Location x COVID interaction on noise complaints

[alexstorer]

What

Follows up the noise finding (the +81% Apr–Jun 2020 lockdown surge) by asking whether that surge is spatially uniform or moderated by location. Adds scripts/noise_location_covid.py, two processed tables, a figure, and a README section.

Result: the surge is strongly spatial

The Apr–Jun 2020-vs-2019 change ranges from −65% to +313% across 27 neighborhoods:

	2019	2020	Change
Bayview Hunters Point	47	194	+313%
Western Addition	40	152	+280%
…
Financial District/South Beach	188	85	−55%
Mission Bay	124	44	−65%

Every residential area rose; the only neighborhoods that fell are the commercial/office/transient cores (Financial District/South Beach, Mission Bay, Chinatown) — exactly the places that emptied out when offices and tourism shut down. That's a clean shelter-at-home signature.

Formal test (with an overdispersion correction)

Difference-in-differences Poisson GLM, calls ~ neighborhood × covid with a log-days offset. The counts are overdispersed (φ ≈ 3.2), so the naïve Poisson LR test (p≈1e-56) is far too liberal — I report the quasi-Poisson F-test instead:

• Neighborhood × COVID interaction: F(26, 108) = 4.07, p ≈ 1.3×10⁻⁷ — highly significant net of each neighborhood's own baseline.
• Negative-binomial cross-check (α estimated) agrees: p ≈ 3×10⁻¹¹.
• Pooled lockdown effect: IRR = 1.74 (+74%).

Caveats (in the README too)

• At supervisor-district resolution (11 units) the interaction attenuates to borderline (p ≈ 0.09) — districts blend residential and commercial blocks, so the contrast averages out; the signal lives at neighborhood scale.
• A 2019-vs-2020 Apr–Jun DiD does not separate the lockdown shock from pre-existing neighborhood trends.
• ~6.6% of noise rows have a null neighborhood (dropped); 14 low-volume neighborhoods (2019 baseline < 20) excluded for stable estimates.

Closes #15

🤖 Generated with Claude Code.