Reasonable and timely prediction of earthquake casualties has become an important factor for government emergency rescue decision. Due to the sudden occurrence and uncertainty of earthquake disaster, the categories of death, serious injury, and minor injury are quite different. A joint Poisson mixed modeling approach can handle these categorical data for early prediction of earthquake casualties. As an area of frequent earthquakes, Yunnan province of China shows the characteristics that minor and moderate earthquakes may cause major disasters. Based on the data of the destructive earthquakes of Yunnan in 1992-2018, we conducted an empirical study using a joint Poisson mixed modeling approach for clustered multinomial data with random cluster sizes. We considered the effects on casualties of earthquake magnitude, depth of epicenter, time of earthquake and location of epicenter, which can be monitored by government departments at early time. The study found that death, serious injury, and minor injury caused by the Yunnan earthquake were positively related to the magnitude of the earthquake, negatively correlated with the earthquake depth and the epicenter distance from county center. In addition, the Yunnan earthquakes showed the sinusoidal periodicity of the month.

The above empirical study has found that in addition to being able to well predict the number of death, serious injury and minor injury, the new model is also very effective in detecting outliers.