Local Finance & Extreme Weather — Replication Code
Stata code replicating the analysis on community banks and economic resilience to extreme weather events.
Data & Replication
The datasets I build from public records, and the code and data behind published and working papers.
Most of what American local governments and nonprofits disclose is technically public and practically unusable — a PDF financial report, minutes and video, a scanned return. Each is available to anyone who asks for that one document, and almost none of it can be read across thousands of organizations at once, which is what an empirical question actually requires. So I build the datasets my research needs out of those records. Two are in active use.
Local Government Meeting Text
A corpus of the documentary record of American local government — agendas, minutes, staff reports, and transcripts — assembled from the platforms cities and counties publish through, currently about 2.1 million records from roughly 4,600 verified portals, of which some 178,000 are substantial meeting documents of 750 words or more. Only a minority of the roughly 35,000 U.S. municipalities publish through any discoverable platform at all, and adoption tracks population and staff capacity — a finding in its own right about who leaves a searchable public record. Every analysis carries the corresponding selection weights.
Local Government Financial Statements
A panel of city and county finances built by machine extraction from Annual Comprehensive Financial Reports — the audited statements governments publish each year and almost no one reads comparably. Existing sources for this kind of question are coarse and lagged; ACFRs are annual, detailed, and specific to the reporting government, which makes them the right instrument for asking what a shock does to a balance sheet: which funds absorb it, what happens to unrestricted net position, and how long the effect persists. The panel is linked to hazard and property-value data for that purpose.
Access
Both datasets are in active use and being documented for publication; once the methods papers are out, the derived data will be released to the extent the underlying sources permit. Some inputs cannot be redistributed — hazard-loss data used in several projects comes from a licensed subscription source and is not mine to share — but the public-record documents and the code that assembles them are a different matter. If you are working on something these would help with, get in touch.