Risk of Bias Assessment Tool

Risk of bias concerns whether the way a study was designed, conducted, analysed or reported could systematically distort a particular finding. It is different from asking whether a manuscript mentions every reporting item. A well-written report can describe a biased study, and a short report can omit information about sound methods. Start with the design and the question your review needs to answer, then use the instrument developed for that situation.

RoB 2 here is the 22 August 2019 instrument for individually randomized parallel-group trials. Cluster and crossover variants are outside this page’s scope. Specify a result or outcome, a time point and whether you want the effect of assignment or adherence. These choices change which departures from intervention matter and which Domain 2 questions apply. Assessing mortality at 30 days does not automatically assess pain at six months in the same trial.

ROBINS-I 2016 evaluates non-randomized intervention effects against a target randomized trial. Describe eligibility, intervention, comparator, follow-up, important confounders and the causal effect in the target-trial field. It is not a generic instrument for every observational exposure study. The newer ROBINS-I V2, revised in November 2025, is a draft and is not implemented here. QUADAS-2 is pinned to 2011 for diagnostic accuracy: describe the intended patients, index test, reference standard and target condition. QUADAS-3 is now the recommended newer version; this page does not substitute it.

RoB 2 decision tables deterministically convert the proposed signalling answers to draft domain judgements. The overall suggestion takes the worst domain; several domains with Some concerns can also justify High when their combined effect substantially reduces confidence. A reviewer can record that reasoning as an overall override. ROBINS-I 2016 uses its published judgement guidance. Some branches require information about the magnitude of a problem that is not captured by a yes/no signalling answer. Those branches remain unresolved until a reviewer supplies a justified domain judgement.

QUADAS-2 signalling questions inform judgement rather than generate a score. All favourable answers can support a Low draft, but an adverse answer needs review of its potential to introduce bias. That domain remains Unclear until a reviewer resolves it. Its four risk-of-bias domains and three applicability judgements remain separate. There is no official whole-study numeric score or overall category in this output. The categorical figure includes text and symbols with colour, and distinguishes unknown information from Some concerns, Moderate and Critical.

NOS has separate cohort and case-control forms. Specify the most important and additional comparability factors; cohort appraisal also needs a follow-up duration and completeness threshold. Native awards produce up to four Selection stars, two Comparability stars and three Outcome or Exposure stars. Nine stars is a maximum, not a validated threshold for a low-risk band. JBI preserves design-specific item responses without adding a total. Cohort uses the verified 2017 archive and qualitative uses the 2020 archive, each explicitly labelled. Export the JSON or CSV with the figure so readers can inspect answers, evidence, target, edition and reviewer override reasons.

For example, “There were no deviations arising because of the experimental context” supports No for RoB 2 question 2.3 about trial-context departures. Missed sessions alone do not establish that cause. That answer informs domain 2 alongside the analysis questions; it cannot establish an overall Low judgement by itself. A reviewer can change the answer or judgement with a recorded reason after checking the study and companion materials.

Review the OpenAI privacy policy before submitting study text.