ROBINS-I V2 (2025): Explicit Cut-off Lines for Risk-of-Bias Judgement
- Mayta
- Feb 4
- 3 min read
Link to ROBIN-I V2 Website A Domain-by-Domain Operational Guide
Core Principle of ROBINS-I V2
ROBINS-I V2 is anchored on the identification of material bias:
Material bias = bias that is large enough to meaningfully distort the estimated effect or invalidate causal interpretation.
All cut-offs below are therefore defined by impact on causal interpretability, not by the mere presence of bias.
Domain 1 — Bias Due to Confounding
Low Risk (Low except for residual confounding)
Cut-off line
All important confounders are identified, measured with acceptable validity, and appropriately controlled using suitable methods (e.g., regression, matching, weighting).
Any remaining confounding is plausibly too small to materially alter the effect estimate.
Note: In non-randomized studies, “pure” Low is rarely achievable; hence Low except for residual confounding.
Moderate Risk
Cut-off line
Most important confounders are controlled.
Any uncontrolled confounding is unlikely to be substantial or to change the direction or magnitude of the effect in a meaningful way.
Supporting evidence (e.g., sensitivity analyses, negative controls) suggests robustness.
Serious Risk
Cut-off line
≥1 important confounder that should have been controlled was not controlled, and
There is strong a priori or empirical rationale that this failure has a material impact on the estimated effect.
Key discriminator: Bias likely changes the size of the effect, but the effect is still interpretable with caution.
Critical Risk
Cut-off line
Confounding is so severe that the observed comparison cannot be interpreted as a causal contrast, or
The study effectively compares groups defined by prognostic severity or indication that directly determines the outcome, without adequate control, or
Preliminary ROBINS-I V2 screening question B2 = “Yes” (confounding too severe to proceed).
At this level, results should not be synthesized as evidence of effect.
Domain 2 — Bias in Classification of Interventions
Low Risk
Cut-off line
Intervention groups are clearly distinguishable at the start of follow-up.
Classification is independent of outcome information.
Moderate Risk
Cut-off line
Some misclassification exists but is non-differential and unlikely to materially bias the effect estimate.
Serious Risk
Cut-off line
Differential misclassification related to outcome occurrence, timing, or prognosis.
Examples include exposure defined using post-baseline information (e.g., immortal time not addressed).
Critical Risk
Cut-off line
Misclassification is so severe that the intended intervention contrast no longer exists in practice.
Groups do not represent the interventions as defined.
Domain 3 — Bias in Selection of Participants
Low Risk
Cut-off line
Participants are included from the start of intervention or exposure.
No selection based on post-intervention variables.
Moderate Risk
Cut-off line
Some post-intervention selection occurs, but sensitivity analyses indicate minimal impact.
Serious Risk
Cut-off line
Selection into analysis depends on post-intervention variables associated with both intervention and outcome.
Typical example: inappropriate per-protocol or responder-only analyses.
Critical Risk
Cut-off line
Selection produces irreversible structural bias (e.g., survivor-only cohorts, missing entire exposure periods).
Bias cannot be corrected analytically.
Domain 4 — Bias Due to Missing Data
Low Risk
Cut-off line
Outcome data are nearly complete (>90–95%), or
Missingness is plausibly MCAR/MAR and handled with appropriate methods (e.g., multiple imputation).
Moderate Risk
Cut-off line
Missing data exist, but sensitivity analyses show no meaningful change in effect estimates.
Serious Risk
Cut-off line
Missingness is plausibly MNAR and related to the true outcome value.
Analysis relies primarily on complete-case methods without correction.
Critical Risk
Cut-off line
Large amounts of outcome data are missing and strongly outcome-dependent, such that the direction or existence of the effect is uncertain.
Domain 5 — Bias in Measurement of Outcomes
Low Risk
Cut-off line
Outcomes are objective or
Outcome assessors are blinded and measurement methods are identical across groups.
Moderate Risk
Cut-off line
Assessors are unblinded, but outcomes are minimally influenced by judgment.
Serious Risk
Cut-off line
Outcomes are subjective and
Assessment is unblinded, with plausible influence of knowledge of intervention on measurement.
Critical Risk
Cut-off line
Outcome measurement methods differ fundamentally between groups or
Measurement is so flawed that outcomes are not comparable.
Domain 6 — Bias in Selection of the Reported Result
Low Risk
Cut-off line
Results are reported according to a pre-specified protocol or SAP.
Moderate Risk
Cut-off line
No pre-registration, but there is little or no analytical flexibility and no indication of selective reporting.
Serious Risk
Cut-off line
Evidence of selective reporting from multiple analyses, scales, time points, or models without prespecification.
Critical Risk
Cut-off line
Selective reporting is so extreme that conclusions would likely reverse if all analyses were reported.
Overall Risk of Bias (Outcome-Specific)
Cut-off rules
Low: All domains Low
Moderate: ≥1 Moderate, none Serious/Critical
Serious: ≥1 Serious, none Critical
Critical: ≥1 Critical or multiple Serious domains rendering the effect non-credible
Final Calibration Rule (Across All Domains)
Serious = biased but still interpretable with cautionCritical = causal interpretation no longer defensible
This distinction—not terminology—is the decisive cut-off line in ROBINS-I V2.
Comments