
Key takeaways
A causal diagram (DAG) writes down what every regression silently assumes: which variable affects which. Missing arrows are the strong assumptions.
Every diagram is built from three structures: chains, common causes and colliders. Controlling for a collider creates a spurious association instead of removing one.
The back-door criterion turns this into a rule for choosing control variables; "add every variable you have" can make an estimate worse.
Patterns of independence in data can sometimes reveal the direction of an arrow, and algorithms such as PC automate this, but only under strong assumptions.
The greatest value of a diagram is the conversation it forces before modeling: what do we really believe causes what?
A pharmaceutical company wants to know how to split its marketing budget for a new antibiotic among channels: sales representatives visiting physicians, other contacts, advertising in medical journals, and free samples. The analysts have 71 months of data on new prescriptions and 11 marketing variables. The fastest route is a regression with new prescriptions as the outcome and all 11 variables as predictors. The result is puzzling: journal advertising has a negative coefficient, as if more advertising made physicians prescribe less, and four variables are not significant because the variables are highly correlated with one another. Haughton et al. (2026) remark that this approach "substitutes data for thought", so a hard-to-interpret result should come as no surprise.
What is missing is an explicit assumption about which variable affects which. A causal diagram, most commonly in the form of a directed acyclic graph (DAG), is the tool for putting that assumption on paper. This article introduces causal diagrams: how to read them, how to use them to choose control variables, when data can suggest the direction of an arrow, and how algorithms learn diagrams from data.
The article has eight parts. Section 1 explains what a causal diagram is. Section 2 presents the three building blocks of every diagram. Section 3 uses them to answer the most practical question: which variables should go into a model. Section 4 shows that data can sometimes speak to causal direction, and Section 5 introduces structure-learning algorithms. Section 6 returns to the pharmaceutical example. Section 7 proposes a process for using causal diagrams in a project, and Section 8 discusses limitations.
1. What is a causal diagram?
A causal diagram consists of nodes (variables) and arrows between them. An arrow from A to B means we assume A has a direct effect on B. "Directed" means each arrow points one way; "acyclic" means you cannot follow the arrows back to where you started, since no variable can be its own cause. To describe feedback such as price affecting sales and sales affecting next period's price, we split variables by time (price in month 1, sales in month 1, price in month 2) so the diagram stays acyclic.
The most important thing to understand about causal diagrams is that the absence of an arrow is the strong assumption. Drawing an arrow from A to B only says there may be an effect, large or small, for the data to decide. Drawing no arrow between A and B asserts that A has no direct effect on B, and that assertion shapes the entire analysis that follows.
Causal diagrams are closely associated with Judea Pearl, who received the 2011 Turing Award for foundational contributions to probabilistic and causal reasoning in artificial intelligence. Pearl also coined the term Bayesian network for this kind of graph; a Bayesian network becomes a causal network when its arrows are read as causal relations rather than merely statistical dependence. The Book of Why (Pearl & Mackenzie, 2018) is the most readable entry point for newcomers.
2. Three building blocks
Every causal diagram, however complex, is assembled from three basic three-variable structures. Understanding how information "flows" through them is enough to read most diagrams in practice. Figure 1 illustrates the three structures with business examples.

Chain (A → C → B): advertising increases website visits, and visits increase sales. Advertising and sales are correlated. But among days with the same number of visits, that is, holding (conditioning on) C fixed, advertising no longer relates to sales, because its whole effect runs through C. C is a mediator.
Common cause (A ← C → B): the holiday season makes the company run more promotions and also makes customers buy more. Promotions and sales are correlated even if promotions do nothing. Holding C fixed (comparing within the same season) removes the spurious association. C is a confounder.
Common effect (A → C ← B): skills and personal connections both help a candidate get hired. In the population, the two may be independent. But looking only at people who were hired, that is, holding C fixed, we see a negative association: hires with weaker skills tend to have better connections, and vice versa. C is a collider, and conditioning on it creates a spurious association rather than removing one.
Table 1 summarizes how the three structures behave. The first two behave alike statistically; the third behaves the opposite way, and that is the source of many mistakes.
3. Choosing control variables with a diagram
The most practical question a causal diagram answers is: when estimating the effect of X on Y with a regression, which variables should be included? The habit of "the more variables the better" fails because it does not distinguish the three structures. This section presents the rule for choosing variables and the variables that should be left out.
3.1. Back-door paths
Between X and Y there may be several paths in the diagram. The path that follows the arrows from X to Y is the causal path, the thing we want to measure. Paths that begin with an arrow into X are called back-door paths, because they sneak into X from behind through a common cause. A small bubble tea shop estimating price sensitivity has the back-door path price ← opening day → sales: the shop opened with low prices and heavy promotion, and both pushed sales up. If this path is left open, the price coefficient absorbs the effect of the opening.
Pearl's back-door criterion says that to estimate the effect of X on Y correctly, we must control for a set of variables that blocks every back-door path, and that set must contain no variable that is a consequence of X. A path is blocked if it passes through a controlled confounder or mediator, or through a collider that is not controlled (Pearl et al., 2016).
3.2. Good and bad controls
The back-door criterion yields practical rules about which variables to control. Cinelli et al. (2024) call this "a crash course in good and bad controls". Table 2 summarizes cases common in business analysis.
Table 2 explains a common phenomenon: adding variables to a model can make the coefficient of interest worse, not better. Controlling for "promoted" when estimating the effect of training creates a spurious association between training and productivity; controlling for "skills after the course" erases the very effect we want to measure.
3.3. One model, one question
A further consequence is that each regression model answers exactly one causal question. In the backpack campaign example of Haughton et al. (2026), gender affects both who the campaign reaches and the demand for backpacks. A model with both campaign and gender correctly estimates the campaign's effect, but the gender coefficient in the same model is only the direct effect of gender, excluding the part that runs through the campaign. Reading every coefficient in a regression table as a total effect is common enough that epidemiologists have named it the "Table 2 fallacy" (Westreich & Greenland, 2013). Haughton et al. (2026) put it simply: regression coefficients have no causal meaning without an explicitly stated causal structure.
4. Can data tell us the direction of an arrow?
So far the diagram has been drawn from domain knowledge. A natural question is whether data can help determine the direction of an arrow. The answer is yes, in some cases, and the reason lies in the collider structure.
4.1. A third variable reveals the direction
With only two correlated variables A and B, data alone cannot tell whether A causes B or B causes A. But with a third variable C, the pattern of independence among the three can rule some possibilities out. Haughton et al. (2026) give an example: thin people tend to drink more diet soda. Are they thin because they drink diet soda, or do they drink it because they are thin? Suppose we also have age. If age is related to diet soda consumption (younger people drink more) but not to thinness, then diet soda cannot be what makes people thin; if it did, age would be indirectly related to thinness through diet soda. This logic is exactly the behavior of the collider and chain structures in Table 1.
A real example cited in the book is a study by Bryant and colleagues of alcohol and road deaths in the United States from 1947 to 1993 (as cited in Haughton et al., 2026). Alcohol consumption per head is correlated with road fatalities (coefficient 0.341, significant). Does alcohol cause accidents, or do accidents drive people to drink? The authors also had average highway speed: speed is correlated with deaths (0.386, significant) but not with alcohol consumption (0.136, not significant). As Figure 2 shows, the only consistent way to arrange these three relationships is that alcohol and speed both lead to deaths, making deaths a collider, so the causal direction runs from alcohol to deaths.

4.2. Three conditions for trusting the result
This reasoning holds only when three conditions are met (Spirtes et al., 2000). The first is causal sufficiency: no omitted variable causes two of the variables in the model. Taken literally this can never be guaranteed, so in practice the analyst must argue convincingly that nothing important has been left out. The second is the causal Markov condition: each variable depends only on its direct causes (its "parents"), with no need to know about "grandparents" or "siblings". The third is faithfulness: two variables are dependent if and only if there is a path connecting them, so no two opposing effects cancel out so perfectly that the data look independent.
5. Learning diagrams from data
With dozens of variables, reasoning as in Section 4 by hand is impossible. Structure-learning algorithms do it systematically. This section introduces the idea behind the most popular algorithm, the available software and an example.
5.1. The PC algorithm
The PC algorithm (named after its authors Peter Spirtes and Clark Glymour) belongs to the family of constraint-based algorithms. Step one: connect every pair of variables with an undirected edge. Step two: remove edges between pairs that are independent, or independent conditional on some other set of variables, using correlation tests (continuous variables) or contingency-table tests (categorical variables). Step three: orient the remaining edges, starting with colliders. If X is linked to Y, Y to Z, but X and Z are independent, then X → Y ← Z. From the identified colliders, the algorithm infers the direction of further edges (Spirtes et al., 2000).
A second family is score-based algorithms: search the space of possible diagrams for the one that fits the data best under a criterion such as BIC. This is a hard optimization problem because the number of diagrams explodes with the number of variables. Different algorithms often produce different diagrams, and finding good algorithms remains an active research field (Haughton et al., 2026).
5.2. Software and a churn example
The main software packages are Tetrad (Carnegie Mellon University, free), GeNIe, the bnlearn package in R, and BayesiaLab (commercial). In Python there is the causal-learn library. Haughton et al. (2026) ran the PC algorithm on data from a wireless provider wanting to know what makes customers switch. The data included many perception variables (reliability, innovativeness, sense of community), demographic variables (income, household size) and service-quality variables. The Tetrad result was surprising: the likelihood of switching was directly influenced only by satisfaction, and no perception variable had an arrow straight to churn. Rerunning in GeNIe, after imposing obvious constraints such as "age may affect spending, but spending cannot change age", added one more cause: the likelihood of the household acquiring more phones.
The example shows two things. The result does not immediately point to a churn solution, but it is a useful reality check on the common belief that brand image directly retains customers. And feeding background knowledge into the algorithm (which variables are certainly exogenous, which arrows are impossible) makes the diagram more sensible and useful. BayesiaLab also implements Pearl's "do-operator", which simulates what happens when a company actively changes a variable, for example increasing spending on one marketing channel, as opposed to merely observing that variable increase.
6. Back to the pharmaceutical company
Return to the opening example. The dataset in the book was simulated from the correlation matrix and summary statistics of a real study of a marketing campaign for a new antibiotic, because the original data were not published (Haughton et al., 2026). The regression with all 11 variables produced a negative coefficient for journal advertising and many insignificant variables, partly because the marketing variables are very highly correlated: months with many rep visits also have many samples and high contact costs.
When a causal diagram was built from the data and from knowledge of the industry, the picture became clearer: journal advertising most likely followed rather than led the growth in new prescriptions. In other words, the company advertised more once the drug was already selling well, rather than advertising being the main driver of sales. Haughton et al. (2026) suggest this finding might lead the marketing department to rethink how it spends its budget. It is also an example of reverse causation: a regression coefficient cannot tell the direction of an arrow, but a diagram forces the analyst to ask about it.
7. A process for using causal diagrams in a project
Causal diagrams need no expensive software or complex algorithms to get started. Much of their value comes from drawing them together with people who understand the business. Table 3 proposes a six-step process for an analytics project with a causal question.
Step 2 in Table 3 is often the most valuable. When marketing, sales and data teams draw the diagram together, hidden disagreements about "what causes what" surface before anyone runs a model, and many later arguments about numbers are settled right there. The free online tool DAGitty lets you draw a diagram and automatically identifies the adjustment set. When the diagram is fully specified in advance, the estimation step can use a structural equation model.
8. Limitations
A causal diagram is not magic that turns observational data into an experiment. It only makes assumptions transparent, and conclusions are only as good as the assumptions. If an unmeasured confounder affects both X and Y, no choice of controls can rescue the estimate, and designs such as instrumental variables or regression discontinuity become necessary.
Structure-learning algorithms have their own limits: results depend on the algorithm, the test thresholds and how variables are discretized, and many edges cannot be oriented from data at all. Haughton et al. (2026) stress that getting diagrams to "speak" involves some art and cannot be done mechanically. They also note that causal diagrams are still rarely used among business analysts, perhaps because regression is so quick and easy to run. That is a pity, because with a moderate number of variables, diagrams are easy to draw and read and can prevent investment decisions based on a coefficient with the wrong sign.
Conclusion
A causal diagram writes down what every regression implicitly assumes: which variable affects which. Three building blocks (chain, common cause, common effect) are enough to see why controlling for a confounder is right while controlling for a mediator or collider is wrong. The back-door criterion turns that insight into a rule for choosing control variables, replacing the habit of throwing every available variable into the model.
In some cases, data can even suggest the direction of an arrow through patterns of independence, as in the example of alcohol, speed and road deaths; algorithms such as PC automate that reasoning for many variables. But diagrams learned from data need checking by people who know the business, and every conclusion still depends on the assumption that no important variable has been left out. The greatest value of a causal diagram may lie not in the algorithm but in the conversation it forces a team to have before running a model: what do we really believe causes what?
References
Cinelli, C., Forney, A., & Pearl, J. (2024). A crash course in good and bad controls. Sociological Methods & Research, 53(3), 1071–1104. https://doi.org/10.1177/00491241221099552
Haughton, D., Haughton, J., & Lo, V. S. Y. (2026). Cause and effect business analytics and data science. CRC Press. https://doi.org/10.1201/9780429172588
Pearl, J., Glymour, M., & Jewell, N. P. (2016). Causal inference in statistics: A primer. Wiley. https://www.wiley.com/en-us/Causal+Inference+in+Statistics:+A+Primer-p-9781119186847
Pearl, J., & Mackenzie, D. (2018). The book of why: The new science of cause and effect. Basic Books. https://www.hachettebookgroup.com/titles/judea-pearl/the-book-of-why/9781541608986/
Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, prediction, and search (2nd ed.). MIT Press. https://mitpress.mit.edu/9780262194402/causation-prediction-and-search/
Westreich, D., & Greenland, S. (2013). The Table 2 fallacy: Presenting and interpreting confounder and modifier coefficients. American Journal of Epidemiology, 177(4), 292–298. https://doi.org/10.1093/aje/kws412

