01 / Activity
Exposure combines road and rail traffic.
Exposure is the reported vehicles per day multiplied by the reported trains per day. It is an activity index, not a count of encounters or an estimate of crashes.
Daily trains are the sum of daytime through trains, nighttime through trains and switching movements, following the Atlas train total. Passenger counts are not added again. If any part is missing or invalid, the total and exposure are unavailable. A reported zero remains zero.
The road and train counts may come from different years. A new inventory download does not make those counts new.
02 / Position
Percentiles put a value in context.
We compare exposure with all crossings that have a known exposure in the same published inventory, then with those in the same state. We also calculate separate state percentiles for road traffic and daily trains.
Percentile = 100 × (number below + half the number equal) ÷ number with a known value.
This is the midrank method. For values of 10, 20, 20 and 30, both 20s have a percentile of 50. The 10 is at 12.5 and the 30 is at 87.5. Missing values are excluded. A group with one known value has a percentile of 50.
This example uses the four values above.
Exposure index
20
This value sits at the middle of the group after ties are shared.
How this is calculatedA percentile is a relative position. A value at 90 is near the top of its comparison group; it does not mean a 90% chance of an incident. With ties, it is also not the exact percentage of crossings with a lower value.
03 / Record quality
Confidence reflects freshness and completeness.
The score starts at 100. We subtract the fixed deductions below and round once to the nearest whole number. Every score includes reasons with a code, a plain-English explanation and the points deducted. These are Atlas rules, with no machine learning.
| Input | Maximum deduction | Rule |
|---|---|---|
| Road count age | 25 points | Lose 1.25 points per year, up to 25 points after 20 years. |
| Train count age | 20 points | Lose 2 points per year, up to 20 points after 10 years. |
| Key fields | 20 points | Lose 5 points each for missing or invalid road traffic, daily trains, warning device class and track count. |
| Inventory record age | 20 points | Lose 2 points per year, up to 20 points after 10 years. |
| Coordinates | 15 points | Lose 15 points for missing or invalid coordinates, or a point outside its state. Lose 5 points if only a latitude and longitude range check is available. |
Age is the inventory source year minus the input year. Missing years, years before 1900 and years after the source year lose their full age weight. Inventory record age uses the record’s own revision date, not the date the Atlas downloaded it.
Confidence = round(100 − road age deduction − train age deduction − missing field deductions − record age deduction − coordinate deduction).
For example, a road count that is 20 years old loses 25 points. An unknown train-count year loses 20. An inventory record that is five years old loses 10. If all key fields are present and only the coordinate range can be checked, another 5 points are lost. The score is 40.
How the coordinate check works
Where a state polygon is available, an even/odd ray crossing check tests whether the point falls inside it. Holes are excluded, boundary points are included and longitude is adjusted across the date line. These simplified boundaries can flag valid points near a border. Without a polygon, latitude must be between −90 and 90 and longitude between −180 and 180. Passing that range check cannot confirm the state.
A score of 100 means the listed checks passed. It does not confirm the record on the ground. Missing train-count years lower confidence even when the train count may be recent.
04 / Comparison
Peers share reported equipment and activity.
Each group combines warning device class, a train band, a road traffic band and a setting. We calculate the exposure midrank within that group across the published inventory.
- Warning devices
- Gates take precedence, followed by flashing lights, then crossbucks or stop signs. Other or unreported equipment is grouped as unknown.
- Trains per day
- Below 2; 2 to below 10; 10 to below 30; or 30 and above.
- Vehicles per day
- Below 1,000; 1,000 to below 10,000; or 10,000 and above.
- Setting
- A reported built-up land cover indicates an urban setting. Otherwise, a population density of at least 1,000 people per square kilometre is used as an urban proxy; lower density is rural. Missing context is kept in an unknown group.
For example, a crossing with gates, 8 trains per day, 1,200 vehicles per day and an urban setting is compared with other crossings in those same four bands. Its peer percentile uses the exposure formula and tie rule above.
The current builder uses available regional population density. This is an approximate setting, not a Census urban designation or a street-level measurement. Small peer groups can give coarse percentiles, and similar activity does not establish similar safety.
05 / Limits
Source dates still matter.
Measures use the FRA values in the selected published inventory. HPMS traffic is not substituted: an accepted crossing-to-road match is not yet available. Any future supplement must show its own source and count year alongside the FRA value.
No FRA accident prediction estimate is included. We have not verified the complete published formula and coefficients for an exact implementation.
State summaries contain fixed distribution bins and a missing-value count. Bin starts are included and the next bin’s start is excluded. The final bin is open-ended. Summaries do not contain crossing records.
06 / Sources
Check the original records.
- FRA Highway-Rail Crossing Inventory
The source for road counts, train counts, warning equipment, record dates and coordinates. Count years describe the counts, while the API’s source date describes the inventory snapshot.
- Atlas data sources and methods
Source dates, regional population context and Census state boundaries used for the available checks.
- FHWA Highway Performance Monitoring System
A possible separate traffic source. No HPMS replacement is used in these measures.
- FRA report DOT/FRA/ORD-20/40
The accident prediction research report. An FRA model estimate has not been implemented here.