Causal mediation
Description
Causal mediation names the mechanism-bearing topology A→M→B. The upstream cause changes a mediator; the mediator in turn changes the outcome. There may also be a direct A→B path, so the observed total effect combines an indirect contribution through M and a direct contribution that bypasses it. The useful question is not merely “what happened in between?” It is: if the mediator were blocked, held fixed, or restored, how would the effect of A on B change? That intervention license distinguishes a causal mediator from a chronological waypoint or explanatory story. It also suggests different actions: intervening on A prevents the whole pathway, intervening on M interrupts one mechanism, and intervening downstream may treat the outcome without changing its cause.Aliases
Mediation and indirect effect are the statistical and causal-inference names. Causal pathway is broader ordinary language. The explicitcausal- prefix prevents collision with the catalog’s mediator design pattern, which coordinates peers rather than transmitting causal influence.
Triggers
- An intervention works, but the mechanism producing its downstream benefit is unclear.
- A treatment changes several intermediate biomarkers and the question is which carries the clinical effect.
- Removing one middle link abolishes a longer causal response.
- A direct association shrinks after an intermediate is controlled, but the causal assumptions behind that adjustment need examination.
- Two interventions reach the same outcome through different pathways and may therefore combine or interfere.
Exclusions
- Sequence without mechanism — “A, then M, then B” is not sufficient evidence that M transmits the effect.
- confounding — a common cause sits upstream of both variables; it is not caused by A.
- collider-bias — the arrows converge on the conditioned variable rather than passing through it.
- Pure direct causation — no intermediate means there is no mediated effect to decompose.
- Unjustified adjustment — a randomized treatment does not make a measured mediator randomized; mediator-outcome confounding remains a live threat.
Structure
Relationships
- confounding — downstream mechanism versus upstream common cause.
- collider-bias — arrows pass through a mediator but converge into a collider.
- root-cause-analysis — causal descent discovers intermediate mechanisms as well as origins.
- cascade — a cascade supplies candidate chains; mediation tests which link carries the effect.
- mediator — both interpose something between endpoints, but one transmits causation and the other coordinates interaction.
Examples
Judea Pearl, “Direct and Indirect Effects,” Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, 2001, 411–420; arXiv:1301.2300 · computer-science
Judea Pearl, “Direct and Indirect Effects,” Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, 2001, 411–420; arXiv:1301.2300 · computer-science
Pearl formalized direct and indirect causal effects using counterfactual interventions rather than treating regression adjustment as sufficient. In the paper’s hiring-discrimination example, gender can affect a hiring decision directly and indirectly through qualifications. Holding qualifications at an arbitrary fixed value asks a different question from changing gender while preserving the qualifications that would naturally obtain for the person under comparison. The distinction makes the mediator’s causal position—not merely its statistical association—explicit.Inference: Before “controlling for” an intermediate variable, state which path-specific question the adjustment is meant to answer. Blocking a mediator can erase part of the very effect under investigation.
James J. Heckman, Rodrigo Pinto, and Peter Savelyev, “Understanding the Mechanisms through Which an Influential Early Childhood Program Boosted Adult Outcomes,” American Economic Review 103(6), 2013, 2052–2086, doi:10.1257/aer.103.6.2052 · economics
James J. Heckman, Rodrigo Pinto, and Peter Savelyev, “Understanding the Mechanisms through Which an Influential Early Childhood Program Boosted Adult Outcomes,” American Economic Review 103(6), 2013, 2052–2086, doi:10.1257/aer.103.6.2052 · economics
Heckman, Pinto, and Savelyev examined how the Perry Preschool intervention produced adult outcomes long after its early IQ effects faded. Their analysis identifies changes in externalizing behavior and academic motivation as important pathways connecting the randomized program to later outcomes, with different contributions across outcomes and sexes. The treatment was randomized, but the mediating capacities were not, so path-specific causal conclusions depend on additional modeling and no-unmeasured-confounding assumptions.Inference: A durable outcome can be mediated by a variable different from the intervention’s most visible short-term effect. When the headline metric fades, look for intermediate capacities the intervention changed persistently.
Mounica Vallurupalli et al., “Effects of Interleukin-1β Inhibition on Incident Anemia: Exploratory Analyses From a Randomized Trial,” Annals of Internal Medicine 172(8), 2020, 523–532, doi:10.7326/M19-2945 · medicine-and-health
Mounica Vallurupalli et al., “Effects of Interleukin-1β Inhibition on Incident Anemia: Exploratory Analyses From a Randomized Trial,” Annals of Internal Medicine 172(8), 2020, 523–532, doi:10.7326/M19-2945 · medicine-and-health
In an exploratory analysis of the randomized CANTOS trial, canakinumab treatment reduced incident anemia. Early reduction in high-sensitivity C-reactive protein, a marker of inflammatory response, accounted for a substantial estimated portion of the treatment effect in the authors’ causal-mediation analysis. Treatment assignment was randomized and the mediator preceded later anemia outcomes, but the inflammatory response itself was not randomized; the mediator-to-outcome claim therefore still depends on assumptions about unmeasured confounding.Inference: Randomizing the upstream treatment strengthens the total-effect claim without automatically proving the intermediate mechanism. A mediation result should carry its extra identification assumptions alongside its pathway estimate.
Jiaxi Wu et al., “Cyclic GMP-AMP Is an Endogenous Second Messenger in Innate Immune Signaling by Cytosolic DNA,” Science 339(6121), 2013, 826–830, doi:10.1126/science.1229963 · biology
Jiaxi Wu et al., “Cyclic GMP-AMP Is an Endogenous Second Messenger in Innate Immune Signaling by Cytosolic DNA,” Science 339(6121), 2013, 826–830, doi:10.1126/science.1229963 · biology
Wu and colleagues identified cyclic GMP-AMP (cGAMP) as the intermediate messenger connecting cytosolic DNA detection to STING-dependent innate immune signaling. Cytosolic DNA induced production of cGAMP; introducing cGAMP was sufficient to activate downstream IRF3 and interferon signaling without the original DNA trigger, while disruption of the receiving STING pathway blocked the response. The experiments therefore did more than order events: they intervened on the intermediate and the machinery through which it acts.Inference: A persuasive mechanism often combines necessity and sufficiency tests at the mediator: block the intermediate pathway to stop transmission, then restore or introduce the intermediate to recover the downstream effect.