Pharos— County Mortality Explorer

County-level mortality outcomes, every ingested covariate as a map layer, spatial hot/cold spots, and model attributions — from the Pharos pipeline
Read this first — ecological data. Every number here describes a county population, not any person in it. County-level associations do not transfer to individuals (the ecological fallacy), and rates in low-population counties are statistically noisy: year-to-year swings are often not meaningful. If you or someone you know needs support: call or text 988.

🖱 Scroll to zoom · drag to pan · double-click to reset · click a county for its trend. Use the toolbar (top-right of the map) to zoom or reset.

Rate vs covariate

County trend — click a county on the map

Sources are separate series — NVSS suicide (blue) and the IHME self-harm + violence aggregate (orange) measure different things. Small-county swings are mostly noise.

What each outcome is

Suicide (CHR/NVSS): age-adjusted death-certificate rates, pooled ~5-year windows, with 95% CIs, deaths and person-years. CDC small-count suppression applies. IHME self-harm + interpersonal violence: modeled small-area estimates with 95% uncertainty intervals, 2000–2019. IHME publishes only this combined aggregate at county level — it includes homicide, and the suicide/homicide mix differs sharply across race/ethnicity groups. IHME substance use disorders: same model family, included as a related outcome.

Spatial statistics

Hot/cold spots are Local Moran's I (LISA) labels on the IHME self-harm + violence rate, queen contiguity, 999 permutations, Benjamini-Hochberg corrected per year. Global Moran's I ≈ 0.63 (2019, p=0.001). A hot spot is where high rates cluster spatially — a pattern, not an explanation.

Model

XGBoost on 33 ElasticNet-selected features, 62,220 county-years, 5-fold CV R² shown in the attribution panel. Fit to the IHME aggregate (the longest usable outcome series) — not to suicide alone. SHAP values are per county-year and stored in the pipeline DB.

Denominators & color

CHR hover shows pooled deaths and person-years; IHME publishes modeled rates without denominators — hover shows ACS population for scale only. Rates: viridis, 2–98 percentile bounds per view. Red is reserved for intervals entirely above national; purple/teal outlines are LISA hot/cold spots; gray is insufficient data. Series colors follow the outcome, never rank.

Collinearity caution

Poverty, education, race, and rurality are highly collinear in US county data (46 of 102 features have VIF ≥ 5; the climate temperature trio exceeds 20,000). SHAP and GWR attributions for collinear features are unstable — shared variance cannot be assigned to one member of a correlated set. Check /api/diagnostics/vif before interpreting.

Still absent

Suicide-only county series beyond CHR's window (needs IHME's 1980–2014 county dataset — drop its MX files in data/raw/ihme_1980_2014/ — or CDC WONDER).