Fitting Bayesian Hierarchical Models in Python with PyMC

• Activity Number: 586
• Type: Invited
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Christopher Fonnesbeck*+ and John Salvatier and Thomas Wiecki
• Companies: Vanderbilt University and Amazon.com and Brown University
• Keywords: Bayesian ; Software ; Python ; MCMC ; Computing

Abstract

Advanced statistical computing algorithms, such as newer gradient based MCMC methods, exacerbate computing tradeoffs inherent in making useful statistical software. On one hand, acceptable performance typically requires a compiled language; on the other, the flexibility to easily implement our domain-specific models, custom statistical distributions and algorithms are best satisfied by high-level scripting languages. We will describe how Python currently offers statisticians the sweet spot in this tradeoff, using PyMC as a case study. Python's intuitive syntax is helpful for new users, and has allowed developers to keep the PyMC code base simple, making it easy to extend the software to meet analytic needs. PyMC itself extends Python's powerful "scientific stack" of development tools, which provide symbolic differentiation, fast and efficient data structures, parallel processing, and interfaces for describing statistical models.