SPOILER ALERT — I never got the bottom of this…but I wanted to publish and get it out there. Hopefully can do data replication experiments some day.
MAG puts out a report that seeks to quantify intersection risk for each of the (1,300 was it?) intersections in the MAG region; the final score is a number between zero and 1, from no-risk to most-risky. It uses the most recently available 5 calendar years of ADOT crash data, which was for this round 2017-2021.
I first became aware of the report’s existence, though it’s been put out before with older data, this past June:
- June 2023, Arizona Republic Proceed with caution: These are the Phoenix area’s most dangerous intersections; (probably paywalled, but is best media version out there that I’ve found, it also has a nice heatmap)
Then this more recent blurb drew my attention:
- September 2023, 12news: Phoenix roundabout ranked riskiest intersection in the county
In traffic engineering circles — oops, possibly confusing pun — modern roundabouts are considered to be significantly safer than other types of intersection control (i.e. signals, 2-way stops, all-way stops) because their physical configuration, e.g:
Overall, there is an observed reduction of 35% and 76% in total and injury crashes, respectively, following conversion to a roundabout. These values are consistent with results from international studies, as shown in Exhibit 5-10.
The findings of these studies all show that injury crashes are reduced more dramatically than crashes involving property damage only. This is in part due to the configuration of roundabouts, which eliminates severe crashes such as leftturn, head on, and right angle crashes.
— NCHRP Program Report 672 Roundabouts: An Informational Guide, 2nd Edition
(I should add here: there are entire other sections which examine specific safety impacts on other-than motorists; we’ll see a little later how the MAG report adjusts for bike/ped crashes)
So here’s the MAG source material:
- MAG’s Safety Programs homepage; under ‘Materials’ right sidebar: 100 Intersections .pdf list, and 100 Intersections (static) map
- In the list document, there’s a brief description, and link to the methodology: Network Screening Methodology for Intersections (NSM-I)
- This 2014 document also has possibly more details on the methodology Technical Memorandum No. 4 – Network Screening Methodologies for Prioritization of Road Safety Needs
- (A copy of these documents is archived here in this shard folder with filenames beginning with MAG-Top100. Document links have a way of frequently breaking )
The methodology document is pretty straightforward, called the Network Screening Methodology for Intersections (NSM-I), the specific were footnoted to a paper published in 2009, and apparently adopted for this purpose by MAG in 2010. It computes three factors CF, CS, and CT for crash Frequency, Severity and Type respectively. It then weights them together (with Severity getting a double-weight) and is mathematically guaranteed to produce a score of between 0 (safest) and 1 (riskiest) in any particular area, in this case Maricopa county (or possibly the MAG region).
Frequency, CF, is the most straightforward: the number of reported crashes at the intersection; it should be directly extractable from the Incident Table of ADOT’s data by looking at OnRoad and CrossingFeature and discriminating by Junction Relation (i.e. being either intersection or Intersection_related types). Sidenote: There is a field in the data called IntersectionATISCode but it doesn’t seem to be populated(?).
CS is also straightforward: Table 1 lists a relative weight of InjurySeverity (KABCO scale; where K is killed and O is prop damage Only. In the data, it’s a number from 5 to 1, with 99 being unknown). E.g. a fatal incident has a weight of 1,450 whereas a suspected injury (C) is 11. InjurySeverity is the highest level of injury suffered by anyone involved regardless of person type (driver, ped, etc) or number of units involved. InjurySeverity is a direct extract from the Incident Table in the ADOT crash data.
CT is a little more massaged, it’s mostly the Manner of Collision (e.g. rear-end, angle, sideswipe, etc) and there’s a table derived somehow that assigns a dollar amount to each severity, Table 2 (which looks nearly identical to Table 1, except instead of weights, it’s the weight multiplied by $4,000; there appears to be a type in Possible Injury, it should be $44,000, not 42,000); and then backs into distributing it to each Manner of Collision… with one caveat, Ped and Bike crashes are not a Manner of Collision in the database. For the purposes of computing CT, they add Bicycle and Pedestrian as a manner of collision. This last bit makes the data extraction more intricate — the Incident Table does differentiate between the number of motorist units (i.e. number of vehicles) and nonmotorists (i.e. the number of peds and/or bicyclists) but nonmotorists are lumped together. You need to look at the Person Table to differentiate between number of pedestrians vs. bicyclists.
In any event, then, for any given crash you take the dollars from Table 4 and multiply it times the unit.
So that’s the CF,CS, and CT are computed and normalized to the maximum value for the “worst” intersection in the entire sample; and then weighted at 1/4, 1/2, and 1/4 respectively; in other words the Severity score is weighted twice as much as Frequency or Type
I never got to the bottom of this
After delving into this for a couple of hours, I never got to the bottom of this. I have all the (presumably same) raw data that MAG has — it all should come from ADOT.
I ran out of time before I could develop any queries to try to replicate their results.
Here’s what’s troubling me: According to their newest results the #1 most dangerous intersecting in MAG is a roundabout, not signalized.
And curiously, there’s a KJZZ news item from March 2021 referring to a prior version, using 2 year older data, so 2015-2019, curiously did NOT have 99th & Lower Buckeye as head-and-shoulders riskiest, rather it was #4, behind what are now (in the newest report) #3,4 and 2 respectively. Another oddity is the scores for the top 4 were closely grouped (about 1.2), and in the newer report the #1 is WAY higher, and 2,3,4 are distant 2nd, 3rd and fourth places. [added a little later: we could posit perhaps that although the time periods overlap, 2015-19, vs. 2017-21; perhaps the crazy nature of pandemic era crashes in the newer period — excessive speeding, say — caused the shift??]
This would seem to be at odds with the accepted traffic engineering theories, mentioned above, that roundabouts mitigate crash severity. And as you can see by the weights, 50% of the final score should be derived from severity, which as you can also see, increase exponentially as severity goes up; and another 25% due to CT, crash type. A roundabout should produce more rear-ends and sideswipe-same-direction which are a couple of the cheapest. Whereas something like angle-opposite-direction (think left turn error), common at signalized intersections, should be impossible at a roundabout.(?).
The huge jump in CF, frequency, compared to the 2 year prior data also seems suspect. A high CF accounts for 25% of the final rating; and according to the newer data, that particular roundabout is the highest of all intersections in Maricopa by a large margin, which wasn’t the case with the two year ago data.
 Qin X., Laracuante L., Noyce D.A., Chitturi M. Systemwide Intersection Safety Prioritization Development and Assessment. (full .pdf)
[2 ] Campbell J.R., Knapp K Alternative Crash Severity Ranking Measures and the Implications on Crash Severity Ranking Procedures. (full .pdf)