8  Mediation and Front-Door Identification

NoteLearning Objectives

By the end of this chapter, students should be able to:

  1. Explain the distinction between the total causal effect of \(T\) on \(Y\) and a pathway-specific effect that operates through a mediator \(M\), and describe why this distinction matters scientifically.
  2. Draw the prototype mediation DAG, write its structural equations, and identify the direct and indirect pathways.
  3. Define the controlled direct effect (CDE) using the do-operator, identify it via the back-door formula for the joint intervention \((T, M)\), and explain why it depends on the fixed level \(m\).
  4. Define the natural direct and indirect effects (NDE, NIE) using the potential outcomes notation \(Y(t, M(t'))\), state the NDE \(+\) NIE \(=\) TE decomposition, and explain why these quantities involve cross-world counterfactuals.
  5. State the sequential ignorability assumptions for identification of natural effects, write the mediation formula, and explain why a treatment-induced mediator–outcome confounder, whether measured or not, generally invalidates the cross-world independence.
  6. Set up the Baron–Kenny three-equation system, derive the product and difference formulas for the indirect effect, and explain why the decomposition fails in nonlinear or interaction models.
  7. State the three front-door conditions, derive the front-door formula using the do-calculus, and explain why the front-door graph enables identification despite unobserved \(T\)\(Y\) confounding.
  8. Contrast mediation analysis and instrumental variables on the dimensions of variable position, identification goal, and key assumption.
NoteHow to Read This Chapter

The chapter is long, and not all of it carries equal weight on a first pass.

Core first reading. Section 8.1Section 8.5 (motivation, setup, TE, CDE, natural effects); Section 8.6 through the mediation formula (Theorem 8.2) and its proof; Section 8.8 through the failure examples; and the summary in Section 8.10.

Second pass. The plug-in estimation recipe in Section 8.6; the Baron–Kenny inference details (Sobel test and bootstrap) in Section 8.7; the front-door numerical example; and the mediation–IV comparison in Section 8.9.

Optional (\(\star\)). The starred remarks on alternative causal frameworks (NPSEM-IE vs. FFRCISTG, separable and interventional effects) and on principal stratification, and the problems marked advanced or second pass.

8.1 Motivation: Mechanisms

The identification results of Chapters 5–7 all answer the same question: what is the total causal effect of \(T\) on \(Y\)? Mediation analysis asks a finer question: through what mechanism does that effect operate?

More concretely, the total effect of \(T\) on \(Y\) may flow along multiple causal pathways. Some of this effect passes through an intermediate variable \(M\) — the mediator — along the path \(T \to M \to Y\). The remainder flows directly along \(T \to Y\), bypassing the mediator entirely. Mediation analysis aims to study mechanisms by defining direct and indirect effect concepts that target each pathway; as we will see, only some of these concepts yield an additive decomposition of the total effect.

This mechanism question matters for scientific and policy reasons. In a clinical trial of a behavioral intervention (\(T\)) on depression (\(Y\)), a researcher may want to know how much of the benefit operates through improved sleep quality (\(M\)) versus other pathways — because if sleep is the main channel, targeting sleep directly may be a more efficient intervention. In an economics study of education (\(T\)) on wages (\(Y\)), how much operates through occupation (\(M\)) versus cognitive skills? The answer determines whether a policy should target educational attainment or occupational access.

To keep the abstract formulas of this chapter tethered to a single concrete scenario, we adopt the depression intervention as a running example: \(T\) is a randomized behavioral intervention, \(M\) is self-reported sleep quality measured mid-trial, and \(Y\) is a depression score (e.g. on the PHQ-9 scale). We return to this example at each key estimand — TE, CDE, NDE/NIE, the Baron–Kenny decomposition — and note in Section 8.8 why the front-door criterion requires a different scenario.

The challenge is that mediators are post-treatment variables: they are affected by the treatment, and may themselves be confounded with the outcome. Conditioning on a post-treatment variable creates exactly the collider and selection-bias problems studied in Chapters 2 and 3. A naive approach — simply including \(M\) as a covariate in a regression of \(Y\) on \(T\) — conflates adjustment with mediation and can introduce bias even in a randomized experiment.

This chapter also develops the front-door criterion, a distinct identification strategy that uses the mediation structure of the DAG to identify causal effects even when treatment and outcome are confounded by an unobserved variable. This makes mediation analysis relevant not only to mechanism research but also to the core identification problem of earlier chapters.

8.2 The Mediation DAG

8.2.1 The Prototype Graph

Throughout this chapter we work with the following prototype graph.

T M Y X U treatment mediator outcome covariates unobserved
The prototype mediation DAG. The causal effect of $T$ on $Y$ operates through two pathways: the direct path $T \to Y$ and the indirect path $T \to M \to Y$. The unobserved variable $U$ confounds the treatment--outcome relationship; $\mathbf{X}$ denotes observed pre-treatment covariates. This graph is the conceptual prototype; the identification sections work with the reduced graph below.

The graph encodes two causal pathways: the direct pathway \(T \to Y\), in which the treatment affects the outcome without passing through the mediator; and the indirect pathway \(T \to M \to Y\), in which the treatment first shifts the mediator, which in turn shifts the outcome.

8.2.2 Structural Equations

The prototype graph corresponds to the following nonparametric structural equation model: \[T = f_T(\mathbf{X},\, U,\, \varepsilon_T), \tag{8.1}\] \[M = f_M(T,\, \mathbf{X},\, \varepsilon_M), \tag{8.2}\] \[Y = f_Y(T,\, M,\, \mathbf{X},\, U,\, \varepsilon_Y), \tag{8.3}\] where each arrow corresponds to the presence of the parent in the child’s structural equation. In particular, \(U\) enters both \(T\)’s and \(Y\)’s equations, making the confounding paths explicit; \(U\) is absent from \(M\)’s equation, reflecting the absence of an arrow \(U \to M\).

Disturbance structure. Writing a recursive system compatible with the graph is not yet a complete causal model: one must also say how the disturbances are jointly distributed. Unless stated otherwise, we adopt the NPSEM-IE reading of Equation 8.1Equation 8.3: all shared causes represented in the model appear explicitly as variables (here \(\mathbf{X}\) and \(U\)), and the equation-specific disturbances \(\varepsilon_T\), \(\varepsilon_M\), \(\varepsilon_Y\) are mutually independent given those explicitly represented background variables. This independent-errors assumption is innocuous for the total-effect results of Section 8.3, but it becomes load-bearing when cross-world counterfactuals appear in Section 8.5Section 8.6.

8.2.3 What Makes Mediation Harder Than Total Effect Estimation

The key difficulty is that \(M\) is a post-treatment variable. This creates two interrelated problems.

Collider bias. Conditioning on \(M\) can open collider paths. Suppose \(U \to T\) and \(V \to M\) and \(V \to Y\), with \(V\) unobserved. The path \(T \to M \leftarrow V \to Y\) is blocked when \(M\) is not conditioned on, but opens as soon as \(M\) is included as a covariate, inducing spurious \(T\)\(Y\) association through \(V\). This is precisely why naively regressing \(Y\) on \((T, M)\) does not isolate the direct effect.

Mediator–outcome confounding. Even when treatment is randomized — eliminating confounding on the \(T\) side — the mediator \(M\) is never randomized. An unobserved variable \(V\) with \(V \to M\) and \(V \to Y\) creates a back-door path from \(M\) to \(Y\) that randomization of \(T\) does not close. This is the central challenge of mediation analysis.

WarningPost-Treatment Variables Are Dangerous to Condition On

The mediator \(M\) is caused by the treatment \(T\). This single fact makes every operation involving \(M\) structurally different from operations involving pre-treatment covariates \(\mathbf{X}\).

Conditioning on a post-treatment variable in a regression or matching procedure is not a neutral act. It can open collider paths that were previously blocked, introduce selection bias, and produce estimates of neither the total effect nor any well-defined direct effect. The appropriate response is not to avoid \(M\) entirely — mediation analysis requires reasoning about \(M\) — but to be precise about which estimand one is targeting and which structural assumptions justify the operation being performed.

This chapter studies three regimes in which conditioning on or marginalizing over \(M\) is justified:

  1. CDE (Section 8.4): both \(T\) and \(M\) are intervened on simultaneously via \(\doop(T, M)\), so no conditioning on an observed \(M\) occurs.
  2. NDE/NIE (Section 8.5Section 8.6): \(M\) is marginalized over using a distribution from a different treatment arm, justified by sequential ignorability.
  3. Front-door (Section 8.8): \(M\) is summed over in the front-door formula, justified by the graphical conditions of the front-door criterion.

Outside these three regimes, conditioning on a post-treatment variable should be treated as an error until proven otherwise.

8.2.4 A Working Graph for the Identification Sections

The prototype graph retains the unobserved confounder \(U\) on purpose: the chapter ultimately demonstrates, in Section 8.8, that the causal effect of \(T\) on \(Y\) can be identified even when \(U\) is hidden. In the intervening sections, however, we rely on back-door adjustment formulas that require every \(T\)\(Y\) and \(T\)\(M\) confounder to be captured by the observed covariate set \(\mathbf{X}\).

NoteWorking Assumption (Section 8.3Section 8.7)

Throughout Section 8.3Section 8.7, we work with the reduced prototype graph shown below. Equivalently, the unobserved confounder \(U\) is assumed either (i) absent or (ii) observed and included in \(\mathbf{X}\), so that \(\mathbf{X}\) blocks every \(T\)\(Y\) and \(T\)\(M\) back-door path. The front-door section Section 8.8 restores the full prototype with \(U\) unobserved, and derives identification of the total effect without access to \(U\).

T M Y X treatment mediator outcome covariates
The reduced prototype graph, obtained by absorbing all $T$--$Y$ and $T$--$M$ confounding into the observed $\mathbf{X}$. This graph is in force throughout @sec-w8-te--@sec-w8-linear.
NoteRemark: Notation for Unobserved Confounders

Three latent variables recur in this chapter, each playing a distinct structural role: \(U\) is a pre-treatment \(T\)\(Y\) confounder (\(U \to T\), \(U \to Y\)); \(V\) is a pre-treatment \(M\)\(Y\) confounder (\(V \to M\), \(V \to Y\)); and \(L\) is a treatment-induced \(M\)\(Y\) confounder (\(T \to L\), \(L \to M\), \(L \to Y\)). Keeping the three roles separate is essential: they break different identification arguments.

NoteThe Chapter’s Estimands at a Glance
Estimand Intervention Decomposes TE? Main identification burden
TE \(\doop(T{=}t)\) (reference) \(T\)\(Y\) confounding (Chs. 4–6)
\(\mathrm{CDE}(m)\) \(\doop(T{=}t)\), \(\doop(M{=}m)\) No Joint-intervention adjustment (Section 8.4)
NDE, NIE Nested, cross-world Yes Sequential ignorability, incl. a cross-world independence (Section 8.5Section 8.6)
TE via front-door \(\doop(T{=}t)\) Not a decomposition Front-door graphical conditions (Section 8.8)

This matrix is the chapter’s road map. A fuller version, the Framework Map, returns at the end of Section 8.7 as a synthesis.

8.3 Total Causal Effect

Before introducing direct and indirect effect concepts, we fix the reference quantity: the total causal effect of \(T\) on \(Y\).

NoteDefinition: Total Effect

The total effect (TE) of \(T\) on \(Y\) is \[\mathrm{TE} = \E\!\bigl[Y(1) - Y(0)\bigr] = \E\!\bigl[Y \mid \doop(T{=}1)\bigr] - \E\!\bigl[Y \mid \doop(T{=}0)\bigr].\]

The total effect captures the combined impact of all causal pathways from \(T\) to \(Y\). It is identified by the back-door formula whenever \(\mathbf{X}\) satisfies the back-door criterion: \[\mathrm{TE} = \sum_{\mathbf{x}} \bigl[\E[Y \mid T{=}1, \mathbf{X}{=}\mathbf{x}] - \E[Y \mid T{=}0, \mathbf{X}{=}\mathbf{x}]\bigr] P(\mathbf{X}{=}\mathbf{x}).\] As throughout the book, the formula additionally presumes the intervention semantics and consistency established in Chapters 3–5, together with treatment positivity, so that both conditional means are well-defined wherever the outer average visits.

Running example. For the depression trial, the TE is the expected change in depression score if the entire study population were assigned to the intervention versus control. It combines the effect operating through sleep improvement with every other pathway by which the intervention acts on depression — cognitive restructuring, behavioral activation, therapeutic alliance, and so on. Mediation analysis asks: of this total, how much is due to sleep?

The goal of mediation analysis is to define direct and indirect effect estimands that target the separate pathways within this total. The next two sections introduce two such families, each formalizing a different notion of what “direct” and “indirect” mean; of the two, only the second — the natural direct and indirect effects — is constructed to yield an additive decomposition of the total effect.

8.4 Controlled Direct Effect

8.4.1 Definition

The do-operator provides a direct definition of the controlled direct effect: intervene on both \(T\) and \(M\) simultaneously, fixing \(M\) at a specified level \(m\). Fixing \(M = m\) shuts down causal pathways that pass through \(M\); any remaining effect of \(T\) on \(Y\) operates through pathways that bypass \(M\) — in the prototype graph, precisely the direct edge \(T \to Y\).

NoteDefinition: Controlled Direct Effect (CDE)

The controlled direct effect of changing \(T\) from 0 to 1 while fixing \(M = m\) by intervention is \[\mathrm{CDE}(m) = \E\!\left[Y \mid \doop(T{=}1),\, \doop(M{=}m)\right] - \E\!\left[Y \mid \doop(T{=}0),\, \doop(M{=}m)\right]. \tag{8.4}\]

Because Equation 8.4 involves two simultaneous interventions, it corresponds to the mutilated graph \(\Gcal_{\overline{T}\,\overline{M}}\) in which all edges into both \(T\) and \(M\) are deleted.

Running example. \(\mathrm{CDE}(m)\) at \(m =\) “poor sleep” is the expected change in depression score comparing intervention to control if every participant’s sleep quality were externally held at the poor-sleep level, regardless of assignment. Any reduction in depression that survives this manipulation must come from non-sleep mechanisms. Whether sleep can be externally fixed in a clinical trial is a separate question — sleep is notoriously difficult to control — which is why the policy interpretation of the CDE in this scenario is strained even when the estimand is statistically identified.

NoteRemark: The CDE Depends on \(m\)

The CDE depends on the level \(m\) at which \(M\) is fixed. In general, the direct effect may vary across values of \(m\) — this is effect modification by the mediator. In a linear additive model, \(\mathrm{CDE}(m) = \tau'\) for all \(m\), which is a special property of linearity. When \(T\) and \(M\) interact in their effect on \(Y\), the CDEs at different values of \(m\) differ, and no single number summarizes the direct effect.

8.4.2 Identification of the CDE

Since the CDE involves a joint intervention \(\doop(T, M)\), its identification in this chapter’s baseline-covariate setting reduces to a back-door-type adjustment applied to the pair \((T, M)\) jointly.

Theorem 8.1 (Baseline Adjustment for the CDE) Suppose \(\mathbf{Z}\) is a set of baseline (pre-treatment) covariates satisfying, for the joint intervention \(\doop(T, M)\) on \(Y\): (i) \(\mathbf{Z}\) contains no descendant of \(T\) or \(M\), and (ii) \(\mathbf{Z}\) d-separates the joint exposure set \(\{T, M\}\) from \(Y\) in \(\Gcal_{\underline{T}\,\underline{M}}\), the graph obtained by deleting all arrows out of \(T\) and out of \(M\). Suppose further that (iii) joint positivity holds at \((t, m)\): \[P(T{=}t,\, M{=}m \mid \mathbf{Z}{=}\mathbf{z}) > 0 \quad \text{for $P$-almost every $\mathbf{z}$ in the support of } \mathbf{Z}.\] Then \[\E[Y \mid \doop(T{=}t),\, \doop(M{=}m)] = \sum_{\mathbf{z}} \E[Y \mid t,\, m,\, \mathbf{z}]\, P(\mathbf{z}). \tag{8.5}\]

Condition (iii) ensures that the conditional expectation is defined at every \(\mathbf{z}\) that receives positive weight under \(P(\mathbf{z})\). Joint positivity is the CDE counterpart of the overlap conditions (P1)–(P2) introduced for the mediation formula in Section 8.6. Even when \(T\) is randomized, (iii) is not automatic: it additionally requires that every level \(m\) of the mediator has positive conditional probability given the treatment arm and \(\mathbf{Z}\). When some \((t, m, \mathbf{z})\) cells are never observed, \(\mathrm{CDE}(m)\) is not empirically identified at that \(m\).

Formula Equation 8.5 is the standard back-door adjustment formula applied to the pair \((T, M)\) as a joint treatment. Under the working assumption of Section 8.2.4, \(\mathbf{Z} = \mathbf{X}\) satisfies condition (ii): deleting the arrows out of \(T\) and \(M\) removes \(T \to M\), \(T \to Y\), and \(M \to Y\), and the only remaining paths from \(\{T, M\}\) to \(Y\), namely \(T \leftarrow \mathbf{X} \to Y\) and \(M \leftarrow \mathbf{X} \to Y\), are blocked by \(\mathbf{X}\).

Graph surgery and back-door adjustment are distinct steps. It is worth pausing on what the theorem does and does not say, because the two steps of identification can be conflated. Graph surgery defines the interventional target by deleting arrows into \(T\) and \(M\); this specifies what we want to compute. Expressing that target as a functional of the observed distribution is a separate step that requires a valid adjustment set \(\mathbf{Z}\) in the original graph. The mutilated graph does not by itself justify the right-hand side of Equation 8.5. Consider what happens if we set aside the working assumption and return to the full prototype with \(U\) unobserved: the back-door path \(T \leftarrow U \to Y\) is present in the original graph and is not blocked by any observed covariate, so \(\mathbf{X}\) alone is not a valid adjustment set. Adjustment identifies the CDE in that setting only when \(\mathbf{Z}\) actually blocks the \(U\)-path. When \(U\) remains unobserved, alternative strategies are required (e.g. the front-door formula of Section 8.8, or an instrumental variable as in Chapter 7).

NoteRemark: Scope of the CDE Theorem — Sufficient, Not Necessary

Theorem 8.1 is a baseline-adjustment result: it treats \((T, M)\) as a single joint exposure and standardizes over a covariate set that precedes both. In the point-treatment mediation graphs of this chapter, condition (ii) can equivalently be checked sequentially: \(\mathbf{Z}\) blocks all back-door paths from \(T\) to \(Y\), and \((T, \mathbf{Z})\) blocks all back-door paths from \(M\) to \(Y\). The theorem’s conditions are sufficient for this setting but stronger than necessary for time-ordered interventions in general. In particular, when a variable \(L\) with \(T \to L\), \(L \to M\), and \(L \to Y\) is observed, \(L\) is a legitimate confounder for the later intervention on \(M\) even though it is a descendant of \(T\); condition (i) excludes it, yet the CDE remains identified by the sequential g-formula, which adjusts for \(L\) at the second stage without treating it as a baseline covariate. Time-ordered joint interventions of this kind require the longitudinal g-formula and the dynamic treatment regime machinery treated later in the book.

8.4.3 Physical Manipulability and the Meaning of the CDE

The CDE answers the question: what is the effect of \(T\) on \(Y\) when the mediator is prevented from changing? This question has a clean causal answer precisely because \(\doop(M{=}m)\) is a genuine intervention: it severs all edges into \(M\), placing \(M\) at \(m\) regardless of what \(T\) does.

NoteRemark: Well-Defined Interventions on the Mediator

The CDE has its clearest interpretation when “set \(M = m\)” corresponds to a sufficiently well-defined intervention — or to a clearly specified class of interventions with equivalent effects on \(Y\). Literal physical control is neither strictly necessary nor sufficient for this: hypothetical or stochastic interventions with clear consistency semantics can define a coherent CDE, while a physically performable manipulation may still admit multiple “versions” of \(M = m\) with different effects on \(Y\), leaving the estimand ambiguous. Settings where the intervention is well-defined include a drug trial where both the assigned treatment and a downstream biomarker can be externally controlled. In many substantive settings, however, no well-defined independent manipulation of the mediator is available:

  • In an education study, one cannot intervene to hold occupation fixed while varying years of schooling.
  • In a behavioral trial, one cannot set sleep quality to a predetermined level while varying the intervention.

When no sufficiently well-defined intervention on the mediator can be articulated, the symbols \(Y(t, m)\) and \(\doop(M{=}m)\) can still be manipulated formally, but the causal estimand itself is ambiguous: different interventions producing the same recorded mediator value may have different effects on \(Y\), so “the” CDE at level \(m\) is not a single well-posed quantity, and consistency has no unique referent. In those cases the natural direct effect (Section 8.5), which asks what would happen if \(M\) were allowed to take its natural value under the reference treatment, avoids specifying an external mediator level; note, however, that it does not resolve the underlying ambiguity, since the nested counterfactual \(Y(t, M(0))\) presupposes the same well-defined mediator input — and it comes at the cost of stronger identification assumptions.

8.4.4 The CDE Does Not Have a Complementary Indirect Effect

A common misconception is that the CDE and some complementary “controlled indirect effect” sum to the total effect, analogous to the NDE \(+\) NIE \(=\) TE decomposition. This is not generally true.

Fixing \(M = m\) by intervention removes treatment-induced variation in the mediator, but the residual quantity \(\mathrm{TE} - \mathrm{CDE}(m)\) is not a controlled indirect effect. It is a well-defined causal contrast — every term below is an interventional mean — but it depends on the selected mediator level \(m\), it does not correspond to any intervention that isolates only the indirect pathway, and it cannot be interpreted as the effect transmitted through \(M\) in a complementary pathway decomposition. Formally: \[\begin{aligned} \mathrm{TE} - \mathrm{CDE}(m) ={}& \Bigl(\E\!\left[Y \mid \doop(T{=}1)\right] - \E\!\left[Y \mid \doop(T{=}1),\,\doop(M{=}m)\right]\Bigr) \\ &- \Bigl(\E\!\left[Y \mid \doop(T{=}0)\right] - \E\!\left[Y \mid \doop(T{=}0),\,\doop(M{=}m)\right]\Bigr), \end{aligned} \tag{8.6}\] which mixes the effect of intervening on \(M\) under \(T=1\) and \(T=0\) in a way that has no simple pathway interpretation. The decomposition TE \(=\) NDE \(+\) NIE holds because the natural effects involve cross-world counterfactuals that are constructed precisely to be complementary; the CDE has no analogous cross-world partner.

The practical implication is that if the research goal is to decompose the total effect into direct and indirect components, the correct estimands are the NDE and NIE, not \(\mathrm{CDE}(m)\) and \(\mathrm{TE} - \mathrm{CDE}(m)\). The CDE is the right estimand when the question is specifically about the effect of \(T\) with \(M\) controlled at a given level — a different and more limited question.

WarningNatural Effects Are a Strictly Harder Estimand Than the CDE

The step from the CDE to the NDE and NIE is not merely a notational upgrade — it is a conceptual escalation with serious identification consequences.

The CDE is a single-world estimand. The quantity \(\E[Y \mid \doop(T{=}t), \doop(M{=}m)]\) involves a single joint intervention. Every individual lives in one hypothetical world where both \(T\) and \(M\) are externally fixed. The do-operator handles this directly, and identification reduces to the standard back-door criterion for the pair \((T, M)\).

The NDE and NIE are cross-world estimands. The nested counterfactual \(Y(t, M(t'))\) requires an individual to simultaneously inhabit two worlds: the world where \(T = t\) (which determines \(Y\)) and the world where \(T = t'\) (which determines what \(M\) would naturally be). When \(t \neq t'\), these are logically distinct states of the world. No single experimental intervention can realize both at once for the same unit. The do-operator alone cannot express this quantity.

The identification price is steep. The CDE requires a back-door-type adjustment condition (plus positivity and consistency) — assumptions about observed confounding that can in principle be satisfied by design. The NDE and NIE require the sequential ignorability assumptions of Section 8.6, including a genuinely cross-world independence (Assumption 3 there), which cannot be satisfied by randomizing \(T\) and cannot be verified from data. Its structural sufficient condition — no treatment-induced mediator–outcome confounder — must be defended on subject-matter grounds.

Practical guidance. When the goal is a pathway-specific effect and the no-treatment-induced-confounder assumption is plausible, NDE and NIE are the right estimands. When that assumption is suspect, the CDE — or a sensitivity analysis around it — is the safer choice.

8.5 Natural Direct and Indirect Effects

8.5.1 Motivation: Cross-World Counterfactuals

The CDE fixes the mediator by external intervention. A more scientifically natural question is: what is the effect of \(T\) on \(Y\) that bypasses \(M\) when \(M\) is held at the value it would naturally take under the reference treatment \(T = 0\)?

Answering this requires a comparison across two intervention worlds: the world where \(T = 1\) but \(M\) is held at the value it would naturally take under \(T = 0\), versus the world where \(T = 0\). The first of these cannot be realized by any single experiment on the same unit — \(T\) would have to be set to 1 (to determine \(Y\)) and to 0 (to fix \(M\) at its natural \(T=0\) value) simultaneously. It requires the nested potential outcomes notation \(Y(t, M(t'))\), which denotes the outcome that would be observed if \(T\) were set to \(t\) and \(M\) were simultaneously set to the value it would naturally take if \(T\) were \(t'\). These are called cross-world counterfactuals.

8.5.2 Definitions

NoteDefinition: Natural Direct and Indirect Effects (Pearl 2001)

For binary \(T \in \{0,1\}\): \[\mathrm{NDE} = \E\!\left[Y(1, M(0)) - Y(0, M(0))\right], \tag{8.7}\] \[\mathrm{NIE} = \E\!\left[Y(1, M(1)) - Y(1, M(0))\right]. \tag{8.8}\] The natural direct effect (NDE) is the expected change in the outcome when \(T\) shifts from 0 to 1, holding the mediator at the value it would naturally take under \(T = 0\). The natural indirect effect (NIE) is the expected change in the outcome due to the shift in the mediator from \(M(0)\) to \(M(1)\), holding \(T = 1\) fixed.

Running example. The NDE is the expected reduction in depression score when the intervention is delivered but each participant’s sleep is held at the value it would have had without the intervention — the part of the benefit that comes from cognitive, behavioral, or therapeutic-alliance pathways, not from sleep. The NIE is the complementary piece. By construction \(\mathrm{NDE} + \mathrm{NIE} = \mathrm{TE}\). One caveat carries over from the CDE discussion: the nested counterfactual \(Y(t, m)\) presupposes that the mediator value is a well-defined causal input. If two different ways of arriving at the same sleep score would have different effects on depression, then “holding sleep at the value it would have had without the intervention” is ambiguous, and the NDE inherits that ambiguity. Throughout, we read \(Y(t, m)\) structurally, as the output of the equation \(f_Y\) in Equation 8.3 evaluated at \((t, m)\).

Decomposition. The total effect decomposes as \[\mathrm{TE} = \mathrm{NDE} + \mathrm{NIE}. \tag{8.9}\]

Proof. \[\begin{aligned} \mathrm{NDE} + \mathrm{NIE} &= \E[Y(1, M(0)) - Y(0, M(0))] + \E[Y(1, M(1)) - Y(1, M(0))] \\ &= \E[Y(1, M(1)) - Y(0, M(0))] = \E[Y(1) - Y(0)] = \mathrm{TE}. \qquad\square \end{aligned}\] The final equality is not pure algebra: it uses composition, \(Y(t, M(t)) = Y(t)\), which holds under the recursive structural interpretation adopted in Section 8.2 — setting \(T = t\) and simultaneously setting \(M\) to the very value it would take under \(T = t\) reproduces the single intervention \(\doop(T{=}t)\).

NoteRemark: Alternative Natural-Effects Decomposition

The decomposition Equation 8.9 is not the only one. The definition above fixes \(M\) at \(M(0)\) in the NDE (the “pure” or “reference” direct effect) and contrasts \(M(1)\) vs. \(M(0)\) while holding \(T = 1\) in the NIE (the “total” indirect effect). The mirror choice fixes \(M\) at \(M(1)\) in the direct effect: \[\mathrm{NDE}^{\star} = \E\!\left[Y(1, M(1)) - Y(0, M(1))\right] \quad \text{("total direct")},\] \[\mathrm{NIE}^{\star} = \E\!\left[Y(0, M(1)) - Y(0, M(0))\right] \quad \text{("pure indirect")}.\] This pair also sums to the total effect. The choice between the two decompositions is conventional, and the two agree when there is no \(T \times M\) interaction in the structural model; when interaction is present, they generally differ, and the gap \(\mathrm{NDE} - \mathrm{NDE}^{\star} = \mathrm{NIE}^{\star} - \mathrm{NIE}\) quantifies the discrepancy. In this textbook we use the Equation 8.9 decomposition throughout.

8.5.3 Interpreting the Cross-World Nature

The NDE involves the counterfactual \(Y(1, M(0))\): the outcome when \(T\) is set to 1 but \(M\) is held at the value it would naturally take under \(T = 0\). This is a cross-world quantity because the two coordinates of the argument refer to different intervention worlds. It cannot be observed for any individual, and in general it cannot be written as a do-expression.

This is not merely a philosophical subtlety. The cross-world nature has direct implications for identification: as shown in the next section, identifying the NDE and NIE requires assumptions that are strictly stronger than those needed for the CDE or the total effect.

NoteRemark: CDE and NDE Coincide Only Without Interaction

In the linear additive Baron–Kenny SEM of Section 8.7 with no \(T \times M\) interaction, the CDE and NDE coincide for all \(m\): \(\mathrm{NDE} = \tau'\) and \(\mathrm{NIE} = ab\). The Baron–Kenny framework therefore implicitly satisfies the cross-world independence condition needed for this equality. When a \(T \times M\) interaction \(\delta\, T M\) is added, the two concepts separate: \(\mathrm{CDE}(m) = \tau' + \delta m\) varies with \(m\), while \(\mathrm{NDE} = \tau' + \delta\, \E[M(0)]\) is a single number. They then agree only at the one level \(m = \E[M(0)]\), and in general answer different scientific questions (Problem 3 works this out).

This remark compares causal frameworks and surveys alternatives to natural effects; it may be skipped on first reading. The main identification argument for the NDE and NIE does not depend on resolving the NPSEM-IE/FFRCISTG distinction.

The status of the cross-world counterfactual \(Y(t', M(t))\) depends on the causal framework adopted, and this has direct consequences for the identification of the NDE and NIE.

Under Pearl’s NPSEM-IE framework (Chapter 1), all relevant counterfactuals are jointly defined as functions of the same exogenous errors. With mutually independent errors and the graphical no-confounding restrictions, the cross-world independence used by the mediation formula (Assumption 3 of Section 8.6) follows. In particular, \(Y(t', M(t))\) is a well-defined random variable for \(t \neq t'\), and the NDE and NIE are identified under sequential ignorability.

Under the Finest Fully Randomized Causally Interpretable Structured Tree Graph (FFRCISTG) framework of Robins (1986) — developed graphically via single world intervention graphs by Richardson and Robins (2014) — the situation is different. An FFRCISTG is a single-world model: its assumptions concern counterfactuals within one intervention world at a time, and are, in principle, the assumptions that randomized experiments can enforce.

The key point is not primarily notational. It is that the single-world independences do not imply the cross-world independence \(Y(t, m) \indep M(t') \mid \mathbf{X}\): two models can agree on every single-world quantity and disagree on the joint law of \((Y(t, m), M(t'))\), and no experiment can distinguish them. Consequently, under FFRCISTG assumptions alone the NDE and NIE are not identified. Identification of natural effects requires additional cross-world structure, such as that supplied by an NPSEM-IE.

Two single-world alternatives deserve mention. First, separable effects (Stensrud et al. 2022, 2023) apply when the treatment can meaningfully be decomposed into distinct components — say \(T_M\), acting on \(Y\) only through \(M\), and \(T_Y\), acting only directly — that could in principle be intervened on separately, together with dismissible-component (isolation) conditions on the graph. Separable effects answer a different scientific question, and they are not available merely by relabeling an ordinary treatment. In the canonical point-treatment setting, under full isolation, their identifying functional coincides with the mediation formula — another instance of the same-formula, different-interpretation principle encountered in the IV context (Chapter 7) — but in richer settings the formulas and assumptions diverge.

Second, interventional (also called randomized) direct and indirect effects replace the individual-level nested counterfactual \(M(t')\) by a random draw from the conditional distribution of \(M(t')\) given \(\mathbf{X}\) (VanderWeele et al. 2014; Vansteelandt and Daniel 2017). Because no unit is required to inhabit two worlds at once, these estimands can be identified under single-world exchangeability assumptions — essentially Assumptions 1–2 of Section 8.6, without the cross-world Assumption 3 — and they remain well-defined even in the presence of a treatment-induced mediator–outcome confounder \(L\). The price is that they decompose an “overall” effect that in general differs from the total effect, and they describe population-level mediator shifts rather than individual-level mechanisms.

For the purposes of this course, we work within the NPSEM-IE framework, where the NDE and NIE are well-defined and identified under sequential ignorability. Students who pursue the single-world literature will find that the estimation machinery of Part III transfers, since the observed-data functionals are the same or closely analogous.

8.6 Identification of Natural Effects

8.6.1 Sequential Ignorability

Identification of the NDE and NIE requires the conditions known collectively as sequential ignorability (Imai et al. 2010), together with consistency, composition, and positivity. We state sequential ignorability as three assumptions of two logically different kinds: Assumptions 1 and 2 below are observed-world statements — conditional independencies of the type met in Chapters 4–6 — while Assumption 3 is a cross-world statement relating counterfactuals from two different intervention worlds.

Despite the neutral-sounding name, these conditions are substantially stronger than the ignorability assumptions used for the total effect or the CDE. Assumptions 2 and 3 govern the mediator–outcome relationship, which is never directly controlled by the investigator; Assumption 3, moreover, is untestable in principle: no experiment, however elaborate, can place one unit in two intervention worlds at once. Students who have internalized the back-door criterion should resist the temptation to read sequential ignorability as a mild extension of it: the assumptions address different sub-problems, and the hardest one has nothing to do with how \(T\) was assigned.

NoteSequential Ignorability Assumptions

For all \(t, t', m\):

  1. Treatment ignorability (joint form). \[\{Y(t', m),\, M(t)\} \;\indep\; T \;\Big|\; \mathbf{X}.\] Graphically: \(\mathbf{X}\) blocks all back-door paths from \(T\) to \(Y\) and from \(T\) to \(M\). Taking the \(M(t)\) marginal gives the treatment–mediator component \(M(t) \indep T \mid \mathbf{X}\), which Step 5 of the mediation-formula proof uses directly; it is a consequence of the joint statement, not a separate assumption.

  2. Mediator–outcome ignorability given treatment. \[Y(t', m) \;\indep\; M \;\Big|\; T, \mathbf{X}.\] Graphically: \((T, \mathbf{X})\) blocks all back-door paths from \(M\) to \(Y\).

  3. Cross-world independence. \[Y(t, m) \;\indep\; M(t') \;\Big|\; \mathbf{X}, \quad \text{including } t \neq t'.\] Unlike Assumptions 1–2, this statement pairs counterfactuals from two different intervention worlds. It has no observed-world analogue and cannot be enforced or tested by any experimental design.

In addition, the observed data are linked to the counterfactuals by consistency (\(Y = Y(t, m)\) on \(\{T{=}t, M{=}m\}\) and \(M = M(t)\) on \(\{T{=}t\}\)) and composition (\(Y(t, M(t)) = Y(t)\)), both automatic under the recursive structural reading of Section 8.2.

NoteWhere Does the Cross-World Assumption Come From?

Assumption 3 does not come for free, but the NPSEM-IE structural model of Section 8.2 supplies a sufficient condition for it. Suppose the working graph and disturbance structure satisfy:

  1. \(\mathbf{X}\) contains the pre-treatment common causes of \(M\) and \(Y\) represented in the model;
  2. there is no treatment-induced mediator–outcome confounder: no variable \(L\) with \(T \to L\), \(L \to M\), and \(L \to Y\);
  3. the equation-specific disturbances are mutually independent given the explicitly represented background variables.

Then, conditional on \(\mathbf{X}\), the counterfactuals \(M(t') = f_M(t', \mathbf{X}, \varepsilon_M)\) and \(Y(t, m) = f_Y(t, m, \mathbf{X}, U, \varepsilon_Y)\) are functions of disjoint disturbances (recall \(U\) does not enter \(f_M\) in the working graph, and any such \(U\) is absorbed into \(\mathbf{X}\)), so independent disturbances deliver \(Y(t, m) \indep M(t') \mid \mathbf{X}\).

If a treatment-induced confounder \(L\) exists, then \(M(t')\) and \(Y(t, m)\) both inherit dependence on \(L\)’s disturbance, and the cross-world independence generically fails — whether or not \(L\) is observed. (As with d-connection throughout the book, a displayed path licenses dependence generically, not necessarily: a structural equation could ignore its \(L\) argument, or effects could cancel. What the presence of \(L\) destroys is the sufficient condition; the independence then holds only by such non-generic coincidence.) Measuring \(L\) does not rescue the mediation formula, because \(L\) is a descendant of \(T\) and cannot be adjusted for as a baseline covariate. In this sense the familiar “no treatment-induced mediator–outcome confounder” condition is the graphical face of Assumption 3, not a freestanding fourth assumption.

NotePositivity (Overlap) Conditions

The sequential ignorability assumptions above are conditional independence statements. A separate requirement — necessary for the mediation formula to be empirically identified — is overlap: every \((t, m, \mathbf{x})\) cell that enters the formula must occur with positive probability in the observed distribution.

(P1) Treatment overlap. \(P(T = t \mid \mathbf{X} = \mathbf{x}) > 0\) for all \(t \in \{0, 1\}\) and for \(P\)-almost every \(\mathbf{x}\).

(P2) Mediator overlap across treatment arms. \(P(M = m \mid T = t, \mathbf{X} = \mathbf{x}) > 0\) whenever \(P(M = m \mid T = t', \mathbf{X} = \mathbf{x}) > 0\), for the pairs \((t, t')\) appearing in the mediation formula and for \(P\)-almost every \(\mathbf{x}\).

Condition (P1) ensures that \(\E[Y \mid T = t, \mathbf{X}]\) and \(P(M \mid T = t, \mathbf{X})\) are defined at every \(\mathbf{x}\) that the outer average visits. Condition (P2) ensures that the outcome regression \(\E[Y \mid T = t, M = m, \mathbf{X} = \mathbf{x}]\) is defined at every \(m\) that receives positive weight under the mediator distribution \(P(M \mid T = t', \mathbf{X} = \mathbf{x})\). Conceptually, (P2) is a support-inclusion requirement, \[\mathrm{supp}\{M \mid T{=}t', \mathbf{X}{=}\mathbf{x}\} \;\subseteq\; \mathrm{supp}\{M \mid T{=}t, \mathbf{X}{=}\mathbf{x}\},\] where \(t\) is the outcome-regression arm and \(t'\) the mediator-distribution arm. For continuous \(M\), positive point probabilities are the wrong currency; the requirement becomes conditional-support inclusion together with density regularity, and Theorem 8.2 below is accordingly stated for discrete \(M\) and \(\mathbf{X}\). Randomization of \(T\) secures (P1) by design, but (P2) is a data-dependent requirement about the mediator’s conditional distribution that no aspect of the experimental assignment guarantees.

Assumptions 1–2 are the natural extensions of the Baron–Kenny conditions of Section 8.7 to the potential outcomes setting: observed-world exchangeability for the treatment and for the mediator. Assumption 3 is the genuinely new requirement, and the second box explains its structural pedigree: it holds under the NPSEM-IE when, in addition to the pre-treatment confounding control already demanded by Assumptions 1–2, there is no treatment-induced mediator–outcome confounder \(L\).

What randomization of \(T\) does and does not provide. Randomizing the treatment \(T\) satisfies Assumption 1 (both its \(T\)\(Y\) and \(T\)\(M\) components) by design. It does not satisfy Assumption 2 or Assumption 3. The mediator \(M\) is a post-treatment variable that is never randomized; any unobserved variable \(V\) with \(V \to M\) and \(V \to Y\) violates Assumption 2 regardless of how \(T\) was assigned. The structural condition behind Assumption 3 is even more demanding: it can be defeated by a variable \(L\) that is itself caused by the treatment, so randomizing \(T\) actually creates the conditions under which such violations can arise. The upshot is that a randomized experiment identifies the total effect and the first-stage \(T \to M\) effect cleanly, but it does not by itself identify the second-stage \(M \to Y\) effect, and therefore it does not by itself identify the NDE or NIE.

WarningRandomization of \(T\) Does Not Secure the Cross-World Assumption

Unlike unmeasured \(T\)\(Y\) confounders, which randomization of \(T\) can often eliminate, a treatment-induced mediator–outcome confounder \(L\) is itself caused by the treatment: the path \(T \to L\) is set in motion by the intervention, so no aspect of how \(T\) is assigned closes the door on this violation of Assumption 3. The absence of such an \(L\) must instead be argued from design, timing, measurement, or subject-matter knowledge; when that argument is not credible, natural effects are fragile. This is why identification of natural effects is considered strictly harder than identification of the total effect or the CDE.

8.6.2 The Mediation Formula

Under the sequential ignorability assumptions, consistency and composition, and the positivity conditions (P1)–(P2), the cross-world expectation \(\E[Y(t, M(t'))]\) is identified from observational data.

Theorem 8.2 (Mediation Formula (Pearl 2001)) Under sequential ignorability (Assumptions 1–3), consistency and composition, positivity (P1)–(P2), and with discrete \(M\) and \(\mathbf{X}\): \[\E\!\left[Y(t, M(t'))\right] = \sum_{m}\sum_{\mathbf{x}} \E\!\left[Y \mid T{=}t,\, M{=}m,\, \mathbf{X}{=}\mathbf{x}\right] P(M{=}m \mid T{=}t',\, \mathbf{X}{=}\mathbf{x})\, P(\mathbf{X}{=}\mathbf{x}). \tag{8.10}\]

Proof sketch. We derive the formula in five labeled steps, each invoking a specific identification assumption.

Step 1 (law of total expectation and composition). Condition on \(\mathbf{X}\) and marginalize over \(M(t')\): \[\E[Y(t, M(t'))] = \sum_{m,\mathbf{x}} \E[Y(t, m) \mid M(t'){=}m,\, \mathbf{X}{=}\mathbf{x}]\, P(M(t'){=}m \mid \mathbf{X}{=}\mathbf{x})\, P(\mathbf{X}{=}\mathbf{x}),\] using the composition axiom \(Y(t, M(t')) = Y(t, m)\) on the event \(\{M(t') = m\}\).

Step 2 (cross-world independence supplied by the causal model). Invoke Assumption 3: \(Y(t, m) \indep M(t') \mid \mathbf{X}\). This is the step where the structural model earns its keep. Under an NPSEM-IE compatible with the working graph, the counterfactuals \(M(t')\) and \(Y(t, m)\) are functions of disjoint disturbances conditional on \(\mathbf{X}\). It is a cross-world restriction: strictly stronger than anything randomization of \(T\) provides, and not testable from the observed-data distribution. With the cross-world independence in hand, the conditioning on \(M(t')\) in Step 1 can be dropped: \[\E[Y(t, m) \mid M(t'){=}m,\, \mathbf{X}{=}\mathbf{x}] = \E[Y(t, m) \mid \mathbf{X}{=}\mathbf{x}].\] This is the move that distinguishes the mediation formula from an ordinary back-door adjustment: the two counterfactual worlds must be decoupled before the observed-data substitutions that follow.

Step 3 (Assumptions 1 and 2). By \(Y(t, m) \indep T \mid \mathbf{X}\) (the \(Y(t,m)\) marginal of Assumption 1) and \(Y(t, m) \indep M \mid T, \mathbf{X}\) (Assumption 2), conditioning on the observables \(T\) and \(M\) leaves the potential-outcome mean unchanged: \[\E[Y(t, m) \mid \mathbf{X}{=}\mathbf{x}] = \E[Y(t, m) \mid T{=}t,\, M{=}m,\, \mathbf{X}{=}\mathbf{x}].\]

Step 4 (consistency for \(Y\)). On the event \(\{T = t,\, M = m\}\), \(Y(t, m) = Y\), and so the left side equals \(\E[Y \mid T{=}t, M{=}m, \mathbf{X}{=}\mathbf{x}]\).

Step 5 (treatment–mediator component of Assumption 1, and consistency for \(M\)). Rewrite the mediator distribution using \(M(t') \indep T \mid \mathbf{X}\) and consistency (\(M(t') = M\) when \(T = t'\)): \[P(M(t'){=}m \mid \mathbf{X}{=}\mathbf{x}) = P(M{=}m \mid T{=}t',\, \mathbf{X}{=}\mathbf{x}).\]

Substituting Steps 2–5 into Step 1 yields Equation 8.10. The positivity conditions ensure that every conditional expectation and every conditional distribution on the right-hand side is well-defined at each cell that receives positive weight. \(\square\)

Interpretation. The mediation formula “mixes” the outcome regression under \(T = t\) with the mediator distribution under \(T = t'\). To compute the NDE, set \(t = 1\) and \(t' = 0\): take the conditional mean of \(Y\) evaluated at treatment 1, but weight the mediator by its distribution under treatment 0. This counterfactual reweighting is what makes the formula non-trivial. Unlike an ordinary back-door adjustment, which is a functional of a single treatment arm, the mediation formula combines two arms, and the bridge between them is the cross-world independence used in Step 2. Whenever a student is tempted to read Equation 8.10 as a routine standardization formula, the presence of the second treatment index \(t'\) on the right-hand side is the visible trace of that cross-world step.

The NDE and NIE from the formula. Applying Equation 8.10 with \((t, t') = (1, 0)\) and \((0, 0)\): \[\begin{aligned} \mathrm{NDE} &= \sum_{m,\mathbf{x}} \bigl[\E[Y \mid T{=}1, M{=}m, \mathbf{X}{=}\mathbf{x}] - \E[Y \mid T{=}0, M{=}m, \mathbf{X}{=}\mathbf{x}]\bigr] \\ &\qquad\quad \times P(M{=}m \mid T{=}0, \mathbf{X}{=}\mathbf{x})\, P(\mathbf{X}{=}\mathbf{x}), \\[4pt] \mathrm{NIE} &= \sum_{m,\mathbf{x}} \E[Y \mid T{=}1, M{=}m, \mathbf{X}{=}\mathbf{x}] \\ &\qquad\quad \times \bigl[P(M{=}m \mid T{=}1, \mathbf{X}{=}\mathbf{x}) - P(M{=}m \mid T{=}0, \mathbf{X}{=}\mathbf{x})\bigr] P(\mathbf{X}{=}\mathbf{x}). \end{aligned}\]

The mediation formula identifies \(\E[Y(t, M(t'))]\) as a functional of the observed distribution, but it is not itself an estimator. Given data \(\{(Y_i, T_i, M_i, \mathbf{X}_i)\}_{i=1}^n\), a plug-in estimator requires two working models and an averaging step:

  1. Fit an outcome model. Regress \(Y\) on \((T, M, \mathbf{X})\) to obtain \(\hat\mu(t, m, \mathbf{x})\).
  2. Fit a mediator model. Regress \(M\) on \((T, \mathbf{X})\) to obtain \(\hat p(m \mid t, \mathbf{x})\). For continuous \(M\) this may be a conditional density or a conditional sampler.
  3. Predict outcomes at the evaluation arm. For each unit \(i\) and each mediator value \(m\), compute \(\hat\mu(t, m, \mathbf{X}_i)\) at the evaluation treatment level \(t\) (the first argument of \(Y(t, M(t'))\)).
  4. Reweight or simulate the mediator at the reference arm. The cross-world ingredient pairs outcome predictions at \(t\) with the mediator distribution at the reference level \(t'\). For discrete \(M\), weight \(\hat\mu(t, m, \mathbf{X}_i)\) by \(\hat p(m \mid t', \mathbf{X}_i)\); for continuous \(M\), draw Monte Carlo samples and average.
  5. Average over the empirical distribution of \(\mathbf{X}\). The plug-in estimate is \[\frac{1}{n}\sum_{i=1}^n \sum_m \hat\mu(t, m, \mathbf{X}_i)\, \hat p(m \mid t', \mathbf{X}_i), \tag{8.11}\] with the Monte Carlo analogue for continuous \(M\). The NDE and NIE are then differences of such terms at \((t, t') \in \{(1, 0),\, (0, 0),\, (1, 1)\}\).
  6. Quantify uncertainty. Under correctly specified regular parametric mediator and outcome models and the usual smoothness and nonsingularity conditions, a nonparametric bootstrap over units provides asymptotically valid uncertainty quantification; boundary parameters and sparse cells can invalidate the ordinary bootstrap approximation. The influence-function methodology of Chapters 10–11 extends to mediation functionals, but the efficient influence function and the robustness structure are parameter-specific: the mediation functional has its own nuisance components and, depending on the parameterization, a multiply rather than doubly robust structure, so the ATE formulas of those chapters cannot simply be applied as-is.

Caveats. The plug-in estimator Equation 8.11 is consistent only when both \(\hat\mu\) and \(\hat p\) are correctly specified; misspecification in either produces bias. Identification is a separate matter from estimation: a correct mediation formula does not salvage a misspecified regression.

NoteRemark: Three Cautions Before Applying the Mediation Formula
  1. Randomization handles one of the three assumptions, not all three. Randomizing \(T\) secures Assumption 1 by design. It leaves Assumptions 2 and 3 entirely open: whether there is unmeasured \(M\)\(Y\) confounding, and whether the treatment has induced a variable that confounds \(M\) and \(Y\), are questions that the experimental design does not answer.
  2. The formula looks like routine adjustment but is not. The expression \(\sum_m \E[Y \mid t, m, \mathbf{x}]\, P(M{=}m \mid t', \mathbf{x})\) resembles a standardization formula, but it is computing a cross-world expectation \(\E[Y(t, M(t'))]\), not a do-expression. Its validity rests on Assumptions 2 and 3; the latter cannot be read off the observed-data distribution or verified by any experiment.
  3. Sequential ignorability is an untestable assumption bundle. Unlike treatment ignorability, which can sometimes be made credible by design or by a rich covariate set, the mediator–outcome ignorability in Assumption 2 and the cross-world independence in Assumption 3 must be defended on subject-matter grounds in every application. When that defense is weak, the CDE — which requires only a back-door-type adjustment condition — is the more credible estimand.

This remark introduces an alternative decomposition framework used in clinical trials with truncation by death; it may be skipped on first reading.

The NDE and NIE decompose the total effect by causal pathway. An alternative approach — particularly relevant when the mediator is an event that can preclude observation of the outcome, such as death — is principal stratification (Frangakis and Rubin 2002).

In this framework, units are classified by their potential mediator values under each treatment arm: the principal stratum \(\{M(1) = m_1, M(0) = m_0\}\) is the subpopulation that would experience mediator level \(m_1\) under \(T=1\) and \(m_0\) under \(T=0\). This classification is latent — just as compliance types in the IV framework (Chapter 7) are unobservable — but it is well-defined within the potential outcomes language.

The leading application is truncation by death: when the mediator \(M\) is a survival indicator, units with \(M(1) = M(0) = 0\) cannot be observed for any outcome \(Y\). The estimand of interest is then the Survival Average Causal Effect (SACE): \[\mathrm{SACE} = \E\!\bigl[Y(1) - Y(0) \;\big|\; M(1) = M(0) = 1\bigr],\] the average treatment effect among the stratum of units who would survive under both treatments. This estimand is analogous to the LATE in the IV framework: just as LATE conditions on the complier stratum, the SACE conditions on the always-survivor stratum.

Identification of the SACE is more delicate than identification of the total effect. Because the always-survivor stratum is a latent subpopulation whose membership is not observed for any unit, randomization of \(T\) alone does not point-identify the SACE. A monotonicity assumption — \(M(1) \geq M(0)\) for all units — partitions the population into always-survivors, treatment-induced survivors, and never-survivors, and yields informative bounds on the SACE; it does not on its own deliver point identification. Further structure is needed, and several strategies are used: principal ignorability of the outcome given covariates, exclusion-style restrictions on stratum-specific effects, parametric models within strata, or sensitivity analyses. Unlike LATE, where a short standard list of IV assumptions point-identifies the stratum-specific effect, there is no single canonical bundle that does the same for the SACE. In every case, the SACE is not identified by the mediation formula: it requires separate identification arguments tailored to the principal stratification structure.

The principal stratification framework is conceptually distinct from the NDE/NIE framework. Where NDE and NIE decompose the pathway of the effect, the SACE asks about the effect within a subpopulation defined by potential mediator values.

Identification versus estimation. The last three sections have been entirely about causal estimands and their identification: what quantities are we trying to learn, and under what structural assumptions are they expressible as functionals of the observed data distribution? No specific model for \(P(Y \mid T, M, \mathbf{X})\) or \(P(M \mid T, \mathbf{X})\) was required. Section 8.7 turns to a parametric linear model that delivers familiar closed-form estimators under additional modeling assumptions — linearity, additivity, and no \(T \times M\) interaction — that go beyond the identification conditions above. Fitting the regression system does not solve the identification problem; it assumes it has already been solved.

8.7 The Linear Mediation Model: A Historical Special Case

The framework developed above — potential outcomes, cross-world counterfactuals, sequential ignorability — is the modern conceptual foundation of mediation analysis. Before that framework existed, practitioners used a simpler regression-based approach that works cleanly under linearity and no interaction. That approach, due to Baron and Kenny (1986), was enormously influential and remains widely cited. We study it here for three reasons: it builds intuition for the two-stage structure of mediation, its coefficients \(a\), \(b\), \(\tau'\) reappear as special cases of the NDE and NIE under linearity, and it is the dominant approach in many applied literatures that students will encounter.

It is not, however, the general framework. The algebraic decomposition \(\tau = \tau' + ab\) is a consequence of linearity, not a causal identity. Outside linear, no-interaction models it fails, and the product and difference methods it produces are not estimates of the NDE or NIE.

8.7.1 The Baron–Kenny Three-Equation System

The regression-based approach restricts the reduced prototype graph to a linear SEM, conditioning throughout on \(\mathbf{X}\): \[Y = \alpha_1 + \tau T + \boldsymbol{\gamma}_1^{\top}\mathbf{X} + \varepsilon_1, \tag{8.12}\] \[M = \alpha_2 + a T + \boldsymbol{\gamma}_2^{\top}\mathbf{X} + \varepsilon_2, \tag{8.13}\] \[Y = \alpha_3 + \tau' T + b M + \boldsymbol{\gamma}_3^{\top}\mathbf{X} + \varepsilon_3. \tag{8.14}\]

The four coefficients of interest are: \(\tau\), the total effect of \(T\) on \(Y\); \(a\), the first-stage effect of \(T\) on \(M\); \(\tau'\), the direct effect of \(T\) on \(Y\) controlling for \(M\); and \(b\), the second-stage effect of \(M\) on \(Y\) controlling for \(T\).

8.7.2 The Component Pathways

Equations Equation 8.12Equation 8.14 identify three distinct causal sub-problems, each by the back-door formula.

First stage (\(T \to M\)). Equation Equation 8.13 is the parametric implementation of the back-door formula for the effect of \(T\) on \(M\): under linearity and Condition 2 below, conditioning on \(\mathbf{X}\) blocks all back-door paths from \(T\) to \(M\), and the OLS coefficient \(a\) identifies \(\E[M(1)] - \E[M(0)]\). In a randomized experiment, \(a\) is identified without conditioning on anything.

Second stage (\(M \to Y\) given \(T\)). Equation Equation 8.14 implements the back-door formula for the effect of \(M\) on \(Y\) given \(T\). In the reduced prototype graph, \(M\) has parents \(T\) and \(\mathbf{X}\), so every back-door path from \(M\) begins with \(M \leftarrow T\) or \(M \leftarrow \mathbf{X}\). The four such paths are \(M \leftarrow T \to Y\) (blocked by \(T\)); \(M \leftarrow T \leftarrow \mathbf{X} \to Y\) (blocked by \(T\) or \(\mathbf{X}\)); \(M \leftarrow \mathbf{X} \to Y\) (blocked by \(\mathbf{X}\)); and \(M \leftarrow \mathbf{X} \to T \to Y\) (blocked by \(T\) or \(\mathbf{X}\)). Conditioning on \((T, \mathbf{X})\) therefore blocks every back-door path from \(M\) to \(Y\) in this graph, provided Assumption 2 of Section 8.6 holds.

WarningThe Second Stage Is Harder Than It Looks

Even in a randomized experiment where \(T\) is randomized, the mediator \(M\) is never randomized. An unobserved variable \(V\) with \(V \to M\) and \(V \to Y\) creates a back-door path from \(M\) to \(Y\) that conditioning on \((T, \mathbf{X})\) cannot block. This is the central challenge of mediation analysis: identifying the second-stage effect \(b\) requires a no-unmeasured-confounding assumption for the mediator–outcome relationship, an assumption that randomization of the treatment does not provide.

T M Y X V M–Y confounder
Violation of Assumption 2 (mediator--outcome ignorability) under a randomized trial. The treatment $T$ is randomized, so $\mathbf{X} \to T$ is absent; the pre-treatment covariates still affect the mediator and the outcome. The unobserved $V$ creates a back-door path $M \leftarrow V \to Y$ that cannot be blocked by conditioning on $(T, \mathbf{X})$. Neither randomization of $T$ nor adjustment for $\mathbf{X}$ removes this bias.

8.7.3 The Product and Difference Formulas

Proposition 8.1 (Mediation Decomposition in the Linear Model) Under Equation 8.12Equation 8.14, \[\tau \;=\; \tau' \;+\; a b. \tag{8.15}\] The indirect and direct effects are therefore identified by \(\tau_{\mathrm{ind}} = ab\) (product method) and \(\tau_{\mathrm{dir}} = \tau - ab = \tau'\) (difference method).

Proof. Substitute Equation 8.13 into Equation 8.14: \[\begin{aligned} Y &= \alpha_3 + \tau' T + b(\alpha_2 + aT + \boldsymbol{\gamma}_2^{\top}\mathbf{X} + \varepsilon_2) + \boldsymbol{\gamma}_3^{\top}\mathbf{X} + \varepsilon_3 \\ &= (\alpha_3 + b\alpha_2) + (\tau' + ab)\,T + (\boldsymbol{\gamma}_3 + b\boldsymbol{\gamma}_2)^{\top}\mathbf{X} + (b\varepsilon_2 + \varepsilon_3). \end{aligned}\] Comparing with Equation 8.12 gives \(\tau = \tau' + ab\). \(\square\)

WarningThe Decomposition \(\tau = \tau' + ab\) Is an Algebraic Identity, Not a Causal Theorem

The equality \(\tau - \tau' = ab\) is a purely algebraic consequence of substituting one linear equation into another. It holds because linearity makes the indirect effect separable and additive. It does not hold in general:

  • In nonlinear models (binary outcomes, count outcomes, survival models), the product and difference methods yield numerically different estimates. Neither equals the NIE in general.
  • When \(T\) and \(M\) interact in their effect on \(Y\), the CDE depends on \(m\), the NDE and CDE diverge, and the indirect effect cannot be summarized by a single number \(ab\).
  • For non-continuous mediators, the product \(ab\) has no simple causal interpretation outside the linear normal model.

Students who learn the Baron–Kenny decomposition first often overgeneralize it. The correct generalization is the mediation formula (Theorem 8.2), which reduces to \(ab\) and \(\tau'\) only in the linear, no-interaction special case.

Mediation decomposition in the Baron–Kenny linear model.
Effect Formula Path(s)
Total \(\tau\) \(T \to Y\) and \(T \to M \to Y\) combined
Direct \(\tau' = \tau - ab\) \(T \to Y\) only
Indirect \(ab\) \(T \to M \to Y\) only
Proportion mediated \(ab/\tau\) Share of total effect via \(M\)

Running example: a numerical walk-through. Suppose a randomized trial of the behavioral intervention (\(T\)) on depression (\(Y\)), with sleep quality (\(M\)) as the mediator and baseline covariates \(\mathbf{X}\), yields \[\hat\tau = 0.50, \qquad \hat a = 0.40, \qquad \hat b = 0.60, \qquad \hat\tau' = 0.26.\] For concreteness, orient the outcome so that larger \(Y\) means greater improvement (e.g. \(Y\) is the reduction in PHQ-9 score from baseline); a beneficial intervention then has positive coefficients throughout. Then:

  • Indirect effect via sleep (product method): \(\hat a \hat b = 0.40 \times 0.60 = 0.24\).
  • Direct effect (difference method): \(\hat\tau - \hat a \hat b = 0.50 - 0.24 = 0.26 = \hat\tau'\), which matches the fitted \(\hat\tau'\), illustrating the algebraic identity.
  • Proportion mediated: \(0.24/0.50 = 0.48\), so roughly 48% of the total effect operates through sleep improvement in this sample.

These numbers serve the illustration only; Problem 4 asks students to recompute them and to construct a Sobel confidence interval.

8.7.4 Inference: The Sobel Test and Bootstrap

The delta method gives an approximate variance for the product \(\hat a \hat b\): \[\widehat{\mathrm{Var}}(\hat a \hat b) \approx \hat b^2\, \widehat{\mathrm{Var}}(\hat a) + \hat a^2\, \widehat{\mathrm{Var}}(\hat b) + 2\, \hat a\, \hat b\, \widehat{\mathrm{Cov}}(\hat a, \hat b). \tag{8.16}\] The Sobel test (Sobel 1982) uses the simplified approximation that drops the cross-covariance: \[\widehat{\mathrm{Var}}_{\mathrm{Sobel}}(\hat a \hat b) = \hat b^2\, \widehat{\mathrm{Var}}(\hat a) + \hat a^2\, \widehat{\mathrm{Var}}(\hat b), \tag{8.17}\] yielding the \(z\)-statistic \(z = \hat a \hat b / \sqrt{\widehat{\mathrm{Var}}_{\mathrm{Sobel}}(\hat a \hat b)}\). Because \(\hat a\) and \(\hat b\) come from separate regressions, their cross-covariance is often treated as negligible, but it is not zero in general and dropping it is a convention rather than an exact simplification.

Statistical validity versus causal interpretation. The Sobel test is a statistical inference procedure for the product of two fitted regression coefficients. Under standard regression conditions, Equation 8.17 is an approximation to the sampling variance of \(\hat a \hat b\), and the resulting \(z\)-statistic tests whether the regression product \(ab\) is zero. That is a property of the fitted model. Interpreting the tested product as an indirect causal effect is a separate claim that additionally requires the causal identification conditions and the structural restrictions below. When those assumptions fail, the Sobel approximation may still be a reasonable statistical test of whether the regression product is zero, but that product no longer corresponds to the indirect effect of \(T\) on \(Y\) through \(M\).

In practice, bootstrap confidence intervals for \(ab\) are preferred over the Sobel test for inference, because the distribution of a product of estimates is skewed in finite samples — especially when either \(\hat a\) or \(\hat b\) is small.

WarningThe “Significance of Both Paths” Criterion Is Not a Test for Mediation

A common misuse of the Baron–Kenny framework, sometimes called the “causal steps” approach, declares mediation to be present when: (i) \(T \to Y\) is significant in Equation 8.12; (ii) \(T \to M\) is significant in Equation 8.13; (iii) \(M \to Y\) is significant in Equation 8.14. This approach has three serious defects.

  1. Statistical significance \(\neq\) mediation. A significant \(\hat a\) and \(\hat b\) does not imply that the product \(ab\) is meaningfully large, nor that it is identified. Low power for \(\hat a\) or \(\hat b\) individually does not imply low power for \(ab\) (and vice versa).
  2. The Sobel test tests a regression product, not a causal quantity. Under standard regression conditions, Equation 8.17 is a valid statistical approximation regardless of whether the Baron–Kenny causal assumptions hold. What those assumptions control is the interpretation of the tested product as an indirect effect.
  3. Zero total effect does not preclude indirect effects. It is possible for \(\tau \approx 0\) while \(ab \neq 0\) and \(\tau' \neq 0\) (direct and indirect effects of opposite sign that cancel). Requiring significance of \(\hat\tau\) as a precondition will miss these cases entirely.

The modern alternative is to estimate \(ab\) (or the NIE directly via the mediation formula), construct bootstrap or nonparametric confidence intervals, and interpret the result as a point estimate with uncertainty — not as a significance test.

8.7.5 The Baron–Kenny Assumptions: Two Distinct Categories

The conditions required for the Baron–Kenny decomposition fall into two fundamentally different categories that should not be conflated.

NoteBaron–Kenny Causal Ignorability Conditions

These are causal identification assumptions. They concern unmeasured confounding and have graphical interpretations. Violating them introduces bias that no amount of additional data can remove given the observed variables.

  1. No unmeasured \(T\)\(Y\) confounding. \(\varepsilon_1 \indep T \mid \mathbf{X}\). Graphically: \(\mathbf{X}\) blocks all back-door paths from \(T\) to \(Y\).
  2. No unmeasured \(T\)\(M\) confounding. \(\varepsilon_2 \indep T \mid \mathbf{X}\). Graphically: \(\mathbf{X}\) blocks all back-door paths from \(T\) to \(M\).
  3. No unmeasured \(M\)\(Y\) confounding given \(T\). \(\varepsilon_3 \indep M \mid T, \mathbf{X}\). Graphically: \((T, \mathbf{X})\) blocks all back-door paths from \(M\) to \(Y\). This condition is not implied by randomization of \(T\).

A caution on reading these displays: the \(\varepsilon_j\) are structural errors of the SEM, not fitted residuals. Population least-squares residuals are orthogonal to the regressors by construction, so the displayed independences are substantive restrictions only under the structural reading; they cannot be checked by inspecting regression output.

NoteBaron–Kenny Structural Modeling Restrictions

These are parametric modeling assumptions. They concern the functional form of the structural equations, not unmeasured confounding. Violating them does not introduce identification bias in the causal sense, but it does mean that the product formula \(ab\) and the coefficient \(\tau'\) no longer equal the NDE and NIE.

  1. Linearity and additivity. The structural equations Equation 8.12Equation 8.14 are correctly specified as linear and additive in their arguments.
  2. No \(T \times M\) interaction. The effect of \(M\) on \(Y\) does not depend on the level of \(T\).

Restrictions 4 and 5 can be probed by residual diagnostics and interaction terms, but model adequacy cannot be conclusively established from observed data alone. Conditions 1–3 cannot be tested from the observed data at all.

In the reduced prototype graph, Conditions 1 and 2 hold because \(\mathbf{X}\) blocks every \(T\)\(Y\) and \(T\)\(M\) back-door path by construction. Condition 3 is different: it requires the absence of any variable \(V \to M\) and \(V \to Y\) not captured by \((T, \mathbf{X})\), and it is never guaranteed by randomization of \(T\) because \(M\) is a post-treatment variable that is not itself randomized. Restrictions 4 and 5 have no graphical counterpart.

Baron–Kenny as a special case of the modern framework. Under Conditions 1–3 and Restrictions 4–5, the Baron–Kenny decomposition delivers valid identification formulas for the NDE and NIE: linearity and no interaction guarantee that the population coefficients satisfy \(\mathrm{NDE} = \tau'\) and \(\mathrm{NIE} = ab\). Translating these identification formulas into consistent estimators requires a separate set of conditions. Outside Conditions 1–3 and Restrictions 4–5, \(\tau'\) and \(ab\) need not identify the CDE or the natural effects: the listed conditions are sufficient rather than necessary, and special structures can restore particular equalities, but in general the coefficients of a misspecified structural system carry no causal interpretation.

NoteFramework Map
Estimand Framework needed Section
Total effect \(P(y \mid \doop(t))\) Do-calculus Ch. 5 (back-door)
Controlled direct effect (CDE) Do-calculus Section 8.4
Natural direct effect (NDE) Potential outcomes Section 8.5
Natural indirect effect (NIE) Potential outcomes Section 8.5
Total effect via front-door Do-calculus Section 8.8

The first row and the last row name the same estimand — the total effect of \(T\) on \(Y\) — but identified by different routes: standard back-door adjustment in the first row, and the front-door formula in the last. The CDE, NDE, and NIE are distinct estimands.

Chapter pivot: from decomposing an identified total effect to identifying a confounded one. Everything up to this point has taken identification of the total effect for granted — the working assumption of Section 8.2.4 absorbed all \(T\)\(Y\) confounding into observed covariates — and asked how that effect decomposes across pathways. The remainder of the chapter reverses the direction of travel. In Section 8.8, the mediator stops being the object of study and becomes an identification device: the total effect itself is unidentified by adjustment, and the mediation structure of the graph is what rescues it.

8.8 Front-Door Identification

8.8.1 The Front-Door DAG

Every identification strategy above assumed that an observed covariate set \(\mathbf{X}\) exists that blocks the back-door paths from \(T\) to \(Y\) through \(U\). What if \(U\) is wholly unobserved and no such adjustment set exists? In the prototype mediation DAG, the back-door criterion then fails for the total effect, and the methods of earlier sections are unavailable. Chapter 3 previewed the front-door strategy for exactly this situation; this section gives the full development — the identification theorem with its support conditions, the do-calculus proof, the failure modes, and a numerical illustration — and connects it to mediation.

The front-door criterion turns this obstacle into an opportunity: under two additional restrictions on the prototype mediation graph, the mediation structure itself provides identification of the total effect without conditioning on \(U\). The two restrictions are: remove the direct \(T \to Y\) edge, so that \(M\) fully mediates the effect of \(T\) on \(Y\); and require that \(U\) has no arrow into \(M\), so that the \(T \to M\) sub-effect is unconfounded.

Why the running example does not apply here. The depression/sleep scenario fails Condition 1 (full mediation): a behavioral intervention on depression plausibly operates through several non-sleep channels — cognitive restructuring, behavioral activation, therapeutic alliance — so the direct edge \(T \to Y\) is present. The front-door formula is therefore unavailable for that setting. The canonical example in which the three front-door conditions are all plausible is Pearl’s smoking–tar–cancer graph: \(T\) denotes smoking, \(M\) the amount of tar deposited in the lungs, \(Y\) a lung-cancer outcome, and \(U\) an unobserved genetic susceptibility. If all of smoking’s carcinogenic effect flows through tar, and the genetic susceptibility does not act on tar directly, the front-door formula identifies the causal effect of smoking on cancer without ever observing \(U\). This is a stylized identification example, not a claim that measured tar burden satisfies the front-door conditions in actual epidemiologic data: a genetic susceptibility could plausibly affect inhalation behavior, tar metabolism, or the measurement of tar itself.

T M Y U treatment mediator outcome unobserved
The front-door graph. Compared with the prototype mediation DAG, two edges are absent: the direct $T \to Y$ edge is removed, and $U$ has no arrow into $M$. These two omissions are what enable identification: $M$ fully mediates the effect of $T$ on $Y$, and the $T \to M$ sub-effect is free of confounding. The front-door formula identifies $P(y \mid \doop(t))$ without observing or conditioning on $U$.
NoteTwo Different Questions About a Mediator-Like Variable

The front-door graph looks like the mediation graph with two edges removed, and \(M\) sits between \(T\) and \(Y\) in both. The resemblance is structural, but the questions being asked are fundamentally different.

Ordinary mediation Front-door identification
Question How much of the total effect of \(T\) on \(Y\) flows through \(M\)? Can \(M\) be used to identify the total effect despite unmeasured \(T\)\(Y\) confounding?
Role of \(M\) Pathway: carries part of the causal effect Instrument-like relay: routes around unmeasured confounding
Target estimand NDE, NIE (decomposition) \(P(y \mid \doop(t))\) (the total effect itself)
What \(M\) must satisfy Sequential ignorability for the \(M\)\(Y\) sub-problem Conditions 1–3 below (graphical, no potential outcomes required)
Direct \(T \to Y\) edge Present Absent (required)

The front-door formula does not decompose the total effect into direct and indirect components. It identifies the total effect as a whole, using \(M\) as a relay that is unconfounded on the \(T\)-side; on the \(Y\)-side, the back-door path through \(T\) is closed by conditioning on \(T\), after which the result is averaged over the observed marginal distribution of \(T\). A researcher who wants to know how much of the effect goes through \(M\) needs ordinary mediation analysis.

Warning“Instrument-Like” Does Not Mean Instrument

The phrase “instrument-like” above is a loose analogy for one feature: \(M\) helps route around unmeasured \(T\)\(Y\) confounding, which is also what an instrumental variable does. \(M\) is not an instrument in the IV sense. An instrument \(Z\) satisfies the exclusion restriction — it affects \(Y\) only through \(T\) — so \(Z\) is excluded from the \(Y\) structural equation. By contrast, the front-door mediator \(M\) lies on the causal path from \(T\) to \(Y\) and enters the \(Y\) structural equation directly. The two strategies share the goal of identifying \(T\)’s total effect under unobserved confounding, but they place the third variable in structurally opposite positions.

8.8.2 The Three Front-Door Conditions

NoteFront-Door Conditions

A variable \(M\) satisfies the front-door criterion for the effect of \(T\) on \(Y\) if:

  1. Full mediation. All directed paths from \(T\) to \(Y\) pass through \(M\). (In the prototype graph this is equivalent to the absence of a direct \(T \to Y\) edge; in richer graphs, deleting that edge is not sufficient if some other directed path bypasses \(M\).)
  2. No unblocked back-door path from \(T\) to \(M\). (In the prototype graph, which carries no adjustment covariates, this means there is no latent common cause of \(T\) and \(M\); in richer graphs a common cause may exist provided observed covariates block the resulting paths.)
  3. No unblocked back-door path from \(M\) to \(Y\) given \(T\). All back-door paths from \(M\) to \(Y\) are blocked by conditioning on \(T\).

The criterion applies to the causal mediator represented in the graph: replacing \(M\) by a mismeasured proxy need not preserve any of the three conditions — a proxy may fail to intercept all directed \(T\)-to-\(Y\) paths, and measurement error can open both stages of the argument. Measurement error in \(M\) is thus an identification threat, not merely an estimation nuisance.

In the front-door graph: Condition 1 holds because there is no \(T \to Y\) edge. Condition 2 holds, but for a subtler reason than “no back-door path exists”: there is a back-door path from \(T\) to \(M\), namely \(T \leftarrow U \to Y \leftarrow M\), but it is blocked by the unconditioned collider \(Y\). Because \(U\) has no arrow into \(M\), no open back-door path remains — and the criterion requires only the absence of open paths. Condition 3 holds because the only back-door path from \(M\) to \(Y\) is \(M \leftarrow T \leftarrow U \to Y\), which is blocked by conditioning on \(T\).

NoteRemark: Single Mediator vs. a Set of Mediators

Pearl (1995) states the front-door criterion for a set of intermediate variables \(\mathbf{M} = \{M_1, \ldots, M_k\}\): the set must intercept all directed paths from \(T\) to \(Y\), have no unblocked back-door from \(T\), and have all back-door paths to \(Y\) blocked by \(T\). The formula generalizes accordingly, with the inner sum running over realizations of \(\mathbf{M}\). For pedagogical clarity, this chapter treats the single-mediator case throughout; the generalization is straightforward conceptually but the multivariate \(\mathbf{M}\) introduces additional positivity and estimation challenges.

Example 8.1 (When the Front-Door Criterion Fails) Each of the three front-door conditions is load-bearing. We illustrate what breaks by modifying the front-door graph in two separate ways. As Case (a) shows, a single added edge can violate more than one condition at once.

Case (a): Conditions 2 and 3 fail — \(U\) also causes \(M\). Start with the front-door graph and add the edge \(U \to M\), so the unobserved confounder directly affects the mediator as well as the treatment and the outcome.

Condition 1 still holds: every directed path from \(T\) to \(Y\) passes through \(M\). Condition 2, however, now fails: the path \(T \leftarrow U \to M\) is an unblocked back-door path from \(T\) to \(M\). Because \(U\) is unobserved, there is no observed variable that can block this path. Condition 3 also fails: the path \(M \leftarrow U \to Y\) is an open back-door path from \(M\) to \(Y\), and conditioning on \(T\) does not block it, since \(T\) does not lie on the path.

Why the formula breaks. Both stages collapse. Stage 1 uses \(P(M{=}m \mid T{=}t)\) as if the \(T \to M\) link were unconfounded. But with \(U \to M\), the association between \(T\) and \(M\) is partly due to their shared cause \(U\): the quantity \(P(m \mid t)\) mixes cause and confounding and no longer identifies \(P(m \mid \doop(T{=}t))\). Likewise, because the open path \(M \leftarrow U \to Y\) survives conditioning on \(T\), the Stage 2 functional does not generally equal \(P(y \mid \doop(M{=}m))\). Failure of either condition alone would already suffice to invalidate the formula.

Case (b): Condition 3 fails — the \(M \to Y\) link has an extra unobserved confounder. Return to the prototypical graph and add a second unobserved variable \(V\) that affects both \(M\) and \(Y\).

Conditions 1 and 2 still hold. But Condition 3 now fails: the path \(M \leftarrow V \to Y\) is a back-door path from \(M\) to \(Y\) that is not blocked by conditioning on \(T\). The path does not pass through \(T\), so fixing \(T\) does nothing to close it.

Why the formula breaks. Stage 2 uses \(\sum_{t'} P(y \mid t', m)\,P(t')\) as the back-door adjustment formula for \(P(y \mid \doop(M{=}m))\), with \(T\) as the adjustment variable. But \(T\) only satisfies the back-door criterion for \(M \to Y\) when it blocks all back-door paths from \(M\) to \(Y\). The new path \(M \leftarrow V \to Y\) is not blocked by \(T\), so \(T\) is no longer a valid adjustment set for Stage 2.

Summary.

Modification Condition(s) violated Consequence
Add \(T \to Y\) Cond. 1 A causal path bypasses \(M\)
Add \(U \to M\) Conds. 2 and 3 Stages 1 and 2 both fail
Add \(V \to M\) and \(V \to Y\) Cond. 3 Stage 2 fails

In each case the graphical check detects the failure before any formula is written down. The first row’s modification — adding a direct \(T \to Y\) edge — shows that the absence of such an edge in the front-door graph is not an independent requirement: it is simply one way of satisfying Condition 1, which demands that all directed paths from \(T\) to \(Y\) pass through \(M\).

8.8.3 Derivation of the Front-Door Formula

Theorem 8.3 (Front-Door Formula (Pearl 1995)) Suppose \(M\) satisfies the front-door criterion for the effect of \(T\) on \(Y\). Assume further the positivity conditions:

(F1) Treatment positivity. \(P(T = t) > 0\) at the evaluation level \(t\), so that \(P(m \mid t)\) is well-defined; every \(t'\) entering the inner sum lies in the support of \(T\).

(F2) Mediator overlap across treatment arms. For every mediator value \(m\) with \(P(M{=}m \mid T{=}t) > 0\) and every \(t'\) with \(P(T{=}t') > 0\), \(P(M{=}m \mid T{=}t') > 0\).

Then \[P\!\left(y \mid \doop(T{=}t)\right) = \sum_m P(m \mid t) \sum_{t'} P(y \mid m, t')\, P(t'). \tag{8.18}\]

Condition (F2) is the front-door analogue of the mediator overlap (P2): it ensures that \(P(y \mid m, t')\) is empirically defined at every \((m, t')\) cell that receives positive weight in Equation 8.18.

Proof. We apply the do-calculus in three steps, each exploiting one of the front-door conditions.

Step 1: Identify the effect of \(T\) on \(M\). Apply Rule 2 with \(X = \varnothing\), \(Z = T\), \(W = \varnothing\); the required graph is \(\Gcal_{\underline{T}}\), obtained by deleting \(T\)’s outgoing arrow. The only remaining path between \(T\) and \(M\) is \(T \leftarrow U \to Y \leftarrow M\), blocked at the unconditioned collider \(Y\); Condition 2 is what guarantees that no open path survives (an edge \(U \to M\) would leave \(T \leftarrow U \to M\) open). Hence \((M \indep T)_{\Gcal_{\underline{T}}}\), and Rule 2 gives \[P\!\left(m \mid \doop(T{=}t)\right) = P(m \mid t).\]

Step 2: Identify the effect of \(M\) on \(Y\). By Condition 3, conditioning on \(T\) blocks all back-door paths from \(M\) to \(Y\), so \(\{T\}\) is a valid back-door adjustment set: \[P\!\left(y \mid \doop(M{=}m)\right) = \sum_{t'} P(y \mid m, t')\, P(t').\]

Step 3: Combine via full mediation. The law of total probability applied under the intervention \(\doop(T{=}t)\) gives \[P\!\left(y \mid \doop(T{=}t)\right) = \sum_m P\!\left(m \mid \doop(T{=}t)\right)\, P\!\left(y \mid \doop(T{=}t),\, m\right), \tag{8.19}\] where the second factor conditions on \(M{=}m\). Converting that conditioning into the intervention \(\doop(M{=}m)\) is the delicate step, and it is here that Conditions 3 and 1 do their work, in that order.

First, apply Rule 2 with \(X = T\), \(Z = M\), \(W = \varnothing\); the required graph is \(\Gcal_{\overline{T}\,\underline{M}}\). Deleting the arrows into \(T\) and out of \(M\) leaves only the edges \(T \to M\) and \(U \to Y\), so no path connects \(M\) to \(Y\) and \((Y \indep M \mid T)_{\Gcal_{\overline{T}\,\underline{M}}}\) holds — Condition 3 secures this in general, since any surviving path would be a back-door path from \(M\) to \(Y\) not blocked by \(T\). Rule 2 gives \[P\!\left(y \mid \doop(T{=}t),\, m\right) = P\!\left(y \mid \doop(T{=}t),\, \doop(M{=}m)\right).\] Second, apply Rule 3 with \(X = M\), \(Z = T\), \(W = \varnothing\); the required graph is \(\Gcal_{\overline{M}\,\overline{T}}\). Deleting the arrows into both \(M\) and \(T\) isolates \(T\), so the now-redundant \(\doop(T{=}t)\) can be dropped: \[P\!\left(y \mid \doop(T{=}t),\, \doop(M{=}m)\right) = P\!\left(y \mid \doop(M{=}m)\right).\] Condition 1 secures this second step in general: every directed path from \(T\) to \(Y\) passes through \(M\), whose incoming arrows have been deleted, so no route from \(T\) to \(Y\) survives. This is the formal content of “the total effect of \(T\) factors through \(M\).” Substituting Steps 1 and 2 into Equation 8.19 yields Equation 8.18. \(\square\)

NoteRemark: Two Routes to the Same Two-Stage Logic

The two-stage structure of the proof — first identify the \(T \to M\) effect, then the \(M \to Y\) effect, and finally compose — is the do-calculus analogue of the two-stage reasoning in the mediation formula (Theorem 8.2). The mediation formula accomplishes this via sequential ignorability — conditional independence assumptions on potential outcomes. The front-door formula accomplishes it via the assumed causal graph and its intervention semantics: Condition 2 makes Step 1 unconfounded by construction, and Condition 3 makes Step 2 unconfounded because conditioning on \(T\) blocks the back-door path \(M \leftarrow T \leftarrow U \to Y\). No covariate set \(\mathbf{X}\) is needed, because the structural restrictions on the graph substitute for the ignorability assumptions — given, as always, the modularity of the structural equations, positivity (F1)–(F2), and a correctly measured mediator.

8.8.4 Intuition: Routing Around Confounding

The front-door formula achieves identification in two steps that each use only unconfounded variation.

  1. \(T\) to \(M\): There is no confounding on the \(T \to M\) edge (\(U\) does not affect \(M\)), so \(P(m \mid t)\) is the causal effect of \(T\) on \(M\).
  2. \(M\) to \(Y\): There is back-door confounding on \(M \to Y\) through the path \(M \leftarrow T \leftarrow U \to Y\), but \(T\) is a non-collider on this path, so conditioning on \(T\) closes it. The resulting conditional \(P(y \mid m, t')\) is then averaged over the observed marginal distribution of \(T\).

The key insight is that \(U\) confounds \(T\) and \(Y\) but not the \(T \to M\) edge. The front-door formula exploits this asymmetry to identify the total causal effect without ever observing or conditioning on \(U\).

NoteExample: Front-Door Formula — Binary Numerical Computation

Let \(T, M, Y \in \{0, 1\}\) with \(U\) unobserved and the graph \(T \to M \to Y\), \(U \to T\), \(U \to Y\). The observed joint distribution \(P(T, M, Y)\) is given in the following population table.

\(T\) \(M\) \(Y\) \(P(T,M,Y)\)
0 0 0 0.360
0 0 1 0.040
0 1 0 0.050
0 1 1 0.050
1 0 0 0.070
1 0 1 0.030
1 1 0 0.120
1 1 1 0.280

All quantities needed for the front-door formula can be read off by marginalization. For example, \(P(T{=}1) = 0.070 + 0.030 + 0.120 + 0.280 = 0.5\), and \(P(M{=}1 \mid T{=}1) = (0.120 + 0.280)/0.5 = 0.8\).

The inner sum (Stage 2) does not depend on the intervention value \(t\), so compute it once for each value of \(M\): \[S_0 = \sum_{t'} P(Y{=}1 \mid t', M{=}0)\,P(t') = 0.1 \times 0.5 + 0.3 \times 0.5 = 0.20,\] \[S_1 = \sum_{t'} P(Y{=}1 \mid t', M{=}1)\,P(t') = 0.5 \times 0.5 + 0.7 \times 0.5 = 0.60.\] Now apply the outer sum (Stage 1) for each intervention value: \[\begin{aligned} P(Y{=}1 \mid \doop(T{=}1)) &= 0.2 \times 0.20 + 0.8 \times 0.60 = 0.52, \\ P(Y{=}1 \mid \doop(T{=}0)) &= 0.8 \times 0.20 + 0.2 \times 0.60 = 0.28. \end{aligned}\] Under the assumed front-door graph, the identified average causal effect is \(0.52 - 0.28 = 0.24\).

Comparison with the naive approach. A naive analyst ignoring the unobserved confounder \(U\) would estimate the effect by the observational contrast: \[P(Y{=}1 \mid T{=}1) = \frac{0.030 + 0.280}{0.5} = 0.62, \qquad P(Y{=}1 \mid T{=}0) = \frac{0.040 + 0.050}{0.5} = 0.18.\] The naive estimate is \(0.62 - 0.18 = 0.44\), almost twice the effect of \(0.24\) identified under the assumed graph. The excess association of \(0.20\) is transmitted along the open back-door path \(T \leftarrow U \to Y\): the graph determines that this non-causal route exists, and the observed table determines its net contribution, but the graph alone says nothing about the sign or magnitude of \(U\)’s individual effects. The front-door formula removes this bias by routing the effect estimate through the unconfounded \(T \to M\) link and the \(T\)-adjusted \(M \to Y\) link, never directly comparing the \(T{=}1\) and \(T{=}0\) groups on \(Y\).

Note how Stage 2 averages \(P(Y \mid t', m)\) over the marginal distribution \(P(t')\), not the treatment-conditional distribution \(P(t' \mid m)\) — the re-weighting that removes the confounding introduced by \(U\).

A final caveat: the observed table does not itself validate the assumed graph. Causal models on the front-door graph that reproduce this table do exist, so the computation is internally coherent, but the calculation is conditional on assuming that the data arose from some model satisfying the front-door graph. The graphical conditions are assumptions supplied from outside the data; the formula converts them into a number, it does not confirm them.

8.8.5 Front-Door vs. Prototype Mediation

Feature Prototype mediation DAG Front-door DAG
Direct \(T \to Y\) edge Present Absent
\(U\) confounds \(T\)\(Y\) Yes Yes
Goal Mechanism analysis (decomposition under natural effects; level-specific direct effect under CDE) Identify total effect despite \(T\)\(Y\) confounding
Requires Assumption 2 (\(M\)\(Y\) no confounding given \(T\)) Yes Satisfied by front-door Condition 3, not by randomization or by default
Identified by front-door formula No (direct \(T \to Y\) breaks Condition 1) Yes

8.9 Mediation vs. Instrumental Variables

Mediation analysis and instrumental variables both involve a third variable connected to the treatment–outcome relationship, but the causal role of that variable is fundamentally different in the two frameworks.

8.9.1 Structural Comparison

Feature Instrumental Variables Mediation Analysis
Position of third variable Pre-treatment (\(Z\) precedes \(T\)) Post-treatment (\(M\) follows \(T\))
Causal role Exogenous source of variation in \(T\) Pathway through which \(T\) affects \(Y\)
Primary goal Identification of \(T \to Y\) effect under confounding Mechanism analysis of \(T \to Y\) effect
Key assumption Exclusion: \(Z\) affects \(Y\) only through \(T\) Sequential ignorability: mediator–outcome ignorability plus the cross-world independence
Estimand LATE (Wald, under monotonicity); under a homogeneous linear effect, the common structural coefficient (Ch. 7) NDE, NIE, or CDE
Unobserved \(T\)\(Y\) confounders Permitted (IV routes around them) Must be addressed separately (Assumption 1)
Testability Relevance testable; exclusion untestable Sequential ignorability untestable

8.9.2 The Conceptual Contrast

The contrast is sharpest in terms of what the third variable does in each framework.

In IV, \(Z\) is a handle: it generates exogenous variation in \(T\) that is free of the back-door confounding path \(T \leftarrow U \rightarrow Y\). The instrument is valuable precisely because it is not on the causal path from \(T\) to \(Y\) — the exclusion restriction says that \(Z\) cannot directly affect \(Y\).

In mediation analysis, \(M\) is a pathway: it transmits part of the causal effect of \(T\) to \(Y\). The mediator is valuable precisely because it is on the causal path. The analysis goal depends on the estimand: under natural effects, the question is how much of the total effect flows through this pathway; under the CDE, the question is what part of \(T\)’s effect on \(Y\) remains when \(M\) is held fixed at a chosen level.

NoteThe Key Conceptual Distinction
IV Mediation
What the third variable does Generates clean variation in \(T\) (exogenous source) Carries part of \(T\)’s causal effect (pathway)
Exclusion vs. inclusion \(Z\) excluded from \(Y\)’s structural equation \(M\) included in \(Y\)’s structural equation
Question answered Does \(T\) cause \(Y\)? How does \(T\) cause \(Y\)?

8.9.3 Can the Same Variable Be Both?

It is worth asking whether the same variable \(M\) could serve as both a mediator and an instrument. The answer is: not for the same treatment–outcome relation. Within a single causal question of how \(T\) affects \(Y\), the mediator role places \(M\) on the causal path (inclusion required), whereas the IV role demands the exclusion restriction — that \(M\) affects \(Y\) only through \(T\). These are mutually incompatible structural assumptions about the same intermediate variable.

The qualifier for the same treatment–outcome relation matters. A variable that mediates one causal question may, in a differently oriented question, serve as an instrument-like exogenous source of variation. For example, occupation might mediate the effect of education on wages while functioning as an exogenous source of variation in some unrelated downstream analysis where neither education nor wages is the target. The incompatibility is between the two roles within one fixed \((T, Y)\) pair, not across different scientific questions.

The front-door identification formula is the closest bridge between the two within a single \((T, Y)\) analysis: it uses the mediator \(M\) to identify the total effect of \(T\) on \(Y\) even when \(T\) is confounded — a goal that IV also pursues. But, as the warning box in Section 8.8 stresses, the front-door \(M\) is not an instrument: it lies on the causal path, satisfies a graphical front-door condition, and substitutes structural restrictions on the graph for the exclusion restriction that defines an IV.

8.10 Summary

  1. Mediation studies mechanisms. Mediation analysis aims to study the mechanisms by which \(T\) affects \(Y\) by defining direct and indirect effect concepts that target the pathways \(T \to Y\) and \(T \to M \to Y\). Of the two direct-effect families developed in this chapter, only the natural effects are constructed to yield an additive decomposition of the total effect; the controlled direct effect is not. The challenge throughout is that the mediator is a post-treatment variable that may be confounded.

  2. The total effect is the baseline estimand. The TE \(= \E[Y(1) - Y(0)]\) captures all pathways combined. It may be identified by randomization, by back-door adjustment on a sufficient covariate set, by front-door identification, or by other identifying structures; which strategy applies depends on the available causal structure.

  3. The CDE uses do-calculus. The controlled direct effect fixes \(M = m\) by joint intervention \(\doop(T, M)\). Under consistency, a well-defined joint intervention, an appropriate adjustment condition for the pair \((T, M)\), and joint positivity, it is identified by standardization (Theorem 8.1). The CDE depends on the fixed level \(m\) and does not have a natural “indirect” complement.

  4. Natural effects require potential outcomes. The NDE and NIE involve cross-world counterfactuals \(Y(t, M(t'))\) that cannot be expressed with the do-operator alone. They decompose the total effect as TE \(=\) NDE \(+\) NIE, and are identified by the mediation formula under sequential ignorability. The critical cross-world independence (Assumption 3) is not secured by randomization of \(T\). Under the NPSEM-IE used in these notes, it follows from the independent-error structural model together with adequate control of pre-treatment common causes and the absence of a treatment-induced mediator–outcome confounder.

  5. The linear model simplifies but restricts. The Baron–Kenny three-equation system gives the clean formulas \(\tau_{\mathrm{ind}} = ab\) and \(\tau_{\mathrm{dir}} = \tau'\), with \(\tau = \tau' + ab\). This decomposition is purely algebraic and holds only under linearity and no interaction. In nonlinear or interaction settings, the product and difference methods disagree and NDE \(\neq\) CDE.

  6. Front-door identification uses mediation structure. When \(M\) fully mediates \(T \to Y\), no \(T \to M\) confounding exists, and \(T\) blocks the back-door paths from \(M\) to \(Y\), the front-door formula identifies the total effect despite unobserved \(T\)\(Y\) confounding. It achieves this by composing two unconfounded sub-effects.

  7. Mediation and IV are complementary, not equivalent. IV uses a pre-treatment variable to generate exogenous variation in \(T\); mediation uses a post-treatment variable to study how the causal effect operates. The same variable cannot simultaneously serve as a mediator and a valid IV for the same treatment–outcome relation.

8.11 Problems

1. Identifying the CDE. Consider the DAG: \(T \to M\), \(T \to Y\), \(M \to Y\), \(X \to T\), \(X \to M\), \(X \to Y\), with all variables observed.

  1. Write the identification formula for \(\E[Y \mid \doop(T{=}1), \doop(M{=}m)]\) using baseline adjustment for the joint intervention \((T, M)\) as in Theorem 8.1.
  2. Add an unobserved \(U\) with \(U \to T\) and \(U \to Y\). Is the CDE identified by any adjustment set consisting of observed variables? In particular, does standardization over \(X\) remain valid? Explain which condition of Theorem 8.1 fails.
  3. Instead add \(U\) with \(U \to M\) and \(U \to Y\). Answer the same questions.

2. CDE vs. total effect. In the reduced prototype graph, let \(\mathbf{Z}\) satisfy the back-door criterion for both the total effect and the joint intervention \((T, M)\).

  1. Write expressions for the total effect and the \(\mathrm{CDE}(m)\) using the back-door formula.
  2. Give a sufficient graphical condition under which fixing \(M\) removes no causal pathway from \(T\) to \(Y\) — for example, that no directed path from \(T\) to \(Y\) passes through \(M\). Interpret this condition.
  3. (second pass) Show by example that the graphical condition of part (b) alone does not make \(\mathrm{CDE}(m)\) equal to the total effect for all \(m\): exhibit a structural equation for \(Y\) with a \(T \times M\) interaction in which the two quantities differ even though \(M\) is not affected by \(T\). What additional functional condition closes the gap?

3. Natural direct and indirect effects. Verify the NDE \(+\) NIE \(=\) TE decomposition algebraically for the linear SEM \(M = \alpha T + \eta\), \(Y = \beta T + \gamma M + \varepsilon\) (no interaction).

  1. Compute \(Y(t, M(t'))\) in the linear model, writing the noise terms explicitly: \(Y(t, M(t')) = \beta t + \gamma \alpha t' + \gamma\eta + \varepsilon\). Note that the same \(\eta\) appears for every value of \(t'\) — the structural construction shares the mediator disturbance across intervention worlds, and this is exactly what makes the nested counterfactual meaningful.
  2. Derive \(\mathrm{NDE} = \beta\) and \(\mathrm{NIE} = \alpha\gamma\) from Equation 8.7Equation 8.8.
  3. Confirm \(\mathrm{NDE} + \mathrm{NIE} = \beta + \alpha\gamma = \mathrm{TE}\).
  4. Now suppose a \(T \times M\) interaction is added: \(Y = \beta T + \gamma M + \delta (T \cdot M) + \varepsilon\). Show that \(\mathrm{CDE}(m) = \beta + \delta m\) and \(\mathrm{NDE} = \beta + \delta\, \E[M(0)]\). Conclude that the two generally differ when \(\delta \neq 0\), and identify the one mediator level at which they coincide.

4. The Baron–Kenny three-equation system. In the reduced prototype graph with the linear SEM Equation 8.12Equation 8.14:

  1. State the three identification assumptions. For each, give the graphical condition in terms of back-door paths.
  2. Derive the equality \(\tau = \tau' + ab\) algebraically.
  3. Suppose estimated coefficients are \(\hat\tau = 0.50\), \(\hat a = 0.40\), \(\hat b = 0.60\), \(\hat\tau' = 0.26\). Compute the indirect effect by both the product and difference methods. Do they agree? Compute the proportion mediated, and state the conditions — direct and indirect effects of the same sign, and a total effect bounded away from zero — under which calling this ratio a “proportion” is meaningful. What happens to \(ab/\tau\) when \(\tau' = -ab\)?
  4. (optional computational practice) An analyst reports \(\hat a = 0.40\) with \(\widehat{\mathrm{SE}}(\hat a) = 0.08\), and \(\hat b = 0.60\) with \(\widehat{\mathrm{SE}}(\hat b) = 0.10\). Compute the Sobel standard error for \(\hat a \hat b\) using Equation 8.17 and construct an approximate 95% confidence interval.

5. The critical role of Assumption 2 (mediator–outcome ignorability). Consider the graph in which an unobserved \(V\) has \(V \to M\) and \(V \to Y\), with \(T\) randomized.

  1. Identify all back-door paths from \(M\) to \(Y\) in this graph.
  2. Can any combination of observed variables \((T, \mathbf{X})\) block all of these paths? Explain using d-separation.
  3. Suppose an analyst fits Equation 8.14 ignoring \(V\) and obtains \(\hat b = 0.80\). Under a linear additive structural model in which \(V\) has positive effects on both \(M\) and \(Y\), in which direction is \(\hat b\) biased? Explain why the DAG alone, without the linearity and sign restrictions, does not determine the direction.
  4. State two additional data structures or design features that would permit identification of the second-stage effect — for example: (i) measuring a covariate set sufficient for the \(M\)\(Y\) back-door paths, or (ii) an experiment that directly assigns the mediator. For each, say which assumption of Section 8.6 it restores.

6. Front-door identification. (Emphasis: the do-calculus derivation and diagnosing failure.) Consider the front-door graph.

  1. Verify that the three front-door conditions hold.
  2. Walk through the three-step proof of Theorem 8.3 for this graph: identify which do-calculus rule justifies each step.
  3. Add a direct edge \(T \to Y\) to the graph. Which front-door condition is violated? Does Equation 8.18 still hold?
  4. Explain how front-door Condition 3 corresponds to the Baron–Kenny requirement that the \(M \to Y\) relation be unconfounded after conditioning on \(T\) (equivalently, Assumption 2 of Section 8.6). The front-door graph does not dispense with this requirement — it satisfies it by construction of the graph. Why does the condition hold in the front-door graph but fail in the graph with an \(M\)\(Y\) confounder \(V\)?

7. Mediation vs. instrumental variables. A researcher studies the effect of a job training program (\(T\)) on wages (\(Y\)). She proposes two intermediate variables: (A) motivation (\(M_A\)), measured after the program starts; (B) a lottery that randomly selects applicants for admission (\(Z\)), measured before the program.

  1. For variable (A): draw the mediation DAG including \(M_A\), \(T\), \(Y\), and an unobserved ability variable \(U\). State the complete set of identification assumptions from Section 8.6 needed to identify the NIE through \(M_A\), and explain why randomization of \(T\) secures only Assumption 1.
  2. For variable (B): draw the IV DAG with \(Z\), \(T\), \(Y\), and \(U\). State the three IV assumptions. Explain why the exclusion restriction and the “mediator inclusion” of mediation analysis are mutually incompatible conditions for the same intermediate variable.
  3. The researcher argues that \(M_A\) and \(Z\) are both “intermediate” variables and that the analyses are interchangeable. Write a one-paragraph critique of this argument.
  4. Can the front-door formula be applied if motivation \(M_A\) fully mediates the effect of \(T\) on \(Y\) and \(U\) does not directly affect \(M_A\)? State the three conditions and assess whether they hold.

8. Front-door identification with baseline covariates. (Second pass; builds on Problem 6.) Consider the DAG: \(T \to M \to Y\), \(\mathbf{X} \to T\), \(\mathbf{X} \to M\), \(\mathbf{X} \to Y\), \(U \to T\), \(U \to Y\), with \(\mathbf{X}\) observed and pre-treatment, \(U\) unobserved, and no direct \(T \to Y\) edge.

  1. Show that the three front-door conditions hold conditionally on \(\mathbf{X}\): all directed paths from \(T\) to \(Y\) pass through \(M\); every back-door path from \(T\) to \(M\) is blocked given \(\mathbf{X}\) (enumerate the paths through \(\mathbf{X}\) and through \(U\), and say which node blocks each); and \((T, \mathbf{X})\) blocks every back-door path from \(M\) to \(Y\).
  2. Derive the covariate-adjusted front-door formula \[P\!\left(y \mid \doop(T{=}t)\right) = \sum_{\mathbf{x}} P(\mathbf{x}) \sum_m P(m \mid t, \mathbf{x}) \sum_{t'} P(y \mid m, t', \mathbf{x})\, P(t' \mid \mathbf{x})\] by repeating the three-step do-calculus argument of Theorem 8.3 with \(W = \{\mathbf{X}\}\) carried through each rule application. (You will also need \(P(\mathbf{x} \mid \doop(t)) = P(\mathbf{x})\); which rule justifies it, and why does it require \(\mathbf{X}\) to be pre-treatment?)
  3. Explain why the inner average uses the covariate-conditional distribution \(P(t' \mid \mathbf{x})\) rather than the marginal \(P(t')\) used in Equation 8.18.
  4. Now add the edge \(U \to M\). Which conditional front-door condition fails, and which stage of the two-stage argument collapses?

9. Nonidentifiability by construction (advanced / second pass). This problem practices the technique of the bow-graph example in Chapter 3: prove nonidentifiability by exhibiting two observationally equivalent models with different causal effects.

Consider the graph from Case (a) of Example 8.1: \(T \to M \to Y\), \(U \to T\), \(U \to M\), \(U \to Y\) (front-door Conditions 2 and 3 are violated). Let \(T, M, Y, U \in \{0,1\}\) with \(U \sim \mathrm{Bern}(1/2)\).

  1. Construct two SEMs \(\mathcal{M}_1\) and \(\mathcal{M}_2\), each compatible with the graph, that induce the same observed distribution of \((T, M, Y)\) but different interventional distributions: specifically, \(P_{\mathcal{M}_1}(Y{=}1 \mid \doop(T{=}1)) \ne P_{\mathcal{M}_2}(Y{=}1 \mid \doop(T{=}1))\). (Hint: set \(T = U\) in both models. In \(\mathcal{M}_1\) set \(M = T \wedge U\) and \(Y = M \vee U\); in \(\mathcal{M}_2\) set \(M = T \vee U\) and \(Y = M \vee U\), where \(\wedge\) and \(\vee\) denote Boolean AND and OR. Observationally \(T = M = Y = U\) in both models, so both produce \(P(T{=}0, M{=}0, Y{=}0) = P(T{=}1, M{=}1, Y{=}1) = \tfrac{1}{2}\). Then compute \(P(Y{=}1 \mid \doop(T{=}1))\) and \(P(Y{=}1 \mid \doop(T{=}0))\) in each model, and take their difference.)
  2. Explain precisely which identification strategy fails in this graph and why. Your explanation should reference (i) the back-door criterion (is there any observed variable that blocks \(T \leftarrow U \to Y\)?), (ii) the front-door criterion (which conditions fail and which stages collapse), and (iii) the completeness theorem (what does the existence of your two models imply about any identification method?).
Baron, Reuben M., and David A. Kenny. 1986. “The Moderator–Mediator Variable Distinction in Social Psychological Research: Conceptual, Strategic, and Statistical Considerations.” Journal of Personality and Social Psychology 51 (6): 1173–82.
Frangakis, Constantine E., and Donald B. Rubin. 2002. “Principal Stratification in Causal Inference.” Biometrics 58 (1): 21–29.
Imai, Kosuke, Luke Keele, and Teppei Yamamoto. 2010. “Identification, Inference and Sensitivity Analysis for Causal Mediation Effects.” Statistical Science 25 (1): 51–71.
Pearl, Judea. 1995. “Causal Diagrams for Empirical Research.” Biometrika 82 (4): 669–88.
Pearl, Judea. 2001. “Direct and Indirect Effects.” Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (UAI 2001), 411–20.
Richardson, Thomas S., and James M. Robins. 2014. ACE Bounds; Single World Intervention Graphs (SWIGs) and Identification of Causal Effects. University of Washington.
Robins, James M. 1986. “A New Approach to Causal Inference in Mortality Studies with a Sustained Exposure Period—Application to Control of the Healthy Worker Survivor Effect.” Mathematical Modelling 7 (9–12): 1393–512.
Sobel, Michael E. 1982. “Asymptotic Confidence Intervals for Indirect Effects in Structural Equation Models.” Sociological Methodology 13: 290–312.
Stensrud, Mats J., James M. Robins, Torben Martinussen, and Jessica G. Young. 2023. “Generalized Separable Effects for Mediation Analysis with a Mediator–Outcome Confounder.” Journal of the American Statistical Association 118 (543): 1736–48.
Stensrud, Mats J., Jessica G. Young, Vanessa Didelez, James M. Robins, and Miguel A. Hernán. 2022. “Separable Effects for Causal Inference in the Presence of Competing Events.” Journal of the American Statistical Association 117 (537): 175–83.
VanderWeele, Tyler J., Stijn Vansteelandt, and James M. Robins. 2014. “Effect Decomposition in the Presence of an Exposure-Induced Mediator-Outcome Confounder.” Epidemiology 25 (2): 300–306.
Vansteelandt, Stijn, and Rhian M. Daniel. 2017. “Interventional Effects for Mediation Analysis with Multiple Mediators.” Epidemiology 28 (2): 258–65.