Causal Inference: Proving Cause, Not Just Correlation

WHY IT EXISTS: The world is full of correlations. A city's crime rate might correlate with its ice cream sales, but one doesn't cause the other. Businesses and scientists need to know which levers actually work. Causal inference provides the statistical framework to move from observing associations to proving cause-and-effect, enabling confident, high-stakes decisions.

THE MENTAL MODEL: Think of it as a controlled scientific experiment, but for messy, real-world data where you can't always run a perfect A/B test. Instead of just plotting two lines on a graph to see if they move together (correlation), causal inference tries to create a "counterfactual"—what would have happened if the "cause" (e.g., a new feature) had not been introduced? The difference between the actual outcome and the estimated counterfactual outcome is the causal effect.

HOW IT WORKS: The gold standard is a Randomized Controlled Trial (RCT), or A/B test, where subjects are randomly assigned to a treatment or control group. Randomization ensures the only systematic difference between groups is the treatment itself, isolating its effect. When an RCT isn't feasible, statisticians use quasi-experimental methods on observational data. Techniques like Difference-in-Differences, Regression Discontinuity, and Instrumental Variables use clever ways to mimic an experiment by finding natural "shocks" or thresholds in the data to isolate a variable's impact.

WHEN TO USE IT: Use causal inference when you need to know if an intervention truly works and the cost of being wrong is high. It's for answering questions like, "How much did our new pricing model increase revenue?" or "Did this public health policy reduce hospital admissions?" It's about measuring the impact of a change.

WHEN NOT TO USE IT: Avoid these complex methods for simple descriptive analytics or forecasting where association is sufficient. If you just need to know that sales are higher in the summer to stock more inventory, you don't need to prove the ultimate cause. Applying causal methods to small datasets or without deep statistical expertise is a common footgun that can produce misleading results.

ONE CANONICAL EXAMPLE: To find the effect of a new fertilizer on crop yield, a simple correlation is not enough; fields with fertilizer might also get more sun. Causal inference would use an RCT: randomly assign plots of land to receive the fertilizer (treatment) or not (control). By comparing the average yield between the two groups, you can isolate the fertilizer's true effect, as randomization ensures other factors like sun and soil quality are, on average, balanced out.