Background

Systematic reviews, which assess the risk of bias in included studies, are increasingly used to develop environmental hazard assessments and public health guidelines. These research areas typically rely on evidence from human observational studies of exposures, yet there are currently no universally accepted standards for assessing risk of bias in such studies. The risk of bias in non-randomised studies of exposures (ROBINS-E) tool has been developed by building upon tools for risk of bias assessment of randomised trials, diagnostic test accuracy studies and observational studies of interventions. This paper reports our experience with the application of the ROBINS-E tool.

Methods

We applied ROBINS-E to 74 exposure studies (60 cohort studies, 14 case-control studies) in 3 areas: environmental risk, dietary exposure and drug harm. All investigators provided written feedback, and we documented verbal discussion of the tool. We inductively and iteratively classified the feedback into 7 themes based on commonalities and differences until all the feedback was accounted for in the themes. We present a description of each theme.

Results

We identified practical concerns with the premise that ROBINS-E is a structured comparison of the observational study being rated to the 'ideal' randomised controlled trial. ROBINS-E assesses 7 domains of bias, but relevant questions related to some critical sources of bias, such as exposure and funding source, are not assessed. ROBINS-E fails to discriminate between studies with a single risk of bias or multiple risks of bias. ROBINS-E is severely limited at determining whether confounders will bias study outcomes. The construct of co-exposures was difficult to distinguish from confounders. Applying ROBINS-E was time-consuming and confusing.

Conclusions

Our experience suggests that the ROBINS-E tool does not meet the need for an international standard for evaluating human observational studies for questions of harm relevant to public and environmental health. We propose that a simpler tool, based on empirical evidence of bias, would provide accurate measures of risk of bias and is more likely to meet the needs of the environmental and public health community.