Research methods guide
How to select confounders with a DAG
Start with a time-ordered graph of variables and assumed causal arrows, then identify the paths that require adjustment. The set depends on whether the target is a total or direct effect.
Consider the total effect of remote monitoring on 90-day admission. Specify the effect, order variables by time and draw paths to see what should be blocked.
Confounders, mediators and colliders
E denotes exposure and Y outcome. Arrows express substantive assumptions.
| Item | Worked example | Verification |
|---|---|---|
| Baseline severity S | S → E and S → Y. Sicker patients are more likely to receive monitoring and be admitted. | Consider adjusting for pre-exposure severity and examine proxy quality. |
| Post-exposure adherence M | E → M → Y. Monitoring may change adherence and then admissions. | Preserve this path for total effects. A direct effect needs a separate question and assumptions. |
| Follow-up inclusion V | E → V ← U → Y. Exposure and unmeasured motivation influence follow-up. | Conditioning on completers can open a path. Address follow-up in design and sensitivity analysis. |
Apply the method to your materials
Put variables in temporal order
List pre-exposure, exposure and follow-up nodes, and review arrows, common causes and measurement times with domain experts.
Compare plausible graphs
Document adjustment sets and failures of identification under competing assumptions. Record data source and measurement quality for each covariate.
Questions that arise in practice
Drop a confounder with a large univariable P value?
Confounding is a causal and measurement question. Document the path and assess measurement before considering empirical variation.
Verified sources
The official sources below were checked on 2026-11-12. Access conditions and update dates are rechecked before publication.
RigorMed workspace
Start with your research materials
Bring your research question, available materials, and the steps you want to check.
