logarithms are used to convert multiplicative effects into additive ones, enabling the direct application of MMS and the degree of suppression to log-ratio effects This transformation ensures that opposing directions on the ratio scale are represented by opposite signs in the log-scale MMS is best viewed as a complement to PM rather than a replacement, coinciding with PM under consistent mediation while remaining bounded under inconsistent mediation The degree of suppression quantifies the specific attenuation that PM cannot convey when pathways oppose each other Together, these two measures provide a simple characterization of both consistent and inconsistent mediation for causal analysis The authors conclude that MMS and the degree of suppression can accompany PM in applied reporting rather than displace it REFERENCES Alwin, D.
Tom: F.
Jane: and Hauser, R.
Tom: M.
Paper summary: Tom: So to recap, this paper introduces two measures, MMS and degree of suppression, to describe mediation when the direct effect and indirect effect have opposite signs. Their central thesis is that standard proportion mediated or PM is unreliable in these cases because it can fall outside its normal range or become unstable as the total effect shrinks.
Jane: They propose MMS, which calculates the absolute magnitude of an indirect pathway relative to the sum of its own and its direct effect's magnitudes, keeping this measure between zero and one. They also have degree of suppression, which measures what fraction of that combined magnitude gets cancelled out because those pathways are going in opposite directions.
Lu: The motivation is clear: when DE and IE oppose each other, the total effect magnitude is actually smaller than the sum of their individual magnitudes, meaning TE equals DE plus IE is less than or equal to DE plus IE. This inequality shows why PM doesn't always work as expected.
Meng: So for practical application in AI systems, this means we can stop relying on a single proportion value and instead use these two complementary measures to get a bounded, interpretable picture of how much effect is being channeled through the indirect route versus how much is going straight through.
Lalam: The authors use an example from the ATLAS program where they show that even with an overall total effect of negative zero point one three nine, inconsistent mediation occurs because the individual direct and indirect effects were pulling in opposite directions.
Tom: That's a concrete example from real data, which is what I like to see. It grounds the theory in something tangible instead of just abstract math. What they claim is that these measures provide a simple characterization of both consistent and inconsistent mediation simultaneously, which is quite useful for causal analysis.
Jane: It’s about providing context for the effect size so that we don't misinterpret a result just because the sign is unexpected. The paper shows how to quantify not just the direction, but also the magnitude interplay when things are fighting each other.
Lu: And they extend this to binary outcomes on a ratio scale by using logarithms, which lets them apply these measures to log-ratio effects, ensuring that opposing directions in the original scale show up as opposite signs in the log scale.
Meng: I worry about the practical deployment of these measures at scale. If we're running millions of inferences per second, calculating these two new metrics instead of one simple proportion might add computational overhead we can't afford.
Conclusion: Tom: So looking at this whole thing, the paper "Quantifying Inconsistent Mediation: The Mediated Magnitude Share and the Degree of Suppression" is essentially saying that when pathways oppose each other, you can’t just use the standard proportion mediated figure because it gets messy.
Jane: They offer MMS and degree of suppression as better tools for that situation. Think of MMS as telling you what slice of the combined effect belongs to the indirect path itself, bounded between zero and one.
Lu: And degree of suppression tells you how much is lost in that fight between the direct and indirect routes, quantifying the cancellation due to opposing directions. It's a way to see attenuation when things aren't aligned.
Meng: For me, what this means for deployment is that if we build a system where we expect some paths to conflict, these measures give us a bounded metric instead of an unstable one that could lead to bad decisions on the scale.
Lalam: Lalam confirms that MMS and suppression work together to give you a full picture of the magnitude interplay, which PM alone can't provide when things are inconsistent.
Tom: So the authors are suggesting we keep PM as a good starting point under consistent conditions but switch to these two measures when inconsistency is suspected to keep our analysis sound. It’s about choosing the right tool for the job instead of trying to force one tool onto every scenario.
Jane: That’s a fair way to put it. The paper gives us concrete, quantifiable ways to handle the complexity that arises when direct and indirect effects operate in opposite directions in causal models, which is a real problem in many fields.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates