.jpg){kind=link}
- Published
- Updated
Randomisation does not guarantee identical characteristics in every realised sample. Baseline differences should be assessed for process problems, chance and measurement inconsistency; a single p-value is not a substitute for understanding their origin and relevance.
Define the question and assumptions
First check the allocation process and underlying data sources. Establish the prognostic importance of an apparently imbalanced characteristic and its role in the planned analysis. Testing many baseline variables individually does not provide sufficient justification for selecting a new adjustment model after outcomes have been examined.
Hypothetical study example
In a hypothetical small study, one arm has greater baseline severity. The team examines randomisation and measurement records while retaining the planned main analysis. If a scientifically justified additional model is used, its purpose, timing and consequences for interpretation are clearly identified in the report.
Design review checks
- Check whether imbalance originates in source observations or analytical derivations.
- Avoid selecting clinically important variables solely through significance labels.
- Report planned analyses separately from subsequently added supportive adjustment models.
Record the decision and its limits
Accepted assessment neither automatically declares randomisation failure nor dismisses real process defects as chance. Use process evidence, clinical relevance and uncertainty together. Do not rewrite the main conclusion solely around a more favourable adjusted estimate; explain the relationship between alternative analyses and the original scientific question.
Frequently asked questions
- Does every baseline difference indicate an error?
- No. Chance differences occur, while process evidence still requires examination.
- Must a significant baseline difference enter the model?
- Selection needs scientific rationale and an analysis plan rather than a p-value alone.