Expand support for spatial models in PyMC
- Mentors
- Chris, Bill Engels
- Organization
- NumFOCUS
- Technologies
- python, Pytensor
- Topics
- Probabilistic Programming, Bayesian, Spatial statistics
This project would implement the Besag, York, Mollie (BYM) model in PyMC. BYM is a highly flexible model for studying spatial data and is used widely in epidemiology, agriculture, and ecology. The strategy behind the BYM model is to treat the outcome of interest as the result of three factors: some set of predictor variables, spatial covariance between neighboring regions, and random effects that represent non-spatial heterogeneity.
Although very flexible, the BYM model can be difficult to specify in a way that is simultaneously computationally efficient, interpretable, and identifiable. Recently, Morris et al (2019) demonstrated an alternative specification of the BYM model that is significantly more efficient, interpretable, and can be identified with Monte Carlo Markov Chain (MCMC) samplers.
Developing an implementation in Python with PyMC would make the model more accessible to a greater variety of users. Furthermore, BYM models are only available in a Bayesian framework. As a leading Bayesian statistics package, PyMC should support this extremely useful model in spatial statistics.