ROBINS-E: Risk of Bias in Non-Randomised Studies of Exposures

ROBINS-E (“Risk Of Bias In Non-randomised Studies – of Exposures”) is the standard instrument for judging internal validity in a single non-randomised study of an exposure effect — a cohort study, or a comparable design, estimating the causal effect of an exposure (environmental, occupational, dietary, behavioural, or similar) on an outcome. It was developed by Higgins, Morgan, Rooney, Taylor, Thayer, Silva and colleagues and published as “A tool to assess risk of bias in non-randomized follow-up studies of exposure effects (ROBINS-E)” (Environment International 2024;186:108602). It shares its structure and much of its vocabulary with ROBINS-I, the equivalent tool for non-randomised studies of interventions, and with RoB 2 for randomised trials — the three are best understood as one family applied at different points on the randomisation spectrum, a distinction covered at the synthesis level in Risk of Bias Assessment: RoB 2, ROBINS-I and the Traffic-Light Plot. This guide focuses on what ROBINS-E adds and where it genuinely diverges from ROBINS-I, not a repeat of that synthesis-level material.

ROBINS-E vs ROBINS-I: exposures, not interventions

The distinction sounds like terminology, but it changes how you apply the tool. ROBINS-I assumes an intervention: something decided and administered at a definable point in time — a drug started, a policy implemented, a procedure performed — where a comparator arm could, in principle, have been randomly assigned instead. ROBINS-E assumes an exposure: often a continuous or ongoing state (air pollution level, occupational chemical exposure, dietary pattern, screen time) that a person does not switch into at a single decision point the way they start a drug, and that frequently could not be randomised even hypothetically at the individual level for ethical or practical reasons.

That difference shows up directly in two of the seven domain names. Where ROBINS-I asks about classification of interventions and deviations from intended interventions, ROBINS-E asks about classification of exposures and departures from intended exposures — the same structural slot in the tool, but built around measuring and defining a state rather than confirming a discrete action was performed as assigned. Everything else in the two tools maps closely, which is deliberate: ROBINS-E’s developers built it as a sibling to ROBINS-I and note it is informing ROBINS-I’s own further development, not a competing framework.

Start with the target experiment, not the study you have

Like ROBINS-I’s target trial, ROBINS-E asks you to specify — before touching the seven domains — the causal effect the result under consideration is actually estimating. Concretely, write down:

• The exposure contrast being estimated (e.g., continuous exposure per unit increase, or a defined high-vs-low contrast), specific enough that it could in principle have been randomly assigned in a hypothetical experiment.
• Eligibility and the point in time (“time zero”) at which a participant’s exposure status and follow-up begin, so exposure that happens before eligibility isn’t silently mixed with exposure that happens after.
• The outcome and follow-up period the hypothetical experiment would measure.
• The confounders and co-exposures expected to differ by exposure level for reasons other than the exposure itself, and that also affect the outcome — listed before reviewing Domain 1’s evidence, not derived after the fact from whatever the data happen to show.

Doing this step first is what keeps a ROBINS-E assessment from becoming a vague description of “the study’s limitations” — every domain judgement that follows is made relative to this hypothetical experiment, not against a generic standard.

The seven domains

1. Bias due to confounding. Whether the exposure groups differ on factors, other than the exposure itself, that also affect the outcome. For most environmental and occupational exposures, this is the dominant domain — exposure is rarely close to random, and confounding by socioeconomic status, geography, or co-exposure is common and often only partially measured.
2. Bias in selection of participants into the study (or into the analysis). Whether inclusion depended on both exposure and a factor related to the outcome (e.g., healthier or more health-conscious people being differentially retained at high exposure levels — a pattern related to, though not identical to, healthy-worker-type selection).
3. Bias arising from measurement of the exposure. Whether the exposure was measured accurately, consistently across groups, and without knowledge of the outcome. Exposure misclassification is frequently the single largest source of ROBINS-E-flagged bias, especially for self-reported or recall-based exposure measures, or biomarkers with a short half-life relative to the exposure window of interest.
4. Bias due to post-exposure interventions or departures from intended exposures. Whether something that happened after exposure assignment — a co-intervention, a behaviour change prompted by awareness of the exposure, exposure switching — differs by exposure group in a way that affects the outcome through a path other than the exposure itself.
5. Bias due to missing data. Whether the outcome or covariate data actually analysed differ systematically from the complete data because of how missingness relates to both exposure and outcome.
6. Bias arising from measurement of the outcome. Whether outcome ascertainment was equally accurate and equally blinded to exposure status across groups.
7. Bias in selection of the reported result. Whether the specific result presented was selected, from among multiple analyses that were performed or could have been performed, on the basis of the finding.

The judgement scale and direction of bias

For each domain, and then overall, ROBINS-E produces three things, not one: a risk-of-bias level (Low, Some concerns, High, or Very High), a predicted direction of bias (does the domain plausibly push the estimate away from the null, toward the null, or is the direction unpredictable), and a judgement on whether that risk of bias is severe enough to threaten the paper’s conclusions about whether the exposure has an important effect on the outcome. Recording direction alongside severity is what lets a systematic review author reason about whether biases across included studies are likely to reinforce or offset each other at the synthesis stage — a severity rating alone can’t support that reasoning.

When to reach for ROBINS-E instead of ROBINS-I

Use ROBINS-E when the causal question is framed around an exposure that a participant is in, rather than an action performed on or by them at a defined moment: air quality and respiratory outcomes, occupational chemical exposure and long-term health, dietary pattern and disease risk, screen time and developmental outcomes. Use ROBINS-I when the study compares an intervention against a comparator — a treatment, a programme, a clinical procedure — something that could have been assigned by a coin flip in an ethical trial. Some topics sit close to the boundary (a preventive behaviour someone chooses to adopt can look like either); when that happens, the deciding question is whether your target experiment would plausibly randomise participants to a single decision at a point in time (ROBINS-I) or to an exposure level or contrast sustained over a period (ROBINS-E). For appraising a systematic review’s overall use of these tools, rather than a single included study, see Choosing a Critical Appraisal Tool for Your Study Design; for the closely related concept of specifying a causal question against a hypothetical randomised design before assessing an observational study, see Target Trial Emulation: Applying RCT Design Principles to Real-World Data.

Frequently asked questions

Is ROBINS-E a Cochrane tool?

ROBINS-E was developed by a large international working group including researchers associated with the Cochrane and NTP/EPA-adjacent risk-of-bias methodology community, in the same lineage as RoB 2 and ROBINS-I, and is hosted publicly at riskofbias.info alongside those tools. It is not itself a Cochrane-branded tool the way RoB 2 is, but it is designed to be used the same way — within a systematic review of observational exposure studies — and its authors state it is informing the ongoing development of ROBINS-I.

Can I use ROBINS-E for a single study outside a systematic review?

Yes — the domain structure and target-experiment framing are equally useful for critiquing or designing one standalone exposure study. The tool’s guidance materials are written primarily with systematic review use in mind, but nothing in the seven domains requires a review context.

What is the difference between ROBINS-E’s domains and ROBINS-I’s domains?

Five of the seven domains are identical in substance: confounding, selection into the study, missing data, measurement of the outcome, and selection of the reported result. The two that differ are worded around exposure rather than intervention — “classification of exposures” instead of “classification of interventions,” and “departures from intended exposures” instead of “deviations from intended interventions” — reflecting that an exposure is usually an ongoing state to be measured rather than a discrete action to be confirmed.

Does a “Very High” risk-of-bias judgement mean the study should be excluded from a review?

Not automatically. ROBINS-E is designed to inform, not replace, the review team’s judgement about whether to exclude a result, downweight it in a narrative or quantitative synthesis, or report it separately with the risk-of-bias judgement made explicit — the same principle GRADE applies when risk of bias lowers certainty in a body of evidence rather than eliminating a single study outright.