Skip to main content
Public Health· 18-page report· 3 figures

Streamflow & West Nile Virus Prediction

Test whether antecedent streamflow improves California county-level WNV outbreak forecasts, finding negligible predictive gain beyond climate while validating 10 surveillance priorities.

What this research found

Spring water conditions are thought to shape mosquito breeding, so this study tested whether April–June streamflow anomalies improve county-level West Nile virus outbreak forecasts in California. Eighteen years of case surveillance, 703 audited stream gauges, and Daymet climate data were combined into a mixed-effects logistic model validated by holding out one year at a time. The answer was no: a climate-only baseline already reached 0.813 average precision, and adding hydrology moved it by 0.0007 — far short of the pre-set 0.05 bar for operational use.

  • Streamflow added almost nothing to prediction. The climate-only baseline reached an out-of-sample AUPRC of 0.813 and ROC-AUC of 0.940; the hydrology-augmented model reached 0.814, a gain of 0.0007 against a pre-set gate of 0.05.
  • Spring warmth dominates the signal. Cumulative April–June growing degree days above 10 °C carry an odds ratio of roughly 8.5 per standard deviation (p ≈ 6×10⁻⁶⁵) for a county-season outbreak.
  • Two streamflow features are statistically detectable yet not predictively useful: dry-to-wet pulse occurrence (odds ratio 0.65) and pulse magnitude (odds ratio 1.54), both surviving false-discovery correction at p = 0.0038.
  • Outbreaks are frequent in the record. 240 of 1,044 county-seasons from 2006 to 2023 saw at least five human cases between July and December — a 23% base rate across 6,360 reported positive cases statewide.
  • The 10 prioritized surveillance counties — Stanislaus, Kern, Fresno, Los Angeles, Tulare, Orange, San Joaquin, Sacramento, Riverside, and Butte — closely track observed history. Stanislaus carries a mean predicted probability of 0.985 against an outbreak in all 18 years.

How it was done

California Department of Public Health arbovirus surveillance was reduced to a binary target — at least five human cases in a county between July and December — over 58 counties and 18 seasons. USGS daily discharge records were audited down to 703 gauges with at least 80% coverage of the 6,574-day window (4,545,375 daily observations), standardized to day-of-year z-scores, and turned into April–June predictors for the proportion of low-flow days and for dry-to-wet flow pulses. Gauges were assigned to counties by point-in-polygon join, with a nearest-neighbour fallback for the four counties holding no gauge, while Daymet single-pixel climate at county centroids supplied spring growing degree days and precipitation. Two mixed-effects logistic regressions — climate-only and climate-plus-hydrology, each with county and year random intercepts — were compared under leave-one-year-out cross-validation with the scaler refitted on training years only, and the result was written up as an 18-page paper with 42 verified references.

Data sources

  • California Department of Public Health arbovirus surveillance — West Nile virus cases 2006–2023, 6,360 reported positive cases
  • USGS National Water Information System — daily mean discharge; 2,397 California gauges enumerated, 703 retained, 4,545,375 daily observations
  • Daymet single-pixel climate API (ORNL DAAC) — daily temperature and precipitation at 58 county centroids, 2006–2023
  • US Census Bureau 2023 gazetteer centroids and 2023 county boundary polygons
  • 42 DOI-verified references on West Nile virus ecology, degree-day models, and hydrological drivers

Limitations

The mixed model was fitted by variational Bayes, which underestimates posterior variance and leaves both versions slightly overconfident, with calibration slopes of 0.73 to 0.74 instead of the ideal 1.0. Four ungauged counties borrow hydrology from neighbours 42–61 km away, climate comes from a single pixel per county rather than an areal average, and human case counts carry reporting and testing biases that vary by county and year.

Figures from this analysis

How this research was produced

K-Dense Web planned and ran this public health investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.

Share:
Public Health

Hantavirus: The Virus in the Dust

13-slide presentation on hantavirus covering global epidemiology, ecology-climate drivers, U.S. surveillance with provisional-data caveats, clinical course, and prevention.

Public Health

Compound Heat & Health Grant Targeting

Rank 3,109 U.S. counties by compound heat, chronic disease, and population vulnerability to allocate a hypothetical $10M research portfolio under uncertainty.

Public Health

Construction Worker PM2.5 Exposure

Combine EPA air monitoring with BLS labor data to estimate 18.95M construction worker-days above the PM2.5 standard and prioritize 20 counties for protection studies.

Run this kind of analysis on your own question

Try K-Dense Web free and see how an AI co-scientist accelerates your research.