DAG-informed covariate selection

Description

Chooses which variables to include as controls in a model by reading them off a causal diagram, rather than by significance tests or data availability.

Covariate selection by stepwise regression or 'include everything' can introduce bias by adjusting for mediators or colliders, or leave confounding by omitting unmeasured causes. DAG-informed selection uses the backdoor criterion to pick a minimal sufficient set, adds precision variables that predict the outcome only, and excludes instruments, mediators and colliders. When the full DAG is uncertain, VanderWeele's disjunctive cause criterion offers a practical rule: adjust for any pre-treatment variable that causes the treatment, the outcome or both, excluding known instruments. Rigour comes from deciding covariates before seeing outcome data and showing robustness across plausible adjustment sets.

Attributes

Loss type
Unmigrated
Effort
Hours

Steps

  1. List all available pre-treatment variables and place them in the DAG.
  2. Derive minimal sufficient adjustment sets for the effect of interest using DAGitty.
  3. Add outcome-only predictors for precision, and remove variables affected by treatment, colliders and pure instruments.
  4. Where the DAG is uncertain, apply the disjunctive cause criterion and record which variables were included on that basis.
  5. Pre-register or time-stamp the covariate list before running outcome models.
  6. Run the analysis with two or three plausible adjustment sets and report how much the estimate changes.

Sources

  • DAGs+SEM Playbook (Vivek Gopinath)
  • Causal Inference: What If (Miguel Hernán & James Robins)
  • Statistical Rethinking (Richard McElreath)
  • VanderWeele (2019) Principles of confounder selection
  • 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 covariate selection
  • Principled adjustment

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference