SF 311 Graffiti & COVID

San Francisco 311 service-request data used to evaluate one hypothesis:

Did graffiti calls decrease in 2020 as a result of COVID?

Data source: DataSF "311 Cases", Socrata dataset vw6y-z8j6. Graffiti requests are identified by service_name in (Graffiti, Graffiti Public, Graffiti Private). This project analyzes calls from 2017–2022.

Findings

Short answer: nuanced — yes for the lockdown, no for the year.

Graffiti calls fell sharply during San Francisco's strict spring-2020 shelter-in-place, but rebounded so strongly afterward that the full-year total ended above 2019.

Window	2020	2019	Change
Apr–Jun (strict lockdown)	9,828	14,362	−31.6%
Jul–Dec (rebound)	32,174	26,785	+20.1%
Full year	59,051	57,035	+3.5%

The shaded lockdown window (Mar 17 – Jun 2020) is the deepest trough in the entire 2017–2022 series — calls dropped to ~3,150/month, well below the usual 4,500–6,500 — then recovered to above-2019 levels by late 2020.

Conclusion: COVID produced a clear but short-lived decline in graffiti 311 calls during the strict lockdown; it did not reduce the full-year total. So the hypothesis holds only for the lockdown window, not for 2020 as a whole.

Caveat: 311 counts measure reports, not graffiti incidence. With far fewer people out during the strict lockdown, the drop plausibly reflects reduced reporting (and less-frequented public space) rather than less graffiti being created. These data show when calls fell, not why.

Seasonality

Short answer: yes, but it's modest. Graffiti reporting has a mild repeating within-year cycle — higher in late winter/early spring, lower in fall — but it is weak relative to the multi-year trend and the COVID shock. (scripts/seasonality.py; monthly series analyzed as calls per day to remove the 28–31-day month-length artifact, with the Apr–Jun 2020 lockdown controlled for.)

Three lenses, deliberately triangulated because they can disagree:

Method	Result	Reading
STL decomposition (period 12)	seasonal strength F_S ≈ 0.48	moderate
Calendar-month profile	peak Feb (214/day), trough Oct (168/day), 1.27×	clear shape
OLS two-way ANOVA (year FE + month dummies)	F = 1.72, p ≈ 0.09	month block not significant at 0.05
SARIMA(1,0,0)(1,0,0)₁₂ + trend + lockdown	seasonal AR₁₂ = 0.29, p ≈ 0.03, ΔAIC +2.6	seasonal term significant but weak

Why the tests disagree, and what's honest: the trend is U-shaped (high 2017 → trough 2019 → rising through 2022), so a naive linear trend misspecifies it and both masks seasonality (OLS) and lets a seasonal term soak up trend (SARIMA). Using year fixed effects in the OLS and a trend + convergent, stationarity-constrained SARIMA fixes this — an earlier non-convergent SARIMA reported a spurious ΔAIC of ~127, which we discarded. The reconciled verdict: seasonality is real but weak — the single seasonal parameter is significant (p≈0.03) while the month block is only borderline (p≈0.09). Trend and COVID explain more of the variation than season does.

Noise complaints & the graffiti–noise relationship

Extends the analysis to noise 311 complaints (service_name like '%Noise%', 45,440 records 2017–2022) and asks whether noise and graffiti move together. (scripts/noise_analysis.py.)

COVID hit the two in opposite directions. During the Apr–Jun 2020 lockdown, graffiti calls fell −31.6% vs 2019 while noise calls rose +81.2% — and for the full year noise was +49.4% vs graffiti's +3.5%. Stuck-at-home life meant fewer people out to notice/report street graffiti but far more friction with neighbors' noise.

They are inversely related, not just during COVID. The monthly per-day rates are negatively correlated — Pearson r = −0.43 raw, −0.45 after removing each series' trend and 12-month seasonality (STL remainders), and still −0.27 with 2020 excluded. So the inverse pattern is modest but real and survives dropping the pandemic year. A plausible common driver — how much time people spend outside vs at home — is exactly what the weather experiment below probes.

Weather

Does weather move complaint volume? Merged daily SF weather (Open-Meteo archive: temperature, precipitation) with each per-day complaint rate and regressed on temperature + precipitation, controlling for the multi-year trend (year FE) and the lockdown. (scripts/download_weather.py, scripts/weather_analysis.py.)

	vs temperature	vs precipitation	model adj R²
Noise	+9.2 calls/day per +10 °F (p < 0.001)	n.s. after controlling temp	0.74
Graffiti	−0.25 bivariate, not significant net of trend/COVID	not significant	0.43

Noise complaints track temperature strongly; graffiti reports don't. Warmer months bring markedly more noise complaints (open windows, outdoor socializing), explaining ~three-quarters of the monthly variation together with trend and lockdown. Graffiti reporting has no robust weather relationship. Noise's raw negative correlation with rain (r = −0.44) disappears once temperature is included — in SF, rain simply coincides with the cold season, so it was the temperature signal in disguise. This temperature sensitivity of noise (and its absence for graffiti) is one reason the two series are inversely related.

Caveat: weather and season are confounded — a warm month is also a summer month — so these coefficients describe a weather/time-of-year association, not a weather effect cleanly separated from the calendar.

Weekly & day-of-week seasonality

The monthly seasonality above was weak; looking at finer weekly resolution (and the raw day-of-week pattern) for both categories tells a sharper, and contrasting, story. (scripts/seasonality_weekly.py.)

	weekly STL strength	week-of-year peak → trough	amplitude	weekend/weekday
Graffiti	0.38	wk 9 (early Mar) → wk 52	1.41×	0.79× (lower on weekends)
Noise	0.35	wk 27 (early Jul) → wk 52	3.48×	1.15× (Saturday peak)

The two are out of phase, on both time scales. Within the year, graffiti reports run high in late winter/early spring and fade through summer, while noise is the mirror image — low in winter, with a big summer peak (and a sharp spike around week 27, the July 4th week). Noise's within-year swing is far larger (3.5× peak-to-trough vs graffiti's 1.4×), consistent with its temperature sensitivity.

Within the week they diverge too: graffiti reports fall on weekends (highest Monday — consistent with commute/ops/business reporting), while noise complaints rise on weekends, peaking Saturday (nightlife and gatherings). Both dip in the final week of the year — partly a holiday reporting lull.

Together the three experiments paint a consistent picture: graffiti and noise are different phenomena — inversely correlated, opposite COVID responses, opposite weather sensitivity, and out-of-phase seasonal and weekly timing.

Reproduce

pip install -r requirements.txt
python scripts/download_data.py     # all categories -> data/raw/sf311_<cat>_2017_2022_<date>.csv
python scripts/download_weather.py  # SF daily weather -> data/raw/sf_weather_2017_2022_<date>.csv
python scripts/analyze.py           # -> data/processed/*.csv, figures/*.png (COVID question)
python scripts/seasonality.py       # -> STL + SARIMA monthly seasonality, figures
python scripts/noise_analysis.py    # -> noise/COVID + graffiti-noise relationship
python scripts/weather_analysis.py  # -> weather vs complaints regressions, figure
python scripts/seasonality_weekly.py # -> weekly + day-of-week seasonality for both
jupyter notebook notebooks/graffiti_covid.ipynb

An optional DataSF app token avoids API throttling: export SODA_APP_TOKEN=...

Layout

data/raw/          NOT committed — regenerate with download_data.py (git-ignored)
data/processed/    aggregated monthly/yearly counts (committed)
scripts/           download_data.py, analyze.py
notebooks/         graffiti_covid.ipynb (narrative)
figures/           generated plots

Raw data is not stored in git. It is fully regenerable from DataSF (vw6y-z8j6) — run python scripts/download_data.py to recreate data/raw/sf311_graffiti_2017_2022_<date>.csv before running the analysis. This keeps the repo lean (git already compresses the CSV, so there's nothing to gain by committing or zipping it) and treats DataSF as the authoritative source.

Notes on the data

• 311 counts measure reports, not graffiti incidence. A drop in calls can reflect fewer people out reporting rather than less graffiti — an important caveat for the COVID question.
• Analysis starts in 2017 to avoid the 2008–2012 reporting ramp-up (mobile/Open311 adoption) that distorts long-run trend comparisons.