Nothing So Practical

do-Calculus for Humans

Pearl’s structural approach to causal inference is built around DAGs (directed acyclic graphs). In many cases, applying this framework boils down to measuring and adjusting for confounding variables. A key insight is that when working with purely observational data, controlling for all confounders can be equivalent to assigning the treatment level, as we would in an experiment. This claim rests on a curious concept: the do-operator.

At first glance, $\text{do()}$ looks like an esoteric way of saying control for confounds. However, that interpretation only holds when key confounders are measured. When we can’t measure all confounds, the role of the do-operator becomes clear. In such cases, the rules of do-calculus provide a system for determining whether a causal effect is identifiable from data given a set of assumptions about the data generating process. If identification is possible, do-calculus produces a formula for estimating the treatment effect. It can also tell you when the data are not capable of recovering the target estimand.

The goal of this post is to demystify the do-operator using a concrete example to show what it does. We start with the simple case when all confounds are measured. Then we’ll look at the more complicated situation where a key confound goes unmeasured. We’ll show how the rules of do-calculus help us deal with this situation using the front-door formula.

Seeing vs. Doing

Suppose a university wants to know whether taking a prep course improves scores on the SAT. They’ve built a dataset on about 10,000 students to try to answer this question. Here is what the simulated data looks like.

Students who took a prep course scored 138 points higher on average. But students who take a prep course are different from those who don’t. They tend to come from households with higher incomes, are more academically motivated, and get better grades. The 138-point gap reflects the prep course plus all the other ways prep course takers were going to outscore the non-takers.

The DAG makes these confounding pathways visible.

The arrows encode how background factors drive both the decision to take a prep course and SAT performance itself.

The naive comparison between the two prep groups can be expressed in notation as:

$E[\text{SAT} \mid \text{Prep} = 1] - E[\text{SAT} \mid \text{Prep} = 0]$.

The comparison conditions on choosing to take a prep course. Because that choice was driven by motivation, income, and academic performance, those factors vary systematically between the two groups. The backdoor paths remain open, and the estimate is biased.

Answering the causal question requires a different condition:

$E[\text{SAT} \mid \text{do}(\text{Prep} = 1)] - E[\text{SAT} \mid \text{do}(\text{Prep} = 0)]$.

The $\text{do()}$ means we need to estimate what would happen if we assigned students to prep (or not), the way a randomized experiment would. Forcing prep to take the value of 1 or 0 neutralizes all the reasons students select into the course. The do-operator is how we figure out whether that counterfactual is hiding in the data.

To apply the $\text{do}$ operator, we perform graph surgery by deleting every arrow pointing into the SAT Prep Course node. In a randomized experiment, nothing about a student’s background determines whether they take prep. The incoming arrows represent those background causes, and deleting them is the graphical way of saying these variables are irrelevant.

In the updated graph, Prep Course sits alone with no parents. The confounders haven’t disappeared, but their influence on prep has. Any remaining association between Prep and SAT Score flows through the causal arrow. That’s exactly what the quantity $P(\text{SAT} \mid \text{do}(\text{Prep}))$ represents.

The Adjustment Formula

We didn’t run an experiment, so the surgery removing the paths is theoretical. But under the right conditions, we can use the data to estimate what an experiment would have shown. The most common approach is backdoor adjustment.

Instead of comparing all prep-takers to non-takers, compare them within groups that share the same background. Using income as an example, compute the effect of prep within each level (low, middle, high) and average those effects together, weighted by each group’s share of the sample. The confounding from income washes out.

Here’s what the adjustment looks like when we stratify by income alone.

The new estimate is 110 points. We’ve reduced the bias, but we’re still way off. Income alone doesn’t capture the confounding through motivation and GPA.

Epidemiologists will recognize the adjustment formula as stratification by another name. You can also accomplish backdoor adjustment by adding the confounders to a regression of Y on X.¹

With no controls, a regression of SAT on Prep gives us exactly the naive comparison:

Add income as a control and the coefficient on Prep drops:

Add GPA and it drops to 74. And if we could somehow measure motivation and include it too:

As we add controls, the coefficient for SAT prep approaches the true effect of 55 points. The last model is hypothetical, since motivation is unmeasured. That gap between 74 and 56 is the bias from the unmeasured confounder.

The adjustment formula, graph surgery, and regression are all doing the same thing. The graph surgery tells you what confounding to remove. The formula tells you how to accomplish that removal. Regression does the removing.

When the Backdoor Fails

This works cleanly when the confounders are measured. In the SAT example, income and GPA are in the dataset. Motivation isn’t, and that gap is exactly where backdoor adjustment comes up short. Because motivation affects both whether a student takes a prep course and how well they score, there’s a backdoor path we can’t block. This is where do-calculus becomes useful.

It turns out we also have data on how many hours each student studied per week. Prep courses increase study time, and study time improves scores. For the sake of simplicity, if we assume the entire causal effect of prep on SAT scores flows through hours studied — a very strong assumption — we have a front-door path.

The front-door criterion exploits the fact that Hours Studied sits between Prep and SAT. That position lets us split the problem into two pieces and chain them together.

Piece one: Prep → Hours Studied. Motivation makes some students more likely to take prep, and motivation also makes them study more. But in this DAG, motivation only reaches hours through a prep course, which means there’s no backdoor between prep and hours. This is one of three conditions the front-door criterion requires. Students who take the course study about 11 more hours on average. No adjustment needed.

Piece two: Hours Studied → SAT Score. This one is trickier. Motivation affects SAT scores directly, and it reaches hours studied through prep course completion. That opens a backdoor: Hours ← Prep ← Motivation → SAT. To close it, we need to control for Prep. Among students who took a prep course, variation in hours studied is no longer driven by motivation. Within each group, hours and scores tell a clean story.

The Front-Door in Practice

Here’s what those two pieces look like as regressions.

Step 1. Regress Hours Studied on Prep Course. No controls needed:

SAT Prep adds 11 hours of study.

Step 2. Regress SAT Score on Hours Studied, controlling for Prep to block the backdoor:

Each additional hour of studying adds 4.9 SAT points.

Notice the Prep coefficient is 84 points. That coefficient absorbs the association between Prep and SAT that runs through the backdoor (Prep ← Motivation → SAT) rather than through Hours. By soaking up that confounding, it frees the Hours coefficient to reflect only the causal effect of studying. The front-door formula encodes exactly this logic, expressed as nested summations instead of two fitted models.

Multiply the two coefficients and you have the front-door causal estimate: $11.0 \times 4.9 \approx 54$ points. Close to the true effect of 55, nowhere near the naive 138, and all without ever measuring motivation.²

This front-door estimate relies on the strong assumption that motivation affects study hours only through the decision to take prep. Motivated students are going to study more regardless of whether they take a prep course, which makes that assumption unlikely to hold. The point is to illustrate the method, not to defend this particular DAG. That said, the front-door adjustment is rarely used in practice because the required conditions are difficult to find in real-world data.

Time To Follow The Rules

Up to this point, we’ve taken the front-door formula as a given. Seeing how it is derived makes clear what do-calculus is actually doing. This is where the three rules come in. Each rule describes a condition under which we can replace a hypothetical “experiment” with a plain observation. That’s do-calculus in a nutshell: figure out how to turn “what if I had assigned treatment?” into “here’s what I measured.”³

Our goal is to estimate P(SAT | do(Prep)), the causal effect of taking a prep course on SAT. To do that, we need to eliminate do(Prep) and replace it with quantities observed in the data. We’ll start with Rule 1 but note that the derivation only requires Rules 2 and 3.

Rule 1 lets you drop an observation from a probability expression when it carries no information about the outcome (Y). If a variable has no open path to Y once the incoming arrows to a variable are removed, conditioning on it changes nothing, so it can come out. It’s the bookkeeping rule for trimming irrelevant terms. This doesn’t apply to the SAT example because even after cutting the incoming paths to Prep, the confounders still have direct paths to SAT score.

For the front-door adjustment scenario, the only route from Prep to SAT runs through Hours, so we need to expand through the mediator using the law of total probability.⁴

Now do(Prep) appears in two places, one for each piece we need to estimate.

Rule 2 lets you swap a do() intervention for an observation (or vice versa) when the two are equivalent in the modified graph. The condition is that all backdoor paths into the intervened-on node are closed, leaving only forward-flowing paths. When that holds, observing the variable gives the same answer as assigning it externally. Note that Rule 2 also runs in the other direction. If you have an observed variable but realize a backdoor is still open, you can insert a do() to represent the intervention needed to close it.

Piece one: does prep affect hours? Apply the do-operator to Prep and close the incoming paths. Now the only path from Prep to Hours runs forward through the causal arrow. Motivation used to flow into Prep, but that path is now cut. No backdoor survives. Seeing who took prep tells us the same thing as assigning prep ourselves. According to Rule 2, we can swap the Prep intervention for the version observed in the data.

Piece two: does hours affect SAT? We start with P(SAT | do(Prep), H=h): Prep is intervened on, Hours is observed. There’s a backdoor from Hours to SAT running through Prep and Motivation. We can’t simply observe hours and call it causal. We need to close the incoming path to Hours. Rule 2: insert do(H), converting the observation into an intervention:

With those paths closed, the influence of Hours only flows forward into SAT, so observing H gives the same result as intervening on H. Now both Prep and Hours are intervened on. Do we still need the do(Prep)? With Hours fixed, Prep has no remaining path to SAT. Here’s where Rule 3 comes in.

Rule 3 lets you drop a do() intervention entirely when the intervened-on variable has no remaining path to the outcome given the other terms in the probability expression. Just as Rule 1 removes an irrelevant observation, Rule 3 removes an irrelevant do(). In the SAT derivation, once both Prep and Hours are intervened on, we have P(SAT | do(Prep), do(H=h)). Prep’s only route to SAT ran through Hours, but Hours is now set. With that path blocked, do(Prep) can’t reach SAT through any channel. Rule 3 says drop do(Prep). We’re left with

One backdoor still runs from Hours back through Prep to Motivation to SAT, but Prep is observed, so we can adjust for it. Apply Rule 2 one more time: replace do(H=h) with an observation, controlling for Prep.

Every intervention is gone and we have the two required pieces, which are estimable from our data.

When Does It Matter

Most applied causal inference doesn’t require thinking about do-calculus. If you can measure the confounders, the backdoor criterion handles it. Draw a DAG, identify the adjustment set, and run the regression. The do-operator is just shorthand for those steps.

When a confounder is unmeasured, the rules of do-calculus show their value. The front-door criterion is one example, but the same framework covers mediation-based strategies and more exotic identification paths. Without do-calculus, each strategy requires its own insight.

The most valuable result isn’t the formula derived from the rules. It’s the negative case, when do-calculus tells you a causal effect cannot be identified from available data, given your assumptions. That stops us from running analyses that look rigorous but aren’t. Knowing you’re stuck is an important result. It tells you precisely what questions the data can’t answer. That’s more useful than a plausible-looking estimate built on unverifiable assumptions.

For most applied work, the do-operator really is just careful bookkeeping. But the three rules are what turn that bookkeeping into guarantees. If identification is possible from the data, the rules will find a formula. If it isn’t, they’ll tell you that too.

References

Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669-688.

Pearl, J. (2012). The do-calculus revisited. Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence (UAI), 3-11.

Glymour, M., Pearl, J., & Jewell, N. P. (2016). Causal Inference in Statistics: A Primer. Wiley.

Huang, Y., & Valtorta, M. (2006). Pearl’s calculus of intervention is complete. Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence (UAI), 217-224.

Shpitser, I., & Pearl, J. (2006). Identification of joint interventional distributions in recursive semi-Markovian causal models. Proceedings of the 21st National Conference on Artificial Intelligence (AAAI), 1219-1226.