Spatial Crash Typology and Risk Prediction

2024–2025 · Interpretable ML to classify crash typologies across census tracts and forecast risk

Roadway safety programs usually target individual hotspots: an intersection here, a corridor there. This project asks a different question — do whole neighborhoods share recognizable crash profiles? Using one year of state-level crash records, I classified 2,480 census tracts across Massachusetts, Connecticut, and Vermont into crash typologies with UMAP and Gaussian mixture models.

The results now live in an interactive dashboard at crashtypology.narslab.org, where you can explore the region tract by tract.

Research Objectives

The goal was a data-driven way to understand how crash patterns vary across space, and how they relate to roadway characteristics and socioeconomic conditions. If tracts fall into a small number of recognizable types, agencies can match interventions to types instead of treating every location as unique.

Methodology

Data collection and processing

I assembled one year of crash records from the three states, joined them to their census tracts, and integrated demographic, roadway, and built environment variables for each tract.

Dimensionality reduction

UMAP (Uniform Manifold Approximation and Projection) compressed the high-dimensional crash data into a small set of latent dimensions while preserving the local and global structure that matters for clustering. This kept the downstream typologies interpretable rather than a black box.

Classification and prediction

Gaussian mixture modeling assigned each tract to a crash typology. I then trained gradient boosting models (XGBoost) with SHAP interpretation to predict crash risk within each typology and show which roadway and demographic factors drive risk where.

Key Results

  • The typologies are distinct and interpretable: each one links to a specific combination of roadway characteristics and socioeconomic factors
  • Classifying tracts into types turns a sprawling crash dataset into actionable risk categories that point toward specific interventions
  • The framework only needs crash records and public census data, so other regions can apply it without new data collection

Impact

This work moves safety analysis away from one-size-fits-all approaches and toward context-specific interventions: knowing a tract’s type tells you something concrete about what is likely to work there. It’s a way for agencies to prioritize safety investments systematically rather than one hotspot at a time.

Funding

New England University Transportation Center (NEUTC), 2024-2025 (PI: J. Oke)

Publications

A Roadway Crash Typology of Census Tracts Enables Targeted Interventions via Interpretable Machine Learning (2025)

  • Published in Data Science for Transportation, Volume 7, Article 14
  • doi: 10.1007/s42421-025-00128-2

Conference Presentations

  • “Enhancing Road Safety: A Data-Driven Spatial Typology of Crashes in New England” - INFORMS Annual Meeting, Seattle, WA (October 2024)
  • “A Spatial Typology Analysis of Crash Characteristics across 2480 Census Tracts” - TRB Annual Meeting, Washington, D.C. (January 2025, poster)
  • “A Roadway Crash Typology Enables Targeted Interventions” - NEUTC Symposium & Leadership Summit, Norwich University, VT (February 2025, poster)