Background
When measurement is costly, fixed sampling schedules can be inefficient because they do not adapt to current uncertainty.
We developed a sequential Bayesian framework for adaptive data-collection design in mechanistic state-space models. It selects the minimum amount of new data required to meet a prespecified precision target.
Poster
Select the preview to open the print-ready PDF. The editable PowerPoint is available above.
Methods
Posterior update and prediction
At each decision time, a particle filtering algorithm updates the joint distribution of latent states and parameters using all observations collected so far. The particle system is propagated forward to obtain a posterior predictive distribution for a future estimand or observable quantity.
Bayesian assurance
For each candidate data-collection intensity, the method computes the predictive probability that the resulting posterior estimator will satisfy a defined precision criterion. The smallest sample size exceeding the assurance threshold is selected for the next observation period.
- Update latent states and parameters
- Propagate particles to the next decision time
- Evaluate candidate collection intensities
- Select the first intensity that reaches assurance q
Results
We demonstrate the approach using Canadian SARS-CoV-2 serological surveillance data. Repeated anti-N and anti-S antibody test results from blood donors inform a compartmental transmission model that estimates population seroprevalence.
Across four regions, adaptive policies reduced the number of blood samples tested for antibodies by 85 to 98 percent relative to observed testing. The median weekly credible interval width for estimated seroprevalence was 0.08. A target width of 0.10 was achieved in 80 to 85 percent of time points.
Conclusion
Sequential Bayesian assurance provides a general approach for converting mechanistic forecasts into adaptive data-collection decisions.
References
- Doucet A, de Freitas N, Gordon N. An introduction to sequential Monte Carlo methods. In Sequential Monte Carlo Methods in Practice. Springer, 2001. Journal link
- Gordon NJ, Salmond DJ, Smith AFM. Novel approach to nonlinear and non-Gaussian Bayesian state estimation. IEE Proceedings F. 1993. Journal link
- Daza-Torres ML, Capistrán MA, Capella A, Christen JA. Bayesian sequential data assimilation for COVID-19 forecasting. Epidemics. 2022. Journal link
- Müller SA, Agweyu A, Akanbi OA, et al. Learning from serosurveillance for SARS-CoV-2 to inform pandemic preparedness and response. The Lancet. 2023. Journal link
- Fosgate GT. Practical sample size calculations for surveillance and diagnostic investigations. Journal of Veterinary Diagnostic Investigation. 2009. Journal link
- King AA, Nguyen D, Ionides EL. Statistical inference for partially observed Markov processes via the R package pomp. Journal of Statistical Software. 2016. Journal link