Four-Way Decomposition
- Introduced by
- Tyler J. VanderWeele
- Year
- 2014
- Type
- Counterfactual decomposition pipeline for total effects
- Family
- Causal inference / mediation and interaction
What does it do?
The four-way decomposition, introduced by Tyler VanderWeele in 2014, unifies the two great themes of effect analysis — mediation and interaction — into a single, exhaustive partition of a total causal effect.
Any total effect of an exposure on an outcome can be split into exactly four pieces: a controlled direct effect (neither mediation nor interaction), a reference interaction (interaction but no mediation), a mediated interaction (both mediation and interaction at once), and a pure indirect effect (mediation but no interaction). These four components are mutually exclusive and add up to the total effect, and they nest the familiar two-way and three-way decompositions as special cases. Formalized in counterfactual notation and developed at book length in VanderWeele's 2015 Explanation in Causal Inference, the method gives social epidemiologists a precise vocabulary for asking how much of an exposure's effect runs through a mediator, how much depends on the exposure and mediator acting together, and how much is direct.
The intuition behind the method
Suppose smoking raises disease risk and you wonder how. Part of the effect might be direct — smoking harms you even if it never changed the intermediate factor you're studying, say a genetic-expression marker. Part might work because smoking changes the marker, which then changes risk — pure mediation. But there is a subtler possibility: maybe the marker only matters for people who smoke, so the effect depends on smoking and the marker being present together. That joint action is interaction, and it can come in two flavors: interaction that operates whether or not smoking actually moved the marker (reference interaction), and interaction that requires smoking to have moved the marker (mediated interaction, which is genuinely both at once). Older approaches forced you to choose between a 'mediation' story and an 'interaction' story. The four-way decomposition says you don't have to choose: it carves the total effect into these four cleanly defined, additive parts so you can see exactly how much each mechanism contributes.
History
Mediation analysis and effect-modification (interaction) analysis developed largely as separate traditions in epidemiology and the social sciences — one asking how an effect is transmitted, the other asking for whom or under what conditions it operates. Tyler VanderWeele's 2014 Epidemiology paper 'A unification of mediation and interaction: a four-way decomposition' brought them together, showing that any total effect partitions into a controlled direct effect, a reference interaction, a mediated interaction, and a pure indirect effect. The result generalized earlier two-way and three-way decompositions and clarified the counterfactual meaning of each component. VanderWeele's 2015 Oxford monograph Explanation in Causal Inference developed the framework comprehensively alongside the broader theory of mediation and interaction, and the approach has since become a standard mechanistic tool, supported by software such as the SAS and R mediation macros and the CMAverse and mediation packages.
When to use it
Use the four-way decomposition when you have an exposure, a single mediator, and an outcome, and you want to understand the mechanism of an effect more finely than a simple direct-versus-indirect split allows — in particular when you suspect the exposure and mediator may interact. It is the right tool when distinguishing 'effect through the mediator' from 'effect that requires the exposure and mediator together' is scientifically or policy-relevant, for example to judge whether intervening on the mediator would help and for whom. It requires the standard mediation identification assumptions: no unmeasured confounding of the exposure-outcome, mediator-outcome, and exposure-mediator relationships, and no mediator-outcome confounder affected by exposure. Prefer simpler two-way natural-effect decomposition when interaction is implausible or you lack power to estimate it. Avoid it when there are multiple causally ordered mediators or strong exposure-induced mediator-outcome confounding, which call for interventional or path-specific extensions.
Limitations and considerations
- Relies on strong, partly untestable no-unmeasured-confounding assumptions for exposure-outcome, mediator-outcome, and exposure-mediator relations.
- Invalid when a mediator-outcome confounder is itself affected by the exposure, which requires interventional or path-specific methods.
- Designed for a single mediator; multiple causally ordered mediators need extensions and the simple four-way split no longer applies.
- Interpretation of the controlled direct effect requires a chosen reference level for the mediator, and results can depend on that choice.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Unifies mediation and interaction in one exhaustive, additive decomposition rather than treating them as competing analyses.
- Yields four mutually exclusive components that sum exactly to the total effect, giving a complete mechanistic accounting.
- Nests the standard two-way (direct/indirect) and three-way decompositions as special cases, with a clear interpretation of each piece.
- Estimable from familiar regression models with an interaction term, and supported by standard mediation software.
Applications
- Disentangling how much of a social or behavioral exposure's effect on health runs through a specific biological or behavioral mediator.
- Identifying mediated interaction to determine whether an intervention on the mediator would benefit primarily the exposed.
- Quantifying gene-environment mechanisms where a genetic factor and an environmental exposure act jointly.
- Refining policy analysis by separating the controlled direct effect (remaining if the mediator were fixed) from mediated pathways.
- Clarifying mechanistic claims in life-course epidemiology where mediators and effect modifiers coincide.
How it works
How it works
The decomposition is built on potential outcomes. The analyst defines the outcome a person would have under a given exposure level and a given mediator value, and the mediator value a person would have under a given exposure level. These counterfactuals make the causal questions precise: comparing outcomes while holding the mediator fixed isolates direct effects, while letting the mediator respond to exposure brings mediation into play. Setting up this notation is what allows mediation and interaction — usually treated as separate ideas — to be expressed in one common language.
Formula
Y_{am} = \text{outcome under exposure } a \text{ and mediator } m;\quad M_a = \text{mediator under exposure } aThe central result writes the total effect (the contrast of outcomes under exposure versus no exposure, with the mediator taking its natural value each time) as the sum of four additive terms. Algebraically rearranging the counterfactual contrast yields the controlled direct effect, the portion attributable to interaction, and the portion attributable to pure mediation, with the interaction portion itself splitting into a reference and a mediated part. Because the four terms are constructed to be mutually exclusive and to sum exactly to the total effect, the decomposition is both exhaustive and non-overlapping — every bit of the effect is assigned to exactly one mechanism.
Formula
Y_1 - Y_0 = \underbrace{(Y_{1m^*}-Y_{0m^*})}_{\text{CDE}} + \underbrace{(Y_{1M_0}-Y_{1m^*})-(Y_{0M_0}-Y_{0m^*})}_{\text{ref. + med. interaction}} + \underbrace{(Y_{1M_1}-Y_{1M_0})}_{\text{pure indirect}}The controlled direct effect measures how much the outcome would change with exposure if the mediator were fixed, by intervention, at a reference level for everyone. It captures the part of the effect that does not operate through the mediator and does not require any interaction — the 'pure' direct channel. This is the component most directly relevant to interventions that could block the mediating pathway, since it tells you how much harm would remain even if the mediator were held at its reference value.
Formula
\text{CDE}(m^*) = E[Y_{1m^*} - Y_{0m^*} \mid c]The two interaction terms quantify the effect attributable to the exposure and mediator working jointly. The reference interaction is the interactive effect that would operate even if the exposure did not change the mediator — it depends on the mediator's natural reference distribution. The mediated interaction is the part that requires both that the exposure changed the mediator and that the exposure and mediator interact, so it is simultaneously mediation and interaction. Separating these makes explicit that interaction can contribute to an effect with or without the mediator actually being moved by exposure.
Formula
\text{INT}_{\text{ref}} = (\theta_3)(E[M_0\mid a^*,c]-m^*);\ \ \text{INT}_{\text{med}} = (\theta_3)(E[M_1\mid a,c]-E[M_0\mid a^*,c])In practice the four components are estimated from two regression models: an outcome model including the exposure, the mediator, their product term, and confounders, and a mediator model regressing the mediator on the exposure and confounders. VanderWeele provides closed-form expressions that combine the outcome-model coefficients (including the crucial exposure-mediator interaction coefficient) with the mediator model to produce each of the four components, with confidence intervals from the delta method or bootstrap. The same logic extends to nonlinear models and is implemented in standard mediation software.
Formula
E[Y\mid a,m,c]=\theta_0+\theta_1 a+\theta_2 m+\theta_3 am+\theta_4' c;\quad E[M\mid a,c]=\beta_0+\beta_1 a+\beta_2' c
Common pitfalls
- Omitting the exposure-mediator interaction term from the outcome model, which collapses the decomposition and hides the interaction components.
- Ignoring exposure-induced mediator-outcome confounding, which biases all four components.
- Interpreting a large mediated proportion causally without addressing unmeasured mediator-outcome confounding.
- Reporting only a single direct/indirect split when interaction is present, thereby misattributing joint effects to pure mediation.
Frequently asked questions
What are the four components and how do they relate to mediation and interaction?
The four are: the controlled direct effect (neither mediation nor interaction — the effect with the mediator fixed at a reference), the reference interaction (interaction but no mediation — the joint effect that operates even without the exposure changing the mediator), the mediated interaction (both mediation and interaction — requires the exposure to change the mediator and the two to interact), and the pure indirect effect (mediation but no interaction — the effect transmitted through the mediator absent any interaction). They are mutually exclusive and sum to the total effect, so together they account for every part of the exposure's impact through these two lenses.
How does the four-way decomposition relate to natural direct and indirect effects?
It refines them. The familiar two-way split into natural direct and natural indirect effects can be recovered by grouping the four components: the natural direct effect equals the controlled direct effect plus the reference interaction, and the natural indirect effect equals the pure indirect effect plus the mediated interaction. The four-way version goes further by revealing how much of each natural effect is due to interaction versus 'pure' mechanism. This is why VanderWeele describes it as a unification — the standard mediation quantities are sums of its building blocks.
What assumptions are needed to interpret the components causally?
Beyond consistency and positivity, the decomposition requires four no-unmeasured-confounding conditions: no unmeasured confounding of the exposure-outcome relationship, of the mediator-outcome relationship, of the exposure-mediator relationship, and — critically — no mediator-outcome confounder that is itself affected by the exposure. The last condition is the one most often violated in observational social-epidemiologic data; when it fails, the natural-effect components are not identified and interventional or path-specific decompositions are needed instead. Sensitivity analysis, such as E-values for the mediation assumptions, is strongly recommended.
Related methods
Sources
VanderWeele, T. J. (2014). A unification of mediation and interaction: a four-way decomposition. Epidemiology, 25(5), 749-761.
Open DOI referenceCopied.VanderWeele, T. J. (2015). Explanation in Causal Inference: Methods for Mediation and Interaction. New York: Oxford University Press.
ISBN: 9780199325870Copied.