Mapping Traffic Fatalities | Hacker News Reader

Mapping Traffic Fatalities

lucaspuente.github.io

71 pointssndean18 comments

I'm very pleased (from a data insight standpoint) that the map doesn't look like a population heatmap [1].

As a next step, I suggest calculating the actual rate of fatalities in a particular location vs. some naive estimate for that area, based on traffic counts, car utilization, or nearby population. This will highlight places that have a higher-than-expected number of fatalities.

Some of these places can already be seen visually. For example, Florida is blotted in red even though the population is concentrated around the coasts and bays and the Orlando area; therefore US 301 and other 'interior' portions are more dangerous than population alone would explain.

The Carolinas are similarly covered, well-outside of the Crescent from Raleigh-Durham through the Triad to the Charlotte suburbs. California's Central Valley looks particularly dangerous too, despite most of the traffic sticking to the far western edge with I-5, or the eastern valley floor where the built-up areas are.

[1] https://xkcd.com/1138/

I'd agree that there it would be very nice to correlate with a traffic load graph. For areas like California, where you have pretty much the I-5 or the Highway 1, knowing the fatality per car travelled would be interesting view.

It looks exactly like a population heatmap.

More like a traffic heatmap. California seems to have a lot of red dots on the I5 between LA and SF. Not a lot of people live there.

I wonder how much "proximity to a good hospital" changes it, as well. If I could zoom in far enough, and traffic can be assumed to be constant, would there be more fatalities on the remote parts of I-5?

This looks interesting: https://i2.wp.com/metrocosm.com/wp-content/uploads/2016/05/u...

>Some of these places can already be seen visually. For example, Florida.

I think if this is true at all (I'm not sure it is) it's just an artifact of the large blobs used on the traffic fatalities map.

I agree with one of the posters below: this is a population heatmap, exactly as identified by XKCD (https://xkcd.com/1138/). Florida, which you mention, is highlighted in red only because this is a small-scale map with poor density representation. Here is Florida with the exact same data, as plotted by me in R with a little more resolution:

https://imgur.com/a/WU3JI

A population map, in other words. As a very basic demonstration of the capabilities of R, this is OK, but as a visualization, this is 100% useless. You would need something like "percentage of expected fatalities for a given stretch of road, given the type of road" to actually have something interesting. And that takes a lot more work.

Would it help to use data like yearly per capita consumption of gasoline by state [1] (as a proxy for miles driven per person) in addition to the existing data? Would that make a difference and shed any more insight?

[1] http://www.statemaster.com/graph/ene_gas_con_percap-energy-g...

You can get vehicle miles driven by state directly, and also such data as fatalities per road user and per miles driven: http://www.iihs.org/iihs/topics/t/general-statistics/fatalit...

However, you won't get much insight from plotting this data geographically: a table (as in that link) is enough. Basically you get exactly what is expected: Daniel Patrick Moynihan's "Law of the Canadian Border" - southern states have more social problems.

If you want high-resolution geographical data for purposes such as finding problem roads, that would be a lot more difficult - you need to have a graph of the road network (from e.g. Open Street Map), with traffic data attached to the arcs of the network (I don't know of any public access databases like that, although such data is commercially available for road planning) plus accident locations accurately mapped to the roads. I suspect it wouldn't be a quick project to do.

There is something clearly very wrong with this data, since is suggests that California has less than 1/10th the per capita consumption of New York, and uses less fuel on an absolute basis than Connecticut. The dynamic range in fuel consumption differs only 2 fold for the top 46 (of 51) states, but then drops more than 5 fold. Those last 5 numbers must be on a different scale.

I am struggling with what will be the purpose of this. Any project that I work on with traffic analysis, PD&E, or roadway design includes a crash analysis.

The crash analysis includes an analysis of every accident within a five year period, identification of high crash locations and what caused the crash (high speed, inebriation, or a list of other options), and then what strategies can be adopted for the most impact in a benefit/cost analysis.

Just looking at heat maps of crashes without inclusion of traffic volumes (one crash per million miles is nothing compared to an area of one crash per thousand miles traveled) becomes meaningless.

have a love-hate relationship with the NHTSA database. On one hand, it covers 40 years of data, and unlike virtually every other long-running dataset in existence, it keeps not only a relatively standardized schema, it publishes a all-in-one user manual that covers every year (and any changes between them); most places publish a new PDF for every year, which means you have to read through each one. Even better is that when they make changes or additions to their methodology, they publish an auxiliary dataset that you can use to retroactively update old datasets.

But then for some reason, they throw out all rationality when it comes to naming files and fields. It's as if there was employee who was old-school and was stuck in mindset of naming things under the 8-character all-caps limitation. And then someone else thought, "Fuck this", and went the entirely opposite direction by naming files with long titles and spaces, and the two employees had a passive-aggressive argument that never got resolved over the decades.

This makes scraping and collating the data a real pain in the ass. But compiling the data got a lot easier when I realized that wget works on FTP sites. Here's a mirror of the FTP site (around 1.7GB zipped) that I put up on Github (still working on getting everything together in one database): https://github.com/wgetsnaps/ftp.nhtsa.dot.gov--fars

Already available for the UK, see:

http://www.crashmap.co.uk/

Is it possible to make such a map for homicides? Does the data exist with sufficient location information?

The FBI Uniform Crime Report and/or NIBRS may be the most comprehensive/granular data there is on homicides nationwide, but I don't think there's incident-level geography:

https://www.icpsr.umich.edu/icpsrweb/NACJD/NIBRS/

Counting homicides, and even determining what exactly is a homicide, is not as clear as one might think:

http://contentsmagazine.com/articles/homicide-watch-an-inter...

Not that many agencies post the data themselves. Surprisingly, the NYPD does post it as part of their felonies dataset: https://data.cityofnewyork.us/Public-Safety/NYPD-7-Major-Fel...

Good find with the NYPD link. I've seen New York data before that just went precinct by precinct, but this has exact latitude and longitude.

Isn't it pretty much a US population density map (if it's not, I can't quite see the difference)? If there are differences, it would have been interesting to display the number of accidents in an area as a proportion of the population.

Reply on news.ycombinator.com