Prevention & care

Bias in health research: selection, measurement, missing results and uncertain direction

10 sources · Latest source check 2026-10-20 UTC

On this page

Bias is a systematic departure from the intervention effect a sound comparison would estimate. The RoB 2 framework evaluates a particular trial result through several domains. Calling a study “randomized” does not settle every domain or every outcome.[1]

Different mechanisms can threaten different results

Different mechanisms can threaten different results
Mechanism	Illustrative concern	Interpretation limit
Confounding	Treatment groups differ in causes of the outcome	Adjustment depends on relevant information and assumptions[2]
Selection	Entry or inclusion depends on relevant post-intervention factors	Study size does not itself remove selection problems[2]
Missing outcome data	Observed participants may differ from those missing	Amount alone does not settle the bias judgment[1]
Outcome measurement	Assessment differs with knowledge of treatment	The outcome and measurement process matter[1]
Selective reporting	A result is selected from several possible analyses	A published result may not represent all results[4]

A risk-of-bias judgment concerns a result and its question

RoB 2 separates randomization, deviations from intended interventions, missing outcomes, outcome measurement and selection of the reported result. Different outcomes or analyses from one trial can therefore need different judgments. Its signaling questions structure assessment, rather than replace methodological reasoning. Incomplete reporting and actual flawed methods are related but different issues; uncertainty about what happened should be identified. A journal label or general reputation cannot substitute for examining the result.[1]

Confounding is one mechanism rather than a synonym for all bias

The original ROBINS-I framework compares an observed intervention study with the hypothetical trial it seeks to emulate. It includes confounding, participant selection, intervention classification, deviations, missing data, measurement and reporting. A measured association can be distorted through several pathways, including ones that ordinary baseline adjustment does not solve. The framework’s target comparison and assumptions should be specified. It does not turn observational data into actual randomization or prove that a large adjusted analysis has removed every important difference.[2]

Assessment instruments have versions and stated limits

The developers’ November 2025 ROBINS-I V2 page identifies a revised draft for specific results from follow-up cohort studies. It describes changed domains, signaling questions and algorithms, and says refinements remain likely. That version should not be silently treated as identical to the 2016 instrument or as a final universal tool for every observational design. A report of a formal assessment needs the instrument version and rationale. This guide explains concepts without claiming to have performed that assessment.[3]

Missing evidence can concern a whole study or one result

ROB-ME addresses bias when entire studies or particular results are unavailable to a meta-analysis because of their P values, direction or magnitude. This is different from every instance of incomplete data within a trial. A search that finds many published reports can still miss unfavorable or unreported results. The tool structures judgments about the missing-evidence mechanism rather than assuming that every unavailable report has a known finding. Funding and disclosure information can prompt scrutiny, but an allegation alone cannot quantify this mechanism or replace result-specific assessment.[4]

A known tendency does not determine one observed result

Jurek and colleagues distinguish bias, defined across repeated studies, from an individual estimate that also varies through random error. Their methodological analysis shows why nondifferential exposure misclassification alone does not guarantee an association is underestimated. Even when additional conditions imply average bias toward no association, a particular observed estimate can overshoot because of chance. “The error affects both groups” is therefore insufficient to declare a result conservative. Bias direction and random uncertainty need separate reasoning tied to the actual measurement process.[5]

Three meanings behind missing data

Missing-data labels concern why a value is unavailable, not simply how many records are missing. In these terms, “random” has a statistical meaning.[10]

Mechanisms depend on relationships with the data
Term	Meaning
Missing completely at random (MCAR)	Missingness is unrelated to observed or unobserved values.[10]
Missing at random (MAR)	After conditioning on observed variables, missingness no longer depends on the unobserved value.[10]
Missing not at random (MNAR)	Missingness still depends on unobserved values after accounting for observed variables.[10]

Evidence & uncertainty

The large MetaBLIND analysis did not detect an average treatment-effect difference according to blinding status. This null comparison does not establish that blinding is irrelevant in every trial; the authors identify methodological uncertainty and retain its importance.[8]

Selected-source overview · Independent review pending

Published questions

Does randomization eliminate every risk of bias?

Selected-source overview · Independent review pending. Read the evidence context.

No. RoB 2 also considers deviations, missing outcomes, measurement and result selection. Assessment concerns a specified result rather than the design label alone.[1]

Evidence behind this answer

RoB 2: a revised tool for assessing risk of bias in randomised trials
original risk of bias framework · 2019
Original indexed abstract and selected original indexed passages inspected; direct original page unavailable or challenge blocked; full text not independently inspected · Systematic deviation, result-level assessment, five domains and reporting-versus-conduct judgments