Bookshelf

RCT = randomized clinical trial

Details on categories, definitions, and items can be found in Sterne JA, Hernan MA, Reeves BC, et al. ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions. The BMJ. 2016;355:i4919. doi:10.1136/bmj.i4919. Note that the first 3 categories of biases presented in the Table occur before or at the time of the intervention or exposure. The remaining categories of biases occur after the intervention. Adapted table used with permission.

Bias arising from design occur before or at the intervention and include bias due to confounding, selection, and classification of interventions. Bias arising from conduct occur after the intervention and can arise from departures from intended interventions, missing data, measurement of outcomes, and reporting.

Bias arising from design for RCTs arise principally from randomization flaws, as contrasted with other designs that have multiple potential sources of bias before or at intervention. Bias arising from conduct for RCTs are similar to bias for other designs.

lncludes nonrandomized controlled studies with investigator-allocated treatment and observational studies of prospective or retrospective cohorts with comparison arms

Relevant only for nonrandomized experimental studies where the investigator allocates treatment

Cases and controls should be similar in all factors known to be associated with the disease of interest, but they should not be so uniform as to be matched for the exposure of interest.

Refers to biases that are internal to the study only, and does not refer to issues of applicability (e.g., restricting the sample to a specific clinical population). Selection bias results when the study design results in a biased estimate of the effect because the design of the study resulted in the exclusion of some participants or their data. For example, studies that evaluate the effect of folic acid supplementation on neural tube on live births only selectively exclude outcomes from pregnancies resulting in fetal deaths. Selection bias can also arise in retrospective studies that do not have complete data for all potential participants at inception or do not restrict their design to “naïve” drug users – by design, these designs potentially exclude eligible participants.

Although we do not expect selection bias to occur routinely in trials, informative censoring in trials with different baseline times could potentially result in selection bias.

This category is relevant only when the review is evaluating the effect of starting and adhering to interventions.

There are no established rules for determining a threshold for appropriate completeness of outcome data. Reviewers should establish what is meant by “Reasonably complete” based on the specific topic and outcome.

Blinding of outcome assessors is especially important with subjective outcome assessments.

Reviewers do not need to evaluate inferential statistics used in studies that report results in a manner that permits meta-analyses or other independent analyses. When reviewers need to rely solely on the results as presented by authors, they may elect to review the use of inferential statistics in the study.