Conference Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 111 - Application and Development of Statistical Methods for Spatio-Temporal Data
Type: Contributed
Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 AM
Sponsor: Section on Bayesian Statistical Science
Abstract #323432
Title: Using Dirichlet Processes and Machine Learning to Estimate Crash Risk on Roadways
Author(s): Benjamin K. Dahl* and Matthew Heaton and Richard Warr and Philip White and Grant G. Schultz and Caleb Dayley
Companies: Brigham Young University and BYU and Brigham Young University and BYU and Brigham Young University and Brigham Young University
Keywords: Point pattern; Poisson process; Log-Gaussian Cox process; Bayesian nonparametrics; Dirichlet process; Traffic

Historically, specifying models for point pattern data has had to balance flexibility with interpretability. On the one hand, mixture model specifications for Poisson process intensity surfaces can flexibly capture the non-linear nature of the intensity surface, but do not yield interpretable regression parameters. On the other hand, log-Gaussian Cox processes can give interpretable regression coefficients for the intensity surface but can be computationally costly to implement. In this project we provide a partial solution to this balancing act by using Dirichlet processes to flexibly model an intensity surface for a Poisson process. We then project the resulting Dirichlet process fit onto a set of basis functions using penalized regression to obtain an estimate of a corresponding log-Gaussian Cox process fit. We demonstrate this process by estimating the intensity surface and associated effects of roadway characteristics on the frequency of crashes along I-15 in Utah from 2019-2020.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2022 program