Risk-of-bias assessment examines how study methods may distort a result used in a review. Choose the tool and version for the design and question: RoB 2 for randomized intervention trials, ROBINS-I for non-randomized intervention effects, and QUADAS tools for diagnostic accuracy. QUADAS-3 is now the Bristol team’s recommended version. Keep each judgment tied to its evidence and to the result or estimate being assessed.
Use the working tables below to record the result being assessed, the chosen tool and version, source support, reviewer judgments and resolution of disagreement. They help organize the work but do not replace the official questions, decision rules or guidance for that tool.
Why Risk of Bias Assessment Matters in Systematic Reviews
Risk of bias is about systematic error in a study result, not about whether authors wrote the paper clearly, and not about random error or imprecision. A study can be precise but biased, or imprecise but methodologically sound. Cochrane distinguishes bias from both imprecision and external validity, and PRISMA separates reporting of study risk of bias from reporting bias due to missing results in a synthesis.
That distinction matters because modern RoB tools are designed to assess the result being used in the review, not merely a study “as a whole.” RoB 2 explicitly focuses on a specific result from a randomized trial. ROBINS-I does the same for a specific result from a non-randomized intervention study, and QUADAS-3 now moves diagnostic accuracy appraisal toward estimate-level assessment as well. This result-level orientation is one of the biggest practical shifts in contemporary evidence synthesis, because one outcome or timepoint from the same study may be at low risk of bias while another is not.
For intervention reviews, risk-of-bias assessment is a core part of the methods described in the Cochrane Handbook. PRISMA 2020 asks authors to report the assessment methods and findings; completing that reporting task does not validate the judgments. GRADE considers risk of bias when assessing certainty for an outcome across the body of evidence.
The assessment should help explain limitations in the included results and inform the planned synthesis. For example, a team may restrict an analysis, compare results across risk-of-bias groups or run a sensitivity analysis. Explain the rationale and report the consequences rather than assuming that a completed template prevents biased conclusions.
Which Risk of Bias Tool Should You Use
Match the tool to both the study design and the question. Then record the exact version, variant and effect of interest. The working tables here follow RoB 2 for individually randomized parallel-group trials, ROBINS-I (2016) and QUADAS-2. These structures are not interchangeable with newer versions or with tools for reviews, missing evidence or prediction models.
Concise comparison table
| Tool | Best use case | Structure or scope | Decision process | Output |
|---|---|---|---|---|
| RoB 2 (22 August 2019) | Individually randomized parallel-group trials | Randomization; deviations; missing outcome data; outcome measurement; reported result selection | Full signaling questions and algorithms | Low risk / Some concerns / High risk |
| RoB 2 cluster variant | Cluster-randomized trials | Includes cluster-specific identification and recruitment issues | Use the official cluster tool, revised 18 March 2021 | Low risk / Some concerns / High risk |
| RoB 2 crossover variant | Crossover randomized trials within the tool’s scope | Includes crossover-specific issues such as period and carryover effects | Use the official crossover tool, revised 18 March 2021 | Low risk / Some concerns / High risk |
| ROBINS-I (2016) | Non-randomized intervention studies using this version | Seven domains, including deviations from intended interventions | Signaling questions and judgment guidance; not the V2 algorithms | Low / Moderate / Serious / Critical / No information |
| ROBINS-I V2 (November 2025 draft) | Follow-up cohort studies within the draft’s scope | Six domains with revised questions and ordering | Draft algorithms; consult the exact release and its guidance | Version-specific proposed judgments; still subject to change |
| QUADAS-2 | Projects whose documented methods use this earlier version | Patient selection; index test; reference standard; flow and timing. Applicability for the first three | Tailored questions and domain judgments | Low / High / Unclear for each relevant judgment |
| QUADAS-3 (v1.2 checked 18 October 2026) | Current recommended QUADAS tool for test accuracy reviews | Estimate-level assessment, ideal test accuracy trial and revised domains | Use the current official tool and guidance | Domain judgments plus a formal overall assessment |
| ROB-ME | Missing evidence in a synthesis with meta-analysis | Missing studies and missing results relevant to the synthesis | Structured assessment distinct from individual-result appraisal | Risk of bias due to missing evidence |
| ROBIS | Risk of bias in systematic reviews | Review-level assessment of the review process | Structured phases and questions | Review-level risk of bias |
| PROBAST family | Prediction-model development or evaluation | PROBAST-2019 uses participants, predictors, outcome and analysis domains; PROBAST+AI updates and extends that tool | Consult the selected version’s official guidance | Version-specific quality, risk-of-bias and applicability assessments |
| Newcastle–Ottawa Scale | Cohort and case-control appraisal where this scale is specified | Selection, comparability and exposure/outcome | Star-based item assessment | Stars; not interchangeable with RoB 2 or ROBINS-I judgments |
Tool sources: RoB 2 and its variants, ROBINS-I (2016), ROBINS-I V2 draft, QUADAS-2, QUADAS-3, ROB-ME, ROBIS, PROBAST and Newcastle–Ottawa Scale. Version information was checked on 18 October 2026.
For a non-randomized intervention review, choose and document the ROBINS-I version before starting. The official site still labels the November 2025 V2 release as a draft; its six-domain structure is not the seven-domain 2016 table below. For a new diagnostic accuracy review, consult QUADAS-3. If an existing protocol uses QUADAS-2, explain that choice and do not silently relabel its assessments.
Downloadable and Printable Risk of Bias Assessment Template
Two simplified CSV record sheets remain available: the RCT sheet follows the five-domain RoB 2 structure of 22 August 2019 for individually randomized parallel-group trials, and the NRSI sheet follows the seven-domain ROBINS-I (2016) structure. Enter the exact version and result identifier before using them. Their prompts are reminders, not the complete official questions or an automated assessment.
For diagnostic accuracy assessments, use the official QUADAS-3 tool and guidance for a new project, or the official QUADAS-2 materials where that version is specified in the protocol. The earlier mixed-version DTA download and combined workbook have been withdrawn from this page because their QUADAS-2 fields were also labeled as QUADAS-3.
The embedded working tables below preserve the distinction between evidence notes and completed judgments. Their fictional fragments illustrate what to record; they do not contain enough source evidence to establish domain or overall judgments. Complete the official assessment and retain it alongside these summary records.
Master fields to include in every template
| Field | What to enter |
|---|---|
| Review ID | Review name or registration identifier |
| Study and report IDs | Unique study identifier and each supporting report, registry or supplement |
| Study design | Design features needed to select the tool and variant |
| Result identifier | Outcome, time point, comparison, analysis population, effect measure and result location |
| Tool and version | Exact name, release date/version, design variant and effect of interest |
| Reviewer 1 and reviewer 2 | Separate initial responses and judgments, with dates |
| Supporting evidence | Quotation or evidence note with page, table, registry field or correspondence date |
| Domain judgment | The selected tool’s category, kept separate from the supporting prose |
| Disagreement and resolution | Original decisions, unresolved issue, resolution and person responsible |
| Overall judgment or summary | Use the selected tool’s overall rules; label a narrative summary separately |
| Notes | Missing information, author contact, assumptions and links to the full official assessment |
Template for randomized controlled trials
RoB 2 has five mandatory domains for individually randomized parallel-group trials. Its full signaling questions feed algorithms proposing domain judgments of Low risk, Some concerns or High risk. Apply the overall rules as well: several domains with Some concerns can justify High risk overall when they substantially lower confidence in the result. Document reasons for any override of an algorithm proposal. A short evidence fragment cannot determine the judgment by itself.
| Section | Working field | Illustrative entry, not an assessment |
|---|---|---|
| Study details | Review ID | Diabetes-SR-2026 |
| Study details | Study ID / citation | Fictional case A |
| Study details | Outcome / timepoint | HbA1c at 24 weeks |
| Study details | Effect of interest | Effect of assignment to intervention |
| Domain 1 | Bias arising from the randomization process | To assess with the full official tool |
| Domain 1 support | Illustrative evidence fragment; add source location | Allocation sequence computer-generated; central allocation concealment |
| Domain 2 | Bias due to deviations from intended interventions | To assess with the full official tool |
| Domain 2 support | Illustrative evidence fragment; add source location | Participants unblinded; limited evidence of differential co-interventions |
| Domain 3 | Bias due to missing outcome data | To assess with the full official tool |
| Domain 3 support | Illustrative evidence fragment; add source location | 11% attrition; reasons partially balanced but outcome related to dropout plausible |
| Domain 4 | Bias in measurement of the outcome | To assess with the full official tool |
| Domain 4 support | Illustrative evidence fragment; add source location | Lab-based HbA1c measured identically across groups |
| Domain 5 | Bias in selection of the reported result | To assess with the full official tool |
| Domain 5 support | Illustrative evidence fragment; add source location | No publicly accessible SAP; multiple timepoints mentioned |
| Overall | Overall RoB 2 judgment | To assess with the full official tool |
| Overall | Likely direction of bias | To assess where supported |
Enter the exact result and effect of interest, then complete the full official RoB 2 tool. Transfer the judgment and its supporting source location into the record. For cluster or crossover trials, use the corresponding official variant instead of adding a homemade question to the parallel-group worksheet.
ROBINS-I (2016) working table
This seven-domain table follows ROBINS-I (2016). Define the target trial, causal contrast, important confounders and relevant co-interventions before assessment, using methodological and subject expertise. The V2 draft changes the domains and questions; this table does not implement it.
| Section | Working field | Illustrative entry, not an assessment |
|---|---|---|
| Study details | Review ID | Asthma-SR-2026 |
| Study details | Study ID / citation | Fictional case B |
| Study details | Target trial specification | Adults with uncontrolled asthma initiating biologic A vs biologic B |
| Study details | Outcome / timepoint / estimate | Exacerbations at 12 months; adjusted HR |
| Pre-intervention | Bias due to confounding | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Disease severity and eosinophil count incompletely adjusted |
| Pre-intervention | Bias in selection of participants into the study | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | New-user design used, but index date capture imperfect |
| At intervention | Bias in classification of interventions | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Pharmacy and prescribing records concordant |
| Post-intervention | Bias due to deviations from intended interventions | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Differential add-on therapy possible after start of follow-up |
| Post-intervention | Bias due to missing data | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Outcome capture via claims database near-complete |
| Post-intervention | Bias in measurement of outcomes | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Hospitalized exacerbation objectively recorded |
| Post-intervention | Bias in selection of the reported result | To assess with ROBINS-I (2016) guidance |
| Support | Illustrative evidence fragment; add source location | Multiple adjustment sets reported; analysis plan unavailable |
| Overall | Overall ROBINS-I judgment | To assess with ROBINS-I (2016) guidance |
Start with the result and target trial, then use the 2016 guidance to assess confounding, selection, intervention classification, deviations, missing data, outcome measurement and reported-result selection. Its categories are Low, Moderate, Serious, Critical and No information. A Critical assessment indicates that the result is too problematic to provide useful evidence for synthesis under that guidance; retain the assessment and explain how it affects the planned analysis. Do not generate that judgment from an isolated example note.
QUADAS-2 working table for reviews using that version
The table below follows QUADAS-2 and is intended for projects whose documented methods use that version. QUADAS-3 is the current version recommended by the Bristol QUADAS team; its official page listed v1.2 when checked on 18 October 2026. For a new review, consult its official tool and guidance. Changing a heading does not convert a QUADAS-2 assessment into QUADAS-3.
| Section | Working field | Illustrative entry, not an assessment |
|---|---|---|
| Study details | Review ID | Sepsis-DTA-2026 |
| Study details | Study ID / citation | Fictional case C |
| Study details | Index test | Plasma biomarker X |
| Study details | Reference standard | Clinical diagnosis by adjudication committee |
| Study details | Accuracy estimate assessed | Sensitivity and specificity at prespecified threshold |
| Risk of bias | Patient selection | To assess: Low / High / Unclear |
| Support | Illustrative evidence fragment; add source location | Case-control sampling with exclusions of indeterminate cases |
| Risk of bias | Index test | To assess: Low / High / Unclear |
| Support | Illustrative evidence fragment; add source location | Threshold prespecified before analysis |
| Risk of bias | Reference standard | To assess: Low / High / Unclear |
| Support | Illustrative evidence fragment; add source location | Adjudication may have partial awareness of biomarker data |
| Risk of bias | Flow and timing | To assess: Low / High / Unclear |
| Support | Illustrative evidence fragment; add source location | Not all participants received the same reference standard; delayed verification |
| Applicability | Patient selection applicability | To assess: Low / High / Unclear |
| Applicability | Index test applicability | To assess: Low / High / Unclear |
| Applicability | Reference standard applicability | To assess: Low / High / Unclear |
| Summary | Narrative summary of domain findings | Complete after assessment; not a formal QUADAS-3 overall judgment |
QUADAS-2 assesses risk of bias in patient selection, index test, reference standard, and flow and timing, with applicability assessed for the first three domains. Its judgment labels are Low, High and Unclear; Some concerns belongs to a different tool’s vocabulary. Tailor the official questions and guidance to the review before appraisal. Keep any prose summary distinct from QUADAS-3’s formal overall assessment.
How to Apply the Template and Turn It Into Better Evidence Synthesis
Plan the tool, version, relevant results and reviewer process before assessment. In the report, describe the methods actually used, including whether reviewers worked independently and how disagreements were handled. Keep the full assessments accessible so the reported judgments can be traced back to evidence.
For each judgment, retain the result identifier, outcome, time point, tool version, supporting source location, reviewer and resolution of disagreement. Keep a domain judgment separate from a prose summary so the original reasoning remains available after export.

Structured assessment still requires judgment. Apply the selected version’s questions and decision rules, record missing information and explain the reasoning. Where the tool provides algorithm-generated proposals, preserve both the proposal and the reason for any override. Do not assume that ROBINS-I (2016) uses the algorithms introduced in V2.
Common pitfalls to avoid
The most common error is assessing the study in general instead of the specific result used in the review. A second frequent error is mixing up risk of bias with reporting quality, certainty, or applicability. A third is inventing a numeric total score from multiple domains. Official tools are domain-based and designed to preserve the mechanism of bias rather than collapse diverse problems into a single homemade score. For diagnostic reviews, the parallel mistake is to ignore applicability concerns even though QUADAS explicitly separates risk of bias from applicability.
For the ROBINS-I (2016) workflow shown here, define the target trial, important confounders and co-interventions before reviewing the example fields. Use the different instructions supplied with V2 if that is your selected version. Agreement between reviewers cannot compensate for an undefined causal question.
For cluster and crossover trials, use the appropriate official RoB 2 variant. Identify recruitment timing and other cluster-specific issues, or period and carryover issues, as directed by that tool. Also document how those designs were handled in the analysis; a correct tool label does not resolve a unit-of-analysis problem.
Tips for better inter-rater reliability
Calibrate on a small, varied set of results before the full assessment. Have reviewers work independently, compare their evidence and interpretations, and resolve recurring ambiguities in a written decision guide. Choose enough pilot cases to cover the designs and difficulties in the review; there is no universal required number.
Follow the selected tool’s response options and branching instructions. When reviewers disagree, check whether they used the same result, source passage and version before debating the final category. Preserve the original responses as well as the resolved judgment.
Good reliability also depends on a clean evidence trail. Store the quotation, registry entry, protocol excerpt, supplement, or author correspondence that supports each judgment. That is consistent with Cochrane’s insistence on “support for judgment” and it makes arbitration faster, more transparent, and much easier to defend during peer review.
How to incorporate RoB into GRADE and meta-analysis
RoB should directly influence synthesis, not sit in an appendix. Cochrane Chapter 7 recommends several strategies when risks of bias vary across studies: restrict the primary analysis to low-risk studies; stratify or subgroup analyses by overall RoB or specific domains; present all studies but explicitly account for RoB in certainty judgments; or, less commonly, use statistical bias adjustment methods. Cochrane does not currently recommend formal weighting schemes that mechanically assign study weights based on RoB categories, because those methods are not sufficiently developed for routine review use.
Prespecify how risk-of-bias judgments affect the main analysis and sensitivity analyses, using the applicable methods guidance. A restriction, stratified analysis or sensitivity analysis may be appropriate, but an all-studies analysis followed by a low-risk subset is not a universal default. Explain which results enter each analysis. A non-significant difference between risk-of-bias groups does not demonstrate absence of bias.
Carry the result-level judgments and their reasons into the synthesis plan and certainty assessment. A risk-of-bias label alone does not determine a universal numerical downgrade or an analysis weight. Explain how the limitations affect the particular outcome and use the selected framework’s guidance for the body of evidence. Cochrane Chapter 14 describes the connection to GRADE.
Assess missing evidence at the synthesis level as well. Individual-result tools do not replace an assessment of results or studies missing from the synthesis. ROB-ME addresses that separate question for systematic reviews with meta-analysis.
A plot can help readers connect judgments to the results being synthesized. robvis produces traffic-light and bar plots for supported risk-of-bias formats. Verify the expected tool version and category labels before import. A plot displays the judgments supplied to it; it does not validate them.
When the assessed results enter an analysis, use a meta-analysis data and reproducibility check to retain their identity, settings and inclusion decisions.
Conclusion
The useful output is a traceable assessment for the result being used. Select the right tool and version, complete its official questions, and retain the source evidence and reviewer decisions.
Use the working tables to carry those decisions into reporting, synthesis and certainty assessment without losing their identifiers or reasons. Keep uncertainty explicit and leave judgments uncompleted until the evidence and assessment support them.
References
Sterne JAC and colleagues. Cochrane Handbook Chapter 25, including ROBINS-I (2016) guidance. See the separate official original version and V2 draft release for version-specific materials.
Page MJ, McKenzie JE, Bossuyt PM, et al. PRISMA 2020 statement and checklist resources.
ROB-ME resources from Cochrane Bias.
McGuinness LA, Higgins JPT. robvis: risk-of-bias visualization tool and package.