Learn · Intermediate
Causal graphs: deciding what to adjust for before interpreting an AI result
A causal graph is a diagram of assumptions about which variables directly influence others. It helps researchers decide whether an observed association can identify the effect of an intervention and which variables should be adjusted for. Its value is preventing plausible statistical corrections from answering the wrong question or creating bias.
Imagine studying whether an AI tutor improves exam performance. Students choose whether to use it. Motivation influences both that choice and later exam scores. A graph draws arrows from motivation to use and from motivation to performance, alongside the hypothesized arrow from use to performance. The extra route through motivation mixes selection with the tutor’s effect.
Judea Pearl’s introduction to causal inference develops the distinction between observing a variable and intervening on it. Observing tutor users tells us how students who chose it performed. Intervening means changing access or use through a specified external action. The two distributions need not match because ordinary use carries information about the reasons students selected it.
The basic diagram is a directed acyclic graph: arrows point from causes to effects, and no directed path loops back into itself. Acyclic does not mean the real world lacks feedback. A feedback process can sometimes be represented by separate variables at successive times. The diagram must state an appropriate time scale instead of pretending that one timeless arrow captures every interaction.
A confounder is a common cause of treatment and outcome. Motivation in the tutor example creates a backdoor path from use to score: follow an arrow backward from use to motivation, then forward to performance. If motivation is adequately measured and the assumed graph has no other relevant open paths, adjustment can compare students within motivation levels before combining those comparisons.
The backdoor criterion formalizes a sufficient adjustment rule. For a treatment’s total effect, the adjustment variables must contain no descendants of treatment and must block every path between treatment and outcome that begins with an arrow into treatment. This is a sufficient criterion under the graph, not a guarantee that the graph is correct or that the data contain enough comparable students.
Adjustment is like comparing runners on similar courses rather than crediting their shoes for differences in the hills. But measuring more characteristics is not automatically safer. Consider completed practice exercises after tutor use. If the tutor increases practice and practice improves scores, practice is a mediator. Controlling for it removes an intended channel of the tutor’s benefit. The resulting question concerns a different effect from the total effect initially proposed.
A collider is more surprising. Suppose both a model’s true quality and a large marketing budget increase the chance of appearing in a curated product roundup. Roundup inclusion is their common effect. Restrict attention to included models and a low-budget model may need unusually high quality to appear, while a well-marketed one can appear with less. Selection can create an association between quality and budget even if neither causes the other.
Conditioning on a collider opens a path that was otherwise blocked. That is why evaluating only products selected for attention can distort comparisons. The problem is different from ordinary benchmark contamination: the test need not appear in training data for the evaluated sample to be selected in a causally misleading way.
The distinction also helps read biological claims. OpenAI’s mouse-brain atlas announcement discusses measured responses after gene perturbations and agent-generated explanations. Perturbing a gene provides experimental leverage, but it does not automatically prove every proposed downstream disease mechanism. Cell composition, measurement choices and intermediate pathways still determine which causal question an analysis can answer.
Likewise, Chapman’s survey reports lower fear among frequent AI users while warning that the association is not causal. Prior attitudes could affect use, and background characteristics could affect both. A graph makes these competing explanations explicit before a regression coefficient is interpreted as reassurance caused by experience.
Randomized trials address treatment assignment through design. Causal graphs address the assumptions connecting assignment, measurement, selection and outcomes, including observational settings. They complement experiments without replacing them. Preregistration can record the graph and target effect before results encourage convenient adjustment choices.
The practical sequence is to define the intervention and outcome, draw plausible causes, distinguish variables measured before and after treatment, and identify a justified adjustment set. Then examine hidden confounding, overlap between compared groups and measurement error. A graph earns its usefulness by exposing assumptions that could be wrong. It does not turn a correlational dataset into causal evidence by decoration.
Key questions
Why can adjusting for more variables make a causal estimate worse?
What does the backdoor criterion identify?
Does drawing a causal graph prove the arrows are correct?
Cite this
APA
Ground Truth. (2026, October 22). Causal graphs: deciding what to adjust for before interpreting an AI result. Ground Truth. https://groundtruth.day/learn/causal-graphs-backdoor-adjustment-and-colliders.html
BibTeX
@misc{groundtruth:causal-graphs-backdoor-adjustment-and-colliders,
title = {Causal graphs: deciding what to adjust for before interpreting an AI result},
author = {{Ground Truth}},
year = {2026},
month = {oct},
url = {https://groundtruth.day/learn/causal-graphs-backdoor-adjustment-and-colliders.html}
}