Together with my University of Michigan colleague Sarah Mills, a few years ago I was involved in the pilot project to launch what has become the Michigan Zoning Map, an effort to create systematic variables describing the zoning districts in communities in our state. Drawing on the methodology created by the pioneering National Zoning Atlas, and originally launched in collaboration with them, the project has become a fully independent project expanding from the original three counties mapped by our team to now including zoning data for 15 of the state’s 83 counties.

During our project, I realized that some prominent national researchers had also analyzed the zoning codes for some of the jurisdictions whose codes we were painstakingly studying. So, I thought it may be an interesting research project to compare our data with theirs. Well, as is often the case, it proved to be much more difficult than I expected. However, with the assistance of PhD student Maina Wachira, this summer we published the results in the journal Cityscape, a leading housing journal published by the U.S. Department of Housing and Urban Development, in an article titled “Comparing Natural Language and Manual Zoning Data: A Data Quality Analysis for 14 Michigan Jurisdictions.” Thank you to the editors and reviewers for their help improving and disseminating this work.

If you’re curious to know a bit more about the work, the National Zoning Atlas had an excellent LinkedIn post, pointing out it addresses a question they are frequently asked: why bother having people read zoning codes when you can use automated methods? They summarize our findings as follows:

“Simple yes-or-no questions fared best: whether a jurisdiction allows accessory dwelling units anywhere matched across datasets in 64% of cases. But once the comparison turned to precise numbers, such as maximum density and height limits, agreement broke down. The gap makes sense. Zoning codes are full of exceptions and conditions that shift by housing type, district, or bedroom count — nuance a trained analyst can track, but that pattern-matching tends to miss. Goodspeed and Wachira conclude that NLP tools can offer a useful broad-strokes picture of zoning nationwide, but research and policy decisions that need precision still call for the district-by-district work the NZA method is built on.”

I know that NLP methods continue to evolve, but we felt that it was worth making a careful comparison with our data. Hopefully others will do similar work to improve our knowledge of the best ways to understand and analyze zoning. Take a look at our article here.