You are on page 1of 24

Geophys. J. R . astr. SOC.

(1986) 86,815-838

Evidence for precursory changes in the frequency-


magnitude b-value

Warwick D . Smith Geophysics Division, Department of Scientific and Industrial

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Research, Box 1320, Wellington, New Zealand

Accepted 1986 February 21. Received 1986February 20; in original form 1985July 19

Summary. A clear case has been established for a significant increase in the
slope of the frequency-magnitude relation for periods of up to six years
before a number of earthquakes in New Zealand and California. In a region
of about 200 km diameter in central New Zealand the b-value increased
from near unity to more than 2.0 at the end of 1968, and remained
abnormally high until the beginning of 1975. An earthquake of magnitude
6.0 followed in 1977 January, within 50 km of the location which had
registered the highest b-value. The anomaly was detected in contour maps
of b-values windowed in space and time, then resolved more precisely in
time using the CUSUM technique. It was statistically significant at a level
higher than 99.99 per cent. Compared with the average rate of moment
release in this region, there was a deficit of 1.3 x 10'*Nm during the
anomalous period, which is equivalent to an earthquake of magnitude 6.08.
This phenomenon of a high b-value has been found before all shallow New
Zealand earthquakes of magnitude 6.0 or more since 1965, wherever the
seismicity rate and detection threshold were adequate to resolve it. Further,
the only statistically significant occurrences of high b-value were those
which were followed by large earthquakes, two of magnitude 6.8 and 6.0 and
a swarm of equivalent magnitude 5.3. The phenomenon has also been found
before the San Fernando and Coalinga earthquakes in California, both of
which were thrust events, but it is not apparent before two large Californian
earthquakes with strike-slip mechanisms. Observations of high b-value are
consistent with reported instances of quiescence in background seismicity.
In fact, what has been interpreted as quiescence may probably be more
accurately described as a localized high b-value. The duration of the anomaly
is not a simple function of magnitude. Different strain rates in the regions
examined may account for this. The available data suggest that careful
monitoring of the b-value may provide a tool for medium-term earthquake
prediction, although perhaps not in all tectonic environments.

Key words: b-value, frequency-magnitude, precursor, prediction, quiescence


816 W. D.Smith
1 Introduction
Ever since Gutenberg & Richter (1944) expressed the magnitude distribution in the form
logN = a bM, the evaluation of the parameter b has been the most frequently performed
~

statistical calculation in seismology. It is an important parameter for the estimation of


earthquake hazard, and must be determined with known confidence limits for extrapolation
to larger earthquakes for which the recurrence intervals are longer than the period of
observation .
There have been many reports of temporal changes in the b-value: that it is significantly
different for foreshocks and aftershocks (e.g. Suyehiro 1966; Utsu 1970, 1971; Wyss & Lee
1973); that it changes throughout an aftershock sequence (e.g. Gibowicz 1973); and that it

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


can vary over large areas and long time intervals, not apparently related to any particular
large event (e.g. Gibowicz 1974; Robinson 1979). The occurrence of high b-values in the
immediate vicinity of, and for a period of some years preceding, large earthquakes has been
reported by Fiedler (1974) in Venezuela, Li et al. (1978) in China, and Kiek (1984) in Papua
New Guinea.
A preliminary note (Smith 1981) reported a high b-value before a large earthquake in
New Zealand. In a relatively small area of about 3 x lo4 km2 a high b-value was observed for
a period of five years before an earthquake of magnitude 6 occurred, and is assumed to have
been related to that earthquake. That note paid inadequate attention to the statistical
aspects of the problem. The present paper, in enlarging upon it, reports occurrences of high
b-values before a number of large earthquakes, shows how these occurrences may be resolved
in time, develops a statistical test which shows that some were highly significant, and
measures the spatial extent of the anomaly in each case. It also notes that the duration of
the anomaly and the spatial extent may be related t o the magnitude of the following earth-
quake, and provides some constraints on the time of occurrence, given the times of onset
and ending of the anomaly.
In a systematic study of the New Zealand catalogue, high b-values have been found before
all shallow earthquakes (depth less than 50 km) of magnitude 6.0 or greater since 1965 for
which there are adequate data. The exceptions are several for which low data rates offer a
plausible explanation for the non-observance of the anomaly. Further, the only significant
occurrences of high b-values are those which preceded large earthquakes.
The same phenomenon has been found before two thrust earthquakes in California, but
it is not apparent before two strike-slip events. These observations add the interesting
suggestion that the b-value precursor may be limited to earthquakes with dip-slip
mechanisms. All the New Zealand examples were in this class.

2 Recognition
During the early stages of a major seismic hazard evaluation in New Zealand, it was
discovered that certain small areas of the country had experienced periods of 6-values much
higher than unity. Fig. 1 lllustrates the development of one such occurrence in central New
Zealand. Each map shows contours of b-values determined at a grid of points 0.25" apart in
latitude and longitude, each using all earthquakes of depth less than 50 km in the previous
four years within 1" (1 11 km) of the grid point. Successive maps are one year apart in time,
from 1970 January to 1975 January.
Aki ( I 965) gave the maximum likelihood estimate of b as
Changes in the f ~ e q u e n c y ~ i n a g i i i t l l db-value
e 817

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Figure I. Contour maps of b-values for shallow e;uthquakes from 1970 (a) to 1975 (0. For parameters
see text. The epicentre of the Cape Campbell earthquake of 1977 January 18 is shown in (0.

where 1; is the mean excess magnitude above the detection threshold. Utsu (1966) tabulated
corrections applicable when magnitude classes are discrete. Weichert (1980) gave the relevant
formula. This correction has not been used here because the classes were small (0.1).
Reducing the threshold by 0.05 in the calculation of b was adequate. Page's (1968) formula,
allowing for a finite maximum magnitude, has not been used either because it was not
judged to be important for seismicity samples taken from at most a few decades (but see
Sections 4 and 8). In the following the term 'magnitude' will be taken to mean excess
magnitude above the threshold, unless indicated otherwise. Zhang & Song (198 1) pointed
out that equation (1) gives a biased estimate of 6 . An unbiased estimator from a sample of
i z events is
818 W. D. Smith
which is the formula used for b-values in Fig. 1. Equation (2) arises because although the
m
expectation of is log L o e/b, that of l / m is nb/(n - I) log,, e.
The magnitude threshold is 4.0. All magnitudes used are &chter’s local magnitude M L ,
ensured by the continued operation of Wood-Anderson seismographs at a few locations.
Haines (I98 I ) has derived distance-attenuation functions appropriate to New Zealand for
this purpose. These functions have been in use since 1977, and earlier magnitudes which
were determined using a modified form of the California Wood-Anderson relation had been
revised back t o 1970 at the time of this study. No evidence of this procedural change at
1970 was found in the b-values calculated, so it has been assumed not to have affected the
results.

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


The b-value in Fig. 1 reached a peak exceeding 2.5 in 1975, which had intensified over
five years and migrated southeastward to a point very close to where the Cape Campbell
earthquake (magnitude 6.0) occurred on 1977 January 18. The peak dissipated from 1975
onwards. Note that considerable smoothing has resulted from the windowing in time (four
years) and space (1 11 km radius). Data rates drop off t o both the NW and SE, but each
b-value in the middle of the map is determined from about 60 events. None represents less
than 20 events.
While the manuscript for this paper was being prepared it was learned that the same
contour map technique for recognizing high b-values has been used in China (Li e t al. 1978).
These observations are totally consistent with Mogi’s (1979) ‘seismic gap of the second
kind’, i.e. a region in which the rate of occurrence for small and moderate earthquakes is
zero or much less than in the surrounding area. Habermann, McCann & Nishenko (1983)
prefer the term ‘quiescence’, which I shall use although I believe that it may misrepresent the
physical characteristics of the phenomenon. Consider a small region where the b-value is
higher than in the surrounding area and where the detection threshold is 3.0, for example.
Imagine further that the rate of occurrence of magnitude 3.0 events is approximately
uniform throughout the whole area. Earthquakes of magnitude 4 w ill thus occur less often in
the central region than in the surroundings, and magnitude 5 possibly not at all, within an
observation period of a few years. But if the detection threshold were not as low, say 4 or 5 ,
the central region would appear to be quiescent. Many studies of quiescence have used the
NOAA catalogue, e.g. Kelleher & Savino (1975), Ohtake, Matumoto & Latham (1977),
Khattri & Wyss (1978), Wyss & Habermann (1979), Habermann (1981), Kanamori (1981),
Habermann & Wyss (1984a). I contend that its threshold, 4.5 or greater for most of these
studies, is so high that a high b-value would appear as a period of quiescence.
Mogi (1979) does note that at a lower threshold quiescent areas do contain activity, as do
Sadovsky et al. (1972) who report ‘variation in seismic regime of a definite region with the
appearance of a seismic “calm” (decrease of seismic activity) replacing earthquakes of
middle classes of energy but conservation of activity of microearthquakes’. This latter
observation is essentially that of increased b-values. Mogi also comments that seismic activity
has sometimes been observed to increase again, filling the quiescent area, prior t o the large
earthquake. Ohtake et al. (1977) and McNally (1982) quote a number of examples. This is
exactly what was observed in the data presented in this study. In each case the b-value peak
dissipated with a return to near normal seismicity before the main event occurred.
As a systematic study of b-value changes throughout New Zealand, maps as in Fig. 1 were
prepared for the whole country, at six-month intervals from 1960 to 1984. The time
window (four years) and sampling radius (1”) were chosen t o be appropriate t o the
seismicity rate, large enough t o provide sufficient data above the detection threshold for
each b-value calculation and small enough t o give some temporal and spatial resolution. A
magnitude threshold of 4.0 was used throughout. This was too low in the early 1960s but
Changes in the frequency-magnihtde b-value 819
Table 1. Shallow New Zealand earthquakes of magnitude 6.0 and greater since
1965, with the Wanganui swarm of 1982. The magnitude given for the swarm
is theequivalent value determined by adding the moments of the largest events.

Earthquake Date Epicentre Magnitude

Gisborne 1966 March 5 38.755 178.30E 6.1


Seddon 1966 April 23 41.745 174.61E 6.0

Inangahua 1968 May 23 41.775 172.015 6.8

Opunake 1974 November 5 39.545 173.46E 6.0

nilford Sound 1976 May 4 44.675 167.38E 6.5

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Cape Campbell 1977 January 18 41.845 174.58E 6.0
Puysegur Bank 1979 October 12 46.705 166.03E 6.5
Wanqanui 1982 October 13 40.175 175.00E (5.3)

nacaulay River 1984 June 24 43.505 170.50E 6.0

appropriate thereafter. A number of occurrences of local high b-values was detected, to


which the significance test developed in the next section was applied. The only significant
occurrences were those before the Inangahua, Cape Campbell and Wanganui earthquakes
(Table 1).
For the purposes of associating high 6-values with large following events, the following
empirical criteria were adopted. The epicentre must lie within the circle defined by the
6-value peak and the confidence limit of the assessed radius of the anomaly. The interval
between the end of the anomaly and the earthquake must be n o greater than half the
duration of the anomaly.
All the shallow earthquakes of magnitude 6 or greater since 1965 are listed in Table 1,
together with a swarm in 1982. Epicentres are shown in Fig. 2. For five of these a b-value
anomaly has been detected.
(i) Gisborne, 1966. The detection threshold was too high in this part of the country to
allow recognition of a local high b-value.
(ii) Seddon, 1966. The detection threshold near Wellington was much lower than else-
where in the country during the early 1960s. Hence, a local high b-value was recognizable
here (see Fig. 3) but not at Gisborne (i). However, the probability that this could have
occurred by chance alone (Table 3) is too high for it to rate as a significant occurrence.
(iii) Inangahua, 1968. The largest earthquake examined. Clear b-value anomaly (Fig. 4).
(iv) Opunake, 1974. Not recognized. The seismicity rate is fairly low in this part of New
Zealand, at the western extremity of the main seismic region. Also, the very strong anomaly
for event (viii) was developing, and may have masked other effects.
(v) Milford Sound, 1976. Although this was a large event its source volume was very
small.Evison(1977a)identifiesa quiescent area of radius about 3 0 k m . With such a small area
affected and a threshold certainly no lower than 4.0, possibly higher, insufficient events
were available to resolve any b-value change.
(vi) Cape Campbell, 1977. The clearest example found, shown in Fig. 1 and used as a
case-study in Sections 3-5. This earthquake occurred at a focal depth of 40 km, on the
surface of the dipping plate and deeper than would normally be classified as a shallow
earthquake in New Zealand. But the anomaly shows up in the routine computation of
b-values using all events of depth less than 50 km.
820 W.D. Smith

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024

Figure 2. Epicentres of eight large earthquakes and one swarm listed in Table 1.

(vii) Puysegur Bank, 1979. This event was offshore t o the south where the detection
threshold was too high for a local b-value change t o be recognized.
(viii) Wanganui Swarm, 1982. This area was identified during epicentre location
procedures throughout the late 1970s as an area where many small events were occurring
(Garrick & Gibowicz 1983). The peak in the b-value contour map was very pronounced
(Fig. 5 ) . It migrated south-eastwards to the final location from 1974 to 1980. The 1982
swarm contained six events over magnitude 4.0.
Changes in the frequency-magnitude b-value 821

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Figure 3. b-value contour map and the Seddon earthquake epicentre, 1966 May. Map date 1965 January,
radius l o , time window 4 yr, magnitude threshold 4.0.

Table 2. Detection thresholds used for analysis of the precursors to the earthquakes listed.
Years 40-44 45-49 50-54 55-59 60-64 65-69 70-74 1 5 - 7 9 80-84

1. Seddon 4.0 4.0 3.8 3.9 3.8 3.8 - - -


2. Inangahua 4.3 4.2 4.2 4.2 4.0 4.0 - - -
3 . Cape Campbell 4.5 4.4 4.1 4.1 4.0 3.9 3.9 3.9 -
4. wanganui 4.5 4.3 4.0 4.0 3.8 3.8 3.7 3.7 3.7

5 . Macaulay River - 4.3 4.3 4.3 4.3 4.0 4.0 4.0 3.9

6 . San Fernando 3.0 3.0 3.0 3.0 3.0 3.0 3.0 - -


7 . Coalinga 3.1 3.1 3.1 3.0 3.0 3.0 3.0 3.0 3.0

Figure4. 6-value contour map and the Inangahua earthquake epicentre, 1968 May. Map date 1967
January, radius 1S o ,time window 4 yr, magnitude threshold 4.0.
822 W. D. Smith

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Figure 5. b-value contour map and a representative epicentre for the Wanganui swarm, 1982 October.
Map date 1979 January,radius l o , time window 4 yr, magnitude threshold 4.0.

(ix) Macaulay River, 1984. Although the seismicity rate is low in this region an anomaly
is recognizable. It is not significant, however (Table 3).
In the light of the compatibility between local high b-values and areas of quiescence, it is
significant that Evison (1977a, b) identified quiescent periods prior t o earthquakes (i), (ii),
(iii) and (v).
Figs 6 and 7 show the b-value contour maps prior to the San Fernando and Coalinga
earthquakes. These maps were derived directly from the catalogues from the California
Institute of Technology and the University of California, Berkeley, respectively. The peak in
Fig. 6 almost reaches the contour level of 2.0 and that in Fig. 7 is just less than 1.5.

Figure 6. b-value contour map and the San Fernando earthquake epicentre, 1971 February. Map date
1969 January, radius lo, time window 4 yr, magnitude threshold 3.0.
Changes in the frequency-magnitude b-value 823

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Figure 7. b-value contour map and the Coalinga earthquake epicentre, 1983 May. Map date 1981 January,
radius lo, time window 4 yr, magnitude threshold 3.0.

3 Resolution
The b-value contour map provides a means of surveillance, whereby the occurrence of a high
b-value may be detected, but the time and space windowing smooths out the anomaly. The
important point in terms of earthquake prediction is that the anomaly in Fig. l(f) reached
its peak two years before the large earthquake, then began its return to normal. Good
resolution in time is obviously essential. The question of resolution in space will be
addressed in Section 5.
Note that an examination of the b-value is essentially an examination of the mean magni-
tude (equations 1 and 2). A low mean magnitude is equivalent to a high b-value.
Statistically what is involved is a change-point process. The hypothesis is that the para-
meters of the distribution change at some unknown point, and it is desired to determine
both that point and the statistical significance of the change. One approach is to use the
CUSUM (Page 1954). For a sample of N earthquakes, magnitudes mi,form

where m is the mean of the entire sample. Thus C,, is the cumulative sum (CUSUM) of the
differences between the individual values and the mean. Clearly CN = 0. The CUSUM is like
an integral in that it is the gradient of C,, which indicates whether the mi are predominantly
greater than or less than the mean m, for i near n . More importantly, a change in gradient
indicates a change in mean magnitude at that point. The CUSUM also has a seismological
analogue in reduced travel-time curves, in that extracting the mean allows detail to be
recognized. Actually the reducing term need not be exactly the mean, but only a constant
value and it is convenient that i t approximate the mean. Fig. 8(a) illustrates the magnitude
CUSUM for earthquakes near 41.5"S, 174"E. This is the location of the b-value peak in
Fig. l(0. Data are plotted event by event, so the time-scale is variable. The CUSUM gradient
becomes negative at about 1968-69 and the trend reverses in 1975, indicating that the mean
magnitude between those dates was less than that of the entire sample. These change points
824 W. D.Smith
rl 0 0
In co YEAR
I?
2I 2I 3
I I l l II I I I I I I I I I I I I I I I I I I /

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


L
7 7 - 7 -
0 50 100 150 200
NO. OF EVENTS
Figure 8. (a) CUSUM of magnitudes for shallow earthquakes of magnitude 4.0 and greater within 1"
radius of 41.5"S, 174°F. Data are plotted event by event. The variable time-scale is shown at the top of
the figure, and the event count a t the bottom. (b) Bartlett's statisticzk for two subsets of the data in (a),
calculnted as described in the text. (c) CUSUM of magnitudes from a synthetic earthquake population,
b = 1.0.

are here resolved much more clearly than a sequence of windowed b-values or mean
magnitudes would allow, A CUSUM of magnitudes from a random distribution with b = 1.O
is shown in Fig. 8(c). It also has a section with negative gradient. The significance of this is
examined in Section 4. In Fig. 8(a) there is another section with negative gradient in the
early 1960s. This is more pronounced at another location, closer to the epicentre of the
Seddon earthquake (Fig. 3).
The two change points in Fig. 8(a) can be identified by eye, but a more objective
procedure is desirable. Worsley (1986) has applied a likelihood ratio test t o the problem of
finding a change-point in a sequence of variables drawn from an exponential distribution. He
calculates Bartlett's statistic z k in the form
Zk = 2 * [HI log (YM/i&) -t n2 log (%/.12 11, (4)
where for a change-point at k in a sample of n events

The statistic z k is calculated for all possible change-points k . It takes a maximum value at
the best estimate of the change-point. Worsley also considers the significance of an identified
change-point. This will be addressed in Section 4.
Changes in the frequency-magnitude b-value 825
For lack of a simultaneous test for two change-points, the data set in Fig. S(a) was
divided by eye into two subsets, the first from the beginning until what appeared t o be the
second change-point, and the second from the first apparent change-point until the end, Fig.
8(b) shows the z k values with solid and open symbols for the two subsets, respectively. The
width of the peaks in Fig. 8(b) suggests that the changes in mean magnitude to a lower value
then back again may not be abrupt, as in Worsley's model. Because it is necessary to identify
specific beginning and end-points, the sample was selected as all events after 1968/743 (1968
November 1 , magnitude 5.5), up to but not including 1975/204 (1975 April 13, magnitude
5.1). Event numbers are those quoted in the New Zealand Seismological Report, an annual
publication of the New Zealand Seismological Observatory. These events mark clear changes

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


in the CUSUM (Fig. 8a), and are in the middle of the broad maxima in Fig. 8(b). It is also
physically plausible that a period of high b-value begins and ends with events of significant
magnitude, such as these. This choice was later confirmed when the spatial extent of the
anomaly had been estimated (Section 5). Parameter z k for the random data in Fig. 8 ( c ) has
very small peaks, much less significant than in Fig. 8(b).
the sample between the two identified change points. Distributions estimated using equation
(2) are also shown. Earthquakes of magnitude 3.9, which were apparently completely
detected in 1969-75, have been added. The threshold generally claimed for the catalogue is
4.0 (since the middle 1960s), within the land area of New Zealand, but a slightly lower
threshold has been achieved close to Wellington, the central station of the network.
Subsequent analysis added further events by using a lower threshold where appropriate
(Table 2 ) but removed others if it was determined that the radius of the anomaly was not 1"
but only lOOkm (Section 5 and Table 3). An exponential distribution is clearly a good fit to
each of the two samples.

4 Discrimination
Knopoff, Kagan & Knopoff (1982) claim, with some justification, that inadequate attention
has usually been paid to the statistical significance of apparent departures from normal
behaviour in the earthquake population. Aki (1965), in proposing a significance test for the
b-value of a sample of earthquakes, invoked the Central Limit Theorem under which, given a
large enough sample, the distribution of the mean magnitude can be assumed to be Normal.
Shi & Bolt (1982) obtained the exact distribution of the b-value, but it will be convenient
here t o use the work of Utsu (1966), who showed that for a population of b-value bo and a
sample of n events with b-value b , the variable 2nbo/b is distributed as x2 with 2n degrees of
freedom. The estimate of b in equation (2) corresponds to the mode of the x2 distribution.
Utsu (1966) also pointed out that in comparing two samples, sizes n and n2 and b-values
b1 and b 2 , the variable b z / b , is distributed as the F-distribution with ( 2 n l , 2 n 2 )degrees of
freedom. Because he determined b-values by the maximum likelihood formula (equation l),
the variable b2/bl is equal to GI/&. So the F-distribution provides a means of testing the
mean of a randomly selected sample against that of another. Values of F for various prob-
ability levels may be found from tables, or the probability may be calculated as the sum of a
fmite series (Kendall & Stuart 1977).
But this is not sufficient to establish the significance of a low mean magnitude for a
sample selected as in Section 3. The sample was not selected randomly, but as the result of a
careful selection procedure (all events between the two change-points) and by virtue of the
fact that its mean was markedly lower than normal. Fig. 8(c) shows the CUSUM of a
synthetic population of earthquakes with b = 1.0. The sample between events 170 and 203,
not including either, clearly has a low mean magnitude. Its b-value is 1.65 compared with
826 W. D.Smith
0.95 for the first 170 events. Such occurrences are to be expected because of the asymmetry
of the exponential distribution of magnitudes: there are more small earthquakes than large
ones so negative slopes in the CUSUM are likely t o be longer than positive ones. The
F-distribution test for this sample gives the probability that the low mean magnitude could
have occurred by chance alone as 0.0032. This is apparently very significant, and from a
random distribution! One approach to the problem is to evaluate a conditional probability
as the ratio of the two probabilities:

1
the low mean of a the sample was
given sample could selected because
Probability
have occurred by its mean was

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


chance alone markedly low

[the low mean could have occurred by


Probability
chance alone, for a random sample]
-
- - __
[a randomly chosen sample can satisfy
objective criteria for a ‘low mean’]

The numerator of this expression is given by the F-distribution test developed by Utsu and
described above. The chief difficulty in evaluating the denominator is that the sample size is
not determined a priori, but by use of the CUSUM.
Write the probability determined from the F-test with degrees of freedom ( p , 4 ) as

Gp,*(X) = F,,,(x) d x . (7)


Jx-

If criteria for a low mean magnitude are defined as


n , z 20 events
m,c; 0.6 m,
then equation ( 6 ) can be expressed as

The denominator is the probability that a random sample of 20 events has mean magnitude
less than 0.6 ml.
This is an ad hoc procedure, but it is nevertheless derived from realistic
selection criteria and does apparently discriminate between random samples and genuine
precursory effects. In the case of Fig. 8(c), which was found to be a representative CUSUM
for random data, the use of equation (9) increases the probability figure from 0.0032 to
0.13, so the high b-value is not significant at the 5 per cent level. For the data samples
treated the denominator in equation (9) was about 0.02. The criteria in equation (8) were
used as a basis for selecting occurrences of high b-values that were worthy of study.
For the earthquake data of Fig. 9, the 141 events 1951/2 tc 1968/743 have mean excess
magnitude 0.375 (threshold 4.0) and the 91 events 1968/753 to 1975/60 have mean 0.182
(threshold 3.9). These correspond to b-values of 1.OO and 1.84. The probability that the
latter could have occurred by chance alone is 0.0003, using equation (9). As an alternative
procedure, the mean magnitude can be determined as a weighted mean, with weights
assigned to individual magnitudes according t o the epicentral distance r of the earthquake
from the central point. Using weights 1 - r / r m a Xwhere
, rmaxis the radius of the circle within
Changes in the frequency-magnitude b-value 827
100

0 NOV.2,1968 to APR.12,1975

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


X

7
4.0 5.0 6.0
Magnitude
Figure 9. Frequency distribution of magnitudes for earthquakes within 1" (111 km) of 41.5"S, 174"E
from 1951 to 1968/743 (crosses) and from 1968/753 to 1975/60 (circles) with the maximum likelihood
lines shown.

which events are used, - 1" in this case, reduces the mean magnitude for events 1968/753 to
1975/60 from 0.182 to 0.147. However, this does not decrease the probability much
because the number of degrees of freedom in the F-test is reduced (see Appendix). But use
of the weighted mean does assist in the identification of high b-values, by sharpening up
small anomalies limited in space. Figs 1 and 3-7 were prepared in this way. Other weighting
functions could be used, but the linear weight is commensurate with the model developed
later (Fig. 1Oc).
Worsley's (1986) work looks attractive for assessing the significance of such a change of
b-value in time. He develops a two-sided test, t o detect either an increase or a decrease in
mean magnitude, and tabulates significance levels of parameter z k as a function of the
number of observations and the probability level desired, and for these same earthquake data
notes that the peak value of z k in Fig. 8(b) (11.49) corresponds to a probability of
0.0269. For a one-sided test this value should be approximately halved. It is still rather
larger than that obtained by the ad hoc procedure above, so apparently represents a more
conservative test. But it does require the a priori selection of the second change point, and
this is a disadvantage.
A third approach would be to determine the probability as the numerator of equation
(9), multiplied by (nl + n2 - 40), as this factor is the total number of ways in which the first
change point can be chosen (assuming that n l and nz must both be 20 or more). This is
rigorous and too conservative, but even so raises the 0.0003 probability figure t o only 0.001.
It may well be that other appropriate tests can be developed, but the formulation in
equation (9) has been retained for this paper.
Periods of high b-value, apparently random, do tend to occur at the edges of seismic
regions. Seismicity rates are low, so uncertainties in b-values are large. The effects of low
maximum magnitudes in these regions may also be significant. Without Page's (1968)
correction, equations (1) and (2) will be systematically in error if the maximum magnitude is
small.
828 W.D. Smirk
4.21
a.
1

4.21
4.1
b.

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


4.01

3.91

5.54 c.
. ... .
$ 4
.n
. . . . ... ." .. . ." . .
-....
4.5L; ... . . .. .. :
.
t .

. .. .. .... .. ._...
2.

.>-. . .---:.~ .,
I . . : "

. I-.L..y- .-...--
4.0
. . .-. 4.....
. . -. ---- " +* ......
/*.*.
- . '1-..... .,.:. .*..r.:..
.I.
.-I.
..It .b
, ,
I " " l " " I
0 50 100 150 200
RADIUS (krn)
Figure 10. Earthquakes near 41.5"S, 174"E from 1968 to 1975, showing: (a) menn magnitudes, with
standard errors in the mean, for events withln circles of radii from 10 to 200 k m ; (b) mean magnitudes,
with standard errors in the mean, for events within annuli 20 k m in width, plotted at the mean radius;
(c) individual magnitudes and the two lines of th2 mean magnitude model.

5 Mensuration
The spatial extent of the anomaly is not clear from the contour maps in Fig. 1, because of
the smoothing effect of the circular window. The migration over about 150km in five years
(Fig. la+) also makes it difficult to determine the size of the area affected, but a b-value
peak is identified in Fig. l(f) at 41.5"S, 174'E. A time window appropriate to thislocation
can be identified from Fig. 8. Using the events within this time window it is possible to
estimate the areal extent of the anomaly, then relate this size to the magnitude of the
following earthquake.
Fig. lO(a) shows the mean magnitudes for earthquakes within radii of 10-200 km, and
standard errors in the mean. It is not clear from this figure where the limits of the anomaly
are, but in Fig. 10(b), which shows the mean magnitudes of earthquakes w i t h annuli 20
km across, there is a clear change at about 80-100 km mean radius. Standard errors are large
because of the few events in each distance range. The transition point could no doubt be
better determined if the data rate were higher and/or the threshold lower, but it turns out
that precision in this estimation is not essential.
Fig. 1O(c) further illustrates the increase in mean magnitude with radius, out to about
100 krn. The individual magnitudes are plotted as a function of distance from the centre of
the anomaly. The lack of medium and large events at short distances is apparent. The
Changes in the frequency-magnitude b-value 829
variation of mean magnitude with radius is modelled by two line segments. They are fitted
simultaneously by least squares, solving for four parameters: the coordinates of the point of
intersection and the two gradients. This is a non-linear procedure but it converges iteratively
with some damping. It is appropriate here because minimizing the sum of squares of the
residuals in magnitude estimates the mean magnitude. The two lines intersect at a radius of
100 i 32 km. This distance is the least well constrained of the four parameters but it may be
taken to represent the radius from 41.5"S, 174'E which includes the entirety of the
anomaly. Clearly a different choice of central point could have resulted in a different radius
obtained, but recall that the central point chosen was the b-value peak in Fig. l(f).
Repeating the CUSUM plot of Fig. 8(a), using only those events within 100 km of the

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


central point (cf. 111 km in Fig. 8a) results in sharper transitions in gradient at the two
change points previously identified.
An estimate of the size of the forthcoming earthquake can be obtained by noting that the
total moment during the anomalous period is much less than the mean rate would predict.
In fact, because a high b-value implies a lack of moderate events, the total moment during
the anomalous period is negligible, and the moment deficit can be computed by summing
the contribution of all events within the affected area, from the beginning of the catalogue
up to the onset of the anomaly, and scaling this total t o the duration of the anomaly. A
convenient moment-magnitude conversion is used, here that of Hyndman & Weichert
(1983), viz.
logM, = 9.0 + 1S M ,
where M o is in SI units N m .
Attempts t o refine the calculation by subtracting the moment during the anomalous
period were unsuccessful, not only because this is small but also because such a high b-value
as was observed (near 2.0) implies that the moment-magnitude relationship no longer holds.
Equation (10) can only hold if b < 1.5, for otherwise the integrated moment does not
converge. So the moment deficit was computed simply by scaling the observed moment rate.
Hyndman & Weichert's (1983) method for obtaining the total moment rate by integration
was not used either, because it is extremely difficult to make a reliable estimate of the
maximum magnitude for a region only 100 km in radius, and it involves assumptions which
are beyond the scope of this paper. But that is certainly a possible alternative.
In Fig. 11 the moment deficit is evaluated for a succession of radii, and plotted on a
quadratic scale. The relationship is approximately linear, as expected for a location near the
edge of a broad seismic region. For a linear seismic region such as a transform fault, or for
other geometries, this would take a different form. The moment deficit for a radius of 100
km is 1.3 x 10"Nm, which (10) equates with an earthquake of magnitude 6.08. The parti-
cular choice of radius in Fig. 10 is not critical because no large change in magnitude results if
the radius is varied within reasonable limits.
In order to make the best use of the available earthquake data, with their varying levels
of completeness, a study of the detection thresholds was made for the five New Zealand
earthquakes with observed b-value precursors. Weichert ( 1980) has addressed the problem of
a change in detection levels, but a different formulation is appropriate here. Note that
equation (9) uses not the actual magnitudes but the excess above the threshold. It is thus
equally true if the mean excess magnitude for n events is expressed as
n
m= l / n 1 (mi - rnmin), ( 1 1)
i= 1
where the threshold m,,, is that value of the threshold appropriate to the time of
830 W.D.Smith
r

h
7 /* I
5
m 16.5
E 5
k
g 4-
W
3-
!-

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


2-
0 t
= 1- 16.0

I
0 50 100 150 200
RADIUS (km)
Figure 11. Moment deficit, computed as described in the text for a succession of radii from 41.5"S,
174'E, and a mean rate established using all events from 1940 t o the beginning of the anomaly. Radius is
on a quadratic scale. Equivalent magnitude is shown at right.

occurrence of event i. Table 2 shows the thresholds determined for five-year periods in the
vicinity of the five New Zealand earthquakes studied and the two in California. These were
evaluated by examining the frequency distribution and cumulative distribution of earth-
quakes by magnitude, and also the b-values computed at successive thresholds, in each five-
year period. In the case of the Macaulay River event there were barely any events in the
period 1940-44, so this period was not used. By adopting the thresholds shown (reduced by
0.05) for the computation of mean excess magnitude, maximum use was made of available
data.
Analysis details for the seven earthquakes are given in Table 3 : the location of the b-value
peak, the determined onset and end dates of the anomaly, b-values and numbers of events in
the background population and between the two dates given, the probability that the
anomaly could have occurred by chance alone, the assessed radius and the magnitude inferred
from the moment deficit. In some cases the probability figure given was found by increasing
the radius from that listed, and using a linear weight to compute the mean magnitude, as was
described in Section 4. It is the lowest value of the probability found from a range of radii.
This procedure is justified from considerations of Fig. 10, where the mean magnitude is seen
to increase continuously to a normal value from a minimum near the analysis centre. The
mean magnitude as used in equation (1 l), for instance, ignores spatial dependence.
Increasing the radius includes more events, and this may increase the mean magnitude, but it
can also decrease the uncertainty in the value obtained. High b-values before earthquakes 2,
3 and 4 (Table 3 ) were the only significant occurrences found in the New Zealand catalogue.
For the Macaulay River earthquake the preliminary epicentre was used as the analysis
centre, because the low data rate did not allow a b-value contour map to be prepared. The
preliminary epicentre is close to the final one, available after the study was completed (Table
1). The inferred magnitude of the Inangahua earthquake was determined using two historical
windows: the first back to 1940 and the second (bracketed) including the 1929 Buller earth-
quake of magnitude 7.7. It is obviously important to use a long period to establish the mean
moment rate when dealing with a large earthquake. The difference between the two values
Changes in the frequency-magnitude b-value 831
Table 3. Precursory data for five New Zealand earthquakes and two in California,
showing date, magnitude, location of b-value peak, onsct and end dates of anomaly.
number of events and h-value for background population and anomalous period,
probability that it could have occurred by chance alone, assessed radiu5 (km) and
magnitude inferred from moment deficit.

b-value Change N1 bl Inferred


Earthquake Date Nag. Peak points N2 b2 Prob. Rad. Mag.
1. Seddon 660423 6.0 41.755 590522 166 1.00 (0.14) 100 5.97
174.75E 650312 52 1.48
2. Inangahua 680523 6.8 41.258 620722 2 36 0.94 0.032 180 6.33
172.00E 670711 29 1.83 (7.1)

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


3. Cape Campbell 770118 6.0 41.505 681101 183 0.99 0.00003 100 6.08
174.00E 750414 80 1.82
4. Wanganui 821013 5.3) 40.258 740523 122 1.22 0.008 50 5.40
175.00E 801116 60 1.86
5. Macaulay R. 840624 6.0 43.508 771026 76 0.98 (0.47) 120 5.94
170.50E 830515 25 1.50
6. San Fernando 710209 6.4 35.501 640531 840 0.94 <0.00001 110 6.08
118.50W 690228 94 1.79
7. Coalinga 830502 6.7 35.90N 760114 350 0.94 (0.075) 90 5.47
121.008 820619 50 1.32

for the inferred magnitude illustrates the problem of determining a reliable estimate of the
mean rate. The Wanganui swarm contained six earthquakes over magnitude 4.0: 4.9,4.8,
4.8,4.7,4.6,4.6. The total moment of these is equivalent to a single event of magnitude 5.3.
The analysis centre was taken t o be a representative epicentre for the swarm earthquakes.
This was close t o the 6-value peak in the four-year contour map. The radius was difficult to
assess because it was small, and therefore strongly dependent on the chosen location of the
analysis centre. The Seddon and Macaulay River precursors are not statistically significant in
terms of equation (9) above. They are included here because these excursions to high 6-value
were followed by large events. In neither case are there enough data during the period of
high b-value to provide a significant result.
Note the correlation in the New Zealand data between the radius of the anomaly and the
magnitude. This is to be expected, at least for earthquakes in similar tectonic regimes. But
the 1976 Milford Sound earthquake (magnitude 6.5) was very rich in high frequencies and
apparently had a very small source volume. The precursory quiescent area identified by
Evison (1977a) was only 3 0 k m in radius, and the detection threshold of 4.0 does not
provide enough data within this small region to resolve the anomaly adequately.
Magnitudes in the Caltech catalogue have been computed to 0.1 precision since 1944, and
in the Berkeley catalogue since 1942. No data prior to these two dates have been used. For
the San Fernando data, event counts, b-values and probability figures in Table 3 use a back-
ground data set which excludes the 1952 Kern County earthquake of magnitude 7.7 and its
aftershocks, as this sequence would otherwise have dominated the calculation (and lowered
the probability still further). The 1952 event is included for the determination of the
moment rate, however. The Coalinga data do not quite reach the 5 per cent significance
level, and the moment deficit calculation gives a very low estimate of the magnitude.
Moments for the two California earthquakes were also determined using equation (lo),
because it is identical with that found by Thatcher & Hanks (1973) for southern California
28
832 W. D.Smith
and Bakun (1984) for central California. Use of the Johnson & McEvilly (1974) formula, viz.

logMo= 10.60+ 1.16M, (12)


increases these estimated magnitudes to 6.35 and 5.69. The variety of earthquake
mechanisms within each of the two regions renders less appropriate the summation of the
scalar moment for an estimate of the magnitude in these cases. Some other scalar quantity
may be better. The calculated total moment rate at San Fernando depends quite strongly on
the Kern County earthquake. The b-value anomaly before the San Fernando earthquake is
consistent with the findings of Ishida & Kanamori (1977): quiescence near the epicentre
from 1965 t o 1968, followed by a resumption of activity. Fig. 6 indicates, however, that the

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


b-value anomaly was centred 120 km to the north of the epicentre.
Both of these earthquakes were thrust events (Whitcomb 19‘71; Eaton, Cockerham &
Lester 1983). It is therefore interesting to note that with a sampling radius of 1.0” and a
four-year time window, no anomaly is detectable before the 1968 Borrego Mountain earth-
quake or the 1979 Imperial Valley shock, both of which had strike-slip mechanisms (Allen &
Nordquist 1972; Johnson, Rojahn & Sharp 1982). Habermann & Wyss (198413) found no
quiescence before the Imperial Valley earthquake either. The b-valuelquiescence
phenomenon may be limited to earthquakes with dip-slip mechanisms, or to thrust regimes.
A systematic examination of the California catalogues is under way.

6 Prediction?
The necessary elements of an earthquake prediction are estimates of the location, magnitude
and time of occurrence of the ensuing event, and the uncertainties in these (Allen 1976). It
appears that the observed phenomenon of high b-values can provide all these elements, for a
viable medium-term prediction.
In all the cases studied the epicentre of the eventual earthquake did not generally
coincide with the mapped b-value peak, but it did lie within the region defined by the peak
and the confidence limit of the assessed radius. Three reasons are apparent for the non-
coincidence: (1) migration of the peak as in Fig l(a-f) clearly introduces systematic error in
locating the peak at any particular epoch, since data must be windowed in time; (2) if the
seismicity rate varies in space, such as near the edge of a seismic region, there is a systematic
bias of the b-value peak in the direction of the lesser seismicity; ( 3 ) if the observed occur-
rences of high b-value can be related toMogi’s( 1969) ‘doughnut’pattern, then his observation
that earthquake epicentres often lie at the edge of the quiescent region should have an
analogue with the region of high b-value. These three reasons together imply that there is no
reason t o suspect coincidence between the epicentre and the apparent location of the b-value
peak.
The moment deficit apparently gives an estimate of the magnitude. There is an
uncertainty in the determination of the mean moment rate throughout historical time. For
the Cape Campbell earthquake the catalogue contained two other previous events of
magnitude 6 within the anomalous area, and the estimate of the magnitude was accurate.
For the larger Inangahua earthquake, going back further in the catalogue provided what was
probably an upper bound to the mean rate. Thus the uncertainty in the magnitude depends
on a good knowledge of the previous seismicity, and over a long period. The larger the
magnitude, the longer the averaging period necessary. For the Coalinga data, however, the
magnitude estimate is not as accurate. This may indicate that too short an interval was used
to establish the mean moment rate, that the moment deficit procedure must be calibrated
for each particular region, that the appropriateness of the moment deficit calculation
Changes in the frequency- magnitude b-value 833
-high b-value- El:t
6. 4.
3.

T210j 2. *7

1
L ,
I I I I I I

20 40 60 80 100

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Tl (months)
Figure 12. Onset and end times of b-value anomalies, plotted as time before the main event in each case.
Numbers identify events in Tables 2 and 3 . For the Macaulay River earthquake (no. 5 ) the end of the
anomaly is not well defined.

depends on the mechanism, or that the agreement with the actual magnitudes in the other
cases is fortuitous. Further examples should resolve this.
The timing of the ensuing earthquake is not well constrained by the available data. The
critical observation is clearly that of the end of the anomalous period, but this does not usually
occur long enough before the main event to be useful. Fig. 12 shows the interval between
the observed end of the anomaly and the main event, plotted against the time from the
beginning of the anomaly to the main event. The data are presented in this way because
abscissa and ordinate are independent. There is no simple relationship between duration of
anomaly and magnitude. Why should there be? Should not the strain rate in a particular area
also affect the rate at which the precursory sequence progresses? For these observations, T2
lies between ten months and two years. Fitting a model in which T2 is proportional to T I
gives T2 = 0.19 T,. In its practical form this becomes T2= 0.24 (TI- T 2 )with a standard
error in the residuals of five months.
The procedure developed above could be summarized as follows:
(i) Recognition. Identify occurrences of high b-values in contour maps prepared at regular
intervals. Map parameters (time window, radius, threshold) must be chosen so that there are
sufficient data at each grid point to compute a b-value (20 events is an absolute minimum).
(ii)ResoEution. Compute the CUSUM centred on a b-value peak, and identify change-
points.
(iii) Discrimination. With a knowledge of detection thresholds throughout the catalogue,
evaluate the significance of the high b-value occurrence (equation 9). This should be done
for a succession of radii about the b-value peak.
(iv) Mensuration. Assess the radius of the anomaly. Repeat step (ii) at this stage to check
the resolution of the anomaly in time. Evaluate the moment deficit.
(v) Prediction. Estimate the date of earthquake occurrence from the dates of onset and
ending of the anomaly.

7 Discussion
The occurrence of periods of high b-value before the Cape Campbell earthquake and the
Wanganui swarm is striking. Other examples quoted are less so, but the fact that the only
statistically significant occurrences of high b-value within New Zealand preceded large
834 W. D. Smith
earthquakes is important. My contention is that a high b-value may be a more accurate
description of the phenomenon than the commonly used term 'quiescence'. The two
descriptions could be confused if the detection threshold is not low enough.
This study is an empirical one. The physical processes involved have therefore not been
examined. What physical mechanism manifests itself in a doubling of the 6-value for a period
of years, then a return t o normal (or less than normal, if foreshocks have low 6-values)
before the earthquake occurs? Is there significance in the fact that this was not observed for
two major strike-slip earthquakes? A better understanding of the b-value itself may help to
answer these questions.
In the case of the Coalinga data, the background seismicity contributing t o the high

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


b-value is largely on the nearby San Andreas Fault. It could be objected that such seismicity
should not be related to earthquakes in the Coalinga area, of which there were very few in
the period represented in Fig. 7. The empirical observation, however, is that this seismicity
did have a b-value of almost twice the normal value there, over an area 90 km in radius. The
conclusion I draw is that the stress buildup in the Coalinga area affected the seismicity out
to that distance, including that on the San Andreas Fault.
And what of the migration of the b-value peak towards the eventual epicentre (Fig. l)?
The migration velocity in Fig. l(a-f) is about 30 km yr-' . Can this be related to the buildup
of strain? The migration of the b-value peak is shown in Fig. 13. The ordinate is the distance
along the line AB shown in the inset map, which passes through the peaks in Fig. l(a and 9.
The abscissa is the date at the centre of the four-year sampling window. The approximate
location of the b-value peak is shown at six month intervals, and the epicentre of the Cape
Campbell earthquake in 1977 January is marked also. Some variation in the migration
velocity is apparent. Such a plot may provide another tool for predicting the location of the
earthquake.
Little significance has been attached to the actual b-value attained during the anomalous
period. This is because of the difficulty in measuring variations in b in both time and space.
A model such as proposed in Fig. 1O(c) may be the key to further developments here.
The use of the moment deficit during the period of the anomaly t o evaluate the magni-
tude of the forthcoming earthquake is also interesting. Why should the moment be only that
accumulated during such a short period and not during the entire interval since the last large

m
Q

9
a
100

Figure 13. Migration of the Cape Campbell b-value peak, along the line AB shown in the inset map. The
star represents the epicentre of the earthquake in 1977.
Changes in the frequency-magnitude b-value 835
earthquake? Perhaps this procedure is appropriate only for moderate earthquakes and not
for great earthquakes, which are the chief contributors to the overall moment rate in a given
region. The implication of a very high b-value, that the moment-magnitude relationship no
longer holds, is also important. Molnar (1979) has commented on some possibilities.
These unanswered questions notwithstanding, the indication from the data presented here
is that the b-value is a medium-term precursor, at least for earthquakes with a significant
dip-slip component. The systematic analysis of the New Zealand catalogue showed a high
b-value before all shallow earthquakes of magnitude 6 and greater since 1965, except for
some where a plausible reason for its non-detection is offered. Further, the only significant
occurrences of high b-values were those which preceded large earthquakes. On a routine

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


basis, recognition, resolution, discrimination and mensuration of the anomaly demand
complete detection of background seismicity to a low magnitude threshold. Discrimination
also requires that a long data base be established. It is hoped that the development of a
reliable procedure for estimating the time of occurrence of the large event which follows a
period of high b-value will result when further observations are to hand.

Acknowledgments
Most of this work was done while 1 was on study leave at the Pacific Geoscience Centre,
Sidney, British Columbia, and I am grateful t o Roy Hyndman, Director of that Centre, for
facilities generously provided and for his helpful comments. Financial support while in
Canada came from the New Zealand Department of Scientific and Industrial Research.
Colleagues in Sidney and in Wellington have been very patient and helpful, in particular
Dieter Weichert, Carry Rogers, Trevor Lewis, Euan Smith, Russell Robinson and John
Haines. For the statistical work I leaned on advice from David Brillinger and Keith Worsley.
Keith Hastings acquainted me with the procedure described in the Appendix. I am also
grateful t o Bob Uhrhammer and Steve Wesnousky for the Berkeley and Caltech catalogues
respectively, to Bill Price who drafted the figures, and to M a x Wyss whose review identified
areas of the text which needed clarification. A final word of gratitude is due to many
anonymous seismologists and technical staff who, over more than 40 years, located
thousands of small earthquakes.

References
Aki, K . , 1965. Maximum likelihood estimate of b in the formula log N = 4 - hM and its confidence limits,
Bull. Earthq. Res. inst. Univ. Tokyo, 43, 237-239.
Allen, C. R., 1976. Responsibilities in earthquake prediction, Bull. seism. SOC.Am., 66, 2069-2074.
Allen, C. R. & Nordquist, J . M., 1972. Foreshock, main shock and larger aftershocks of the Borrego
Mountain earthquake, Prof. Pap. USgeol. Surv., 787, 16-23.
Bakun, W. H., 1984. Seismic moments, local magnitudes and codaduration magnitudes for earthquakes
in central California, Bull. seism. SOC.Am., 74, 439-458.
Eaton, J., Cockerham, R. & Lester, F., 1983. Study of the May 2, 1983 Coalinga earthquake and its after-
shocks, based on the USGS seismic network in northern California, CDMG Spec. Publ. 66.
Evison, F. F., 1977a. Fluctuations of seismicity before major earthquakes, Nature, 266,710-712.
Evison, F. F., 1977b. Precursory seismic sequences in New Zealand, iVZ J. Geol. Geophys., 20, 129-141.
Fiedler, G., 1974. Local b-values related to seismicity, Tectonophys., 23, 277-282.
Garrick, R. A . & Gibowicz, S. J., 1983. Continuous swarm-like seismicity: the Wanganui, New Zealand
earthquakes, Geophys. J. R . 4sfr. SOC., 75,493-512.
Gibowicz, S. J., 1973. Variation of the frequency-magnitude relation during earthquake sequences in
New Zealand, Bull. seism. SOC. Am., 63,517-528.
836 W. D. Smith
Gibowicz, S. J., 1974. Frequency-magnitude, depth and time relations for earthquakes in an island arc:
North Island, New Zealand, Tectonophys., 23, 283-297.
Gutenberg, B . & Richter, C. F., 1944. Frequency of earthquakes in California, Bull. seism. Soc. Am., 34,
185 -1 88.
Habermann, R. E., 1981. Precursory seismicity patterns: stalking the mature gap, in Earthquake
Prediction - an International Review, Maurice Ewing series, 4 , 29-42.
Habermann, R. E., McCann, W. R. & Nishenko, S. P., 1983. A gap is .. ., Bull. seism. Soc. Am., 73,
1485-1486.
Habermann, R. E., & Wyss, M., 1984a. Earthquake triggering during preparation for great earthquakes,
Geophys. Res. Lett., 1 1 , 2 9 1-294.
Habermann, R. E. & Wyss, M., 1984b. Background seismicity rates and precursory seismic quiescence:
Imperial Valley, California, Bull. seism. Soc. Am., 74, 1743-1755.

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


Haines, A. J., 1981. A local magnitude scale for New Zealand earthquakes, Bull. seism. Soc.Am., 25,
1-32.
Hyndman, R. D. & Weichert, D. H., 1983. Seismicity and rates of relative motion on the plate boundaries
of Western North America, Geophys. J. R. astr. Soc., 72, 59-82.
Ishida, M . & Kanamori, H., 1977. The spatio-temporal variation of seismicity before the 1971 San
Fernando earthquake, California, Geophys. Res. Lett., 4 , 345-346.
Johnson, C. E., Rojahn, C. & Sharp, R. V. (eds), 1982. The Imperial Valley, California, earthquake,
October 15,1979,ProK Pap. USgeol. Surv. 1 2 5 4 , 4 5 1 pp.
Johnson, L. R. & McEvilly, T. V., 1974. Near-field observations and source parameters of Central
California earthquakes, Bull. seism. Soc. Am., 64,1855-1886.
Kanamori, H _,1981. The nature of seismicity patterns before large earthquakes, in Earthquake Prediction
- an International Review, Maurice Ewing Series, 4 , 1-19.
Kelleher, J. & Savino, J., 1975. Distribution of seismicity before large strike-slip and thrust-type earth-
quakes, J. geophys. Res., 80,260-271.
Kendall, M. G . & Stuart, A., 1977. 7'he Advanced Theory of Statistics, 4th ed, Chapter 16, Vol. 1, Mac-
millan, New York.
Khattri, I . & Wyss, M., 1978. Precursory variation of seismicity rate in the Assam area, India, Geology, 6 ,
685.
Kiek, S . N., 1984. Earthquake prediction from b-values, Nature, 310, 456.
Knopoff, L., Kagan, Y. Y. & Knopoff, R., 1982. b-values for foreshocks and aftershocks in real and
simulated earthquake sequences, Bull. seism. Soc. Am., 72, 1663-1676.
Li Quan-lin, Chen Jin-biao, Yu Lu & Hao Bai-lin, 1978. Time and space scanning of the b-value: a method
for monitoring the development of catastrophic earthquakes, Acta geophys. Sinica, 21, 101-1 25.
Translated in Chinese Geophys., 2,225-255.
McNally, K., 1982. Variations in seismicity as a fundamental tool in earthquake prediction, Bull. seisrn.
Soc. Am., 72, S 3 5 1 4 3 6 6 .
Mogi, K., 1969. Some features of recent seismic activity in and near Japan (2). Activity before and after
great earthquakes, Bull. Earthq. Inst. Univ. Tokyo, 47, 395-417.
Mogi, K., 1979. Two kinds of seismic gaps,&re appl. Geophys., 117, 1172-1 186.
Molnar, P., 1979. Earthquake recurrence intervals and plate tectonics, Bull. seism. Soc. Am., 69, 115-
133.
Ohtake, M., Matumoto, T. & Latham, G., 1977. Temporal changes in seismicity preceding some shallow
earthquakes in Mexico and central America, Bull. Int. Inst. Seisrn. Earthq. Eng., 15, 105-123.
Page, E. S., 1954. Continuous inspection schemes, Biometrika, 41, 100-1 14.
Page, R., 1968. Aftershocks and microaftershocks of the great Alaska earthquake of 1964, Bull. seism.
Soc. Am., 58,1131-1168.
Robinson, R., 1979. The variation of energy release, rate of occurrence and b-value of earthquakes in the
main seismic region, New Zealand,Phys. Earth planet. Int., 18, 209-220.
Sadovsky, M., Nersesov, I., Nigmatullaev, S., Latynina, L., Lukk, A., Semenov, A., Simbireva, 1. &
Ulomov, V., 1972. The processes preceding strong earthquakes in some regions of Middle Asia, in
Forerunners of Strong Earthquakes, eds Savarensky, E. F. & Rikitake, T., Tectonophys., 14,
295 -307.
Shi, Yaolin & Bolt, B. A., 1982. The standard error of the magnitude-frequency b-value, Bull. seism.
Soc. Am., 72: 1677-1687.
Smith, W. D., 1981. The b-value as an earthquake precursor,Nature, 289, 136-139.
Changes in the frequency -- magriitude b-value 837
Suyehiro, S., 1966. Differences between aftershocks and foreshocks in the relationship of magnitude to
frequency of occurrence for the great Chilean earthquake of 1960, Bull. seism. SOC. Am., 56,
185 -200.
Thatcher, W. & Hanks, T. C., 1973. Source parameters of southern California earthquakes, J. geophys.
Res., 78, 8 5 4 1-85 16.
Utsu, T., 1966. Statistical significance test of the difference in b-value between two earthquake groups,
J. Phys. Earth, 14,37-40.
Utsu, T., 1970. Aftershocks and earthquake statistics, J. Faculty Sci. Hokkaido Univ. Japan, Ser. VII. 3
(Part II), 197-266.
Utsu, T., 1971. Aftershocks and earthquake statistics, J. Faculty Sci. Hokkaido Univ. Japan, Ser. VII, 3
(Part III), 379-441.
Weichert, D. H., 1980. Estimation of the earthquake recurrence parameters for unequal observation

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


periods for different magnitudes, Bull. seism. SOC.Am., 70,1337-1346.
Wkitcomb, 1. H., 1 9 7 1 . Fault-plane solutions of the February 9 , 1971 San Fernando earthquake and some
aftershocks, ProJ Pap. USgeol. Surv. 733, 30-32.
Worsley, K . J., 1986. Confidence regions and tests for a change-point in a sequence of exponential family
random variables, in Proceedings Pacific Statistical Congress, pp. 266--272, eds Francis, I . S.,
Manley, B . J. F. & Lon], F. C., Elsevier (North-Holland).
Wyss, M. & Habermann, R. E., 1979. Seismic quiescence precursory t o a past and a future Kuril Island
earthquake, Pureappl. Geophys., 117, 1195-121 1.
Wyss, M. & Lee, W. H. K., 1973. Time variation of the average earthquake magnitude in Central
California, in Proc. Conf. Tectonic Problems of the San Andreas Fault System, 24-42, School of
Earth Science, Stanford University.
Zhang, J. Z. & Song, L.-Y. 1981. On the method of estimating b-value and its standard error, Actu seism.
Sinica, 3, 292-301.

Appendix: significance of weighted means


Figs 1 and 3-7 were prepared using b-values weighted in space, in order to better detect
anomalies of limited spatial extent. In the statistical formulation of Section 4 a weighted
mean magnitude can be used, instead of a simple mean. Let individual exponential variables
m , i = 1 , N be assigned weights wi. The weighted mean can be formed as
N
C wimi
i= 1
r72=
C wi
i= 1
The F-test which expresses the probabilities in equation (9) has degrees of freedom 2nl and
2 n 2 , for sample sizes n and n2 . These should be replaced by their effective values when
weights are used.
The procedure adopted in stratified sampling theory is t o note that an exponential
distribution is x;, and that a linear combination of x2 variables is also x '. So if x is
distributed as c&, where c is some constant and K the effective number of degrees of
freedom, then this distribution has mean p = cK and variance u2 = 2c2 K . So if the mean and
variance are known, we have K = 2 p2 /a2. In practice, however, the mean and variance must
be estimated, giving

N
s2 wi'
i= 1
838 W. D. Smith
where is the unweighted mean and s2 the sample variance of the mi. However this is not a
stable estimate, particularly for small samples. It often gives an estimate of K greater than
2N. For a stable estimate of K , use can be made of the fact that for an exponential
distribution (of magnitudes, in this case) the variance is equal to the square of the mean.
Equation (A2) then reduces t o

(A31
N
c4

Downloaded from https://academic.oup.com/gji/article/86/3/815/615025 by guest on 02 April 2024


i= 1
So degrees of freedom 2n and 2n2 (equation 9) are replaced by K 1 and K 2 determined as
in equation (A3) for each sample.

You might also like