What is Causal Inference?
Causal Inference is the process of defining and estimating how an outcome would change under a specified intervention or treatment, using experimental design or explicit assumptions that connect counterfactual quantities to observed data.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Design the target experiment before fitting a model
Write the hypothetical experiment that would answer the question: eligibility, assignment, treatment strategies, outcome, follow-up, censoring, and analysis population. This prevents vague claims such as “the effect of engagement” when exposure versions and timing differ. Hernán and Robins' Causal Inference: What If organizes observational analyses around explicit causal questions and target-trial thinking.
Identification is not an algorithm setting
Observed data identify a causal estimand only under defensible design and structural assumptions. Randomization supports exchangeability by design; observational studies usually need measured pre-treatment covariates sufficient to block confounding paths, plus consistency and positivity. A flexible learner can reduce functional-form error, but it cannot discover missing counterfactuals, repair impossible interventions, or prove that no confounder was omitted.
Validate the design, estimator, and decision separately
Check treatment and outcome timestamps, assignment integrity, overlap, covariate balance, missingness, interference, attrition, and estimator sensitivity. Compare multiple justified estimators and negative or placebo controls where available, then report uncertainty and transportability limits. Refutation tests can reveal contradictions; passing them does not prove the causal assumptions true or guarantee that a future rollout matches the study population.
Key Characteristics
- Targets intervention contrasts rather than predictive association alone
- Requires an explicit population, treatment, comparator, outcome, and horizon
- Separates estimand definition, identification, estimation, and validation
- Uses Potential Outcomes, causal graphs, or compatible structural frameworks
- Depends on design assumptions that cannot be certified by fit metrics alone
- Reports uncertainty, sensitivity, interference, and transportability limits
Common Use Cases
- Estimating whether a product intervention changes retention
- Evaluating a treatment when randomized assignment is unavailable
- Separating recommendation-policy effects from user-selection effects
- Auditing which pre-treatment variables belong in an adjustment set
- Planning a target trial before analyzing historical operational data
Example
Loading code...Frequently Asked Questions
How is Causal Inference different from prediction?
Prediction estimates outcomes under the observed data-generating process and may rely on any stable association. Causal Inference estimates contrasts under interventions. A highly accurate predictor can use variables that are useless or harmful for deciding what action will change an outcome.
Does Causal Inference require a randomized experiment?
Randomization is often the strongest design for identifying an effect, but observational designs can support causal estimates under explicit assumptions and appropriate methods. Those assumptions are not granted by the algorithm and require domain, timing, and design evidence.
What is the difference between identification and estimation?
Identification establishes that a causal estimand can be written as a unique function of the observed-data distribution under stated assumptions. Estimation uses finite data to approximate that function. Better estimation cannot rescue an estimand that is not identified.
Can a causal graph prove that a relationship is causal?
No. A graph makes causal assumptions explicit and can reveal valid adjustment sets or contradictions implied by those assumptions. Its arrows come from design and domain knowledge; fitting a graph to observational data alone does not certify the causal structure.
What should be reviewed before acting on a causal estimate?
Review the target population and estimand, treatment and outcome timing, assignment mechanism, consistency, interference, exchangeability, positivity, missing data, model and fold design, uncertainty, sensitivity analyses, and whether the deployment population matches the evidence.