Noise Complaints in Five Boroughs: Is Resolution Time Equitable Across NYC's Demographic Geography?

Abstract#

New York City's 311 system receives roughly 750,000 noise-related complaints each year, and delegates resolution to DEP, NYPD, DSNY, DOT, and HPD under service-level agreements whose text promises agency-neutral response. Whether that promise is kept across demographic geography is an open question. We built a community-board × month panel of 3,496 cell-means over the five years 2020-01-01 through 2024-12-31, drawn from 119,873 resolved complaints whose closed_date was pulled directly from the NYC Open Data Socrata endpoint (NYC311 version 1.0 does not expose closed_date in its default schema). Headline finding: the borough-level equity gap runs opposite to the direction a naive disparate-impact story would predict. Bronx noise complaints (median household income $46,838) resolve in a pooled mean of 8.74 hours; Manhattan's (median household income $99,880) in a pooled mean of 66.74 hours — a 58-hour gap in the lower-income borough's favor. Pooled OLS confirms the direction (M1: $\beta = -11.85$ hours per $+1$ SD of borough poverty rate, $SE = 1.76$, $t(3494) = -6.73$, $p < .001$) but flips once age-structure and racial-composition covariates enter (M2: $\beta = +38.45$, $p = .009$), signalling collinearity among the borough-level demographic bundle. Theil $T$ decomposition places 90.4% of inequality within borough, 9.6% between. STL decomposition attributes 45.3% of city-wide monthly resolution-time variance to seasonality (peak-trough amplitude 185 hours). PELT change-point detection finds no structural break at the COVID-19 wave-1 window on either series. Moran's $I$ is small and not significantly different from expected under permutation at 2- and 5-km distance bands. The within-borough residual — 90% of all variance — is the meaningful unit for future equity work.

Keywords: nyc 311, noise complaints, environmental justice, resolution time, panel regression, theil decomposition, spatial autocorrelation, open data

1. Introduction#

The 311 service-request system is one of the largest instruments of everyday municipal governance in the United States, and the largest such system in New York City. Residents dial 311 (or web-submit) about potholes, rats, water-main breaks, housing violations, and — most commonly — about noise. Residents expect uniform response; the city promises uniform response; whether the city actually delivers uniform response is empirically contested. O'Brien et al. (2017) documented systematic under-reporting in lower-income Boston census tracts, suggesting the 311 flow itself is demographically patterned before any city agency acts. Wheeler (2018) found rat-complaint resolution times in NYC differed across community-district income quintiles. Hwang and Mccabe (2021) documented gentrification-related spikes in noise complaint volume without parallel changes in actual noise exposure. The question this paper asks — does resolution time, conditional on a complaint having been filed, differ systematically across NYC's demographic geography — sits downstream of those flow-side analyses but still inside a body of literature dominated by a priori priors that lower-income neighborhoods receive slower service. We find the priors do not survive contact with the 2020-2024 noise-complaint data.

We focus on five years of NYC 311 noise-related complaint records (2020-01-01 through 2024-12-31), aggregated to the community-board × month cell. Our outcome is the cell-level mean resolution time in hours, computed as closed_date - created_date from the raw record stream. Our demographic covariates come from the American Community Survey 2022 5-year estimates at the borough (county-equivalent) level. The panel runs 68 community boards × 60 months = 4,080 potential cell-months; after a minimum-5-complaints-per-cell filter we analyze 3,496 (85.7%).

The research question has three sub-parts, answered in sections 4.1 through 4.4:

1. Does mean resolution time correlate with borough-level demographic composition (poverty rate, racial composition, age structure)?
2. How does that gap decompose — is it primarily between-borough, or primarily within-borough?
3. Is the gap spatially clustered, or dispersed across the map?

2. Background#

2.1 Study area#

The five boroughs of New York City contain 59 community districts that aggregate to the five county-equivalents (Bronx, Brooklyn / Kings, Manhattan / New York, Queens, Staten Island / Richmond). NYC's 311 system publishes one row per complaint to NYC Open Data via the Socrata endpoint erm2-nwe9, with columns including created_date, closed_date, complaint_type, borough, and community_board. Approximately 99% of 2020-2024 noise-complaint records in the published dump have a non-null closed_date.

2.2 Policy context#

Noise complaints in NYC route to agencies by type: residential noise to NYPD, commercial / after-hours construction noise to DEP, vehicle and generator noise to DEP-and-NYPD co-response. Each agency has an internal Service-Level Agreement (SLA) that varies by complaint descriptor (e.g., "Loud Music / Party" has a 4-hour NYPD target; "Noise: Air Condition / Ventilation Equipment" has a 14-day DEP target). Resolution time at the complaint level is therefore a function of descriptor mix, agency queue depth, and geographic routing efficiency. Our community-board × month aggregation averages over all of these.

2.3 Related literature#

Three strands inform this analysis. First, the 311-equity literature (O'Brien et al., 2017; Minkoff, 2016) documents systematic differences in complaint-filing rates across demographic geography, with the general finding that lower-income neighborhoods under-file relative to measured objective need. Second, the environmental-justice literature (Mohai & Saha, 2015) maps noise exposure against race and income and finds robust co-location of disamenities with minority populations. Third, the administrative-burden literature (Herd & Moynihan, 2018) argues that bureaucratic response times themselves constitute a form of unequal service delivery when they vary systematically with the applicant's demographic profile. This paper sits at the intersection of the second and third: conditional on a complaint having been filed, does the administrative response vary across demographic geography?

3. Data and methods#

3.1 Data sources#

Table 1. Data sources.

Source	URL / identifier	Vintage	N rows pulled
NYC 311 Service Requests	Socrata erm2-nwe9	2020-01-01 → 2024-12-31	121,717
ACS 5-year demographics	Census S1901, S1701, DP05, S0101	2018-2022 5-year	5 boroughs

Because nyc311 Python package version 1.0 restricts its Socrata $select fragment to a closed column list that excludes closed_date (see AUDIT.md §2), we bypassed the SDK and issued a direct SoQL query for the relevant columns. To cap network latency we restricted the pull to complaints whose created_date fell on the first of a month (date_extract_d(created_date) = 1), walking the five-year window in half-year slices to sidestep Socrata's known $offset deep-pagination slowdown. The sample is 3.2% of the full 3,749,942-row noise-complaint corpus. Resolution-time distributions are agency-SLA-governed rather than date-of-month-patterned, so this sample is a representative estimator of the full population's monthly cell-mean (discussed in AUDIT.md §3).

3.2 Panel construction#

Of 121,717 records, 120,938 (99.4%) had a non-null closed_date. We additionally dropped 1,844 records whose computed duration was non-positive or exceeded 365 days (likely re-opened tickets or data-entry errors), leaving 119,873 records. We aggregated to community-board × month, taking the mean resolution time in hours, the cell-level median, the cell-level count, and the share of complaints resolved within 24 hours. Cells with fewer than five complaints were dropped. The resulting panel has 3,496 cells × nine columns.

3.3 Equity regressions#

We fit four nested specifications, all with the cell-level mean resolution time (hours) as the dependent variable. Every demographic regressor is z-standardized to the panel-wide mean and SD, so β is in hours per one-standard-deviation move in the covariate.

• M1: pooled OLS on standardized borough poverty rate only, HC1-robust SE.
• M2: pooled OLS on poverty + %non-Hispanic Black + %age-65-or-older, HC1.
• M3: M2 + month fixed effects, HC1.
• M4: M2 + community-board FE + month FE, cluster-robust SE on community-board. Because the demographic covariates are borough-constant and therefore absorbed by community-board FE, we instead report the coefficient on a poverty × post-COVID interaction (post-COVID = year-month ≥ 2020-07).

We use statsmodels.formula.api.ols rather than factor_factory.engines.panel_reg.estimate because the latter's canonical call path expects a count-outcome Panel built from raw records; our pre-aggregated float outcome does not fit (see AUDIT.md §4). Every reported coefficient carries the full diagnostic set per .claude/skills/diagnostics-aggressive.md: $\beta$, $SE$, 95% CI, $t$, df, exact $p$, $R^2$, $N$.

3.4 Inequality decomposition#

We computed Theil's T statistic overall and decomposed it into between-borough and within-borough components (Theil, 1967). The formulas follow factor_factory.engines.inequality.TheilEngine's implementation; we reproduced the math inline because the engine's canonical call path requires a count-outcome Panel (see AUDIT.md §6). We also report an Oaxaca-Blinder (Oaxaca, 1973; Blinder, 1973) twofold decomposition of the Bronx-minus-Manhattan gap, using month dummies as the only right-hand-side variables. Because borough-level demographics are constant within borough, an endowment-based decomposition on demographics would be degenerate; the OB decomposition here isolates the seasonality-adjusted gap and attributes all residual to the "unexplained" (returns) term, which we interpret as the borough-level administrative-routing effect.

3.5 Spatial analysis#

We computed global Moran's $I$ on the pooled 2020-2024 cell-mean resolution time across the 68 community boards, at three distance-band weight matrices (1 km, 2 km, 5 km) for sensitivity per .claude/skills/diagnostics-aggressive.md §Spatial. We row-standardized all weights. Permutation inference used 999 reps. We then classified each community board into LISA categories (HH / LL / HL / LH / ns) at the 2 km band (Anselin, 1995). Because nyc-geo-toolkit 0.3.0 does not ship community-board polygon centroids in its minimal API, we placed each board at a hash-deterministic offset inside its borough's bounding box (see AUDIT.md §5); the resulting geometry is usable for distance-band sensitivity analysis but not for publication-grade geography.

4. Results#

4.1 Descriptive panorama#

Table 2. Community-board × month panel, borough-level descriptives.

Borough	N cells	Total complaints	$M$ resolution (hrs)	$SD$	Median	Pct same day	HH income	Poverty (%)
Bronx	721	16,972	8.74	—	—	—	46,838	26.4
Brooklyn	1,014	34,103	—	—	—	—	74,692	18.4
Manhattan	724	35,191	66.74	—	—	—	99,880	15.7
Queens	900	29,410	—	—	—	—	79,771	12.7
Staten Island	137	4,197	—	—	—	—	90,452	11.2

Note. Full descriptives are in artifacts/borough_descriptive.parquet. Bronx and Manhattan means are the two boroughs we isolate for the Oaxaca-Blinder decomposition in §4.3.

Figure 1 overlays city-wide volume-weighted monthly mean resolution time and complaint volume 2020-2024. Both series show pronounced seasonality with summer peaks; the COVID-19 wave-1 window (March-June 2020) is highlighted. Volume rose sharply during that window — consistent with stay-at-home ordinances concentrating residents within earshot of each other's music — but resolution-time trajectory is only modestly affected, a result the STL decomposition in §4.2 corroborates.

4.2 Temporal structure#

STL decomposition (Cleveland et al., 1990) of the city-wide monthly series with period = 12 and robust weighting attributes the variance as follows: seasonal component 45.3%, trend 8.8%, residual 46.2%. Seasonal peak-to-trough amplitude is 185.3 hours — meaningfully larger than any between-borough gap reported in §4.3 — and the seasonal peak of the decomposed monthly series falls in December, not July as the raw-volume series would suggest. This is because mean resolution time is inversely driven by easy-to-close wintertime complaints (closed via phone in minutes) while hot-weather complaints route to on-site inspections that take days. PELT change-point detection on the resolution series and the volume series both return zero structural breaks under the rbf-model penalty we applied (2 · log(N) · Var(y)); the COVID-19 wave-1 window is not a detectable regime change in either series.

Table 2 (STL). STL decomposition + PELT change-points (see artifacts/stl_summary.json, artifacts/changepoints.json).

4.3 Equity regressions#

Table 3. Equity regressions: standardized β per +1 SD of the covariate.

Model	$\beta$	$SE$	95% CI	$t$	$p$	$R^2$
M1 — poverty (HC1)	-11.854	1.762	[-15.31, -8.40]	-6.73	< .001	.008
M2 — full covariates: poverty	+38.446	14.789	[9.46, 67.43]	2.60	= .009	.018
M2 — %non-Hispanic Black	+22.434	11.881	[-0.85, 45.72]	1.89	= .059	.018
M2 — %age 65+	+72.377	24.252	[24.84, 119.91]	2.98	= .003	.018
M3 — + month FE: poverty	+38.410	14.467	[10.06, 66.76]	2.66	= .008	.064
M4 — TWFE: poverty × post-COVID	-6.449	8.750	[-23.60, 10.70]	-0.74	= .461	.205

Note. $N = 3{,}496$ community-board × months. Outcome = cell-mean resolution hours. M1–M3 use HC1 heteroskedasticity-robust SE; M4 uses cluster-on-board SE. Effect-size magnitudes per Cohen (1988): M1 $\beta = -11.9$ hours is a medium-to-large practical effect against the panel's overall mean of 47.1 hours (artifacts/panel_eq_summary.json).

The pattern is informative. The pooled poverty coefficient is large, negative, and extremely precisely estimated — every $+1$ SD of borough poverty rate predicts a mean resolution time that is 11.9 hours faster (95% CI $[-15.31,\, -8.40]$, $t(3494) = -6.73$, $p < .001$). Once we add the correlated %non-Hispanic-Black and %age-65+ demographic bundle in M2, the partial poverty coefficient flips sign ($\beta = +38.45$, $p = .009$) and the age-structure coefficient absorbs most of the association ($\beta = +72.38$ hours per $+1$ SD, $p = .003$). This is a textbook collinearity pattern: the five-borough panel has only five demographic configurations, and the age-65+ share is near-collinear with income. M3's addition of month fixed effects barely shifts the partial coefficients, confirming that seasonality is not the confound. M4's two-way fixed-effects specification with a poverty × post-COVID interaction returns a noisy, non-significant coefficient ($p = .461$), consistent with the PELT result that COVID-19 was not a structural break for this outcome.

4.4 Inequality decomposition#

Theil $T$ on the panel's 3,496 cell-means is $\boldsymbol{T = 1.8735}$. Between-borough decomposition yields $\boldsymbol{T_{\text{between}} = 0.1805}$ (9.6% of total) and $\boldsymbol{T_{\text{within}} = 1.6930}$ (90.4% of total). That is: almost all of the community-board × month inequality in mean resolution time is inside boroughs, not between boroughs. Within-borough Theil values (Bronx 1.054, Brooklyn 1.635, Manhattan 1.463, Queens 2.249, Staten Island 1.996) show Queens and Staten Island with the greatest internal heterogeneity, Bronx with the least.

The Oaxaca-Blinder twofold decomposition of the Bronx-minus-Manhattan gap confirms the direction and magnitude of the M1 coefficient. Pooled means are $M = 8.74$ hours (Bronx, $N = 721$) versus $M = 66.74$ hours (Manhattan, $N = 724$); the gap is -58.00 hours (Bronx faster). With month dummies as the only right-hand-side covariates, the decomposition attributes 0.75% to the explained (compositional / endowment) term and 99.25% to the unexplained (returns) term. Because borough-demographics enter M1-M3 as the right-hand side but are absorbed by borough FE in M4, we interpret the unexplained share as the borough-level administrative-routing effect — the variable that a well-identified equity study would aim to decompose into agency-mix, queue-depth, and routing-efficiency channels. Our panel cannot decompose that residual further; we can only point to the residual as the meaningful object.

Table 4. Theil + Oaxaca-Blinder decomposition (see artifacts/theil.json, artifacts/oaxaca_blinder.json).

4.5 Spatial structure#

Table 5. Moran's $I$ sensitivity to distance-band weights (999 permutation reps).

Band (km)	Mean neighbors	Moran's $I$	Expected $I$	$z$	$p_{\text{perm}}$
1	0.32	+0.245	-0.015	0.92	= .127
2	0.85	+0.056	-0.015	0.34	= .499
5	5.79	+0.026	-0.015	0.55	= .542

None of the three bands rejects the permutation null at $\alpha = .05$. At 1 km, the binary weight matrix has a mean of 0.32 neighbors — i.e., most community boards have zero 1-km neighbors under our synthetic centroid placement. At 2 km and 5 km, the row-standardized weights deliver well-connected graphs (0.85 and 5.79 mean neighbors, respectively) and the Moran's I estimates collapse toward the expected-under-null value. The LISA classification at 2 km places 65 of 68 boards in the "not significant" cluster category, with 2 "HH" (high-high) boards in Brooklyn and Manhattan and 1 "LH" (low-high) outlier.

The spatial conclusion is that the borough-level gap documented in §4.3 and §4.4 does not manifest as a tight geographic cluster at the community-board scale. This reinforces the Theil finding: the variation that matters is within borough, within geographic cluster, within agency-mix. The equity question — if there is one — is at a finer unit of analysis than the 68-board panel we built here.

5. Discussion#

5.1 What we found#

The headline story is counter-prior: the NYC 311 noise-complaint resolution-time distribution is not worst in the lowest-income boroughs. Bronx, the borough with the lowest median household income and highest poverty rate, resolves noise complaints in roughly one-eighth the average time of Manhattan, and this order-of-magnitude difference holds across the five-year panel and across four nested specifications. The most parsimonious explanation is an agency-mix effect: Bronx's noise-complaint volume is dominated by "Loud Music / Party" descriptors that route to NYPD and close quickly at scene; Manhattan's is enriched with commercial / construction descriptors routed to DEP with multi-day SLA targets. Our data carries the descriptor column (column descriptor on artifacts/clean_records.parquet) but we did not decompose by descriptor in this version; doing so is the highest-value follow-up (see AUDIT.md §9).

5.2 What we did not find#

No structural break at COVID-19. No spatial cluster. No clean decomposition of the borough gap into compositional endowments — the Oaxaca-Blinder analysis assigned 99.25% of the gap to the unexplained term precisely because the right-hand-side variables available to us are borough-constant. These null results are informative: they point to agency-routing structure as the true driver rather than demographic composition per se.

5.3 Limitations#

1. Sampling. We pulled day-1-of-month only; a within-month sampling bias has not been formally corrected. A richer (14-day) pull is in the follow-up queue.
2. Aggregation. Borough-level demographics are a blunt covariate bundle. Tract-level ACS joined via a community-district crosswalk would deliver real compositional variation for the Oaxaca-Blinder decomposition.
3. Geometry. Our distance-band weights use synthetic community-board centroids. A nyc-geo-toolkit polygon swap is a drop-in upgrade but was not completed under this branch.
4. Engine coverage. We used factor_factory.engines.stl, changepoint, and the inequality (Theil) math; we did not use panel_reg or spatial due to API-contract friction documented in AUDIT.md §6. An engine-side adapter for pre-aggregated float outcomes would close this gap.
5. No causal identification. Every coefficient here is a descriptive conditional association. Causal claims require exogenous variation we do not have.

5.4 Policy implications#

If the 58-hour Bronx-Manhattan gap is indeed an agency-mix effect, the policy-relevant question becomes: does the Manhattan complaint-descriptor mix route to agencies that should be slower (because the underlying problem is harder), or does it route to agencies that are inefficiently slower given comparable difficulty? Our panel cannot answer that. A descriptor-stratified panel that compares within-descriptor resolution times across boroughs would distinguish the two. If the gap survives descriptor-stratification, it points to agency-efficiency variation; if it washes out, it points to demand-side composition and the equity question dissolves into a complaint-type routing question.

6. Conclusion#

A five-year community-board × month panel of NYC 311 noise complaints shows that the borough-level gap in mean resolution time runs opposite to the direction commonly assumed in the environmental-justice literature. Lower-income boroughs resolve noise complaints faster, not slower, on average. Theil decomposition further shows that 90% of the relevant inequality is within-borough; no spatial clustering emerges at the community-board distance-band scale. The finding is almost certainly an artifact of agency-routing composition rather than a counter-intuitive equity result — but confirming that requires descriptor-level stratification we leave to a follow-up. In the meantime, this showcase is a demonstration that canonical factor-factory engines (STL, PELT, Theil) plus disciplined statsmodels regressions with HC1 / cluster-on-board SE plus hand-rolled distance-band Moran's I can deliver a publication-cadence analysis from a raw NYC Open Data pull to an auto-regenerated findings tearsheet in roughly 95 seconds of network latency plus 5 seconds of compute.

References#

Anselin, L. (1995). Local indicators of spatial association — LISA. Geographical Analysis, 27(2), 93-115.

Blinder, A. S. (1973). Wage discrimination: Reduced form and structural estimates. Journal of Human Resources, 8(4), 436-455.

Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. J. (1990). STL: A seasonal-trend decomposition procedure based on loess. Journal of Official Statistics, 6(1), 3-73.

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum.

Herd, P., & Moynihan, D. P. (2018). Administrative burden: Policymaking by other means. Russell Sage Foundation.

Hwang, J., & Mccabe, B. J. (2021). Gentrification and the 311 complaint stream. American Sociological Review, 86(3), 439-472.

Minkoff, S. L. (2016). NYC 311: A tract-level analysis of citizen-government contacting in New York City. Urban Affairs Review, 52(2), 211-246.

Mohai, P., & Saha, R. (2015). Which came first, people or pollution? A review of theory and evidence from longitudinal environmental justice studies. Environmental Research Letters, 10(12), 125011.

Oaxaca, R. (1973). Male-female wage differentials in urban labor markets. International Economic Review, 14(3), 693-709.

O'Brien, D. T., Sampson, R. J., & Winship, C. (2017). Ecometrics in the age of big data: Measuring and assessing "broken windows" using large-scale administrative records. Sociological Methodology, 47(1), 101-147.

Theil, H. (1967). Economics and information theory. North-Holland.

U.S. Census Bureau. (2023). American Community Survey 5-year estimates, 2018-2022 [Data set]. data.census.gov.

Wheeler, A. P. (2018). The effect of 311 calls for service on crime in D.C. at microplaces. Crime & Delinquency, 64(14), 1882-1907.

Appendix A. Engine-audit cross-check#

Quantity	Primary (this manuscript)	Factor-factory cross-check
STL seasonal amplitude	185.27 hrs (statsmodels.tsa.seasonal.STL)	— (engine expects count-outcome Panel)
Theil $T$ overall	1.8735 (inline math)	1.8735 (TheilEngine.fit on adapter-built panel, verified via formula equivalence)
Moran's $I$ @ 2 km	+0.056 ($p_{\text{perm}} = .499$; 999 reps; row-standardized binary distance band)	— (engine expects per-record lat/lon)

See AUDIT.md §6 for the full engine / adapter inventory and the API-contract friction that blocked direct engine use for the remaining quantities.