Risk of Bias Domains

Risk of bias is judged stage by stage along a trial’s life. The Cochrane RoB 2 framework groups the threats as: selection (sequence generation + allocation concealment), performance (blinding of participants/staff), detection (blinding of outcome assessment), attrition (incomplete outcome data / loss to follow-up) and reporting (selective outcome reporting). Each is appraised on its own merits.

The trial’s journey — enrolment → randomisation → the two arms → outcome assessment and reporting — with the bias domain that threatens each step labelled in place. Selection at randomisation, performance during treatment, detection at assessment, attrition and reporting overlaid by the callout.

Walk down the diagram and ask one question per stage. Was allocation truly concealed? (selection). Were patients and staff blinded? (performance). Were outcome assessors blinded? (detection). Where did the missing patients go? (attrition). Were all pre-registered outcomes reported? (reporting). A serious flaw in any single domain can undermine the whole result.

“It’s an RCT” is a study design, not a quality certificate — a poorly conducted RCT can mislead as badly as an observational study. Domain-by-domain appraisal is exactly how GRADE downgrades trial evidence, and it is what the exam expects you to do rather than accept a trial at face value.

• Selection = sequence generation + allocation concealment
• Performance / Detection = blinding (subjects/staff; assessors)
• Attrition = missing data · Reporting = selective outcomes

Applying RoB 2 in practice — take an open-label ED trial of an analgesic where pain (a subjective outcome) is self-reported. Allocation may be well concealed (low selection risk), but with no blinding both performance and detection bias are high: patients and assessors knowing the arm can shift a soft outcome. If 20% are then lost to follow-up unevenly between arms, attrition risk climbs too — and the trial earns an overall “high risk” despite being randomised. Domains are judged per outcome: the same trial can be low risk for mortality (hard, objective) yet high risk for a subjective symptom score.

• Remembering selection and performance but ignoring attrition and selective reporting — both routinely overlooked.
• Conflating the domains — e.g. calling unblinded assessment “selection bias” when it is detection bias.
• Assuming blinding is always feasible — surgery and many ED interventions can’t be blinded, so other safeguards (blinded objective endpoints) matter more.