Proposed priors for the inference of multiple step-stress

accelerated test under type-II censoring and Gamma

distribution

Fernando A. Moala

State University of Sao Paulo, Sao Paulo, Brazil

Abstract

This paper presents the multiple step-stress accelerated life test using type-II
censoring. Assuming that the lifetimes of test item follow the gamma distribution, the
maximum likelihood estimation and Bayesian approaches are used to estimate the
distribution parameters. In this paper, noninformative, empirical and informative priors
are investigated and compared by examining the summaries and frequentist coverage
probabilities of intervals. Bayesian estimators are obtained through a conjunction of
Markov Chain Monte Carlo (MCMC) method and numerical system and integration
techniques. Some numerical illustrations considering simulated and illustrative lifetime
data are presented to illustrate the proposed methodology.

Keywords: Bayesian, acclerated test, step-stess, cumulative exposure, type II censoring,

Fisher information, elicitation, Gamma distribution, MCMC.

1
1. INTRODUCTION

The industries are challenged to shortly launch new products keeping its reliability
in increasing levels. For this, it would be important to conduct tests in an expedited
manner in order to significantly reduce the time required to obtain the results of reliability
for the product and a lower cost of test.
The accelerated test is the most appropriate statistical method to obtain
information about the reliability of new products very fastly than would be possible if the
product was left to fail in normal use. In order to failure occurs earlier, the item to be
tested is placed in operation under stress levels that exceed normal conditions of use. One
way of applying stress is the step-stress accelerated test, denoteb by SSALT, proposed by
Nelson (1980) wherein the items are successively subjected to more than one stress level.
In this test, the stress level is changed to a new value after a prefixed time, and the
procedure is repeated.
The cumulative exposure model proposed by Sedyakin (1966) and generalized by
Bagdonavicius (1978) and Nelson (1980) is one of the most useful models in the analysis
of SSALT. In this model, it is assumed that the residual lifetime of one item depends only
on the current cumulative fraction and the applied stress level, with no memory of how
the stress was accumulated. Moreover, if the level of applied stress is maintained, the
surviving units will fail according to the cumulative distribution at the same stress level
that is currently being tested at, but starting at the previous accumulated stress level.
Several authors have studied this model as Nelson (1980 e 1990), Miller and
Nelson (1983), Alhadeed and Yang (2005), Tang (2003), Gouno, et al. (2003), Xiong and
Ji (2004). Balakrishnan, et al. (2007) and Balakrishnan and Xie (2007) analyzed the
simple SSALT with exponential lifetime distribution with type-II and Type-II hybird
censored data, respectively. Kateri and Balakrishnan (2008) discussed inferential methods
for SSALT under Weibull distributed lifetimes with type-II censoring. Balakrishnan, et al.
(2009) presented exact inference for the simple step-stress model from the exponential
distribution when there is time constraint on the duration of the experiment. Alkhalfan
(2012) analyzed the step-stress model when the lifetimes follow the gamma distribution
by considering different censoring schemes.
Most of paper found in the literature present the estimation of parameters of
SSALT through the maximum likelihood approach. On the other hand, a few papers have
been proposed by using Bayesian approaches to analyze data from SSALT. DeGroot and
Goel (1979) proposed one Bayesian criterion optimality for simple SSALT under the
framework of decision theory. Leu and Shen (2007) presented optimum plans for simple

2
SSALT from a Bayes view-point with the lifetime under exponential distribution. Liu
(2010) considered a SSALT for Weibull distributed lifetimes under Bayesian approach.
Lee and Pan (2011) presented a Bayesian inference model for simple SSALT using type-
II censoring and more recently, Ganguly, et al. (2014) provide a Bayesian analysis of
simple SSALT under Weibull lifetime distribution based on different censoring schemes.
Among the arguments in favour of Bayesian approach we can emphasize.the
possibility of incorporating prior information in the analysis and often require less sample
data.
In this paper we deal with a Bayesian analysis for estimation of parameters for a
multiple SSALT using type-II censoring and assuming that the lifetimes of test item
follow the gamma distribution. The maximum likelihood estimation and Bayesian
approaches are used to obtain the estimators for the distribution parameters.
The application of Bayesian inference in statistical analysis requires the
specification of a prior distribution for the model parameters.
Thus, we also intend to introduce prior distributions that better represent three
types of situations. The first, of subjective type, assuming that there is some kind of
information available from a specialist, the second with noninformative priors and the
third, type of empirical, obtained with the information from the current data.
In the Bayesian context is important to check whether the use of different
parameterization improves the estimation of the parameters of interest or leads to some
computational advantage, or both. In this paper we also present a reparameterization for
the parameters. The effects of the reparameterizaton in the Bayesian estimation are
considered.
Prior knowledge has not been frequently considered in Bayesian analyses due to
difficulty to represent the expert's opinion subjectively or arguing that, using a large
amount of data, prior knowledge tends to have little effect in the final inferences.
However, if there is small data set and/or large number of censored data the use of expert
opinion improves so much the estimation. In this paper we propose a simple approach to
elicit the expert's knowledge.
The rest of the paper is organized as follows. The SSALT under exposure
cumulative model is introduced in Section 2. In Section 3, the maximum likelihood
estimators under Type-II censored data with gamma distribution are obtained. In Section
4, we develop the Bayesian approach for parameter estimation based on the different prior
distributions proposed. In Section 5, a simulation study is performed for comparisons
between the MLE and the Bayesian estimators. An illustrative example is presented in
Section 6. Section 7 is the conclusion.

3
2. DESCRIPTION OF THE MODEL

We assume that the failure time data come from the 7-step-stress model with
ordered stress levels B"  B#  Þ Þ Þ  B7 (7 2) and let 7"  7#  Þ Þ Þ  77" be the
pre-fixed times to change stress used in the SSALT.
According to the cumulative exposure model defined by Nelson (1990), the
cumulative distribution function for the test is given by:

Ú
Ý
Ý
Ý
J" (>) 0  > Ÿ 7"
Ý ã ã
K(>) œ Û J3 (>  73"  &3" )
Ý
73"  > Ÿ 73 , (1)
Ý
Ý
Ý ã ã
Ü J7 (>  77"  &7" ) 77"  >

where J3 (>) is the cumulative distribution function (cdf) of the failure time at stress B3 , 73
is the time to change stress, and the equivalent starting time, &3" , is the solution of the
equation
J3 Ð&3" ) œ J3" (73"  73#  &3# ) ß 3 œ 2ß Þ Þ Þß 7Þ (2)

The equivalent starting time, &3" , is the solution of the equation (2) given by

&3" œ Š73"  73#  &3# ‹ ))3"3 ß 3 œ 2ß Þ Þ Þß 7, (3)

with 70 œ &0 œ 0.
Initially 8 units are tested data lower stress level B" . The test is run until time 7" ,
when the stress level is increased to B# and the life test continues until a pre-specified <#
( Ÿ 8) number of failures are observed.
Suppose that the failure time X at each stress level B3 has a Gamma distribution
with common shape parameter ! and scale parameter )3 , with pdf and df given by.

" ' > " !1  )3

> u
" !1  )3
03 (>Ñ œ >(!))3! > / and J3 (>Ñ œ >(!) 0 )3! ? / .? , 3 œ "ß . . ., 7, (4)

respectively, and and 3 œ 0 denotes the normal operating condition. Thus, the cumulative
distribution function (cdf) of the lifetime X for the SSALT is given by

K(>) œ MKÐ>‡3 Ñ ß 73"  > Ÿ 73 (5)

4
and the corresponding probability density function (pdf) of X4 is given by

‡
1(>) œ "
>(!))3 >‡3 !1 />3 ß 73"  > Ÿ 73 (6)

 ! )44 with ?6 œ 76  76" , for 3 œ 1ß Þ Þ Þß 7 and 70 œ 0.

3"
>73" ?
where >‡3 œ )3
4œ"
" ' B !1 ?
MKÐBÑ œ >(!) 0 ? / .? is the incomplete gamma ratio.
We consider a reparametrization of the m-step-stress model, in which )3 is
assumed to satisfy a log-linear link function of the form

691 )3 œ "0  "" B3 , 3 œ 1ß Þ Þ Þß 7, (7)

where "0 and "" are unknown parameters and we need to develop inference only for these
two parameters instead of for the original m parameters )3 , 3 œ 1, 2, . . . 7. Therefore, the
failure time of this SSALT has the cumulative distribution function KÐ>Ñ given by
K(>) œ MKÐ'3 Ñ ß 73"  > Ÿ 73 (8)

3 œ 1ß Þ Þ Þß 7 and '3 œ Ð>  73" Ñ/Ð"0 "" B3 Ñ  ! ?4 /Ð"0 "" B3 Ñ

3"

4œ"

Once the model has been formulated then the parameters !, "0 and "" can be
estimated using the maximum likelihood and Bayesian approaches.

3. MAXIMAL LIKELIHOOD ESTIMATION

In the 7-step-stress model with Type-II censoring, we start with 8 identical units
placed simultaneously on the SSALT. Each unit will be subjected to an initial stress level
B" . After that, the experiment is run until a fixed time denoted by 7" , at which time the
stress level is changed to B2 , and the successive failure times are recorded. Then, at the
fixed time 7# , the stress is increased to B3 , and so on. This way, the stress level starts with
B" and changed to B2 , B3 , . . ., B7 at fixed times 7" , 7# , . . ., 77 , respectively. The
experiment is terminated when a fixed number of failures < are observed. Let 85 be the
number of units that fail between 75" and 75 at stress level B5 , for 5 œ ", 2, Þ Þ Þ, 7. The
observed censored sample is given by the ordered failure times denoted by

>"  Þ Þ Þ  >8"  7" Ÿ >8" "  Þ Þ Þ  >8" 82  7# Ÿ Þ Þ Þ  77 Ÿ >8"  Þ Þ Þ87" "  Þ Þ Þ  ><
(9)

5
 Thus, the likelihood function based on this data is a function of the unknown
parameters of !, "! and "1 given by

#Š # 15 Ð'3‡5 Ñ‹Š"  MKÐ'<‡7 Ñ‹

7 <5 8<7
8!
PÐ!, "! ß "1 l >Ñ œ Ð8<7 Ñ! , (10)
5œ" 35 œ<5" +1

where <o œ 0, <5 œ ! 83 ß <7 œ <, > is the vector of observed failure time data and
5

3œ"

'3‡5 œ Ð>35  75" Ñ/Ð"0 "" B5 Ñ  ! Ð74  74" Ñ/Ð"0 "" B4 Ñ ,

5"
(11)
4œ"

for 5 œ "ß 2, . . ., 7.
Let us assume, without loss of generality, that 74  74" œ 7 , that is, the 71 œ 7 ß
7# œ 27 , . . . ., 75" œ Ð5  "Ñ7 .
Since we are considering the cumulative exposure model then the likelihood
function of !, "! and "1 , based on the observed Type-II censored data, is given by

ˆ>Ð!щ<7 /B:š  ! 85 Ð"0  "" B5 Ñ›’ # Š # '3‡5 !" / '35 ‹“

7 7 <5
"
PÐ!,"! ß "1 l >Ñ º
5œ" 5œ" 35 œ<5" +1

‚ ’"  MKÐ'<‡7 Ñ“
8<
, (12)

It is convenient to work with the log-likelihood function rather than the likelihood
function in (12), which is given by

6Ð!, "! ß "1 l >Ñ º  <7 68Ð>Ð!ÑÑ  ! 85 Ð"0  "" B5 Ñ  Ð!  "Ñ! !

7 7 <5
68Ð'3‡5 Ñ 
5œ" 5œ" 35 œ<5" +1

! ! '3‡5  Ð8  <7 Ñ 68Š"  MKÐ'<‡7 Ñ‹.

7 <5
(13)
5œ" 35 œ<5" +1

Differentiating the log-likelihood function in (13) with respect to !, "! and "1 , we
obtain the following likelihood equations which need to be solved for finding the MLE of
!, "! and "1 .
The first partial derivatives are given by the following equations:

l >Ñ œ  <7 <Ð!Ñ  ! ! Š<Ð!ÑM KÐ'<‡7 Ñ  F" Ð'<‡7 Ñ‹ ,

7 <5
` Ð8<7 Ñ
`! 6Ð!,"! ß "1 68Ð'3‡5 Ñ  "MKÐ'<‡7 Ñ (14)
5œ" 35 œ<5" +1

l >Ñ œ  ! 85  Ð!  "Ñ! ! Ð  "Ñ  ! !

7 7 <5 7 <5 '‡ <
` Ð8<7 Ñ'<‡7 !" / 7
`"! 6Ð!,"! ß "1 Ð  '3‡5 Ñ 
5œ" 5œ" 35 œ<5" +1 5œ" 35 œ<5" +1 >Ð!ÑŠ"MK‡ Ð'<‡7 Ñ‹

and (15)

6
 l >Ñ œ ! 85 B5  Ð!  "Ñ! ! Š E" Ð>35 Ñ‹  ! ! Š  E" Ð>35 Ñ‹ 
7 7 <5 7 <5
` "
`"1 68 PÐ!,"! ß "1 '3‡
5œ" 5œ" 35 œ<5" +1 5 5œ" 35 œ<5" +1
‡
Ð8<7 Ñ'<‡7 !" / '<7
 E" Ð><7 Ñ (16)
>Ð!ÑŠ"MKÐ'<‡7 Ñ‹

>w Ð!Ñ " 'B !1 ?

where < Ð!Ñ œ >Ð!Ñ , F" ÐBÑ œ >(!) 0 68Ð?Ñ? / .? and E" Ð>35 Ñ œ Ð>35  75" ÑB5 /Ð"0 "" B5 Ñ 

 7 /"0 ! B4 /"" B4 Þ
5"

4œ"

The maximum likelihood estimators must be derived numerically because there is

no obvious simplication of the non-linear likelihood equations. Here, numerical
likelihood maximization was carried out on the log-likelihood using V software.
Estimating confidence intervals for !, "! and "1 can be obtained by asymptotic
normal approximation of MLE in large samples, that is

(! s! , "
s, " s 1 ) µ R$ ÐÐ!, "! , "1 Ñß M " (! s! , "
s, " s 1 )Ñ , 8 Ä _ß (17)

where M (!, "! , "1 ) is the observed Fisher information matrix given by M34 (!, "! ,
"1 ) œ  Ðf# 6(!, "! , "1 )Ñ. Elements of Fisher information matrix are given in Appendix
A.
Approximate confidence intervals for the individual parameters !, "! and "1 with
confidence coefficient 100(1  # )% are given by ! s„D ## ÈZ"" , " s ! „ D # ÈZ22 and
È
#
s "
" 1 „D # Z33 where V33 is the diagonal element of M (! s s
s, " ! , " 1 ) and D # indicate the ; -th
# #

upper percentile of the standard normal distribution.

4. BAYESIAN ESTIMATION

In a Bayesian framework, the inference is based on the information provided by

the posterior distribution of parameters, denoted as :Ð!, "! , "1 | >), which is obtained as
the product of the likelihood function (9) and the prior 1Ð!, "! , "1 Ñ resulting in:

:Ð!,"! ,"1 |>) º 1Ð!,"! ,"1 Ñ ˆ>Ð!"щ<7 /B:š  ! 85 Ð"0  "" B5 Ñ›’ # Š # ‹“ ‚
7 7 <5
 '3‡
'3‡5 !" / 5

5œ" 5œ" 35 œ<5" +1

‚ ’"  MKÐ'3‡5 Ñ“
8<7
. (18)

To progress with the Bayesian analysis it is necessary to specify a prior

distribution over the parameter space. Different prior distributions can be used in our
study according to all currently available information.

7
 Assuming prior independence among the parameters !, "0 and "1 , we consider
the prior distributions as follows:

! µ >(aß ,Ñ , "0 µ R Ð.! , 5!# Ñ and "1 µ R Ð." , 5"# Ñ , (19)

where the hyperparameters a, ,, .! , 5!# , ." and 51# are known. High values of 53# and
a œ b œ 0.01 or 0.001 suggest little prior information about the parameters "3 and !,
respectively.
When the inferential results are not very accurate, different parameterizations can
improve the estimation. A suitable reparameterization we can consider is given by:

<0 œ /"0 and <1 œ /"" . (20)

and the corresponding priors designed for <0 and <1 are given by

<0 µ >Ð-" ß ." Ñ and <1 µ >Ð-2 ß .2 Ñ , (21)

where the hyperparameters "-3 " and ".3 ", 3 œ "ß #, are known.
The lack of knowledge a priori about the parameters of interest may lead to less
accurate estimators mainly in the presence of large number of censored data.
When there is prior knowledge about the subject in study, a more sensible prior
distribution can be elicited for the parameters. There are numerous different methods
proposed in the literature (see O'Hagan, et al, 2006) for eliciting prior distributions. In this
paper, it is employed an elicitation process based on the predictive prior distribution. We
suppose that an engineer can summarize his knowledge about the reliability of an item
through statements of percentiles. Thus, we can, from the expression

T ÖX Ÿ >: × œ '! : 0 Ð> Ñ.> œ :,

>
(22)

for a given :, express the expert's information.

In order to facilitate the elicitation process, we will consider the prior distributions
under the reparameterization (20) and the set of prior distributions (21). Thus, we obtain
the hyperparameters values a, ,ß -" , ." ß -# and .# of the gamma priors designated to the
parameters !, <0 and <1 by solving the equation

8
 : œ '! '! '! KÐ>: ¸ !, <0 ß <1 Ñ1Ð!,<0 ß <1 ¸a,,ß -" ß -2 ß ." ß .2 ÑÑ. ! . "! . "1 ,
_ _ _
(23)

where KÐ>: ¸ !, <0 ß <1 Ñ is given by (6) and replacing "0 and "1 by 691Ð<0 Ñ and 691Ð<1 Ñ,
respectively.
From the prior distributions (21), the probability in (23) becomes

: œ - '! '! '! KÐ>: ¸ !, <0 ß <1 Ñ!a" <0-" " <1-# " /B:Ö  Ð !, 
_ _ _ <0 <"
."  .# Ñ×. ! . "! . "1 ,
(24)
"
where - œ Þ
>ÐaÑ>Ð-" Ñ>Ð-# Ñ,a ."-" .#-#
In practice, there exists a set of values for the hyperparameters which satisfies (24)
with any value of >: . Therefore, the method we propose requires that the engineer
provides six distinct percentiles that generate simultaneous equations. Given these
information, a nonlinear system with equation (24) is solved numerically to obtain the
values of the hyperparameter a, ,ß -" ß -2 ß ." and .2 .
In a similar way, we can use the elicited mean IÐX Ñ as

IÐX Ñ œ '! '! '! IÐX ¸ !, <0 ß <1 Ñ1Ð!,<0 ß <1 ¸a,,ß -" ß -2 ß ." ß .2 ÑÑ. ! . "! . "1 ,
_ _ _
(25)

where the mean IÐX ¸ !, )Ñ is given in the Appendix B.

We can then use feedback by reporting summaries from the elicited prior to see if
the it is an acceptable representation of the expert's belief.

5. SIMULATION STUDY

To evaluate the performance of the procedures of inference described in this

paper, a Monte Carlo simulation was conducted. The simulation study is carried out with
1000 samples generated from the step-stress model for different sample sizes (8) and
highly censored setting. We present several simulations results from different
sample sizes (8) and number of failure Ð<Ñ.
The values for parameters are chosen to be ! œ 2, "! œ 4 and "1 œ 2 and levels
of stress are B" œ 1, B2 œ 1.5 and B3 œ 2.5.
Firstly, we perform the Bayesian estimation of the parameters assuming lack of
information. Thus, consider the noninformative priors for the parameters !, "0 and "1
given in (16). As discussed in Section 4, we can pre-specify appropriate values for the
hyperparameters "a" and "," such that the gamma prior for ! is essentially

9
noninformative. We use equal low parameter values setting by this way its prior mean
equal to one and its prior variance large. For example, we may use a œ , œ 0.01 which
results I (!) œ 1 and @+<(!) œ 100.
We represent prior ignorance of the values of the parameter "0 by Normal prior.
The usual choice for the prior mean is the zero value and a large variance 52 to represent
high uncertainty, for instance 5 œ 100. For the parameter "1 we consider a Uniform prior
in the interval Ð0, _Ñ, indicating that all possible values for "1 are equally probable.
Therefore, the priors for the parameters !, "! and "1 are given by

! µ >Ð0.01ß 0.01Ñ, "0 µ R Ð0, 100# Ñ and "1 µ Y Ð0, 1000Ñ. (26)

and the joint prior distribution for Ð!, "! , "1 Ñ is obtained as the product of the three priors
above and will denote by 1Ð!, "! , "1 Ñ.
We also test the reparameterization (20) in this simulation. The parameters !, <0
and <1 have a positive support, hence the gamma distribution is an appropriated prior
distribution for them with hyperparameters 0.01, that is,

! µ >Ð0.01ß 0.01Ñ, <0 µ >Ð0.01ß 0.01Ñ and <1 µ >Ð0.01ß 0.01Ñ (27)

Let us denote the joint prior for !, <0 and <1 by 1(!, <0 , <1 ).
To try improve upon the estimation of the parameters we can also apply the
empirical Bayes in which the prior distribution is estimated from the data as proposed in
Section 4. The empirical prior was obtained from the MLE presented in the Tables 1 to 3.
Due to the marginal posterior distributions can not be expressed in closed-form
we apply the Markov Chain Monte Carlo (MCMC) algorithm to obtain the point
estimators and credible intervals for the parameters. The chain is run for 15000 iterations
with a burn-in period of 5000.
The estimates are compared in terms of their bias, mean-squared error (MSE) and
estimated coverage probabilities.
Tables 1-3 show the averages of the 1000 estimators and corresponding standard
deviations for the parameters !, "0 and "1 .
The first column gives the sample size (8) and number of failures (<). In column
MLE, the maximum likelihood estimators, in third column, the Bayes estimators are
obtained by using the prior 1(!, "0 , "1 ), fourth column gives the Bayes estimators under
reparameterization (20) with the prior 1(!, <0 , <1 ) and column Empirical gives the
Empirical Bayes estimators.

10
 70% censored data for α 50% censored data for α

1.2

0.8
0.8
bias

bias

0.4
0.4
0.0

0.0
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β0 50% censored data for β0

0.0 0.5 1.0 1.5 2.0

4 5 6
bias

bias
0 1 2 3

20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β1 50% censored data for β1

0.8
1.0 1.5
bias

bias

0.4
0.5
0.0

0.0

20 40 60 80 100 20 40 60 80 100

n n
MLE Bayes Transf

Figure 1: Bias from estimation of !, "! and "1 under noninformative priors when ! œ 2,
"! œ 4 and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5

11
 70% censored data for α 50% censored data for α

2.0

2.0
MSE

MSE
1.0

1.0
0.0

0.0
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β0 50% censored data for β0

10 20 30 40

6
MSE

MSE

4
2
0

20 40 60 80 100 0 20 40 60 80 100

n n

70% censored data for β1 50% censored data for β1

5

1.5
4

1.0
3
MSE

MSE
2

0.5
1

0.0
0

20 40 60 80 100 20 40 60 80 100

n n
MLE BAYES TRANSF

Figure 2: MSE from estimation of !, "! and "1 under noninformative priors when ! œ 2,
"! œ 4 and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5

12
 70% censored data for α 50% censored data for α

0.95

0.95
CP

CP
0.80

0.80
0.65

0.65
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β0 50% censored data for β0

0.95

0.95
CP

CP
0.80

0.80
0.65

0.65
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β1 50% censored data for β1

0.95

0.95
CP

CP
0.85

0.85

20 40 60 80 100 20 40 60 80 100

n n
MLE Bayes Transf 0.95

Figure 3: Coverage probability from estimation of !, "! and "1 under noninformative priors
when ! œ 2, "! œ 4 and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5

13
**************CONCLUSOES ABAIXO DAS TABELAS DA VERSAO 3**************

It is important to observe from Figures 1 and 2 that the point estimation and
coverage probabilities for the three parameters are worse when the proportion of
censoring increases or the sample size is small. On the other hand, if the sample size 8 is
large there is practically no difference, as expected.
Besides, it is clearly observed that "0 and "1 are underestimated by using
1(!,<0 ,<1 ) while ! is overestimated in oppositive to the others three approaches.
By analysing the point estimation of the parameters we observe that the MLE and
the Empirical Bayes approaches have practically the same performance while the Bayes
approach under reparameterization has slightly better. However, when compare the
estimators for the parameter "0 , the MLE, 1(!,"0 ,"1 ) and Empirical Bayes are quite
affected by the high number of censored data or small sample size while the 1(!,<0 ,<1 ) is
not and hence provides better estimation than the others approaches.
Therefore we can conclude that the application of reparameterization <0 and <1
improves upon the estimation compared with the others approaches for all pair of values
Ð8ß <Ñ.
Unfortunately, the prior 1Ð!,"! ,"1 Ñ performs poorly for the three parameters in
comparison with the others approaches, except when 8 tends to large.
Figure 3 shows the coverage probability of the 95% intervals for the parameters !,
"0 and "1 . The confidence intervals for the parameters based on MLE are obtained by the
asymptoticaly normal approximation.
Now, the comparison of the estimation approaches based on the criterion of
coverage probability we observe from Figure 3 that the estimated coverage probability
from MLE and Bayes approaches are quite similar and they are too low when the
proportion of censored data is large. Empirical Bayes takes to the highest coverage
probability close to 100% and the prior 1(!,<0 ,<1 ) takes to the posterior mean closer to the
corresponding nominal level.

*******************************************************************

14
 An alternative to nonnformative priors and empirical Bayes would be to define the
prior distribution based on the expert's information that is, elicited priors.
To elicit reasonable informative prior distribution based on expert's information,
we apply the elicitation approach as described in section 4.
Thus, to obtain the hyperparameters, the expert's judgment concerning quantities
of interest are elicited and equated to their theoretical expression such as mean and
percentiles. Solving these equations leads to the desired hyperparameter estimators (see
section 4).
This way, we ask six lifetimes >3 , 3 œ "ß Þ Þ Þß 6, such that T ÖX Ÿ >1 × œ 0.10,
T ÖX Ÿ >2 × œ 0.25, T ÖX Ÿ >3 × œ 0.50, T ÖX Ÿ >4 × œ 0.75, T ÖX Ÿ >5 × œ 0.90 and
T ÖX Ÿ >5 × œ 0.99. From the distribution function (6), (7) and parameter values ! œ 2,
"! œ 4 and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5 we have the elicited lifetimes
>" œ 3.93, >2 œ 5.77, >3 œ 7.72, >4 œ 8.49, >5 œ 8.93, >6 œ 9.94. These values provide
equation (20) and the hyperparameters are obtained by solving the system composed by
six equations which solution is a œ 19.76, , œ 0.1057, -" œ 3.2691, ." œ 15.9327,
-# œ 19.2345 and .# œ 0.3827.
By considering the prior distributions for the parameters !, <0 and <1 given in
(21), the elicited prior distributions for these hyperparameters are given by

! µ >Ð19.76, 0.1057Ñ, <0 µ >Ð3.2691, 15.9327Ñ and <1 µ >Ð19.2345, 0.3827Ñ. (28)

The performance of the elicited joint prior distribution is examined by comparing

it with three other prior distributions: empirical and two informative priors under
hyperparameters fixed a priori. The selection of the hyperparameters is made such that the
prior expectation of the parameters is the true values while the variances assume small
values in order to have informative prior distributions.
Authough the information contained in the empirical prior is not subjective since
it originates from observed data, we consider it in the analysis with the informative priors
in order to compare with the elicited prior.
We consider two set of informative (non-elicited) prior distributions given by

! µ >Ð4, 0.5Ñ, <0 µ >Ð7.453, 7.326Ñ and <1 µ >Ð2.184, 3.384Ñ, (29)
which joint prior distribution is denoted by 1" (!, <0 , <1 ) and

! µ >Ð44.44, 0.045Ñ, <0 µ >Ð116.64, 0.4630Ñ and <1 µ >Ð49, 0.14286Ñ,(30)

denoted by 1# (!, <0 , <1 )Þ

15
 Table 5 gives the prior expectances and variances for the parameters !, <0 and <1
and Table 6 the probabilities a priori by considering the three informative priors.

Table 5: Prior mean and variances

expectance and variances
I (!) @+<(!) I (<0 ) @+<(<0 ) I(<1 ) @+<(<1 )
Elicited prior 2.09 0.22 52.08 52.08 7.36 2.82
1" (!, <0 , <1 ) 2 1 54.6 400 7.39 25
1# (!, <0 , <1 ) 2 0.09 54.6 25 7.39 1

Table 6: Estimated prior probabilities

Estimated prior probabilities
T" T# T$ T% T& T'
Elicited prior 0.17 0.32 0.52 0.73 0.83 0.94
1" (!, <0 , <1 ) 0.21 0.34 0.50 0.63 0.69 0.74
1# (!, <0 , <1 ) 0.10 0.246 0.473 0.758 0.87 0.967

The next, we investigated the effect of prior information on the posterior

distribution. Tables 7 to 9 display the posterior mean and standar deviation of parameters
!, <0 and <1 for the elicited prior, empirical Bayes and priors 1" (!,<0 ,<1 ) and
12 (!,<0 ,<1 ). Table 10 displays the 95% coverage probability.

16
 70% censored data for α 50% censored data for α

0.10
0.15

0.06
bias

bias
0.05

0.02
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β0 50% censored data for β0

0.30

0.08
0.20
bias

bias

0.04
0.10

0.00
0.00

20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β1 50% censored data for β1

0.0

-0.02
-0.2
bias

bias

-0.08
-0.14
-0.4

20 40 60 80 100 20 40 60 80 100

n n
prior1 prior2 elicit

Figure 4: Bias from estimation of !, "! and "1 under informative priors when ! œ 2, "! œ 4
and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5

17
 70% censored data for α 50% censored data for α

0.10

0.08
MSE

MSE
0.00

0.00
20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β0 50% censored data for β0

0.010
0.06
MSE

MSE

0.000
0.00

20 40 60 80 100 20 40 60 80 100

n n

70% censored data for β1 50% censored data for β1

0.00 0.02 0.04
0.08
MSE

MSE
0.00

20 40 60 80 100 20 40 60 80 100

n n
Elicit Prior 1 Prior 2

Figure 5: MSE from estimation of !, "! and "1 under informative priors when ! œ 2, "! œ 4
and "1 œ 2, B" œ 1; B2 œ 1.5 and B3 œ 2.5

**************
From Tables 7 to 9, we observe that the elicited prior gives the same posterior
estimations as the informative 1# (!, <0 , <1 ). Therefore, we can conclude that the simple
and efficient process of elicitation proposed with six quantiles provided by the expert is
possible to construct a very informative prior. Tables 6 shows that this is possible even
with reasonable elicited probabilities.

18
 Several other important conclusions we can also do with this simulation study.
First of all, in oppositive to the noninformative priors, the posterior estimations with
informative priors for the parameters ! and "! are not affected by the high percentage of
censored data or small sample size, but tends to slightly underestimate "1 . Secondly, the
posterior estimators do not seem be too sensitive to the increase of the prior variance.
Thirdly, the posterior standard deviation decreases very slow as sample size increases.
Finally, we observe that the coverage probabilities reach practically 100% for all
informative priors.
The empirical prior does not perform well in comparison with the informative
priors since that its information comes just from the data and the MLE does not provide
good estimators. However, it is important to consider it in practical situations when there
is not available prior information.

7. ILLUSTRATIVE EXAMPLE

In this section, we consider the data generated with 8 œ 50 observations tested

with test stopped after 25th failure (< œ 25). We consider the parameters ! œ 3, "0 œ 4
and "1 œ 0.3. There are three different levels of stresses B" œ 3, B2 œ 6 and B3 œ 10.
The pre-specified times for the stress change are 7" œ 15 and 72 œ 25. The simulated
failure times in the first step are [9.93820, 11.72592, 14.43394, 14.47979], in the second
step are [15.94424, 16.36506, 17.00182, 17.81110, 18.68743, 19.63954, 19.97476,
20.38834, 20.99103, 23.66255, 23.72308, 24.35782] and in the third step are [25.46746,
25.50180, 25.67916, 25.98097, 26.02556, 26.08623, 26.27048, 26.27455, 26.60130].
We now compare the performance of the MLE and Bayesian procedures under the
noninformative (1Ð!,"! ,"1 Ñ and 1(!,<0 ,<1 )), elicited priors and empirical Bayes.
The comparison is carried out by examining the point estimators, intervals and
plots of posterior densities and survival functions.
Tables 11 to 13 present the estimator, standard-deviation and 95% credible
interval.

Table 11: Summaries and 95% credible intervals for !.

Approach !
s S.D. 95% C.I.
MLE 2.9332 1.5353 (-0.0759, 5.9423)
Elicited prior 2.6473 0.4136 (1.9193, 3.5470)
1Ð!,"! ,"1 Ñ 3.2830 1.7353 (1.0797, 7.6721)
1(!,<0 ,<1 ) 3.2371 1.2206 (1.5515, 6.1635)

19
 Table 12: Summaries and 95% credible intervals for "0 .
Approach s0
" S.D. 95% C.I.
MLE 3.7036 1.2432 (1.2669, 6.1402)
Elicited prior 3.9622 0.1999 (3.5576, 4.3377)
1Ð!,"! ,"1 Ñ 3.7516 1.2489 (1.5060, 6.3415)
1(!,<0 ,<1 ) 3.5917 0.8231 (2.0252, 5.2386)

Table 13: Summaries and 95% credible intervals for "1 .

Approach s1
" S.D. 95% C.I.
MLE 0.2654 0.1182 (0.0337, 0.4972)
Elicited prior 0.2840 0.0367 (0.2119, 0.3543)
1Ð!,"! ,"1 Ñ 0.2584 0.1186 (0.0362, 0.4965)
1(!,<0 ,<1 ) 0.2462 0.0843 (0.0804, 0.4034)

The results of Tables show that there is little difference between the performances
of the MLE and Bayesian estimators. The elicited prior gives the smallest standard
deviation and a remarkably narrowest 95% credible interval for the parameters. Besides,
the MLE and noninformative priors provide underestimation of the parameters in
oppositive to empirical Bayes. Surprisingly, the lower endpoint of 95% interval for !
obtained by the MLE approach gives a negative value although this parameter assumes
positive values.
For comparison of the different priors proposed in this paper, the marginal
posterior densities for !, "! and "" by using 1Ð!,"! ,"1 Ñ, 1(!,<0 ,<1 ), empirical and
elicited are plot in Figure 1.
From Figure 1, as expected by the Tables, the marginal posterior densities by
using 1Ð!,"! ,"1 Ñ and 1(!,<0 ,<1 ) are close and the posterior under elicited prior shows a
sharp peak than others due to smaller standard deviation.
To visualize the performance of the estimation, the survival functions WÐ>Ñ are
plotted in Figure 2 based on the MLE and Bayesian approaches.

20
 Figure 1: Marginal posterior densities for the parameters !, "! and "" .

From Figure 2, we observe that there is no difference among the plots of survival
functions with MLE and Bayesian estimations. However, the estimated survival functions
do not provide a close fitting to the true function mainly between the change stress
7" œ 15 and 72 œ 25.

8. CONCLUSIONS

The step-stress accelerated life testing under gamma distribution and Type-II
censoring discussed in this paper can be applied to many problems in reliability analysis.
Besides, in many applications it is desirable to use more than two levels of stress and the
reduced SSALT model used here requires the estimation of only three parameters instead
of the various parameters of the complete model, making the statistical analysis simpler
and computationally faster.

21
 Figure 2: Marginal posterior densities for the parameters !, "! and "" .*****

In the statistics analysis of SSALT presented, in general, the Bayesian technique

has provided more accurate estimators than the maximum likelihood estimation.
However, noninformative priors commonly used in Bayesian inference has showed
inaccurate posterior estimators when the number of test units and failure observations
from the experiment is scarce. To overcome this problem, one reparameterization was
proposed and its effects on the posteior estimation were investigated.
The results of our simulation study indicated that the noninformative prior under
the new parameterization has a more superior performance in terms of coverage
probability, close to the nominal level, and lower standard-deviation.
We also propose a elicited prior when expert's information is available. It is a
quick and accurate method to obtain information for the likely values of the
hyperparameters priors assigned to each parameter of the SSALT model. The
hyperparameters are obtained simply by eliciting directly the probabilities provided by the
expert and based on the calculus of prior predictive probability.

22
 Our elicited prior performs at well as the informative priors used in the simulation
study for comparison. This elicited prior can be applied to any practical problem
involving SSALT, for instance, in the engineering applications, generally there are some
knowledge on the failure mechanism useful in a highly censored environment.
When ones prior knowledge about the parameter is not available, it is possible to
make use of the empirical Bayes for the estimation of the parameters of SSALT. This
implies the use of data as a source of information to determine the hyperparameters of the
prior distribution. The simulation results shown that the accuraces and coverage
probabilities of the empirical Bayes estimators are better than the MLE and
noninformative priors. However, due to data-dependence its performance will be
influenced by the non-Bayesian estimators used as hyperparameters.
Larger sample sizes lead to increased precision of the parameter estimation and
hence the choices of prior distributions have minor effects on posterior inferences.

9. REFERENCES

Alkhalfan, L. (2012). Inference for a Gamma step-stres model under censoring. PhD’s
thesis, McMaster University, Canada.

Alhadeed, A.A.; Yang, S.S. (2005). Optimal simple step-stress plan for cumulative
exposure model using log-normal distribution. IEEE Transactions on Reliability. vol. 54,
no. 1, pp. 64-68.

Bagdonavicius, V. (1978). Testing the hypothesis of additive accumulation of damages.

Probability Theory Application. v. 23, 403–408.

Balakrishnan, N., Kundu, D., Teng Ng, H. K. and Kannan, N. (2007) Point and interval
estimation for a simple step-stress model with Type-II censoring. Journal of Quality
Technology, 39, 35-47.

Balakrishnan, N.; Xie, Q (2007). Exact inference for a simple step-stress model with
Type-II hybrid censored data from the exponential distribution. Journal of Statistical
Planning and Inference, 137, 2543-2563.

23
Balakrishnan, N.; Xie Q.; Kundu, D. (2009). Exact inference for a simple stepstress
model from the exponential distribution under time constraint. Annalsl of the institute of
Statistical Mathematics, 61, 251-274.

DeGroot, M.H.; Goel, P.K. (1979). Bayesian estimation and optimal designs in partially
accelerated life testing”, Naval Research Logistics Quarterly, vol 26, 1979 May, pp 223-
235.

Ganguly, A., Kundu, D., Mitra, S. (2014). Bayesian Analysis of Simple Step-stress Model
under Weibull Lifetimes. IEEE Transaction on Reliability. (Accepted for publication).

Gouno E., Sen A.; Balakrishnan N. (2004). Optimal step-stress test under progressive
Type-I censoring, IEEE Transactions on Reliability, 53, 383-393.

Kateri, M.; Balakrishnan, N. (2008). Inference for a simple step-stress model with type-II
censoring and Weibull distributed lifetime. IEEE Transactions on Reliability. v. 57, 616–
626.

Lee, J.; Pan, R. (2012). Bayesian analysis of step-stress accelerated life test with
exponential distribution. Quality and Reliability Engineering International. v. 28, 3, 353–
361.

Leu, L.Y.; Shen, K.F. (2007). Bayesian approach for optimum step-stress accelerated life
testing. Journal of Chinese Statistical Association. vol. 45, 221 - 225

Liu, X. (2010). Bayesian designing and analysis of simple step-stress accelerated life test
with Weibull lifetime distribution. Master’s thesis, Russ College of Engineering and
Technology, Ohio University, Ohio.

Miller, R.; Nelson, W. (1983). Optimum simple step stress plans for accelerated life
testing. IEEE Transactions on Reliability. vol. 32, 59-65.

Murthy, D., Xie, M., & Jiang, R. (2004). Weibull models. John Wiley & Sonc, Inc

Nelson, W. (1980). Accelerated life testing step-stress models and data analysis. IEEE
Transactions on Reliability. vol. 29, 103-108.

24
Nelson, W. (1990). Accelerated Testing: Statistical Models, Test Plans and Data
Analyses. John Wiley&Sons, New York.

O'Hagan, A..; Buck, C.; Daneshkhan, A., Eiser, R., Garthwaite, P., Jenkinson, D., Oakley,
J., and Rakow, T. (2006). Uncertain Judgements: Eliciting Expert Probabilities.
Wiley&Sons, Chichester

Seydyakin, N.M. (1966). On one physical principle in reliability theory. Technical

Cybernatics, 3:80–87.

Tang, L.C. (2003). Multiple steps step-stress accelerated tests. In Hand Book of
Reliability Engineering (Eds., Pham and Hoang), pp. 441-455, Springer-Verlag, New
York.

Xiong, C.; Ji, M. (2004). Analysis of grouped and censored data from step-stress life test.
IEEE Transactions on Reliability. vol. 53, 22–28.

25
 Appendix A

Deriving the observed Fisher matrix

Confidence intervals and tests of hypotheses for parameters can be obtained either
by using the likelihood ratio method or the approximate normality of MLEs in large
samples. In the latter case it is most convenient to use the approximation.
The asymptotic distributions of the MLE is then used to obtain the approximate
confidence intervals for !, "! and "1 .
Let M (!, "! , "1 ) œ [M34 (!, "! , "1 )], for 3; 4 œ 1, 2, 3, denote the observed Fisher
information matrix of !, "! and "1 where

Ô M"" M"# M"" ×

M (!, "! , "1 ) œ M"" M"" M"" .
Õ M"" M"" Ø
(A1)
M""

The observed Fisher information matrix M (!, "! , "1 ) is given by

M (!, "! , "1 ) œ  Ðf# 6(!, "! , "1 )Ñ. (A2)

where f# 6(!, "! , "1 ) denote the matrix of second derivatives of the log-likelihood.

’Š"  MKÐ!,'<3 Ñ‹Š#< Ð!ÑF" Ð'<$ Ñ 

`# Ð8<3 Ñ
M11 œ ` ! # 6 Ð ! , "! ß "1
l >Ñ œ  <$ < w 'Ð!Ñ 
Š"MKÐ!, '<3 Ñ‹
#

 MKÐ<w 'Ð!Ñ  <w # Ð!Ñ,'<3 Ñ  F# Ð'<$ Ñ‹  Š< Ð!ÑMKÐ!,'<$ Ñ  F" Ð'<$ Ñ‹ “ ,

#

(A3)

l >Ñ œ ! ! ’Š"  MKÐ!,'<3 Ñ‹

$ <5 Ð8<$ Ñ'<!$ / '<$
`#
M12 œ ` ! ` " ! 6 Ð ! ," ! ß " 1 Ð  "Ñ 
>Ð!ÑŠ"MKÐ!, '<$ Ñ‹
#
5œ" 35 œ<5" +1

‚ 68Ð'<3 Ñ  <Ð!Ñ  F" Ð'<$ Ñ“, (A4)

l >Ñ œ ! ! # ’< Ð!Ñ 

$ Ð8<$ Ñ'<!$" / '<$ A1 Ð><$ Ñ
`# 1
M13 œ ` !` "1 6Ð!,"! ß "1 A Ð> Ñ 
>Ð!ÑŠ"MKÐ!, '<$ Ñ‹
'< 1 3 5
5œ" 35 œ<5" +1 $

 Š"  MKÐ!,'<3 Ñ‹68Ð'<3 Ñ  F" Ð'<$ Ñ“, (A5)

26
 l >Ñ œ ! ! ’Š"  MKÐ!,'<3 Ñ‹
$ Ð8<$ Ñ'<!$ / '<$ A1 Ð><$ Ñ
`#
M22 œ 6Ð!,"! ß "1 '< $ 
>Ð!ÑŠ"MKÐ!, '<$ Ñ‹
` "!# #
5œ" 35 œ<5" +1

“,
'<!$ / '<$
‚ Ð'< 3  !Ñ  >Ð!Ñ (A6)

l >Ñ œ  ! !
$ Ð8<$ Ñ'<!$ / '<$ A1 Ð><$ Ñ
`#
M23 œ ` "! ` "1 6Ð!,"! ß "1 Ð  A1 Ð>35 ÑÑ 
>Ð!ÑŠ"MKÐ!, '<$ Ñ‹
#
5œ" 35 œ<5" +1

‚ ’Š"  MKÐ!,'<3 Ñ‹Ð '!<  "Ñ  “

'<!$" / '<$
>Ð!Ñ (A7)
3

and

l >Ñ œ (!  ")! ! Š'35 A# Ð>35 Ñ  A#" Ð>35 Ñ‹  ! ! A# Ð>35 Ñ

$ $
`# "
M33 œ ` "1#
6 Ð ! , "! ß "1 '3#
5œ" 35 œ<5" +1 5 5œ" 35 œ<5" +1

šŠ"  MKÐ!,'<3 Ñ‹’Š !'"  "‹A#" Ð><$ Ñ  A# Ð><$ Ñ“  >Ð!Ñ A" Ð><$ Ñ›
Ð8<$ Ñ'<!$" / '<$ '<!$" / '<$ #

>Ð!ÑŠ"MKÐ!, '<$ Ñ‹
#
< 3

(A8)

" 'B !1 ? " 'B # !1 ?

where F" ÐBÑ œ >(!) 0 68Ð?Ñ? / .?, F2 ÐBÑ œ >(!) 0 68 Ð?Ñ? / .?,

E" Ð>35 Ñ œ Ð>35  75" ÑB5 /Ð"0 "" B5 Ñ  7 /"0 ! B4 /"" B4 ,

5"

4œ"

E2 Ð>35 Ñ œ Ð>35  75" ÑB#5 /Ð"0 "" B5 Ñ  7 /"0 ! B4 /"" B4

5"

4œ"

It is known that M2" œ M"# ß M3" œ M"3 and M$2 œ M23 . Now, the variances and
covariances of ! s ! and "
s, " s 1 can be obtained through the observed Fisher information
matrix.
The asymptotic distribution of the maximum likelihood estimators are then given
(! s! , "
s, " s 1 ) µ R$ ÐÐ!, "! , "1 Ñß M " (! s! , "
s, " s 1 )Ñ, which can be used to construct
100(1  # )% confidence intervals for the parameters !, "! and "1 .

27
 Appendix B

Expected lifetime of step-stress accelerated test

In this appendix we show the calculus to derive the expected lifetime of step-stress
accelerated test.
We consider the calculations for just three steps, although it can be extended for
any 7 step-stress.
The integrals below necessary to derive the expectancy of variable X .

-
( B / , .B œ , - ’ Š ‹ / #, [ 23>>+5/<Q Œ ß
,"a - a" -  #a  -
ß 
a1  a" B a a" -
0 a(a  ") , # # ,

Š ‹ / #, [ 23>>+5/<Q Œ  "ß ß “ ß
,#a - a# -  #a  - a a" -
(B1)
a(a  ") , # # ,

where [ 23>>+5/<Q Ð., / , DÑ function is defined as (B2)

[ 23>>+5/<Q Ð., / , DÑ œ / # D # / 2C:/<1/97ˆ "#  /  .ß "  #/ ß D ‰Þ

D "
(B3)

_ >ˆaß ., ‰
(
a1  B,
B / .B œ (B4)
. ,a

-
Š ‹ / #, [ 23>>+5/<Q Œ ß
,- a -  #a  -
( B/ , œ ß 
a  B a a" -
(B5)
0 a" , # # ,

_ >ˆa  "ß ., ‰
(
B
Ba / , .B œ (B6)
. ,a"

By using these integralsthe expectancy is derived as:

7" 7# _
IÐX Ñ œ ( >0" (>).>  ( >0# (>  7"  &" ).>  ( >0$ (>  7#  &# ).> œ
! 7" 7#

7" 7#
( (
" >
!  )1 " >7 &
!  )"# "
œ > / .>  >(>  7 "  & " ) / .> 
>(!))"! ! >(!))#! 7"

28
 _
(
" >7 &
!  )#$ #
 >(>  7 #  & # ) / .>. (B7)
>(!))$! 7#

where each integral is given below.

From (B5), the first integral is given by

7"  !#
)" 7"! 7"
( Œ  [ 23>>+5/<Q Œ ß ß .
>
!  )1
7
 #)" ! !  " 7"
> / .> œ / " (B8)
! !  " )" # # )"

Now, by replacing ? œ >  7"  &" in the second integral we have

7# 7# 7" &"
( .> œ (
>7" &"
  )?
>(>  7"  &" )! / )#
(?  7"  &" )?!" / # .? œ
7" &"

7# 7" &" 7# 7" &"

œ( .?  Ð7"  &" )(
 )?  )?
?! / # ?!" / # .?. (B9)
&" &"

let us solve the two integrals in (B9) separately

7# 7" &" !
)# Ð7 #  7 "  & " Ñ! 7 #  7 "  & "
3Ñ ( Œ  / #)#
? # 7 7 &
!  )#  # " "
? / .? œ
&" !" )#
!

[ 23>>+5/<Q Š !# ß !" ‹
)# &!" &"
Œ  Œ ß ß 
7# 7" &"
# &
 #)" ! !  " &"
# ß )#  / # [ 23>>+5/<Q
!  " )# # # )#
(B10)
and
7# 7" &"
?!" / )# .? œ )#!" (7#  7"  &" )’ #
)"! (7#  7"  &" )!"
33Ñ (
?

&" !(!  ")

Š 7# )7"#&" ‹ [ 23>>+5/<Q Š !# ß !" ‹

 !#

7# 7" &"
7# 7" &" )##! (7#  7"  &" )!#
/ #)#
# ß )# !(!  ")

Š (7# )7"#&" ) ‹ [ 23>>+5/<Q Š !#  "ß !" ‹“  )#!" &" ’

 !# (7# 7" &" )
 (7# 7" &" ) )#"! &"!"
/ #)#
# ß )# !(!")

 !#
Š &)"# ‹ # [ 23>>+5/<Q Š ß
# ß )# ‹
 !# )#! &"!# &"
Œ 
& &
 #)" ! !" &"  #)"
/ #  # / # [ 23>>+5/<Q
!(!  ") )#
Š !#  "ß !"
# ß )# ‹“
&"
(B11)

29
 Now, by replacing ? œ >  72  &2 in the third integral in (B7) we have

_ >Š!  "ß &)2$ ‹ Ð72  &2 Ñ>Š!ß )&2$ ‹

(
?
!"  )$
Ð?  72  &2 Ñ? / .? œ  (B12)
&2 )$!" )$!

Finally, with the integrals (B8), (B10), (B11) and (B12), the expectance IÐX Ñ can
be obtained,
The development of this result can be generalized to a larger number of steps in
the test.

30