Reproduce Kansas City COVID-311 paper on SF data

[alexstorer]

What

Reproduces Tran, Bani-Yaghoub & DeLisle (2023), "Non-emergency responses in the 311 system during the early stage of the COVID-19 pandemic: a case study of Kansas City" (DOI:10.20517/dpr.2022.08, CC-BY 4.0) against SF 311 data, same Mar–Aug 2019-vs-2020 window. Adds scripts/covid_category_reproduction.py, three processed tables, a figure, and a README section.

Scorecard

Paper's claim (Kansas City)	SF result	Reproduces?
Aggregate decline (KC −13%)	−10.9% (328,494 → 292,829)	✅ yes
Decline not universal	7 of 23 interpretable categories rose, 16 fell	✅ yes
Street-condition categories fall	Street Defects −46%, Streetlights −18%, Blocked Street −17%	✅ yes
Text-mine free-text to tag COVID (KC 4.3%)	Not possible — service_details is categorical; 0 of 18 keywords discriminate	❌ no
Channel shift toward phone	Opposite — SF shifted to Mobile/Open311 (56→60%); phone fell (28→25%)	❌ no

Novelty verdict

The paper's robust findings reproduce, which only adds SF as a confirming third city. The two things worth recording:

1. The paper's signature text-mining method does not generalize to SF — the description field is a categorical subtype label, not narrative, so the discriminating pandemic vocabulary (covid/mask/corona/quarantine/social-dist) never appears. The 199 raw keyword hits are generic-word false positives (ppe inside shipped, actually more common in 2019).
2. SF is a strong "noise-up" data point (noise +80%, graffiti −14%) against the paper's NYC-up/Dallas-down split — sharpening this repo's existing inverse graffiti↔noise finding into an explicit COVID-mobility story.

Notes / caveats

• All-category counts come from Socrata aggregation queries (no bulk download); noise/graffiti windows are cross-checked against the committed raw files as a validation assert.
• SF's taxonomy is messier than KC's — relabeling artifacts (Muni Employee/Service Feedback, Abandoned Vehicle −90%) are flagged and excluded from interpretation.
• Channel comparison is SF's overall mix vs the paper's COVID-subset split (suggestive, not like-for-like).
• Reference PDF is git-ignored and cited by DOI.

Closes #13

🤖 Generated with Claude Code.

[alexstorer]

⚠️ Heads-up before merging: the "text-mining not reproducible on SF" conclusion in this PR is now in question — see #17.

The keyword probe here only checked service_details (categorical → 0 hits). It did not check status_notes, which does contain COVID free-text: 13,235 rows mention "covid" in the Mar–Aug 2020 window (plus 2,149 "social dist", 1,272 "pandemic").

Caveat that #17 asks to resolve: those status_notes look like agency disposition text ("Outreach provided per CDC Covid-19 guidelines…"), not citizen request text — so they may measure COVID-handled cases rather than COVID-motivated requests, which is a different thing than the paper's method. Worth resolving #17 before this claim lands as-is.

🤖 Generated with Claude Code.