The 64 structures shown below represent the complete space of possible causal relationships among three variables under ideal conditions: (1) proper measurement — X, M, and Y are each measured without discriminant validity failures or systematic error (Module 1), and (2) no omitted variables — these three variables are the only players in the system (Module 2). In real research, measurement error and hidden confounders expand this space considerably.
Causal inference is not induction — you cannot prove a causal structure from data. It is deduction by elimination: you bring every tool in your methodological toolkit to bear and see which structures survive. The exercise below makes this concrete.
The grid shows all 64 structurally distinct relationships that can exist among three variables X, M, and Y — arranged as a triangle with X at the lower-left, Y at the lower-right, and M at the top. Each cell shows one possible causal structure. Arrows between any two variables can point forward (A → B), backward (A ← B), bidirectionally (A ↔︎ B, representing correlated unmeasured causes or feedback), or not exist at all.
Click the buttons below to apply research findings. Each finding rules out the structures inconsistent with it — those cells will be crossed out in red. After exhausting your toolkit, notice how many structures remain. That irreducible remainder is why causal inference is hard.
After applying all three tools together — regression, experimental manipulation of X, and the indirect-pattern finding — three causal structures survive, not one:
| Structure | Interpretation |
|---|---|
| X → M → Y | Classic mediation: X causes M, M causes Y |
| X → M ← Y | Y causes M (reverse causation on the M–Y link; M is a collider) |
| X → M ↔︎ Y | M and Y are reciprocally linked or share an unmeasured common cause |
All three are consistent with every piece of evidence you can collect in this system without additional assumptions. Causal inference is deduction by elimination — each tool removes some structures — but the elimination rarely converges to one uniquely identified story. Ruling out the two alternatives requires additional design choices (e.g., time-ordering evidence that Y cannot precede M, an instrument for M, or a manipulation of M) that go beyond regression and a single experiment.
The exercise above assumed no hidden variables. Now introduce U — an unobserved variable that may have arrows to X, M, and/or Y but is never measured. No research tool can constrain U directly.
Each of the three U-pairs (U–X, U–M, U–Y) can be absent, forward, backward, or bidirectional — 4³ = 64 U configurations for every one of the 64 X–M–Y structures, giving 4,096 total states. Rather than draw all 4,096 cells, the grid below preserves the familiar 64 X–M–Y layout. Each cell’s blue shading encodes how many of the 64 U configurations remain consistent with your findings: deeper blue = U is highly ambiguous for that structure; pale = almost no U configuration survives; crossed out = the X–M–Y structure is ruled out for every possible U configuration. The small dashed circle labelled U in each cell is a reminder that the confounder is always present but unobserved.
Notice in particular: pressing Manipulate X eliminates U→X (since X is now exogenous), but U’s connections to M and Y remain entirely free — experimental control of X does not eliminate unmeasured confounding of the M–Y path.
Every participant in the AlterEco eco-label study had two potential outcomes:
Their individual treatment effect is \(\tau_i = Y_i(1) - Y_i(0)\).
The problem? We can only observe one of these for any individual. This is called the Fundamental Problem of Causal Inference (Holland, 1986). Everything that follows is a strategy for getting around it.
set.seed(2024)
N <- 2000
env <- rnorm(N, 0, 1)
price_sens <- rnorm(N, 0, 1)
income <- rnorm(N, 0, 1)
Y0 <- 5.50 + 0.80*env - 0.50*price_sens + 0.30*income + rnorm(N, 0, 1.10)
Y0 <- pmax(1, pmin(10, Y0))
ITE <- 0.60 + 0.80*env - 0.55*price_sens + rnorm(N, 0, 0.40)
Y1 <- pmax(1, pmin(10, Y0 + ITE))
ATE_true <- mean(Y1 - Y0)
cat(sprintf("True ATE = %.3f\nTrue ITE SD = %.3f\n", ATE_true, sd(Y1 - Y0)))True ATE = 0.586
True ITE SD = 0.967| Person | Env concern | Y(0) No label | Y(1) With label | True ITE | Assigned to | What we observe |
|---|---|---|---|---|---|---|
| Sophie (high concern) | 1.3 | 8.79 | 10.00 | 1.21 | Control | Y(0)=$8.79 Y(1)=??? |
| Rens (moderate+) | 1.2 | 6.42 | 6.57 | 0.15 | Treatment | Y(0)=??? Y(1)=$6.57 |
| Lotte (neutral) | -0.1 | 5.16 | 5.87 | 0.70 | Control | Y(0)=$5.16 Y(1)=??? |
| Daan (low concern) | -1.2 | 6.73 | 6.62 | -0.11 | Treatment | Y(0)=??? Y(1)=$6.62 |
| Kim (price-sensitive) | -1.7 | 5.12 | 5.46 | 0.34 | Control | Y(0)=$5.12 Y(1)=??? |
A counterfactual is the answer to a “what if” question about the road not taken. For Sophie, who saw the eco-label and paid $8.20, the counterfactual is: what would she have paid if she had not seen it? That number — say, $6.80 — never happened in our world, but it is what the word “caused” points at. The eco-label caused a $1.40 increase in Sophie’s WTP precisely because $8.20 (observed) minus $6.80 (counterfactual) equals $1.40.
This might sound abstract, but counterfactuals are built into everyday causal language:
The table above makes the problem concrete: we assigned Sophie to the control condition and observed her pay $6.80, so her counterfactual Y(1) — what she would have paid with the label — is invisible. We assigned Daan to the treatment and observed Y(1), so his Y(0) is invisible. Both missing cells are counterfactuals. The fundamental problem of causal inference is that every individual appears in exactly one condition, so every individual’s causal effect requires reasoning about a world that did not occur.
Before running any analysis, try writing out the counterfactual claim your paper makes:
“If participant \(i\) had been assigned to condition A instead of condition B, their outcome would have been ___ rather than ___.”
If you cannot fill in the blank — or if filling it in requires assuming things you have not measured — your causal claim is more fragile than it looks. This exercise is the fastest way to spot whether your design actually supports the inference you intend to draw.
The individual treatment effect for person \(i\) is:
\[\tau_i = Y_i(1) - Y_i(0)\]
This is the only truly assumption-free definition of causality: it is the effect on this person, in this moment, caused by this treatment. Did the eco-label change this consumer’s willingness to pay? That is the ITE.
The ITE is simultaneously the most interpretable estimand and the one we can never directly observe — because every person appears in only one condition at one point in time. The histogram below shows the true ITE distribution that exists in reality but is entirely invisible to any researcher. Notice how wide it is: some consumers gain over $2 in WTP from seeing the label; others gain nothing; a few are actually put off by it. This heterogeneity is the central fact that all other estimands must grapple with.
tibble(ITE=ITE,
env_grp=cut(env, c(-Inf,-0.5,0.5,Inf), labels=c("Low","Moderate","High"))) |>
ggplot(aes(x=ITE, fill=env_grp)) +
geom_histogram(bins=50, alpha=0.75, position="identity") +
geom_vline(xintercept=ATE_true, linetype="dashed", linewidth=1.1, colour="grey20") +
annotate("text", x=ATE_true+0.06, y=82,
label=sprintf("ATE = $%.2f", ATE_true), hjust=0, size=3.5, fontface="bold") +
scale_fill_manual(values=c("High"=clr_eco,"Moderate"="#74b9a0","Low"=clr_ctrl)) +
labs(x="Individual Treatment Effect Y(1)−Y(0) ($)", y="Count", fill=NULL,
title="The eco-label effect is wildly different across individuals",
subtitle="In real research you never see this distribution — only the ATE is estimable") +
theme_mod3()
\[\text{ATE} = E[\tau_i] = E[Y_i(1) - Y_i(0)]\]
The ATE is the expectation of the ITE across the entire target population — the answer to “what would happen on average if this treatment were applied to everyone?”
The assumption needed to use ATE as a proxy for ITE: Treatment effect homogeneity — the belief that every unit responds the same way (\(\tau_i = \tau\) for all \(i\)). If the eco-label added exactly the same WTP boost to every consumer, the ATE and every ITE would be identical, and knowing the ATE would fully characterise all individual responses.
In practice, homogeneity is almost never plausible. The ITE histogram above makes this plain: effects range from negative to strongly positive. This means the ATE answers a different question than the ITE. The ATE is the right estimand for population-level policy decisions — “should supermarkets mandate eco-labels on all products?” — because its answer applies in aggregate even when individual effects vary wildly. But it is a poor guide to predicting any individual consumer’s response, because it averages over enormous heterogeneity.
Why the ATE is still the gold standard for experiments: In a randomised trial, random assignment makes the treated and control groups exchangeable in expectation, so the observed mean difference is an unbiased estimate of the ATE without assuming homogeneity. The experiment recovers the ATE honestly; it is the subsequent inference from ATE to individual effect that requires the homogeneity assumption.
That said, the exchangeability assumption is more fragile than it appears — and more fragile than is often acknowledged. Random assignment guarantees balance in expectation across infinite replications, but any single study can produce unlucky covariate imbalance. More importantly, the behavioural conditions for exchangeability extend beyond the draw of assignment: if participants in the treatment condition behave differently because they know they are treated — through demand characteristics, Hawthorne effects, or awareness of experimental purpose — or if the sample differs systematically from the target population, then the observed mean difference estimates something other than the population ATE. Module 2 explored these threats in detail: the design features that protect exchangeability (pre-registration, cover stories, unobtrusive measures, stimulus sampling, representative recruitment) are precisely what make an ATE estimate credible rather than merely randomised.
treat_rct <- rbinom(N, 1, 0.5)
Y_obs_rct <- ifelse(treat_rct==1, Y1, Y0)
df_rct <- tibble(id=1:N, treat=treat_rct, Y_obs=Y_obs_rct, Y0, Y1, ITE, env, price_sens, income)
ATE_est <- mean(Y_obs_rct[treat_rct==1]) - mean(Y_obs_rct[treat_rct==0])
lm(Y_obs ~ treat, data=df_rct) |> tidy() |>
filter(term=="treat") |>
transmute(Estimand="ATE", `True value`=round(ATE_true,3), Estimate=round(estimate,3),
SE=round(std.error,3),
`95% CI`=sprintf("[%.3f, %.3f]", estimate-1.96*std.error, estimate+1.96*std.error)) |>
knitr::kable(caption="ATE estimate from randomised eco-label experiment")| Estimand | True value | Estimate | SE | 95% CI |
|---|---|---|---|---|
| ATE | 0.586 | 0.72 | 0.08 | [0.563, 0.878] |
\[\text{CATE}(x) = E[\tau_i \mid X_i = x]\]
The CATE partitions the population by pre-treatment characteristics \(X\) and estimates a separate average treatment effect within each subgroup. The eco-label may add $1.40 WTP for consumers with high environmental concern but only $0.15 for low-concern consumers — the ATE buries this story.
What CATE relaxes: The global homogeneity assumption. CATE allows different types of consumers to respond differently; it only requires homogeneity within each subgroup (everyone with the same covariate profile \(x\) is assumed to respond identically to their group’s average).
What CATE still assumes — and why the gap to ITE remains:
You have measured the right moderators — and measured them well. You can only condition on variables you have collected. If the strongest driver of heterogeneity is something unmeasured — say, a consumer’s specific shopping occasion or the brand’s existing equity — your CATEs are averages over that unobserved heterogeneity and may not correspond to any real subgroup’s true effect. But even when you have collected the right moderator, measurement quality matters enormously: a construct that suffers from discriminant validity failures, cross-group non-invariance, or systematic response bias (see Module 1) produces distorted CATE estimates even if the causal heterogeneity is real. A CATE model built on “environmental concern” items that actually blend concern with general pro-social identity will attribute the wrong heterogeneity to the wrong mechanism. The entire CATE machinery is predicated on the measurement infrastructure being sound.
Correct functional form when extrapolating. Estimating CATE for a continuous moderator (e.g., environmental concern score) requires assuming a relationship between the moderator and the treatment effect — linearity, correct interactions, no omitted non-linearities. Even a well-estimated linear CATE model will give wrong predictions for individuals whose true effect relationship is non-linear.
Within-cell homogeneity. Even if you carve out fine-grained cells, predicting an individual’s response from their cell’s CATE still assumes every person in that cell responds identically to the cell average. As cells get finer this becomes a stronger claim about fewer people, eventually reducing to the same problem as the ITE itself.
The plot below shows CATE estimates versus the (unobservable) true CATEs across four environmental concern bins. The estimates track the truth well here — but each bar still summarises a distribution of individual effects within the bin.
df_rct |>
mutate(env_grp=cut(env, c(-Inf,-1,0,1,Inf), labels=c("Very Low","Low","High","Very High"))) |>
group_by(env_grp) |>
summarise(CATE_true=mean(Y1-Y0), CATE_est=mean(Y_obs[treat==1])-mean(Y_obs[treat==0]),
n=n(), .groups="drop") |>
pivot_longer(starts_with("CATE"), names_to="type", values_to="val") |>
ggplot(aes(x=env_grp, y=val, fill=type)) +
geom_col(position="dodge", alpha=0.85, width=0.65) +
geom_hline(yintercept=ATE_true, linetype="dashed", colour="grey30") +
annotate("text", x=0.65, y=ATE_true+0.05, label="Overall ATE", size=3.5) +
scale_fill_manual(values=c(CATE_true=clr_eco, CATE_est=clr_ctrl),
labels=c("True CATE","Estimated CATE")) +
labs(x="Environmental Concern", y="Eco-label effect ($)", fill=NULL,
title="CATE: the label works far better for environmentally engaged consumers",
subtitle="Reporting only the ATE hides this story completely") +
theme_mod3()
When the treatment of interest cannot be directly assigned — only encouraged — the population stratifies into four compliance types based on how each unit would respond to assignment under both conditions:
| Type | \(D_i(Z{=}0)\) | \(D_i(Z{=}1)\) | Description |
|---|---|---|---|
| Complier | 0 | 1 | Takes treatment only when assigned |
| Always-taker | 1 | 1 | Takes treatment regardless |
| Never-taker | 0 | 0 | Never takes treatment |
| Defier | 1 | 0 | Takes treatment only when not assigned |
In the eco-label context: a store may be assigned to display eco-labels (\(Z=1\)), but individual consumers may or may not actually look at them (\(D=1\)). Consumers who would read any label they see are always-takers; those who read the label only because the store displays it prominently are compliers; those who never engage with shelf labels are never-takers.
Each estimand below targets a different slice of the ITE distribution — and each translation back to the ITE requires different assumptions:
Intent-to-treat (ITT): The effect of assignment to the treatment condition regardless of actual receipt. In an imperfect-compliance world, ITT underestimates the ATE by a factor equal to the compliance rate, because some assigned units are not actually treated. ITT is the most conservative and assumption-light estimand here, but it answers the wrong question if you care about the treatment’s biological or psychological mechanism rather than the logistical reality of assignment.
Local Average Treatment Effect (LATE): The ITT scaled by the first-stage compliance rate (the Wald estimator). This recovers the average ITE for compliers only — those who take the treatment because they were assigned to, not because they always would have. The LATE is the right estimand when you care about the policy lever of assignment, and it rests on two additional assumptions beyond randomisation:
Even with these assumptions satisfied, LATE still describes an average over compliers’ ITEs, requiring within-complier homogeneity to predict any individual complier’s response.
ATT and ATC: The average treatment effect for units that did (ATT) or did not (ATC) actually receive the treatment. Because treatment receipt may be self-selected — consumers who seek out eco-labels are likely those who care most about sustainability and therefore respond most strongly — ATT \(\neq\) ATE \(\neq\) ATC. Translating from ATT to ITE for a specific treated unit requires assuming that unit is representative of the treated group, which is a homogeneity claim within the treated subpopulation.
Every estimand is a different answer to a different version of the causal question. The further you move from the population-wide ATE toward more targeted estimands, the more specific the claim — but also the more assumptions required to bridge back to the individual level:
| Estimand | Population | Additional assumption(s) to reach ITE |
|---|---|---|
| ATE | All units | Treatment effect homogeneity |
| CATE | Subgroup \(X = x\) | Correct moderator specification; within-cell homogeneity |
| LATE | Compliers | Monotonicity; exclusion restriction; within-complier homogeneity |
| ATT | Treated units | Representativeness of the treated for the target individual |
| ATC | Control units | Representativeness of controls for the target individual |
None of these estimands is the ITE. All of them are averages over a group of individuals. The ITE is the only true unit-level causal effect, and it is the only one that requires no assumptions about other units — but it is also the only one we can never observe. Applied causal inference is the art of choosing the estimand whose assumptions are most defensible given your design and question, while being explicit that a gap always remains between your estimate and the individual-level effect that treatment ultimately acts upon.
set.seed(2024)
Z <- treat_rct
p_comply_if_Z1 <- plogis(0.8 + 0.6*env)
p_always_taker <- plogis(-2.5 + 0.5*env)
D_if_Z1 <- rbinom(N, 1, p_comply_if_Z1)
D_if_Z0 <- rbinom(N, 1, p_always_taker)
# Enforce monotonicity: no defiers (D_i(1) >= D_i(0) for all i)
D_if_Z0 <- pmin(D_if_Z0, D_if_Z1)
D <- ifelse(Z==1, D_if_Z1, D_if_Z0)
compliance_type <- case_when(
D_if_Z1==1 & D_if_Z0==0 ~ "Complier",
D_if_Z1==1 & D_if_Z0==1 ~ "Always-taker",
D_if_Z1==0 & D_if_Z0==0 ~ "Never-taker",
D_if_Z1==0 & D_if_Z0==1 ~ "Defier"
)
knitr::kable(table(compliance_type), col.names=c("Compliance type","Count"),
caption="True compliance breakdown (observable only in simulation)")| Compliance type | Count |
|---|---|
| Always-taker | 133 |
| Complier | 1244 |
| Never-taker | 623 |
Y_obs_iv <- ifelse(D==1, Y1, Y0)
df_iv <- tibble(Z, D, Y_obs_iv, Y0, Y1, ITE, compliance_type, env, price_sens)
ITT <- mean(Y_obs_iv[Z==1]) - mean(Y_obs_iv[Z==0])
FS <- mean(D[Z==1]) - mean(D[Z==0])
LATE_est <- ITT / FS
LATE_true <- mean(ITE[compliance_type=="Complier"])
ATT_est <- mean(ITE[D==1])
ATC_est <- mean(ITE[D==0])
tibble(
Estimand = c("ATE","ITT","LATE (Wald)","ATT","ATC"),
Description = c(
"Effect averaged over all 2,000 consumers",
"Effect of *assigning* the label regardless of compliance",
"Effect for consumers who complied *because* they were assigned",
"Effect for consumers who *actually saw* the label",
"Effect for those who *did not see* it"
),
`True value` = round(c(ATE_true, ITT, LATE_true, ATT_est, ATC_est), 3),
Estimated = round(c(ATE_true, ITT, LATE_est, ATT_est, ATC_est), 3)
) |> knitr::kable(caption="Five ways to summarise 'the effect' of the eco-label")| Estimand | Description | True value | Estimated |
|---|---|---|---|
| ATE | Effect averaged over all 2,000 consumers | 0.586 | 0.586 |
| ITT | Effect of assigning the label regardless of compliance | 0.540 | 0.540 |
| LATE (Wald) | Effect for consumers who complied because they were assigned | 0.716 | 0.887 |
| ATT | Effect for consumers who actually saw the label | 0.788 | 0.788 |
| ATC | Effect for those who did not see it | 0.507 | 0.507 |
Potential outcomes give you a precise language for what you want to estimate. But they are largely silent on how to estimate it — which variables to control for, which to leave alone, and why. Directed acyclic graphs (DAGs), developed by Judea Pearl and colleagues, fill that gap. They provide a visual language for encoding your causal assumptions and a set of rules for reading off, directly from the picture, which adjustment strategies identify causal effects and which produce bias.
Before the formalism, an analogy that keeps the concepts concrete.
Imagine you are standing in front of a sealed wall. On the left side is an input valve, \(X\). On the right side is an output spout, \(Y\). You can see the openings and gauges on the outside, but the plumbing inside the wall is hidden.
A DAG is your proposed blueprint of the pipes behind the wall. Every arrow is a claim about how water can flow. Every missing arrow is equally a claim: it says there is no direct pipe between those two openings.
The causal question is: “If I turn the \(X\) valve myself, how much does the \(Y\) spout change?” That is different from asking: “When the \(X\) gauge reads high, is the \(Y\) gauge also high?” The second is observational — you are watching natural variation. The first is interventional — you are doing something. The pipe analogy will make each path structure, and the distinction between them, easy to see.
A DAG has two ingredients:
The “acyclic” part means no variable can cause itself through any chain of arrows — time must flow forward. You can represent feedback by treating each time-point as a separate node (\(X_t \rightarrow Y_t \rightarrow X_{t+1}\)), but within a cross-sectional DAG, cycles are forbidden.
Every arrow you draw is an assumption. Every missing arrow is also an assumption — it claims the two variables are not directly linked. A DAG made honest is a complete inventory of what you believe about the causal structure of your system.
Three building blocks cover nearly every situation you will encounter:
Chain (mediator): \(X \rightarrow M \rightarrow Y\). \(X\) causes \(M\), \(M\) causes \(Y\). In this simple structure — where there is no direct \(X \rightarrow Y\) path and no confounding — the entire association between \(X\) and \(Y\) runs through \(M\). If you block \(M\) by controlling for it, you sever the only route and the \(X\)–\(Y\) association disappears. In the pipe analogy: the pipe from \(X\) to \(Y\) passes through a visible middle gauge \(M\). Clamping \(M\) shut stops the flow — which is useful if you are asking whether anything bypasses \(M\), but wrong if you just want to know the total effect of opening \(X\).
Fork (confounder): \(X \leftarrow U \rightarrow Y\). A common cause \(U\) produces an association between \(X\) and \(Y\) even if \(X\) does not cause \(Y\) at all. Controlling for \(U\) blocks this spurious path. In the pipe analogy: there is a hidden water source \(U\) feeding both the \(X\) side and the \(Y\) side. The \(X\) and \(Y\) gauges move together not because \(X\) feeds \(Y\) but because both are fed from the same upstream source.
Collider: \(X \rightarrow M \leftarrow Y\). Both \(X\) and \(Y\) cause \(M\), but the arrows flow into \(M\), so the path is blocked — \(X\) and \(Y\) are not associated in the general population. Conditioning on \(M\) (controlling for it, stratifying on it, or selecting a sample where \(M\) is fixed) opens a spurious association that did not previously exist. In the pipe analogy: \(M\) is a warning light triggered by either \(X\) pressure or \(Y\) pressure. In the general system \(X\) and \(Y\) are unrelated. But if you look only at cases where the warning light is on, you create a false connection — low \(X\) pressure implies \(Y\) must have been high enough to trigger it, and vice versa.
A path transmits association between two variables unless it contains a collider or you condition on a non-collider node on the path. Conditioning on a collider opens a path that was closed.
This has an immediate practical implication: not every control variable improves your estimate. If you control for a collider, you introduce bias. Choosing covariates by “putting in everything correlated with Y” is precisely the wrong strategy.
Suppose you study whether conscientiousness (\(X\)) causes job performance (\(Y\)). You restrict your sample to employees who were hired — conditioning on \(M = \text{hired}\). The structure is:
\[\text{Conscientiousness} \rightarrow \text{Hired} \leftarrow \text{Prior performance signals}\]
and prior performance signals also predict later job performance (\(Y\)). Hiring is a collider: both conscientiousness and prior performance independently raise the probability of being hired. In the full applicant pool, conscientiousness and prior performance signals are unrelated. But inside the hired group, they are negatively correlated: a candidate with low conscientiousness who was still hired must have had strong prior performance to compensate, and vice versa. Conditioning on the collider opens the path Conscientiousness \(\leftrightarrow\) Prior signals \(\rightarrow\) Performance, creating a spurious negative association between conscientiousness and performance within your sample. This is why studies restricted to elite samples — top MBA programmes, Fortune 500 firms, clinical trial completers — routinely find attenuated or reversed relationships compared to population-level estimates.
Observational data tell you about the world as it is — how variables co-occur. Causal questions ask about the world as it would be if you intervened. Pearl’s do-operator makes this distinction precise.
These are equal only when \(X\) has no upstream common cause with \(Y\). Suppose environmental values (\(U\)) cause both label exposure (\(X\)) and WTP (\(Y\)). Then \(P(Y \mid X = \text{high})\) partly reflects the kind of person who ends up in high-\(X\) situations — one with high environmental values, and therefore high WTP regardless. \(P(Y \mid \mathit{do}(X = \text{high}))\) removes that selection: you have forced \(X\) to be high, so \(U\) no longer determines who gets which value of \(X\), and the observed \(Y\) difference is purely causal.
In the pipe analogy: observing high \(X\) pressure may just tell you the hidden upstream source \(U\) is running strong. Setting \(X\) yourself — attaching your own controlled pump — cuts \(X\) free from whatever was feeding it naturally. That is why randomisation identifies causal effects: it is a physical implementation of \(\mathit{do}(X)\).
Random assignment is the cleanest implementation of \(\mathit{do}(X = x)\): by forcing \(X\) to the value you choose, you sever all incoming arrows to \(X\). The resulting \(P(Y \mid \mathit{do}(X))\) is identified without further adjustment. When you cannot randomise, identification requires finding an adjustment set that mimics what randomisation would have done.
A backdoor path from \(X\) to \(Y\) is any path that starts by going backwards along an incoming arrow into \(X\). Backdoor paths carry confounding association — they are routes through which \(X\) and \(Y\) move together for reasons that have nothing to do with \(X\) causing \(Y\).
In the pipe analogy: a backdoor path is a hidden pipe that enters \(X\) from behind (upstream contamination) and then reaches \(Y\). The water appears to flow from \(X\) to \(Y\) but is really entering through a hidden joint behind the \(X\) valve. To estimate the true effect of turning the \(X\) valve, you need to close every backdoor route — without closing pipes that are downstream consequences of \(X\).
The backdoor criterion says: a set of variables \(Z\) is sufficient to identify the causal effect of \(X\) on \(Y\) if \(Z\) (1) blocks all backdoor paths from \(X\) to \(Y\), and (2) does not include any descendant of \(X\) — no post-treatment variables or mediators, since those lie downstream and blocking them would remove genuine causal flow.
In the eco-label study, suppose consumers who chose to shop at eco-conscious stores (\(X\) = saw label) differ systematically from other shoppers in their underlying environmental values (\(U\)). \(U\) causes both store choice (and thus label exposure) and WTP (\(Y\)). The path \(X \leftarrow U \rightarrow Y\) is a backdoor path. If we could measure and control for \(U\), we would block this path and identify the causal effect. Randomising store assignment — so \(X\) is no longer caused by \(U\) — severs the backdoor path entirely.
When a valid backdoor adjustment set exists, the causal effect is:
\[P(Y \mid \mathit{do}(X)) = \sum_z P(Y \mid X, Z=z)\, P(Z=z)\]
This is just a weighted average of the \(X \rightarrow Y\) association within strata of \(Z\) — familiar as regression adjustment or stratification.
The mediator structure (\(X \rightarrow M \rightarrow Y\)) creates a choice: do you control for \(M\) or not? The answer depends entirely on your research question — not on convention or habit.
If you want the total effect of \(X\) on \(Y\), do not control for \(M\). Controlling for \(M\) blocks the indirect path \(X \rightarrow M \rightarrow Y\) and returns only whatever direct \(X \rightarrow Y\) path exists outside \(M\). In the pipe analogy: if you clamp the middle gauge shut, you are no longer measuring the total flow caused by opening \(X\) — you have blocked the route through which most of \(X\)’s effect travels.
If you want the direct effect of \(X\) on \(Y\) holding \(M\) fixed, you may control for \(M\) — but only if the \(M \rightarrow Y\) relationship is itself unconfounded. If there is an unmeasured variable that affects both \(M\) and \(Y\), then conditioning on \(M\) no longer identifies the direct effect: the mediator–outcome confounding is left unblocked, and you may introduce additional bias rather than remove it. The requirement is not just that \(M\) is measured, but that you can credibly claim the \(M\)–\(Y\) relationship is clean.
In the eco-label context: suppose the label works by (a) directly signalling product quality and (b) activating environmental guilt (\(M\)), which then raises WTP. If you want to know “does the label raise WTP overall?”, do not control for guilt — you would delete the mechanism. If you want to know “does the label raise WTP beyond the guilt pathway?”, you may control for guilt — but only after establishing that no unmeasured variable (say, general pro-social identity) jointly drives both guilt and WTP.
Sometimes the backdoor path cannot be blocked — \(U\) is unobserved and there is no valid instrument. But if a mediator \(M\) satisfies three conditions, the causal effect is still identifiable through the frontdoor criterion:
The classic example: smoking (\(X\)) causes lung cancer (\(Y\)) through tar deposits (\(M\)). Genetic predisposition (\(U\)) confounds the \(X\)–\(Y\) relationship — heavier smokers may also carry genetic risk factors that independently raise cancer risk. But the \(X \rightarrow M\) path is clean: whether someone smokes determines tar accumulation, and there is no direct route for \(U\) to affect tar independently of smoking. And the backdoor paths from \(M\) to \(Y\) all pass through \(X\), which is observed and can be adjusted for. So we can (1) estimate the effect of smoking on tar using variation in smoking status, then (2) estimate the effect of tar on cancer adjusting for smoking status to close the backdoor, and finally (3) combine these two stages to recover the total causal effect of smoking on cancer — even though \(U\) is never measured.
In the pipe analogy: the back of the wall is inaccessible (you cannot measure \(U\)), but all the water from \(X\) to \(Y\) must pass through a visible meter \(M\). You can read the \(X \rightarrow M\) flow directly (the \(X\)–\(M\) segment is clean), and you can read the \(M \rightarrow Y\) flow after accounting for the \(X\) value (which blocks the only remaining backdoor into \(M\)). Combining both readings reconstructs the total flow from \(X\) to \(Y\) without ever seeing the hidden upstream source.
In management research, the frontdoor is relevant when you have a clear causal chain — policy \(\rightarrow\) behaviour \(\rightarrow\) outcome — and you can measure the behaviour even though the policy’s take-up is confounded by unobservables. It is less commonly exploited than instrumental variables but rests on weaker assumptions when the full mediating structure is genuinely present and measurable.
You do not need to run Pearl’s full do-calculus algebra to benefit from DAGs. Drawing the DAG before your analysis forces three productive habits:
Modules 3.2–3.5 will apply this framework to regression discontinuity, difference-in-differences, and mediation analysis, where the DAG determines whether each design’s identifying assumption is plausible.