Natural Hazards 7: 99-108, 1993.

99
© 1993 Kluwer Academic Publishers'. Printed in the Netherlands.

On Some Problems About Testing Stochastic

Models of the Earthquake Time Series
G E R A S S I M O S A. P A P A D O P O U L O S
Section of Geology and Geophysics, Department of Civil Engineering, Hellenic Air Force Academy,
Dekelia, Attika, Greece, and
Department of Seismoteetonics, Earthquake Planning and Protection Organization, 226 Messogion
Ave., 15561 Holargos, Athens, Greece

(Received: 27 June 1991; in final form: 28 July 1992)

Abstract. Decision on whether or not a specific stochastic model is an adequate description of the
earthquake time series depends on (a) the characteristics of the earthquake sample (total time inter')al,
magnitude range, inclusion or not of dependent events) and (b) the selection of counting interval, r,
which seems to be of crucial importance. Certain improvements are proposed for handling earthquake
samples when tested against stochastic models. Particularly, it is proposed that r should be nonarbi-
trarily selected as equal to the mean return period of events of magnitude equal to or larger than the
lower magnitude, Mmin, for which the data are complete. A method0,1ogical experiment has been
performed using three data sets of South Aegean intermediate depth main shocks of variable Mmin. It
is clear that the distribution gradually approaches the simple Poisson model as the counting interval
decreasingly approaches a nonarbitrarily selected value, which indicates the important misleading effect
of using arbitrarily selected counting intervals. The efficiency of the proposed improvements in stochas-
tic modelling of volcanic eruption and tsunami time series is discussed.

Key words. Counting interval, Poisson model, point processes, earthquake statistics, South Aegean,
intermediate depth earthquakes.

1. Introduction
T h e o c c u r r e n c e o f e a r t h q u a k e s can be m o d e l l e d as a p o i n t p r o c e s s t h a t is a d i s c r e t e
stochastic p r o c e s s d e s c r i b i n g the p o s i t i o n s o f t h e e v e n t s in t i m e . T h e stochastic
m o d e l l i n g o f t h e e a r t h q u a k e o c c u r r e n c e p r o c e s s has p r o v e d v e r y useful in e a r t h -
q u a k e p r e d i c t i o n studies, in u n d e r s t a n d i n g the n a t u r e o f the e a r t h q u a k e p h e n o m -
e n a , a n d in assessing s e i s m i c i t y a n d seismic h a z a r d . T h u s , decisions o n the v a l i d i t y
o f specific s t o c h a s t i c m o d e l s a r e o f crucial i m p o r t a n c e in d e s c r i b i n g seismic activity.
P r e v i o u s studies s h o w t h a t t h e r e j e c t i o n o r a d o p t i o n o f a specific stochastic
model depends upon the selection of the earthquake sample and handling of the
d a t a . R e l a t e d p r o b l e m s a r e h e r e e x a m i n e d with t h e m a i n p u r p o s e of i m p r o v i n g
t h e e v a l u a t i o n p r o c e d u r e for testing s t o c h a s t i c m o d e l s a g a i n s t a c t u a l e a r t h q u a k e
d a t a . E m p h a s i s is given o n t h e n o n a r b i t r a r y s e l e c t i o n of t h e c o u n t i n g i n t e r v a l , a
p r o b l e m e x a m i n e d in a p r e l i m i n a r y f o r m b y P a p a d o p o u l o s (1989a). Tests with
[00 GERASSIMOS A. PAPADOPOULOS
samples of subcrustal shocks occurring in the South Aegean have been performed
as a methodological experiment. Finally, the efficiency of the proposed criteria in
stochastic modelling of other geophysical events, such as volcanic e(uptions and
tsunamis, is discussed.
r

2. Some Theoretical Considerations

Let us start with a few theoretical considerations concerning some basic aspects
of statistical seismology useful for the remaining part of the paper.
It is known that the magnitude-frequency relation of Gutenberg and Richter
(1944) can be expressed as

log N k = ak - b M , (1)

where Nk is the cumulative frequency of earthquakes with magnitude equal to or

larger than M, k is the length of the time interval covered by the data, and a~
and b are parameters. For k = 1 year, the relation (1) becomes

log N = a - b M , (2)

where

a = ak - log k. (3)

From Equation (2) we get

T x = 10 bM-a, (4)

where T x is the mean return period of earthquakes with magnitude, X, equal to

or larger than a certain value M.
On the basis of the assumptions that the number of earthquakes in a year is a
Poisson random variable with mean /x, and X is distributed with the cumulative
distribution function

F ( x ) = P r ( X < ~ x) = 1 - exp(-/3x), x ~> 0. (5)

Epstein and Lomnitz (1966) have shown that the probability R r ( M ) , of occurrence
of one or more earthquakes of magnitude x ~> M in a T-year period, can be
expressed as
Rr(X) = 1 - e x p [ - / x T exp(-/3X)], (6)

where
/3 = b In 10. (7)

From (6) and (7), we easily get

Rr(X) = 1 - exp[-lO~Texp(-bXln 10)], (8)

where
TESTING STOCHASTIC MODELS 101

a = In/x/ln 10. (9)

The prototype of all point processes is the simple Poisson process. This is the
process which has commonly been considered in seismic literature. The notion
that seismic events occur according to the Poisson process is further discussed
in the next section. The Poisson process is defined by the basic conditions of
independence, stationarity and orderliness (for detailed theoretical explanations,
see Lomnitz (1974)). The probability density function of z, the number of events
per unit of time, is the Poisson distribution

f(z) = exp(-/x)~Z/z!, (10)

where/x, the mean value, is the parameter of the distribution expressing the rate
of seismic activity. The probability of observing z events in time, T, is
P(z) = exp(-/xT)( txT)Z/z! (11)

3. Problems in Testing Stochastic Models

Many authors have tested their earthquake catalogs of several regions of the world
against the Poisson process (see short reviews in Shlien and Toks6z (1970) and
Rao and Kaila (1986)). Results of these tests indicate that, in some cases, the
Poisson model does not explain the time series of earthquakes, whereas in some
others, this model is an adequate description of the earthquake process in time.
In addition, some authors showed that the process with the negative binomial
entries as a model for describing the occurrence of shallow earthquakes in Europe
and the Alpine-Himalayan belt is usually better than the Poisson process (Schen-
kovfi, 1973, 1982; Rao and Kaila, 1986).
A review of the existing literature revealed that the main factors affecting the
decision on whether or not a specific stochastic model is an adequate description
of the earthquake time series in a given region are: (a) the total time interval
considered, (b) the earthquake magnitude range chosen, (c) the inclusion or not
of dependent events, i.e., foreshocks and mainly aftershocks, and (d) the counting
interval, that is the time unit selected for the test performance. These problems
are examined below.
(a) In studies of the seismic process, we utilize one or more samples of obser-
vation to obtain estimations which can be considered as representative of the
whole seismic process. It is generally assumed that the larger the period of obser-
vation used, the smaller the error of estimation. To secure that the selected length
of the observation is adequate to substantiate conclusive results, I suggest that this
length should be taken to be no shorter than the time interval which incorporates a
high probability of observing at least one event with magnitude X i> M where
. . . .

Mma x is the largest magnitude involved in the sample. Formula (8) provides an
estimate of such a probability. However, derivation of this formula is based on
the assuption that the process is Poissonian which means a-priori adoption of the
102 GERASSIMOS A. PAPADOPOULOS
Poisson model. An equivalent, independent solution is to consider a time interval
no shorter than Tx, as determined by (4), with X = Mmax.
(b) The selection of the earthquake magnitude range depends on the complete-
ness of the data and the purposes of the study. Obviously, the lower magnitude
threshold in the sample cannot be smaller than the lower magnitude for which the
data are complete. According to Shlien and Toks6z (1970) and Schenkovfi (1973)
there is a tendency for better fits to the simple Poisson process when the low
magnitude events are removed. To interpret this tendency, Schenkovfi (1973)
suggested that when the magnitude threshold is raised, the mean time interval
between events becomes longer, which decreases the likelihood of a possible
interaction between successive earthquakes. However, Papadopoulos and Voido-
matis (1987) concluded that the time series of strong main earthquake occurrences
in the back-arc Aegean region consists of a random and a nonrandom component
depending on the magnitude range: random for M, = 6.5-6.8, nonrandom for
Ms ~> 6.9. Time clustering of the largest events is in agreement with a model
according to which seismic energy release within 5-year time windows approxi-
mates a harmonic curve with a period of about 50 years. A model of regional
stationary accumulation of thermal stresses along certain seismic belts and their
cyclic relaxation may explain this periodicity.
Any change in the lower magnitude threshold has two important consequences.
The first is its influence on the selection of the counting interval, a problem
examined below. The second is the influence on the percentage error involved in
the sample. The theory of counting processes indicates that the percentage error
decreases with increasing mean value. In the Poisson counting process the standard
deviation is equal to /x, the mean value. Thus, the larger the mean value, the
smaller the error (e.g., Singh and Sanford, 1972). Therefore, any change in the
lower earthquake magnitude threshold in the sample causes increase or decrease
of ix, i.e., decrease or increase of the fractional error of the sample, respectively.
(c) The role of dependent events, particularly of aftershocks, in the stochastic
modelling of the earthquake process has been examined by many investigators.
Ben-Menahem (1960) and Schenkovfi and Kfirnfk (1971) concluded that the data
are consistent with the Poisson process after ruling out some events as clear
aftershocks. Knopoff (1964) suggested that deviations from the Poisson distribu-
tion is due to the presence of aftershocks, while Aki (1956) concluded that the
deviation cannot be entirely attributed to the presence of aftershocks but may be
due to the existence of a positive type of clustering of shocks. Shlien and Toks6z
(1970) believe that the occurrence of aftershocks can be modelled by generalizing
the Poisson process to allow for the occurrence of more than one event in a time
unit. Ultimately, the problem of including or not including dependent events in
tests of earthquake samples against probabilistic models can be viewed on the
basis of the purposes of the study.
(d) The most important problem is the nonarbitrary selection of counting in-
terval, ~-. The effect of the counting interval in testing the hypothesis that seismic
TESTING STOCHASTIC MODELS 103
time series follow this or that stochastic model has beel~ mentioned by some
authors (Shlien and Toks6z, 1970, p. 1968). A good example is given by Kfirnik
et al. (1983) who examined the time distribution of the aftershocks of the iV/, =
6.5, 20 July 1978 event in northern Greece, and concluded that the negative
binomial distribution is rejected for r = 1 day. On the contrary, this is not rejected,
at the same significance level, for r taken as equal to only one hour.
The problem is very clear: how could a nonarbitrary length of r be selected?
What is proposed here is to select a counting interval as long as the mean return
period, Tx, of events having magnitude equal to or larger than the lower magni-
tude threshold, Mmin, in the sample. This means that the counting interval is taken
as long as the time interval during which one event of X > Mmin, is expected as
an average. Formula (4) provides a statistical estimate of Tx. In particular, for
great or even large earthquakes occurring in subduction zones, probabilistic, sta-
tistical, and geotectonic estimates of the repeat times have been proposed (e.g.,
Rikitake, 1976; Molnar, 1979; Sykes and Quittmeyer, 1981; Singh et al., 1983).
According to Knopoff and Kagan (1977), statistical methods which utilize all
available data of earthquake catalogs give superior estimates of the parameters of
seismicity, including return times, than do extreme value methods.

4. Methodological Tests in South Aegean Data Sets

Samples of intermediate depth shocks (h i> 70 kin) associated with the South Ae-
gean Wadati-Benioff zone have been used to put into an actual test the suggestions
discussed earlier. Dealing with subcrustal shocks has the advantage that the
number of dependent events is very small compared to the number of dependent
events associated with shallow main shocks. This means that we can evaluate only
the main events by easily removing the dependent ones, thus providing favourable
conditions for examining how our decision on the validity of the Poisson or other
stochastic model is influenced by the factors analyzed earlier emphasizing the
effect of the counting interval selection.
Three data sets of intermediate depth earthquakes occurring in the South Ae-
gean area have been examined. They cover the time intervals of 1964-1985, 1950-
1985 and 1911-1985 and have been derived from a common data source, the
earthquake list of Comninakis and Papazachos (1986). The magnitude-frequency
relationship indicates that the three data sets may be complete for Ms >/4.5,
4.8, and 5.2, respectively. Table I supplies information on several parameters
characterizing the data sets; the time length, T, maximum and minimum magni-
tudes, Mma× and Mmin, number of main shocks, n, and dependent events, n',
probability of RT(X) determined from (8) for X ~> M . . . . as well as the mean
return period, Tx, determined from (4) for X = Mmax and X = Mmin. All magni-
tudes are expressed in surface-wave magnitude. Although only main shocks were
considered, Table I provides the numbers of both main and dependent events for
comparison. Dependent events include foreshocks and aftershocks.
104 GERASSIMOS A. PAPADOPOULOS
Table I. Parameters of the three data sets. Key: T - length of the time interval; Mm~x= X1 and
Mmin= X2 are the maximum and minimum magnitudes in the sample, respectively; n = number of
main shocks; n' = number of dependent events; Rr(M) = probability determined from formula (8);
Txi (i - 1, 2) = mean return period determined from formula (4).
Time interval T(yr) M .... Mmin gt H' Rr(M = Mma×) Tx1 (yr) Tx2 (yr)
1964-1985 22 6.8 4.5 92 i0 0.85 11.6 0.32
1950-1985 36 6.8 4.8 62 4 0.96 11.6 0.51
1911-i985 75 8.0 5.2 64 3 0.63 75.9 0.95

Parameters ak, a, and b have been determined for South Aegean subcrustal
shocks by Papadopoulos (1989b) in the least-squares sense: a~ = 5.83, a = 3.56,
and b = 0.68. These values were introduced in the relations (4) and (8).
Table I shows that the probability, R r ( X ) , of observing an earthquake of
X > 6.8 in time intervals of T = 22 yr and T = 36 yr is 0.85 and 0.96, respectively,
while we get R r ( X ) = 0.63 for X = 8.0, T = 75 yr. On the other hand, the mean
return period of earthquakes with X >/6.8 is equal to 11.6 yr, which is much less
than 22 yr and 36 yr, the time intervals of the first two data sets. For the third set,
we have Tso = 75.9 yr, which is almost equal to the corresponding T = 75 yr.
Consequently, the lengths of the time intervals, especially the first two, are satisfac-
torily selected with respect to the objective and unified criteria proposed in the
present paper.
Increase of the lower magnitude threshold from Mmin = 4.5 to 4.8 and then to
5.2 increases the mean return period TMmin from 0.32 to 0.51 and then to 0.95 yr,
which means that, according to the criteria imposed here, the interval r should
be selected as T 1 = 3.84, r2 -- 6.12, and 7' 3 = 11.4 months for the 1964-1985, 1950-
1985, and 1911-1985 data sets, respectively. For convenience, the values 7"1 = 4 ,
r2 = 6, and 7"3 = 12 months have been considered.
Tests of the data against the Poisson process have been performed for intervals,
r, as explained earlier as well as for several counting intervals greater than 7". By
this way we can control the influence of changing r on the frequency distribution
of the number of intervals with z earthquakes (z = 0, 1, 2 . . . . ). Figure 1 shows
this distribution for T = 1964-1985, Ms 1> 4.5 and r is equal to 2 yr (open circles),
l y r (triangles), and 0.5 yr (solid circles). Distribution for the same earthquake
sample and nonarbitrarily selected counting interval (7" = 4 months) is presented
in Figure 2.
As one may observe, the distribution gradually approaches the Poisson model
as the interval r decreasingly approaches the mean return period, T4.5. The X2-test
indicates that the hypothesis about the compatability of the data with the Poisson
process is not rejected at the 5 and 1% significance levels when the interval r is
taken equal to T4.s, i.e., the mean return period of earthquakes with magnitude
equal to or larger than the lower magnitude involved in the sample. P a r a m e t e r / x
of the Poisson distribution is/xt = 1.39 events/4 months.
Similar results were obtained for the data sets of 1950-1985 and 1911-1985.
TESTING STOCHASTIC MODELS 105

12
@

4
A A
A g A A
o ~ @ o o o

2 4 6 8 10
z >
Fig. t. Frequency distribution of the number of intervals with z earthquakes for the time interval of
1964-1985, Ms >! 4.5, and unit of time equal to 2 yr (open circles), I yr (triangles), and 0.5 yr (solid
circles).

25

20

8
15

10

f (z)
0 I ! | I I

0 1 2 3 4
z >
Fig. 2. Frequency distribution of the number of intervals with z earthquakes for the time interval of
1964-1985, Ms/> 4.5, and unit of time equal to the objective one (r = 4 months). Observed (open
circles) frequencies fit the theoretical ones (solid circles) determined for the pure Poisson model.

T h e data fit the Poisson distribution when the value of r is r2 = T4.s and r3 =
Ts.2, respectively. P a r a m e t e r of the Poisson distribution is /x2 = 0.86 events/6
m o n t h s and/*3 = 0.85 events/12 months.

5. Conclusions and Discussion

Quantitative estimations o f the seismic activity are usually based on statistical and
probabilistic models. As a c o n s e q u e n c e , the decision that a specific probabilistic
m o d e l is an a d e q u a t e description of the e a r t h q u a k e process at a particular region
106 GERASSIMOS A. PAPADOPOULOS
should be based on certain evaluation procedures. A review of the existing litera-
ture shows that decisions of this sort are influenced by some parameters of the
earthquake sample: (a) the total time interval, T, (b) the earthquake magnitude
range, (c) the inclusion or not of foreshocks and aftershocks, and (d) the counting
interval, r. The most important seems to be the last one. Tests with actual
earthquake data have shown that the validity of a stochastic model directly depends
on the selected length of r. A common practice, however, is the arbitrary selection
of r. To select a nonarbitrary length of r, it is suggested that r should be taken
equal to the mean return period of events having magnitude equal to or larger
than the lower magnitude bound in the sample. Nonarbitrary handling of the
sample parameters (a), (b), and (c) are also suggested.
To examine, in a practical way, the effect of changing r , an experiment was
performed with three samples of South Aegean intermediate depth main shocks
which were tested against the pure Poisson process for several values of r. It was
found that the distribution gradually approaches the Poisson model as r decreas-
ingly approaches the nonarbitrary value r = TMm~,. For TMmin , the data fitted the
Poisson distribution at 5 and 1% significance levels.
Probabilistic models are useful for describing time series of other geophysical
events, such as volcanic eruptions and tsunamis, not only for understanding related
physical processes but also for assessing associated hazards. It seems that the
suggested handling of sample parmeters is applicable to tsunami time series be-
cause it is possible to establish from tsunami data relations similar to the relations
(1), (4), and (8). Moreover, tsunami data eliminate the problem of dependent
events examined earlier.
Dealing with stochastic modelling of volcanic eruption time series presumes that
the points in the process, i.e., the eruption onset times, are considered regardless
of the eruption total duration. Difficulties appear in applying the proposed hand-
ling because it seems very hard to establish reliable relations of the form of (1)
and, consequently, of (4) and (8). Another important problem is the selection
of the data sample by discriminating between significant volcanic eruptions and
nonsignificant, minor volcanic events, a problem examined by Papadopoulos
(1987a, b).
Tests with actual tsunami and volcanic eruption data would be of crucial impor-
tance in examining the efficiency of the suggested nonarbitrary handling of sample
parameters. Very recently, Papadopoulos (1990) examined the randomness of the
time distribution of the post-1500 Santorini (Thera) volcano eruptions. Lacking a
relation of the form of (1), and, consequently, of (4), the nonarbitrary selection
of the counting interval has been roughly based on the inter-arrival time of the
eruption onset. It has been shown that Thera is a simple Poissonian volcano,
which is consistent with similar results obtained for other volcanoes of the world
(see the short review in Scandone, 1983).
TESTING STOCHASTIC MODELS 107

Acknowledgements
I am grateful to K. Aki (California), L. Knopoff (California), and Z. Schenkovfi
(Prague) for sending reprints and to two anonymous reviewers for their construc-
tive comments.

References
Aki, K.: 1956, Some problems in statistical seismology, Zisin 8, 205-228.
Ben-Menahem, A.: 1960, Some consequences of earthquake statistics for the years 1819-1958, Beitr.
Geophys. 69, 68-72.
Comninakis, P. E. and Papazachos, B. C." 1986, A catalogue of earthquakes in Greece and the
surrounding area for the period of 1901-1985, Geophys. Lab. Univ. of Thessaloniki PuN. No. 1.
Epstein, B. and Lomnitz, C.: 1966, A model for the occurrence of large earthquakes, Nature 211,
954-956.
Gutenberg, B. and Richter, C. F.: 1944, Frequency of earthquakes in California, Bull. Seism, Soc.
Am. 34, 185-i88.
Kfirnfk, V., Schenkovfi, Zd., and Schenk, VI.: 1983, Time pattern of the Thessaloniki sequence of
1978, in B. C. Papazachos and P. G. Carydis (eds.), The Thessaloniki, Northern Greece, Earthquake
of June 20, 1978 and its Seismic Sequence, Technical Chamber of Greece, Thessaloniki, pp. 151-
158.
Knopoff, L.: 1964, The statistics of earthquakes in Southern California, Bull. Seism. Soc. Am. 54,
1871-1873.
Knopoff, L. and Kagan, Y.: 1977, Analysis of the theory of extremes as applied to earthquake
problems, J. Geophys. Res. 82, 5647-5657.
Lomnitz, C.: 1974, Global Tectonics and Earthquake Risk, Elsevier, Amsterdam.
Molnar, P.: 1979, Earthquake recurrence intervals and plate tectonics, Bull. Seism. Soc. Am. 69, 115-
133.
Page, R.: 1968, Aftershocks and microaftershocks of the great Alaska earthquake of 1964, Bull. Seism.
Soc. Am. 58, 1131-1168.
Papadopoulos, G. A.: 1987a, Sequences and clusterings of significant volcanic eruptions in convergent
plate margins during 1900-1980, Tectonophysics 138, 211-222.
Papadopoulos, G. A.: 1987b, Large deep-focus shocks and significant volcanic eruptions in convergent
plate boundaries during 1900-1980, Tectonophysics 138, 223-233.
Papadopoulos, G. A.: 1989a. Objective selection of time units in testing stochastic models of the
earthquake occurrence process, Europ. Geophys. Soc. XIV General Assembly, Barcelona, 13-i7
March, Ann. Geophys., Special Issue, 277 (abstr.).
Papadopoulos, G. A.: 1989b, Forecasting large intermediate depth earthquakes in the South Aegean,
Phys. Earth Planet. Inter. 57, 192-198.
Papadopoulos, G. A.: 1990, Deterministic and stochastic models of the seismic and volcanic events in
the Santorini volcano, in D. A. Hardy, J. Keller, V. P. Galanopoulos, N. C. Flemming, & T. H.
Druitt (eds.), Thera and the Aegean WorM, 3rd Internat. Congr. Proc., Santorini, September 1989,
The Thera Foundation, London, Vol. 2, pp. 151-159.
Papadopoulos, G. A. and Voidomatis, Ph.: 1987, Evidence for periodic seismicity in the inner Aegean
seismic zone, Pure Appl. Geophys. 125, 613-628.
Rao, N. M. and Kaila, K. L.: 1986, Application of the negative binomial to earthquake occurrences
in the Alpide-Himalayan belt, Geophys. J. Roy. Astr. Soc. 85, 283-290.
Rikitake, T.: 1976, Recurrence of great earthquakes at subduction zones, Tectonophyics 35, 335-362.
Scandone, R.: 1983, Problems related with the evaluation of volcanic risk, in H. Tazieff and J. C.
Sabroux (eds.), Forecasting Volcanic Events', Elsevier, Amsterdam, pp. 57-67.
Schenkov& Z.: 1973, Model of earthquake occurrence in the European area, Pure Appl. Geophys.
109, 1700-1711.
Schenkovfi, Z.: 1982, Statistical models of earthquake occurrences, Acta Univ. Carolin.-Math. Phys.
23, 81-84.
108 GERASSIMOS A. PAPADOPOULOS
Schenkov~, Z. & K~rnfk, V.: 1971, Statistical and other characteristics of earthquakes in the Aegean
region, Travaux Inst. Gdophys. Acad. Tchdcosl. Sci. 19, I49-166.
Shlien, S. and Toks6z, N.: 1970, A clustering model for earthquake occurrences, Bull. Seism. Soc.
A m . 60, 1765-1787.
Singh, S. and Sanford, A.: 1972, Statistical analysis of microearthquakes near Socorro, New Mexico,
Bull. Seisrn. Soc. A m . 62, 917-926.
Singh, S. K., Rodriguez, M., and Esteva, L.: 1983, Statistics of small earthquakes and frequency of
occurrence of large earthquakes along the Mexican subduction zone, Bull. Seism. Soc. A m . 73,
1779-1796.
Sykes, L. R. and Quittmeyer, R. C.: 1981, Repeat times of great earthquakes along simple plate
boundaries, in D. W. Simpson and P. G. Richards (eds.), Earthquake Prediction - A n international
Review, Am. Geophys. Union, Washington, D.C., pp. 217-247.