Prediction, Description, and Causation
The same dataset can support very different questions. A descriptive analysis summarizes what happened, a predictive analysis forecasts an unobserved outcome, and a causal analysis compares outcomes under interventions. A useful model for one job is not automatically valid for another.
Prerequisites
Review Correlation and Causation. That lesson explains association measures; this path starts where it stops, by defining intervention effects and the assumptions needed to estimate them.
Learning goals
- Classify a question as descriptive, predictive, or causal.
- Define a causal contrast before selecting a model.
- Attach population, intervention, outcome, assumptions, and limits to a causal conclusion.
Running scenario
A college offers Study Sprint, a six-week coaching program before a final exam. Let A = 1 mean assignment to program access at the start of the term, A = 0 mean usual resources, and Y mean the final score. We will reuse this scenario throughout the nine lessons.
Three questions sound similar but target different quantities:
| Goal | Question | Useful evidence | What the answer does not establish |
|---|---|---|---|
| Description | What was the mean score among last term's participants? | Accurate records for that group and period | What would have happened without the program |
| Prediction | Which current students are likely to score below 60? | Representative labeled data and out-of-sample validation | That changing a predictive feature changes the score |
| Causation | How would scores change if eligible students were assigned access rather than usual resources? | A defined intervention, comparison strategy, and identifying assumptions | Effects for other programs, populations, or versions |
Prior grades may predict final scores extremely well. Setting a student's prior grade is not the Study Sprint intervention, so predictive importance is not a causal effect. Conversely, random assignment can identify an average effect even when individual scores remain difficult to predict.
From question to estimand
Write the causal question before fitting a model:
Among students eligible at term start, what is the six-week average effect on final score of assigning Study Sprint access rather than usual resources?
This sentence fixes the population, treatment alternatives, outcome, follow-up, and summary. The corresponding average treatment effect will later be written as . An estimator—such as a randomized difference in means—is a procedure for learning about that estimand. The two are not synonyms.
Assumptions and limits
The causal wording is justified only if the treatment versions and outcome are well defined, assignment groups are comparable for the target contrast, every relevant type of student could receive either assignment, one student's assignment does not change another's score (or spillovers are modeled), and the variables are measured well enough for the intended interpretation. Later lessons unpack each condition.
Even a valid estimate is local to its population, implementation, comparison, and time. “Study Sprint works” is therefore too broad. “In this eligible population, assigning this version rather than usual resources changed the six-week mean score under the stated assumptions” is an interpretable conclusion.
Common mistakes
- Calling a regression coefficient causal because many covariates were included.
- Evaluating an effect estimator only by predictive accuracy.
- Treating a feature-importance score as the effect of intervening on that feature.
- Reporting a group mean or before-after change as the missing counterfactual.
- Asking whether a program “works” without defining for whom, compared with what, and when.
Practice
Classify each question as descriptive, predictive, or causal: “How many students enrolled?”, “Who will miss the final?”, and “Would sending reminders increase attendance?” Then specify the population, time, and outcome missing from each vague version.
Next steps
Counterfactual Thinking formalizes the comparison hidden inside a causal question.
Primary sources
Checked 2026-10-15: Shmueli, “To Explain or to Predict?” and Hernán and Robins, Causal Inference: What If.