Whether the coefficient is close to the true causal effect or wildly different depends entirely on which variables you included, omitted, and conditioned on incorrectly.
A pricing team at a B2B SaaS company wants to know: if we raise the mid-tier plan from ninety-nine to one-twenty-nine dollars per month, what happens to revenue? The analyst fits a standard regression of renewal probability on price, controlling for everything she can think of—customer characteristics, engagement metrics, legacy plan status. She presents the coefficient to the CFO, who uses it to forecast the revenue impact. This is the standard workflow. In most cases, it is fundamentally wrong. That regression coefficient is not the causal effect of price on renewal. It is a statistical association, conditional on the other variables in the model, contaminated by whatever confounding the regression failed to address. Whether it is close to the true causal effect or wildly different depends on which variables were included, which were omitted, and which are being conditioned on incorrectly. The difference between association and causation is not academic—it shapes every forecast and every decision that follows.
The CFO's forecast was accordingly an underestimate—or overestimate, depending on what was omitted. The analyst gave the wrong number, and the business made the wrong decision based on it.
In the pricing example, the analyst controlled for engagement metrics—login frequency, feature adoption—along with customer characteristics and legacy plan status. Engagement is downstream of price: customers who pay higher prices might engage less or more, and that engagement then influences renewal. By conditioning on engagement in the regression, the analyst partially closed the very causal pathway she was trying to measure. The result is a biased estimate, pushed toward zero. The CFO's forecast underestimated the true causal effect. The price increase might have had a larger negative impact on renewals than the regression suggested, or a smaller impact, or even positive. We do not know. But the analyst's number was wrong. This is not a problem with sample size or statistical power. It is a fundamental mismatch between what the regression estimates—a partially closed effect—and what the decision-maker needs—the full causal effect of price on renewal.
The backdoor criterion answers one question: given a causal diagram and a target effect, which variables must you adjust for to estimate that effect from observational data?
The backdoor criterion is a graphical method for answering a precise question: given a causal diagram—a picture of which variables influence which—and a target causal query—what is the effect of X on Y?—which set of variables, if any, is sufficient to adjust for in order to estimate that effect? The procedure has three steps. First, identify all backdoor paths from X to Y. A backdoor path is any path from X to Y that starts with an arrow pointing into X. These are the paths along which confounding—spurious correlation—can flow. Second, find a set of variables Z that blocks all the backdoor paths. Every backdoor path must contain at least one variable from Z, and that variable must not be a collider on the path or a descendant of a collider. Call this Z a deconfounding set. Third, Z must not contain any descendants of X, because conditioning on a descendant of X partially or completely closes the causal pathway you are trying to estimate. If such a set exists, the causal effect can be estimated by adjusting for Z.
Customer characteristics—industry, company size—influence both the price offered by sales and the customer's natural renewal probability. Engagement sits downstream as a mediator, not a confounder. Adjust for characteristics; leave engagement alone.
Apply the backdoor criterion to the pricing problem. The causal diagram has this structure: customer characteristics—industry, company size—influence both the price the customer is offered, because the sales team gives discounts to certain industries and sizes, and the customer's renewal probability, because some industries are stickier than others. Engagement metrics are downstream of price: they are mediators between price and renewal. The diagram has exactly one backdoor path: price receives an arrow from customer characteristics, and customer characteristics have a path to renewal. This is the confounding. This path can be blocked by adjusting for customer characteristics—controlling for industry and company size in the regression. Engagement should not be adjusted for. Engagement is a mediator on the causal pathway from price to renewal. Adjusting for it would partially close the very effect we are trying to estimate. This diagnosis tells the analyst exactly what to do: regress renewal on price, controlling for customer characteristics but not engagement. The result is the causal effect of price on renewal.
The choice of which variables to include is made on grounds that have nothing to do with causal structure. None of these intuitive criteria reliably produces a deconfounding set.
Standard regression is fundamentally cause-blind. The coefficient on a predictor depends on what other predictors are in the model—but in typical practice, the choice of which variables to include is made on grounds that have nothing to do with causal structure. Analysts select variables based on availability: what data do we have? Statistical significance: which variables are correlated with the outcome? Intuition: what feels relevant? Or precedent: what did other studies control for? None of these criteria reliably produces a deconfounding set. None of them account for colliders or mediators. None of them distinguish between variables that block confounding and variables that close causal pathways. This is why the standard workflow fails. The regression that controls for engagement, customer characteristics, legacy plan status, and engagement metrics is built on statistical intuition, not causal logic. It happens to be wrong because engagement is a mediator. If the analyst had included a different variable—one that is also a mediator, or a collider—the estimate would be wrong in a different way. The specific direction of the bias depends on the causal structure, not on how the variables were selected.
The analyst wants the total effect of price—the full impact operating through all channels. Controlling for engagement strips out exactly that indirect path, producing an underestimate biased toward zero.
The first common error is controlling for a mediator. In the pricing problem, engagement is downstream of price and upstream of renewal. The analyst controls for engagement in the regression, thinking it improves the model by explaining variation in the outcome. In fact, it blocks part of the causal pathway. Price affects engagement; engagement affects renewal. When you condition on engagement, you hold it constant across different price levels. You are no longer estimating the full effect of price on renewal—you are estimating the effect of price on renewal, net of the engagement channel. This is sometimes called the direct effect, but it is not what the CFO asked for. The CFO wants the total effect: if we raise the price, what happens to renewal, considering everything—including the fact that higher prices might reduce engagement, which then reduces renewal further? Controlling for engagement gives the wrong answer. It removes exactly the indirect path that constitutes part of the true causal effect. The result is an estimate biased toward zero, underestimating the magnitude of the true effect.
The direction and magnitude of the bias depend on the strength of the confounder's effects on X and Y, and the direction of correlation with the outcome.
The second common error is failing to control for a confounder. Suppose the analyst omits customer characteristics from the regression—maybe because the data is unavailable, or because she did not realize it was relevant. Customer characteristics affect both price—sales gives discounts to certain industries—and renewal—some industries churn more. This is a confounder, and it creates a backdoor path from price to renewal. By omitting customer characteristics, the analyst leaves this path open. The regression coefficient on price now reflects both the causal effect of price on renewal and the spurious correlation flowing through the confounder. If customers in stickier industries tend to pay higher prices, the coefficient will be biased upward—the true effect of price looks smaller than it really is, or even positive when the true effect is negative. If customers in stickier industries tend to pay lower prices, the coefficient will be biased downward. The direction and magnitude of the bias are determined by the causal structure, not by sample size or model fit. This is why confounders matter so much in observational data: they are often invisible in the regression output.
The backdoor criterion is powerful, but not universal. When no observable variables can block all confounding, other methods become necessary.
The backdoor criterion offers a complete answer to one question: given a causal diagram, which set of variables, if any, can identify the causal effect by adjustment? Sometimes the answer is none. There may be a backdoor path that cannot be blocked by any observable variables—either because the confounder is unobserved, or because conditioning on it would require conditioning on a collider. In these cases, the effect cannot be estimated from observational data using the backdoor criterion alone. Other approaches become necessary. The front-door criterion is a graphical method that sometimes works when the backdoor criterion fails—it exploits the structure of mediators to identify effects despite unobserved confounding. Instrumental variables exploit external randomness that influences X but not Y, allowing the analyst to estimate the effect of X on Y without conditioning on confounders. A randomized experiment directly breaks all confounding by randomly assigning the treatment X, removing the need for adjustment entirely. Each method works in specific causal structures. The backdoor criterion tells you which one to use.
These methods augment the backdoor criterion for complex, high-dimensional settings where linear regression is too restrictive, handling the difficult cases the graphical criterion identifies.
The backdoor criterion identifies which variables to adjust for. But it does not specify how to adjust—whether to use linear regression, logistic regression, or something else. In low-dimensional settings, standard regression works. But when the number of confounders is large, or the relationships between variables are nonlinear, standard regression can fail in new ways. Two modern methods address these challenges while preserving the causal logic of the backdoor criterion. Double machine learning uses machine learning to estimate the relationships between X and confounders, and between Y and confounders, then combines these estimates to recover an unbiased estimate of the causal effect. It sidesteps the problem of overfitting to confounders while maintaining the causal interpretation. Causal forests extend decision trees to estimate treatment effects that vary across units—how the effect of price on renewal depends on customer characteristics, for example. Instead of a single coefficient, you get a heterogeneous effect: the causal effect is different for different customers. Both methods are built on the backdoor criterion. They simply allow the adjustment process to be more flexible than linear regression alone.
Whether a regression coefficient is the causal effect depends entirely on whether the variables in the model are the right variables—a question that cannot be answered by looking at data alone. It requires explicit reasoning about causal structure.
The arc from pricing problem to modern methods reveals a fundamental principle. Standard regression assumes that if you include enough variables and the model fits the data well, the coefficients are causal. This assumption is false. Whether a regression coefficient is the causal effect depends entirely on whether the variables in the model are the right variables—a question that cannot be answered by looking at data alone. It requires explicit reasoning about the causal structure. The backdoor criterion makes this explicit. You must draw a causal diagram. You must identify which variables are confounders, which are mediators, and which are irrelevant. Only then can you decide what to adjust for. The diagram is not derived from data. It is a prior claim about how the world works. If the diagram is wrong, the estimate is wrong, no matter how much data you have or how sophisticated your regression method is. Modern methods like double machine learning and causal forests extend the logic to high-dimensional and nonlinear settings, but they cannot overcome a misspecified diagram. The essential requirement remains: you must know the causal structure to estimate causal effects.
Regression is a tool for estimation, but it derives its causal meaning from the graph. Without a causal diagram that correctly describes the data-generating process, standard regression is cause-blind, and its coefficients are merely statistical associations contaminated by confounding.
The central claim threading through this chapter is this: the causal effect of X on Y is not determined by what you regress on. It is determined by the causal structure—the graph. The backdoor criterion translates the graph into a recipe: adjust for the deconfounding set, and you identify the causal effect. Omit a confounder, and you leave confounding in place. Include a mediator, and you partially close the causal pathway. Neither error is corrected by more data, better model fit, or higher statistical power. The error is causal, not statistical. This principle applies to every observational study that aims to estimate causal effects: pricing decisions, marketing attribution, treatment effects in medicine, policy impacts in economics. The standard workflow—gather data, fit a regression, report the coefficient—is widespread because it is simple. But it is frequently wrong. The correct workflow starts upstream, with a causal diagram that makes your assumptions explicit. Only then can you apply the backdoor criterion, choose the right set of variables, and estimate the effect reliably. The diagram is not guaranteed to be correct. It depends on domain knowledge and subject-matter expertise. But making it explicit is the first step toward getting the causal effect right.
Living Models · Chapter 8 · Estimating Effects
Chapter Eight establishes the bridge from causal theory to practical estimation. The pricing example shows why standard regression fails when applied blindly. The backdoor criterion provides a graphical method to identify which variables to adjust for and which to leave alone. Two common errors—controlling for mediators and omitting confounders—arise when causal structure is ignored. Modern methods extend the logic to complex, high-dimensional settings. But the fundamental principle remains: causation is read from the graph, not from the fit of the model. Every observational analysis that claims to estimate a causal effect must begin with explicit reasoning about the causal structure. Only then can regression, machine learning, or any other estimation method deliver a reliable answer. In the next chapter, we move forward to causal discovery: how to learn the causal structure from data when it is not known in advance.