USGS PAGER fatality probabilities for 9,130 earthquakes in 2013-2026 gave 9 times too many events with deaths, mainly because PAGER 2.x sets zero modelled deaths to 0.5

We tested the fatality probabilities that the USGS PAGER system issued for 9,130 earthquakes from 2013 to mid-2026 against deaths in the NCEI/WDS Global Significant Earthquake Database. We used the first archived version (median 22.6 minutes after origin) and later versions from the ComCat archive. The first versions gave 2,399.6 expected events with at least 1 death; NCEI records 264 (ratio 0.11). The ratios were 0.25 for at least 10 deaths, 0.66 for at least 100 (34 against 51.5, p = 0.0015) and 0.63 for at least 1,000 (10 against 15.9, p = 0.08). Most of the error comes from one rule in the PAGER 2.x code (in use from April 2017): when the model gives fewer than 0.001 deaths, PAGER sets the expected count to 0.5. Thus 94% of PAGER 2.x forecasts (6,373 events) stated P(at least 1 death) = 0.24-0.35 (mean 0.327), but only 1.7% (95% CI 1.4-2.1%) of these events had deaths. The rule also makes events with no modelled deaths get 4 times the probability of events with 0.01-0.1 modelled deaths. PAGER 1.x had no such floor. Forecasts with positive model estimates also gave about 2 times too many events at all thresholds, and later versions were not better at 100 or more deaths. The issued forecasts had a negative ranked probability skill against climatology (-1.58). When we replaced only the floor forecasts by leave-one-year-out frequencies, the skill became +0.06 (CI -0.05 to +0.15) for first versions and +0.14 (0.04 to 0.24) by 24 hours, and for magnitude 6 or more it was better than a magnitude-depth model. Limitations: NCEI misses some events with 1-9 deaths (a Wikipedia union check changes 264 to 310 deadly events, with the same conclusions), counts include some non-shaking deaths, and only 34 events had 100 or more deaths.

earthquake loss estimation · PAGER · forecast calibration · fatalities · probabilistic forecast verification · ranked probability score · NCEI significant earthquakes · ComCat · rapid impact assessment · reliability