Causal Inference
Start with prediction versus intervention
Read an introduction to potential outcomes and try its examples to see why the alternative outcome matters.
Start with: Causal Inference
RedLectern topic
Causal AI
Causal questions · interventions · counterfactuals · assumptions behind causal learning.
For: Analysts · researchers.
After this page you can: Separate prediction from a defensible causal claim.
Page reviewed
Part of: Artificial Intelligence › Machine Learning
About Causal AI
A predictive model estimates outcomes among cases with observed features. A causal question asks what would happen under an action or condition, such as changing a policy or offering a program. Those are different targets: an association in observed data does not by itself show what an intervention would change. A useful starting habit is to state the action, outcome, and population before choosing an estimator.
Causal inference combines a well-defined question with assumptions about how the data arose. Randomized experiments can support causal comparisons, while observational analyses rely on assumptions about selection, confounding, and measurement. Potential outcomes make the missing comparison explicit: for each unit, only the outcome under the condition that occurred is observed. A defensible analysis therefore explains its target effect, the assumptions needed to identify it, how it estimates that effect, and what checks probe the result. Causal machine learning uses flexible learning methods within this reasoning; it does not make the assumptions disappear.
Start here
Effect inference
Follow an explicit causal workflow
See how causal assumptions, effect identification, estimation, and refutation fit together.
Start with: Effect inference
Learn the foundations
1.What makes a claim causal?Prediction describes outcomes among cases with observed features; intervention asks about outcomes if an action were set.
Prediction describes outcomes among cases with observed features; intervention asks about outcomes if an action were set. They can differ when factors influence both the proposed cause and the outcome. For example, a common cause can create an association even when changing the proposed cause would not change the outcome. A causal inference is therefore an estimate of an intervention effect under assumptions, not a stronger-sounding label for a predictive association. Start by naming the cause, outcome, population, and comparison that the question requires.
- Effect inference
Summarizes a workflow that makes the causal question and target effect explicit before estimation and refutation.
Causal inference
Estimating effects of interventions under explicit assumptions and evidence.
Confounder
A factor influencing both a proposed cause and its measured outcome.
Counterfactual
An alternative outcome considered under a different action or condition.
Try it
A model finds that students who join tutoring have lower average grades than those who do not. Give one possible reason this association may not be the effect of tutoring.
Prior grades could influence both tutoring enrollment and later grades. If students with lower prior grades are more likely to join tutoring, the observed difference need not be tutoring’s effect.
2.Ask what changes under an interventionAn intervention question compares outcomes under specified actions, rather than merely comparing people who happened to receive different treatments.
An intervention question compares outcomes under specified actions, rather than merely comparing people who happened to receive different treatments. Define the treatment or action, outcome, population, and time frame before selecting a method. In the potential-outcomes view, each unit has an outcome under each possible treatment condition, but only the outcome for the condition that occurred is observed. This missing comparison is why a simple treated-versus-untreated difference can mix an intervention effect with pre-existing differences. Practice by stating the counterfactual comparison your question needs.
- Causal Analysis: Impact Evaluation and Causal Machine Learning with Applications in R
A deeper treatment of causal analysis methods, including causal machine learning and empirical applications.
- Effect inference
Introduces modeling a causal problem, identifying a target estimand, estimating the effect, and refuting the estimate.
Try it
A team wants to know whether offering a study app improves exam scores. Write the intervention and comparison in one sentence.
Estimate the difference in exam scores for the target students if they are offered the study app versus if those same students are not offered it, over the specified exam period.
3.Make assumptions visibleA causal estimate depends on assumptions connecting observed data to the intervention effect of interest.
A causal estimate depends on assumptions connecting observed data to the intervention effect of interest. A causal graph can record beliefs about how variables influence one another and help explain which adjustment choices are being made. For observational data, assumptions such as having measured the relevant confounders may not be testable from the data alone. Estimation is only one part of the analysis: define the target effect, explain why the available evidence identifies it under the assumptions, and use sensitivity or robustness checks to probe how conclusions respond when assumptions are challenged. These checks inform confidence; they do not prove every assumption true.
- Modeling Causal Relations
Explains how a causal graph makes assumptions explicit and how domain knowledge can inform the graph.
Try it
A study estimates the effect of a job-training program from observational data. Name one assumption to state and one check that could probe sensitivity to it.
State whether measured variables capture the factors that influence both enrollment and later employment. Sensitivity analysis can examine how robust the estimate is to a violation such as unobserved confounding.
Latest
No recent updates are available.
Books
- Causality: Models, Reasoning, and Inference
Causality: Models, Reasoning, and Inference
Judea Pearl · 2009
A formal account connecting causal graphs to intervention effects and counterfactual reasoning, with attention to the assumptions behind identification.
Causal Inference: What If
Miguel A. Hernán and James M. Robins · 2020
A practical framework for defining causal effects and estimating them from observational or experimental data. The authors make the full text freely available online.
Causal Inference: The Mixtape
Scott Cunningham · 2021
A worked introduction to potential outcomes and applied causal designs, with practical coding examples. A free online version is available from the author.