First, the key terms.
- Confounder
- A variable that helps create an association between an exposure and an outcome.
- Intervention
- A deliberate change whose effect is being studied.
The idea, explained
A variable can predict an outcome because of common causes, selection effects or reverse causality. Randomized experiments can help isolate causal effects when implemented correctly. Observational methods require assumptions about confounding, measurement and which variables should be controlled. A causal diagram makes some assumptions explicit; it does not prove them. Before using a prediction to justify action, ask whether changing that input would actually change the outcome.
State the causal question
Predicting exam results from study habits is different from estimating the effect of a new tutoring program. Specify the treatment, outcome, population and time period before selecting a method.
Explain why groups differ
People choosing tutoring may already be struggling or especially motivated. A comparison with nonparticipants mixes those differences with the program’s effect. Randomization or justified observational assumptions are needed to address the comparison.
Follow a worked example.
Students who buy more exam-preparation books may receive lower scores because struggling students seek more help. That association does not establish that buying books reduces performance. Prior preparation could influence both purchasing and the final result.
Walk it through, one step at a time
- Students who attend tutoring score lower on average than those who do not.
- Before claiming harm, check whether weaker initial performance made tutoring more likely. Baseline differences can explain the observed association.
- A well-designed randomized study, where appropriate, would compare outcomes under assignment rather than relying only on self-selection.
Where people get stuck
Controlling for every available variable can introduce bias, including when a variable is affected by the treatment.
Now try it yourself.
1. Check the core idea
Does a model’s strong predictive feature necessarily represent a useful intervention?
Reveal the answer & reasoning
No. It may be a proxy, a consequence of the outcome or a marker of another cause. Prediction and intervention answer different questions.
2. Apply what you learned
Would controlling for every recorded variable always remove causal bias?
Reveal the answer & reasoning
No. Some controls can introduce bias, and important common causes may be unmeasured. The choice of controls should follow the causal question and justified assumptions.