Confounder/mediator/collider identification

Description

Classifies each variable relative to a cause and outcome as a confounder, mediator or collider, which decides whether to adjust for it.

Whether to control for a variable depends on its causal role, not its correlation. A confounder causes both the treatment and the outcome and must be adjusted for. A mediator lies on the causal path and adjusting for it removes part of the effect you want to measure. A collider is caused by both treatment and outcome (or their causes); adjusting for or selecting on it creates spurious associations. In UX, selection on colliders is common: analysing only users who completed onboarding, or only survey respondents, can manufacture effects such as a trust badge appearing to harm conversion. Rigour comes from reading roles off an explicit DAG rather than intuition.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. Draw the DAG for the treatment and outcome of interest, including key unmeasured variables.
  2. For each other variable, trace its paths: if it causes both treatment and outcome it is a confounder; if it sits between them it is a mediator; if two arrows point into it it is a collider on that path.
  3. Check sample filters and inclusion rules as potential collider selection (e.g. analysing only active users).
  4. Decide adjustment: include confounders, exclude mediators for total effects, and never condition on colliders or their descendants.
  5. Document the role of every variable in a table alongside the evidence for its placement.

Sources

  • DAGs+SEM Playbook (Vivek Gopinath)
  • Causal Inference: What If (Miguel Hernán & James Robins)
  • Statistical Rethinking (Richard McElreath)
  • Cole et al. (2010) Illustrating bias due to conditioning on a collider
  • DAGs+SEM Playbook (Gopinath)
  • Causal Inference (Hernán & Robins)
  • Statistical Rethinking (McElreath)
  • The Inquiry Series, Ch 46 — Directed Acyclic Graphs — Making Assumptions Explicit
  • Cinelli, C., Forney, A. & Pearl, J., A Crash Course in Good and Bad Controls (2022)

Also known as

  • Causal role classification
  • Variable role analysis

Categories

  • Research Operations
  • Research Design & Rigor