RoB 2 Risk of Bias Tool: Five Domains, Signalling Questions and a Worked Example

Published October 29, 2026  |  Last updated October 31, 2026

.evrob-wrap{overflow-x:auto !important;margin:22px 0 !important;}.evrob-table{border-collapse:collapse !important;width:100% !important;min-width:520px !important;}.evrob-table th{background-color:#0d2137 !important;color:#ffffff !important;padding:10px 12px !important;text-align:left !important;border:1px solid #cddaea !important;font-size:0.92rem !important;}.evrob-table td{background-color:#ffffff !important;color:#1a1a1a !important;padding:9px 12px !important;border:1px solid #cddaea !important;vertical-align:top !important;font-size:0.92rem !important;}.evrob-table tr:nth-child(even) td{background-color:#f0f5fa !important;}.evrob-cap{margin:6px 0 0 0 !important;font-size:0.85rem !important;color:#1a1a1a !important;}

The RoB 2 risk of bias tool is the revised Cochrane instrument for judging whether the result of a randomized trial can be trusted. Sterne and colleagues published it in the BMJ in 2019 (BMJ 2019;366:l4898). It looks at five domains, asks a fixed set of questions in each, and converts the answers into one of three judgements, Low risk of bias, Some concerns or High risk of bias. This guide explains the RoB 2 signalling questions domain by domain, shows how the algorithms reach a judgement, walks one hypothetical trial from first answer to overall rating, and sets out how to report the result in a systematic review.

What the RoB 2 Risk of Bias Tool Assesses

RoB 2 assesses a single trial result, not a whole trial. The Cochrane guidance defines that result as an estimate of the relative effect of two interventions on a particular outcome. A trial that reports pain, function and adverse events therefore needs several assessments. The guidance adds that these decisions are likely to line up with the outcomes in your Summary of Findings table, although it has not yet issued a formal rule on how many results to assess (riskofbias.info).

Three design features are worth fixing before you open the first paper.

• The domains are mandatory. The guidance states that the five domains cover all types of bias that can affect randomized trials, so each is required and no further domains should be added.
• The names describe causes. The authors avoided the version 1 labels selection bias, performance bias, attrition bias and detection bias because those labels did not describe the specific issues assessed.
• The effect of interest must be declared. You choose between the effect of assignment to intervention (the intention-to-treat effect) and the effect of adhering to intervention (the per-protocol effect). Only domain 2 changes with this choice, and the choice is unrelated to whether you pool odds ratios or risk differences.

Handbook chapter 8 warns that naive per-protocol and as-treated analyses may be seriously biased, and valid per-protocol estimation needs specialised methods, so most reviews specify the assignment effect. Our guide to intention-to-treat versus per-protocol analysis explains why that distinction matters before you start answering questions.

The evidence behind the domains is empirical. In the ROBES meta-epidemiological study, which covered 2,443 randomized trials in 228 Cochrane meta-analyses, intervention effect estimates were on average exaggerated in trials with high or unclear (versus low) risk-of-bias judgements for sequence generation, with a ratio of odds ratios of 0.91 (95% credible interval 0.86 to 0.98) (Savović et al., Am J Epidemiol 2018). That study used the older judgements rather than RoB 2, but it illustrates why the tool asks about the randomization process at all.

The Five Domains and Their RoB 2 Signalling Questions

The table below gives the five domains for an individually randomized parallel-group trial, the core issue each one tests, and the number of signalling questions in the template for the effect of assignment to intervention.

How RoB 2 Signalling Questions Are Answered

Each of the RoB 2 signalling questions is written to be reasonably factual. You answer with one of five options, Yes, Probably yes, Probably no, No or No information. Yes and Probably yes carry the same implication for risk of bias, as do No and Probably no. Questions are phrased so that a Yes can point to either a low or a high risk of bias, which is why you must read the algorithm and never assume that Yes is the good answer.

Four rules prevent most answering errors.

1. Use No information sparingly. The guidance reserves it for cases where the report is too thin to support Probably yes or Probably no and where a guess would be unreasonable. A large trial run by an experienced trials unit that does not describe sequence generation in a word-limited paper is more likely a Probably yes.
2. Know what a missing detail means. If a question looks for evidence of a problem, No information means no evidence of that problem. If it concerns something that should be reported, such as losses to follow-up, silence raises concern.
3. Answer each question independently. A baseline imbalance must not change your answers to the sequence generation or concealment questions. It is recorded under question 1.3 only.
4. Quote the report. The free-text box beside each question should hold brief direct quotations whenever possible.

The algorithms then map answers to a proposed domain judgement. Treat that output as a proposal. The guidance says users should verify it, may change it, and must explain any departure. It also tells assessors to read “risk of bias” as “risk of material bias”, so concerns belong only where an issue is likely to affect the ability to draw reliable conclusions.

Reliability needs attention in practice. In an inter-rater reliability study by Minozzi and colleagues, four raters applied RoB 2 to 80 results covering seven outcomes from 16 trials in a Cochrane review. During the calibration exercise they reached no agreement on the overall judgement (Fleiss kappa of -0.15), and the mean time to apply the tool was 168.5 minutes per study. The team then wrote a topic-specific implementation document and measured agreement again, and the paper reports that reliability improved (Minozzi et al., J Clin Epidemiol 2022). The practical lesson is to agree your interpretation rules, such as what counts as “nearly all” outcome data, before extraction starts.

Domain 1 RoB 2 Signalling Questions, Randomization Process

The RoB 2 signalling questions in domain 1 ask whether the allocation sequence was random (1.1), whether it was concealed until participants were enrolled and assigned (1.2), and whether baseline differences suggest a problem with randomization (1.3). Differences compatible with chance do not create risk, and a small number of differences significant at the 0.05 level are usually compatible with chance. A Low judgement requires adequate concealment and no concern about sequence generation.

Domain 2 RoB 2 Signalling Questions, Deviations from Intended Interventions

The RoB 2 signalling questions in domain 2 are the most complex, and they split into two parts for the effect of assignment. Part 1 covers questions 2.1 to 2.5. It asks whether participants (2.1) and carers or people delivering the intervention (2.2) were aware of assignment, and, if so, whether deviations arose because of the trial context (2.3), whether they affected the outcome (2.4) and whether they were balanced between groups (2.5). “Trial context” means effects of recruitment and engagement, such as a participant who feels unlucky after being assigned to the comparator and seeks the experimental treatment. Non-adherence that mirrors what would happen outside the trial is answered No or Probably no.

Part 2 covers questions 2.6 and 2.7 and asks whether an appropriate analysis estimated the effect of assignment. Intention-to-treat analyses and modified intention-to-treat analyses that exclude only participants with missing outcome data are appropriate. Naive per-protocol and as-treated analyses are inappropriate. Question 2.7 then asks whether analysing people in the wrong group could have had a substantial impact, and the guidance notes that this can be true even when fewer than 5% were affected if the outcome is rare.

Domain 3 RoB 2 Signalling Questions, Missing Outcome Data

The RoB 2 signalling questions in domain 3 begin with a simple test. Were data for this outcome available for all, or nearly all, randomized participants (3.1)? The guidance says availability from 95% of participants will often be sufficient for a continuous outcome, while for a dichotomous outcome the threshold depends on event risk. Imputed values count as missing data. If data are incomplete, you ask whether there is evidence that the result was not biased by the missingness (3.2), whether missingness could depend on the true value (3.3), and whether it is likely that it did (3.4). Reasons for a Yes at 3.4 include different proportions missing between arms, different reasons between arms, and censoring at treatment discontinuation in time-to-event analyses.

Domain 4 RoB 2 Signalling Questions, Measurement of the Outcome

The RoB 2 signalling questions in domain 4 ask whether the measurement method was inappropriate (4.1), whether measurement or ascertainment could have differed between groups (4.2), whether assessors knew the intervention received (4.3), whether knowledge could have influenced the assessment (4.4), and whether it likely did (4.5). For participant-reported outcomes the assessor is the participant. Knowledge of assignment can influence self-reported symptoms and judgement-based ratings, but the guidance notes it is unlikely to influence outcomes such as all-cause mortality.

Domain 5 RoB 2 Signalling Questions, Selection of the Reported Result

The RoB 2 signalling questions in domain 5 ask whether the data were analysed in line with a pre-specified plan finalized before unblinded outcome data were available (5.1), and whether the result is likely to have been selected from multiple eligible outcome measurements (5.2) or multiple eligible analyses (5.3). Examples of multiple measurements include several scales or time points for one outcome. Examples of multiple analyses include adjusted and unadjusted models or different cut-points. You need the protocol or statistical analysis plan to answer with confidence.

From Domain Judgements to an Overall Rating

The overall judgement uses the same three labels as the domains. A result is Low only if all domains are Low. It is Some concerns if at least one domain raises some concerns and none is High. It is High if at least one domain is High, or if there are Some concerns in multiple domains in a way that substantially lowers confidence in the result. That final clause is a judgement call, so record your reasoning whenever you use it.

The guidance also says a High judgement in any one domain should carry the same implication for the result as a whole, whichever domain it comes from. It discourages stopping once one domain is High when you are running a full systematic review, because transparent domain-level records help readers and future trial design. Stopping rules are acceptable only when your protocol states that High risk trials will play no role in the synthesis.

RoB 2 Signalling Questions in Practice, a Worked Example

The following trial is hypothetical, and the RoB 2 signalling questions below are answered for it in full. It exists only to show the mechanics. It is an open-label, two-arm randomized trial of a structured exercise programme versus usual care in adults with chronic low back pain. A total of 240 adults were randomized, 120 per arm. The outcome is pain intensity on a 0 to 10 numeric rating scale at 12 weeks, reported by the participant. The review asks about the effect of assignment to intervention.

The overall rating is High because domain 4 is High, whatever the other four domains say. Notice that the verdict hinges on one answer, question 4.5. Reviewers should record why they chose Probably yes, quoting the report where possible. The next table shows how the rating moves when one answer or one design feature changes.

Plots like Figure 2 are conventional in systematic reviews. McGuinness and Higgins built the robvis R package and Shiny web app so that review teams can produce publication-quality risk-of-bias figures quickly (Res Synth Methods 2021). For a broader look at drawing trial and meta-analysis plots, see our guide to clinical research graphs.

A Real Example, RoB 2 Signalling Questions Applied to the ARTHROL Trial

The invented trial above shows the mechanics, but it cannot show how the RoB 2 signalling questions behave against a real report. This section applies the tool to a published randomized trial, ARTHROL, which tested oral resveratrol against placebo in adults with painful knee osteoarthritis. The effect of interest is the effect of assignment to intervention (intention-to-treat), and the result assessed is the change in knee pain on a 0 to 100 numeric rating scale at 3 months. Tables 9 to 11 and Figure 3 show the appraisal.

The table below follows the trial report, the protocol and the analysis plan in the order of the five domains. Each row names where the evidence sits, so a second reviewer can check the answer in a minute. The domain judgements come from the algorithm tables in the Cochrane guidance of 22 August 2019, and the algorithm is only a proposal that an assessor may override with a stated reason.

The summary below collects the judgements and the answers that decided each one.

What a Low Rating Does and Does Not Tell You

The primary result was a mean change in knee pain of -15.7 points with resveratrol and -15.2 with placebo, a difference of -0.6 (95% CI -8.0 to 6.9). RoB 2 rates this result Low because the trial was run and reported in a way that protects against bias. It says nothing about precision. The authors randomized 142 of a planned 164 participants and describe the study as under-powered, and that limitation belongs to the imprecision judgement in a GRADE assessment, not to a RoB 2 domain.

Why Non-Adherence Did Not Count Against the Trial

Only 120 of 142 participants (85%) completed the allocated intervention, which looks worrying until you recall the question being asked. For the effect of assignment, departures from the protocol only matter when they arise from the trial context, and an analysis of everyone as randomized is appropriate. A review that asked about the effect of adhering to intervention would use a different set of questions in domain 2 and would need a more specialised analysis than the one reported here.

Where Two Assessors Might Differ

Four rows carry a judgement-call tag. Row 1.3 depends on how much baseline difference you accept as chance, row 3.1 sits on the 95% rule of thumb, and rows 5.1 and 5.3 depend on how you treat an analysis plan dated after follow-up and a change in the centre term. Under the algorithm, a less favourable answer to 1.3 (PY), 5.1 (PN or NI) or 5.3 (NI) would lower that domain, and with it the overall rating, to Some concerns, and a Y or PY on 5.3 would give High. An N on 3.1 would send the assessor to question 3.2 and the sensitivity analyses. Whichever way you answer, record the quotation and section that justified your choice. That record is what makes a RoB 2 assessment checkable by a second reviewer.

Compare this outcome with the hypothetical open-label trial earlier in this guide. There, a participant-reported outcome in an unblinded trial drove domain 4 to High. Here, the same kind of outcome sits in a masked, placebo-controlled design, so domain 4 stays Low. The outcome type did not change the rating, the masking did. The full signalling-question set is in the Cochrane RoB 2 guidance and the trial details are in the original article.

Cluster and Crossover Variants in Brief

The 2019 guidance supplies three templates, one each for individually randomized parallel-group trials, cluster-randomized parallel-group trials (including stepped-wedge designs) and randomized crossover trials and other matched designs. The five domains above apply to individually randomized trials including crossover trials. Cluster-randomized trials receive an additional domain, 1b, which covers bias arising from identification or recruitment of individual participants within clusters. The guidance states that no version exists yet for cluster crossover trials. The Cochrane Handbook summarises both variants in its Chapter 23, and the official templates sit on riskofbias.info. If you meet a crossover trial, our article on crossover design and washout periods explains the design features you will need to judge.

Reporting RoB 2 Signalling Questions and Judgements

Reporting should let a reader reproduce your judgements. The guidance recommends presenting consensus domain-level judgements in the main review, as a table, a figure or within a forest plot, and placing the signalling answers with their supporting text in an appendix. Individual assessor records belong in a supplement and should not be prominent. Cover the following points in your methods.

1. Name the tool and its version, and state which design template you used for each trial.
2. State the effect of interest, and for adherence, which deviations you assessed.
3. List the outcomes and time points assessed, and say which result from each trial you chose when several were reported.
4. Describe how assessors worked and how disagreements were resolved.
5. Explain every override of an algorithm proposal and any use of the multiple Some concerns clause.
6. Cite the sources you used for each trial, including protocols and registry entries, because domain 5 depends on them.

Report the review itself against PRISMA 2020, and place your risk-of-bias figure beside the forest plot. Our PRISMA 2020 flow diagram generator handles the flow diagram, and the GRADE evidence table calculator supports the step that follows. If you are writing this section inside a thesis, the article on matching reporting guidelines to thesis chapters shows where each item belongs.

Common Mistakes When Applying RoB 2

• Rating the trial instead of the result. Randomization problems affect the whole study, but deviations, measurement and selection of the reported result attach to the specific outcome.
• Trusting the label “intention-to-treat”. The guidance notes that the term has no consistent definition in study reports. Check the actual analysis against question 2.6.
• Treating covariate adjustment as a fix for broken randomization. The guidance says a trial that controlled for baseline imbalances to mitigate failures of randomization should be assessed with ROBINS-I instead (Sterne et al., BMJ 2016).
• Letting one answer drive another. Answer each question on its own evidence.
• Defaulting to No information. Use Probably yes or Probably no when the circumstances support a reasonable judgement.
• Skipping the implementation note. The reliability results above show why a short, topic-specific rule sheet is worth the time.
• Presenting algorithm output without reasoning. A bare colour grid invites doubt. Add abridged justifications.

Several of these mistakes also appear in the wider list in our guide to common mistakes in meta-analysis. If you are unsure how many trials your synthesis needs once the High risk ones are set aside, read the minimum number of studies for a meta-analysis. For the study-design hierarchy behind these decisions, see levels of evidence in clinical research, and for the next step after risk of bias, see our GRADE methodology guide. Our earlier overview of the Cochrane Risk of Bias Tool 2.0 gives a broader introduction, and our meta-analysis masterclass covers pooling, heterogeneity and publication bias.

Tools and Resources

Three free Editverse tools support the work around a risk-of-bias assessment. Use them to draw figures, choose tests and build flow diagrams.

Professional Support for Risk-of-Bias Assessment and Systematic Reviews

A defensible RoB 2 assessment starts with the review protocol, because the effect of interest, the outcomes and the rule sheet are all decided there. Editverse supports systematic reviews from protocol to manuscript, with reporting work aligned with COPE, ICMJE and GPP 2022. The packages below are the closest fit to the risk-of-bias workflow in this article.

Not sure which package fits your question? Our systematic review writing service page explains how the process works, and our manuscript writing services cover the write-up once your assessments are complete.

Conclusion

The RoB 2 risk of bias tool rewards a disciplined routine. Decide the effect of interest and the results to assess, write a short rule sheet, answer each of the RoB 2 signaling questions from the report and the protocol, and let the algorithms propose domain ratings that you then verify. The overall rating follows the worst domain, with one judgement-based exception for several Some concerns. Report consensus judgements in the main text and the supporting answers in a supplement. Done this way, the assessment becomes a transparent record that readers, peer reviewers and GRADE panels can follow.

References

1. Sterne JAC, Savović J, Page MJ, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ. 2019;366:l4898. DOI 10.1136/bmj.l4898. PMID 31462531. Open source
2. Higgins JPT, Savović J, Page MJ, Elbers RG, Sterne JAC, editors. Revised Cochrane risk-of-bias tool for randomized trials (RoB 2). Guidance document, 22 August 2019. Open source
3. Cochrane Handbook for Systematic Reviews of Interventions, Chapter 8, Assessing risk of bias in a randomized trial. Open source
4. Sterne JA, Hernán MA, Reeves BC, et al. ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions. BMJ. 2016;355:i4919. DOI 10.1136/bmj.i4919. PMID 27733354. Open source
5. Savović J, Turner RM, Mawdsley D, et al. Association between risk-of-bias assessments and results of randomized trials in Cochrane reviews: the ROBES meta-epidemiologic study. Am J Epidemiol. 2018;187(5):1113-1122. DOI 10.1093/aje/kwx344. PMID 29126260. Open source
6. Minozzi S, Dwan K, Borrelli F, Filippini G. Reliability of the revised Cochrane risk-of-bias tool for randomised trials (RoB2) improved with the use of implementation instruction. J Clin Epidemiol. 2022;141:99-105. DOI 10.1016/j.jclinepi.2021.09.021. PMID 34537386. Open source
7. McGuinness LA, Higgins JPT. Risk-of-bias VISualization (robvis): an R package and Shiny web app for visualizing risk-of-bias assessments. Res Synth Methods. 2021;12(1):55-61. DOI 10.1002/jrsm.1411. PMID 32336025. Open source
8. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. DOI 10.1136/bmj.n71. PMID 33782057. Open source
9. Nguyen C, Coudeyre E, Boutron I, Baron G, Daste C, Lefèvre-Colau MM, et al. Oral resveratrol in adults with knee osteoarthritis: a randomized placebo-controlled trial (ARTHROL). PLoS Med. 2024;21(8):e1004440. DOI 10.1371/journal.pmed.1004440. PMID 39137167. Open source
10. Nguyen C, Boutron I, Baron G, Coudeyre E, Berenbaum F, Poiraudeau S, et al. Evolution of pain at 3 months by oral resveratrol in knee osteoarthritis (ARTHROL): protocol for a multicentre randomised double-blind placebo-controlled trial. BMJ Open. 2017;7(9):e017652. DOI 10.1136/bmjopen-2017-017652. PMID 28965100. Open source