← All systems

Ped & Bike Danger Score

Methodology

Data Sources

NHTSA FARS
Fatality Analysis Reporting System — all fatal motor vehicle crashes in the US
Format: csv · Source
EPA National Walkability Index
Block-group level walkability index, population density, and built-environment measures
Format: csv · Source
California SWITRS
Statewide Integrated Traffic Records System — California fatal and injury collisions
Format: csv · Source

Units of Analysis

Sites (Intersection-Level)

Crashes are assigned to spatial bins using H3 hexagonal cells at resolution 10 (~15m edge length), approximating intersection-level grouping. When H3 is not available, a 0.001-degree grid (~111m) is used as a fallback. Each bin with at least one pedestrian or bicycle crash becomes a scored site.

Corridors

Sites within approximately 500 meters of each other are grouped into corridor segments. Corridor scores aggregate the site-level metrics to identify stretches of road with systemic danger patterns rather than isolated incidents.

Feature Engineering

Crash Severity Score

Each crash is weighted by KABCO severity: Fatal (5.0), Serious Injury (3.0), Minor Injury (1.5), Possible Injury (0.5), No Injury/PDO (0.1). The sum gives each site a severity-weighted crash history.

Crash Frequency

Annual crash rate: total crashes divided by years with recorded crashes. Captures persistent danger patterns vs one-time events.

Walkability / Pedestrian Exposure

EPA National Walkability Index (1–20) at the census block group level, joined to crash sites via spatial binning. Higher walkability implies greater likely pedestrian exposure, increasing the population at risk.

Equity Index

Rank-based composite of population density and walkability index as proxies for transit-dependent and underserved populations. Sites scoring higher serve communities with fewer transportation alternatives.

School & Transit Proximity

When local data is available: distance to nearest school (within 400m buffer) and aggregate transit stop frequency (within 200m buffer). These increase the vulnerability weight of a site.

Composite Risk Score

Each feature is min-max normalized to [0, 1] across all sites, then combined using a weighted sum. The default weights are:

Feature Weight
Crash Severity 0.35
Crash Frequency 0.15
Walkability Exposure 0.15
Equity Index 0.15
School Proximity 0.10
Transit Proximity 0.10

Peer-Group Frontier Estimation

In addition to the composite score, the system applies the shared PDL peer-frontier scoring method. Sites are grouped by state, and for each peer group the 25th percentile crash severity defines the "attainable frontier" — the level achievable by the better-performing quarter of peers. Sites are then ranked by recoverable impact: the gap between their severity and the frontier, multiplied by their crash frequency. This prioritizes sites where intervention would yield the largest real-world improvement.

Portfolio Optimizer

Given a fixed budget, the optimizer selects the set of intervention sites that maximizes total risk reduction. Each site's intervention cost is estimated based on its risk level (ranging from $50K speed reduction to $350K raised intersection). Expected risk reduction is modeled as 30% of current risk weighted by crash frequency.

The optimizer uses a greedy benefit-cost ratio heuristic: sites are ranked by risk-reduction-per-dollar and selected in order until the budget is exhausted. This provides a practical approximation to the 0-1 knapsack problem.

Rolling Backtest

The model is validated using a rolling temporal backtest:

  1. Train on all crash data from years < t
  2. Generate predicted danger rankings for year t
  3. Compare against actual crash sites in year t
  4. Repeat for each available test year

Metrics: Precision@100 (fraction of top 100 predicted sites that had actual crashes), Recall@100 (fraction of actual crash sites captured), and overall hit rate. Results are compared against a naive hotspot baseline that ranks sites only by raw crash count.

Limitations

  • FARS only records fatal crashes; non-fatal ped/bike crashes are underrepresented unless SWITRS is loaded
  • Spatial binning approximates true intersection locations; local centerline data would improve precision
  • Walkability is joined at coarse (~1km) grid resolution due to lack of point-level data
  • Equity index uses proxy measures; CDC SVI would be more rigorous
  • Intervention costs are estimated using national averages, not local construction costs
  • The 30% risk reduction assumption is a rough average across intervention types