Depression and Self-Harm Emergency Care Clusters Mapped Across San Francisco Youth

Mental health emergency care among San Francisco teenagers is not spread evenly across the city. A cross-sectional spatial analysis of service utilization among transitional age youth (TAY), defined in the study as residents aged 15 to 19 years, found that rates of depression-related and self-harm-related emergency department visits were significantly clustered across the city’s 18 ZIP code areas between fiscal years 2017 and 2021. The analysis, published in Discover Pediatrics, used pre-aggregated data from the San Francisco Department of Public Health Maternal, Child and Adolescent Health dataset and formal spatial statistics to test whether service rates were randomly distributed or organized into recognizable geographic patterns. The answer, the authors report, was clearly non-random, with a consistent gradient running from high-utilization neighborhoods in the southeast of the city toward low-utilization areas in the northwest.

The study examined five outcome domains: depression-related emergency department visits, depression-related hospitalizations, self-harm-related emergency department visits, self-harm-related hospitalizations, and a combined measure of total mental health service utilization summing all four. Rates were calculated per 1,000 TAY residents in each ZIP code area. Total mental health service rates ranged widely, from 10.8 to 37.0 per 1,000 residents, with a mean of 23.31. Depression-related emergency visit rates ranged from 2.0 to 14.0 per 1,000, self-harm emergency visits from 2.5 to 13.4, depression hospitalizations from 2.1 to 10.0, and self-harm hospitalizations from 1.8 to 6.2. The 18 ZIP code areas analyzed represented the complete set of 2010 Census ZCTA boundaries for San Francisco that also appeared in the health department dataset; no areas with available data were excluded.

To test for clustering, the researchers applied Moran’s I, a statistic that measures global spatial autocorrelation, or whether values that are geographically close tend to be similar. Under complete spatial randomness with 18 units, the expected value of Moran’s I is −0.059. Three of the five outcomes showed statistically significant positive autocorrelation. Depression-related emergency department visit rates showed the strongest clustering, with Moran’s I of 0.424, a Z score of 3.34 and p less than 0.001. Self-harm emergency visit rates had Moran’s I of 0.292 (p = 0.008), and total service utilization rates Moran’s I of 0.229 (p = 0.021). In contrast, depression hospitalization rates (Moran’s I = 0.111, p = 0.127) and self-harm hospitalization rates (Moran’s I = 0.135, p = 0.141) did not reach statistical significance.

Local clustering was assessed with Local Indicators of Spatial Association and Getis-Ord Gi* statistics, using Queen contiguity weights with row standardization. The Gi* hot spot analysis identified five significant hot spots for total mental health service utilization. Potrero Hill (94107, Z = +2.36) and South of Market (94103, Z = +2.34) were hot spots at the 99 percent confidence level, Mission (94110, Z = +1.82) at 95 percent, and Bayview–Hunters Point (94124, Z = +1.62) and Visitacion Valley (94134, Z = +1.59) at 90 percent. The LISA analysis independently confirmed South of Market, with a rate of 25.8 per 1,000, and Potrero Hill, at 23.6 per 1,000, as the only two statistically significant high-high clusters; the remaining 16 ZIP code areas showed no significant local clustering by that method.

Cold spots, where disproportionately few residents per capita reached emergency services, were concentrated in the western and Richmond neighborhoods. Sunset/Parkside (94116, Z = −1.76), Inner Richmond (94118, Z = −1.67) and Inner Sunset (94122, Z = −1.66) were significant at 95 percent confidence, while West of Twin Peaks (94127, Z = −1.51) and Outer Richmond (94121, Z = −1.48) reached 90 percent. Eight ZIP code areas, 44 percent of the total, showed no statistically significant clustering at all. Sensitivity analyses using k-nearest neighbor weighting schemes with k of 4, 5 and 6 produced results consistent with the primary specification: depression and self-harm emergency visit rates and total service rates remained significantly clustered under all four weighting methods, and Potrero Hill, South of Market, Mission and Bayview–Hunters Point were hot spots in every specification.

The divergence between emergency department and hospitalization outcomes is a notable technical feature of the results. The authors suggest that California’s per-capita psychiatric inpatient bed capacity was below the national average during the study period, and that because psychiatric bed availability correlates negatively with hospitalization rates, bed supply may have constrained the hospitalization signal relatively uniformly across ZIP code areas. Emergency department visits, by contrast, face no such capacity ceiling and may therefore act as more sensitive geographic indicators of mental health burden. The authors caution, however, that they cannot determine whether the divergence reflects the true prevalence of severe psychiatric illness in the areas concerned.

A supplementary analysis using American Community Survey sociodemographic variables found only two statistically significant Spearman rank correlations across the 18 ZIP code areas. The uninsured rate among adults aged 35 to 64 was positively correlated with depression emergency visit rates (ρ = 0.50, p = 0.033), and Asian racial composition was negatively correlated with depression hospitalization rates (ρ = −0.53, p = 0.029). Several other associations approached but did not reach significance, including the uninsured rate and bachelor’s degree attainment against total service rates, and poverty, unemployment, median household income, rent burden, median age and foreign-born proportion showed no meaningful associations with any outcome. The authors describe most of these supplementary results as underpowered.

The study carries an important interpretive constraint that the authors emphasize repeatedly: ZIP code-level service utilization rates result from a combination of underlying need and access to care, and the cross-sectional, aggregate design cannot disentangle the two. A high-utilization area could signal genuinely elevated mental health burden, relatively better geographic or financial access to services, or both simultaneously. Conversely, a cold spot may reflect lower true prevalence, better-functioning primary and outpatient care, or substantial unmet need suppressed by access barriers, stigma or service scarcity. Because the dataset captures only people who accessed services, populations facing structural barriers may not appear in the records at all, meaning the true burden of mental health need across San Francisco neighborhoods is likely greater than the reported figures.

Several further limitations shape how the findings should be read. This is an ecological analysis, so group-level associations cannot be directly inferred to individuals without risking the ecological fallacy. ZIP codes are administrative units not designed to reflect meaningful social environments, and Moran’s I values and cluster designations are sensitive to the scale and boundaries chosen. With only 18 spatial units, statistical power is limited, and the authors anticipate roughly 0.9 false discoveries by chance at the α = 0.05 threshold. The dataset spans the onset of the COVID-19 pandemic, which altered both mental health burden and care access, and because the data were pre-aggregated across the full five-year period, year-level disaggregation was not possible. The restriction to 15-to-19-year-olds also limits generalizability, and the cross-sectional design precludes any causal inference.

Within those constraints, the authors argue the analysis fills a documented gap: no prior study had formalized spatial clustering of mental health service utilization across San Francisco ZIP code areas, despite well-documented equity differences in the city, which has concentrated pockets of homelessness, poverty and ethnoracial segregation and one of the nation’s highest income inequality levels. The authors note that California county behavioral health plans have traditionally organized services by program or population type rather than geographic targeting, and that the new Behavioral Health Services Act calls for equity metrics but does not yet require neighborhood-level disaggregation of planned resources. They position the mapped hot spot corridor of South of Market, Potrero Hill, Mission, Bayview–Hunters Point and Visitacion Valley, and the western cold spots, as an empirical foundation for spatially explicit service planning, whether the goal is reducing high utilization in hot spots or investigating potentially suppressed utilization in cold spot areas.

Subject of Research: Spatial clustering of mental health service utilization among transitional age youth in San Francisco

Article Title: Spatial clustering of depression and self-harm related service utilization among transitional age youth in San Francisco

Article References: Umer, A., & Shamim, A. (2026). Spatial clustering of depression and self-harm related service utilization among transitional age youth in San Francisco. Discover Pediatrics, 1(1), Article 3. https://doi.org/10.1007/s44571-026-00001-9

Image Credits: AI Generated

DOI: 10.1007/s44571-026-00001-9

Keywords: spatial epidemiology, mental health services, transitional age youth, San Francisco, emergency department visits, self-harm, depression, Moran's I, hot spot analysis, health geography, ZIP code areas, service utilization