HOSTED BY Available online at www.sciencedirect.

com

ScienceDirect
Soils and Foundations 57 (2017) 412–422
www.elsevier.com/locate/sandf

Statistical analysis for evaluating cyclic strength

of cement-treated soils
Tsutomu Namikawa a,⇑, Yoshio Suzuki b, Sadatomo Onimaru c, Takuya Tsukahara a,1,
Ryo Kurosawa a,1, Kentaro Shimada a,1
a
Department of Civil Engineering, Shibaura Institute of Technology, Tokyo, Japan
b
Mind Inc., Chiba, Japan
c
School of Policy Studies, Kwansei Gakuin University, Hyogo, Japan

Received 28 July 2016; received in revised form 9 December 2016; accepted 13 February 2017
Available online 12 May 2017

Abstract

This study presents a statistical analysis for evaluating the cyclic unconfined compressive strength of cement-treated soils. In labora-
tory tests on cement-treated soils, the strength of the specimens varies to some extent even though they are prepared in the laboratory.
Therefore, the influence of the variation in specimen strength on the cyclic loading test results should be assessed in order to obtain the
cyclic strength precisely. Moreover, the variability of the specimen strength results in non-failure specimen data even at a large number of
loading cycles. In the present study, the binary regression approach is adopted for analyzing the cyclic loading test results to take into
account the non-failure specimen data. In the binary regression approach, the cyclic loading test results are treated as having one of two
possible forms, namely, ‘failure’ or ‘non-failure’, at prescribed cycles. The parameters in the binary regression model are related to the
variability of the cyclic strength and the ratio of the cyclic strength to the unconfined compressive strength. The binary regression analysis
results reveal that the variability of the cyclic strength approximately corresponds to the variability of the specimen strength. The non-
failure specimen data is properly treated in evaluating the cyclic strength by adopting the binary regression analysis. Using the binary
regression analysis approach proposed in this study, an unnecessarily large number of cycles to failure is not required in the cyclic loading
tests. This reduces the experimental cost for obtaining the deterioration property of the materials.
Ó 2017 Production and hosting by Elsevier B.V. on behalf of The Japanese Geotechnical Society. This is an open access article under the CC BY-
NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Cement-treated soil; Strength; Cyclic loading; Variability; Statistical analysis (IGC: D6)

1. Introduction
Ground improvement by cement mixing has been widely
used for structural foundations and liquefaction mitigation
methods. When employing cement mixing for structural
foundations and liquefaction mitigation methods, cyclic
loadings induced by the inertia forces of the structure
Peer review under responsibility of The Japanese Geotechnical Society.
⇑ Corresponding author.
E-mail address: namikawa@shibaura-it.ac.jp (T. Namikawa).
1
Formerly.
and the soil mass are applied to the improved ground dur-
ing an earthquake (Namikawa et al., 2007; Khosravi et al.,
2016). These cyclic loadings are likely to degrade cement-
treated soils. Therefore, the deterioration of strength is
one of the major issues in the seismic design of ground
improvement by cement mixing, and the mechanical prop-
erties of cement-treated soils subjected to cyclic loadings
should be investigated to assess the internal stability of
ground improvement by cement mixing.
Several researchers have conducted laboratory cyclic
loading tests on cement-treated soils (Terashi et al., 1983;
Nakajima et al., 1984; Sharma and Fahey, 2003; Viana

http://dx.doi.org/10.1016/j.sandf.2017.05.008
0038-0806/Ó 2017 Production and hosting by Elsevier B.V. on behalf of The Japanese Geotechnical Society.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422
da Fonseca et al., 2013). In the cyclic loading test results,
the deterioration of the material strength with cyclic load-
ing is usually represented as a plot of applied stress against
the number of cycles to failure, which is known as an S–N
curve. Terashi et al. (1983) conducted cyclic unconfined
compression tests for cement-treated soil specimens pre-
pared in a laboratory and provided the S–N curve based
on the experimental results. The S–N curve shows a linear
relationship between the applied stress and the logarithm
of the number of cycles to failure. In the test results, the
data varies widely and a few specimens do not fail in
2
105 cycles which was set as the number of cycles to run-
out. In their study, the data on the non-failure specimen
was not considered when evaluating the S–N curve.
Nakajima et al. (1984) also conducted cyclic unconfined
compressive tests for cement-treated soil specimens pre-
pared in a laboratory. As in the results provided by
Terashi et al. (1983), the experimental results provided by
Nakajima et al. (1984) show that there is a large variation
in the relationship between the applied stress and the num-
ber of cycles to failure.
Previous experimental evidence suggests that the data
obtained from cyclic loading tests for cement-treated soils
varies widely even though the specimens are prepared in
a laboratory. Such variability in the cyclic loading test
results might be mainly induced by the variability of the
unconfined compressive strength qu of the specimens
because the qu of the specimens prepared using the same
laboratory procedure varies to some extent. Kitazume
et al. (2015) examined the influence of applying different
molding techniques to the laboratory preparation of spec-
imens on the strength of cement-treated soils. They showed
that the coefficient of the variation values for qu of the spec-
imens prepared by the tamping method lies in the range of
0.05–0.18. This indicates that it is difficult to carry out cyc-
lic loading tests with specimens having exactly the same
strength. However, previous studies have never assessed
the influence of the variation in qu on the variation in the
cyclic loading test results. An assessment of that influence
is required to analyze the factors inducing the variation
in the cyclic loading test results.
When the qu of specimens varies in the cyclic loading
tests, the specimens are likely not to fail under the cyclic
stress that is determined from the mean of qu. This indi-
cates that the variation in qu yields non-failure specimen
data which was not considered when evaluating the S–N
curve in previous studies. This implies that the S–N curve
is likely to have been evaluated with bias. In particular,
when the number of cycles to runout is not large, many
non-failure specimens are yielded in the cyclic loading tests.
Japanese Geotechnical Standard JGS2562 (JGS, 2012) rec-
ommends that the number of cyclic loadings to runout be
200 in order to determine the fatigue properties of rocks
in cyclic triaxial compression tests. Although the recom-
mended cyclic number is not large, that standard does
not describe how to treat the data on runout specimens
when determining the fatigue properties.
413
This study presents a statistical analysis for evaluating
the cyclic strength qcyc of cement-treated soils. The cyclic
stress is defined here as the axial compression stress applied
cyclically to a specimen under an unconfined condition.
qcyc is defined as the maximum value of cyclic axial com-
pression stress rmax causing failure at a specific number
of cycles in the cyclic unconfined compression tests. In
other words, qcyc represents the cyclic unconfined compres-
sive strength for a specific number of cyclic loadings. qcyc is
likely to become smaller than qu because cyclic loading
degrades the strength. Strength ratio Rs, defined as the
ratio of the mean of qcyc to that of qu, is used as the dete-
rioration index in this study.
In the binary regression approach, the experimental
results are treated as having one of two possible forms,
‘failure’ or ‘non-failure’, at prescribed cycles to take into
account the non-failure specimen data. In the analysis,
rmax is treated as an explanatory variable, and the param-
eters are calculated by the maximum likelihood method.
The parameters in the binary regression model are theoret-
ically related to the variations in qcyc and Rs. Moreover, the
error of the estimator obtained from the binary regression
analysis is evaluated on the basis of the asymptotic normal-
ity of the maximum likelihood estimator, and the uncer-
tainty involved in the estimated parameters is examined.
In order to demonstrate the benefits of the proposed
method, the binary regression approach is adopted for ana-
lyzing the results of cyclic unconfined compression tests.
The analysis results give Rs for denoting the deterioration
of strength due to cyclic loading and the interpretation of
the factors on the variability of qcyc. The S–N curve is also
evaluated as the sample regression line using only the fail-
ure specimen data obtained from the tests. Comparing the
sample regression line to the binary regression analysis
results, the reliability of the S–N curve obtained from the
cyclic loading test results is examined without the non-
failure specimen data.
This study suggests that the non-failure specimen data
can be utilized to evaluate the cyclic strength by adopting
the binary regression analysis. Using the binary regression
analysis proposed in this study, an unnecessarily large
number of cycles to failure can be avoided in the cyclic
loading tests and the experimental cost to obtain the dete-
rioration property of the materials can be drastically
reduced.
2. Binary regression approach
2.1. Binary regression analysis
Using the binary regression approach (Nawata 1992),
the experimental data can be treated as binary data and
can take into account the results of the non-failure speci-
men data yielded in the cyclic loading tests. The binary
regression analysis method for evaluating qcyc at a specific
number of cycles is described in this section. The ratio of
maximum axial compression stress rmaxi, applied in cyclic
414 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422
unconfined compression tests, to the mean of unconfined
compressive strength lqu, obtained from monotonic load-
ing tests, is assumed to be an explanatory variable for fail-
ure Xi as
X i ¼ rmaxi =lqu
The cyclic loading test results are treated as having one
of two possible forms, ‘failure’ or ‘non-failure’, at pre-
scribed cycles. These two possible values are represented
by a random variable Yi. A ‘failure’ is defined as Yi = 1,
while a ‘non-failure’ is defined as Yi = 0. Factor Y*i , deter-
mining the value of Yi, is defined as
Y i ¼ b0 þ b1 X i  ei
where b0 and b1 are the parameters, and eI is an error term.
Yi is defined as
Y i ¼ 1; if Y i =0
Y i ¼ 0; if Y i < 0
Eqs. (2) and (3) show that if ei 5 b0 + b1Xi, Yi = 1.
Therefore, the conditional probability of Yi = 1 given Xi
is expressed as
PðY i ¼ 1jX i Þ ¼ Fðb0 þ b1 X i Þ
where F is the cumulative distribution function of ei. In this
study, a normal distribution is assumed for the probability
distribution function of ei; the probit model is adopted in
the binary regression analysis.
Parameters b0 and b1 are determined from the experi-
mental results using the maximum likelihood estimation.
Likelihood function L(b0, b1) is given as
Lðb0 ; b1 Þ ¼ PYi¼1 Fðb0 þ b1 X i ÞPYi¼0 f1  Fðb0 þ b1 X i Þg
The maximum likelihood estimators for b0 and b1 are
defined as the values giving the maximum value for L(b0,
b1). The log likelihood function log L(b0, b1) is maximized
in the estimating procedure.
Parameters b0 and b1 can be interpreted as meaningful
indicators describing the properties of qcyc. Those parame-
ters are associated with Xi = rmaxi/lqu. Eq. (2) is rewritten
as
Y i ¼ b0 ð1  b1 =b0  rmaxi =lqu Þ  ei
When the term in the round brackets on the right-hand
side of Eq. (6) becomes 0, Y*i depends only on ei. Therefore,
when b1/b0rmaxi/lqu becomes 1, the effect of cyclic loading
on the strength is eliminated. Considering that the maxi-
mum stress represents the cyclic strength at a specific num-
ber of cycles, b0/b1 is identified as being the ratio of mean
cyclic strength lqcyc to lqu. Here, strength ratio Rs is
defined as
Rs ¼ lqcyc =lqu ffi b0 =b1
Rs denotes the ratio of the cyclic loading strength to the
monotonic loading strength and can be used as the deteri-
oration index due to cyclic loading. Rs is b0/b1 and is
expressed as Rs. Substituting Eq. (7) into Eq. (6) yields
the following expression:
ð1Þ
Y i ¼ b0 ð1  1=Rs  rmaxi =lqu Þ  e1
¼ b0 ð1  rmaxi =lqcyc Þ  eI ð8Þ
When rmaxi is replaced with qcyc, denoting the random
variable with mean value lqcyc, the probability of failure
becomes 50% at the prescribed cycle, that is, Y*i = 0. Then,
Eq. (8) is written as
ð2Þ ei ¼ b0 ð1  qcyc =lqcyc Þ ¼ ðqcyc  lqcyc Þ=ðlqcyc =b0 Þ ð9Þ
Eq. (9) can be interpreted as the equation in which qcyc is
normalized by the mean and the standard deviation. There-
fore, lqcyc/b0 is identified as the standard deviation of qcyc,
and 1/b0 is identified as the coefficient of the variation in
ð3Þ
cyclic strength Vqcyc. This indicates that 1/b0, evaluated
from the cyclic loading test results, is interpreted as Vqcyc.
Comparing 1/b0 to the coefficient of the variation in uncon-
fined compressive strength Vqu, the factor inducing the
variability of qcyc, obtained from the cyclic loading tests,
can be examined.
ð4Þ
2.2. Errors in estimated parameters
The maximum likelihood estimators for b0 and b1
involve errors. The errors of the estimators could be evalu-
ated on the basis of the asymptotic normality of the maxi-
mum likelihood estimator (Amemiya, 1985). The
maximum likelihood estimators for b0 and b1 are expressed
as the normal distribution with the variance-covariance
matrix Cov(bi, bj) defined as 1/A(b0, b1)ns. Here, A(b0,
ð5Þ b1) is defined as the mean of the second order differential
matrix of log L(b0, b1) and ns is the number of data. For
the binary data, if ns is sufficiently large, estimated matrix
b 0 ; b1 Þ is given as (Amemiya, 1985)
Aðb
bf 2  
b 1X ns
1 Xi
Aðb0 ; b1 Þ ffi i
ð10Þ
b i ð1  F
ns i¼1 F b i Þ X i X 2i
b i ¼ Fðb0 þ b1 X i Þ; bf i ¼ fðb þ b X i Þ
F
0 1
where f is the probability density function. The asymptotic
variance-covariance matrix Cov(bi, bj) of the maximum
ð6Þ
likelihood estimators for b0 and b1 can be calculated from
b ; b Þ. The errors of the estimators for b0 and b1 are
Aðb
0 1
quantitatively evaluated from Cov(bi, bj).
3. Example application
3.1. Cyclic loading tests
In order to demonstrate the benefits of the proposed
method, the binary regression approach is adopted for ana-
ð7Þ lyzing the results of cyclic unconfined compression tests.
 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422 415

The authors conducted cyclic unconfined compression tests

on cement-treated sand specimens prepared in a labora-
tory. In that study, unconfined compression tests were also
conducted on specimens to which a monotonic loading was
applied in order to evaluate the mean and the variability in
qu of the specimens. The experimental program and the
results are described herein.
The composition of the cement-treated sands used in the
tests is described in Table 1. This table shows the weight
ratio of the materials composing the specimens. The speci-
mens were prepared with three different cement contents
and consisted mainly of Toyoura sand, cement, and water.
The cyclic strength of the cement-treated sands is examined
in this study. A small amount of Kaolinite clay was added
to prevent the separating of the cement-water paste from
the aggregate. The specimens were prepared by following
JGS0812 (JGS, 2009). All the materials shown in Table 1
were mixed for 10 min. Thereafter, the mixed cement-
treated soil surrey was put in a cylindrical mold and the
air in the surrey was removed by vibrating compaction.
The curing times were 7 days for Case 1 and Case 2, and
28 days for Case 3. A cylindrical specimen, 50 mm in diam-
eter and 100 mm in height, was used in the tests. The load Fig. 1. Cyclic loading test appratus.
in the vertical direction was measured by an inner load cell
that was free from the effects of friction at the bearing. The
Table 2
displacement of the loading shaft was measured by a dis- Summary of unconfined compressive strength qu.
placement transducer. The apparatus used for the cyclic
Case Curing time (days) ns mqu (kN/m2) squ (kN/m2) Vqu
compression tests is shown in Fig. 1.
1 7 50 587 42.9 0.0731
A summary of the unconfined compression test results is
2 7 30 823 53.9 0.0655
shown in Table 2. The loading rate in the unconfined com- 3 28 50 687 56.3 0.0819
pression tests for Case 1 and Case 2 is 1 mm/min. That for
Note: ns, number of specimens; mqu, mean of qu; squ, standard deviation of
Case 3 is 0.2 mm/min owing to the limitation of the testing qu; Vqu, coefficient of variation in qu.
apparatus used in that test. Although it is known that the
loading rate affects the strength of cement-treated soil
(Onimaru et al., 2012), the five-times-greater loading rate from these tests lie in the range of those obtained by
does not largely affect the value of qu. In this study, the dif- Kitazume et al. (2015).
ference in the loading rate is ignored when analyzing the In the cyclic loading tests, a cyclic axial stress is applied
test results. Table 2 shows the number of specimens, ns, to a specimen under an unconfined condition. The axial
the sample mean of qu, mqu, and the standard deviation cyclic load is applied by an air-activated Bellofram piston.
of qu, squ. The coefficient of variation, Vqu = squ/mqu, is A schematic picture of the applied stress is shown in Fig. 2.
also shown in this table. In the unconfined compression Mean axial stress rm, which is the mean between the max-
tests, a large number of specimens were tested for each case imum and the minimum stress, is applied to the specimen
to obtain accurate values for mqu and squ. Some variability
in qu can be seen between the specimens prepared in the
same manner in the laboratory. The values for Vqu are
First peak stress
0.0731, 0.0655, and 0.0819 for Case 1, Case 2, and Case
3, respectively. It seems that the variation in qu increases
as the cement content decreases. The Vqu values obtained
σmax
Table 1
Mixing properties of cement-treated soil. σm
Case Water Cement Sand Kaolin
1 18.6 5.0 70.9 5.5
2 18.6 6.0 69.9 5.5
3 18.7 4.5 71.2 5.6
Note: weight (%) Fig. 2. Schematic picture of cyclic load.
416 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422

before cyclic loading. Thereafter, two-way cyclic loading is

applied to the specimen. A sinusoidal excitation of 0.4 Hz
is adopted for the cyclic loading. This value of the cyclic
loading frequency lies in the range of 0.1–1.0 Hz, which
is prescribed in JGS2562 (JGS, 2012). It should be noted
that constant-amplitude uniform cyclic loading is adopted
in this study, while an irregular seismic loading is applied
for ground improvement during an earthquake. From a
practical point of view, some coefficients correcting the cyc-
lic strength under uniform cyclic loading should be intro-
duced to evaluate the cyclic strength under irregular
seismic loadings. The correcting coefficients for the cyclic
strength under irregular seismic loadings is a topic for
future research. Moreover, multi-directional stresses are
applied for ground improvement under in-situ three-
dimensional stress conditions. Further research is also
required on the influence of such a complicated loading
environment in the seismic design of ground improvement Fig. 3. Cyclic unconfined compressive test results at rmax/mqu = 0.731 for
works. Case 2: (a) Axial strain time history and (b) stress-strain relationship.
The values for rmax are set on the basis of the mqu
obtained from the unconfined compression test results.
Considering the cyclic loading induced by earthquakes,
the tests are terminated at 1000 cycles; this is because the
number of loading cycles to runout is 1000. The number
of loading cycles to runout in this study is more than
200, which is the prescribed number of cyclic loadings at
the end of a test in JGS2562 (JGS, 2012) targeting the cyc-
lic loading induced by earthquakes. However, the number
of loading cycles to runout in this study is smaller than that
in previous studies (e.g., Terashi et al., 1983). A smaller
number of cycles to runout is beneficial in that the test cycle
can be shortened and the experimental cost can be reduced.
Due to the smaller number of cycles, however, many spec-
imens may not fail during the cyclic loading tests. There-
fore, in this study, the binary linear regression analysis is
adopted to consider the non-failure data in evaluating qcyc.
It should be noted that the number of loading cycles to
Fig. 4. Cyclic unconfined compressive test results at rmax/mqu = 0.898 for
runout must be determined from the targeted phenomena. Case 2: (a) Axial strain time history and (b) stress-strain relationship.
For example, a larger number of cycles to runout is
required to investigate the deterioration of strength with
the cyclic loading caused by the transportation movement. erence. When the specimen fails at the first cycle in the cyc-
Examples of the time history of the axial strain and the lic loading tests, the failure is likely to occur around the
stress-strain relationship during cyclic loading are shown in first peak stress state shown in Fig. 2. Such a loading con-
Figs. 3 and 4. The results are obtained from the tests for dition is considered as the monotonic loading condition.
Case 2. Fig. 3 shows the results for a non-failure specimen, Thus, the results obtained from the unconfined compres-
while Fig. 4 shows the results for a failure specimen. Cyclic sion tests, regarded as monotonic loading tests, are plotted
ratcheting and the following shake-down phenomena are at 1 cycle in Fig. 5. Here, the qcyc for nf is defined as the
observed during the initial loading cycles in both test rmax value for nf. In all cases, it seems that qcyc decreases
results. The ratcheting behavior implies that the cyclic with the logarithm of nf and that there is a large scatter
loading induces the accumulation of strain. Thereafter, in the test data. Although the relationship between qcyc
while the strain extension seems to decrease in the non- and nf might be approximately calculated from the failure
failure specimen results, the strain extension increases at specimen data, many non-failure specimens were yielded
around 40 cycles in the failure specimen results. in these tests. Therefore, an elaborate analysis, in which
The plot for the ratio of rmax to mqu against the loga- the non-failure specimen data is considered and the factor
rithm of the number of cycles to failure nf is shown in of the data variability is examined, is required to clarify the
Fig. 5. The non-failure data is plotted at 1000 cycles. The deterioration of the strength of the cement-treated soils
unconfined compression test results are also plotted for ref- subjected to cyclic loadings.
 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422
Fig. 5. Ratio of rmax to mqu versus logarithm of number of cycles to
failure nf: (a) Case 1, (b) Case 2, and (c) Case 3.
3.2. Analysis results
A binary regression analysis of the cyclic loading test
results was conducted. In the analysis, mqu is adopted for
lqu in Eq. (1). mqu is a random variable with a probability
distribution function. When qu follows the normal distribu-
tion, the standard deviation in mqu, denoting the variability
pffiffiffiffi
in the sample mean, is given by squ = ns . The values for the
standard deviation of mqu are 6.07 for Case 1, 9.84 for Case
2, and 7.96 for Case 3. These values are around 1% of mqu,
indicating that the error in substituting mqu for lqu is small.
An example of log likelihood function log L(b0, b1) at
1000 cycles for Case 1 is shown in Fig. 6(a) and (b). In these
figures, the values for b0 and b1, giving the maximum value
for log L(b0, b1), are approximately identified as the opti-
417
β1 = 15.8
(a)
β0 = 12.7
(b)
Fig. 6. Log likelihood function log L(b0, b1) at 1000 cycles for Case 1: (a)
b0 versus log L(b0, b1) and (b) b1 versus log L(b0, b1).
mum values for b0 and b1. Accurate optimum values for
b0 and b1 are calculated using the statistical computing
software R. The calculated b0 and b1 values at each data
point in Case 1 are shown in Fig. 7. The fluctuations in
the values for b0 and b1 are due to a scatter in the test data
in the S–N plot shown in Fig. 5. In particular, the variation
in those values becomes large under 30 cycles because there
are few test results at low cycles. The values for b0/b1 and
1/b0, denoting Rs and Vqcyc, are also given in this figure. It
can be seen that b0/b1 decreases with nf and its inclination
changes from point to point. 1/b0 changes irregularly with
nf.
The fluctuation in the values for b0 and b1 suggests that
it is not worthwhile to examine the binary regression anal-
ysis results at each data point. Therefore, the values for b0
and b1 are examined at specified cycles in this study. The
number of cycles ncyc, at which the test results are treated
as having one of two possible forms, ‘failure’ or ‘non-
failure’, are selected to be 100 and 1000. It has been cus-
tomary to consider 10 or 20 cycles in view of the typical
number of significant cycles during an earthquake
(Ishihara, 1996). Considering the fact that ground improve-
ment works are subjected to several earthquakes during
their life period, ncyc = 100 was selected as the representing
cycle. Moreover, ncyc = 1000 was selected as the represent-
ing cycle since small earthquakes occur several times a year.
In the analysis, mqu, shown in Table 2, was used as lqu in
Eq. (1).
The calculated optimum values for b0 and b1 are shown
in Table 3. The values for b0/b1 shown in this table
418 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422

140 cyclic loading induced by earthquakes, are also shown in

βo
120 Table 3. In Case 2, the values for b0 and b1 at 200 cycles
β1
100 are identical to those at 1000 cycles because there is no data
80
between 200 and 1000 cycles. In the other cases, the value
β0 , β1

for b0/b1 at 200 cycles is approximately the same as that at

60
100 cycles.
40
The value for 1/b0, which corresponds to the value for
20 Vqcyc, is also calculated from the value for b0. The values
0 for 1/b0 agree reasonably well with Vqu, as shown in
(a) 1 10 100 1000
Table 2, indicating that the variability in qcyc approxi-
The number of cycles nf
mately corresponds to that of qu. Table 3 shows the deter-
1.10 ministic values for b0/b1 (=Rs) and 1/b0 (=Vqcyc), and those
values include errors. The errors in the estimated values
1.00 will be evaluated quantitatively in the following section.
The cumulative distribution function of the standard
0.90 normal distribution is shown in Fig. 8. Cumulative distri-
β 0 /β 1

bution function F(b0 + b1Xi) represents the probability

0.80 of failure P(Yi = 1|Xi). The experimental results at 1000
cycles in Case 1 are superimposed in this figure. The value
0.70 for b0 + b1Xi for each data point is calculated using the
(b) 1 10 100 1000 estimated values for b0 and b1 shown in Table 3. The fail-
The number of cycles nf ure specimen data is located on the line of P(Yi = 1|Xi) = 1,
while the non-failure specimen data is located on the line of
0.10
P(Yi = 1|Xi) = 0. It can be seen in the experimental results
0.08 that the amount of failure data increases as the value of
b0 + b1Xi increases. Fig. 8 illustrates the fit of the probit
0.06
model assumed in the binary regression analysis to the
䠍 /β 0

0.04 experimental results.

0.02

0.00
(c) 1 10 100 1000
The number of cycles nf

Fig. 7. Binary regression analysis results for Case 1: (a) b0, b1, (b) b0/b1,
and (c) 1/b0,

represent Rs. The values for b0/b1 indicate that the values
for Rs lie in the range of 86–91% at 100 cycles and in the
range of 80–83% at 1000 cycles. The values for b0 and b1
at 200 cycles, the prescribed number of cyclic loadings at Fig. 8. Cumulative distribution function and test results at 1000 cycles for
the end of a test in JGS2562 (JGS, 2012) targeting the Case 1.

Table 3
Estimator of parameters.
Case ns ncyc b0 b1 b0/b1 1/b0
1 58 100 23.1 26.8 0.864 0.0432
200 21.6 25.5 0.848 0.0462
1000 12.7 15.8 0.806 0.0786
2 27 100 12.6 14.3 0.884 0.0791
200 12.2 14.7 0.828 0.0823
1000 12.2 14.7 0.828 0.0823
3 52 100 16.1 17.7 0.910 0.0620
200 9.18 10.3 0.895 0.109
1000 6.37 7.95 0.801 0.157
Note: ns, number of specimens; ncyc, number of loading cycles.
 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422 419

3.3. Errors in estimated parameters

The Cov(bi, bj) of the maximum likelihood estimators

for b0 and b1 were calculated from Aðb b 0 ; b1 Þ defined in
Eq. (10). The calculated values for Cov(bi, bj) are shown
in Table 4. The standard errors in estimators r(b0) and r
(b1), calculated from the square roots of Cov(b0, b0) and Vqu
Cov (b1, b1), are also shown in Table 4. It can be seen that
the standard errors of the maximum likelihood estimators
for b0 and b1 are not small. In particular, for Case 2, the (a)
ratio of the standard error to the estimators (see Table 3)
is large, indicating that the confidential interval of the esti-
mator becomes large. The coefficient of variation 1/b0,
denoting Vqcyc, and its 90% confidence interval are shown
in Fig. 9. For Case 1 and Case 3, the maximum likelihood
estimators for b0 and the standard errors at 30 and 300
cycles are additionally calculated, and the 1/b0 values cal-
culated from those results are also plotted in Fig. 9. While
the estimated confidence intervals are not significant for Vqu
Case 1, the estimated confidence intervals for Case 2 are
relatively large. In Eq. (10), the variance of the estimators
is inversely proportion to ns. For Case 2, the smaller num-
(b)
ber of specimens leads to the larger confidence interval of
the estimator. Vqu is superimposed in Fig. 9. In almost
all cases, Vqu lies in the range of the 90% confidence inter-
vals of 1/b0 denoting Vqcyc. This consistency implies that
the variability of qcyc is mainly attributed to the variability
of qu.
The standard error of b0/b1, representing Rs, is evalu-
ated from the estimators for b0 and b1. The estimators
Vqu
for b0 and b1 follow the bivariate normal distribution with
Cov(bi, bj). In this study, the distribution of b0/b1 was cal-
culated by a Monte Carlo simulation; the realizations of b0
and b1 were sampled from the bivariate normal distribu- (c)
tion using the statistical computing software R. The num-
ber of realizations is 1000. The mean and the standard Fig. 9. Estimated value of 1/b0 and its 90% confidence interval: (a) Case 1,
deviation of b0/b1 for Case 1 and Case 3 are shown in (b) Case 2, and (c) Case 3.
Table 4. The values for b0/b1 and its 90% confidence inter-
vals are shown in Fig. 10. The values for b0/b1 at 30 and 1000 cycles, the analysis provides the value for Rs without
300 cycles are additionally calculated here. For Case 1, a considerable error at the other cycles. These results sug-
the 90% confidence intervals are sufficiently narrow, sug- gest that the reliability of Rs obtained in the binary regres-
gesting that the proposed binary regression analysis pro- sion analysis can be assessed quantitatively on the basis of
vides the value for Rs with a small error. For Case 3, the asymptotic normality of the maximum likelihood
although the 90% confidence interval becomes wide at estimators.

Table 4
Variance-covariance matrix of estimators.
Case ncyc Cov(b0, b0) Cov(b1, b1) Cov(b0, b1) r(b0) r(b1) r(b0/b1)
1 100 25.7 33.0 -29.1 5.79 6.77 0.0106
1000 17.8 24.4 -20.8 3.40 4.12 0.0170
2 100 33.5 45.8 -39.2 5.07 5.75 –
1000 11.5 17.0 -14.0 4.22 4.94 –
3 100 18.4 22.3 -20.2 4.29 4.72 0.0177
1000 4.13 5.56 -4.76 2.03 2.36 0.0455
Note: ncyc, number of loading cycles.
420 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422

r2 = 0.519

(a)

r2 = 0.0702

Fig. 10. Estimated b0/b1 and its 90% confidence interval: (a) Case 1 and (b)
(b) Case 3.

It should be noted that b0/b1 was not calculated for Case

2 because the realization of b0 involves negative values.
Negative values are generated for b0 owing to the large
standard error induced from an insufficient number of
specimens for Case 2. These results indicate that the num-
ber of specimens needs to be more than about 50 in order
to evaluate the error of Rs.

r2 = 0.551
4. Discussions
(c)
4.1. Assessment of S–N curve
Fig. 11. Sample regression line: (a) Case 1, (b) Case 2, and (c) Case 3.
As mentioned in the Introduction, the deterioration of
the material strength with cyclic loading is usually repre-
sented as a plot of stress versus the number of cycles to fail-
sion line. The numbers of the failure specimens nsf used to
ure nf, which is known as an S–N curve. Here, the S – N
calculate the sample regression line are 37, 16, and 37 for
curve is evaluated from the experimental results presented
Case 1, Case 2, and Case 3, respectively.
in the previous section. The S–N curve is obtained as the
The sample regression line calculated by the least
sample regression line applied to the experimental relation-
squares method is shown in Fig. 11. Sample regression
ship between the ratio of rmax to mqu and the logarithm of
coefficients ba 0 and ba 1 are shown in Table 5. The coefficient
nf. The term rmax represents qcyc for nf. The following equa-
of determination r2, which indicates how well the data fits a
tion is adopted as the sample regression line:
regression line, is also shown in this table. In all cases, the
qcyc =lqu ffi rmax =mqu ¼ a0 þ a1 log nf ð11Þ values for r2 are far smaller than 1.0, indicating that the
sample regression lines are not likely to approximate the
in which a0 and a1 are the regression coefficients. In this data very well.
study, the sample regression line is evaluated from only The values for Rs (=lqcyc/lqu), evaluated by the binary
the failure specimen data; that is, the non-failure specimen regression analysis, are superimposed in Fig. 11. For Case
data is excluded from the calculation of the sample regres- 1, although there is a difference between the sample
 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422
Table 5
Estimated coefficients of sample regression line.
Case nsf b
a0 b
a1
1 37 0.970 0.0557
2 16 0.926 0.0214
3 37 1.076 0.0909
Note: nsf, number of specimens; r2, coefficient of determination.
regression line and the binary regression analysis results at
30 cycles, the values for Rs evaluated by the binary regres-
sion analysis lie approximately on the sample regression
line in the range of 100–1000 cycles. In this range, the bin-
ary regression analysis results support the validity of the
sample regression line, indicating that the sample regres-
sion line without the non-failure specimen data is likely
to provide the proper value for Rs.
For Case 2, the binary regression analysis results lie far
from the sample regression line at 1000 cycles. In this case,
no failure specimen data is found in the range of 200–1000
cycles, resulting in the difference between the two analysis
results at 1000 cycles. Therefore, the reliability of the sam-
ple regression line for Case 2 is suspiciously low in the
range of 200–1000 cycles.
For Case 3, there is some difference between the binary
regression analysis results and the sample regression line at
300 cycles. The large variation in the experimental results
around 300 cycles is likely to cause that difference. At the
other cycles, the binary regression analysis results lie
approximately on the sample regression line.
These comparisons suggest that the binary regression
analysis results can be utilized to assess the influence of
excluding the non-failure specimen data in the sample
regression line. When non-failure specimens are generated
at a runout cycle in cyclic loading tests, a binary regression
analysis should be conducted to assess the reliability of the
sample regression line.
4.2. Cyclic strength in design procedure
On the basis of the analysis results of the cyclic loading
tests, the cyclic strength in practical designs is discussed
here. The current design practices for ground improvement
by cement mixing are based on unconfined compressive
strength criteria. It is well known that the strength of
cement-treated soil constructed in practical projects varies
widely. Such variability in strength cannot be ignored in
practical designs. In some design procedures, specified
strength Fc is determined by considering the variability in
strength in a cement-treated ground (CDIT, 2002; Futaki
and Tamura, 2002). In those design procedures, Fc is eval-
uated as
F c ¼ ð1  K  V qu Þlqu
where K is the reduction coefficient. Following Eq. (12), the
specified strength for cyclic loading Fcyc is defined as
F cyc ¼ ð1  K  V qcyc Þlqcyc
421
As mentioned above, the statistical analysis results indi-
cate that Vqcyc corresponds to Vqu. This suggests that Eq.
r2 (13) can be rewritten as
0.519
0.070 F cyc ¼ ð1  K  V qu Þlqcyc ¼ Rs ð1  K  V qu Þlqu ¼ Rs  F c
0.551
ð14Þ
Eq. (14) shows that, when Rs is known, Fcyc can be easily
determined from Fc. This study shows that the values for
Rs can be evaluated from the cyclic loading test results
using a binary regression analysis. Using Eq. (14), the val-
ues obtained for Rs can be easily adopted into the frame-
work of the current design procedure in which the
variability in strength in the cement-treated ground is taken
into account.
5. Conclusions
A statistical analysis approach has been presented for
evaluating the qcyc of cement-treated soil specimens with
strength variability. A binary regression analysis was
adopted to treat the data of non-failure specimens. In this
study, the parameters in the binary regression analysis were
related to Vqcyc and Rs defined as the ratio of lqcyc to lqu.
To demonstrate the applicability of the binary regression
analysis, the results of cyclic unconfined compression tests,
conducted with cement-treated soil specimens prepared in
the laboratory, were analyzed as binary data having either
a ‘failure’ or a ‘non-failure’ form, at prescribed loading
cycles.
The binary regression analysis results revealed that Vqcyc
approximately corresponds to Vqu. This indicates that the
variability of the specimen strength mainly caused the vari-
ability appearing in the cyclic loading test results. On the
basis of this consistency, a design procedure has been pro-
posed in the present study for determining the specified
cyclic strength with variability. The value for Rs was also
evaluated from the test results involving the non-failure
specimen data using the binary regression analysis. More-
over, the estimated errors in the parameters were calculated
on the basis of the asymptotic normality of the maximum
likelihood estimator. It was demonstrated that the reliabil-
ity of Vqcyc and Rs can be assessed quantitatively by calcu-
lating the asymptotic variance-covariance matrix of the
estimators.
The sample regression line, denoting an S–N curve, was
also calculated from the test results from which the non-
failure specimen data was excluded. The comparison
between the binary regression analysis results and the sam-
ple regression lines showed that the sample regression line
estimated from the experimental results without the non-
ð12Þ
failure specimen data was able to provide a reliable value
for Rs in some cases, even though the coefficient of
determination r2 was far smaller than 1.0. This examina-
tion suggested that a binary regression analysis should be
ð13Þ conducted when ascertaining the validity of the sample
422 T. Namikawa et al. / Soils and Foundations 57 (2017) 412–422
regression line from which the non-failure specimen data
has been excluded.
The binary regression analysis adopted in this study has
allowed us to properly set the number of cycles to runout
for a considered phenomenon regardless of whether the
specimen fails or not. Accordingly, the cyclic strength of
a material with variability can be obtained without the
need for a long-term test, involving a large number of cyclic
loadings to failure, and at a reduced experimental cost.
References
Amemiya, T., 1985. Advanced Econometrics. Harvard University Press,
Cambridge, Massachusetts.
Coastal Development Institute of Technology (CDIT), Japan, 2002. The
Deep Mixing Method; Principle, Design and Construction. Balkema,
Lisse, The Netherlands, pp. 93–107, Chapter 6.
Futaki, M., Tamura, M., 2002. The quality control in deep mixing method
for the building foundation ground in Japan. In: Kitazume, M.,
Terashi, M. (Eds.), Proceedings of Tokyo Workshop 2002 on Deep
Mixing. Port and Airport Research Institute, and Coastal Develop-
ment Institute of Technology, Tokyo, pp. 139–149.
Ishihara, K., 1996. Soil behaviour in earthquake geotechnics. Resistance
of Sand to Cyclic Loading. Claredon Press, Oxford, pp. 208–246,
Chapter 10.
Japanese Geotechnical Society (JGS), 2009. Japanese Geotechnical
Standard JGS0821-2009. Practice for making and curing stabilized
soil specimens without compaction.
Japanese Geotechnical Society (JGS), 2012. Japanese Geotechnical
Standard JGS2562-2012. Method for cyclic undrained triaxial com-
pression test to determine fatigue properties of rock.
Kitazume, M., Grisolia, M., Leder, E., Marzano, I.P., Correia, A.A.S.,

Oliveira, P.J.V., Ahnberg, H., Andersson, M., 2015. Applicability of
molding procedures in laboratory mix tests for quality control and
assurance of the deep mixing. Soils Found. 55 (4), 761–777.
Khosravi, M., Boulanger, R., Tamura, S., Wilson, D., Olgun, C., Wang,
Y., 2016. Dynamic centrifuge tests of soft clay reinforced by soil-
cement grids. J. Geotech. Geoenviron. Eng. 142 (7), 04016027.
Nakajima, H., Saito, S., Babasaki, R., 1984. Studies on deep mixing
method using cement hardening agent (No. 23) – fatigue life and creep
rapture of improved soil. Proc. 19th Annual Conference of Japanese
Geotechnical Society, pp. 1631–1634 (in Japanese).
Namikawa, T., Koseki, J., Suzuki, Y., 2007. Finite element analysis of
lattice-shaped ground improvement by cement-mixing for liquefaction
mitigation. Soils Found. 47 (3), 559–576.
Nawata, K., 1992. Analysis of Scientific Data (Basic Statistics 3).
University of Tokyo Press, pp. 231–250, Chapter 8, (in Japanese).
Onimaru, S., Koseki, K., Miyashita, Y., Mikami, T., Suzuki, Y., 2012.
Dynamic shear strength of improved soil with B-type blast-furnace
cement. J. Soc. Mater. Sci. Jpn. 61 (1), 64–67 (in Japanese).
Sharma, S.S., Fahey, M., 2003. Evaluation of cyclic shear strength of two
cemented calcareous soils. J. Geotech. Geoenviron. Eng., ASCE 129
(7), 608–618.
Terashi, M., Tanaka, H., Mitsumoto, T., Honma, S., Ohhashi, T., 1983.
Fundamental properties of lime and cement treated soils 3rd Report,
22(1). Port and Harbour Research Institute (in Japanese).
Viana da Fonseca, A., Rios, S., Amaral, M.F., Panico, F., 2013. Fatigue
cyclic tests on artificially cemented soil. Geotech. Test. J. 36 (2), 1–9.