_(6008495935).jpg){kind=link}
- Published
- Updated
Baseline variables may improve precision or account for design features in analysis. Choosing which covariates enter the model after examining treatment outcomes can weaken the credibility of the resulting interpretation, particularly when only favourable models are retained.
Define the question and assumptions
Assess clinical and prognostic rationale, measurement time and data quality. Functional form, category definitions and handling missing values matter as much as the variable list. Discuss why adjusting for a post-treatment variable can address a different question from adjustment for a genuine pretreatment characteristic.
Hypothetical study example
In a hypothetical study, the team proposes making the covariate combination with the smallest p-value the main model. The statistician distinguishes the prespecified model from this search. Additional models may be reported with transparent selection rationale and exploratory status rather than retaining only the most attractive result.
Design review checks
- Record each covariate together with its scientific rationale and data source.
- Specify transformations and category rules for continuous variables before outcome review.
- Test how missing covariates alter the population contributing to each analysis.
Record the decision and its limits
An accepted plan includes model structure and special-case rules, not just variable names. Feasibility of the planned model can be assessed using blinded data. Disclose timing and consequences of outcome-informed changes. Good model fit alone does not establish correspondence to the scientific question or freedom from systematic bias.
Frequently asked questions
- Are more covariates always better?
- No. Unnecessary complexity and limited information can affect model stability.
- Can a post-treatment variable be adjusted for?
- Its purpose and causal interpretation need separate assessment; it is not ordinary baseline adjustment.