the IARC confounding bias group: David B Richardson; Sadie Costello, Sander Greenland, Jay S Kaufman, Kaitlin Kelly-Reif, Kyle Steenland, and Eric J Tchetgen Tchetgen

Abstract

Objective. Present methods to detect and evaluate confounding bias in epidemiologic studies. Methods. Confounding is bias that arises when the exposure and outcome share a common cause. Confounding can lead to spurious associations away from or toward the null. Potential confounders can be identified a priori by looking at the literature, to see what others have identified as confounders. Confounders must not be confused with intermediate variables which are on the pathway between exposure and disease. Results. Once identified, potential confounders can be controlled in the study design by matching (in either case–control or cohort studies), by stratifying on them in the analysis, or by their inclusion in a model where the outcome is regressed on exposure and potential confounders. Reviewers will need to assess how well potential confounders have been controlled. If there is likely uncontrolled confounding, a reviewer should assess the direction and magnitude of possible confounding bias. This can be done by 1) using negative control outcomes or negative control exposures, 2) triangulation, ie., considering different types of evidence from studies with different designs, or 3) quantitative bias adjustment, using a priori knowledge of the effect of a confounder on outcomes, as well as the likely prevalence of the exposure among exposed and non-exposed, to judge the likely effect of the confounder. Conclusion. In some studies, investigators will have measured potential confounders and controlled for them in the design or analysis. Reviewers will want to consider whether residual confounding is likely to be present. If there are potentially important uncontrolled confounders, reviewers will want to assess the likely impact of such confounding.

of the IARC selection bias group: Neil Pearce, Laura Beane Freeman, Manolis Kogevinas, Richard MacLehose, Ellen Aagaard Nohr, Marie-Elise Parent, Lorenzo Richiardi

Abstract

Objective: Present methods to detect and evaluate selection bias in epidemiologic studies. Methods: In epidemiological studies, there is usually a target population to which we wish to draw inferences. The study is then conducted in a source population followed over a clearly-defined period of time (the risk period). There is also a (perhaps smaller) study population, i.e. the group of people who actually take part in the study. Selection bias refers to differences between the source population and the study population. In cohort studies, important mechanisms for selection bias include non-response at baseline, loss-to-follow-up, left truncation, right truncation, and immortal person-time bias. In case-control studies, all of these biases are possible, but bias can also occur due to inappropriate selection of controls. Results: Qualitative methods for assessing the existence, direction, and strength of selection bias, include the use of negative control exposures, negative control outcomes, ad hoc re-analyses of published data, comparisons with external data, and the use of multiple control groups. Quantitative methods also exist for sensitivity analyses in which “adjustments” are made for hypothesized selection bias. Conclusions: One of the primary questions posed to reviewers involved in an IARC Monograph is, “Can we reasonably rule out selection bias as an explanation for an observed exposure-cancer association?” This can be particularly difficult to assess, since most published studies have little or no discussion of the potential for selection bias. It is therefore important that authors, and editors, are encouraged to report the information that is required for a valid assessment of the potential, direction and strength of possible selection bias.

Andrew Kunzmann, Elisa Pasqual

Abstract

Objective: Summarize the framework for incorporating bias appraisal for observational epidemiology studies into evidence synthesis for cancer hazard identification, as recommended in the recently published IARC volume. Describe how bias assessment has been incorporated into recent Monographs evaluations. Methods: The recommended framework for incorporating bias assessments into evidence synthesis focuses on identification of key biases among informative studies, including direction and (where possible) magnitude, assessing their impact on study estimates, and triangulating findings among studies with different sources of bias. Scenarios are described for situations with few vs. many informative studies. We examined use of the bias assessment approaches laid out in the book for confounding, information bias, and selection bias in Monographs Volumes 131–138 (evaluations conducted in 2022–2025). Results: Each of the eight volumes considered has incorporated bias assessment tools into the evidence synthesis. For example, in an evaluation of antimony (Vol. 131), an assessment of bias resulting from co-exposure to arsenic among smelter workers led to a conclusion that arsenic could not entirely explain the observed lung cancer risk. In evaluations of occupational exposure as a firefighter (Vol. 132) and several pharmaceuticals (Vol. 137), meta-analyses stratified by major sources of bias were used in evidence triangulation. In the evaluation of PFOA (Vol. 135), information bias from single-timepoint exposure measurement was quantified using repeated-measurement data. For acrylonitrile (Vol. 136), externally conducted bias adjustment for healthy worker survivor bias was important for the determination of sufficient evidence for lung cancer. Conclusions: Incorporation of bias appraisal tools into evidence synthesis has strengthened recent IARC Monographs cancer hazard identification, giving tools to inform and complement Working Groups’ expert judgment. Acknowledgements: The framework for incorporating bias assessments into evidence synthesis is from a chapter co-authored by Amy Berrington de González, Nathan DeBono, Alexander Keil, Deborah Lawlor, Ruth Lunn, and David Savitz.