About two weeks ago, the LAB issued a report based on data they collected from the everybicyclistcounts.org site that was setup and run by Elizabeth Kiker (whom I guessed was an intern; but i really don’t know).
I had been in contact with Elizabeth a few times over the course of her data collection in 2012 with some Arizona fatalities. It seems like a perfectly reasonable project; especially after reading a perfectly sensible preview of the project why-every-bicycle-counts-and-what-we-can-learn-fatal-crashes dated May 2012, especially when it described the limitation of methodology which i will quote at some length:
Limitations: Despite the enormous value and new analysis we believe this project will bring to the bicycling community, we recognize that there are (at least) several limitations.
No coverage of injuries and near-misses. As it is, we have taken on the significant task of trying to track down what amounts to nearly two crashes per day. It would be nearly impossible to do the same for the tens of thousands of non-fatal crashes.
No exposure data. Without knowing how many people are riding under difference conditions, it’s impossible to know the relative risk of different circumstances. This problem haunts other bicycle risk analysis as well.
Not scientific. It is not a census of all fatal crashes (though we are trying to make it as comprehensive as possible), nor is it a random sample. We report only the ones we find out about. In addition, it requires a person enter the information and make determinations about which categories best apply. (Links to all sources are available on EveryBicyclistCounts.org.)
Dependence on public sources. Every Bicyclist Counts still depends, to some extent, on police reporting and media accounts. These are often flawed, incomplete, or biased. We believe that additional information from cyclists and families can help improve our data. The project may, in the long run, help improve the quality of future reporting.
However, the final report and 5/21/2014 blog announcement (direct link to .pdf) is another story altogether. Now, Andy Clarke (The LAB’s Executive director at the time; see comment below for another example of Clarke’s thinking, that one involving supporting a mandatory-use law) is running all over the place, see also e.g. bicycleretailer.com interview, saying essentially that the LAB has uncovered some sort of new proof;
For the longest time it’s been an article of faith that we should be taking the lane, and that separated bike facilities are unnecessary … well, I think we are grown up enough now to say that’s not the case. Most people feel more comfortable actually having a paved shoulder or a cycle track or having a buffered or protected bike lane, and those things will reduce the fear and the incidence of being hit from behind. And we shouldn’t feel bad or awkward about saying that.
Without mentioning what those of us who have studied the stats already know, for decades: 1) rear-ends collisions are over-represented in the fatal collisions, versus the much more numerous non-fatal collisions (100:1. I.e. there are about 100 non-fatals for each fatal collision. In AZ for example, the rate is about 2,000:20 per year), and 2) rear-end are over-represented when any of the following factors are present: rural; high-speed; nighttime. So Andy draws conclusions from the non-scientific, non-random sample of fatals to whip up fears that rear-ends are more common than they actually are according to, e.g. PBCAT to apparently engender support for “separated facilities”.
The Report Itself
So, the report’s dataset is 628 anecdotally collected reports of cyclist fatalities; mainly in the year 2012. They note that they collected reports on “552 in 2012 alone. In 2012, FARS reported 726 bicyclist deaths”. It is quite likely that the LAB’s 552 is almost but not quite a perfect subset of the 726 since FARS counts strictly traffic fatalities involving an in-transport motor vehicle.
The report’s methodology is, as noted above, and pretty much by definition, non-scientific and non-random. The report is also non-rigorous statistically — there are no statistical measures (e.g. no confidence intervals) of any kind. This is, however, not mentioned in the report.
One error / omission is table 1: the list of crash types and percentages did not “add up”. It listed 567 incidents, rather than 628. The author, Ken McLeod was kind enough to supply the missing 61 incidents and now the results can be calculated. See the worksheet “LAB EBC” for the corrected Table 1 I have on google docs.
I encourage LAB to release their data; Ken told me that they are considering this. There’s really no excuse not to. (update 5/2015 still no data).
The choice to calculate the fraction of crash types (e.g. rear-end = 40%) is probably inflated by the fact that it’s calculated as a fraction of “of known crash types” and they have a very high number of unknown crash types. As previously mentioned there was no confidence interval presented — it would presumably be very large. FARS in 2010 (the most recent year PBCAT was completed) found motorist-overtaking crash group to be only 25% of all bicyclist fatalities — FARS is a complete census; there is no sampling error.
Rather than note that, e.g. the 40% might be incorrect, due to methodology limitations, Andy (in particular) rather uses the report as a propaganda tool citing it as “proof” of the need for more segregated facilities.
The report oddly does not mention rural/urban which tends to be rather important in rear-end collisions. It is also highly unlikely segregated facilities will ever be built on the vast majority of rural roads.
The report (correctly) mentions there is no exposure data — and thus exposure risk can’t be quantified; they also present (via FARS data) the number of 2012 fatalities by road class (rural / urban and freeway, arterial, collector, local; etc). So the finding that “most fatalities occur on urban arterial roads” is somewhat self-evident. (note to self: check TYP_INT, Type of Intersection and RELJCT1/2, Relation to Junction in the incident table against pbcat data).
see also: fars-and-pbcat