Bias in fitting the ETAS model: a case study based on New Zealand seismicity

David S. Harte

Published December 19, 2012, SCEC Contribution #6082

We fit various forms of the ETAS model to a large region that includes all of the most seismically active areas of New Zealand. The ETAS model contains two components: a component describing background or immigrant events, and a part describing aftershocks of the background events and aftershocks of the aftershocks. We refer to the first part as the background part and the second as the ETAS part. Generally all of the sophistication, and the bulk of the model parameters, lies in the ETAS part of the model. The background component is generally treated as a nuisance component and is often very simplistic. While the main interest lies in the ETAS part of the model, the poor model description of the background part imposes considerable bias on the ETAS part of the model. For example, a poorly specified spatial density of the background events causes many of the background events to be seen as ETAS events. It can also cause the estimated Omori power-law decay p to be too small, and hence the aftershock sequences appear to continue for too long. On the other hand, the boundary of the observation region can impose a reverse bias which causes aftershocks that are close but within the boundary to be seen as background events. In almost all of the large NZ event sequences since 1965, the model consistently under-fits these sequences. Consequently, it over-fits those space-time regions where there is ‘normal’ seismicity with no major events present. This may indicate that the space-time region of a major event sequence is much closer to criticality, in that aftershock events appear to be much more easily initiated. The standard ETAS model does not reflect this observation.

Citation
Harte, D. S. (2012). Bias in fitting the ETAS model: a case study based on New Zealand seismicity. Geophysical Journal International, 192(1), 390-412. doi: 10.1093/gji/ggs026.