Detection of change points in underlying earthquake rates, with application to global mega-earthquakes. Issue 2 (8th December 2015)

• Record Type: Journal Article
• Authors: Touati, Sarah; Naylor, Mark; Main, Ian
• Is Part Of: Geophysical journal international. Volume 204:Issue 2(2016:Feb.)
• Page Start: 753
• Page End: 767
• Publication Date: 2015-12-08
• Subjects: Probabilistic forecasting -- Earthquake interaction, forecasting and prediction -- Seismicity and tectonics -- Statistical seismology; Geophysics -- Periodicals; 550
• DOI: 10.1093/gji/ggv398
• Languages: English
• ISSNs: 0956-540X

Abstract

Abstract: The recent spate of mega-earthquakes since 2004 has led to speculation of an underlying change in the global 'background' rate of large events. At a regional scale, detecting changes in background rate is also an important practical problem for operational forecasting and risk calculation, for example due to volcanic processes, seismicity induced by fluid injection or withdrawal, or due to redistribution of Coulomb stress after natural large events. Here we examine the general problem of detecting changes in background rate in earthquake catalogues with and without correlated events, for the first time using the Bayes factor as a discriminant for models of varying complexity. First we use synthetic Poisson (purely random) and Epidemic-Type Aftershock Sequence (ETAS) models (which also allow for earthquake triggering) to test the effectiveness of many standard methods of addressing this question. These fall into two classes: those that evaluate the relative likelihood of different models, for example using Information Criteria or the Bayes Factor; and those that evaluate the probability of the observations (including extreme events or clusters of events) under a single null hypothesis, for example by applying the Kolmogorov–Smirnov and 'runs' tests, and a variety of Z -score tests. The results demonstrate that the effectiveness among these tests varies widely. Information Criteria worked at least as well as the more computationally expensive Bayes factor method, and the Kolmogorov–Smirnov and runs tests proved to be the relatively ineffective in reliably detecting a change point. We then apply the methods tested to events at different thresholds above magnitude M ≥ 7 in the global earthquake catalogue since 1918, after first declustering the catalogue. This is most effectively done by removing likely correlated events using a much lower magnitude threshold ( M ≥ 5), where triggering is much more obvious. We find no strong evidence that the background rate of large events worldwide has increased in recent years.