Front-Door Adjustment

Description

Estimate the effect of a feature on an outcome when unmeasured factors confound both, by going through a measured step that the feature works through.

Backdoor adjustment needs every confounder to be measured. Often in product data it cannot be: the same unmeasured motivation makes people adopt a feature and also makes them retain. The front-door criterion offers another route when the feature affects the outcome only through a measurable intermediate step (a mediator) and that step is not itself driven by the hidden confounder except through the feature. You estimate two links that are each unconfounded, feature to mediator and mediator to outcome (adjusting for the feature), and combine them to recover the full causal effect. It is the diagram in which D causes M, M causes Y, and an unmeasured U affects both D and Y. The assumptions are strong and must be argued from how the product works, which is exactly why drawing the DAG with the team is part of the method.

Attributes

Loss type
Siloed
Stage
Post-launch
Effort
Days

Steps

  1. Draw a DAG of the feature (D), the outcome (Y), plausible unmeasured confounders (U) and the mechanism the feature works through (M).
  2. Check the front-door conditions: D affects Y only through M; no unblocked backdoor path from D to M; every backdoor path from M to Y is blocked by D.
  3. Make sure the mediator is logged reliably for every user (for example, the number of saved searches after adopting alerts).
  4. Estimate the effect of D on M, then the effect of M on Y while adjusting for D.
  5. Combine the two with the front-door formula, using bootstrap for intervals.
  6. Stress-test the result: argue each assumption with product and engineering colleagues and compare with any experiment you can run.

Sources

  • Cinelli, C., Forney, A. & Pearl, J., A Crash Course in Good and Bad Controls (2022)
  • Pearl, J., Causality: Models, Reasoning, and Inference (2009)

Also known as

  • Front-door criterion
  • Mediator-based causal identification

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference