causal-identification

Causal Identification

Overview

A regression coefficient is a correlation with good posture. It becomes a causal effect only when a design rules out the other explanations — and that design rests on assumptions no amount of clean data or tight standard errors can supply. The fatal error is silent: the code runs, the coefficient is significant, the sign is plausible, and it's still confounding wearing the costume of an effect.

Core principle: State the identification assumptions before you estimate, and test the ones that are testable. The estimate is only as credible as the assumption you can't test.

Before any model, answer the Angrist–Pischke question: if you could run the ideal randomized experiment, what would it be — and what real-world variation are you using as a stand-in? Name the source of variation in one sentence and argue it's as good as random, or you have a regression hoping to be an identification strategy.

The Design Card — sign-off before estimation

A causal claim locks a Design Card — invoke analysis-state-management, write it to docs/analysis/ (or point to the PAP), get sign-off before estimating. (Structural → model card, prediction → Prediction Spec — same gate, different document; mechanics owned by analysis-checkpoints.) Fields:

• Causal question + estimand — ATT/ATE/LATE, for which population.
• Design + source of variation — the "what's your experiment?" answer, one sentence.
• The untestable assumption, in plain language, and why it's plausible here.
• Diagnostics planned — the design-specific tests you'll run before reading the estimate.
• Robustness shortlist — the ~3 checks aimed at the main threat (approval-gated, see below).
• Primary spec — outcome, treatment, FE/controls (each control with its confounding story), SEs/clustering.

Entering mid-stream ("just run the DiD") does not waive the card: "the user already said regress Y on X" names the spec, not the design — reconstruct the card from context in ≤10 lines and confirm before estimating.