ROBINS-I V2 tool

We are pleased to announce a revised version of Version 2 of the ROBINS-I tool, posed on 20 November 2025. Note that this is still a draft version and is subject to change.

ROBINS-I V2 Nov 2025 full (PDF document) or ROBINS-I V2 Nov 2025 template  (editable document)

The Risk Of Bias In Non-randomized Studies – of Interventions, Version 2 (ROBINS-I V2) aims to assess the risk of bias in a specific result from an individual non-randomized study that examines the effect of an intervention on an outcome. The document describes the ROBINS-I V2 tool for follow-up (cohort) studies. Please note that refinements to this version are likely as we respond to feedback on this release.

The development group for ROBINS-I V2 was led by Jonathan Sterne and Julian Higgins. 

Summary of key changes from ROBINS-I (2016) to ROBINS-I V2 (Nov 2025)

• A key change is the addition of algorithms, which map answers to signalling questions onto proposed risk-of-bias judgements for each domain. A lot of the signalling questions have been revised to make them work better with these algorithms.

• Some signalling questions now have more levels of response option: we introduced the idea of a “strong” yes (or no) versus a “weak” yes (or no) to replace the single "(Probably) yes" (or "no") response – these then lead to different risk-of-bias judgements.

• A new ‘triage’ section (part B) provides a quick mapping to 'Critical risk of bias' when this is appropriate.

• We refined specification of the effect of interest. We now determine that the study is estimating the per protocol effect if the analysis accounts for protocol deviations. 

• We dropped pre-specification of important co-interventions.

• We dropped the domain on "deviations from intended intervention"

• Domain 1 (confounding): 

  • We improved the table for evaluation of confounding factors

  • The domain is split into two variants: one for addressing the intention-to-treat effect (when only baseline confounding needs to be addressed); and one for addressing the per-protocol effect (when both baseline and time-varying confounding need to be addressed)

• Domain 2 (classification of intervention):

  • Renumbered from domain 3 to domain 2

  • We added new questions to address classification bias arising from immortal time

• Domain 3 (selection into the study):

  • Renumbered from domain 2 to domain 3

  • We added new questions to address selection bias arising from immortal time

• Domain 4 (missing data)

  • Renumbered from domain 5 to domain 4

  • We have reconceived this domain and much expanded it as a result of extensive discussions with missing data experts

• Domain 5 (measurement of the outcome)

  • Renumbered from domain 6 to domain 5

• Domain 6 (selection of reported result)

  • Renumbered from domain 7 to domain 6

  • We added a question on the availability of a pre-specified analysis plan

Archived version from November 2024

ROBINS-I V2 Nov 2024 (editable document) or ROBINS-I V2 Nov 2024  (PDF document)

Summary of key changes from ROBINS-I (2016) to ROBINS-I V2 (Nov 2024)

• A key change is the addition of algorithms, which map answers to signalling questions onto proposed risk-of-bias judgements for each domain. A lot of the signalling questions have been revised to make them work better with these algorithms.

• Some signalling questions now have more levels of response option: we introduced the idea of a “strong” yes (or no) versus a “weak” yes (or no) to replace the single "(Probably) yes" (or "no") response – these then lead to different risk-of-bias judgements.

• A new ‘triage’ section (part B) provides a quick mapping to 'Critical risk of bias' when this is appropriate.

• We refined specification of the effect of interest. We now refer to “a per protocol effect” rather than “the per protocol effect” and require specification of what that per-protocol effect is.

• We dropped pre-specification of important co-interventions

• Domain 1 (confounding): 

  • We improved the table for evaluation of confounding factors

  • The domain is split into two variants (one for addressing baseline confounding only; and one for addressing both baseline and time-varying confounding)

• Domain 2 (classification of intervention):

  • We moved this domain up (it used to be Domain 3)

  • We added new questions to address classification bias arising from immortal time

• Domain 3 (selection into the study):

  • We moved this domain down (it used to be Domain 2)

  • We added new questions to address selection bias arising from immortal time

• Domain 4 (deviations from intended intervention)

  • The domain is now split into two variants (one when interest is in the intention-to-treat effect; and one for when interest is in a per-protocol effect)

  • We expanded the first variant (ITT effect) expanded so that it matches the RoB 2 tool

  • We dropped the question about co-interventions when addressing a per-protocol effect

• Domain 5 (missing data)

  • We have reconceived this domain and much expanded it as a result of extensive discussions with missing data experts

• Domain 7 (selection of reported result)

  • We added a question on the availability of a pre-specified analysis plan

Development of ROBINS-I  V2  was funded in part by Medical Research Council (MRC) grant MR/M025209/1. Development involved members of the Cochrane Bias Methods Group and the Cochrane Non-Randomised Studies Methods Group.