Abstract
Bayesian phylodynamic models have revolutionized epidemiology by enabling researchers to infer key aspects of the geographic history of disease outbreaks. These models contain many parameters that must be estimated from minimal information (the area from which each pathogen was sampled), rendering inferences under this approach inherently sensitive to the choice of priors on the model parameters. Here, we demonstrate that: (1) the priors assumed in ≈ 93% of surveyed phylodynamic studies make strong and biologically unrealistic assumptions, and; (2) these priors distort the conclusions of epidemiological studies. We offer strategies and tools to specify more reasonable priors that will enhance our ability to understand pathogen biology and, thereby, to mitigate disease.
Introduction
Phylodynamics has transformed epidemiology by integrating genetic data with spatial information to reconstruct the geographic history of pathogen dispersal. By analyzing the "family tree" (phylogeny) of viral samples alongside their collection locations, researchers can infer the pathways through which diseases spread. However, these models are inherently "data-poor" regarding geography; while a viral genome contains thousands of informative sites, the geographic data for each sample is typically a single discrete location. Consequently, Bayesian phylodynamic inferences are highly sensitive to the "priors"—the initial assumptions researchers provide before data analysis.
The Prevalence of Default Priors
A comprehensive literature survey of 749 published empirical studies revealed a concerning trend: approximately 93% of studies rely on the "default" geographic priors provided by popular software like BEAST. These defaults are often perceived as neutral or "uninformative," but this study demonstrates they are actually strongly informative in ways that bias results toward overly simplistic dispersal histories.
Biologically Unrealistic Assumptions
The default priors assume that pathogens move through the absolute minimum number of routes possible, regardless of the outbreak's duration or the number of geographic areas involved. For instance, the default settings often imply an average of only 0.5 dispersal events, even when a virus has spread across ten different countries—a scenario that logically requires at least nine dispersal events.
As shown in Figure 4, the default prior distributions for the number of dispersal routes (Δ) are heavily skewed toward zero. This "minimalist" bias forces the model to favor a sparse history of spread, which can lead to high-certainty but incorrect conclusions about the source of an outbreak.
Empirical Impact on Outbreak Reconstruction
The researchers re-analyzed 14 published datasets, including SARS-CoV-2, Influenza (H3N2, H5N1), and Dengue virus. In every case, alternative priors that were more diffuse and scaled naturally with outbreak duration were decisively preferred over the defaults (Bayes factors > 10).
The choice of prior significantly altered conclusions regarding outbreak origins:
- Washington State COVID-19: Default priors suggested a 90% probability that the virus originated from elsewhere in North America. Alternative priors revealed a more complex picture, with nearly equal likelihoods of origin from China (~39%) or North America (~36%).
- Brazilian COVID-19 Spread: The default model identified São Paulo as the source of 84.6% of domestic spread. The alternative model showed São Paulo accounted for only 19.5%, highlighting the major roles of five other regions that were previously obscured.
Conclusion and Recommendations
The sensitivity of geographic inference to prior assumptions is the rule, not the exception. To mitigate these biases, the authors have developed PrioriTree, an interactive tool designed to help researchers visualize and specify biologically reasonable priors. Moving forward, the epidemiological community must move away from "black-box" defaults and adopt rigorous sensitivity testing to ensure that our understanding of disease dynamics is driven by data rather than unrealistic mathematical assumptions.
Summary updated