Science Communication & Media
Leveraging semi-Bayesian data methods to communicate study results that inform policy Simone Wien* Simone Wien Timothy L. Lash Melvin Livingston Whitney S. Rice Hannah L.F. Cooper Michael R. Kramer
Policymakers often informally incorporate their prior beliefs about a policy’s effect when presented with new evidence from study data. This semi-formal Bayesian approach can conflict with the frequentist approach often used to guide inference. This discrepancy may make it difficult for epidemiologists to anticipate how policymakers will interpret study results. For example, a policy advocate may interpret protective results that are not statistically significant as confirmation that a policy improves health, while a policy skeptic may view those results as evidence that the policy is ineffective.
We adapt Greenland’s semi-Bayesian data augmentation method to demonstrate how study results can be interpreted by different policymakers and to aid epidemiologists in developing arguments for varying prior beliefs. Using the effect of a policy on receipt of prenatal care as an example, we constructed priors reflecting different policymaker views and calculated their updated beliefs (posteriors) given hypothetical study data indicating a small but harmful policy effect (RD -5, 95% CL -7, -3 per 100 pregnancies). Per the method, priors and study data were operationalized as weighted estimates and pooled in lieu of a formal version of Bayes’ theorem to calculate the posterior distribution.
This approach helps develop arguments tailored to a policymaker’s posterior distribution (Figure 1). For instance, while Policymaker 3 may remain unconvinced that the policy is harmful given their prior belief in its protective effect, their posterior suggests that they may be more open to a discussion of how there is insufficient evidence that the policy is protective (RD 0, 95% CL -1.4, 1.4). Policymaker 1’s posterior indicates that the data are sufficient to convince them that the policy is harmful; Policymaker 2 and 4’s priors are reinforced given the data. We argue that this method helps epidemiologists frame results for policy audiences, including manuscript discussion sections.
