Designed & builtR Shiny · railway safety engineering

HIGHWAY-RAIL GRADE-CROSSING DECISION SUPPORT

Crash Predictor at HRGCs

An interactive safety application that turns a competing-risk survival model into a practical comparison of crossing controls, geometry, crash occurrence, and severity over time.

Launch the app U.S. DOT reportSee the interface ↓
Crossing-risk comparisonComplete application recording · shown without cropping or zoom
29 yearscumulative prediction horizon
28control-device options
4 outcomescrash · PDO · injury · fatal
R Shinyinteractive model delivery

01 / PRODUCT OVERVIEW

Compare safety consequences before choosing a treatment.

Crossing-prioritization methods often compress risk into a single score or study crash frequency and severity separately. That can hide important tradeoffs.

This application exposes the fitted survival relationships directly. A user can change the warning-device combination and crossing geometry, then inspect how predicted crash occurrence and severity evolve—not only whether one scenario receives a higher or lower rank.

My roleTechnical Project Lead · Lead Developer · Key ResearcherTechnical workflow, competing-risk analysis, safety interpretation, application logic, R Shiny development, and decision-support delivery.

02 / USER WORKFLOW

One continuous path from crossing inputs to a severity-aware comparison.

The interface keeps assumptions visible and connects every input to multiple complementary outputs.

  1. 01

    Select the outcome

    Choose total crash occurrence or a severity-specific outcome: property-damage-only, injury, or fatal crash likelihood.

  2. 02

    Build two scenarios

    Select a reference and an alternative from 28 control-device configurations for a direct safety comparison.

  3. 03

    Set crossing geometry

    Adjust roadway lanes, crossing-to-intersection distance, and the acute highway-rail crossing angle.

  4. 04

    Choose the horizon

    Move the year control across the 29-year model horizon to inspect short- and long-term differences.

  5. 05

    Interpret the change

    Read cumulative-probability curves, a year-specific bar comparison, and severity-share pie charts together.

03 / MODEL-TO-APPLICATION PIPELINE

Survival modeling translated into an inspectable decision system.

Competing risks means that property-damage-only, injury, and fatal outcomes are treated as mutually exclusive event types whose relationships should not be analyzed in isolation.

Crossing data01

Assemble longitudinal safety records

North Dakota crossing inventory, crash, control-device, traffic, train-service, and geometric information form a common time-to-event dataset.

Survival model02

Represent competing crash outcomes

A competing-risk framework with Cox proportional-hazard regression models crash occurrence and severity using a shared set of predictors.

Scenario engine03

Recalculate cumulative likelihood

The application applies fitted relationships to user-selected device and geometry combinations across the forecast horizon.

Decision view04

Compare consequences, not labels

Linked curves, bars, and severity shares reveal when a countermeasure produces different effects across outcomes or time.

04 / APPLICATION INTERFACE

Inputs and predictions stay visible together.

Figures reproduced from the U.S. DOT report show the linked input and prediction panels and the year-specific scenario controls.

HRGC crash predictor with prediction panel and crossing input panel
Prediction panel: cumulative curves, year-specific bars, and severity shares. Input panel: outcome, two device scenarios, lanes, distance, and angle.
HRGC crash predictor severity, device, angle, and year controls
Scenario controls and a year-16 comparison of crossbucks against crossbucks plus stop signs.

05 / VERIFIED MODEL EVIDENCE

The interface makes nonlinear and outcome-specific effects visible.

1.61% → 3.64%PDO likelihood by year 16

The report’s example compares crossbucks-only with crossbucks plus stop signs and shows a 2.03-percentage-point modeled difference.

5% → 12%modeled crash likelihood

Across the studied conditions, increasing roadway lanes from one to four raised the reported cumulative crash likelihood.

3.3% → 21%modeled crash likelihood

Increasing main tracks from one to three produced a substantial modeled increase, illustrating why geometry cannot be reduced to a generic crossing label.

06 / DECISION VALUE & LIMITS

A comparison tool for screening—not a universal treatment rule.

The report demonstrates that adding a traffic-control device does not automatically improve every safety outcome. Effectiveness depends on the existing control configuration, crossing geometry, time horizon, and the outcome being evaluated.

The model was estimated from North Dakota data. The authors recommend before-and-after evaluation, interaction-effect analysis, and life-cycle cost analysis before a treatment is selected for a real crossing.

Read the complete methodology and limitations

TRY THE PRODUCT

Explore a crossing scenario.

Open the live R Shiny application or review the complete U.S. DOT research report.

Launch the app U.S. DOT report