Causal Inference

A predictive association asks what outcomes accompany an observed variable; a causal effect asks what outcomes would follow an intervention. Potential-outcomes reasoning compares alternative outcomes for the same unit, while causal graphs describe relationships and possible confounding. Because both alternatives usually cannot be observed, identification depends on a design or assumptions such as exchangeability, overlap and a consistent treatment definition. Randomized experiments can support identification through assignment; observational studies require additional reasoning. Estimation then turns the identified quantity into a numerical result. A sophisticated regression or propensity model cannot, by itself, supply the missing causal assumptions.

Start by defining the intervention, outcome, target population and time horizon. Draw or document a causal model before deciding which variables to adjust for; controlling for a mediator or collider can change or bias the question. Assess whether treatment groups have comparable support and select an estimation strategy consistent with the design. Report uncertainty and examine alternative specifications, placebo tests and sensitivity to unmeasured confounding. The deliverable is an effect estimate with a defensible identification argument and limitations, rather than a coefficient labeled causal because it is statistically significant.

Causal conclusions can fail when key confounders are unmeasured, treatment groups lack overlap or one person's treatment affects another's outcome. Refutation tests can expose weaknesses, but passing them does not prove the graph or assumptions true. Average effects may conceal substantial variation, and results need not transport to another population. Causal inference is distinct from forecasting and feature importance: a variable useful for predicting an outcome is not necessarily an effective intervention target.