Pearl's do-calculus

Description

A set of three formal rules for deciding whether an interventional quantity P(Y | do(X)) can be computed from observational data given a causal graph.

Do-calculus distinguishes seeing from doing: P(Y | X) is what we observe among users who happened to use a feature, while P(Y | do(X)) is what would happen if we made everyone use it. Given a causal DAG, Pearl's three rules (insertion or deletion of observations, action/observation exchange, insertion or deletion of actions) transform interventional expressions into observational ones where possible. The calculus is complete: if an effect is identifiable from the graph, the rules will find the formula, and if not, no adjustment can rescue it. For UX, its value is less in hand derivation and more in using software that applies it to say whether log data can answer a causal question at all.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Translate the product question into an interventional query, e.g. the effect on retention of do(enable auto-invest).
  2. Draw the causal DAG including unmeasured variables as latent nodes, drawing on qualitative research and domain experts.
  3. Use an identification tool (e.g. DAGitty, DoWhy or the R package causaleffect) to apply do-calculus and return an identifying formula or a non-identifiability result.
  4. If identifiable, implement the estimand (backdoor, front-door or a more general formula) with appropriate estimators.
  5. If not identifiable, state which additional measurement or experiment would make it so.
  6. Document the graph and derivation so reviewers can challenge the assumptions rather than the arithmetic.

Sources

  • Causal Inference: What If (Miguel Hernán & James Robins)
  • DAGs+SEM Playbook (Vivek Gopinath)
  • Nash Equilibrium Playbook (Vivek Gopinath)
  • Causality (Judea Pearl)
  • The Book of Why (Judea Pearl & Dana Mackenzie)
  • Causal Inference (Hernán & Robins)
  • DAGs+SEM Playbook (Gopinath)
  • Nash Equilibrium Playbook (Gopinath)
  • The Inquiry Series, Ch 45 — The Causal Revolution — Pearl's Ladder & the Fundamental Problem

Also known as

  • Do-calculus
  • Causal calculus
  • do-operator

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference