Front-Door Adjustment for Causal Mediation

The Hidden Confounder and the New Clue

Let's update our case file. The causal diagram we used previously showed Package Type as a measurable confounder. Now, it's an unobserved variable, which we'll label $U$. This variable creates a spurious backdoor path between our treatment $X$ (Studio Choice) and our outcome $Y$ (Satisfaction).

However, your investigation has turned up a new piece of evidence: data on whether a client received a "Personalized Style Consultation" ($Z$). Velvet & Vellum offered these consultations more frequently as part of their standard service, while Satin & Steel reserved them for specific, un-documented circumstances. This consultation is a mediating variable; it lies on the causal pathway between the studio choice and the client's final satisfaction.

This gives us the following causal structure:

The path $X\to Z\to Y$ is the "front door" — the direct causal mechanism we want to measure. The path $X\leftarrow U\to Y$ is the un-blockable "back door." The front-door adjustment allows us to isolate the flow of causality through the front door, even as spurious association leaks through the back.

The Front-Door Criterion: Conditions for Entry

To use this method, the mediating variable $Z$ must satisfy three strict conditions, known as the Front-Door Criterion.

A variable $Z$ (or set of variables) satisfies the front-door criterion relative to $X$ and $Y$ if:

1. $Z$ intercepts all directed paths from $X$ to $Y$. This means the entire causal effect of $X$ on $Y$ is mediated through $Z$. In our case, the studio's influence on satisfaction works entirely through the quality and nature of its services, which we are summarizing with the "Personalized Consultation" variable.
2. There is no unblocked backdoor path from $X$ to $Z$. The confounder $U$ must not directly cause $Z$. Our unobserved confounder (client's budget/enthusiasm) affects their studio choice and overall satisfaction, but we assume it doesn't independently make them eligible for a consultation except via their choice of studio.
3. All backdoor paths from $Z$ to $Y$ are blocked by $X$. This is the crucial step. Any confounding path between the mediator $Z$ and the outcome $Y$ must pass through $X$. In our diagram, the path $Z\leftarrow X\leftarrow U\to Y$ is a backdoor path from $Z$ to $Y$, and it is blocked if we condition on $X$.

The following video provides an excellent conceptual breakdown of these conditions and the overall logic of the adjustment.

Deconstructing the Front-Door Formula

With the conditions met, we can deploy the formula. It looks complex, but it's a logical extension of what you already know. The intuition is that we are chaining together two separate, identifiable causal effects.

The front-door adjustment formula is:

Let's break this down into two parts:

1. Effect of $X$ on $Z$: The first part, $P(Z=z|X=x)$, measures the effect of the treatment on the mediator. Because Condition 2 states there is no confounding between $X$ and $Z$, the causal effect is simply the observed conditional probability: $P(Z=z|do(X=x))=P(Z=z|X=x)$.

2. Effect of $Z$ on $Y$: The second part, ${\sum }_{x^{\prime }}P(Y=y|X=x^{\prime },Z=z)P(X=x^{\prime })$, is the causal effect of the mediator on the outcome. Notice the backdoor path from $Z$ to $Y$ is $Z\leftarrow X\leftarrow U\to Y$. To block this path, we must adjust for $X$. This expression is the backdoor adjustment formula for the effect of $Z$ on $Y$, where we adjust for $X$!

So, the front-door adjustment cleverly identifies the total effect by first finding the effect of the studio choice on getting a consultation, and then combining it with the (adjusted) effect of the consultation on satisfaction.

This blog post provides a superb intuitive proof of how the formula is derived by chaining two backdoor adjustments.

Solving the Case with the Front-Door Adjustment

Your team has managed to salvage the following data from the studio's records, involving 1000 clients. The "Package Type" data is gone, but the "Personalized Consultation" data is intact.

Let's calculate the true causal effect of choosing Velvet & Vellum ($X=V\&V$) on satisfaction ($Y=1$). The following article walks through an almost identical calculation by hand. We will follow its structure to solve our case.

Our goal is to compute P(Y=1 | do(X=\text{V&V})). The formula is:

Where $Z=1$ means "Yes" for Consultation, and $Z=0$ means "No".

Step 1: Calculate all necessary probabilities from the table.

• Marginal Probabilities of $X$:

  • $P(X=\text{V\&V})=(100+330)/1000=430/1000=0.43$
  • $P(X=\text{S\&S})=(480+90)/1000=570/1000=0.57$

• Conditional Probabilities $P(Z|X)$:

  • $P(Z=1|X=\text{V\&V})=330/430\approx 0.767$
  • $P(Z=0|X=\text{V\&V})=100/430\approx 0.233$

• Conditional Probabilities $P(Y|X,Z)$:

  • $P(Y=1|X=\text{V\&V},Z=1)=225/330\approx 0.682$
  • $P(Y=1|X=\text{V\&V},Z=0)=40/100=0.400$
  • $P(Y=1|X=\text{S\&S},Z=1)=57/90\approx 0.633$
  • $P(Y=1|X=\text{S\&S},Z=0)=252/480=0.525$

Step 2: Calculate the inner sum (the effect of $Z$ on $Y$).

This is the backdoor adjustment of $Z\to Y$ controlling for $X$.
Let's call this term $P(Y=1|do(Z=z))$.
$P(Y=1|do(Z=z))=P(Y=1|X=\text{V\&V},Z=z)P(X=\text{V\&V})+P(Y=1|X=\text{S\&S},Z=z)P(X=\text{S\&S})$

• For $Z=1$ (Consultation = Yes):
  $P(Y=1|do(Z=1))=(0.682\times 0.43)+(0.633\times 0.57)\approx 0.293+0.361=0.654$
• For $Z=0$ (Consultation = No):
  $P(Y=1|do(Z=0))=(0.400\times 0.43)+(0.525\times 0.57)\approx 0.172+0.299=0.471$

Step 3: Calculate the final outer sum.

This combines the effect of $X\to Z$ with the effect of $Z\to Y$.
$P(Y=1|do(X=\text{V\&V}))=P(Z=1|X=\text{V\&V})P(Y=1|do(Z=1))+P(Z=0|X=\text{V\&V})P(Y=1|do(Z=0))$

The true causal effect of choosing Velvet & Vellum is a 61.1% satisfaction rate.

You can repeat the calculation for Satin & Steel as an exercise. The result is $P(Y=1|do(X=\text{S\&S}))\approx 52.1\%$. Once again, even with the primary confounder hidden, our advanced techniques reveal that Velvet & Vellum is causally superior. Case closed.

Conclusion

You have now mastered one of the most elegant tools in the causal inference toolkit. While the backdoor adjustment is a frontal assault on confounding, the front-door adjustment is a subtle flanking maneuver. It proves that even with critical information missing, a causal effect can be identified by carefully tracing the mechanism through which it operates.

Key Takeaways:

• When a primary confounder $U$ is unobserved, backdoor adjustment is not possible.
• If a mediating variable $Z$ exists that satisfies the three conditions of the Front-Door Criterion, the causal effect is still identifiable.
• The Front-Door Adjustment Formula estimates the effect of $X$ on $Y$ by chaining two identifiable effects: the unconfounded effect of $X$ on $Z$, and the backdoor-adjusted effect of $Z$ on $Y$ (controlling for $X$).

Preview of the Next Module: The Coach's Miracle Cure

Your success in the Studio Paradox case has brought you a new client. A relationship coach is making waves with a bold claim: a 100% success rate in improving client satisfaction. This sounds too good to be true. In our next module, we will shift focus from comparing groups to scrutinizing the claims of a single entity. You will need new methods to determine if this coach has a "miracle cure" or if the numbers are merely a statistical illusion. Our first task will be to assess the likelihood of such perfect results occurring by chance alone.