Bulletin of the Seismological Society of America, Vol. 100, No. 4, pp. 1551–1567, August 2010, doi: 10.

1785/0120080354

Multivariate Bayesian Regression Analysis Applied to Ground-Motion

Prediction Equations, Part 1: Theory and Synthetic Example
by Danny Arroyo and Mario Ordaz

Abstract An application of a linear multivariate Bayesian regression model to

compute pseudoacceleration (SA) ground-motion prediction equations (GMPEs) is
presented. The model is able to include the correlation between observations for a
given earthquake, the correlation between SA ordinates at different periods, and the
correlation between regression coefficients of the ground-motion prediction model.
We evaluate the advantages of the Bayesian approach over the traditional regression
methods, and we discuss the differences between univariate and multivariate analyses.
Because the application of the Bayesian method is in general complex and implies an
increase in the numerical effort with respect to the traditional methods, our computer
code to perform linear Bayesian analyses is freely available on request.

Introduction
In the past, empirical pseudoacceleration (SA) ground-
motion prediction models were constructed by fitting avail-
able data to certain functional forms, using the least-squares
method. Several authors observed that in some cases the
decay of SA with distance could not be correctly determined
with this method because it disregarded the correlation
between observations recorded at different sites for a given
earthquake (Campbell, 1981, 1985; Joyner and Boore, 1993,
1994). The two-stage regression method and the one-stage
maximum-likelihood method were then developed to solve
this problem (Brillinger and Priesler, 1984; Abrahamson
and Youngs, 1992; Joyner and Boore, 1993, 1994). The one-
stage maximum-likelihood approach was introduced by
Brillinger and Priesler (1984), and Abrahamson and Youngs
(1992) and Joyner and Boore (1993, 1994) proposed com-
putational algorithms to implement it.
In many cases, however, the information contained in
data sets is not enough to properly constrain all regression
coefficients for a given functional form and, in practice, cer-
tain coefficients are fixed in order to stabilize the regression
analysis. In this approach, which can be considered as the
constrained version of the maximum-likelihood method,
the fixed values are defined by careful reviews of individual
terms of the adopted functional form.
Some ground-motion prediction equations (GMPEs) have
been derived using univariate Bayesian analysis (Veneziano
and Heidari, 1985; Ordaz et al., 1994; Reyes, 1999; Sibilio,
2006; Wang and Takada, 2009). Ordaz et al. (1994) discussed
the advantages of the Bayesian analysis with respect to the
least-squares method. However, the correlation between
observations recorded at different sites for a given earthquake
was not included in the model presented by Ordaz et al.
(1994), nor in other similar studies. The model that we present
in this article can be considered an extension of the original
work of Ordaz et al. (1994). Nevertheless, our model is more
general and is able to include the correlation between obser-
vations recorded at different sites for a given earthquake, the
correlation between SA ordinates at different periods, and the
correlation between regression coefficients of the GMPE.
In this article we discuss the theory of the Bayesian
model, and we compare a GMPE obtained through the pro-
posed technique with GMPEs obtained with the least-squares
and the one-stage maximum-likelihood methods. For the
comparisons, we use as a benchmark a set of synthetic SA
spectra with predefined statistical parameters. For the pre-
sented examples we used only the one-stage maximum-
likelihood method because it has been documented that this
method and the two-stage method lead essentially to the
same results (Joyner and Boore 1993, 1994). Finally, in the
last part of the article we discuss the differences between
multivariate and univariate analyses.
The Regression Model
For a given period T, a standard shape shown in equa-
tion (1) was adopted as the GMPE,
y
T  α1
T  α2
T
Mw  6  α3
T
Mw  62
 α4
T ln
R  α5
TR; (1)
where y
T is the natural logarithm of SA
T, Mw is the mo-
ment magnitude, R is the closest distance to the rupture area,
and αi
T are the coefficients to be determined by regression
analysis. Although in this article we have used the functional
form shown in equation (1), the procedure presented can be
readily applied to other linear functional forms.

1551
1552 D. Arroyo and M. Ordaz

The multivariate regression model is defined in equa- According to these assumptions, matrix Φ is a block
tion (2): diagonal matrix,
Y  XαT  E; (2) 2 3
ϕ1 0    0
where the superscript T stands for transpose, Y is a known 6 0 ϕ2    0 7
6 7
no × nT matrix that includes no observations of y
T for nT Φ  6 .. .. . . .. 7; (5)
4 . . . . 5
periods, X is a known no × np matrix that comprises no ob- 0 0    ϕne
servations of np parameters in the model (note that according
to equation 1 the elements of the first column of the matrix X where ne is the number of earthquakes and the squared sub-
are equal to unity), α is an unknown nT × np matrix that com- matrix ϕi related to earthquake i is given by
prises the coefficients determined by regression analysis (each 2 3
1 γe    γe
row of α contains the αi
T coefficients for a given T), and E 6 γe 1    γe 7
is an unknown no × nT matrix that is comprised of the regres- 6 7
ϕi  6 .. .. . . .. 7: (6)
sion residuals. 4 . . . . 5
It is assumed that the elements of E are correlated, γe γe    1
normally distributed random variables, with zero mean.
The rank of ϕi is equal to the number of records of the earth-
The correlation between elements of E is defined through
quake i. Note that γ e is equal to the γ parameter in Joyner
an unknown no nT × no nT matrix Ω, which is defined in
and Boore (1993, 1994), and ρT 1 ;T 2 is the coefficient of cor-
equation (3):
relation between spectral ordinates SA
T for a given pair of
Ω  Φ⊗Σ; (3) periods, namely T 1 and T 2 . In summary, we used for the mul-
tivariate case the same structure of matrix Φ that was used by
where Φ is an unknown no × no matrix that accounts for the
Joyner and Boore (1993) for the univariate case.
correlation between the rows of Y, Σ is an unknown nT × nT
Some authors have identified that the intraevent correla-
matrix that accounts for the correlation between spectral
tion is a function of the distance between stations (Boore
ordinates, and the symbol ⊗ stands for Kronecker product.
et al., 2003; Kawakami and Mogi, 2003; Wang and Tanaka,
2005). The model can be extended to include the spatial cor-
The One-Stage Maximum-Likelihood Method relation between observations by using a value of γ e that
depends on the distance between stations and performing the
In this section we extend the equations proposed by Bayesian analysis consistently. This variant, however, has
Joyner and Boore (1993, 1994) for the univariate model to not been pursued in this article.
the multivariate case. For a given γ e , the values of α and Σ that maximize the
For the model described in equation (2), the likelihood likelihood are the well-known weighted least-squares estima-
of Y is defined in equation (4): tors, defined in equations (7) and (8):
L
Yjα; Σ; Φ; X ∝ jΣjno =2 jΦjnT =2

 ^ T 
XT Φ1 X1 XT Φ1 Y;
α (7)
1 1 1
× exp  TrΦ
Y  Xα Σ
Y  Xα   ;
T T T (4)
2
^ T T Φ1
Y  Xα
^ 
Y  Xα ^ T
where Tr denotes trace and the symbol ∝ stands for pro- Σ : (8)
no
portionality because we have omitted the normalization
constant. In the one-stage maximum-likelihood method, the value of
Following Joyner and Boore (1993, 1994) we consid- γ e that maximizes the likelihood is found iteratively. The
ered that the elements of E, εij , can be expressed as the sum values of α and Σ for the regression analysis are then com-
of earthquake-to-earthquake variability (εe ) and record-to- puted from equations (7) and (8). Normally, instead of
record variability (εr ). In addition, the following considera- directly maximizing equation (4) the maximization is per-
tions were made: formed over its natural logarithm, given by
1. For a given earthquake and a given site, the coefficient of no n
correlation between residuals for two different periods, ln
L
Yjα; Σ; Φ; X ∝ ln
jΣj  T ln
jΦj
2 2
say T 1 and T 2 , is equal to ρT 1 ;T 2 . 1 1
2. For a given earthquake, the coefficient of correlation 
TrΣ
Y  Xα ^ T T
2
between residuals for the same period at different sites × Φ1
Y  Xα ^ T : (9)
is equal to γ e .
3. For a given earthquake, the coefficient of correlation Although not shown, it can be demonstrated that the prob-
between residuals for two different periods, say T 1 and ability distribution of α is a matrix Student t-distibution. The
T 2 , at different sites is equal to γ e ρT 1 ;T 2 . mean value of α is α, ^ and the covariance matrix of α ^V 
4. Residuals related to different earthquakes are independent. vec
α^ (note, that vec stands for stack) is given by
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1
1 ν=2
COV
α
^ V  f
XT Φ1 X1 ⊗
Y  Xα
^ T T
no  np  2
× Φ1
Y  Xα
^ T g: (10)
Note that COV
α ^ V  exists only if the number of degrees of
freedom of the distribution (i.e., no  np ) is greater than 2. It
is also worth noting that, even in the multivariate case, the
least-squares method is a particular case of the one-stage
maximum-likelihood method. The well-known least-squares
estimators can be found setting γ e  0:0 (i.e., Φ  I) in
equations (4)–(10).
The Bayesian Model
In the Bayesian approach α, Σ, and Φ are regarded as
matrix random variables with known joint prior density
p
α; Σ; Φ. This prior density is updated through Bayes
theorem, and the posterior density is given by the product
between the likelihood and the prior density,
p
α; Σ; ΦjX; Y ∝ L
Yjα; Σ; Φ; Xp
α; Σ; Φ: (11)
In standard Bayesian analysis, three types of p
α; Σ; Φ are
commonly used: vague or noninformative densities, conju-
gate densities, and generalized conjugate densities. Vague
densities are used when prior knowledge about parameters
is diffuse, and conjugate and generalized conjugate den-
sities are used when prior information about parameters is
available. A more detailed description of each family of
probability density functions and their implications in the
regression analysis can be found elsewhere (see, for example,
Broemeling, 1985; Rowe 2002). In this article, we adopted
a generalized conjugate probability density function as
basic density. In order to keep the structure of Φ shown in
equation (5), we used a scalar beta density for γ e. Thus, the
prior density of Φ is not of standard form. The prior joint
probability density used in our analysis is given by
p
αV ; Σ; Φ  p
αV p
Σp
Φ: (12)
According to equation (12), the regression model considers
that, a priori, α, Σ, and Φ are independent; note that in
equation (12) αV  vec
α.
Following Rowe (2002), we assume that the prior den-
sity of αV is the normal density defined in equation (13) with
mean αV0 and covariance matrix Δ. Thus, αV0  vec
α0 ,
where α0 is the prior mean value of α, and the positive
nT nP × nT nP matrix Δ is the prior covariance matrix of
αV0 . In other words,


1=2 1 1
p
αV  ∝ jΔj exp 
αV  αV0  Δ
αV  αV0  :
T
2
(13)
For Σ we used as prior density the inverted Wishart (Rowe,
2002) shown in equation (14) with parameters ν and Q,
1553


1 1
p
Σ ∝ jΣj exp  TrΣ Q : (14)
2
According to the properties of the inverted Wishart density,
the positive nT × nT matrix Q can be computed from the
prior mean value of Σ as follows:
Q 
ν  2nT  2Σ0 ; (15)
where Σ0 is the prior mean value of Σ and the scalar ν is a
measure of our degree of certainty on Σ0 . In order to give a
finite value to the variance of the elements of Σ, the value of
ν should be greater than 2nT  4; the larger the value of ν,
the greater the degree of certainty on Σ0 .
In Bayesian analysis, usually an inverted Wishart den-
sity is also used for Φ (Rowe, 2002). However, if it is desired
that Φ has the structure shown in equation (5), an inverted
Wishart density cannot be used. After noticing that Φ is a
function only of γ e , we decided to use a scalar beta density
for γ e,
p
γ e  ∝ γ ea1
1  γ e b1 ; (16)
where parameters a and b can be computed from the prior
mean value and standard deviation of γ e .
In summary, the prior information about regression
parameters is included in the analysis through αV0 , Δ, Q,
ν, a, and b, which are known as hyperparameters, and equa-
tions (12)–(16). Substituting equations (4), (13), (14), and
(16) into equation (11), we obtain the posterior joint density
of the regression parameters,
p
α; Σ; γ e jX; Y ∝ jΣj
no ν=2 jΦjnT =2 γ ea1
1  γ e b1


1
× exp 
αV  αV0 T Δ1
αV  αV0 
2

1
× exp  TrfΣ1 
Y  XαT T
2

1
× Φ
Y  Xα   Qg :
T
(17)
This joint density should be marginalized in order to obtain
the posterior marginal mean values of α, Σ, and γ e . However,
for this density, it is not possible to obtain marginal distribu-
tions in an analytical closed form. Posterior marginal mean
values can only be numerically computed, for which we use
the stochastic integration method known as Gibbs sampling.
Gibbs Sampling Method
Given the posterior joint density defined in equa-
tion (17), posterior marginal mean values can be estimated
by averaging random variates generated from the posterior
conditional densities of α, Σ, and γ e , given in equa-
tions (18)–(20),
1554 D. Arroyo and M. Ordaz


1
p
αV jΣ; Φ; X; Y ∝ exp 
αV  α
~ V T According to equation (18), α ~ V is the mean of the
2 posterior conditional density of α, and it is computed as the
×
Δ1  XT Φ1 X⊗Σ1  weighted average between the prior mean value and the con-
 ditional weighted least-squares estimate (see equation 21).
×
αV  α
~ V ; (18) Hence, it is interesting to assess the contribution to the final
estimate of α of the prior information, in comparison with
the contribution of the data. These contributions can be eval-
p
Σjα; Φ; X; Y ∝ jΣj
no ν=2 uated with vectors Wp and Wd , defined in equations (27) and

1 (28), respectively,
× exp  TrfΣ1 
Y  XαT T
2
 Wp  diag
Δ1  Δ1 1 1
d  Δ ; (27)
1
× Φ
Y  Xα   Qg ;
T
(19)

Wd  diag
Δ1  Δ1 1 1
d  Δd ; (28)
nT =2
p
γ e  ∝ γ ea1
1  γe
b1
jΦj
  where Δ1 T 1
d  X Φ X⊗Σ
1
is the inverse of the covari-
1 1 1
exp  TrfΦ 
Y  Xα Σ
Y  Xα  g ;
T T T ance matrix of the conditional weighted least-squares esti-
2 mate of α. The posterior marginal mean values of Wp and
(20) Wd can be computed in a way similar to the one used to
where compute α,

Σ,

and γ
e .
~ V  Δ1  XT Φ1 X⊗Σ1 1 Δ1 αV0
α

XT Φ1 X⊗Σ1 α
^ V ; (21) Synthetic Data
In order to assess the performance of the least-squares
method, the one-stage maximum-likelihood method, and the
^ V  vecYT Φ1 X
XT Φ1 X1 :
α (22) Bayesian technique, we generated different sets of synthetic
SA
T spectra with predefined statistical properties. We con-
In order to compute posterior marginal mean values, starting sidered 25 structural periods ranging between 0 and 5.0 sec.
values of Σ and γ e must be assumed, say Σ
0 and γ
e0 , and
We generated six sets of synthetic spectra assuming the num-
then one has to cycle through, bers of earthquakes shown in Table 1, and we assumed that
1. α

l1  a random variate from equation (18) with Σ  each earthquake was recorded at the number of sites shown
Σ

l and Φ  Φ

l , in Table 1. Thus, the number of records for a given set is the
2. Σ
l1  a random variate from equation (19) with α 

product between the number of earthquakes and the number
α

l1 and Φ  Φ

l , of sites. In order to obtain a reasonable event sample, we
3. γ
e
l1  a random variate from equation (20) with α  considered that Mw followed a modified Gutenberg–Richter
α

l1 and Σ  Σ

l1 , distribution, with a minimum value of Mw equal to 6, a max-
imum value of Mw equal to 8.2, and β  2, where β is the
where Φ

l is the value of Φ related to γ
e
l .
parameter controlling the relative frequencies of earthquakes
The first s random variates, called the burn in sample, of different sizes. Also, we assumed that R followed a uni-
are discarded, and the following K terms are averaged in form distribution between 250 and 400 km.
order to compute the marginal mean values. In addition, The predefined statistical properties of the set of ground
the covariance matrix of α
can be computed by averaging motions were αp , Σp , and Φp . We set αp as the value of the
the covariance matrix related to each term of the Gibbs sam- GMPE proposed by Reyes (1999) for station CU of Mexico
pling method. Techniques to generate random variates from City, whose coefficients are shown in Table 2. Because the
the densities shown in equations (18)–(20) can be found else-
where (Rowe, 2002). The Gibbs sampling method almost
surely converges to the mean value of the population param-
eter (Rowe, 2002) regardless of the values of Σ and Φ used Table 1
as starting values. A more detailed discussion about the con- Synthetic Sets of SA
T Spectra Used
vergence of the Gibbs sampling can be found in Geman and Set Earthquakes Sites no
Geman (1984). The value of K required to attain conver- 1 3 4 12
gence of the Gibbs sampling depends on the correlation 2 5 10 50
between observations and on the covariance of the regression 3 10 10 100
parameters, so it has to be defined iteratively. In the compu- 4 20 10 200
tations presented in this article we attained convergence of 5 50 10 500
6 100 10 1000
the Gibbs sampling with K ranging from 200 to 500.
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1555

correlation between the residuals of the predictive model is
equal to the correlation between the logarithm of spectral
ordinates (Baker and Cornell, 2006), the diagonal terms of
Σp were set as the variances of the residuals (σ2 ) related to
the Reyes (1999) model (which are also presented in
Table 2), while the off-diagonal terms were computed with
the equation proposed by Baker and Cornell (2006) to esti-
mate the coefficient of correlation ρT 1 ;T 2 . For Φp, we used
the structure shown in equation (5) with γ e  0:2234, which
is the value that we infer from the results presented by Joyner
and Boore (1993, 1994). Given the number of earthquakes
and the number of records shown in Table 1, an no × nT
matrix random variate from a matrix normal distribution
was generated. The mean value of the distribution was set
equal to αp , and the covariance matrix was obtained from
Φp ⊗Σp . The attenuation of synthetic SA values with R is
presented in Figure 1 for the case of T  0 (i.e., PGA). Note
that, regardless of the set considered, αp , Σp , and Φp repre-
sent the statistical properties of the entire population of SA
T
spectra; hence these parameters were used as benchmark for
the regression analysis presented in the following sections.
Results for the Least-Squares Method
A comparison between regression parameters obtained
with the least-squares method and the benchmark is pre-
sented in Figure 2. The least-squares method is able to attain
the benchmark values only for the coefficient of the magni-
tude (α2
T) and for σ. For no greater than or equal to 100,
reasonable values of α2
T and σ are observed. On the other
hand, very unrealistic values for α3
T, α4
T, and α5
T are
observed. Positive values of α4
T and α5
T are observed
even for no  1000, while positive values of α3
T are
observed for no < 500. In some cases we have shortened
the vertical axis in order to improve the clarity of the plots;
therefore, although not shown, very unrealistic values were
observed for no  12 and no  50. This is due to the fact
that least-squares results are based only on the information
contained in the data set, and for no  12 we have only 3
earthquakes recorded at 4 different stations and for no  50
we have only 5 earthquakes recorded at 10 different stations.
In addition, in Figure 3 we have plotted the standard devia-
tion of the regression coefficients, that is, the square root of
the diagonal elements of the covariance matrix defined in
equation (10), as a function of T. As no increases, the scatter
of regression parameters decreases.
Results for the One-Stage Maximum-
Likelihood Method
Figure 4 shows a comparison of regression parameters
obtained with the one-stage maximum-likelihood method
and the benchmark. The results are very similar to those
observed for the least-squares method. In general, the one-
stage maximum-likelihood method is not able to attain the
benchmark values, except for α2
T and σ. The greater dif-
ferences between least-squares and one-stage maximum-
likelihood methods are observed for α1
T and α4
T, for

Table 2
αp and σ2 Used to Generate the Synthetic SA Spectra
T (sec) α1 α2 α3 α4 α5 σ2

0.0 5.8929 1.2457 0:09757 0:5 0:00632 0.17626

0.1 6.0831 1.1954 0:09668 0:5 0:00643 0.18784
0.2 6.7942 1.0675 0:09858 0:5 0:00732 0.18494
0.3 6.9623 1.1303 0:10357 0:5 0:00768 0.15107
0.4 6.7632 1.2513 0:09682 0:5 0:00727 0.19549
0.5 6.9039 1.2236 0:08753 0:5 0:00753 0.17397
0.6 6.5941 1.2748 0:06768 0:5 0:00693 0.18936
0.7 6.7755 1.3445 0:04662 0:5 0:0076 0.20463
0.8 6.5941 1.3676 0:03662 0:5 0:00705 0.20044
0.9 6.4534 1.347 0:0244 0:5 0:00648 0.18376
1.0 6.5638 1.3387 0:05429 0:5 0:00665 0.18297
1.2 6.6903 1.3167 0:05225 0:5 0:00723 0.19870
1.4 6.4825 1.3203 0:06347 0:5 0:00662 0.21055
1.6 6.4614 1.4268 0:10542 0:5 0:00632 0.23412
1.8 6.0912 1.4088 0:09393 0:5 0:00516 0.27786
2.0 5.8698 1.3854 0:05267 0:5 0:00505 0.28847
2.2 5.8367 1.4032 0:04392 0:5 0:00547 0.35229
2.4 5.838 1.4032 0:07922 0:5 0:00559 0.33916
2.6 5.858 1.3951 0:06917 0:5 0:00601 0.34523
2.8 5.6616 1.3937 0:0717 0:5 0:00583 0.33564
3.0 5.4214 1.4344 0:0608 0:5 0:00568 0.34968
3.5 4.6026 1.4899 0:06365 0:5 0:00438 0.33114
4.0 3.6367 1.5601 0:05951 0:5 0:00283 0.35594
4.5 3.098 1.4864 0:05698 0:5 0:00187 0.40088
5.0 3.2887 1.5282 0:03953 0:5 0:00364 0.44147
1556
Figure 1. PGA attenuation with distance for the synthetic data
used in this study (sixth synthetic set, Table 1).
which slightly better estimates are obtained with the least-
squares method. In Figure 5, the standard deviation of the
regression coefficients is shown; the trends are similar to
those observed for the least-squares method. Slightly larger
values of standard deviation are observed for the maximum-
likelihood method than those obtained with the least-squares
Figure 2. Results for least-squares method (benchmark values are plotted with a thick line).
D. Arroyo and M. Ordaz
method. In Table 3 estimates of γ e related to different sets are
shown; very accurate estimates are observed with no ≥ 200.
Prior Information for the Bayesian Method
In this section we present a short discussion about how
to define the prior information for the Bayesian analysis. We
have not included a sound discussion because a synthetic
example is presented. In a companion article (Arroyo and
Ordaz, 2010) we use a set of actual ground-motion records
and we present a complete discussion on how the prior in-
formation can be defined.
The elements of α0 were set as follows. With the
amplitude Fourier spectra defined by Brune’s model (Brune,
1970) and common attenuation factors, and using random
vibration theory, we constructed a set of SA spectra related
to several values of Mw and R (McGuire and Hanks, 1980;
Boore, 1983). Then, we fit the values so computed to the
functional form using the least-squares method in order to
compute α0 ; the elements of α0 obtained are shown in
Table 4. This implies that a priori we believe that the attenua-
tion of SA spectra can be properly characterized by Brune’s
model and common attenuation factors.
The block-symmetric matrix Δ is the prior covariance
matrix of αV and can be written as
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1557

Figure 3. Standard deviation of the regression coefficients for the least-squares method (the symbols are the same as in Fig. 2).

2 3 coefficient α1
T in equation (1) depends on site effects

δ11 δ12 
δ1np
(Ordaz et al., 1994), we assigned a large value, with respect
6 δ22 
δ2np 7
6 7 to the prior mean value of α1
T, to the diagonal elements of
Δ6 .... 7; (29)
4 . 5
. δ11 . This implies that α1
T is not controlled by its prior
δnp np mean value, so it is free to attain the value that yields the
where δij is an nT × nT symmetric matrix defined in equa- best fit in the regression analysis. In the computations we
tion (30): used cov
α1T K ; α1T K   10; 000. For δ22, δ33 , and δ44 we

2 3
cov
αiT 1 ; αjT 1  cov
αiT 1 ; αjT 2   cov
αiT 1 ; αjT nT 
6 cov
αiT 2 ; αjT 2   cov
αiT 2 ; αjT nT  7
6 7
δij  6 .. .. 7: (30)
4 . . 5
cov
αiT nT ; αjT nT 

In equation (30) cov
αiT k ; αjT l  is the prior covariance
between the coefficient αi
T for the period T k and the
coefficient αj
T for the period T l . The matrix Δ was set
as diagonal; therefore, only the diagonal elements of δ11 ,
δ22 , δ33 , δ44 , and δ55 are different from zero. This implies
that a priori we believe that the different αi
T random vari-
ables are uncorrelated. We set that structure for Δ because,
from the presented example, we do not have information
about correlation between the different αi
T. Because the
assigned a value that implies a coefficient of variation of
0.59 to their diagonal elements, as it was done in a previous
study (Ordaz et al., 1994). Similarly to δ11 , for the diagonal
elements of δ55 , we also used a large value with respect to the
prior mean value of α5
T; in the computations we used
cov
α5T K ; α5T K   1.
We set the diagonal elements of Σ0 equal to 0.49, which
means that a priori we believe that the expected standard
deviation of the residuals is equal to 0.7, independently of
1558 D. Arroyo and M. Ordaz

Figure 4. Results for one-stage maximum-likelihood method (benchmark values are plotted with a thick line).

Figure 5. Standard deviation of the regression coefficients for the one-stage maximum-likelihood method (the symbols are the same
as in Fig. 4).
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1559

Table 3 the equations of Baker and Cornell (2006) as prior informa-

Values of γ e Obtained through One-Stage Maximum- tion because they were used to generate the synthetic
Likelihood Method and Bayesian Method data.
Benchmark One-Stage Maximum-Likelihood Bayesian The degree of certainty of Σ0 depends on ν. As it has
γ e  0:2234 Method Method been discussed, the larger the value of ν, the larger the degree
no  12 0.7490 0.3892 of certainty on Σ0 . The posterior mean value of Σ can be
no  50 0.0000 0.2336 expressed as a weighted average between the prior mean
no  100 0.1148 0.2596 value and the conditional weighted least-squares estimator.
no  200 0.1872 0.2351
no  500 0.2114 0.2382
The weighting factors are ν=
no  ν for the prior mean
no  1000 0.2235 – value and no =
no  ν for the conditional weighted least-
squares estimate. In the computations we have used a value
equal to 55 (the minimum required for nT  25 in order to
give a finite value to the covariance of Σ0 ) because we be-
T. In addition, the off-diagonal terms were defined through lieve that Σ0 is very uncertain. For no  100 the weighting
the coefficient of correlation shown in equation (31): factor for prior information is about 0.35.
Finally, our prior information of γ e is vague, so we have
ρ0T 1 ;T 2  exp
qjT 1  T 2 j: (31) set a  b  1:5, which is a very flat density, with mean
value equal to 0.5, in order to force γ e to take the value that
Note that we use a very simple function to define the prior yields the best fit to data.
correlation between spectral ordinates of SA. According to
equation (31), together with q  1:0, the coefficient of cor- Results for the Bayesian Model
relation varies from unity, when T 1 and T 2 are equal, to
nearly 0.05 when the difference between T 1 and T 2 is about In order to exemplify the convergence of the Gibbs sam-
3 sec. It must be acknowledged that equation (31) is quite pling, Figure 6 shows a comparison between estimates of
arbitrary, and it was created only for the synthetic example regression parameters for K ranging between 50 and 200.
presented in this article. When working with actual ground As has been stated, any value of Σ
0 and γ
e0 could be used
motions it could be more reasonable to use the correlation as the starting value; for the computations we set γ
e0  0 and
coefficients defined by Baker and Cornell (2006) as prior Σ
0  I. In addition, the prior mean values used in the anal-
values. In the presented example we decided not to use ysis are plotted in Figure 6. We observed that the Gibbs sam-
pling method converges for relatively small values of K.
Although not shown, similar trends were observed for other
Table 4 values of no . In Table 5 convergence of the Gibbs sampling
α0 Used in the Analysis regarding γ e is presented; γ e also converges for K between
T (sec) α1 α2 α3 α4 α5
100 and 200. We note that these trends are valid only for the
presented example; for other cases of analysis it is recom-
0.0 7.463 0.785 0:026 0:758 0:005
0.1 6.303 0.835 0:035 0:601 0:005
mended to build plots similar to Figure 6 in order to assess
0.2 5.521 0.914 0:050 0:541 0:004 the convergence of the Gibbs sampling. The value of K
0.3 5.114 0.977 0:062 0:525 0:004 should be chosen carefully because computation time re-
0.4 4.826 1.032 0:073 0:518 0:003 quired for the analysis grows with K. In order to minimize
0.5 4.597 1.083 0:082 0:514 0:003 the computational effort, the peak value of K used in the
0.6 4.405 1.130 0:092 0:512 0:003
0.7 4.236 1.176 0:101 0:510 0:003
Bayesian analysis was 500.
0.8 4.084 1.221 0:111 0:509 0:003 In Figure 7 the weighting factors computed through
0.9 3.945 1.266 0:120 0:508 0:002 Gibbs sampling and equations (27) and (28) are shown for
1.0 3.817 1.310 0:130 0:507 0:002 no  200. As expected, for α1
T and α5
T the contribu-
1.2 3.575 1.398 0:150 0:506 0:002 tion of the prior mean values is marginal because we set,
1.4 3.363 1.485 0:170 0:505 0:002
1.6 3.149 1.568 0:189 0:505 0:002
a priori, a very large value for their covariance. In the case
1.8 3.015 1.641 0:205 0:504 0:002 of α2
T the contribution of the prior mean value is also
2.0 2.873 1.699 0:216 0:503 0:002 marginal. However, as can be observed in Figure 6, the pos-
2.2 2.740 1.729 0:216 0:503 0:002 terior mean values of α2
T are similar to the prior mean
2.4 2.542 1.718 0:200 0:503 0:002 values especially for systems in the range between 0.1 and
2.6 2.482 1.616 0:154 0:502 0:002
2.8 2.429 1.641 0:160 0:503 0:002
1.8 sec. This means that prior mean values of α2
T are
3.0 2.183 1.768 0:191 0:503 0:002 similar to the values obtained from data. On the other hand,
3.5 1.938 1.832 0:195 0:503 0:002 greater contribution of prior mean values is observed for
4.0 2.002 1.944 0:236 0:499 0:002 α3
T and α4
T. For α3
T the greater contribution of prior
4.5 1.485 1.995 0:251 0:532 0:001 mean values is observed for T shorter than 0.6 sec, while for
5.0 1.700 1.987 0:207 0:501 0:001
longer periods almost a constant value of 0.4 was observed.
1560 D. Arroyo and M. Ordaz

Figure 6. Convergence of Gibbs sampling for the set with no  200 (prior values are plotted with a thick line).

For α4
T a nearly constant value of 0.9 is observed. We note As has been stated, the Bayesian method was not able to
that in both cases the posterior mean values are close to their attain the correct value of α3
T. According to Figure 8, as
prior counterparts, independent of Wp ; this observation will no increases, slightly better estimates are observed, espe-
be discussed later in the article. cially for T larger than 2 sec. However, in general, posterior
The results obtained with the Bayesian model are pre- mean values are close to prior mean values of α3
T, even if
sented in Figures 8 and 9. Conversely to what happened with the weighting factor decreases. For example, in Figure 7 it
other models, the Bayesian model was able to attain bench- can be seen that for no  200, Wp is about 0.4 in the long-
mark values, except for α3
T. Note that with no between period range, while for no  500 (although not shown) a Wp
100 and 200 very accurate estimates of the regression param-
eters are observed. In general, the scatter of the Bayesian
regression coefficients is smaller than that observed with Table 5
other methods, which is an advantage of the Bayesian ap- Convergence of γ e for no  200
proach over other methods. In Table 3 estimates of γ e related Prior γ e  0:5
to different sets are shown; accurate estimates are observed K  12 γe  0:2367
for no greater than 50. Note that in spite of our use of a prior K  50 γe  0:2338
mean value of γ e  0:5 the data shifted the prior mean value K  100 γe  0:2353
K  200 γe  0:2351
to the correct value of γ e .
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1561

Figure 7. Weighting factors observed for no  200 (black circles, weighting factor for prior information; thick line, weighting factor for
data information).

value about 0.2 was observed. This means that information
contained in the data is not enough to completely define
α3
T; in other words, α3
T has little effect on y and almost
any value could have been used. Hence, the Bayesian
method, instead of leading to any value of α3
T (such as
other methods) leads to values of α3
T that are close to its
prior mean value.
the three methods yield very similar values of σ. Hence,
we decided to assess the global accuracy of the models as
follows. Suppose that a regression analysis is performed with
a set of size no and for a given T we obtain estimates of the
regression coefficients. Because for the synthetic example
the true coefficients are known, the expected value of the
error that would be observed if the estimated coefficients
(related to some value of no ) were applied is given by

Discussion Ez 
α1R  α^ 1  
α2R  α^ 2 E
Mw  6

Although it is clear that, at least for the presented exam-
ple, the Bayesian method usually yields regression param-
eters that are closer to the population parameters than those
obtained with other methods, it must be acknowledged that
Bayesian estimates are not related to the minimum standard
error. Furthermore, as can be observed in Figures 2, 4, and 8,

α3R  α^ 3 E
Mw  62  
α4R  α^ 4 Eln R

α5R  α^ 5 ER  Eε; (32)
where αjR is the true value of the j coefficient of the regres-
sion and α^ j is its corresponding estimate. Note that for a very
large set α^ j tends to αjR ; hence, Ez  Eε  0.
1562
Furthermore, the variance of z is given by
Ez2  
α1R  α^ 1 2 
α2R  α^ 2 2 E
Mw  62 

α3R  α^ 3 2 E
Mw  64 

α4R  α^ 4 2 E
ln R2  
α5R  α^ 5 2 ER2 
 Eε2i   2
α1R  α^ 1 
α2R  α^ 2 E
Mw  6
 2
α1R  α^ 1 
α3R  α^ 3 E
Mw  62 
 2
α1R  α^ 1 
α4R  α^ 4 Eln R
 2
α1R  α^ 1 
α5R  α^ 5 ER
 2
α2R  α^ 2 
α3R  α^ 3 E
Mw  63 
 2
α2R  α^ 2 
α4R  α^ 4 E
Mw  6 ln
R
 2
α2R  α^ 2 
α5R  α^ 5 E
Mw  6R
 2
α3R  α^ 3 
α4R  α^ 4 E
Mw  62 ln
R
 2
α3R  α^ 3 
α5R  α^ 5 E
Mw  62 R
 2
α4R  α^ 4 
α5R  α^ 5 ER ln
R:
Figure 8. Results for Bayesian method (benchmark values are plotted with a thick line).
D. Arroyo and M. Ordaz
For a very large set, α^ j tends to αjR , so Ez2   Eε2 , which
is the value of σ2 shown in Table 2.
In order to assess the global accuracy of the predictions
related to the three methods, in Figure 10 we present a com-
p
parison of σp  Ez2  for different values of no . For small
sets (i.e., no  12 and 50) the global accuracy of the Bayesian
method is better than those observed for the other two
methods. Nevertheless, for no greater than 100 the global
accuracy of the 3 methods is practically the same. Note that
the global accuracy for the least-squares and one-stage
maximum-likelihood methods is very similar in all cases
and that the same level of accuracy is attained with the three
methods regardless of differences observed in the regression
coefficients. As reference, in Table 6 the regression coeffi-
cients related to different methods are compared for
no  200 and T  0 (i.e., PGA). Hence, it can be concluded
that there are multiple solutions for the regression analysis that
are close to the minimum error solution, and as can be
observed in Figures 2, 4, and 8 those solutions might be very
(33) different.
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1563

Figure 9. Standard deviation of the regression coefficients for Bayesian method (the symbols are the same as in Fig. 8).

Based on the last observation one might consider that
the great computational and analytical work required by
the Bayesian model is not warranted. However, the Bayesian
approach has some advantages that make its application
worthwhile:
1. In the case of few data points, the accuracy of the Bayes-
ian method is greater than that obtained with other
methods. Usually, the scatter of the regression coeffi-
cients is smaller for the Bayesian model than for other
methods, so narrower confidence intervals would be
obtained for the Bayesian coefficients.
2. From theoretical grounds α4
T must be negative.
However, the least-squares and one-stage maximum-
likelihood methods yield positive values of this coeffi-
cient even for no  1000, that is, a very large sample.
In order to avoid this problem, some authors have fixed
α4
T during regression analysis, normally at a value of
0:5 for far-field ground motions and 1:0 for near-field
records. But, even without fixing a predefined value of
α4
T, the Bayesian analysis yielded a value for this
coefficient that matches with seismological theory. For
instance, regarding the results presented in Table 6,
even when the three methods yield the same level of
accuracy, most analysts would definitely not use results
obtained with the least-squares and one-stage maximum-
likelihood methods because α2
T and α4
T are very far
from what would be expected from seismological theory.
3. It is true that while values of M and R lie in the ranges
observed in the sample, the three methods yield the same
level of accuracy, even when some coefficients seem the-
oretically unacceptable. However, we decided to compare
the predicted SA spectra, obtained with coefficients
shown in Table 6, with the corresponding benchmark
spectra. For the comparison we choose M  7 and four
1564 D. Arroyo and M. Ordaz

Figure 10. Comparison of σp related to different methods (benchmark values are plotted with a thick line).

values of R (150, 200, 450, and 500 km). These values
are out of the range of the data contained in the synthetic
set; thus, this comparison can be regarded as an evalua-
tion of the possibility of extrapolating the GMPEs. The
results are summarized in Figure 11. The Bayesian
regression yields acceptable results, but the least-squares
and one-stage maximum-likelihood methods might lead
to very inaccurate results. Note that for R  200, large
differences are observed for the least-squares and one-
stage maximum-likelihood methods, in spite of the fact
that this distance is only 20% smaller than the minimum
value of R included in the synthetic set.
Based on the results discussed in this section we con-
clude that the analytical and computational effort required
by the Bayesian method is completely warranted. Moreover,
in the future, the computation time required by the Bayesian
analysis will certainly decrease as more powerful computers
will be developed.

Table 6
Regression Parameters for no  200 and T  0
Method α1
T α2
T α3
T α4
T α5
T

Least-squares 3.7863 1.4263 0:2264 0:1136 0:0065

One-stage maximum-likelihood 9.6409 1.4263 0:2252 1:3340 0:0028
Bayesian 7.5033 1.0779 0:0243 0:8729 0:0042
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1565

performing the analysis period by period. For the case of the

least-squares method it is clear, from equation (6) together
with γ e  0, that the results are equal for both types of anal-
ysis and that only the diagonal elements of matrix Σ can be
estimated through univariate analysis. Conversely, in the case
of the one-stage maximum-likelihood and the Bayesian
methods, the results obtained are different because the cor-
relation structures of E are also different. For the multivariate
analysis the correlations between elements of E are defined
in equation (3) (the value of γ e is the same for all values of
T), while in the univariate analysis the correlation structure
cannot be described by equation (3) because different values
of γ e are obtained for each value of T. That is, the covariance
matrix Ω is coupled and it cannot be expressed as the
Kronecker product of two matrices.
If the matrix Ω cannot be separated, the likelihood func-
tion should be written as

L
Yjα; Ω; X ∝ jΩj1=2


1
× exp  vec
Y  vec
XαT T Ω1
2

× vec
Y  vec
XαT  : (34)

Note that Ω is a function of γ e for each period (γ eTi ), of the

Figure 11. Comparison of SA spectra related to different meth-
ods (the symbols are the same as in Fig. 10).
One question that naturally arises after going through a
Bayesian analysis is how dependent the results are on the prior
information. It has been argued that, because of the inherent
subjectivity of the Bayesian approach, it lends itself to the
perpetuation of unrecognized error. However, in a previous
work of the second author it has been shown that this objection
is not justified (see Ordaz et al., 1994, for further details).
Differences between Univariate
and Multivariate Analysis
To finish the article, we discuss differences between per-
forming a multivariate analysis and the common practice of
coefficients of correlation between residuals for different
periods, and of the variances of the residuals for each period.
Although not shown, the elements of Ω can be defined
according to the following rules:
1. For a given earthquake and a given site, the coefficient
of correlation between residuals for two h different
p
periods, T 1 and T 2 , is equal to ρT 1 ;T 2 γ eT 1 γ eT 2 
pi

1  γ eT 1 
1  γ eT 2  .
2. For a given earthquake the coefficient of correlation be-
tween residuals for the same period (T i ) at different sites
is equal to γ eT i .
3. For a given earthquake the coefficient of correlation be-
tween residuals for two different periods, say T 1 and T 2 ,
p
at different sites is equal to ρT 1 ;T 2 γ eT 1 γ eT 2 .
4. Residuals related to different earthquakes are inde-
pendent.

Table 7
Comparison of Results Obtained through Multivariate and Univariate Analysis, T  0 and
no  200
Method* α1
T α2
T α3
T α4
T α5
T σ γe

Benchmark 5.8929 1.2457 0:09757 0:5 0:00632 0.41983 0.2234

MML 9.6409 1.4263 0:2252 1:3340 0:0028 0.3925 0.1872
UML 11.6577 1.4237 0:2232 1:7608 0:0015 0.3714 0.0362
MB 7.5033 1.0779 0:0243 0:8729 0:0042 0.4131 0.2351
UB 7.1722 1.0962 0:0275 0:80207 0:0044 0.3953 0.1362
*MML, multivariate maximum-likelihood; UML, univariate one-stage maximum-likelihood; MB,
multivariate Bayesian; UB, univariate Bayesian.
1566
The one-stage maximum-likelihood method for this
correlation structure consists of maximizing equation (34).
However, the maximization process is much more complex
than in the case of a single γ e for all periods because
maximization must be carried out considering simulta-
neously all the coefficients of correlation between residuals,
all the γ eT i parameters, and all the variances of the residuals.
Also, the Bayesian model becomes very complex if different
values of γ e are adopted for each period because prior
densities for all γ eT i and those parameters are coupled with
the variances of the residuals and with the coefficients of
correlation between residuals for different periods. Because
we consider that the application of this Bayesian model
would be impractical, the complete analysis is not presented
in this article. However, the complete analysis can be ob-
tained following the procedure presented in the section titled
The Bayesian Model.
Thus, the common practice of performing the analysis
period by period implicitly disregards the correlation be-
tween residuals for different periods. The influence of this
assumption in the accuracy of the GMPE depends on how
large this correlation is. In Table 7 we present a comparison
of regression parameters obtained through multivariate
maximum-likelihood, multivariate Bayesian, univariate one-
stage maximum-likelihood, and univariate Bayesian meth-
ods for no  200 and T  0 (i.e., PGA). Unsurprisingly,
the multivariate results are closer to the benchmark values,
especially for γ e, because the synthetic data were generated
with the correlation structure defined in equation (3).
Nevertheless, in the case of data obtained from actual ground
motions, the true correlation structure is unknown; hence,
we cannot consider that the multivariate analysis is more
accurate than the univariate analysis based only in the results
presented in Table 7. However, from the discussion presented
in this section we consider that the multivariate regression
model is theoretically more robust than the common practice
of performing the analysis period by period.
It is interesting to note that during the computations we
observed that the time required in the Bayesian analysis for
the univariate and the multivariate cases is almost the same
because the rank of matrix Φ is equal in both cases. In
practice, an approach that can be used is to perform the
univariate analysis and include the correlation between spec-
tral ordinates through the copula technique described in
Goda and Atkinson (2009).
Conclusions
We have presented a linear multivariate Bayesian regres-
sion method that includes the correlation between observa-
tions for a given earthquake, the correlation between SA
ordinates at different periods, and the correlation between
regression coefficients of the GMPE. Through comparisons
of GMPEs obtained with the least-squares and the one-stage
maximum-likelihood methods we have shown that multiple
solutions close to minimum error could exist and that the
D. Arroyo and M. Ordaz
Bayesian method could be used to obtain a GMPE consistent
with seismological theory. In addition, for the presented syn-
thetic example, it is shown that GMPEs obtained through
Bayesian analysis yield more accurate results than GMPEs
related to other methods when the GMPEs are extrapolated.
However, the Bayesian method requires significantly more
analytical and computational work than traditional methods.
Hence, our computer code to perform linear Bayesian ana-
lyses is freely available on request.
Data and Resources
No actual data were used in this article. The synthetic
data are available on request.
Acknowledgments
The authors appreciate the efforts of two anonymous reviewers,
who greatly contributed to improve the original manuscript. Constructive
comments by Gail Atkinson are also appreciated.
References
Abrahamson, N. A., and R. R. Youngs (1992). A stable algorithm for regres-
sion analysis using the random effects model, Bull. Seismol. Soc. Am.
82, 505–510.
Arroyo, D., and M. Ordaz (2010). Multivariate Bayesian regression analysis
applied to ground motion prediction equations, Part 2: Numerical ex-
ample with actual data, Bull. Seismol. Soc. Am. 100, 1568–1577.
Baker, J. W., and C. A. Cornell (2006). Correlation of response spectral
values for multicomponent ground motions, Bull. Seismol. Soc. Am.
96, 215–227.
Boore, D. M. (1983). Stochastic simulation of high-frequency ground
motions based on seismological models of the radiated spectra, Bull.
Seismol. Soc. Am. 73, no. 6, 1865–1894.
Boore, D. M., J. F. Gibbs, W. B. Joyner, J. C. Tinsley, and D. J. Ponti (2003).
Estimated ground motion from the 1994 Northridge, California, earth-
quake at the site of the Interstate 10 and La Cienega boulevard bridge
collapse, west Los Angeles, California, Bull. Seismol. Soc. Am. 93,
2737–2751.
Brillinger, D. R., and H. K. Preisler (1984). An exploratory analysis of the
Joyner–Boore attenuation data, Bull. Seismol. Soc. Am. 74, 1441–
1450.
Broemeling, L. D. (1985). Bayesian Analysis of Linear Models, Marcel Dek-
ker, New York, 454 pp.
Brune, J. N. (1970). Tectonic stresses and spectra of seismic waves from
earthquakes, J. Geophys. Res. 75, 4997–5009.
Campbell, K. W. (1981). Near-source attenuation of peak horizontal accel-
eration, Bull. Seismol. Soc. Am. 71, 2039–2070.
Campbell, K. W. (1985). Strong motion attenuation relations: A ten-year
perspective, Earthq. Spectra 1, no. 4, 759–804.
Geman, S., and D. Geman (1984). Stochastic relaxation, Gibbs distributions,
and the Bayesian restoration of images, IEEE Trans. Pattern Anal.
Mach. Intell. 6, 721–741.
Goda, K., and G. M. Atkinson (2009). Interperiod dependence of
ground-motion prediction equations: A copula perspective, Bull.
Seismol. Soc. Am. 99, no. 2A, 922–927.
Joyner, W. B., and D. M. Boore (1993). Methods for regression analysis
of strong-motion data, Bull. Seismol. Soc. Am. 83, no. 2,
469–487.
Joyner, W. B., and D. M. Boore (1994). Errata: Methods for regression
analysis of strong-motion data, Bull. Seismol. Soc. Am. 84, no. 3,
955–956.
Multivariate Bayesian Regression Analysis Applied to Ground-Motion Prediction Equations, Part 1 1567

Kawakami, H, and H. Mogi (2003). Analyzing spatial intraevent variability Wang, M., and T. Takada (2005). Macrospatial correlation model of seismic
of peak ground accelerations as a function of separation distance, Bull. ground motions, Earthq. Spectra 21, 1137–1156.
Seismol. Soc. Am. 93, 1079–1090. Wang, M., and T. Takada (2009). A Bayesian framework for prediction of
McGuire, R. K., and T. C. Hanks (1980). RMS accelerations and seismic ground motion, Bull. Seismol. Soc. Am. 99, no. 4, 2348–2364.
spectral amplitudes of strong ground motion during the San
Fernando, California, earthquake, Bull. Seismol. Soc. Am. 70, no. 5,
1907–1919. Departamento de Materiales
Ordaz, M., S. K. Singh, and A. Arciniega (1994). Bayesian attenuation Universidad Autónoma Metropolitana-Azcapotzalco
Av. San Pablo # 180. Colonia Reynosa Tamaulipas
regressions: An application to Mexico City, Geophys. J. Int. 117,
Azcapotzalco CP 02200
335–344. México D.F., México
Reyes, C. (1999). El estado límite de servicio en el diseño sísmico de aresda@correo.azc.uam.mx
edificios, Ph.D. Thesis, School of Engineering, UNAM. (D.A.)
Rowe, D. B. (2002). Multivariate Bayesian Statistics: Models for Source
Separation and Signal Unmixing, Chapman & Hall/CRC, New York,
329 pp. Instituto de Ingeniería
Sibilio, E. (2006). Seismic risk assessment of structures by means of stochas- UNAM Ciudad Universitaria
tic simulation techniques, Ph.D. Thesis, Braunschweig Technical Coyoacán CP 04510
University and Florence University. México D.F. México
mors@pumas.iingen.unam.mx
Veneziano, D., and M. Heidari (1985). Statistical analysis of attenuation in
(M.O.)
the eastern United States, in Methods of Earthquake Ground-Motion
Estimation for the Eastern United States, EPRI Research Project
RP2556-16, Palo Alto, California. Manuscript received 3 December 2008