New England Crash Typology Dashboard

Interactive dashboard mapping five kinds of crash-prone places across 2,664 New England census tracts

Overview

An interactive companion to my crash typology research. We analyzed more than 222,000 crashes across the census tracts of Massachusetts, Connecticut, and Vermont and used unsupervised machine learning to uncover five recurring crash environments. Knowing a tract’s type points directly to the safety interventions most likely to work there — and this dashboard lets you explore that, tract by tract.

Live dashboard: crashtypology.narslab.org

What It Shows

  • An interactive map of all 2,664 census tracts, colored by crash type or by measures such as crash counts, severity, or truck involvement, with a full crash profile for every tract
  • The five crash types themselves — from fast rural to slow urban environments — uncovered by compressing 25 crash risk features into six latent dimensions with UMAP and clustering with Gaussian mixture models
  • Each type’s fingerprint across the six dimensions, alongside SHAP values from a 96%-accurate LightGBM classifier showing which dimensions the model relies on when assigning tracts
  • 33 evidence-based safety countermeasures mapped to the crash types they address, so selecting a type surfaces its matched interventions
  • Temporal patterns, including crash counts by hour and weekday

The larger point is transferability: any region with crash records can map its own crash-prone environments the same way and match them to cataloged countermeasures, moving safety investment from one-spot-at-a-time toward systematic, typology-guided planning.

Technical Stack

  • Frontend/visualization: React single-page app with Leaflet for the interactive map
  • Data: Tract-level GeoJSON with typology assignments, crash measures, and per-tract profiles
  • A Roadway Crash Typology of Census Tracts Enables Targeted Interventions via Interpretable Machine Learning — Data Science for Transportation, 7, 14 (2025)

Funding and Collaborators

Funded by the US Department of Transportation via the New England University Transportation Center (NEUTC). Built at NARSLab, UMass Amherst, with Prof. Jimi Oke (PI).