Regression analysis of reported earthquake precursors. I. Presentation of data

Around 700 reported precursors of about 350 earthquakes, including the negative observations, have been compiled in 11 categories with 31 subdivisions. The data base is subjected to an initial sorting and screening by imposing three restrictions on the ranges of main shock magnitude (M≥4.0), precursory time (t≤20 years), and the epicentral distance of observation points (X _m≤4.10^0.3M ). Of the 31 subcategories of precursory phenomena, 18 with 9 data points or more are independently studied by regressing their precursory times against magnitude. The preliminary results tend to classify the precursors into three groups:

1. The precursors which show weak or no correlation between time and the magnitude of the eventual main shock. Examples of this group are foreshocks and precursory tilt.
2. The precursors which show clear scaling with magnitude. These include seismic velocity ratio (V _p/V_s), travel time delay, duration of seismic quiescence, and, to some degree, the variation ofb-value, and anomalous seismicity.
3. The precursors which display clustering of precursory times around a mean value, which differs for different precursors from a few hours to a few years. Examples include the conductivity rate, geoelectric current and potential, strain, water well level, geochemical anomalies, change of focal mechanism, and the enhancement of seismicity reported only for larger earthquakes. Some of the precursors in this category, such as leveling changes and the occurrence of microseismicity, show bimodal patterns of precursory times and may partially be coseismic.

In addition, each category with a sufficient number of reported estimates of distance and signal amplitude is subjected to multiple linear regression. The usefulness of these regressions at this stage appears to be limited to specifying which of the parameters shows a more significant correlation. Standard deviations of residuals of precursory time against magnitude are generally reduced when observation distance enters as a second independent variable. The effect is more pronounced for water well level and conductivity rate changes. While a substantial portion of the data seem to suffer from personal bias, hence should be regarded as noise, the observations of a number of strain sensitive phenomena such as strain, water well level, and conductivity rate changes, appear to be internally more consistent. For instance, their precursory times suggest a scaling relationship with the strain energy surface density associated with the main shock. The scaling is not identical for all three phenomena so that they may constitute the imminent, short- and intermediate-term manifestation of the same process, i.e. strain loading, respectively.