Using Forecast and Observed Weather Data to Assess Performance of Forecast Products in Identifying Heat Waves and Estimating Heat Wave Effects on Mortality

BACKGROUND Heat wave and health warning systems are activated based on forecasts of health-threatening hot weather. OBJECTIVE We estimated heat-mortality associations based on forecast and observed weather data in Detroit, Michigan, and compared the accuracy of forecast products for predicting heat waves. METHODS We derived and compared apparent temperature (AT) and heat wave days (with heat waves defined as ≥ 2 days of daily mean AT ≥ 95th percentile of warm-season average) from weather observations and six different forecast products. We used Poisson regression with and without adjustment for ozone and/or PM10 (particulate matter with aerodynamic diameter ≤ 10 μm) to estimate and compare associations of daily all-cause mortality with observed and predicted AT and heat wave days. RESULTS The 1-day-ahead forecast of a local operational product, Revised Digital Forecast, had about half the number of false positives compared with all other forecasts. On average, controlling for heat waves, days with observed AT = 25.3°C were associated with 3.5% higher mortality (95% CI: -1.6, 8.8%) than days with AT = 8.5°C. Observed heat wave days were associated with 6.2% higher mortality (95% CI: -0.4, 13.2%) than non-heat wave days. The accuracy of predictions varied, but associations between mortality and forecast heat generally tended to overestimate heat effects, whereas associations with forecast heat waves tended to underestimate heat wave effects, relative to associations based on observed weather metrics. CONCLUSIONS Our findings suggest that incorporating knowledge of local conditions may improve the accuracy of predictions used to activate heat wave and health warning systems.