Alan L. Kafka

Does the Spatial Distribution of Smaller Earthquakes Delineate Areas Where Larger Earthquakes Are Likely to Occur

In the development of the latest generation of National Seismic Hazard Maps, one of the component models used the spatial distribution of smaller earthquakes in the eastern United States to forecast the locations of larger earthquakes (Frankel, 1995). Variations of this hypothesis, that smaller earthquakes indicate where larger earthquakes are likely to occur, are found throughout earthquake studies. In a previous study (Kafka and Walcott, 1998), we tested this hypothesis for earthquakes in the northeastern United States (NEUS) to see how well the spatial distribution of smaller earthquakes recorded by seismic networks in the NEUS “forecasts” the locations of larger earthquakes that have already occurred. The essence of our procedure is to systematically analyze how often previously occurring smaller earthquakes occurred in the vicinity of larger earthquakes. The purpose of this study is to extend that investigation to other areas of the world to obtain a more global perspective on this issue. Here we report on an extension of that investigation to the southeastern United States, the New Madrid seismic zone, southern California, northern California, Israel, Turkey, and the entire eastern United States. Our results to date do, in fact, suggest that (in a variety of tectonic environments) the spatial distribution of smaller earthquakes delineates areas in which larger earthquakes are likely to occur. In a number of cases where locations of larger earthquakes were not forecast based on this approach, we suspect that the misses are, at least in part, due to incompleteness and quirks in the earthquake catalogs.

A Non-Poissonian Element in the Seismicity of the Northeastern United States

Earthquakes of M ≥ 2.7 (1975-2000) in the accreted-terranes region of the northeastern United States are more temporally clustered than expected from a random process. This clustering is evident even when aftershocks have been removed from the earthquake catalog. The distances between clustered event pairs are uniformly distributed between 20 km and over 400 km. It is not clear why this clustering is occurring. Curiously, statistically significant temporal clustering was not found for earthquakes from nearby Quebec on the North American craton. Manuscript received 24 July 2001.

Exaggerated claims about earthquake predictions

The perennial promise of successful earthquake prediction captures the imagination of a public hungry for certainty in an uncertain world. Yet, given the lack of any reliable method of predicting earthquakes [e.g., Geller, 1997; Kagan and Jackson, 1996; Evans, 1997], seismologists regularly have to explain news stories of a supposedly successful earthquake prediction when it is far from clear just how successful that prediction actually was. When journalists and public relations offices report the latest ‘great discovery’ regarding the prediction of earthquakes, seismologists are left with the much less glamorous task of explaining to the public the gap between the claimed success and the sober reality that there is no scientifically proven method of predicting earthquakes. A striking example of this situation occurred when NASA posted a feature article on its Web site in 2004 in which an earthquake prediction project it funded was heralded as an “amazing success” (see http://www.nasa.gov/vision/earth/environment/0930_earthquake.html). Because this kind of hyperbole is a constant source of frustration for scientists at Weston Observatory (Boston College, Weston, Mass.), where seismologists try to accurately report the state of the art of research on earthquake prediction to the public, we decided to test just how amazing this particular success was.

Major geological features and lateral variation of crustal structure in southern New England

Abstract Rg dispersion studies in southern New England (SNE) provide a basis for comparing lateral variation in crustal structure with geological features. Rg waves reveal lateral variations at depths ranging from near the surface down to a few kilometers. The similarity between Rg results and those of other geophysical studies suggests that variations revealed by Rg studies penetrate into the middle of the crust, if not deeper. The Waterbury Dispersion Region (WDR), named after a local geological structure, is characterized by the highest Rg group velocities in SNE (∼2.9–3.1 km/s at a period of 1 s). These higher velocities may be due to rocks enriched in oceanic type crust and/or granulite facies metamorphic rocks. The Hartford and New Haven dispersion regions (HDR and NHDR) comprise the northern and southern parts of the Mesozoic Hartford Rift Basin, respectively. The HDR is characterized by the lowest group velocities in SNE (∼2.1–2.4 km/s at a period of 1 s). These lower velocities appear to be related to relatively thick Pleistocene glacial sediments overlying a thick sequence of Mesozoic sedimentary rocks. Higher velocities in the NHDR are attributed to a thinner cover of glacial sediments and a thinner package of sedimentary rocks. These younger rocks overlie metamorphic basement which is, on average, closer to the surface in the NHDR than in the HDR. The Bronson-Avalon Dispersion Region (BADR), east of the Hartford Rift Basin, represents several terranes that are geologically diverse but, at the scale of this study, not differentiated by Rg dispersion. The seismic structure beneath the BADR appears to be quite homogeneous; group velocities in this region range from about 2.6 to 2.8 km/s at a period of 1 s. The BADR is interpreted to consist of stacked layers of gently dipping to nearly flat-lying thrust sheets. The Torrington Dispersion Region (TDR) lies north of the WDR and west of the HDR and has group velocities similar to those of the BADR.

Proximity to past earthquakes as a least-astonishing hypothesis for forecasting locations of future earthquakes

The cellular seismology (CS) method of Kafka (2002, 2007) is presented as a least-astonishing null hypothesis that serves as a useful standard of comparison for other, more complex, spatial forecast methods (i.e., methods that forecast the loca- tions, but not the times, of earthquakes). Spatial forecast methods based on analyses of earthquakes in California, such as that of Ebel et al. (2007) and the pattern informatics (PI) method of Rundle et al. (2002, 2007) provide opportunities for comparing meth- ods that incorporate information about rates of seismicity with a method (i.e., CS) that only assumes that future earthquakes will occur near epicenters of past earthquakes. The Ebel et al. (2007) five-year-forecast method (E07) maps the spatial distribution of rates of seismicity, and the PI method not only considers rates of seismicity but also incorporates temporal changes in local rates of seismicity as a measure of the potential for future earthquakes to occur at some location. Our comparison of success rates of the E07 method and the PI method with CS for earthquakes in California has yet to reveal any compelling evidence that inclusion of seismicity rates or temporal changes in local seismicity rates in a spatial forecast model improves the ability to forecast locations of earthquakes.

Short‐term non‐poissonian temporal clustering of magnitude 4+ earthquakes in california and western nevada

The M4+ mainshocks throughout California and western Nevada from 1932 to 2004 show non‐Poissonian temporal clustering over time periods of a few days. The short‐term clustering is independent of the distance between earthquake epicenters. It implies that some of the M4+ mainshocks are mutually triggered by some unknown regional cause. In southern California, more short‐term clustering is found for M4+ earthquakes east of the San Andreas Fault. In central California, most M4+ mainshocks at Long Valley, CA have occurred within 10 days of M4+ mainshocks around the San Francisco Bay area. The clustering implies predictable behavior in the occurrences of M4+ mainshocks. We propose a hidden Markov model (HMM) as an earthquake forecast method for the region. Our HMM assumes a hidden sequence of interevent time states associated with observations of earthquake occurrences (times, locations, and magnitudes) with transition probabilities between states determined with the Baum‐Welch algorithm and the past earthquak...