When can a study support a causal explanation?

For: Readers evaluating claims that one behaviour, exposure or intervention changes another outcome.

An association tells you that observations vary together. A causal question asks what would change under a specified intervention. Those questions can use the same variables while requiring different evidence. Begin with the scatterplot lecture: describe the pattern before supplying an explanation. Then ask what decision or action the causal claim is about. A slope is not yet an answer to that question.

In an original example, students who spend more time revising tend to obtain higher marks. More revision might help. Prior knowledge, motivation or available study time might also affect both variables. Exam difficulty could change the pattern across classes. Sketch these possibilities before choosing an analysis. A correlation is useful descriptive information, but selecting the explanation requires knowledge of how the data were generated. NPTEL scatterplots and correlation sequence.

The counterfactual approach compares outcomes under alternative interventions. An individual is observed under one actual course of events; the other outcome is missing. Randomisation can support comparisons of average effects, while the realised study still requires attention to its conduct and analysis. Hernán and Robins develop this distinction in Chapters 1 and 2 of Causal Inference: What If. The book is a deeper continuation, not a claim that every causal question has one simple formula.

For one common approach to observational data, Chapter 3 makes the needed assumptions explicit. The intervention must be sufficiently well defined and connected to the observed exposure. The measured information must support an appropriate comparison without remaining confounding, and the relevant groups must have a positive chance of each treatment being compared. These are commonly called consistency, exchangeability and positivity. They are substantive requirements, not boxes certified by a small p-value. Inspected book version, Chapter 3.

Return to revision time. “Make everyone study more” is underspecified. Compare offering a defined revision session with not offering it, in a particular population over a particular period. That is a different intervention from compelling an extra hour of solitary study. Write down which effect you want before deciding whether existing records can inform it. If the records contain no comparable students receiving one of the alternatives, a neat regression output cannot manufacture that comparison.

The regression videos contribute statistical mechanics and a useful warning about overfitting; they do not teach a complete causal-identification strategy. Keep prediction, description and causal estimation as separate aims when reading their examples. A model can predict well without explaining an intervention’s effect, and a causal estimate can remain uncertain even when the association is strong.

End with a short assumption audit. Name the intervention, comparison, population, outcome and follow-up period. Draw two plausible causal stories, identify an observation that would distinguish them, and state one assumption that the available data cannot verify. The aim is a more explicit claim with visible limits, not a stronger conclusion obtained by adding the word “causal”.

Next step

Work through Chapters 1–3 of What If, then identify the assumptions behind one observational claim in your own subject.

Limits of this route

The video descriptions were checked against identified official-transcript excerpts, not complete viewing. An accessible official media endpoint is not evidence that every frame played. Teaching examples are not independent validation of the underlying scientific or clinical claims. The causal account is based on selected book passages; it does not cover all identification strategies, time-varying treatment or selection bias. The revision example is hypothetical.