Most Dangerous Highway in the US: 25 Deadliest Stretches

The most dangerous highway in the US is a 10-mile stretch of U.S. 19 in Pasco County, Florida, between Port Richey and Hudson. From 2019 to 2023 it had 61 fatal crashes, in which 64 people died. A typical road with the same traffic and design would have about 15 fatal crashes in five years. U.S. 19 had about four times that number.

Wherobots ran this analysis on public federal data. Wherobots is the AI Context Engine for Physical World Data: a platform that joins location data, such as crash records, road maps and traffic counts, at national scale.

We ranked every 10-mile stretch of Interstate and U.S. route with traffic counts in 48 states by how many more fatal crashes it had than a typical road like it. The crash records come from the Fatality Analysis Reporting System (FARS), the national record of every fatal traffic crash. Road maps and traffic counts come from the data states report to the Federal Highway Administration. The method is the network screening method in the Highway Safety Manual, the reference that U.S. road agencies use to find dangerous roads.

How we ranked the stretches

We asked two questions about each stretch. Did this road have more fatal crashes than it should? And is that a lasting pattern, or a bad year or two? We followed six steps:

1. Place each crash on the road. A crash counts on a road when it lies within 100 meters (about 330 feet) of a road section with the same route number. That placed 47,739 fatal crashes from 2019 to 2023.
2. Join each route into one continuous road. Some states record one highway as a separate piece for each county. We joined those pieces end to end. A road ends where the pavement is shared with a different route number, because the crashes there are recorded under that other route.
3. Slide a 10-mile window along each road. A new window starts at every mile, so every possible stretch gets a count. That gave 145,592 windows on 2,167 roads.
4. Work out the expected number of fatal crashes. We fitted a model to all of these roads. For each window, it gives the fatal crashes expected on a typical road like it. The model matches roads on daily traffic, number of lanes, speed limit, access control, truck traffic and urban or rural setting.
5. Give less weight to small counts. A window with 3 crashes could easily have had 1 or 6 by chance. The Highway Safety Manual's Empirical Bayes step blends the actual count with the expected count. The fewer crashes a window has, the closer the result moves to the expected count.
6. Rank by excess crashes. The excess is the adjusted count minus the expected count. We ranked windows by their excess and kept only windows that do not overlap.

Here is how the steps work on U.S. 19 in Pasco County. The window had 61 fatal crashes. The model expected 14.7. The adjusted count is 53.6, so the excess is 38.8 fatal crashes, the most of any stretch in the country.

The 25 most dangerous highway stretches in America

Ranked by excess fatal crashes, 2019 to 2023. Every stretch is 10 miles long (one is 9.9). "Expected" is the number of fatal crashes on a typical road with the same traffic and design. Routes that start with "I-" are Interstates, so I-35 is Interstate 35.

The map below shows all 25 stretches. Click a stretch to see its fatal crashes, deaths, expected fatal crashes and excess. These five had the most excess fatal crashes:

1. U.S. 19 in Pasco County, Florida, from Port Richey to Hudson: 61 fatal crashes, where 14.7 were expected. The excess is 38.8.
2. U.S. 441 in Broward County, Florida, from North Lauderdale to Davie: 50 fatal crashes, where 15.5 were expected. The excess is 29.2.
3. I-35 in Dallas, Texas: 51 fatal crashes, where 16.4 were expected. The excess is 27.0.
4. I-30 from Dallas to Mesquite, Texas: 48 fatal crashes, where 13.3 were expected. The excess is 25.8.
5. I-75 from Forest Park to Atlanta, Georgia: 51 fatal crashes, where 20.2 were expected. The excess is 25.1.

The excess fatal crashes come from the Empirical Bayes adjustment, so they are smaller than the fatal crashes minus the expected crashes.

What the 25 most dangerous stretches have in common

All 25 stretches are in urban areas. Fifteen are Interstates and 10 are U.S. routes. Eight states hold all 25: Florida has 7, Texas 6, Georgia 4, California 3 and Louisiana 2. Colorado, Illinois and Missouri have 1 each.

Together the 25 stretches had 1,062 fatal crashes and 1,139 deaths from 2019 to 2023. A typical road with the same traffic and design would have had about 352 fatal crashes. These stretches had about three times as many.

The U.S. routes on the list are busy roads lined with shops, homes and side streets. Examples are U.S. 19 in Pasco and Pinellas counties, U.S. 441 and U.S. 1 in Broward County, and Colfax Avenue, which carries U.S. 40 through Denver. The Interstates run through Dallas, Fort Worth, Houston, Austin, Atlanta, Miami, New Orleans, St. Louis, Chicago and the Los Angeles area.

The ranking holds when the window moves. For each of the 25 stretches, we checked every window that overlaps it by at least half. Each of those windows still had at least 4.8 more fatal crashes than expected.

Which types of road are most dangerous per mile driven

The ranking above ranks places. A death rate measures something else: the risk of each mile a driver travels on a type of road. A vehicle mile is one vehicle driving one mile.

These are the totals for all 14,570 Interstate and U.S. route stretches of at least 10 miles with traffic counts, from 2019 to 2023. Together they had 49,050 deaths, or 0.83 deaths per 100 million vehicle miles.

Road type	Share of vehicle miles	Share of deaths	Deaths per 100 million vehicle miles
Interstates	66.8%	46.2%	0.57
Urban U.S. routes	16.8%	24.6%	1.21
Rural U.S. routes with 4 or more lanes	11.4%	17.8%	1.29
Rural two-lane U.S. routes	5.0%	11.4%	1.91

Rural two-lane U.S. routes are the most dangerous type of road per mile driven. Their rate is 3.3 times the Interstate rate. They do not lead the ranking above because they carry little traffic. A 10-mile stretch of rural two-lane road rarely has enough fatal crashes to stand out from chance. A busy urban road that is more dangerous than it should be has many more crashes to count.

How weather and the calendar change the risk

Deaths on these roads peak in late summer and fall. From 2019 to 2023, August had the most deaths, 4,551, and February had the fewest, 3,415. More driving explains part of that gap. Per mile driven, the death rate was lowest in March, at 0.76 deaths per 100 million vehicle miles. It was highest in September, at 0.87, and stayed above 0.85 from August through December.

To see where weather matters, we matched each state's weather to its fatal crashes, day by day and month by month. The weather comes from hourly reports at about 2,000 airports in the 48 states, from 2019 to 2022, the last year in the data. Monthly rain and temperature for each state come from the National Oceanic and Atmospheric Administration. Crash reports on these roads list rain, snow, ice or fog at 10.8% of deaths. The Federal Highway Administration counts 9% of deaths on all U.S. roads as weather-related for the same years.

Where the weather is worst each month

Michigan has the most hours of snow of any state from November through February. Its airports reported snow for about 121 hours in an average February. Wyoming has the most snow from March to May, and Montana in September and October. Oregon and Washington get the most hours of rain from November to January, and Rhode Island and Massachusetts in April and October. Heavy fog is most common in Oregon and Washington in winter, and in New England from May to October.

Heavy rain raises the number of fatal crashes

Heavy rain is the clearest danger in the data. We called a day a heavy-rain day when a quarter or more of a state's airports reported moderate or heavy rain. We compared each one with the same state and weekday one week earlier or later, on a dry day. Heavy-rain days had 297 fatal crashes on these roads, and the matched dry days had 236. That is 1.26 times as many.

The wider the rain, the bigger the increase. When half or more of a state's airports reported it, there were 1.42 times as many fatal crashes as on the matched dry days, though that rests on fewer days. Of the heavy-rain days, 54% were in the South, where the increase was 1.27 times. This design follows Andrey and Yagar (1993), who compared crashes in rain with crashes in matched dry periods.

Snow and fog days show no clear increase

Snow days had 437 fatal crashes, and the matched dry days had 444. The two counts are within the range that chance alone produces. Heavy-fog days had 1.12 times as many fatal crashes as their dry days, which is also within that range. The data hold no traffic counts for each day, so they cannot show whether fewer people drove in snow.

Over a whole month, region matters more than weather

Over a whole month, the weather does not line up with excess fatal crashes. A month with more snow or rain than usual for that state had about as many fatal crashes as expected. Eisenberg (2004) studied the 48 states from 1975 to 2000 and found that wetter months had slightly fewer fatal crashes, while rainy days had more. The region matters more. The South as a whole had more fatal crashes than expected in every month, and the Northeast and Midwest had fewer, snowy or not.

The most dangerous stretch in each state

The map below shows the 10-mile stretch with the most excess fatal crashes in each of 48 states, one state at a time. It opens on Florida. Use the arrows to move through the states in order of excess. Each view shows the same road for up to 25 miles on each side and every fatal crash on it from 2019 to 2023.

Alaska, Hawaii and the District of Columbia are left out. Alaska's Interstates have no route number in the federal road data. Hawaii and the District of Columbia have no stretch that meets the rules. Tennessee's U.S. routes are not in the data, so its stretches are all Interstates.

Florida's worst stretch is also the worst in the country: U.S. 19 in Pasco County, with 38.8 excess fatal crashes. Eight of the 48 state stretches are in the national top 25. Of the 48, 21 are Interstates and 27 are U.S. routes.

In states with few fatal crashes, the excess is small. In 17 states the worst stretch has fewer than 5 excess fatal crashes. The smallest is I-94 in North Dakota, at 0.6. For those states, the stretch on the map is the worst in the state, but it may not differ much from a typical road.

Work zones on the most dangerous stretches

Work zones do not explain the list. We downloaded a snapshot of the public work zone feeds that states publish in the Work Zone Data Exchange (WZDx) format on October 15, 2026. Three of the 25 stretches overlapped an active work zone that day: U.S. 441 in Broward County, I-95 in Miami-Dade County and U.S. 1 between Goulds and Florida City. Crash reports marked 25 of the 1,139 deaths on the 25 stretches as work-zone related.

The snapshot covers 16 of the 25 stretches. The other 9 are in Georgia, California, Colorado and Illinois, which had no WZDx feed that anyone could download without a key on the snapshot date. The snapshot covers one day, so a work zone that started or ended at another time would not appear.

Data and rules

Every step ran in Wherobots. These are the data sources and the rules behind the ranking:

• Crashes. FARS fatal crashes from 2019 through 2023, the five most recent final years. Our FARS totals match the federal counts for each year, for example 42,721 deaths in 2022.
• Crash locations. Crashes with missing or placeholder coordinates are left out. On Interstates and U.S. routes, that removed 170 crashes, and 437 more had no route number we could read. Of 53,187 eligible crashes, 47,739 (89.8%) lie within 100 meters of a section with the same route number.
• Roads and traffic. The road map and traffic counts come from the 2024 Highway Performance Monitoring System (HPMS), the road data states report to the Federal Highway Administration. The 2024 average daily traffic is applied to all five years. Tennessee's U.S. routes are left out because Tennessee's road data does not identify them, and Texas frontage roads are left out.
• Continuous roads. A road is cut where two sections in a row are more than 1 kilometer (0.6 miles) apart. Pieces of the same route number are joined when their ends meet within 200 meters. Every ranked stretch is one continuous piece of road.
• Windows. Each window is 10 miles long, and a new one starts every mile. A window needs traffic counts on at least 95% of its length.
• Months and weather. Each crash takes the month and the weather listed in its FARS record, and crash reports gave no weather for 2,889 of the deaths. Monthly traffic is an estimate. Each stretch's five-year traffic is split by the monthly travel shares for its state's rural or urban main roads, from the Federal Highway Administration's Traffic Volume Trends report. The months add back to the five-year totals for crashes, deaths and traffic.
• Weather by state and day. Airport weather reports from 2019 to 2022 give the hours of snow, rain and heavy fog at each airport. Each report counts on the local date it starts, and no report counts for more than 24 hours. A heavy-rain day is a day when a quarter or more of a state's airports reported moderate or heavy rain for an hour or more. A dry control day is the same weekday one week before, or else after, when no more than 1 in 20 of the state's airports reported rain, snow or fog.
• Weather by month. The monthly comparison counts fatal crashes in each state, month and year from 2019 to 2022. It compares them with the expected number, after allowing for each state's usual level in each calendar month. That leaves the year-to-year change in weather as the thing being tested, as in Eisenberg (2004).
• Expected crashes. The model is a Poisson regression, fitted separately for Interstates, urban U.S. routes, rural U.S. routes with 4 or more lanes, and rural two-lane U.S. routes. It uses daily traffic, lanes, speed limit, access control, truck share and urban share. Where the road data has no speed limit or truck count, the model uses the national median and records that the value was missing.
• Empirical Bayes adjustment. The weight on the expected count is 1 / (1 + k × expected), where k measures how widely real counts vary around the model. This is the formula in Part B of the Highway Safety Manual, published by the American Association of State Highway and Transportation Officials (AASHTO).
• Ranking. Windows are ranked by excess fatal crashes. A window that overlaps a higher-ranked window on the same road is skipped.

What the ranking leaves out

The ranking covers fatal crashes only. FARS records crashes in which someone died within 30 days, so crashes with only injuries are left out. Road agencies usually build their expected-crash models from all injury crashes. With fatal crashes alone, the counts are small everywhere, which is why the adjustment for small counts matters here.

The model uses the road features that the federal road data records on nearly every section. Traffic signals, driveways, curves and hills are recorded on only a sample of sections, so they are not in the model. Some of a stretch's excess may come from those features.

The ranking covers only Interstates and U.S. routes. From 2019 to 2023, 132,604 of the 186,398 fatal crashes in FARS happened on other roads, such as state routes and local streets. A road that is not on this list may still be more dangerous than it should be.

Why AI is blind to the most dangerous roads

AI is blind to the physical world. The AI models people use every day were trained and grounded on text, documents, databases, and the internet. They can describe a highway's reputation from news stories. They cannot place 47,739 crash locations on a road map by themselves.

A state traffic safety engineer who asks an AI assistant where the most dangerous stretch is needs that placement. The answer depends on which road each crash was on, how much traffic each stretch carries and how many fatal crashes a typical road like it has. Without that context, the assistant can only repeat the roads that made the news. The engineer may then spend a safety budget where the headlines were.

Wherobots supplies that context. For this ranking, Wherobots placed each crash on its road, built 145,592 windows on 2,167 roads and computed the expected and excess fatal crashes for each window. A road agency can run the same steps on its own crash records and road data. To screen your state's roads, contact the Wherobots team.