Extending ROBINS-I for assessment of studies reporting instrumental variable analyses

ID: 

2056

Session: 

Poster session 2 Thursday: Evidence synthesis - methods / improving conduct and reporting

Date: 

Thursday 14 September 2017 - 12:30 to 14:00

Location: 

Exhibition Halls 1 & 2

All authors in correct order:

Davies N1, Didelez V2, Elbers R1, Higgins J1, Jackson J3, McAleenan A1, Small D4, Swanson S5, Tchetgen Tchetgen E6, Sterne J1
1 University of Bristol, United Kingdom
2 University of Bremen, Germany
3 Johns Hopkins Bloomberg School of Public Health, USA
4 University of Pennsylvania, USA
5 Erasmus MC, Netherlands
6 Harvard School of Public Health, USA
Presenting author and contact person

Presenting author:

Jonathan Sterne

Contact person:

Jonathan Sterne
Abstract text
Background: Systematic reviews should assess risk of bias in included studies, in order to draw conclusions about the strength of the evidence for causal effects of interventions on health outcomes. The ROBINS-I (Risk Of Bias In Non-randomised Studies - of Interventions) tool provides a structured approach to assessing risk of bias in non-randomised studies of interventions (NRSI) based on comparisons with a hypothetical high-quality pragmatic randomised trial. Answers to signalling questions lead to judgements of risk of bias within bias domains and overall. The published tool and guidance mainly focus on studies with a cohort-type design. Instrumental variable (IV) analyses, which can estimate causal effects of interventions in the presence of unmeasured confounding of the intervention-outcome association but require other assumptions, are not currently addressed.

Objectives: Adapt the ROBINS-I tool, including its signalling questions and accompanying guidance, to assess risk of bias in studies reporting IV analyses.

Methods: An international working group of experts in IV methodology met via teleconferences and face-to-face. The group agreed changes to the ROBINS-I tool through consensus.

Results: We added two new bias domains to ROBINS-I. The first (replacing the ROBINS-I confounding domain) assesses the core IV assumptions that 1) the IV is associated with the intervention; 2) there is no residual confounding of the IV-outcome association; and, 3) the effect of the IV on the outcome is via the intervention. The second assesses the plausibility of the additional ‘point identifying’ assumptions required for the validity of different types of IV estimate. In addition we adapted signalling questions within existing ROBINS-I bias domains, such as the ‘deviations from intended intervention’ domain.

Conclusions: The extended ROBINS-I tool contains signalling questions and guidance to assess risk of bias in studies reporting IV analyses. These will now be piloted, with feedback used to inform further modifications.