It's not entirely perfect yet, but I have a workflow which generates these maps right out of Python for all of my cities, with some provision for longitudinal comparisons. To do still are to deal with projection issues (they're all in "web Mercator" because that's what Leaflet uses), and centering issues caused by inconsistencies between folium, selenium, and PhantomJS. And to improve legends and captions. But it's pretty cool, if you ask me.
I'm working on scaling up the data analysis from the thesis, and I'm making some good progress, thanks to Folium. I'm pretty close to being able to run this on an arbitrary number of cities: just need to make the code a little more robust.
I've been working on generalizing my code so that I can make comparisons for dozens (or hundreds) of cities. And of course, I've found a ton of bugs, mostly related to my own poor understanding of Python data structures and functions. But I also found a fundamental issue with the calculations I used in my thesis.
After looking at the job connectivity maps, I was curious to explore the idea that densities above a certain level led to walking more than cycling. I don't have enough data to make a definitive statement, but I did find an interesting phenomenon related to connectivity in Columbus. Columbus has three disconnected zones where jobs are very close, broken up by areas of low job access.
OSMnx makes it easy to generate statistics on street networks. Again these aggregated stats show very strong correlations, especially between bike mode share and intersection density. One of the problems of cycling advocacy is that there's often not much that can be done to change this indigenous condition.
One striking result from the target area analysis is the correlation between residential density and cycling rates in the target area. For these four data points, the correlation between density and bicycle mode share is dramatic (r=0.97), which seems to speak to the importance of indigenous conditions in people's mode choices. Unfortunately, the effect disappears when examined at the census tract level.
One of the factors I'm trying to measure in valuing facilities is their usefulness; do they actually go where utility cyclists want to ride? One of the factors I used was access to jobs within 1.5 miles along the cycling network. These maps really visualize the differences in density of the street networks.
Combining my concepts of target area and value-added facilities, I developed maps for each city which select the circle of census tracts near the central city which have the great number of commute cyclists. One thing that's visibly notable is the relatively weak connection of value-added facilities to cycling rates; only in Columbus are they well-connected. Part of what I conclude is that cultural factors and the indigenous qualities of the street network are more important than value-added facilities in cycling mode choice. The methodology is potentially interesting for other kinds of analysis; you could easily optimize for any other census data.
I've been too busy with thesis work to post any updates here, but now it's actually done! And I'll have some time to start sharing bits of it. The most common reaction I've gotten from bike-identified people I've shared the work with is, "that's interesting, I haven't thought about it that way before," which I will accept as an indication that the project has been successful.
One problem I noticed while doing field research is that the city extents vary greatly; Austin, for example, has over six times Minneapolis' land area, which means that Austin includes some sprawling, low-density areas with low cycling rates that are excluded from Minneapolis' mode share numbers. I looked at creating a 4-mile radius circle to encompass the highest-cycling area of each city (by ACS mode share data), to be able to develop more comparable metrics from city to city. There are still a lot of issues with the data, but at least I'll be comparing rotten apples to rotten apples.
