05. Use causal assumptions to choose adjustment variables
Lesson
LEARNING OBJECTIVES
Make causal assumptions explicit, distinguish common causes from mediators and colliders, and justify a covariate strategy for a defined causal contrast.
DRAW THE QUESTION, NOT JUST THE AVAILABLE COLUMNS
Begin with a specific exposure or intervention and outcome, in a defined population and period. Draw a directed acyclic graph on paper using substantive knowledge and relevant evidence. Include important unmeasured variables and selection processes when relevant. Arrows represent assumed causal relationships; absent arrows are assumptions too.
A common cause can create a non-causal path between exposure and outcome. A mediator lies on a pathway through which exposure may influence outcome. A collider has two or more causes; conditioning on it can create an association between its causes. Stratification, regression adjustment, restriction and sample selection can all condition on variables.
Do not choose covariates by searching for the model with the smallest exposure p-value. Decide whether the target is a total effect, a direct effect or simply an association, and justify an appropriate strategy. Different targets can require different variables and assumptions.
WORKED FICTIONAL CASE
Suppose household resources affect both water-source choice and childhood illness through other conditions. Household resources may therefore create a confounding path. If water source changes treatment practices that then affect illness, those practices could be mediators of part of the total effect.
Now suppose both illness and household resources influence clinic attendance. Restricting the analysis to clinic attenders can induce a distorted relationship through selection. Recording many clinic variables does not guarantee that this selection problem can be repaired.
These relationships are hypotheses, not proven facts about a particular community. Discuss them with subject specialists and people familiar with the setting; consider plausible alternative diagrams.
ADDITIONAL IDENTIFICATION QUESTIONS
Are compared exposure strategies defined clearly enough that their versions have a meaningful interpretation? Are relevant common causes measured adequately? Do all relevant covariate groups have a realistic opportunity to receive each exposure strategy? These concern consistency, exchangeability and positivity. A diagram alone cannot demonstrate that all such conditions hold.
YOUR TASK
For your own question, draw and date a graph. Write a short role statement for each proposed variable: exposure, outcome, common cause, possible mediator, selection factor or unresolved. Propose an adjustment set and explain why each variable is included or excluded.
Add one alternative plausible graph and state how it could change measurement or analysis. Record which assumptions cannot be checked from the observed dataset alone.
SELF-CHECK
Does adjustment for a mediator necessarily estimate the total effect? Does a causal graph prove its own assumptions?
FEEDBACK
No. Conditioning on a mediator can block part of the pathway of interest and may introduce other problems. Graphs clarify assumptions and potential adjustment choices; they do not validate an incomplete causal story or eliminate unmeasured confounding.
PORTFOLIO OUTPUT
A dated causal-assumptions diagram, a covariate-role table and an explicit adjustment rationale.
FURTHER READING
BMJ: How to use directed acyclic graphs: guide for clinical researchers.
https://www.bmj.com/content/388/bmj-2023-078226
Make causal assumptions explicit, distinguish common causes from mediators and colliders, and justify a covariate strategy for a defined causal contrast.
DRAW THE QUESTION, NOT JUST THE AVAILABLE COLUMNS
Begin with a specific exposure or intervention and outcome, in a defined population and period. Draw a directed acyclic graph on paper using substantive knowledge and relevant evidence. Include important unmeasured variables and selection processes when relevant. Arrows represent assumed causal relationships; absent arrows are assumptions too.
A common cause can create a non-causal path between exposure and outcome. A mediator lies on a pathway through which exposure may influence outcome. A collider has two or more causes; conditioning on it can create an association between its causes. Stratification, regression adjustment, restriction and sample selection can all condition on variables.
Do not choose covariates by searching for the model with the smallest exposure p-value. Decide whether the target is a total effect, a direct effect or simply an association, and justify an appropriate strategy. Different targets can require different variables and assumptions.
WORKED FICTIONAL CASE
Suppose household resources affect both water-source choice and childhood illness through other conditions. Household resources may therefore create a confounding path. If water source changes treatment practices that then affect illness, those practices could be mediators of part of the total effect.
Now suppose both illness and household resources influence clinic attendance. Restricting the analysis to clinic attenders can induce a distorted relationship through selection. Recording many clinic variables does not guarantee that this selection problem can be repaired.
These relationships are hypotheses, not proven facts about a particular community. Discuss them with subject specialists and people familiar with the setting; consider plausible alternative diagrams.
ADDITIONAL IDENTIFICATION QUESTIONS
Are compared exposure strategies defined clearly enough that their versions have a meaningful interpretation? Are relevant common causes measured adequately? Do all relevant covariate groups have a realistic opportunity to receive each exposure strategy? These concern consistency, exchangeability and positivity. A diagram alone cannot demonstrate that all such conditions hold.
YOUR TASK
For your own question, draw and date a graph. Write a short role statement for each proposed variable: exposure, outcome, common cause, possible mediator, selection factor or unresolved. Propose an adjustment set and explain why each variable is included or excluded.
Add one alternative plausible graph and state how it could change measurement or analysis. Record which assumptions cannot be checked from the observed dataset alone.
SELF-CHECK
Does adjustment for a mediator necessarily estimate the total effect? Does a causal graph prove its own assumptions?
FEEDBACK
No. Conditioning on a mediator can block part of the pathway of interest and may introduce other problems. Graphs clarify assumptions and potential adjustment choices; they do not validate an incomplete causal story or eliminate unmeasured confounding.
PORTFOLIO OUTPUT
A dated causal-assumptions diagram, a covariate-role table and an explicit adjustment rationale.
FURTHER READING
BMJ: How to use directed acyclic graphs: guide for clinical researchers.
https://www.bmj.com/content/388/bmj-2023-078226
My study notes
Stored only in this browser. Do not enter patient information.Saved in this browser. No account needed.