Why English? | Appraising Non-Randomised Intervention Results With ROBINS-I Confounding, Selection and Reporting Bias

People searching for the ROBINS-I tool, risk of bias in non-randomised studies or confounding in intervention research need a clear causal question before they need a coloured chart. English matters because intervention, eligibility, time zero, outcome and effect of interest must describe the same comparison. Small wording differences can reveal immortal time or adjustment for the wrong variables.

The official ROBINS-I site and its ROBINS-I V2 page, checked on 5 November 2026, state that the revised V2 document posted on 20 November 2025 remains a draft subject to change. It assesses a specific result from a follow-up cohort study and now uses six domains, algorithms and a critical-risk triage. The site also preserves the established 2016 materials.

This article owns non-randomised intervention-result bias appraisal. STROBE owns observational reporting, and RoB 2 owns randomised trials. Every assessment should record which ROBINS-I version was used; mixing instructions silently can change domains, questions and judgements.

– Specify the target trial and effect of interest. – Identify baseline and time-varying confounding. – Align eligibility, intervention classification and time zero. – Inspect selection, missingness, measurement and reporting. – Record the ROBINS-I version and draft status.

---

Confirm the Study and Result Are in Scope

The current V2 draft covers a specific intervention effect result from a follow-up cohort study.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Record V1 or V2 Explicitly

Do not combine the 2016 tool with the November 2025 V2 draft without documenting the choice.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Define the Target Trial

Specify how participants would have been randomised, treated and followed in the ideal comparison.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Align Eligibility and Time Zero

Eligibility, intervention assignment and follow-up should begin coherently to avoid immortal-time distortions.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Choose the Effect of Interest

Distinguish an intention-to-treat effect from a clearly specified per-protocol effect.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

List Important Confounding Factors

Use subject knowledge to identify variables affecting both intervention choice and outcome.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

List Co-Interventions and Post-Baseline Factors

For relevant effects, identify treatment changes and time-varying confounding without indiscriminate adjustment.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Use the V2 Triage Section

A result may go directly to critical risk when fundamental problems make detailed assessment unhelpful.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 1: Identify Baseline Confounding

Compare groups on causes of intervention choice and outcome present before treatment.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 1: Identify Time-Varying Confounding

For per-protocol effects, consider factors influenced by earlier treatment that also affect later treatment and outcome.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 1: Evaluate Measurement of Confounders

Poorly defined or missing confounders cannot be repaired by a sophisticated model name.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 1: Evaluate Adjustment Methods

Match regression, weighting, standardisation or other methods to the causal structure and effect.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 2: Classify Intervention Status

Treatment classification should be accurate, timed correctly and independent of later outcomes.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 2: Detect Immortal Time

Participants must not receive guaranteed event-free time merely to qualify as exposed.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 3: Examine Selection Into the Study

Inclusion after intervention start or because of future information can distort group comparability.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 3: Examine Prevalent Users

Existing users may differ from new starters because early harms and discontinuations are already filtered out.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 4: Quantify Missing Data

Keep missing participant, variable and outcome information separate.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 4: Explain Missingness

Ask whether missing data depend on treatment, prognosis or unobserved outcomes.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 4: Evaluate Analytical Handling

Complete-case analysis, imputation or weighting each require assumptions and diagnostics.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 5: Define the Outcome

Outcome criteria, timing and ascertainment should be appropriate for the target question.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 5: Examine Assessor Knowledge

Awareness of intervention may influence subjective diagnosis or coding.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 5: Compare Measurement Across Groups

Differential surveillance or record quality can create apparent effects.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 6: Find the Prespecified Analysis Plan

A dated plan narrows opportunities to select the most favourable result.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 6: Check Multiple Measurements

Several definitions, time points or instruments can enable selective reporting.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Domain 6: Check Multiple Analyses

Many adjustment sets or model forms create a result-selection problem.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Use Signalling Algorithms Transparently

Record responses, algorithm proposals and reasons for overrides.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Make Domain and Overall Judgements

Use low, moderate, serious, critical or no information and avoid numerical averaging.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Interpret Low Risk Conservatively

Low ROBINS-I risk means comparable to a well-performed randomised trial for the domain, not proof of no residual confounding.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Did You Know? V2 Is Still a Draft

The official site says the 20 November 2025 V2 document remains subject to change.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Keep ROBINS-I Separate From RoB 2

RoB 2 is for randomised trials; ROBINS-I addresses non-randomised intervention effects.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Help Students Explain Confounding

Write a causal sentence showing how a factor influences both treatment choice and outcome.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Help Families Read Associations Safely

A carefully adjusted association can still differ from the effect an individual would experience.

Build the target-trial comparison

ROBINS-I asks how a non-randomised result compares with a well-performed randomised trial addressing the same causal question. Define the intervention strategies, eligible population, outcome, time zero, follow-up and effect of interest before judging the study. This controlled language exposes immortal time, prevalent-user designs and mismatched eligibility or treatment starts.

Evidence-linked signalling judgement

Use reports, protocols, analysis plans, registries and correspondence where available. Record each signalling response and the passage supporting it. The current V2 draft adds algorithms and a triage route to critical risk, but an algorithm proposes rather than replaces judgement. If reviewers override a proposed result, they should preserve the reason and the document version used.

Current-status boundary

As checked on 5 November 2026, the official ROBINS-I site labels the 20 November 2025 V2 document a draft subject to change and scopes it to follow-up cohort studies. The established judgement language is low, moderate, serious, critical or no information. Date the tool version, avoid mixing V1 and V2 instructions, and never treat a non-randomised association as automatically causal.

---

Frequently Asked Questions

What does ROBINS-I stand for?

Risk Of Bias In Non-randomized Studies—of Interventions.

What does it assess?

Risk of material bias in a specific intervention-effect result.

What is the target-trial idea?

The non-randomised study is compared conceptually with a well-performed randomised trial addressing the same question.

Is V2 final?

No. The official site says the 20 November 2025 V2 document is a draft subject to change.

What is V2 currently scoped to?

Follow-up cohort studies.

What are the established judgement categories?

Low, moderate, serious, critical and no information.

Why is confounding central?

Factors may affect both intervention choice and outcome, creating an association that differs from the intervention effect.

What is immortal time?

A period during which an outcome cannot occur for someone to be classified in an intervention group.

How does ROBINS-I differ from RoB 2?

ROBINS-I is for non-randomised intervention studies; RoB 2 is for randomised trials.

Does adjustment prove causality?

No. Credibility depends on design, measurement, assumptions and unmeasured confounding.

---

Useful Next Reading

• Use the official ROBINS-I resources.
• Check the current ROBINS-I V2 draft status.
• Read observational reports with STROBE.
• Keep adjusted estimates complete with SAMPL.

Closing thought: ROBINS-I is most useful when the causal question, target trial, confounders, evidence sources, tool version and reasons for every judgement remain visible.