Home/Insights/Methodology reference/Experimental design in research methodology
Experimental design in research methodology
The investigator controls the exposure and assigns it at random. That one feature is what licenses a causal claim.
What randomisation actually does
Random assignment distributes both known and unknown confounders between groups by chance. No other design handles unmeasured confounding, which is why the randomised trial sits where it does in the evidence hierarchy.
Randomisation is not the same as alternation, assignment by date of birth, or by clinician preference. Those are predictable and therefore subvertible.
Allocation concealment — keeping the upcoming assignment from whoever enrols participants — is a separate safeguard from randomisation itself, and its absence is associated with exaggerated effect estimates.
Blinding
Blinding protects what happens after assignment: differential co-intervention, differential assessment, and differential dropout. Participants, carers, outcome assessors and analysts can each be blinded independently.
"Double-blind" is ambiguous and should be replaced by stating exactly who was blinded. Where blinding is impossible, blinded outcome assessment is often still achievable and should be.
Common structures
Parallel group: participants receive one intervention. Crossover: each receives both in sequence, which controls between-person variation but requires a stable condition and an adequate washout to avoid carry-over. Factorial: two or more interventions tested simultaneously, efficient when interaction is implausible. Cluster-randomised: groups rather than individuals are assigned, which requires analysis accounting for clustering and a larger sample.
What must be fixed before the data arrive
A single primary outcome, the analysis plan, and the sample size calculation with its assumed effect size. Trials are registered prospectively so that these are on record — ICMJE member journals require registration before the first participant is enrolled.
Analysis is by intention to treat: participants are analysed in the group they were assigned to, regardless of what they received. Per-protocol analysis reintroduces the selection that randomisation removed and is a secondary analysis at best.
In practice
Fix, flag or query.
The distinction that keeps an edit defensible: what an editor may change silently, what must be shown to the author, and what only the author can answer.
- Register: prospectively, with the primary outcome stated.
- Analyse: by intention to treat.
- Describe: who was blinded, rather than writing "double-blind".
Reporting standards
- CONSORT 2010 Statement and its extensions
- SPIRIT Statement, for trial protocols
- ICMJE Recommendations, on prospective trial registration
Reporting guidelines named here are hosted by the EQUATOR Network. Writing to the checklist for your design before submission is the cheapest available improvement to a methods section.
Related