1
Test for homogeneity of lifetimes of several systems under generalized inverted scale family of
distributions based on Type II censored sampling design
Shanubhogue, A1 and Raykundaliya, D.P
1,*.
1Department of Ststistics, Sardar Patel University, Vallbh Vidyangar- 388120, Gujrat-India
*corresponding author: [email protected], +919428435134
Abstract:
In this paper we consider generalized inverted scale family of distributions (GIFD) as
probability models for m independent systems of different make and study the problem of
simultaneous estimation of the unknown parameters under an extension of Type II censoring
scheme. We consider Maximum likelihood estimation (MLEs) of the scale and shape parameters.
Asymptotic confidence intervals for �+1 parameters based on the MLEs are also constructed.
As a special case, we consider generalized inverted exponential distribution (GIED) and
conducted simulation study to investigate performance of the estimates and confidence intervals
under type II censoring. Further, likelihood ratio test is used to test for homogeneity of several
scale parameters and the cost effectiveness of the Type II censoring scheme in planning of
experiment for the study of several populations is done using realistic cost function. We also
study effect of censoring on duration of experiment, total time on experiment, cost of experiment
and on estimates of parameters through simulation.
Keyword: Type II censoring scheme, generalized inverted exponential family of distribution
(GIED), Newton-Raphson method, ML estimation, Monte-Carlo Simulation
technique, likelihood ratio test, total time on test, cost function, proportion of
censoring
Introduction:
The industrial revolution onward man has become more and more dependent on
“Machine” or “System”. The failure of the machine stops the whole system abruptly. Therefore,
it is important to know among the several machines during the same work, which machine
should be faithful or reliable for their daily usage and having minimum maintenance but not
unduly expensive. This requires extensive results from reliability and life tastings. In life testing
2
experiment, the objective is to obtain the estimates of mean time to failure, hazard rate and
reliability at a given point etc. In case of a study of systems from different brands performing
identical task the problems are: (i) simultaneous estimation of the parameters of the assumed
probability models for the life times and (ii) obtaining tests for the homogeneity of life time
distributions or equality of certain parameter of the life time distributions. As continuation of
life testing experiments till all the units put on test fail may lead to not termination of experiment
in a manageable duration of time and increased experimental cost, one has to resort to some kind
censoring scheme leading to enough data information in reasonable duration of time. Often,
Type II censoring is the most commonly used sampling plan. In statistics literature we find lots
of research papers which use the plan for various lifetime models such as normal, exponential
and Weibull. Few noteworthy research papers to cite are Gupta (1952), Cohen (1965), Mann et
al. (1974), Lawless (1982), Sinha (1986) and Hossain et al. (2003) .
In manufacturing setting, the problem of comparing effectiveness of products is
important. In this case after placing several independent samples of units manufactured by
several processes, the reliability engineer would like to make early and efficient decision on the
effectiveness of the products under the life test in terms of mean time to failure, standard hazard
rate functions, reliability function etc. Balakrishnan and Ng (2006) extensively study the
problem of comparing two populations in terms of stochastic ordering. Sharafi et al. (2013)
study a distribution free test for comparison of hazard rates of two distributions under Type II
censoring. Shafay et al. (2013) discuss the Bayesian inference on jointly Type II censored sample
for two exponential populations. Srivastava (1987) discusses the method of generalized Type II
censoring for comparison of life times of several systems under Weibull distribution. Further,
experimenters are also concerned with the cost of experiments, total time under experiments and
duration of experiment for better planning of the experiment. There are several costs associated
with the conduct of the experiment. In this pursuit, Srivastava (1989) studies a realistic cost
function for several systems put on tests under life experiment for the generalized Weibull
distribution.
The scale family of distributions play very important role in lifetime data analysis.
Exponential distribution, Rayleigh distribution and Weibull distribution are such distributions
which are widely used by researchers in reliability studies. Recently, Potdar and Shirke (2013)
3
proposes a generalized inverted scale family of distributions (GIFD) and shows that generalized
inverted exponential distribution (GIED) (for detail one can refer; Abouammoh and Alshingti
(2009)), generalized inverted-half logistic distribution (GIHD) and generalized inverted Rayleigh
distribution (GIRD) are the members of GIFD. Further they obtain ML estimates of parameters
of GIFD and obtain confidence intervals for the parameters involved in the distributions and
discuss the inferential procedures for GIHD.
In this paper we discuss inferential problems on lifetimes of several systems under the
generalized inverted scale family of distributions based on a generalization of Type II censored
scheme for several samples. We further, study the reliability characteristics of the distributions.
Under generalized Type II censoring design we put ‘�’ units, for each ‘�’ types of systems,
simultaneously on test and continue the experiment till �∗ failures in each system are observed
i.e. the total number of units put on test are ‘��’ and the total number of failures we observe at
the end of experiment is � = ��∗. Here we assume that lifetimes of systems follow
generalized inverted scale family of distributions with shape parameter � and scale
parameters; � = 1,2, … ,�. The observed failure times are denoted by ��; � = 1,2, … , �∗; � =1,2, … ,�. Thus the experimental data under this scheme are(�, �, ��; � = 1,2, … , �∗; � =1,2, … ,�). The organization of whole paper is as below.
In Section 2, we give the probability density function, the survival function and the hazard
rate of the generalized inverted scale family of distributions (GIFD), and develop the likelihood
for Type II censored sampling design under GIFD. In Section 3, we derive the expressions for
maximum likelihood equations for the parameters and their asymptotic variance-covariance
matrix for both cases of known and unknown shape parameters. Further, we obtain asymptotic
confidence interval for �+ 1 parameters of the distributions. Section 4 discusses likelihood
ratio test for simultaneous testing of homogeneity of scale parameters when the shape parameter
is known and when it is unknown. In Section 5, we consider GIED as a special case and obtain
ML estimates of the parameters, confidence intervals for the parameters, reliability
characteristics of the distribution, and efficiency measures like mean square error (MSE) and
standard error(SE) for the parameters through Monte-Carlo simulations. Further the likelihood
ratio test for simultaneous testing of homogeneity of scale parameters, when the shape parameter
is known, is discussed and the cut-off points for the test statistics are also obtained through
4
Monte-Carlo simulation. In Section 6, we discuss realistic cost function for the problem of
planning such experiments. We further study the effect of proportion of censoring on cost
function. Some concluding remarks are given in Section 7.
2. Generalized inverted scale family of distributions (GIFD) and likelihood function for
type II censoring design
Consider an item whose life time is denoted by�. The random variable � is said to have
two -parameter generalized inverted scale family of distributions, if its distribution function is
given by
��(�; , �) = �(� ≤ �) = 1 − �� � !"#$%, � ≥ 0; �, > 0 (2.1)
where �(�) is a parameter free survival function. If � has the distribution function (2.1), then
the corresponding density function, reliability function and hazard function are respectively
given by
)(�; , �) = %!"* ℎ � !"# �� � !"#$%, , � ≥ 0; �, > 0 (2.2)
��(�; , �) = �� � !"#$% (2.3)
and
-�(�; , �) = ./(")0/(") = 123*4� 523#�6� 523#$175�6� 523#$1 = %!"* 84� 523#6� 523#9 . (2.4)
The marginal likelihood function for �-th type of system under type II censoring based on
observing G*failures from� units is
: = ;!(;,=∗)!∏ )(��)?�@A (�=∗)B;,=∗=∗�C . (2.5)
Therefore, the joint likelihood of the whole experiment is
: = ∏ :DC = ∏ E ;!(;,=∗)!∏ )(��)?�@A (�=∗)B;,=∗=∗�C FDC . (2.6)
Substituting the density function in (2.2) and survival function in (2.3) in (2.6) we get
: = ∏ G ;!(;,=∗)!∏ %!H"IH* ℎ J !H"IHK L� J !H"IHKM%, �� � !H"N∗H#$%(;,=∗)=∗�C ODC . (2.7)
5
3. Maximum likelihood estimation and confidence interval
The log likelihood function of (2.7) is
P = �PQ � ;!(;,=∗)!# +��∗PQ� − �∗∑ PQ − 2DC ∑ ∑ ln �� +=∗�C DC ∑ ∑ ln Lℎ J !H"IHKM=∗�C DC
+(� − 1)∑ ∑ ln L� J !H"IHKM + �(� − �∗)∑ PQ �� � !H"N∗H#$DC =∗�C DC (3.1)
Differentiating (3.1) with respect to � and (� = 1,2,⋯ ,�) we get
VWV% = D=∗% +∑ ∑ ln L� J !H"IHKM + (� − �∗)∑ PQ �� � !H"N∗H#$DC =∗�C DC (3.2)
VWV!H = − =∗!H − !H*∑ X "IH4′Y 52H3IHZ4Y 52H3IHZ[=∗�C − (%, )!H* ∑ X "IH
6′Y 52H3IHZ6Y 52H3IHZ[ − %(;,=∗)"N∗H!H*6′J 52H3N∗HK6J 52H3N∗HK
=∗�C (3.3)
for, � = 1,2,⋯ ,�
Equating equation (3.2) to zero we have,
�\ = − D=∗∑ ∑ ]^86Y 52H3IHZ9_(;,=∗)∑ W`L6J 52H3N∗HKMaHb5N∗Ib5aHb5 (3.4)
The estimates of parameters (� = 1,2,⋯ ,�) can be obtained in two different cases: (i) shape
parameter � is known and (ii) shape parameter � is unknown.
3.1 Maximum likelihood estimation when shape parameter c is known
The solutions of ML equations (3.3) cannot be obtained explicitly. Hence, we use
Newton-Raphson method for the data (�, �, ��; � = 1,2, … , �∗; � = 1,2, … ,�). The method
requires some initial values of ( , d, e, … , D). The initial values are obtained using the least
square method as suggested in Ng (2005). The method uses the empirical distribution function
for �-th population as given below
�fg��h = 1 − ∏ (1 − ik)�kC
with
ik = ;,k_ , for l = 1,2… , �;� = 1,2… , �∗; � = 1,2, … ,�
6
The parameters’s are estimated by fitting the linear regression through the method of least
square. Here we consider
m� = �� , where
m� = 675n�57opHg3I75h#5/1r�57opHg3Ih#5/1* s � = 1,2… , �∗; � = 1,2, … ,� ,
and �t is such that
�f(�t) = 0
The least square estimates of ; � = 1,2, … ,� are given by
ut = ∑ vIH"IHN∗Ib5∑ "IH*N∗Ib5 ; � = 1,2, … . ,� (3.5)
These values are used as initial solution for the ML estimation by the Newton-Raphson method.
The MLEs of reliability (�(�); � = 1,2, …�) and hazard rate (-(�); � = 1,2, … ,�) can be
evaluated using invariance property of MLEs as
�p(�) = �� � !p"H#$% (3.6)
-g�; u, �h = %!p"* w4J 52p3HK6J 52p3HKx for � = 1,2, … ,� (3.7)
3.1.1 2 Observed Fisher information matrix under the design
The Fisher information matrix is
y = �gzkh# ,
where
zk = −{ J V*WV!HV!|K for �, l = 1,2, … ,�.
7
Now for � = l, V*WV!H* = =∗!H* + ∑ "IH* !H}~�
�4Y 52H3IHZ4′′Y 52H3IHZ,�4′Y 52H3IHZ�*_d!H"IH4Y 52H3IHZ4′Y 52H3IHZ
84Y 52H3IHZ9* ���=∗�C
+(� − 1)∑ "IH* !H}~��6Y 52H3IHZ6′′Y 52H3IHZ,�6′Y 52H3IHZ�
*_d!H"IH6Y 52H3IHZ6′Y 52H3IHZ86Y 52H3IHZ9* ��
�=∗�C
+ %(;,=∗)"N∗H* !H} �6J 52H3N∗HK6′′J 52H3N∗HK,Y6′J 52H3N∗HKZ*_d!H"IH6J 52H3N∗HK6′J 52H3N∗HKL6J 52H3N∗HKM* � for (� = 1,2,⋯ ,�)
(3.8)
where, ℎ′(∙) = VV!H ℎ(∙), �′(∙) = VV!H�(∙), ℎ′′(∙) = V*V!H* ℎ(∙), �′′(∙) = V*V!H*�(∙) for (� = 1,2,⋯ ,�).
Since systems are independently functioning in the experimental set up, we have
V*WV!HV!| = 0 ∀l ≠ � = 1,2, …� (3.9)
Theorem 3.1: Let assume that all the usual regularity conditions needed for the asymptotic
distribution of mle hold true. Then, the limiting distribution of u , as � → ∞ such that =∗;
remains constant, is �-variate normal with mean vector and dispersion matrix y, .
3.2 Maximum likelihood estimation when shape parameter c is unknown
The solutions of ML equations 3.2 and 3.3 cannot be obtained explicitly. Hence, we use Newton-
Raphson iterative method for the data (�, �, ��; � = 1,2, … , �∗; � = 1,2, … ,�). The initial
values of the parameters of ; � = 1,2, … ,� are obtained through the method of least square as
discussed in Section 3.1 and initial value � is obtained by using equation (3.4). These initial
values are used in the Newton-Raphson method to obtain the MLEs of (�, ). The MLE of
reliability �(�)and hazard rate -(�), � = 1,2, … ,� can be evaluated using invariance
property of MLEs as
8
�p(�) = �� � !p"H#$%�
(3.10)
-�g�; u, �\h = %�!p"* w4J 52p3HK6J 52p3HKx for � = 1,2, … ,� (3.11)
3.2.1 Observed Fisher information matrix under the design
The Fisher information matrix is
� = −{ wV*WV%* ∆′∆ yx y = �gzkh# ,
where
zk = V*WV!HV!| for �, l = 1,2, … ,�
and
∆′ = � V*WV%V!5 , V*WV%V!* , … , V*WV%V!a#.
V*WV%* = −D=∗%* (3.12)
V*WV%V!H = − !H*∑ X "IH6′Y 52H3IHZ6Y 52H3IHZ[ − (;,=∗)"N∗H!H*
6′J 52H3N∗HK6J 52H3N∗HK=∗�C For (� = 1,2,⋯ ,�) (3.13)
V*WV!H* = =∗!H* + ∑ "IH* !H}~��4Y 52H3IHZ4′′Y 52H3IHZ,�4′Y 52H3IHZ�
*_d!H"IH4Y 52H3IHZ4′Y 52H3IHZ84Y 52H3IHZ9* ��
�=∗�C
+(� − 1)∑ "IH* !H}~��6Y 52H3IHZ6′′Y 52H3IHZ,�6′Y 52H3IHZ�
*_d!H"IH6Y 52H3IHZ6′Y 52H3IHZ86Y 52H3IHZ9* ��
�=∗�C
9
+ %(;,=∗)"N∗H* !H} �6J 52H3N∗HK6′′J 52H3N∗HK,Y6′J 52H3N∗HKZ*_d!H"IH6J 52H3N∗HK6′J 52H3N∗HKL6J 52H3N∗HKM* � for � = 1,2, … ,�
(3.14)
where, ℎ′(∙) = VV!H ℎ(∙), �′(∙) = VV!H�(∙), ℎ′′(∙) = V*V!H* ℎ(∙), �′′(∙) = V*V!H*�(∙) for � = 1,2, … ,�.
Since the systems are independently functioning in the experimental set up, we have
V*WV!HV!| = 0 ∀l ≠ � = 1,2, …�. (3.15)
Theorem 3.2: Let assume that all the usual regularity conditions needed for the asymptotic
distribution of mle hold true. Then the limiting distribution of (�\, u), as � → ∞ such that =∗;
remains constant, is(� + 1)-variate normal with mean vector (�, ) and dispersion matrix �,
under regularity conditions.
3.3 Confidence interval
Assuming asymptotic normal distribution for the MLEs, CIs for (�, , d, e, … , D) are
constructed. Let (�\, u , ud, ue, … , uD) are MLE of (�, , d, e, … , D) respectively. Let �\d(�\) and �\dguh; � = 1,2, … ,� is the estimated variance of �\and u; � = 1,2, … ,� respectively.
Then, 100(1 − �)% asymptotic CI for � and : � = 1,2, … ,� are respectively given by
J�\ − �� d� ��\d(�\), �\ + �� d� ��\d(�\)K and Yu − �� d� ��\dguh, u + �� d� ��\dguhZ For � = 1,2, … ,� (3.16)
where �� d� is the upper 100(1 − �)"4 percentile of standard normal distribution.
10
4. Testing of hypotheses
The proposed design will have significance only when we are able to ascertain that the m
type of systems are not all have identical life time. This can be done by developing ANOVA
approach for the proposed design. However, we will utilize likelihood approach to develop a test.
The testing of hypothesis problem is to test
�t: = d = ⋯ = D = against � : ≠ kforatleastonepair(�, l), � ≠ l. (4.1)
As we are considering maximum likelihood estimation, the use likelihood ratio test is
much convenient. The test statistic is
¤¥¦ = max%,! :(�, , �)max%,! :(�, , �) The test based on −2PQ(¤¥¦) rejects H0 in support of H1 if it is larger than upper �thcut point of
chi-square distribution (m-1) degrees of freedom.
4.1 Computation of likelihood under H0
The log likelihood equation (3.1) can be written under to as
P = �PQ � ;!(;,=∗)!# + ��∗ ln � − ��∗ ln − 2∑ ∑ ln �� + ∑ ∑ ln ℎ J !"IHK=∗�C DC =∗�C DC
+(� − 1)∑ ∑ ln� J !"IHK=∗�C DC + �(� − �∗) ∑ ln� � !"N∗H#DC (4.2)
Differentiate (4.2) with respect to � and we have
VWV% = D=∗% +∑ ∑ ln� J !"IHK=∗�C DC + (� − �∗)∑ ln� � !"N∗H#DC (4.3)
VWV! = −D=∗! − !*∑ ∑ n "IH ∙ 4′Y
523IHZ4Y 523IHZs=∗�C DC − (%, )!* ∑ ∑ n "IH ∙ 6′Y523IHZ6Y 523IHZs=∗�C DC − %(;,=∗)!* ∑ w "N∗H
6′J 523N∗HK6J 523N∗HKxDC
(4.4)
Differentiate (4.3) and (4.4) with respect to � and we have
11
V*WV%* = −D=∗%* (4.5)
V*WV%V! = − !*∑ n "IH ∙ �′Y523IHZ6Y 523IHZs − (;,=∗)!* ∑ w "N∗H
�′J 523N∗HK6J 523N∗HKxDC =∗�C (4.6)
V*WV!* = D=∗!* + !}∑ ∑ "IH* ª««
¬4Y 523IHZ4′′Y 523IHZ,�4′Y 523IHZ�*_d!"IH4Y 523IHZ4′Y 523IHZ
�4Y 523IHZ�* ®®
=∗�C DC
+ (%, )!} ∑ ∑ "IH* ª««¬6Y 523IHZ�′′Y 523IHZ,��′Y 523IHZ�
*_d!"IH6Y 523IHZ�′Y 523IHZ�6Y 523IHZ�
* ®®=∗�C DC
+ %(;,=∗)!} ∑ "N∗H* °6J 523N∗HK�′′J 523N∗HK,Y6′J 523N∗HKZ*_d!"IH6J 523N∗HK�′J 523N∗HK
Y6J 523N∗HKZ* ±DC (4.7)
Since the likelihood equations 4.3 and 4.4 are not mathematically tractable, we use the Newton-
Raphson method to obtain the estimates of parameters �²Q³. 5. Application of generalized inverted exponential distribution (GIED)
Consider a member of generalized Inverted family of distribution, namely Generalized
Inverted Exponential distribution suggested by Abouammoh and Alshingti (2009). Let, � be
generalized Inverted Exponential random variable. The cdf and pdf of � are respectively,
��(�) = 1 − L1 − ´�, 523#M% � ≥ 0, �, > 0 (5.1)
and
)�(�) = %!"* ´�, 523# L1 − ´�, 523#M%, � ≥ 0, �, > 0 (5.2)
Here � and are scale and shape parameters respectively.
The reliability function of GIED (, �) is given by
�(�) = L1 − e�7523#M% , � ≥ 0, �, > 0 (5.3)
12
The failure rate function of GIED (, �) is given by
-�(�) = .(")¦(") = %!"* µJ7523K ,µJ7523K , � ≥ 0, �, > 0 (5.4)
Since the mean of the distribution has no closed form and it is finite only if� > 1 , we shall
consider here the median time to system failure (¶·�¸�) which is given by
¶·�¸� = n− !W`8 ,(t.¹)519s, �, > 0 (5.5)
It is clearly seen that the distribution belongs to GIFD if we take
� � !"# = L1 − ´�, 523#M (5.6)
and
ℎ � !"# = ´�, 523# (5.7)
and some other results
ℎ′ � !"# = ··! ℎ � !"# = ··! ´�, 523# = −´�, 523#, ℎ′′ � !"# = ´�, 523# �′ � !"# = ´�, 523#, �′′ � !"# = −´�, 523# (5.8)
Substituting the equations 5.6 to 5.8 in equations 3.1 to 3.13 we obtain the log likelihood
function, maximum likelihood equations for parameters, Fisher information of parameters for
generalized inverted exponential distribution. Details are omitted to avoid repetitions.
5.1 Algorithm, Numerical Exploration and Conclusions
In this Section, a Monte Carlo simulation study is conducted to compare the performance
of the estimates. Maximum likelihood estimates are obtained for observations generated through
the Type II censoring design when numbers of systems to be compared are 2 and 3 for known as
well as unknown shape parameter. All calculations are performed with the aid of R- language
version- R.2.12.0. The simulation study is conducted for both known as well as unknown shape
parameter.
13
5.2.1 Known Shape Parameter
In this section, we carryout simulation study for two sets of parameter values � = 2, � = 2.5, = 1.5, d = 1.3 and for� = 3, � = 2.5, = 1.5, d = 1.3, e = 1.4,
and for different values of u and �∗. Here we keep total number of failures in the whole
experiment � = ��∗ fixed. We simulate 1000 samples for each case. The simulation results are
summarized in Table 1 and Table 2. To carry out our objective we proceed through following
algorithm.
Step 1: Taking � = 2, � = 2.5, = 1.5, d = 1.3 we generate � random numbers from �½{¾(�, , d, … , D) for each type of systems. The same is repeated for the
parameters� = 3, � = 2.5, = 1.5, d = 1.3, e = 1.4.
Step 2: Generate �∗ Type II censored observations for each type of systems. The generated �∗ failure times are (� , �d, … , �=∗); � = 1,2, … ,� for each type of systems.
Step 3: Obtain initial estimate of parameters ; � = 1,2, … ,� using the least method discussed
in Section 3.1 i.e. evaluate initial estimates using formulae
ut = ∑ vIH"IHN∗Ib5∑ "IH*N∗Ib5 ; � = 1,2, … . ,�.
Step 4: Obtain initial value for sample information matrix ¿pusing the value obtained in Step 3
and also obtain the score vector ¸ ′ = � VWV!5 , VWV!* , … , VWV!a#. Step 5: Use Newton-Raphson iterative method
uÀÁÂ = uÃW· + ¿p, �uÃW·# ∗ ¸.
Step 6: Repeat the Step 5 until the ∑ ÄuÀÁÂ − uÃW·ÄDC < Æ where Ç is very small predefined
quantity.
Step 7: Repeat the procedures in Steps 1 to Step 6 for Q = 1000 times and obtain the following
quantities.
(a) {y = ∑ !pH|È|b5;
14
(b) Mean Squared Error, ¶¸{ = ∑ (!pH|,!H)*È|b5 ` where; � = 1,2, … ,� the values of
parameters given in Step 1.(c) Average of variance-covariance matrices computed for different simulated samples, say y∗, .(d) Evaluate Reliability functions �pk(�) and hazard rater�k(�); � = 1,2, … ,�; l =1,2, … , n using equations 5.3 and 5.4 for each simulated sample and obtain the average
of all these values.
(e) Obtain MSE(Reliability) = ∑ (0pH|("H),0H("H))*È|b5 ^ and MSE (hazard rate)=
∑ (ÌH|("H),ÌH("H))*Í|b5 `
respectively at median given in equation (5.5).
Step 8: Obtain Standard Error (SE) of estimates by taking square root of diagonal
elementsy∗, .
Table 1
Maximum likelihood estimate of parameters, reliability and hazard rates and their efficiency measures � = 2,� = 2.5, = 1.5, d = 1.3²Q³� = (0.4700,0.5424), �(�) = (0.5,0.5), -(�) = (2.4099,2.0886) � �∗ Ò dÒ � Ò(� ) �dÒ(�d) r �(� ) r\(�d)
12
06
{y 1.5200 1.3068 0.5059 0.5117 2.3851 2.0486 ¶¸{ 0.0917 0.0746 0.0120 0.0129 0.1731 0.1396 ¸{ 0.3164 0.2727 - - - -
24
12
{y 1.5134 1.3107 0.5019 0.5030 2.4016 2.0774 ¶¸{ 0.0450 0.0318 0.0062 0.0069 0.0875 0.0733 ¸{ 0.2200 0.1907 - - - -
36
16
{y 1.5011 1.3007 0.5045 0.5048 2.3924 2.0723 ¶¸{ 0.0320 0.0250 0.0044 0.0047 0.0619 0.0494 ¸{ 0.1774 0.1538 - - - -
48
24
{y 1.4993 1.3011 0.5039 0.5029 2.3949 2.0788 ¶¸{ 0.0235 0.0164 0.0033 0.0031 0.0464 0.0324 ¸{ 0.1532 0.1328 - - - -
60
30
{y 1.4966 1.3071 0.5042 0.5000 2.3939 2.0881 ¶¸{ 0.0187 0.0152 0.0026 0.0028 0.0367 0.0292 ¸{ 0.1365 0.1193 - - - -
72
36
{y 1.5014 1.3045 0.5019 0.5006 2.4025 2.0862 ¶¸{ 0.0158 0.0126 0.0022 0.0024 0.0305 0.0245 ¸{ 0.1249 0.1086 - - - -
84
42
{y 1.4972 1.3007 0.5031 0.5017 2.3983 2.0830 ¶¸{ 0.0128 0.0093 0.0018 0.0017 0.0249 0.0185 ¸{ 0.1120 0.0950 - - - -
96
48
{y 1.5049 1.3036 0.4999 0.5002 2.4098 2.0877 ¶¸{ 0.0116 0.0085 0.0016 0.0015 0.0225 0.0164 ¸{ 0.1083 0.0938 - - - -
15
Table 2
Maximum likelihood estimate of parameters, reliability and hazard rates and their efficiency measures
Ó = Ô,c = Õ. Ö, ×Ø = Ø. Ö,×Õ = Ø. Ô, ×Ô = Ø. Ù,Ú = (Û. ÙÜÛÛ, Û. ÖÙÕÙ, Û. ÖÛÔÝ), �(�) = (Û. Ö, Û. Ö, Û. Ö), Þ(Ú) = (Õ. ÙÛßß, Õ. ÛààÝ, Õ. ÕÙßÔ) � �∗ Ò dÒ eÒ � Ò(� ) �dÒ(�d) �eÒ(�e) r �(� ) r\(�d) -e�(�e) 12 04
{y 1.5173 1.2979 1.4037 0.5077 0.5148 0.5019 2.3782 2.0380 2.2404 ¶¸{ 0.0974 0.0703 0.0849 0.0127 0.0126 0.0124 0.1838 0.1359 0.1555 ¸{ 0.3163 0.2706 0.2981 - - - - - -
24
08
{y 1.5011 1.3073 1.4105 0.5068 0.5040 0.5038 2.3836 2.0747 2.2350 ¶¸{ 0.0468 0.0350 0.0460 0.0065 0.0065 0.0071 0.0917 0.0682 0.0817 ¸{ 0.2183 0.1900 0.2054 - - - - - -
36
12
{y 1.5003 1.3088 1.3985 0.5047 0.5010 0.5051 2.3919 2.0848 2.2307 ¶¸{ 0.0309 0.0234 0.0254 0.0043 0.0044 0.0042 0.0604 0.0457 0.0506 ¸{ 0.1773 0.1547 0.1653 - - - - - -
48
16
{y 1.5018 1.3079 1.4025 0.5031 0.5003 0.5027 2.3978 2.0872 2.2395 ¶¸{ 0.0243 0.0180 0.0204 0.0034 0.0033 0.0033 0.0477 0.0345 0.0405 ¸{ 0.1534 0.1336 0.1433 - - - - - -
60
20
{y 1.4982 1.2989 1.4044 0.5037 0.5034 0.5009 2.3958 2.0774 2.2457 ¶¸{ 0.0193 0.0138 0.0150 0.0027 0.0026 0.0024 0.0378 0.0269 0.0295 ¸{ 0.1367 0.1185 0.1281 - - - - - -
72
24
{y 1.5039 1.3049 1.3981 0.5010 0.5005 0.5030 2.4058 2.0867 2.2383 ¶¸{ 0.0161 0.0127 0.0128 0.0022 0.0024 0.0021 0.0308 0.0246 0.0254 ¸{ 0.1252 0.1086 0.1163 - - - - - -
84
28
{y 1.5059 1.3017 1.3995 0.4997 0.5015 0.5022 2.4108 2.0834 2.2411 ¶¸{ 0.0126 0.0110 0.0115 0.0018 0.0021 0.0017 0.0246 0.0217 0.0228 ¸{ 0.1159 0.1002 0.1077 - - - - - -
96
32
{y 1.4928 1.3048 1.3961 0.5045 0.4998 0.5013 2.3930 2.0891 2.2447 ¶¸{ 0.0113 0.0090 0.0093 0.0016 0.0017 0.0015 0.0224 0.0174 0.0184 ¸{ 0.1065 0.0932 0.1000 - - - - - -
The conclusions for these studies are given below.
We observe that for the known shape parameter α, averages of estimated values of the
parameters are very close to their true values and the averages of mean square errors are relatively
small. Further, we observe that the estimates and standard/mean square errors are decreasing
functions of number of systems � of each type put on test.
5.2 Unknown shape parameter
Similar study, with not much change in the algorithm, one can make simulation study for
the case of unknown shape parameter. We make simulation studies for the set of parameter
values � = 2, � = 2.5, = 1.5, d = 1.3, � = (0.4700,0.5424) and � = 3, � = 2.5, = 1.5, d = 1.3, e = 1.4,� = (0.4700,0.5424,0.5036) by taking n=1000.
16
Table 3
Maximum likelihood estimate of parameters, reliability and hazard rates and their efficiency measures Ó = Õ,c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô, Ú = (Û. ÙÜÛÛ, Û. ÖÙÕÙ), �(�) = (Û. Ö, Û. Ö), Þ(Ú) = (Õ. ÙÛßß, Õ. ÛààÝ) � �∗ �\ Ò dÒ � Ò(� ) �dÒ(�d) r �(� ) r\(�d)
12
06
{y 7.6968 1.2929 1.3116 0.4602 0.4559 3.7798 3.3771 ¶¸{ 483.34 0.2609 0.2044 0.0267 0.0296 10.2981 10.1162 ¸{ 39.47 0.4694 0.4134 - - - -
24
12
{y 3.8356 1.3949 1.2183 0.4813 0.4744 2.9270 2.5705 ¶¸{ 8.7988 0.1435 0.0970 0.0110 0.0112 1.5460 1.2587 ¸{ 2.7442 0.3482 0.3023 - - - -
36
18
{y 3.2538 1.4270 1.2373 0.4866 0.4854 2.7328 2.3744 ¶¸{ 3.2557 0.0869 0.0624 0.0065 0.0070 0.7267 0.5515 ¸{ 1.5771 0.2859 0.2474 - - - -
48
24
{y 3.0131 1.4484 1.2575 0.4907 0.4894 2.6355 2.2883 ¶¸{ 1.6604 0.0678 0.0485 0.0047 0.0048 0.4568 0.3386 ¸{ 1.1334 0.2505 0.2170 - - - -
60
30
{y 2.8370 1.4723 1.2734 0.4927 0.4935 2.5645 2.2208 ¶¸{ 0.9501 0.0519 0.0376 0.0037 0.0038 0.3027 0.2347 ¸{ 0.8982 0.2270 0.1965 - - - -
72
36
{y 2.8082 1.4614 1.2727 0.4959 0.4926 2.5484 2.2211 ¶¸{ 0.8139 0.0477 0.0324 0.0033 0.0033 0.2687 0.2107 ¸{ 0.8030 0.2053 0.1787 - - - -
84
42
{y 2.7971 1.4622 1.2738 0.4943 0.4907 2.5464 2.2219 ¶¸{ 0.6105 0.0384 0.0271 0.0025 0.0029 0.2006 0.1695 ¸{ 0.7198 0.1896 0.1648 - - - -
96
48
{y 2.6923 1.4838 1.2859 0.4956 0.4959 2.5015 2.1663 ¶¸{ 0.4429 0.0306 0.0245 0.0023 0.0023 0.1701 0.1250 ¸{ 0.6354 0.1799 0.1562 - - - -
17
Table 4
Maximum likelihood estimate of parameters, reliability and hazard rates and their efficiency measures Ó = Ô,c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø.Ô, ×Ô = Ø. Ù,Ú = (Û. ÙÜÛÛ, Û. ÖÙÕÙ, Û. ÖÛÔÝ),�(�) = (Û. Ö, Û. Ö, Û. Ö), Þ(Ú) = (Õ. ÙÛßß, Õ. ÛààÝ, Õ. ÕÙßÔ) � �∗ �\ Ò dÒ eÒ � Ò(� ) �dÒ(�d) �eÒ(�e) r �(� ) r\(�d) -e�(�e)
24
08
{y 4.9530 1.3474 1.1634 1.2517 0.4468 0.4478 0.4477 3.4987 3.0095 3.2462 ¶¸{ 32.9732 0.1482 0.1087 0.1251 0.0199 0.0190 0.0192 5.9173 3.7950 4.1568 ¸{ 5.6021 0.3417 0.2946 0.3169 - - - - - -
36
12
{y 3.7464 1.3992 1.1971 1.2955 0.4617 0.4673 0.4660 3.0371 2.6188 2.8132 ¶¸{ 6.0686 0.0969 0.0683 0.0829 0.0110 0.0113 0.0102 1.5790 1.1968 1.2909 ¸{ 2.2504 0.2875 0.2451 0.2657 - - - - - -
48
16
{y 3.3883 1.4173 1.2222 1.3097 0.4696 0.4740 0.4761 2.8825 2.4756 2.6624 ¶¸{ 3.2677 0.0671 0.0550 0.0625 0.0079 0.0074 0.0077 0.9885 0.6801 0.8306 ¸{ 1.5922 0.2503 0.2165 0.2315 - - - - - -
60
20
{y 3.1061 1.4421 1.2493 1.3409 0.4804 0.4800 0.4824 2.7300 2.3677 2.5397 ¶¸{ 1.8389 0.0574 0.0416 0.0484 0.0056 0.0060 0.0058 0.6293 0.4656 0.5389 ¸{ 1.2094 0.2286 0.1979 0.2123 - - - - - -
72
24
{y 3.0412 1.4359 1.2477 1.3457 0.4828 0.4814 0.4851 2.6981 2.3424 2.5212 ¶¸{ 1.3220 0.0448 0.0336 0.0419 0.0047 0.0047 0.0047 0.4747 0.3555 0.4222 ¸{ 1.0363 0.2065 0.1795 0.1939 - - - - - -
84
28
{y 2.8649 1.4628 1.2744 1.3692 0.4894 0.4869 0.4874 2.5988 2.2585 2.4355 ¶¸{ 0.8869 0.0395 0.0311 0.0337 0.0037 0.0036 0.0041 0.3370 0.2440 0.3134 ¸{ 0.8748 0.1956 0.1706 0.1832 - - - - - -
96
32
{y 2.8502 1.4554 1.2650 1.3650 0.4908 0.4898 0.4883 2.5903 2.2454 2.4244 ¶¸{ 0.7773 0.0329 0.0260 0.0307 0.0035 0.0032 0.0032 0.3053 0.2125 0.2584 ¸{ 0.8057 0.1816 0.1579 0.1705 - - - - - -
In the presence of unknown shape parameter�, from Table 3 and Table 4 it is seen that
the MLEs of scale parameters; � = 1,2, … ,�, the reliability characteristics and hazard rate are
getting estimated closely to their true values. However the convergence rate is slow compared to
the convergence rate when shape parameter� is known. Perhaps, it may be the effect of estimate
of unknown shape parameter�. Further we can say, slightly large sample size is required than
what we consider for the estimates to reach their true values.
5.3 Testing of Hypothesis under GIED
The hypothesis given in equation 4.1 can be tested for GIED by simply substituting
equation (5.6-5.7) in (4.2-4.8) we have required equations. Details are omitted to avoid
repetitions. The likelihood equations are not mathematically tractable for known as well as
unknown shape parameter we use the Newton-Raphson method to obtain the estimate of
parameter . Here we deal with only known shape parameter.
18
We demonstrate the test procedure for � = 2 and � = 3 . We generate data under our
design for the parameter values under H1:� = 2.5, = 1.5, d = 1 and
H1:� = 2.5, = 1.9, d = 1.5, e = 1 respectively. Then carryout the test procedure as
suggested above. The procedure is repeated for the different choices of u and �∗. The results are
produced in the table 5 and table 6 respectively.
Table 5
Likelihood Ratio Test for testing áÛ: ×Ø = ×Õ = ×âãáØ: ×Ø ≠ ×Õ when c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô
ä å∗ ×p ×ØÒ ×ÕÒ ææáÛ ææáØ çÕ p-value
12 06 1.3358 1.6602 1.0435 19.89 21.17 2.58 0.1085
24 12 1.2360 1.4820 1.0064 55.40 57.20 3.62 0.0573
36 18 1.2892 1.5217 1.0778 107.65 109.76 4.22 0.0399
48 24 1.2828 1.4946 1.0838 146.05 148.55 4.99 0.0255
60 30 1.2782 1.5000 1.0698 188.27 191.75 6.95 0.0083
72 36 1.2719 1.4907 1.0679 258.68 262.65 7.95 0.0048
84 42 1.2668 1.5241 1.0299 295.13 301.63 13.00 0.0003
96 48 1.2579 1.5139 1.0208 348.38 355.89 15.02 0.0001
Table 6
Likelihood Ratio Test for testing áÛ: ×Ø = ×Õ = e = ×âãáØ: ×è ≠ ×é(è ≠ é = Ø, Õ, Ô) when c = Õ. Ö, ×Ø = Ø. ß, ×Õ = Ø. Ö, ×Ô = Ø
ä å∗ ×p ×ØÒ ×ÕÒ ×ÔÒ ææáÛ ææáØ çÕ p-value
24 08 1.5648 2.0149 1.6065 1.1429 58.28 61.84 7.13 0.0283
36 12 1.4565 1.9386 1.4481 1.4481 102.26 108.26 12.35 0.0021
48 16 1.6085 2.0605 1.6637 1.6637 152.59 159.92 14.68 0.0006
60 20 1.5570 1.9373 1.6765 1.6765 204.43 213.22 17.58 0.0001
72 24 1.5060 1.9868 1.5475 1.5475 253.28 266.30 26.04 0.0000
84 28 1.4195 1.9031 1.4848 1.4848 300.09 317.54 35.00 0.0000
96 32 1.4345 1.9097 1.4978 1.4978 356.75 376.75 40.14 0.0000
From the Table 5, we infer that the power the test is poor for small sizes. However as the
sample size becomes 36(total number failures observed irrespective type of systems) it exhibits
its power in identifying the alternative. It requires more sample size to detect small departure
from homogeneity. From Table 6, it can reveal that for comparing homogeneity of three systems,
as sample size becomes 24 it exhibits its power to identify magnitude of departure from
homogeneity. The consistency of the test is also inferred as sample size tends to 96 the p value
becomes almost zero up to three digits in both Table 5 and Table 6.
19
6. The cost function
Suppose we start life testing experiment with � machines each from m makes or brands.
Further, we shall assume that cost of failure of a system is constant, say ê , irrespective of the
brand. Since the experiment is terminated after observing total G=m G* where G
* is a fixed
number failures being observed on each brand of systems. Then the total cost due to failure
is� = mê G∗ which is fixed and pre-planned. In this scheme there are � separate sub
experiments in the sense that the way we observe the machines of type � does not depend upon
the failures of machines of type �ì, for � ≠ �ì. Let �=∗; � = 1,2… ,� denote the time of failure of �∗th machine of the �th type. Define �Díî by �Díî = �²ï(�=∗ , �=∗d, … , �=∗D), (6.1)
So that �Díî , denotes the duration of experiment for which the whole experiment under Type II
censoring lasts. Under this experimental scheme, the component of cost from the view point of
length of the experiment, is êe�Díî. Now we shall consider the cost related with total experimental time for all the units put
under test. Supposeêd is the cost per unit time associated with testing time of a machine,
irrespective of the brand. Let � , �d, … , �=∗ be the failure times of �∗ systems out of � systems
of type�, � = 1,2, . . . . , �. Then the total time the on test for the ith
sub experiment is �� +(� − 1)(�d − � ) + (� − 2)(�e − �d) + ⋯+ (� − �∗ + 1)g�=∗ − �(=∗, )h. Hence, the total
machine time under Type II censoring is �d where
�d = ∑ � + �d +⋯ .+(� − �∗ + 1)�=∗C (6.2)
The contribution of this cost is �dêd.
The fourth component of the cost relates to the total number of machines used in the
whole experiment, which under Type II censoring equal to �� = ð. This cost deals with the
actual procurement, storage and handling of the machines. We shall assume thatñ(ð), an
increasing functions of ð, is the cost of putting N machines on experiment. In some cases, ñ(ð) may increase slower than ð or ñ may even be discontinuous.
Finally, let êt denote the overhead cost associated with conducting the experiment. We
shall assume that êt represents general costs such as those associated with planning the
experiment, administrative and consultancy costs, etc. However, we make the assumption that êtis independent ofð, the total number of machines used in the experiment; such costs are
already included in ñ(ð). With the above assumptions, the actual cost of the experiment under
20
Type II censoring design is ê = êt + ê � + êd�d + êe�Díî + ñ(ð) (6.3)
We carryout Monte-Carlo simulation study for the choice of cost êt = 100, ê = 5, êd = 10, êe = 10 and ñ(ð) = 0.5ð for the cost effectiveness of Type II censoring scheme with different
censoring proportions. We consider � = 2 and � = 3 and results are given in Table 7 and
Table 8 respectively. Further, we fix total number of failures � and change the number of units
to be tested at each sub experiment. The results are given in Table 9.
It can be seen from Table 7 and Table 8, total time under experiment, duration of test and
cost associated with experiments in Type II experiments are comparatively less than no
censoring for m=2 as well as m=3. Therefore, in terms of cost, Type II censoring is
recommended for comparison of several systems.
From Table 9 and Table 10 we can see that as proportion of censoring increases when
total number of failures is fixed, the total time on test also increases linearly and duration time of
the experiment decreases. The facts can be seen from Figure 1 and Figure 2. Further, from Table
11 we can observe that as proportion of censoring increases when number of units to be tested is
fixed, the total time on test, duration of test and the total cost of experiment are decreasing while
mean square error of estimates are increasing.
21
Table 7
Comparative Table for total time under experiment and duration of experiment
when no censoring and 1/2 Censoring Ó = Õ,c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô
No Censoring ½ Censoring (å/Ó) ä å òÕ òóôõ ö òÕ òóôõ ö
12 12 18.0732 2.9559 442.291 9.9583 0.5272 382.291
24 24 35.4111 3.5768 753.879 19.8305 0.5302 633.879
36 36 53.5503 4.3074 1074.577 29.9269 0.539 894.577
48 48 71.3685 4.8756 1390.441 39.833 0.5388 1150.441
60 60 89.5124 5.4885 1710.009 49.7208 0.5372 1410.009
72 72 107.6014 5.9442 2027.456 59.7131 0.538 1667.456
84 84 125.1195 6.209 2337.285 69.8531 0.5423 1917.285
96 96 143.2841 6.7334 2656.175 79.4327 0.5393 2176.175
Table 8
Comparative Table for total time under experiment and duration of experiment
when no censoring and 1/3 Censoring Ó = Ô,c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô, ×Ô = Ø. Ù
No Censoring 2/3 Censoring (G/m) ä å òÕ òóôõ ö òÕ òóôõ ö
12 12 26.6254 2.6323 590.577 12.1862 0.3952 303.78
24 24 53.7196 3.607 1069.266 24.4007 0.4008 509.386
36 36 80.7912 4.2592 1544.504 36.8109 0.3974 706.117
48 48 106.9302 5.1795 2013.097 48.9774 0.3999 903.276
60 60 134.4159 5.5035 2489.194 61.1037 0.4023 1105.037
72 72 161.7802 6.293 2968.732 73.2893 0.4041 1304.934
84 84 187.9244 6.6052 3431.296 85.653 0.4035 1506.565
96 96 215.6138 6.4423 3904.561 97.9054 0.4036 1707.09
22
Table 9
Total time under experiment, duration of experiment and cost of experiment for fixed failures Ó = Ô,å = ØÕ, c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô, ×Ô = Ø. Ù ä òÕ òóôõ ö ä òÕ òóôõ ö
4 8.718 1.574 228.92 10 11.0219 0.4387 249.606
5 8.6535 0.8871 222.906 11 11.6579 0.4155 257.234
6 8.9423 0.6879 225.302 12 12.1182 0.3897 263.079
7 9.4497 0.5848 230.845 13 12.7585 0.377 270.855
8 10.0057 0.5146 237.203 14 13.1883 0.3582 276.465
9 10.6023 0.4851 244.374 15 13.8141 0.3506 284.147
Table 10
Total time under experiment, duration of experiment and cost of experiment for fixed failures Ó = Ô,å = ÕÙ, c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô, ×Ô = Ø. Ù(÷ = Ø − åÓä : proportion of censoring)
ä ÷ Ò dÒ eÒ � Ò(� ) �dÒ(�d) �eÒ(�e) r � (� ) r\(�d) -e�(�e) òÕ òóôõ ö
24 0.67 EV 1.5015 1.307 1.4182 0.5015 0.5048 0.5004 2.4031 2.0719 2.2465 24.2905 0.3987
502.892
MSE 0.0561 0.039 0.0454 0.0074 0.007 0.0068 0.1049 0.0745 0.0454 - - -
27 0.7 EV 1.5124 1.299 1.4118 0.5023 0.508 0.5025 2.4001 2.0614 2.2395 26.1149 0.3766
525.415
MSE 0.0462 0.0375 0.0416 0.0062 0.0069 0.0064 0.0879 0.0736 0.0792 - - -
30 0.73 EV 1.5011 1.3034 1.4169 0.506 0.5052 0.5 2.3866 2.0709 2.2486 27.7067 0.3537
545.604
MSE 0.0422 0.0326 0.0391 0.0058 0.0061 0.0059 0.0814 0.0641 0.0726 - - -
33 0.76 EV 1.5028 1.3214 1.3988 0.5054 0.4974 0.507 2.3887 2.0964 2.2241 29.3456 0.3339
566.295
MSE 0.0422 0.033 0.037 0.0058 0.0057 0.0059 0.0816 0.0608 0.0723 - - -
36 0.78 EV 1.5147 1.3058 1.4141 0.5004 0.5037 0.5006 2.4073 2.0757 2.2463 30.9336 0.3254
586.59
MSE 0.0395 0.0306 0.0366 0.0053 0.0056 0.0055 0.0745 0.0593 0.0665 - - -
39 0.79 EV 1.5053 1.3137 1.4065 0.5037 0.499 0.5028 2.3153 2.0909 2.2389 32.4637 0.3107
606.244
MSE 0.0374 0.0321 0.0308 0.0051 0.0056 0.0048 0.0716 0.0593 0.0585 - - -
42 0.81 EV 1.504 1.3115 1.4113 0.5042 0.5003 0.5013 2.3935 2.087 2.2442 34.0454 0.3009
626.463
MSE 0.0374 0.0206 0.0328 0.0051 0.0046 0.005 0.0715 0.0485 0.0614 - - -
45 0.82 EV 1.511 1.312 1.4141 0.5009 0.5 0.4999 2.4058 2.0879 2.2501 35.5085 0.2934
645.519
MSE 0.0333 0.0256 0.0299 0.0044 0.0046 0.0046 0.0627 0.0486 0.0563 - - -
48 0.83 EV 1.5126 1.3138 1.4092 0.5006 0.4992 0.5016 2.4068 2.0905 2.243 37.0428 0.2854
665.282
MSE 0.0353 0.0255 0.0305 0.0048 0.0046 0.0046 0.0668 0.0481 0.0502 - - -
23
Fig. 1:÷ vs total time under test
Fig. 2:÷ vs duration of test
Table 11
Maximum likelihood estimate of parameters, reliability and hazard rates and their efficiency measures, total
time under experiment, duration of experiment and cost of experiment for fixed failures Ó = Ô, ä = ÕÙ, c = Õ. Ö, ×Ø = Ø. Ö, ×Õ = Ø. Ô, ×Ô = Ø. Ù(÷ = Ø − å(CÓå∗)Óä : proportion of censoring)
å∗ ÷ Ò dÒ eÒ � Ò(� ) �dÒ(�d) �eÒ(�e) r �(� ) r\(�d) -e�(�e) òÕ òóôõ ö
2 0.92 EV 1.5465 1.3527 1.4369 0.4997 0.4918 0.5015 2.4098 2.1139 2.2425 14.5499 0.2195
343.69
MSE 0.1046 0.0768 0.0999 0.0120 0.0118 0.0106 0.1718 0.1266 0.1661 - - -
3 0.88 EV 1.5281 1.3185 1.4488 0.5014 0.5030 0.4932 2.4030 2.0776 2.2716 16.7091 0.2546
395.64
MSE 0.0817 0.0559 0.0977 0.0101 0.0010 0.0106 0.1446 0.1022 0.1313 - - -
4 0.83 EV 1.5199 1.3141 1.4306 0.5026 0.5045 0.4979 2.3987 2.0730 2.2716 18.4367 0.2820
443.19
MSE 0.0678 0.0527 0.0601 0.0088 0.0093 0.0086 0.1250 0.0994 0.1060 - - -
6 0.75 EV 1.5297 1.3126 1.4111 0.4983 0.5035 0.5045 2.4154 2.0761 2.2321 21.5628 0.3410
535.04
MSE 0.0635 0.0446 0.522 0.0079 0..80 0.0080 0.1114 0.0849 0.0985 - - -
8 0.67 EV 1.5040 1.3133 1.4128 0.5064 0.5023 0.5027 2.3846 2.0801 2.2387 24.3820 0.3986
623.81
MSE 0.0522 0.0403 0.0452 0.0071 0.0071 0.0073 0.1009 0.0757 0.0848 - - -
9 0.63 EV 1.5078 1.3057 1.4107 0.5042 0.5053 0.5032 2.3933 2.0703 2.2371 25.7072 0.4278
667.35
MSE 0.0501 0.0403 0.0452 0.0068 0.0071 0.0072 0.0967 0.0754 0.0841 - - -
10 0.58 EV 1.4993 1.3025 1.3971 0.5077 0.5063 0.5089 2.3798 2.0673 2.2171 27.1351 0.4611
711.96
MSE 0.0489 0.0359 0.0438 0.0068 0.0066 0.0071 0.0962 0.0738 0.0821 - - -
12 0.50 EV 1.5033 1.3120 1.4000 0.5060 0.5017 0.5069 2.3865 2.0825 2.2240 29.9872 0.5369
801.24
MSE 0.0471 0.0340 0.0396 0.0065 0.0060 0.0063 0.0920 0.0636 0.0778 - - -
24
7. Concluding remarks
In this article we obtain maximum likelihood estimates of parameters of m lifetimes
distributions using a generalization of Type II censoring scheme, proposed by Srivastava (1989),
under the assumption of generalized inverted scale family of distributions. We consider
generalized inverted family of exponential distribution (GIED) as an example for the more
general procedure we discuss. Then we obtain the estimates of parameters involved in the
distributions, reliability and hazard rate at a given point. The efficiency measures like MSE and
SE are simulated. The likelihood ration test is discussed to test homogeneity of lifetimes of
several brands. Finally we have studied the cost effectiveness of the Type II censored sampling
scheme, through Monte-Carlo simulation. Expressions given in this article can also be used for
generalized inverted half-logistic distribution and generalized inverted Rayleigh distribution etc.
Note: We have written a subroutine to compute the test statistics in R language. The source code
is available on request.
Acknowledgement:
The authors would like to thank the referee for their valuable comments and suggestions which
helped us to improve the earlier version of this paper. The second author would like to thank
University Grant Commission (UGC) for financial support. This research is funded by
University Grant Commission (UGC) for Minor Research Project (F.37-549/2009(SR)).
References
Abouammoh ,A. M. and Alshingiti ,M. A.. Reliability estimation of generalized inverted
exponential distribution. Journal of Statistical Computation and Simulation, 79:1301-1315, 2009.
Balakrishnan, N. and Ng, H.K.T. (2006). Precedence-type tests and applications. Hoboken,
NJ:John Wiley & Sons
Cohen, A.C. Jr (1965), “Maximum likelihood estimation with Weibull distribution based on
complete and on censored samples”, Technometrics, 7, pp. 559-88.
Gupta, A.K. (1952), “Estimation of mean and standard deviation of a normal population from a
censored sample”, Biometrika, 39, pp. 260-73.
25
Hossain, A. and Zimmer, W. (2003). “Comparison of estimation methods for Weibull
parameters: Completed censored samples, Journal of Statistical Computation and Simulation.
73(2).145-153
Lawless, J.F. (1982). Statistical Models and Methods for Lifetime Data, John Wiley & Sons,
New York, NY.
Mann, N., Schafer, E. and Singpurwalla, N. (1974), Methods for Statistical Analysis of
Reliability and Life Data, John Wiley & Sons, New York, NY.
Ng, H.K.T. (2005). Parameter estimation for a modified Weibull distribution for progressive
Type II censored samples, IEEE Trans.Reliab. 54, 374-380.
Poda, K.G. and Shirke, D.T. (2013). Inference for the parameters of generalized inverted family
of distributions. ProbStat Forum, 06, pp. 18-28.
Shafay, A.R., Balkrsihnan, N. And Abdel-Aty, Y. (2013). “Bayesian inference based on a jointly
Type II censored samples from two exponential population”. Journal of Statistical Computation
and Simulation. http:/dx/xoi.org//10.1080/00949655.2013.813 025
Sharafi, M.m Balkrishnan, N. And Khaledi, B.E. (2013). “Distribution-free comparison of
hazard rates of two distributions under Type II censoring”. Communications in Statistics-Theory
and Methods, 42, 1889-1898
Sinha, S.K. (1986), Reliability and Life Testing, Wiley Eastern Limited, New Delhi.
Srivastava, J.N. (1987). Srivastava,J.N. (1987). More efficient and less time-consuming
censoring designs for life testing, J.Statist.Plann. Inference, Vol. 16, pp. 389-413
Srivastava, J.N.(1989). Advances in the statistical theory of comparison of lifetimes of machines
under the generalized Weibull distribution. Commun.Statist.-Theory Meth. 18 (3), 1031-1045