Prediction and Causal Inference

Prediction asks what outcome is likely given available information. Causal inference asks how an outcome would change under a specified intervention. These questions can use some of the same data and statistical tools, but they require different reasoning. A highly accurate predictor does not automatically identify an effective action. It may rely on signals that are useful for forecasting yet have no causal influence on the outcome.

Imagine a school dataset in which learners receiving more tutoring tend to have lower examination marks. A predictive model might use tutoring hours as a signal of lower expected marks. It would be a mistake to conclude that removing tutoring would improve achievement. Learners may have been offered tutoring because they were already struggling. Prior difficulty influences both tutoring allocation and later marks, creating confounding. The association cannot be interpreted as the effect of tutoring without addressing how the support was assigned.

One way to express a causal question uses potential outcomes: what would happen to the same learner under one intervention and under another? Only one of those outcomes is observed for that learner in the realized situation. Estimating an average effect therefore requires a design and assumptions that make comparisons informative. Hernán and Robins develop this perspective, while Pearl explains causal reasoning through graphical and structural models. Both approaches make clear that causal interpretation needs more than a fitted association.

Random assignment can support a credible comparison when appropriate and feasible, because assignment does not systematically follow prior learner characteristics. Even then, missing outcomes, nonparticipation, spillovers between learners, and changes in implementation require attention. Observational approaches can also support causal inference under explicit assumptions, using designs such as discontinuities, suitable comparison groups, or careful adjustment. Adding many variables to a model is not automatically helpful: adjusting for the wrong variable can introduce bias rather than remove it.

For an AI-supported feedback pilot, define the intervention precisely. Is the question about access to the tool, actual use, a particular feedback routine, or an entire teaching programme? Define the comparison, population, timeframe, and learning outcome. An assisted homework score is different from an unaided transfer test. A diagram of plausible causal relationships can expose assumptions about prior attainment, teacher choices, motivation, and access. The diagram is a statement to justify, not evidence that the assumptions are true.

Educators can practice by writing two questions about the tutoring example: one forecasting a learner's likely result and one estimating the effect of offering support. They then identify what evidence each question needs. Causal estimates also have uncertainty and may not transport to a different school or learner population. Average benefit does not imply equal benefit for every individual. The practical value of the distinction is that an institution can use prediction to organize attention while demanding additional evidence before claiming that acting on the prediction will improve learning.