Background Many population health research questions rely on observational data, where unmeasured confounding is a major source of bias. Sensitivity analyses for unmeasured confounding are increasingly applied, but often without sufficient consistency and transparency. We propose accessible recommendations to guide applied researchers in using two existing sensitivity analyses. 1) Bias Factor (BF), which is derived from the expected strength of associations between unmeasured confounder and exposure/outcome, based on expert knowledge and previous research. The main effect estimate (and confidence intervals, CIs) are adjusted using the BF. 2) E-value (EV), which identifies the strength of associations between unmeasured confounder and exposure/outcome required to entirely attenuate the main effect estimate (or for CIs to contain the null)
Methods We conducted a scoping review for commentaries and reviews discussing the application, strengths, and limitations of the BF and EV. We triangulated these with epidemiological guidance (e.g. STROBE) and informal discussions with quantitative researchers in applied statistics, epidemiology and social policy.
Results The BF was criticised for the potential for authors to selectively pick confounder associations that minimally impact the results. The EV removes the potential for author bias and future-proofs analyses (as knowledge of confounders advances). However, it potentially discourages authors’ rigorous and transparent consideration of unmeasured confounding; and places burden upon the reader to judge whether this degree of confounding would seem feasible. Furthermore, population research typically aims to estimate an effect size (not merely the existence of an effect, which is the focus of the EV). Initial recommendations. Unmeasured confounders are identified at protocol stage. A range of exposure/outcome associations are identified for the confounder(s), from systematic reviews, high-quality individual studies, and expert opinion. At publication stage: 1) the full range of BFs are applied to the main effect size and CIs, reported in full, and the most pertinent highlighted in the discussion; 2) The EV, for the main effect and CIs, is compared with best estimates derived using the BF, observed confounders-exposure/outcome associations, and effect sizes for other important exposure/outcome risk factors; 3) The importance of the effect size after considering potential residual confounding should be assessed; 4) Results are discussed in context of other threats to bias, including measurement error among measured confounders (as applied in primary studies and systematic reviews).
Conclusion These simple recommendations, supported by a real-life research example, can improve sensitivity analyses for unmeasured confounding and reduce the potential for selective reporting, thereby improving the quality of population health research.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.