Estimate an exposure–outcome relationship
Adjust for relevant common causes of the exposure and outcome. Avoid automatically adjusting for variables that occur after the exposure or are consequences of two other variables.
Methods note · regression and confounding
How do you choose covariates?
Decide what each variable is doing before putting it in a model. Use the research question, timing and clinical knowledge, not a list of bivariate p-values.
Use this when: you have a list of possible regression variables.
Bring: the outcome, main exposure, study timeline and purpose of the model.
You will make: a short record explaining why each variable is included or excluded.
Start with purpose
First decide what you want the model to do.
Make useful predictions for new people
Use information that would genuinely be available at the prediction time, then check performance on data not used to fit the model.
Do not apply causal-adjustment rules to a prediction model. Likewise, do not apply prediction-selection procedures to a causal question. Write down the model’s purpose before choosing any covariates.
Practical workflow
Make every covariate decision visible.
- Define the target effect. Name the exposure, comparator, outcome, timeframe, population and whether the target is a total or direct effect.
- List candidates before outcome-led modelling. Use the protocol, clinical knowledge, literature and data dictionary, not a table of bivariate p-values.
- Order variables in time. Mark each as pre-exposure, exposure-time or post-exposure.
- Assign a causal role. Plausible cause of exposure, cause of outcome, common cause, mediator, collider, design variable, precision variable or duplicate/proxy.
- Draw the assumed pathways. Use a simple causal diagram to identify which non-causal paths must be blocked.
- Check the data can support the set. Review measurement quality, missingness, sparse categories, collinearity, outcome events and parameters.
- Pre-specify and justify. Record the chosen primary set and any sensitivity sets before interpreting the adjusted exposure estimate.
| Candidate | Timing and assumed role | Decision | Reason |
|---|---|---|---|
| Variable name | Pre/post exposure; cause, mediator, collider, precision or design variable | Include / exclude / sensitivity | Substantive and causal justification |
Standalone worked example
Exercise-class attendance and falls.
A community service follows older adults for six months. Attendance at a weekly strength-and-balance class is observed, not assigned. The question is whether regular attendance is associated with having a fall. The table shows a plausible causal starting point, not a truth discovered by software.
| Candidate | Working role | Teaching decision | Why |
|---|---|---|---|
| Age | Pre-attendance common cause | Include | May influence both class attendance and fall risk. |
| Falls in the previous year | Prior outcome and common cause | Include | May prompt attendance and strongly predicts another fall. |
| Baseline mobility | Pre-attendance common cause | Include | May affect ability to attend and later fall risk. |
| Long-term conditions | Pre-attendance health status | Include | May affect access, participation and fall risk. |
| Distance from the venue | Access factor | Consider and justify | May affect attendance; include only if the assumed pathways support adjustment. |
| Balance at three months | Possible mediator | Exclude from the total-effect model | The class may improve balance, which may then reduce falls. |
| Enjoyment of the class | Post-attendance consequence | Exclude from the primary model | It is measured after attendance begins and may sit on the pathway to continued attendance. |
A real study must justify this structure from its protocol, setting and evidence. No table or p-value can discover the true causal diagram.
Keep bivariate analysis in its proper role
Crude associations matter, but answer a different question.
Describe the unadjusted exposure–outcome association, its direction, magnitude and uncertainty.
Reveal sparse cells, distributions, coding problems, unexpected patterns and the crude estimate used for comparison.
Decide which variables are confounders or admit covariates to the model according to p<0.05.
Use the target effect, temporal order, subject knowledge and assumed causal structure to select a defensible set.
The analytical sequence
Describe the crude association → select covariates independently → fit the adjusted model → compare crude and adjusted estimates → explain the difference.
Previous falls differ strongly between attendance groups. Is a small bivariate p-value the reason to adjust for them?
No. The comparison documents baseline imbalance. Previous falls belong in the proposed adjustment set because they occur before attendance and plausibly affect both attendance and the chance of another fall.
Go deeper · causal diagrams and covariate selection
DAGitty
Draw assumed causal structures and identify minimally sufficient adjustment sets.
Open DAGittyWhat If
Hernán and Robins explain causal questions, confounding and adjustment through epidemiological examples.
Open the bookSTROBE
Report how confounders were defined, why variables entered models and how adjusted estimates were obtained.
Open STROBE