Causal Inference Basics

• Define the counterfactual and the causal question precisely
• Use DAGs to find confounders and selection bias
• Choose between A/B tests, natural experiments, and regression controls

• Budget follows causality: you only double down on levers that genuinely move outcomes.
• Confounders can fake huge 'effects' in observational data.
• Asking the causal question sharply often saves the cost of a misguided A/B test.

The rain (a common cause) drives both — the exact structure a confounder has, and exactly what causal analysis must untangle.

• State the counterfactual: outcome under treatment vs under no treatment for the same users.
• Draw a small causal diagram (DAG) of the variables to spot confounders and colliders.
• Choose the design: randomize if possible; otherwise use a natural experiment, before/after with a control, or regression controls.
• Estimate the effect and check the assumptions that make it credible.

• Growth decides which activation nudges actually raise D7 retention vs correlate with it.
• Pricing estimates elasticity without a matched-price test using regional variation.
• Platform teams estimate feature roll-out impact with difference-in-differences vs non-rolled countries.

• Write the causal question before touching data — design should not follow the data's hint.
• Draw the DAG and question every arrow that sneaks in.
• Report effect estimates with their assumptions front and center (unconfoundedness, parallel trends).

• Adjusting for variables that are actually outcomes of the treatment (overcontrolling), which biases the effect.
• Reading the regression of Y on X with confounders as a causal estimate without defending the DAG.
• Using a natural experiment whose 'intervention' also changed other things at once.

• If the effect flips sign after adding a variable, return to the DAG — you may have adjusted a collider.
• If the parallel-trends plot shows the two lines diverging before the intervention, the design is broken.