As risk assessments move toward new alternative methodologies and away from animal testing, risk assessors have become increasingly reliant on observational epidemiological evidence, including matched case-control, cohort, and cross-sectional studies. Observational data are informative due to their directness to human responses. However, they have limitations in the context of quantitative risk assessment, inherent to the potential for chance, bias, and confounding. To address these issues, we adapt methods for bias assessment in a Bayesian context to fully quantify uncertainties in observational data for specific use in human health risk assessment, including causal analyses. 

 Using simulated annealing, we develop post-hoc methods that use marginal counts reported in the observational study combined with population-based information regarding population distributions of study attributes and the reported statistical estimates to recreate plausible data sets that would reproduce the analysis. By recreating the analysis dataset, bias assessments can be applied to hypothetical datasets consistent with the original analysis.   Through a series of case study applications, the approach’s utility in assessing the magnitude and direction of potential systemic biases related to confounding and exposure-misclassification is demonstrated. As examples, uncertainties in using food frequency questionnaires to evaluate specific dietary components or recall estimates of exposures to a substance will be used to demonstrate how our novel Bayesian bias assessment approaches can improve confidence in determinations of causal relationships and confidence in quantitative dose-response relationships.  This information is then synthesized with further analysis, including evaluation of negative controls and publication bias.  Given the study’s design, this informs the minimum effect necessary to generate a false positive observation that would result in the study’s publication.  It is anticipated that Annealed Bayesian bias assessment methods will be an essential part of a risk assessment toolbox, as quantifying confidence and uncertainty will allow for more informative integration of observational data in decision making.