You are on page 1of 49

Test (2007) 16: 211–259

DOI 10.1007/s11749-007-0061-y
I N V I T E D PA P E R

Progressive censoring methodology: an appraisal

N. Balakrishnan

Published online: 28 June 2007


© Sociedad de Estadística e Investigación Operativa 2007

Abstract Properties of progressively censored order statistics and inferential proce-


dures based on progressively censored samples have recently attracted considerable
attention in the literature. In this paper, I provide an overview of various developments
that have taken place in this direction and also suggest some potential problems of
interest for further research.

Keywords Progressive censoring · Order statistics · Life-testing experiment ·


Bounds · Generalized order statistics · Characterizations · Markov property ·
Likelihood inference · Reliability sampling plans · Goodness-of-fit tests ·
Prediction · Competing risks · Step-stress test · Hybrid censoring · Permanents ·
Outliers · Robustness

Mathematics Subject Classification (2000) 62G30 · 62E99 · 62F10 · 62N01 ·


62N05

1 Introduction

1.1 History

In his Ph.D. thesis entitled Estimation of the parameters of a population from a


multi-censored sample, Herd (1956) was the first to discuss estimation of the pop-

This invited paper is discussed in the comments available at:


http://dx.doi.org/10.1007/s11749-007-0062-x, http://dx.doi.org/10.1007/s11749-007-0063-9,
http://dx.doi.org/10.1007/s11749-007-0064-8, http://dx.doi.org/10.1007/s11749-007-0065-7,
http://dx.doi.org/10.1007/s11749-007-0066-6, http://dx.doi.org/10.1007/s11749-007-0067-5,
http://dx.doi.org/10.1007/s11749-007-0068-4, http://dx.doi.org/10.1007/s11749-007-0069-3,
http://dx.doi.org/10.1007/s11749-007-0070-x, http://dx.doi.org/10.1007/s11749-007-0071-9.
N. Balakrishnan ()
Department of Mathematics and Statistics, McMaster University, Hamilton, ON, Canada L8S 4K1
e-mail: bala@mcmaster.ca
212 N. Balakrishnan

ulation parameters based on progressively censored samples, even though he referred


to them as “multi-censored samples”. A few years later, Cohen (1963) discussed
the importance of progressive censoring in life-testing reliability experiments. While
these two authors were primarily concerned with progressive censoring as an useful
sampling mechanism in reliability experiments and developing inferential procedures
that could be used to analyze such data, an independent query that appeared in 1966
(Query 18, Technometrics, August 1966) presented another scenario, wherein a pro-
gressively censored sample could arise naturally:
It is not uncommon in our life-testing for items to fail for reasons quite un-
related to the normal failure mechanism. For example, consider a number of
lamps placed simultaneously on life-test. One of the lamps might be acciden-
tally broken after the start of the test but before all the lamps had burned out.
If all lamps but one had burned out and the last were accidentally broken, the
population parameters are easily estimated by techniques designed to deal with
censoring on the right. Breakage of any lamp but the longest lived one in the
sample, however, introduces the problem of how to utilize the information that
this lamp burned the observed number of hours before it was destroyed. What
procedures can be recommended?
While this query enquired as to how such an accidental breakage of an unit under
test can be accommodated into inference, Cohen pointed out (in a subsequent issue
of Technometrics), in response to this query, that progressive censoring methodology
is the ideal methodology for this situation.

1.2 Concepts and notation

Consider a reliability experiment in which n identical units are placed on a life-test.


Let X1 , X2 , . . . , Xn denote the lifetimes of these experimental units. As usual, we
shall assume that these variables are independent and identically distributed from
an absolutely continuous population with cumulative distribution function F (x) and
probability density function f (x). Let X1:n ≤ X2:n ≤ · · · ≤ Xn:n denote the corre-
sponding ordered lifetimes, which are the usual order statistics. In life-testing exper-
iments, it is, of course, natural for the lifetime data to arrive in the form of these order
statistics.
Suppose the experimenter decides to carry out the life-test for a pre-fixed length
of time, say T . Then the data arising from such a time-constrained life-test would be
of the form X1:n ≤ · · · ≤ Xm:n with the remaining n − m lifetimes being more than
T ; here, of course, m is random (0 ≤ m ≤ n) and, in fact, has a binomial distribution
with parameters (n, F (T )). This situation is referred to as Type-I censoring.
If the experimenter decides to carry out the life-test until the time of the mth fail-
ure, then the data arising from such a life-test would be of the form X1:n ≤ · · · ≤ Xm:n
with the remaining n − m lifetimes being more than Xm:n . This situation is referred to
as Type-II censoring. It is clear that there is a fundamental difference between these
two censoring schemes in that in the case of Type-I censoring, the duration of the
life-test is fixed and the number of failures is random, while in the case of Type-II
censoring the duration of the life-test is random and the number of failures is fixed.
Progressive censoring methodology: an appraisal 213

Note that in both these cases the censoring occurs on the right. Though not very
common in reliability experiments (but occurring often in survival analysis), left cen-
soring can easily be introduced into these two censoring models.
After observing that the Type-II censoring, described above, results in m complete
failures and n − m censored observations, a natural extension of it will be a censoring
scheme in which n − m units are withdrawn from the life-test at different time points
(rather than all at the time Xm:n ). Such a versatile censoring scheme is what is re-
ferred to as progressive Type-II right censoring. Specifically, consider the situation in
which n identical units, with lifetime cumulative distribution function (cdf) F (x) and
probability density function (pdf) f (x), are placed on a life-testing experiment. Then,
immediately following the first failure, R1 surviving units are removed from the test
at random; next, immediately following the second observed failure, R2 surviving
units are removed from the test at random, and so on; finally, at the time of the mth
observed failure, all the remaining Rm = n − R1 − · · · − Rm−1 − m surviving units are
removed from the test. In this setup, the number of complete failures to be observed
(m) and the progressive censoring scheme (R1 , . . . , Rm ) to be carried out are all as-
sumed to be pre-fixed. As the censoring times are all random here and the numbers
of items to fail before each censoring time are all fixed, this scheme is said to be pro-
gressive Type-II right censoring, and the ordered values obtained as a consequence
of this type of censoring are referred to as progressively Type-II right censored order
(R1 ,...,Rm ) (R1 ,...,Rm )
statistics. These are denoted by X1:m:n , . . . , Xm:m:n , and sometimes simply
by X1:m:n , . . . , Xm:m:n if there is no confusion in the context of that discussion. A
book-length account of the developments on progressive Type-II censoring is due to
Balakrishnan and Aggarwala (2000).
It is important to mention here that the definition and observational scheme given
above for progressively Type-II right censored order statistics is valid only for contin-
uous parent distributions. In the case of discrete parent distributions, however, when
we censor R1 units immediately following the first failure, these censored units need
not all be surviving units; if two or more units failed at the time of the first failure
(since ties are possible in this discrete case), then the R1 units censored may also
include some of these failed units. This naturally introduces complications in the cor-
responding distribution theory; see Sect. 6 for details.
Just as in the case of ordinary Type-I censoring, time-based censoring can be in-
troduced in the framework of progressive censoring as well. For example, the exper-
imenter may pre-fix m censoring times T1 , T2 , . . . , Tm . At times T1 , T2 , . . . , Tm−1 ,
from the life-test R1 , R2 , . . . , Rm−1 surviving units are randomly removed or cen-
sored, respectively; here, Ri ’s (i = 1, . . . , m − 1) are fixed with the condition that
there are at least Ri surviving units at censoring time Ti . Finally, the life-test is ter-
minated at time Tm with Rm surviving units at that time. This scheme is clearly a
generalization of the Type-I right censoring described earlier, and is, therefore, re-
ferred to as progressive Type-I right censoring. Though this scheme is more practical
and useful from the point of view of designing a life-testing experiment, it poses
great difficulties in studying the theoretical properties of ordered failure times arising
from such a progressive Type-I right censoring scheme. This is due to the fact that
there is a possibility that the life-test may terminate before reaching the mth stage of
censoring due to the random nature of the failures occurring within each time inter-
val. Consequently, most of the inferential work carried out in the literature based on
214 N. Balakrishnan

this progressive Type-I right censoring scheme is numerical in nature and is, in some
sense, conditional on having observed such a failure data.
In this discussion, therefore, I will focus on the progressive Type-II right censoring
situation and will present an overview of various developments relating to this case.
For the sake of completeness, I will also present a brief discussion on the progressive
Type-I right censoring scheme and some related work. Of course, in preparing this
sort of an overview article, I had to be selective on the topics for discussion and, in
doing so, may have omitted some articles containing important results; if so, it should
be considered as a slip on my part or as a consequence of my ignorance, and not due
to personal nonscientific antipathy.

2 Basic distribution theory

1 (R ,...,Rm )
The joint density of the progressively censored order statistics, X1:m:n ,...,
(R1 ,...,Rm )
Xm:m:n , is given by Balakrishnan and Aggarwala (2000, p. 8) as


m
 R
fX1:m:n ,...,Xm:m:n (x1 , . . . , xm ) = C f (xi ) 1 − F (xi ) i , x1 < · · · < xm , (1)
i=1

where

C = n(n − R1 − 1)(n − R1 − R2 − 2) · · · (n − R1 − · · · − Rm−1 − m + 1).

Note that the constant C, in addition to being the normalizing constant, also rep-
resents the number of ways in which the m progressively Type-II right censored
order statistics may occur if the observed failure times are x1 , . . . , xm as there are
n units on test immediately preceding the first observed failure, n − R1 − 1 units
on test immediately preceding the second observed failure, and so on, and finally
n − R1 − · · · − Rm−1 − m + 1 units still on test immediately preceding the mth ob-
served failure.
Suppose the failure times of the first r units to failure were not observed and that
Rr+i surviving units are randomly withdrawn (or censored) from the life-test at the
time of the (r + i)th failure for i = 1, 2, . . . , m − r. The lifetimes of these m − r
(Rr+1 ,...,Rm ) (Rr+1 ,...,Rm )
completely observed failures, denoted by r Xr+1:m:n , . . . , r Xm:m:n , are the
so-called general progressively Type-II censored order statistics. Under this setting, it
(Rr+1 ,...,Rm )
is evident that n = m + Rr+1 + · · · + Rm and r Xr+1:m:n is distributed as the (r +
1)th order statistic from a sample of size n from the distribution F (x). If the failure
times are assumed to arise from a continuous population with cdf F (x) and pdf f (x),
then the joint density of these general progressively Type-II censored order statistics
is given by Balakrishnan and Sandhu (1996) and Aggarwala and Balakrishnan (1998)
as
 r 
m
 R
fr Xr+1:m:n ,...,r Xm:m:n (xr+1 , . . . , xm ) = C ∗ F (xr+1 ) f (xi ) 1 − F (xi ) i ,
i=r+1

xr+1 < · · · < xm , (2)


Progressive censoring methodology: an appraisal 215

where C ∗ is the normalizing constant. Even though this general progressive Type-II
censoring may not be often encountered in practice, many of the theoretical properties
and inferential procedures based on the progressively Type-II right censored order
statistics can be generalized to this general progressive Type-II censoring situation.
So we refrain ourselves from elaborating on these details here in this discussion.
Progressively Type-II right censored order statistics form a special case of gener-
alized order statistics discussed by Kamps (1995a, 1995b).  In this regard, let n be a
natural number, k > 0, m1 , . . . , mn−1 be real, and Mr = n−1 j =r mj for 1 ≤ r ≤ n − 1
be parameters with γr = k + n − r + Mr > 0 for all r ∈ {1, 2, . . . , n − 1}; let us
also denote m̃ = (m1 , . . . , mn−1 ) for n ≥ 2 and m̃ real for n = 1. Then, the random
variables X(r, n, m̃, k) (for r = 1, . . . , n) with joint density function

fX(1,n,m̃,k),...,X(n,n,m̃,k) (x1 , . . . , xn )

n−1 
n−1 
n
∗∗
=C γj {1 − F (xi )} {1 − F (xn )}
mi k−1
f (xi ),
j =1 i=1 i=1

x1 < x2 < · · · < xn , (3)

are called “generalized order statistics”. It is readily seen that the progressively
Type-II right censored order statistics with joint density in (1), is a special case
of the generalized order statistics corresponding to the case n = m, mi = Ri (for
i = 1, . . . , m − 1), and k = Rm + 1. Consequently, the general theoretical properties
and inferential procedures based on these generalized order statistics established by
Kamps and other authors, also hold for the special case of progressively Type-II right
censored order statistics. Note, however, that the general progressively Type-II cen-
sored order statistics with joint density in (2) is not subsumed in the general model
in (3). We, therefore, refrain here from detailing the developments on generalized
order statistics, and just focus on results concerning the progressively Type-II right
censored order statistics only.

3 Dependence properties

The usual order statistics from a continuous population possess several interesting
dependence properties; see, for example, Arnold et al. (1992) and David and Nagaraja
(2003). Many of these properties are also satisfied by progressively Type-II right
censored order statistics, and there are some exceptions as well.

Result 1 Let X1:m:n , . . . , Xm:m:n denote the progressively Type-II right censored or-
der statistics from a continuous population with cdf F (x) and pdf f (x). Then:
(i) The marginal distribution of Xi:m:n , 1 ≤ i ≤ m, is free of (Ri , Ri+1 , . . . , Rm ).
(ii) (X1:m:n , . . . , Xi:m:n ) form a progressively Type-II right censored sample of
size i from n units on a life-test with the progressive censoring scheme
(R1 , . . . , Ri−1 , n − i − R1 − · · · − Ri−1 ).
216 N. Balakrishnan

Proof From the joint density of X1:m:n , . . . , Xm:m:n in (1), we obtain the joint density
of X1:m:n , . . . , Xi:m:n as

fX1:m:n ,...,Xi:m:n (x1 , . . . , xi )



i  ∞ ∞  ∞ 
m
 R  R
=C f (x ) 1 − F (x )  ··· f (xk ) 1 − F (xk ) k
=1 xi xi+1 xm−1 k=i+1

× dxm dxm−1 · · · dxi+1 . (4)

Upon using the fact that


 ∞
 R {1 − F (y)}Rk +1
f (x) 1 − F (x) k dx = ,
y Rk + 1

the expression in (4) reduces to

fX1:m:n ,...,Xi:m:n (x1 , . . . , xi )



i−1
 R f (xi ){1 − F (xi )}m−i+Ri +Ri+1 +···+Rm
=C f (x ) 1 − F (x )  × m−i  k 
=1 k=1 k + j =1 Rm−j +1


i−1
 R
= n(n − R1 − 1) · · · (n − R1 − · · · − Ri−1 − i + 1) f (x ) 1 − F (x ) 
=1
 n−i−R1 −···−Ri−1
× f (xi ) 1 − F (xi ) , x1 < · · · < xi . (5)

The expression of the joint density of (X1:m:n , . . . , Xi:m:n ) in (5) readily reveals that
it is free of (Ri , Ri+1 , . . . , Rm ), which proves Part (i). Part (ii) follows immediately
by comparing the joint density of (X1:m:n , . . . , Xi:m:n ) in (5) with the joint density of
progressively Type-II right censored order statistics in (1). 

Remark 1 Part (ii) of Result 1 readily reveals that


(R ,...,Rm ) d (R ,...,Ri−1 ,n−i−R1 −···−Ri−1 )
1
Xi:m:n = Xi:i:n
1
(6)

for i = 1, . . . , m. This is intuitively clear as what happens after the observation of the
ith progressively Type-II right censored order statistic should not affect its distribu-
tion.

Result 2 The progressively Type-II right censored order statistics (Xi+1:m:n , . . . ,


Xm:m:n ), given (X1:m:n , . . . , Xi:m:n ) = (x1 , . . . , xi ), are distributed exactly as progres-
sively Type-II right censored sample of size m − i from n − i − R1 − · · · − Ri units on
a life-test with density f (x) truncated on the left at xi with the progressive censoring
scheme (Ri+1 , . . . , Rm ).

Proof From the joint density of (X1:m:n , . . . , Xi:m:n ) and (X1:m:n , . . . , Xm:m:n ) in (5)
and (1), respectively, we readily obtain the conditional joint density of (Xi+1:m:n , . . . ,
Progressive censoring methodology: an appraisal 217

Xm:m:n ), given (X1:m:n , . . . , Xi:m:n ) = (x1 , . . . , xi ), as

fXi+1:m:n ,...,Xm:m:n |X1:m:n ,...,Xi:m:n (xi+1 , . . . , xm |x1 , . . . , xi )


fX1:m:n ,...,Xm:m:n (x1 , . . . , xm )
=
fX1:m:n ,...,Xi:m:n (x1 , . . . , xi )
 j −1
m 
m
f (xk ) 1 − F (xk ) Rk
= n− Rj − j + 1 ,
1 − F (xi ) 1 − F (xi )
j =i+1 k=1 k=i+1

xi < xi+1 < · · · < xm . (7)

Upon comparing (7) with (1), we observe that (7) is simply the joint density function
of m − i progressively Type-II right censored order statistics from n − i − R1 − · · · −
f (x)
Ri units on a life-test with density function 1−F (xi ) , x > xi , with the progressive
censoring scheme (Ri+1 , . . . , Rm ). Hence, the result. 

From the conditional joint density function of (Xi+1:m:n , . . . , Xm:m:n ), given


(X1:m:n , . . . , Xi:m:n ) = (x1 , . . . , xi ), derived in (7), we also readily observe the fol-
lowing result.

Result 3 The progressively Type-II right censored order statistics Xi:m:n , 1 ≤ i ≤ m,


from an arbitrary continuous distribution F (x) form a Markov chain.

Remark 2 In the case of usual order statistics, it is also known that the conditional
distribution of the ith order statistic from a sample of size n from a continuous dis-
tribution F (x), given that the j th order statistic equals xj (for i < j ≤ n), is exactly
the same as the distribution of the ith order statistic from a sample of size j − 1 from
the population with distribution F (x) truncated on the right at xj ; see, for example
Arnold et al. (1992) and David and Nagaraja (2003). This result will not hold in the
case of progressively Type-II right censored order statistics, as it is clearly evident
that the units censored before the observation of the j th progressively Type-II right
censored order statistic (given as xj ) may have had their lifetimes to be larger than
xj and so they can not possibly be from the distribution F (x) truncated on the right
at xj .

4 Different representations

Even though the joint density of progressively Type-II right censored order statistics
assumes a simple form as in (1), marginal distributions do become complicated. For
this reason, different representations have been provided in the literature.

4.1 Thomas–Wilson representation

Let X1:n < · · · < Xn:n denote the usual order statistics from a sample of size n
from a continuous population with cdf F (x) and pdf f (x), and X1:m:n < · · · <
218 N. Balakrishnan

Xm:m:n be a progressively Type-II right censored sample of size m from these n


order statistics with (R1 , . . . , Rm ) as the progressive censoring scheme. Then, each
Xj :m:n corresponds to some usual order statistic, i.e., Xj :m:n = XKj :n , where Kj ,
the rank of Xj :m:n in the original set of n order statistics, can take on the values
Kj −1 + 1, Kj −1 + 2, . . . , R1 + · · · + Rj −1 + j for j = 2, . . . , m, and K1 ≡ 1. Then
Thomas and Wilson (1972) showed that the joint probability mass function of the
rank vector (K1 , . . . , Km ) is given by

P (K1 = k1 , . . . , Km = km )

m
= P (K1 = k1 ) P (Ki = ki |K1 = k1 , . . . , Ki−1 = ki−1 ), (8)
i=2

with P (K1 = 1) = 1, where


n−ki 
i+R1 +···+Ri−1 −ki
P (Ki = ki |K1 = k1 , . . . , Ki−1 = ki−1 ) = n−ki−1 , 2 ≤ i ≤ m. (9)
i−1+R1 +···+Ri−1 −ki−1

Hence, if all possible rank vectors N (say) can be listed for the specified pro-
gressive censoring scheme (R1 , . . . , Rm ), then for each rank vector we can define an
m × n indicator matrix D for  = 1, 2, . . . , N whose (r, s)th element is 1 if s = Kr ,
and 0 otherwise; then, evidently,

(X1:m:n , . . . , Xm:m:n )T = D (X1:n , . . . , Xn:n )T (10)

for some  ∈ {1, 2, . . . , N}. The mixture representation in (10) may be used in con-
junction with (8) and (9) in order to compute distributions and moments of pro-
gressively Type-II right censored order statistics as mixtures of the correspond-
ing quantities for the usual order statistics. For example, with XPCOS denoting
(X1:m:n , . . . , Xm:m:n )T , XOS denoting (X1:n , . . . , Xn:n )T , and μPCOS and μOS in place
of the mean vectors of XPCOS and XOS , respectively, we immediately obtain from
(10) that
N
μPCOS = E E(XPCOS |D ) = E(D μOS ) = π D μOS ,
=1
where π is the probability mass function of the rank vector corresponding to the
indicator matrix D (for  = 1, . . . , N ).

4.2 Kamps–Cramer representation

Starting with the joint density of generalized order statistics presented in (3), Kamps
and Cramer (2001) derived a simpler expression for the marginal distribution not by
carrying out the necessary integration, but by exploiting some properties of exponen-
tial generalized order statistics. With
m
γr = m − r + 1 + Ri , for r = 1, . . . , m,
i=r
Progressive censoring methodology: an appraisal 219


r
cr−1 = γi , for r = 1, . . . , m,
i=1

and

r
1
ai,r = , for 1 ≤ i ≤ r ≤ m,
γj − γi
j =1
j =i

their expression for the cdf of the rth progressively Type-II right censored order sta-
tistic becomes
r
ai,r  γ
FXr:m:n (x) = 1 − cr−1 1 − F (x) i , for r = 1, . . . , m. (11)
γi
i=1

Upon differentiating the expression of FXr:m:n (x) with respect to x, we readily obtain
the pdf of Xr:m:n as
r
 γ −1
fXr:m:n (x) = cr−1 ai,r 1 − F (x) i f (x), for r = 1, . . . , m. (12)
i=1

Kamps and Cramer (2001) have also presented a similar expression for the joint den-
sity of Xr:m:n and Xs:m:n , for 1 ≤ r < s ≤ m.

4.3 Guilbaud representation

Guilbaud (2001) presented another mixture representation that expresses the progres-
sively Type-II right censored order statistics as mixtures of the usual order statis-
tics. As in Sect. 4.1, let X1:n < · · · < Xn:n denote the usual order statistics from a
sample of size n from a continuous population with cdf F (x) and pdf f (x), and
X1:m:n < · · · < Xm:m:n be a progressively Type-II right censored sample of size m
from these n order statistics with (R1 , . . . , Rm ) as the progressive censoring scheme.
Then Guilbaud (2001) has shown that, for i = 1, . . . , m,
n
Xi:m:n = ŵij Xj :n (13)
j =1

with probability one, where (ŵi1 , . . . , ŵin ) is a random vector independent of


(X1:n , . . . , Xn:n ) having a multinomial distribution with parameters 1 and
(wi1 , . . . , win ). In particular, the representation in (13) yields
n
P (Xi:m:n ∈ B) = wij P (Xj :n ∈ B), (14)
j =1

for any Borel set B. Guilbaud (2001) has presented a recursive computational proce-
dure for the mixing probabilities wij ’s.
220 N. Balakrishnan

Remark 3 In fact, a mixture representation for the entire progressively Type-II cen-
sored sample (X1:m:n , . . . , Xm:m:n ) in terms of the usual order statistics
(X1:n , . . . , Xn:n ) has been given by Guilbaud (2001) who has also discussed sto-
chastic behavior of the m × n weight matrix Ŵ = (ŵij ).

4.4 Balakrishnan–Childs–Chandrasekar representation

Though the above described mixture representations are intuitive and will enable one
to carry out some theoretical properties of progressively Type-II right censored order
statistics, they may not be efficient from a computational point of view. For instance,
in Thomas–Wilson representation in (10), the value of N (number of all possible
rank vectors) may be very large even for relatively small values of m and n, as it
depends critically on the progressive censoring scheme (R1 , . . . , Rm ). For example,
in the case when n = 20 and m = 10, we have N = 56 for the progressive censoring
scheme (0, 0, 0, 0, 0, 0, 5, 0, 0, 5), while N is as large as 74,457 for the progressive
censoring scheme (2, 3, 0, 4, 1, 0, 0, 0, 0, 0). For this reason, therefore, a direct ap-
proach of Kamps and Cramer (2001) described in Sect. 4.2 or of Balakrishnan et al.
(2002) described below, would be more efficient computationally.
For the purpose of deriving simple explicit expressions for marginal and joint den-
sities of progressively Type-II right censored order statistics, the following lemma
was established in Balakrishnan et al. (2002) using induction.

Lemma 1 Let f (x) and F (x) be the pdf and cdf of an absolutely continuous distri-
bution, and aj > 0 for j = 1, . . . , r. Then, for r ≥ 1,
   x2 
r
xr+1 x3  a −1
··· f (xj ) 1 − F (xj ) j dx1 · · · dxr
T T T j =1
r
 b (a ,...,ar )  a +···+ar−i
= ci,r (a1 , . . . , ar ) 1 − F (xr+1 ) i,r 1 1 − F (T ) 1 , (15)
i=0

where

(−1)i
ci,r (a1 , . . . , ar ) = i r−i+j r−i r−i  ,
j =1 k=r−i+1 a k j =1 k=j ak
r
bi,r (a1 , . . . , ar ) = aj ,
j =r−i+1

0 i−1
and the conventions that j =1 aj ≡ 1 and j =i dj ≡ 0.

Upon starting with the joint density of the progressively Type-II right censored
sample in (1) and carrying out the necessary integration by using Lemma 1, Bal-
akrishnan et al. (2002) derived an expression for the pdf of Xr:m:n (for 1 ≤ r ≤ m)
Progressive censoring methodology: an appraisal 221

as
r−1
 R −1
fXr:m:n (xr ) = C ci,r−1 (R1 + 1, . . . , Rr−1 + 1)f (xr ) 1 − F (xr ) i , (16)
i=0

where
C = n(n − R1 − 1) · · · (n − R1 − · · · − Rr−1 − r + 1),
Rr∗ = n − r − R1 − · · · − Rr−1 ,
and
r−1
Ri = Rr∗ +1+ (Rj + 1).
j =r−i

Proceeding similarly, these authors also derived an expression for the joint density of
Xr:m:n and Xs:m:n (for 1 ≤ r < s ≤ m) as
fXr:m:n ,Xs:m:n (xr , xs )
r−1 s−r−1
=C ci,r−1 (R1 + 1, . . . , Rr−1 + 1)ck,s−r−1 (Rr+1 , . . . , Rs−1 )
i=0 k=0

× f (xr ){1 − F (xr )}Rik −1 f (xs ){1 − F (xs )}Rk −1 , xr < xs , (17)
s−1−k
where Rik = j =r−i (Rj + 1).
It is important to note that Kamps–Cramer representation in (12) and Balakrishnan–
Childs–Chandrasekar representation in (16) are the same in form, with the former
having been obtained by exploiting some properties of exponential generalized order
statistics, while the latter having been obtained by means of direct integration using
Lemma 1.

4.5 Balakrishnan–Cramer representation

All the representations presented so far hold under the assumption that the pro-
gressively Type-II right censored sample arises from n iid variables. Now consider
the case when X1 , . . . , Xn are independent variables with continuous distributions
F1 , . . . , Fn and densities f1 , . . . , fn , respectively. Then Balakrishnan and Cramer
(2007) derived the joint density of the progressively Type-II right censored sample
(R1 ,...,Rm ) (R1 ,...,Rm )
(X1:m:n , . . . , Xm:m:n ). They specifically showed that the joint density is given
by
m
1  m
fX1:m:n ,...,Xm:m:n (x1 , . . . , xm ) = γj fP (j ) (xj )
(n − 1)!
j =2 P ∈Sn j =1
 m+ρj 
  
× 1 − FP (k) (xj ) , x1 < · · · < xm ,
k=m+ρj −1 +1

(18)
222 N. Balakrishnan


where γ1 = n, γj = m i=j (Ri + 1), for j = 2, . . . , m, ρ0 = 0, ρr = R1 + · · · + Rr , for
r = 1, . . . , m, ρm = n − m, Sn is the set of all permutations P of (1, 2, . . . , n), and
P (j ) is the j th component of the permutation vector P ∈ Sn .
Balakrishnan and Cramer (2007) also presented a permanent representation for the
joint density in (18) as

fX1:m:n ,...,Xm:m:n (x1 , . . . , xm )


⎛ ⎞ }1
f1 (x1 ) ··· fn (x1 )
 ⎜ 1 − F1 (x1 ) ··· 1 − Fn (x1 ) ⎟ }R1
1 
m
⎜ ⎟.
Per ⎜ ⎟.
.. .. ..
= γj ⎜ . . . ⎟. , (19)
(n − 1)! ⎝ f (x )
j =2
1 m ··· fn (xm ) ⎠ }1
1 − F1 (xm ) · · · 1 − Fn (xm ) }Rm

where Per(A) denotes the permanent of the matrix A; permanent is exactly like a
determinant except that all the terms in the expansion have a positive sign.
In the homogeneous case, i.e., F1 = · · · = Fn = F and f1 = · · · = fn = f , the
permanent in (19), when expanded, will simply yield

m
 R
n! f (xi ) 1 − F (xi ) i ,
i=1

which, when multiplied with the constant in the front, simply reduces to

m
fX1:m:n ,...,Xm:m:n (x1 , . . . , xm ) = C f (xi ){1 − F (xi )}Ri , x1 < · · · < xm ,
i=1

exactly as in (1).

4.6 Fischer–Balakrishnan–Cramer representation

In the heterogeneous case, Fischer et al. (2007) established recently a mixture repre-
sentation for the joint distribution of the progressively censored order statistics. They
specifically showed that, for any Borel set B ∈ B m ,
(R1 ,...,Rm )  
P X1:m:n (R1 ,...,Rm )
, . . . , Xm:m:n ∈B
1  
= m−1 γ −1 P Xk1 :n , . . . , Xkm :n ∈ B ,
i
i=1 Ri ZR

where the summation is over Z R , defined as


  i−1 
 
i
Z = (l1 , S1 , l2 , S2 , . . . , lm , Sm ) : Si ⊆ {1, . . . , n}
R
Sj ∪ {lj } ,
j =1 j =1
  i−1  
 
|Si | = Ri , li = min {1, 2, . . . , n} Sj ∪ {lj } ∀i = 1, . . . , m .
j =1
Progressive censoring methodology: an appraisal 223

In the homogeneous case, i.e., F1 = · · · = Fn = F , the above mixture form reduces


to
(R ,...,Rm )  
1
P X1:m:n (R1 ,...,Rm )
, . . . , Xm:m:n ∈B
−i
m Rj ∗ −ij −1∗ 
R1 R2∗ −i1∗ R
m−1∗ m−2∗
j =2 Rj
= ··· m−1 γj −1
i1 =0 i2 =0 im−1 =0 j =1 Rj

× P (X1:n , Xi1∗ +2:n , . . . , Xim−1∗ +m:n ) ∈ B ,
j j
where Rj ∗ = ν=1 Rν , 1 ≤ j ≤ m, and ij ∗ = ν=1 iν , 1 ≤ j ≤ m − 1.
It should be mentioned that Guilbaud’s mixture representation in (14) can be
obtained from this form as the one-dimensional marginal distribution. Note, how-
ever, that the above mixture form yields explicit expressions for the weights, while
Guilbaud’s representation involves an iterative procedure in the computation of its
weights.

5 Useful transformations for some distributions

Just as the usual order statistics possess some interesting results concerning transfor-
mations, progressively Type-II right censored order statistics also admit some useful
transformations in the case of distributions such as uniform, exponential and Pareto.

5.1 Uniform distribution

Let the progressively Type-II right censored order statistics U1:m:n , . . . , Um:m:n be
from the Uniform(0,1) distribution. Then, the joint density is readily obtained from
(1) to be

m
fU1:m:n ,...,Um:m:n (u1 , . . . , um ) = C (1 − ui )Ri , 0 < u1 < · · · < um < 1. (20)
i=1

In this case, Balakrishnan and Sandhu (1995) used the Jacobian method to establish
the following result, which is an extension of Malmquist’s (1950) transformation
result for the usual order statistics to the case of progressive Type-II right censoring.

Result 4 With (U1:m:n , . . . , Um:m:n ) denoting a progressively Type-II right censored


sample from the Uniform(0,1) distribution with the progressive censoring scheme
(R1 , . . . , Rm ), the random variables
1 − Um−i+1:m:n
Vi = (i = 1, . . . , m − 1) and Vm = 1 − U1:m:n (21)
1 − Um−i:m:n
are all mutually independent random variables with
d
Vi = Beta(i + Rm−i+1 + · · · + Rm , 1), i = 1, . . . , m. (22)
224 N. Balakrishnan

Remark 4 As mentioned by Balakrishnan and Sandhu (1995), the independent beta


random variables V1 , . . . , Vm in (22) provide a simple algorithm for the computer
generation of a progressively Type-II right censored sample from the Uniform(0,1)
distribution, which, when transformed by F −1 (·), will result in the generation of a
progressively Type-II right censored sample from an arbitrary continuous distribution
F (·).

Remark 5 By inverting the transformation in (21), we can write



m
Ui:m:n = 1 − Vj , i = 1, . . . , m. (23)
j = m−i+1

Since Vj ’s are independent beta random variables, the representation in (23) will
readily yield explicit expressions for the single and the product moments of progres-
sively Type-II right censored order statistics from the Uniform(0,1) distribution. For
example, we have


m 
m
j + Rm−j +1 + · · · + Rm
E(Ui:m:n ) = 1 − E(Vj ) = 1 − .
j + 1 + Rm−j +1 + · · · + Rm
j = m−i+1 j = m−i+1

Such explicit expressions have been used by Balakrishnan and Rao (1997) to develop
series approximations for the single and the product moments of progressively Type-
II right censored order statistics from an arbitrary continuous distribution F (·).

5.2 Exponential distribution

Let the progressively Type-II censored sample X1:m:n , . . . , Xm:m:n be from a standard
exponential distribution. Then the joint density is readily obtained from (1) to be
 m 
fX1:m:n ,...,Xm:m:n (x1 , . . . , xm ) = C exp − (Ri + 1)xi ,
i=1
0 < x1 < · · · < xm < ∞. (24)

Then the following result was established by Thomas and Wilson (1972) and Viveros
and Balakrishnan (1994) regarding normalized spacings, which is an extension of
the corresponding result for the usual order statistics due to Sukhatme (1937), Rényi
(1953), and Epstein and Sobel (1953, 1954).

Result 5 Let X1:m:n , . . . , Xm:m:n denote a progressively Type-II right censored sam-
ple from the standard exponential distribution with the progressive censoring scheme
(R1 , . . . , Rm ). Then the corresponding normalized spacings

Z1 = nX1:m:n and
Zi = (n − i + 1 − R1 − · · · − Ri−1 )(Xi:m:n − Xi−1:m:n ), for i = 2, . . . , m, (25)

are independent and identically distributed as standard exponential.


Progressive censoring methodology: an appraisal 225

Remark 6 The independent standard exponential random variables Z1 , . . . , Zm in


(25) provide a simple algorithm for the simulation of a progressively Type-II right
censored sample from an exponential distribution.

Remark 7 By inverting the transformation in (25), we can express

i
d Zj
Xi:m:n = , i = 1, . . . , m, (26)
n − R1 − · · · − Rj −1 − j + 1
j =1

which reveals readily that the progressively Type-II right censored order statistics
from an exponential distribution form an additive Markov chain.

Remark 8 Since Zj ’s are independent exponential random variables, the represen-


tation in (26) will readily yield explicit expressions for the single and the product
moments of progressively Type-II right censored order statistics from the standard
exponential distribution. For example, we have

i i
E(Zj ) 1
E(Xi:m:n ) = = ,
n − R1 − · · · − Rj −1 − j + 1 n − R1 − · · · − Rj −1 − j + 1
j =1 j =1

i
Var(Zj )
Var(Xi:m:n ) =
(n − R1 − · · · − Rj −1 − j + 1)2
j =1

i
1
= ,
(n − R1 − · · · − Rj −1 − j + 1)2
j =1

and

i
1
Cov(Xi:m:n , Xi :m:n ) = Var(Xi:m:n ) = , (27)
(n − R1 − · · · − Rj −1 − j + 1)2
j =1

for 1 ≤ i < i ≤ m ≤ n.

Remark 9 From (27), we observe that the variance-covariance matrix has a special
product decomposable form, viz., σi,j = ai bj , which enables an explicit inversion
of the variance-covariance matrix; see Balakrishnan and Cohen (1991) and Arnold
et al. (1992). This, in turn, facilitates the derivation of explicit best linear unbiased
estimators for the location and scale parameters of the exponential distribution and
also exact inference based on them, as displayed in Viveros and Balakrishnan (1994),
Balakrishnan and Sandhu (1996) and Balakrishnan et al. (2001).
226 N. Balakrishnan

5.3 Pareto distribution

Let X1:m:n , . . . , Xm:m:n be a progressively Type-II right censored sample from the
Pareto(ν) distribution with pdf

f (x) = νx −ν−1 , x > 1, ν > 0.

Then, from (1), we have the joint density to be


m
fX1:m:n ,...,Xm:m:n (x1 , . . . , xm ) = Cν m xi−ν(Ri +1)−1 ,
i=1
1 < x1 < · · · < xm < ∞. (28)

In this case, the following transformation result can be easily established by Jacobian
method; see Balakrishnan and Aggarwala (2000).

Result 6 Let X1:m:n , . . . , Xm:m:n be a progressively Type-II right censored sample


from the Pareto(ν) distribution with the progressive censoring scheme (R1 , . . . , Rm ).
Then the variables
Xi:m:n
Y1 = X1:m:n and Yi = , for i = 2, . . . , m, (29)
Xi−1:m:n

are mutually independent, and


d 
Yi = Pareto ν(n − i + 1 − R1 − · · · − Ri−1 ) , for i = 1, . . . , m. (30)

Remark 10 Result 6 is a generalization of the corresponding result for the usual order
statistics from the Pareto distribution; see, for example, Arnold (1983) and Johnson
et al. (1994).

Remark 11 By inverting the transformation in (29), we have the representation

d 
i
Xi:m:n = Yj , for i = 1, . . . , m, (31)
j =1

where Yj ’s are independent Pareto random variables, as in (30). As a result, we can


readily obtain explicit expressions for the single and the product moments of pro-
gressively Type-II right censored order statistics from the Pareto(ν) distribution. For
example, we have


i 
i
ν(n − j + 1 − R1 − · · · − Rj −1 )
E(Xi:m:n ) = E(Yj ) = ,
ν(n − j + 1 − R1 − · · · − Rj −1 ) − 1
j =1 j =1

provided it exists.
Progressive censoring methodology: an appraisal 227

Remark 12 One can similarly derive explicit expressions for variances and covari-
ances from the representation in (31). In this case as well, the variance-covariance
matrix has a product decomposable form mentioned earlier in Remark 9, which fa-
cilitates the derivation of explicit best linear unbiased estimators for the location and
scale parameters of the Pareto distribution and also exact inference based on them.

6 Discrete case

All the distributional results and properties of the progressively Type-II right censored
order statistics discussed so far have been for the continuous case. If the underlying
cdf F (x) is discrete, the observed order statistics will be discrete, and so ties may
occur among them with positive probability. Because of these ties, the distribution
theory of progressively Type-II right censored order statistics from a discrete distribu-
tion becomes quite complicated. Balakrishnan and Dembińska (2007a) have recently
studied this case in detail and some of their key results are based on the representation
in Result 7. It is of interest to mention here that Tran (2006)1 has recently discussed
discrete generalized order statistics and obtained some general results along similar
lines. However, as pointed out by Balakrishnan and Dembińska (2007b), the discrete
generalized order statistics do not cover as wide a class of models as in the contin-
uous case, though the case of progressively Type-II right censored order statistics is
indeed a special case.

(R ,...,R ) (R ,...,R )
Result 7 Let X1:m:n1 m
≤ · · · ≤ Xm:m:n
1 m
be progressively Type-II right censored
(R1 ,...,Rm ) (R1 ,...,Rm )
order statistics from any cdf F (x) and let U1:m:n ≤ · · · ≤ Um:m:n denote
progressively Type-II right censored order statistics (with the same censoring scheme)
arising from the uniform distribution on [0, 1]. Define the generalized inverse function
F ← (·) by
 
F ← (u) = inf x : F (x) ≥ u , u ∈ (0, 1).
Then,
 (R1 ,...,Rm )  
(R1 ,...,Rm )
X1:m:n (R1 ,...,Rm ) d
, . . . , Xm:m:n = F ← U1:m:n , . . . , F ← Um:m:n
(R1 ,...,Rm )
, (32)

d
where = stands for equality in distribution.

Remark 13 The representation in Result 7 is a generalization of the corresponding


well-known result for the usual order statistics presented, for example, in Nagaraja
(1992), Nevzorov (2001, p. 10) and David and Nagaraja (2003, p. 15):

(X1:n , . . . , Xn:n ) = F ← (U1:n ), . . . , F ← (Un:n ) ,
d
(33)

where (X1:n , . . . , Xn:n ) and (U1:n , . . . , Un:n ) are the usual order statistics arising from
any cdf F (x) and the usual uniform order statistics, respectively.

1 I thank Professor Erhard Cramer for bringing this work to my attention.


228 N. Balakrishnan

By using the representation in Result 7, Balakrishnan and Dembińska (2007a)


have derived the following expression for the joint probability mass function (pmf) of
the progressively Type-II right censored order statistics from a discrete distribution.

Result 8 The joint pmf of progressively Type-II right censored order statistics
(R1 ,...,Rm ) (R1 ,...,Rm )
(X1:m:n , . . . , Xm:m:n ) from a discrete cdf F (x) is given by
(R ,...,R ) (R ,...,R ) 
1
P X1:m:n m
= x1 , X2:m:n
1 m
= x2 , . . . , Xm:m:n
(R1 ,...,Rm )
= xm
 m
=C (1 − ui )Ri du1 · · · dum , (34)
A i=1

where, as in Sect. 2,

C = n(n − R1 − 1) · · · (n − R1 − R2 − · · · − Rm−1 − m + 1), (35)

and A is the m-dimensional space given by



A = (u1 , . . . , um ) : u1 ≤ u2 ≤ · · · ≤ um and

F (xi− ) < ui ≤ F (xi ), i = 1, 2, . . . , m . (36)

Remark 14 It is important to note that in view of Result 7, the joint cdf of the pro-
gressively Type-II right censored order statistics has the same integral form for both
continuous and discrete cdf F (x), and is given by
(R1 ,...,Rm ) (R1 ,...,Rm ) 
P X1:m:n ≤ x1 , X2:m:n ≤ x2 , . . . , Xm:m:n
(R1 ,...,Rm )
≤ xm

= P F ← (U1:m:n ) ≤ x1 , . . . , F ← (Um:m:n
(R1 ,...,Rm ) (R1 ,...,Rm )
) ≤ xm
(R1 ,...,Rm ) 
= P U1:m:n ≤ F (x1 ), . . . , Um:m:n
(R1 ,...,Rm )
≤ F (xm )
  m
=C (1 − ui )Ri du1 · · · dum , (37)
B i=1

where C is as in (35) and


 
B = (u1 , . . . , um ) : 0 ≤ u1 ≤ u2 ≤ · · · ≤ um and ui ≤ F (xi ), i = 1, 2, . . . , m .
(38)

From the quantile representation in Result 7 and the marginal and joint distrib-
utions of uniform generalized order statistics, Balakrishnan and Dembińska (2007a)
have derived the following expressions for the marginal distributions of progressively
Type-II right censored order statistics.
(R ,...,Rm ) (R ,...,Rm ) (R ,...,Rm )
Result 9 The joint pmf of X1:m:n 1 1
, X2:m:n 1
, . . . , Xi:m:n , 1 ≤ i ≤ m, from
a discrete cdf F (x) is given by
(R ,...,Rm ) (R ,...,Rm ) (R ,...,Rm ) 
1
P X1:m:n = x1 , X2:m:n
1
= x2 , . . . , Xi:m:n
1
= xi
Progressive censoring methodology: an appraisal 229

 
i−1
= Ci (1 − uj )Rj (1 − ui )n−R1 −···−Ri−1 −i du1 · · · dui , (39)
Ai j =1

where

Ci = n(n − R1 − 1)(n − R1 − R2 − 2) · · · (n − R1 − · · · − Ri−1 − i + 1) (40)

and Ai is the i-dimensional space given by



Ai = (u1 , . . . , ui ) : u1 ≤ u2 ≤ · · · ≤ ui and

F (xj− ) < uj ≤ F (xj ), j = 1, 2, . . . , i .

(R ,...,R )
1 m 1 m (R ,...,R )
1 m (R ,...,R )
Result 10 Let (X1:m:n , X2:m:n , . . . , Xm:m:n ) be progressively Type-II
right censored order statistics from a discrete cdf F (x). Then, for 1 ≤ i ≤ m
 i  F (x)
 i  1
(1 − u)γj −1 du
(R1 ,...,Rm )
P Xi:m:n = x = Ci
γ − γj F (x − )
j =1 =1,=j

i 
i
1 {1 − F (x − )}γj − {1 − F (x)}γj
= Ci ,
γ − γj γj
j =1 =1,=j

(41)

or equivalently

(R1 ,...,Rm )  i 
i
1 1 − {1 − F (x)}γj
P Xi:m:n ≤ x = Ci , (42)
γ − γj γj
j =1 =1,=j

where
m
γr = m − r + 1 + Ri , for r = 1, . . . , m.
i=r

(R1 ,...,Rm )
It is important to observe that the expression of the cdf of Xi:m:n in (42) for
the discrete case is exactly the same as the Kamps–Cramer representation in (11) for
the continuous case.
It is well-known that the usual order statistics from a discrete distribution with
at least three points in its support do not form a Markov chain; see, for example
Nagaraja (1982) and Arnold et al. (1992, Sect. 3.4). In a similar vein, Balakrishnan
and Dembińska (2007a) established the following non-Markovian structure of the
progressively Type-II right censored order statistics in this case.

Result 11 Let F (x) be a discrete cdf with at least three points in its support. Then,
(R1 ,...,Rm ) (R1 ,...,Rm ) (R1 ,...,Rm )
X1:m:n , X2:m:n , . . . , Xm:m:n , the progressively Type-II right censored or-
der statistics arising from F (x), do not form a Markov chain.
230 N. Balakrishnan

Remark 15 Though there are many similarities and parallel results between the con-
tinuous and discrete cases for progressively Type-II right censored order statistics as
pointed above, the non-Markovian structure in the discrete case in Result 11 is in
direct contrast with the Markovian structure in the continuous case presented earlier
in Result 3.

Open Problem 1 Even though (as stated in Result 11) the progressively Type-II
(R1 ,...,Rm )
right censored order statistics Xi:m:n , i = 1, . . . , m, from a discrete distribution
do not form a Markov chain, one could expand the state space into two dimensions
(R1 ,...,Rm )
by incorporating to Xi:m:n a variable that describes in some way the number
(R1 ,...,Rm )
of ties with Xi:m:n , and then examine, whether this bivariate sequence forms a
Markov sequence. This has been established for the usual discrete order statistics by
Rüschendorf (1985). Alternatively, one could ask, whether there are some suitable
events conditioned on which the progressively Type-II right censored order statistics
do form a Markov chain, along the lines of Nagaraja (1986) for the usual discrete
order statistics. Some interesting discussion along these lines have already been made
by Tran (2006) and Balakrishnan and Dembińska (2007b).

In what follows, our focus will only be on progressively Type-II right censored
order statistics from the continuous cdf and related issues.

7 Unimodality, bounds and relations

7.1 Unimodality

There are many results in the literature concerning the unimodality properties of the
distributions of the usual order statistics; see, for example, the book by Dharmad-
hikari and Joag-Dev (1988). A distribution F is said to be unimodal with a mode μ if
F is convex on (−∞, μ) and concave on (μ, ∞). There is also a notion of strong uni-
modality, wherein a distribution F is said to be strongly unimodal if the convolution
of distribution F with any unimodal distribution G is also unimodal.
Interesting results in this context have been established by Cramer (2004) re-
garding unimodality of the distributions of progressively Type-II right censored or-
der statistics. Specifically, by considering the joint density of the uniform progres-
sively Type-II censored sample in (20), and noting that the functions (1 − ui )Ri (for
i = 1, . . . , m) are all logconcave, he established that the joint density of the uniform
progressively Type-II censored order statistics is logconcave by simply using the fact
that a product of logconcave functions is also logconcave. From this result, the log-
(R1 ,...,Rm )
concavity of the marginal pdf of Ur:m:n readily follows from the property that
any marginal density of a logconcave joint density is also logconcave; see Brascamp
and Lieb (1975). Now, upon using Ibragimov’s (1956) result that a nondegenerate cdf
F is strongly unimodal if and only if F is absolutely continuous and the correspond-
(R1 ,...,Rm )
ing pdf f is logconcave, the above result readily implies that the cdf of Ur:m:n
is strongly unimodal. It needs to be mentioned here that this result was established
Progressive censoring methodology: an appraisal 231

earlier by Cramer et al. (2004) for the uniform generalized order statistics by using
an inductive argument.
More generally, Cramer (2004) has also proved that if the progressively Type-II
right censored order statistics are from a strongly unimodal cdf F , then the marginal
(R1 ,...,Rm ) (R1 ,...,Rm )
pdf of any Xr:m:n is logconcave and consequently the cdf of Xr:m:n (for
r = 1, . . . , m) is strongly unimodal.

7.2 Universal bounds

Universal bounds for moments of the usual order statistics have been discussed exten-
sively in the literature; see, for example, Arnold and Balakrishnan (1989). In a similar
vein, Balakrishnan et al. (2001) derived universal bounds for means and variances of
progressively Type-II right censored order statistics from an arbitrary continuous dis-
tribution.
For demonstrating these bounds, let us consider Kamps–Cramer representation in
(R1 ,...,Rm )
(12) for the marginal density of Ur:m:n , the rth uniform progressively Type-II
right censored order statistic, given by
r
fUr:m:n (u) = cr−1 ai,r (1 − u)γi −1 , 0 < u < 1, r = 1, . . . , m, (43)
i=1

where γr = m − r + 1 + Rr + · · · + Rm . So,
 1 r r 
2 1
fUr:m:n (u) du = cr−1
2
ai,r aj,r (1 − u)γi +γj −2 du
0 i=1 j =1 0

r
 
r

= 2
cr−1 ai,r (γi + γj − 1) . (44)
i=1 j =1

Then, with μ and σ 2 denoting the mean and variance of the distribution F , and
(R1 ,...,Rm )
Xr:m:n the progressively Type-II right censored order statistic from F , we get
   
 E(Xr:m:n
(R1 ,...,Rm )
) − μ   1 −1 
 = G (u){fUr:m:n (u) − 1} du
 σ 
0
 r
1/2
air
≤ cr−1
2 r −1 (45)
j =1 (γi + γj − 1)
i=1

by means of Cauchy–Schwarz inequality, where G is the cumulative distribution


function of the random variable Y = (X − μ)/σ . As done by Balakrishnan et al.
(2001), improvements to this bound are possible either in the form of interpolation-
type bounds in Balakrishnan (1990) or by the method of greatest convex minorant
discussed in Moriguti (1953) and Balakrishnan (1993). Furthermore, Balakrishnan
et al. (2001) also derived upper bounds for variances of progressively Type-II right
censored order statistics, by utilizing an upper bound for the variance of a function
232 N. Balakrishnan

of a random variable due to Cacoullos and Papathanasiou (1985). Some more gen-
eral results in the context of generalized order statistics have also been established
recently.

Open Problem 2 There are many different types of bounds that exist in the case of
usual order statistics for means, variances, covariances, correlations, means of linear
functions, etc. The question is whether these could be generalized to the case of
progressive Type-II right censoring situation!

7.3 Relations

Several recurrence relations are known for densities and joint densities of the usual
order statistics; see Arnold and Balakrishnan (1989) and Balakrishnan and Sultan
(1998). For example,

rfXr+1:n (x) + (n − r)fXr:n (x) = nfXr:n−1 (x), x ∈ R, 1 ≤ r ≤ n − 1, (46)

and

rfXr+1:n ,Xs+1:n (x, y) + (s − r)fXr:n ,Xs+1:n (x, y) + (n − s)fXr:n ,Xs:n (x, y)


= nfXr:n−1 ,Xs:n−1 (x, y), x < y, 1 ≤ r < s ≤ n − 1, (47)

are the well-known triangle and tetrahedron rules, which have been used quite exten-
sively in the order statistics literature.
These relations have been generalized by Kamps and Cramer (2001) and Balakr-
ishnan et al. (2005) to the case of progressively Type-II right censored order statistics
from an arbitrary continuous distribution F as follows:

γr+1 f (R ,...,Rm )
1 (x) + (γ1 − γr+1 )f (R ,...,R )
1 m (x) = γ1 f (R ,...,R )
2 m (x),
Xr:m:n Xr+1:m:n Xr:m−1:n−R
1 −1

x ∈ R, 1 ≤ r ≤ m − 1, (48)

and

(γ1 − γr+1 )f (R ,...,R )


1 m 1 (R ,...,R )
m (x, y) + (γr+1 − γs+1 )f (R ,...,Rm )
1 (R ,...,R )
1 m (x, y)
Xr+1:m:n ,Xs+1:m:n Xr:m:n ,Xs+1:m:n

+ γs+1 f (R ,...,Rm )
1 (R ,...,Rm )
1 (x, y) = γ1 f (R ,...,R )
2 m (R ,...,R )
2 m (x, y),
Xr:m:n ,Xs:m:n Xr:m−1:n−R ,Xs:m−1:n−R
1 −1 1 −1

x < y, 1 ≤ r < s ≤ m − 1. (49)

Open Problem 3 The proofs of the relations in (48) and (49) provided by Kamps and
Cramer (2001) and Balakrishnan et al. (2005) are algebraic in nature. However, in the
case of relations in (46) and (47) for the usual order statistics, there are simple and
elegant probabilistic proofs. This raises the question, whether similar probabilistic
proofs could be provided for the relations for the progressively Type-II right censored
case!
Progressive censoring methodology: an appraisal 233

Open Problem 4 There are numerous relations satisfied by marginal and joint den-
sities of the usual order statistics, and they have also been utilized to determine upper
bounds on the number of single and product moments to be determined for a sam-
ple of size n given these moments for all sample sizes less than n; see, for example,
Arnold and Balakrishnan (1989). The question here is whether similar results can be
developed for the progressively Type-II right censored case as well!

8 Recurrence relations for specific distributions

In addition to general recurrence relations such as those in (46) and (47) that are
satisfied for an arbitrary distribution, the usual order statistics also satisfy several re-
currence relations for many specific distributions such as exponential, Pareto, power
function, logistic, normal, their truncated forms, etc.; see, for example, Arnold and
Balakrishnan (1989), Arnold et al. (1992, Chap. 4), and Balakrishnan and Sultan
(1998). Some of these results have been generalized to the case of progressive Type-II
right censoring, but the proofs are somewhat involved and are of different type com-
pared to those for the corresponding results for the usual order statistics.

8.1 Exponential and truncated exponential distributions

For the standard exponential distribution with pdf f (x) = e−x , x ≥ 0, we have the
characterizing differential equation as

f (x) = 1 − F (x), for x ≥ 0. (50)

Just as Joshi (1978, 1979, 1982) used (50) to establish recurrence relations for sin-
gle and product moments of the usual order statistics, Aggarwala and Balakrishnan
(1996) established a recursive algorithm for the computation of single and product
moments of progressively Type-II right censored order statistics. Their results for
1 (R ,...,Rm )(k) (R ,...,Rm )k
1
single moments is as follows, where μi:m:n denotes E(Xi:m:n ).

Result 12 For n ≥ 1 and k = 0, 1, . . . ,

(n−1)(k+1) k + 1 (n−1)(k)
μ1:1:n = μ1:1:n ; (51)
n
for 2 ≤ m ≤ n and k = 0, 1, . . . ,
(k+1)
μ(R 1 ,...,Rm )
1:m:n
1  (R1 ,...,Rm )(k) (R1 +R2 +1,R3 ,...,Rm )(k+1) 
= (k + 1)μ1:m:n − (n − R1 − 1)μ1:m−1:n ; (52)
R1 + 1

for 2 ≤ i ≤ m − 1, m ≤ n and k = 0, 1, . . . ,
234 N. Balakrishnan

(R ,...,Rm )(k+1)
1
μi:m:n
1  (R1 ,...,Rm )(k)
= (k + 1)μi:m:n
Ri + 1
(R ,...,R ,Ri +Ri+1 +1,Ri+2 ,...,Rm )(k+1)
− (n − R1 − · · · − Ri − i)μi:m−1:n
1 i−1

(R ,...,R ,Ri−1 +Ri +1,Ri+1 ,...,Rm )(k+1) 


+ (n − R1 − · · · − Ri−1 − i + 1)μi−1:m−1:n
1 i−2
; (53)

for 2 ≤ m ≤ n and k = 0, 1, . . . ,
(k+1) k + 1 (R1 ,...,Rm )(k) (R1 ,...,Rm−2 ,Rm−1 +Rm +1)(k+1)
μ(R 1 ,...,Rm )
m:m:n = μ + μm−1:m−1:n . (54)
Rm + 1 m:m:n

Remark 16 The recurrence relations in (51–54) are complete in the sense that they
will enable one to obtain all the single moments of all progressively Type-II right
censored order statistics for all sample sizes and all progressive censoring schemes
(R1 , . . . , Rm ) in a simple and efficient recursive manner, as displayed by Aggarwala
and Balakrishnan (1996). It should be mentioned here that Balakrishnan et al. (2001)
presented a simpler recursive algorithm in this case.

Remark 17 Proceeding in an analogous manner, Aggarwala and Balakrishnan (1996)


also established a recursive algorithm for the computation of product moments, which
will enable the computation of all the product moments for all sample sizes and all
progressive censoring schemes.

Because of the additive Markov chain representation of the exponential progres-


sive Type-II right censored order statistics in (26), the direct computation of single
and product moments is rather simple and so there is no real necessity for the recur-
sive algorithm in this case. However, the recursive approach is flexible enough to be
adapted to treat truncated exponential distributions (for which the direct computation
becomes quite complicated), as carried out by Aggarwala and Balakrishnan (1996)
and Balakrishnan and Aggarwala (2000) for the right truncated and doubly truncated
exponential densities f (x) = e−x /P for 0 ≤ x ≤ P1 , and f (x) = e−x /(P − Q) for
Q1 ≤ x ≤ P1 , respectively, where P1 = − ln(1 − P ) and Q1 = − ln(1 − Q).

8.2 Pareto and truncated Pareto distributions

Consider the standard Pareto distribution with pdf f (x) = νx −ν−1 , x ≥ 1, ν > 0,
and its doubly truncated form with pdf f (x) = νx −ν−1 /(P − Q), 1 ≤ Q1 < x ≤ P1 ,
ν > 0, where P1 = (1 − P )−1/ν and Q1 = (1 − Q)−1/ν . In these cases, Balakrishnan
and Aggarwala (2000) used the characterizing differential equation
 
1−P ν  ν
f (x) = + 1 − F (x) , for 1 ≤ Q1 < x ≤ P1 , (55)
P −Q x x

to establish a complete recursive algorithm for the computation of all single and prod-
uct moments of all progressively Type-II right censored order statistics for all sample
Progressive censoring methodology: an appraisal 235

sizes and all progressive censoring schemes (R1 , . . . , Rm ). These results are general-
izations of the corresponding results for the usual order statistics due to Balakrishnan
and Joshi (1982).

8.3 Power function and truncated power function distributions

Consider the standard power function distribution with pdf f (x) = νx ν−1 , 0 < x < 1,
ν > 0, and its doubly truncated form with pdf f (x) = νx ν−1 /(P − Q), 0 ≤ Q1 < x <
P1 ≤ 1, ν > 0, where P1 = P 1/ν and Q1 = Q1/ν . In these cases, Balakrishnan and
Aggarwala (2000) used the characterizing differential equation
 
P ν  ν
f (x) = − 1 − F (x) , for 0 ≤ Q1 < x ≤ P1 ≤ 1, (56)
P −Q x x

to establish a complete recursive algorithm for the computation of all single and prod-
uct moments of all progressively Type-II right censored order statistics for all sam-
ple sizes and all progressive censoring schemes. These results are generalizations of
the corresponding results for the usual order statistics due to Balakrishnan and Joshi
(1981).

8.4 Logistic distribution

Consider the standard logistic distribution with pdf and cdf as

e−x 1
f (x) = and F (x) = , for x ∈ R. (57)
(1 + e−x )2 1 + e−x

In this case, the characterizing differential equation can be expressed as


     2
f (x) = F (x) 1 − F (x) = 1 − F (x) − 1 − F (x) , x ∈ R. (58)

Recently, Balakrishnan and Kannan (2007) made use of (58) to develop a complete
recursive algorithm for the computation of all single and product moments of all pro-
gressively Type-II right censored order statistics for all sample sizes and all progres-
sive censoring schemes. However, due to the complexity in characterizing differential
equation in (58), the resulting recurrence relations are also complex and the proofs
are quite involved. Despite this, the recursive scheme is complete and is, therefore,
quite efficient and useful from a computational viewpoint. These results are general-
izations of the corresponding results for the usual order statistics established by Shah
(1966, 1970); see also Balakrishnan (1992).

Open Problem 5 Recurrence relations have been established for the usual order sta-
tistics for many other distributions including normal, Weibull, gamma, half logistic,
and log-logistic. The natural question that arises here is to what extent these results
can be generalized to the case of progressive Type-II censoring!
236 N. Balakrishnan

Open Problem 6 Many of the recurrence relations for the usual order statistics have
also been shown to be characterizations of the underlying distribution. Naturally, we
would like to examine whether the relations established for the progressively Type-II
right censored order statistics also lead to characterization results for the underlying
distribution. One should mention here the works of Keseling (1999), Keseling and
Kamps (2003), and Balakrishnan and Malov (2005) for some characterization results
based on distributional and independence properties concerning progressively Type-
II right censored order statistics, with the first two dealing with generalized order
statistics.

9 Parametric inference

Numerous papers have been written concerning inference based on progressively


Type-II censored samples. Consequently, it will not be practical to aim to detail all
these developments here, and so I shall provide a brief outline of some basic results
and then provide a list of pertinent articles and the cases discussed therein.

9.1 Linear inference


(R1 ,...,Rm ) (R1 ,...,Rm )
Let Y1:m:n , . . . , Ym:m:n be a progressively Type-II right censored sample
from a scale family of distributions with pdf g(x; θ ) = θ1 f ( xθ ), where f (·) is the
(R1 ,...,Rm ) (R1 ,...,Rm )
standard density. Then Xi:m:n = θ1 Yi:m:n , i = 1, . . . , m, is the correspond-
ing standardized progressively Type-II censored sample from f , and let μ and 
denote its mean vector and variance-covariance matrix. Then, the best linear unbi-
ased estimator (BLUE) of θ and its variance are (see Balakrishnan and Cohen 1991,
and David and Nagaraja 2003)
   
θ ∗ = μT  −1 Y μT  −1 μ and Var(θ ∗ ) = σ 2 μT  −1 μ . (59)

In the case of the scaled exponential distribution, when the explicit expressions for
μ and  presented earlier in (27) are used in (59) and simplified, we obtain
m
1 θ2
θ∗ = and Var(θ ∗ ) =
(R ,...,Rm )
(Ri + 1)Yi:m:n
1
. (60)
m m
i=1

∗ 
Remark 18 Since mθθ = m i=1 Zi , where Zi ’s are the normalized spacings, defined
in (25), which are iid standard exponential variables (see Result 5), we readily note
∗ d
θ = χ2m ; so, exact confidence intervals and tests of hypotheses can be con-
that 2mθ 2

structed for θ .

Remark 19 From (60), we observe that the precision of the BLUE of θ depends only
on m and n, and not on the progressive censoring scheme (R1 , . . . , Rm ).

Example 1 Let us consider the progressively Type-II right censored sample of size
m = 8 from n = 19 times to breakdown on insulating fluids tested at 34 kilovolts in
Progressive censoring methodology: an appraisal 237

Table 1 Progressively Type-II right censored data

i 1 2 3 4 5 6 7 8

Ri 0 0 3 0 3 0 0 5
Times 0.19 0.78 0.96 1.31 2.78 4.85 6.50 7.35
Log-times −1.6608 −0.2485 −0.0409 0.2700 1.0224 1.5789 1.8718 1.9947

Table 1, taken from Viveros and Balakrishnan (1994) who produced the data from
Table 6.1 of Nelson (1982).
If we assume a scaled exponential distribution for times to breakdown in Table 1,
(60) immediately yields

θ ∗ = 9.09 and Std. err.(θ ∗ ) = 3.21.

1 (R ,...,R )
m 1 m (R ,...,R )
Now, suppose Y1:m:n , . . . , Ym:m:n is a progressively Type-II right cen-
sored sample from a location-scale family of distributions with pdf
 
1 x −μ
g(x; μ, σ ) = f ,
σ σ

where f (·) is the standard density. Let μ and  denote once again the mean vector
and the variance-covariance matrix of the corresponding standardized progressively
(R1 ,...,Rm ) (R1 ,...,Rm )
Type-II right censored sample Xi:m:n = (Yi:m:n − μ)/σ , i = 1, . . . , m,
from f . Then, the BLUEs of μ and σ and their variances and covariance are (see
Balakrishnan and Cohen 1991, and David and Nagaraja 2003)

μ∗ = −μT Y and σ ∗ = 1T Y, (61)

and

σ2  σ2 
Var(μ∗ ) = μT  −1 μ , Var(σ ∗ ) = 1T  −1 1 , and
(62)
∗ ∗ σ2 −1

Cov(μ , σ ) = − T
μ  1 ,

where 1 = (1, . . . , 1),  =  −1 (1μT − μ1T ) −1 / , and = (1T  −1 1)


T −1
× (μ  μ) − (μ  1) .
T −1 2

In the case of the two-parameter exponential distribution with pdf g(x; μ, σ ) =


1 −(x−μ)/σ
σ e , x ≥ μ, when the explicit expressions for μ and  presented earlier in
(27) are used in (61) and (62) and simplified, we obtain

1
μ∗ = Y1:m:n − σ∗
(R ,...,Rm )
1
and
n
m (63)
∗ 1 (R ,...,Rm ) (R ,...,Rm ) 
σ = (Ri + 1) Yi:m:n
1
− Y1:m:n
1
,
m−1
i=2
238 N. Balakrishnan

and
mσ 2 σ2 σ2
Var(μ∗ ) = , Var(σ ∗ ) = and Cov(μ∗ , σ ∗ ) = − .
(m − 1)n2 m−1 n(m − 1)
(64)
(R1 ,...,Rm )
∗ d 2n(Y −μ) d
Remark 20 In this case, it can be shown that 2(m−1)σ
σ = χ2m−2
2 , 1:m:n
σ =
2
χ2 , and that they are mutually independent. Therefore, as mentioned by Viveros and
Balakrishnan (1994), exact confidence intervals and tests of hypotheses can be carried
out for the parameters μ and σ based on F and χ 2 distributions, respectively.

Remark 21 Once again, we observe from (64) that the precision of the BLUEs of
μ and σ depend only on m and n, and not on the progressive censoring scheme
(R1 , . . . , Rm ).

Some further results have been obtained by Balakrishnan et al. (2001, 2002), Bal-
akrishnan and Lin (2002), and Chandrasekar et al. (2002) regarding properties of
these estimators, some generalizations and inferential procedures based on them for
the scaled and two-parameter exponential distributions.
Best linear unbiased and various other linear estimators based on progressively
Type-II right censored samples have been studied for a variety of distributions by a
number of authors including Mann (1969, 1971), Thomas and Wilson (1972), Cac-
ciari and Montanari (1987), Montanari and Cacciari (1988), Balakrishnan and Sandhu
(1996), Balakrishnan and Rao (1997), Aggarwala and Balakrishnan (1998), Balakr-
ishnan and Aggarwala (2000), Balakrishnan et al. (2001), and Burkschat et al. (2006).

9.2 Likelihood inference

In the case of the scaled exponential distribution, it can be easily verified that the
maximum likelihood estimator (MLE) of θ is exactly the same as the BLUE θ ∗ in
(60). In the case of the two-parameter exponential distribution, the likelihood function
of the progressively Type-II right censored sample readily yields the MLEs of μ and
σ to be
m
(R ,...,Rm ) 1 (R ,...,Rm ) (R ,...,Rm ) 
μ̂ = Y1:m:n
1
and σ̂ = (Ri + 1) Yi:m:n
1
− Y1:m:n
1
. (65)
m
i=2

Remark 22 The BLUEs of μ and σ in (63) are thus unbiased versions of the MLEs
of μ and σ in (65), since
σ (m − 1)σ
E(μ̂) = μ + and E(σ̂ ) = . (66)
n m
(R ,...,R )
1 m (R ,...,R )
1 m
Now let Y1:m:n , . . . , Ym:m:n be a progressively Type-II right censored
sample from a general location-scale family of distributions with pdf g(x; μ, σ ) =
x−μ (R1 ,...,Rm ) (R1 ,...,Rm )
1
σ f ( σ ), where f (·) is the standard density. Let Xi:m:n = (Yi:m:n − μ)/σ ,
Progressive censoring methodology: an appraisal 239

i = 1, . . . , m, be the corresponding standardized progressively Type-II right censored


sample from f . Then the log-likelihood function, obtained from (1), is
m m
 
ln L(μ, σ ) = const. − m ln σ + ln f (xi:m:n ) + Ri ln 1 − F (xi:m:n ) . (67)
i=1 i=1

From (67), we have the likelihood equations for μ and σ


m m
∂ ln L 1 f (xi:m:n ) 1 f (xi:m:n )
=− + Ri = 0, (68)
∂μ σ f (xi:m:n ) σ 1 − F (xi:m:n )
i=1 i=1

m m
∂ ln L m 1 f (xi:m:n ) 1 f (xi:m:n )
=− − xi:m:n + Ri xi:m:n = 0.
∂σ σ σ f (xi:m:n ) σ 1 − F (xi:m:n )
i=1 i=1
(69)

In most cases, these two likelihood equations can be solved only by numerical meth-
ods and, consequently, various properties of the MLEs (such as bias, mean square er-
ror, coverage probabilities of asymptotic confidence intervals, accuracy of observed
Fisher information as an estimate of expected Fisher information, etc.) have been
evaluated only by extensive Monte Carlo simulations. Since the MLEs are to be de-
termined only numerically in these cases, some approximate estimators, which are ex-
plicit in form, have also been derived and their performance has been compared with
that of the MLEs. These results have been developed for a wide array of distributions
by many authors, including Balakrishnan and Kannan (2001), Aggarwala and Bal-
akrishnan (2002), Balakrishnan et al. (2003, 2004), Balakrishnan and Asgharzadeh
(2005), Rezaei (2007), and Balakrishnan and Hossain (2007). The maximum likeli-
hood estimation of parameters in this context has also been handled with the use of
EM algorithm by Ng et al. (2002, 2004) and Lin et al. (2006b). The existence and
uniqueness of the MLEs obtained from (68) and (69) have been established for the
normal and Weibull distributions by Balakrishnan and Mi (2003) and Balakrishnan
and Kateri (2007).

Example 2 Let us consider the data on times to breakdown of fluids, presented in


Table 1, and assume the Weibull (α, θ ) distribution. Then, as the log transformation
leads to an extreme value distribution with location parameter μ = ln θ and scale
parameter σ = 1/α, (68) and (69), when applied for the log-times in Table 1 with
the extreme value distribution, yield the MLEs of μ and σ to be μ̂ = 2.222 and
σ̂ = 1.026; see Viveros and Balakrishnan (1994) and Balakrishnan et al. (2004). The
last authors also found 95% confidence interval for μ to be (1.68, 4.00) and 90%
confidence interval for σ to be (0.71, 2.13) based on these MLEs with the aid of sim-
ulated percentage points of the pivotal quantities. Note that the confidence interval for
σ includes the value 1, which means the hypothesis that the Weibull shape parameter
equals 1 (i.e., the distribution is exponential) may be accepted in this case.
240 N. Balakrishnan

A final mention should be made to the work of Balakrishnan and Kim (2004, 2005)
who have discussed the maximum likelihood estimation of the parameters of the
bivariate normal BN(μX , μY , σX2 , σY2 , ρ) distribution based on progressively Type-II
right censored samples (with respect to one variable and their concomitants on the
other variable) and an EM algorithm implementation for this purpose.

9.3 Conditional inference

Since the MLEs of μ and σ can only be determined numerically from (68) and (69),
the derivation of the exact sampling distributions of the MLEs μ̂ and σ̂ (like in the
exponential case) becomes impossible in the general case. For this reason, Viveros
and Balakrishnan (1994) adopted Fisher’s (1934) conditional inference approach and
derived the exact conditional distributions of the pivotal quantities for μ and σ (based
on any equivariant estimators of μ and σ , such as the MLEs), conditional on ancillary
statistics computed from the given data.

Example 3 Viveros and Balakrishnan (1994) considered the progressively Type-II


right censored data on log-times to breakdown presented in Table 1, assumed the
underlying distribution to be extreme value with location μ and scale σ , and used
the MLEs μ̂ = 2.222 and σ̂ = 1.026 (see Example 2) to determine the exact con-
ditional confidence intervals for μ and σ . After computing the ancillary statistics
(R1 ,...,Rm )
(Yi:m:n − μ̂)/σ̂ , they obtained the 95% conditional confidence interval for μ to
be (1.63, 4.03) and the 90% conditional confidence interval for σ to be (0.72, 2.18).
It is heartening to see that these conditional confidence intervals are very close to the
unconditional confidence intervals presented earlier in Example 2.

The conditional approach has also been used by Childs and Balakrishnan (2000)
and Lin et al. (2004) to develop inference for the parameters of Laplace and log-
gamma distributions, respectively, based on progressively Type-II censored samples.

9.4 Robust inference

Basak and Balakrishnan (2003) studied the robust estimation of the location para-
meter μ based on a progressively Type-II right censored sample. Taking the scale
parameter σ to be 1, we have the likelihood equation for μ to be [see (68)]
m m
∂ ln L f (xi:m:n ) f (xi:m:n )
=− + Ri = 0, (70)
∂μ f (xi:m:n ) 1 − F (xi:m:n )
i=1 i=1

where xi:m:n = yi:m:n − μ. The MLE of μ, determined from (70), is efficient if the
assumed distribution is appropriate for the data, but it has the undesirable property
of being very sensitive to outlying observations. In a robust estimation procedure,
therefore, one solves instead the function
m m
− ψ1 (xi:m:n ) + Ri ψ2 (xi:m:n ) = 0, (71)
i=1 i=1
Progressive censoring methodology: an appraisal 241

Table 2 Values of the ORE and the MLE of μ for the original log-time
data in Table 1 and with one outlier on the right

Data ORE MLE

Original Data 2.307 2.206


One outlier 1.9947 → 3.9894 1.936 3.786
1.9947 → 7.9788 1.936 7.693
1.9947 → 11.9682 1.936 11.681
1.9947 → 19.9470 1.936 19.659

where the functions ψ1 and ψ2 are appropriately chosen to get various robust esti-
mates of μ. The extent by which an outlier influences the estimator is specified by the
“gross-error sensitivity” and the estimator which minimizes it is the “most robust esti-
mator”. Naturally, there has to be a tradeoff between efficiency and protection against
outliers. In this context, for a specified level of protection, the estimator which has the
maximum efficiency amongst all estimators providing that level of protection is the
“optimal robust estimator”. Basak and Balakrishnan (2003) have discussed properties
of the M-estimator, such as the influence function and the breakdown point, and then
derived the most robust estimator and the optimal robust estimator (ORE) and their
properties.

Example 4 By considering the progressively Type-II right censored data on log-times


to breakdown, presented in Table 1, assuming the underlying distribution to be ex-
treme value with location μ and scale σ = 1 (note that the MLE of σ in this case is
σ̂ = 1.026), and creating one outlier by multiplying the largest observation 1.9947
by 2, 4, 6 and 10, Basak and Balakrishnan (2003) obtained the estimates presented in
Table 2, thus displaying the insensitivity of the ORE and the sensitivity of the MLE
to the presence of outliers.

Open Problem 7 While the robust estimation of the location parameter μ, based on
progressively Type-II censored samples, has been addressed by Basak and Balakrish-
nan (2003), the joint robust estimation of the location μ and the scale σ still remains
open!

9.5 Other inferential methods

Sherif and Tan (1978) discussed structural inference for Weibull distribution, while
Lin et al. (2006a) studied Bayesian inference for the linear hazard rate distribution
using Markov Chain Monte Carlo methods based on progressively Type-II right cen-
sored samples. Bootstrap method of constructing confidence intervals for the location
and scale parameters has been discussed by Robinson (1983). Progressive interval
censoring model and associated inference has been discussed by Aggarwala (2001a);
see also Aggarwala (2001b). Bhattacharya (2007) has recently discussed restricted
maximum likelihood estimation and likelihood test procedures in the context of or-
der restricted inference for the exponential case under progressive censoring.
242 N. Balakrishnan

9.6 Prediction

Viveros and Balakrishnan (1994) used the conditional method of inference described
earlier in Sect. 9.3 to develop conditional prediction interval for an observation from
an independent future sample based on an observed progressively Type-II right cen-
sored sample. Balakrishnan and Lin (2002) discussed exact prediction intervals for
last censored failure times in a progressively Type-II right censored sample from an
exponential distribution, based on the BLUE described earlier in Sect. 9.1. Recently,
Basak et al. (2006) presented a detailed discussion on the point prediction of censored
failure times in a progressively Type-II right censored sample. They specifically ex-
amined the best linear unbiased predictors, the maximum likelihood predictors and
the conditional median predictors, and discussed their properties such as unbiased-
ness, consistency and efficiency, and then compared these predictors in terms of mean
squared prediction error.

9.7 Information measures

Chandrasekar and Balakrishnan (2002), continuing the work of Nagaraja (1994), dis-
cussed Tukey’s linear sensitivity measure in the multiparameter situation. This mea-
sure is closely related to the BLUEs and for location-scale family of distributions;
in fact, it turns out to be the inverse of the variance-covariance matrix of the BLUEs
given in (62). For example, for the two-parameter exponential distribution, the linear
(R1 ,...,Rm ) (R1 ,...,Rm )
sensitivity measure based on Y1:m:n , . . . , Ym:m:n turns out to be [using the
expressions in (64)]
 −1  
1
m
(m−1)n2
− n(m−1)
1
1 n2 n
S(Y; μ, σ ) = 2 = . (72)
σ − n(m−1)
1 1 σ2 n m
m−1

Note that the Fisher information matrix does not exist in this case, since μ is a thresh-
old parameter here.
Zheng and Park (2004) presented a direct decomposition form for the Fisher in-
formation in a progressively Type-II right censored sample, by using the Markovian
property, as sums of conditional Fisher information, corresponding to the smallest
usual order statistic from a left truncated distribution (see Result 2). This decom-
position, in conjunction with Balakrishnan–Childs–Chandrasekar representation, de-
scribed earlier in Sect. 4.4, enables the exact computation of the Fisher information.
Another approach has been taken by Ng et al. (2002, 2004) for the computation
of Fisher information in a progressively Type-II right censored sample. In the frame-
work of EM-algorithm for the likelihood estimation, they used the “missing infor-
mation principle” (see, for example, Louis 1982, and Tanner 1993) to compute the
Fisher information in a progressively Type-II right censored sample as the difference
between the complete sample information and the missing information.

Open Problem 8 The Fisher information and its computation has been discussed
only in the case when the progressively Type-II right censored sample is from a con-
tinuous cdf F (x). The natural question that arises here is whether this work can be
Progressive censoring methodology: an appraisal 243

carried out for the discrete case by making use of Balakrishnan-Dembińska represen-
tation, described earlier in Sect. 6!

Open Problem 9 The Fisher information in the usual order statistics as well as the
usual Type-I and Type-II censored data have been used to characterize some families
of distributions; see, for example, Gertsbakh and Kagan (1999), Zheng (2001), Zheng
and Gastwirth (2003), and Hofmann et al. (2005). This raises a question, whether
similar characterizations could be established based on Fisher information in a pro-
gressively Type-II right censored data!

10 Optimal progressive censoring

As seen already in Sects. 9.1 and 9.2, in the case of scaled exponential and two-
parameter exponential distributions, the precision of the BLUEs and the MLEs of
the parameters depend only on m and n and not on the progressive censoring scheme
(R1 , . . . , Rm ). Also, the Fisher information and the linear sensitivity measure in these
two cases, respectively, do not depend on (R1 , . . . , Rm ); see Sect. 9.7. For other dis-
tributions, however, for fixed values of the complete sample size n and the number
of complete failures m, one could look for an optimal choice of (R1 , . . . , Rm ) that
optimizes some criterion.
Balakrishnan and Aggarwala (2000) dealt with this problem and determined the
optimal choice of (R1 , . . . , Rm ), by minimizing the variance of the BLUE in (59)
for a scale family and the trace and determinant of the variance-covariance matrix
of the BLUEs in (62) for a location-scale family. The optimal censoring schemes
were determined numerically by an exhaustive search among all possible choices
of (R1 , . . . , Rm ), and tables were presented for different distributions and various
choices of m and n. But this approach becomes infeasible for large sample sizes.
Burkschat et al. (2006) considered a general family of distributions
  1/q
y −μ
F= F ≡ F (x) = 1 − 1−x− sign(q)x , q = 0 ,
σ 1−e , q =0

with x ≥ 0, 1 − sign(q)x > 0, μ ∈ R and σ > 0, which includes the exponential,


uniform and Pareto distributions. By studying the variances and covariance of the
BLUEs, and minimizing the trace and determinant of this variance-covariance matrix,
these authors proved that the optimal censoring scheme is always (n − m, 0, . . . , 0)
or (0, 0, . . . , n − m), depending on whether q > 0 or q < 0. Interestingly, by exam-
ining numerically the cases of normal, log-normal and extreme value distributions,
Balakrishnan and Aggarwala (2000) also found these extremal censoring schemes to
be optimal almost always.
Ng et al. (2002, 2004) approached this problem from the Fisher information point
of view and the optimality criteria, such as minimal missing information and the trace
and determinant of the inverse of the Fisher information matrix, were all determined
through the missing information principle (see Sect. 9.7).
244 N. Balakrishnan

Table 3 Comparison of optimal


progressive censoring scheme Censoring scheme Trace Determinant
with that in Example 1
Optimal 0.176956 0.006489
Example 1 0.239820 0.010806
Relative efficiency 135.5% 166.5%

Example 5 Consider the progressive Type-II right censored experiment in Ex-


ample 1, in which n = 19, m = 8, and the progressive censoring scheme was
(0, 0, 3, 0, 3, 0, 0, 5). By assuming an extreme value distribution for log-times to
breakdown, as done in Example 2, Ng et al. (2004) determined the optimal progres-
sive censoring schemes that minimize the trace and determinant of the inverse of the
Fisher information matrix to be the same, viz., (0, 11, 0, 0, 0, 0, 0, 0), and reported
the results as in Table 3.
Thus, the optimal progressive censoring scheme is considerably more efficient
than the scheme adopted in Example 1. By showing that the efficiencies change very
little, when the progressive censoring scheme departs slightly from the optimal one,
Ng et al. (2004) also demonstrated that their optimal sampling schemes are naturally
robust.
Balakrishnan et al. (2007) employed the maximum Fisher information as a cri-
terion for determining the optimal progressive censoring scheme for a wide variety
of lifetime distributions. They, in fact, hypothesized, based on numerical results, that
the optimal censoring scheme is always a one-step method, restricting censoring to
exactly one point in time. They showed that this optimal point of censoring can be at
the end (conventional right censoring) or after a certain proportion of observations,
depending on the underlying distribution and the parameter of interest. Balakrishnan
et al. (2007) then compared the Fisher information and the expected time for com-
pletion of the life-testing experiment of such an optimal one-step censoring scheme
with those of the conventional Type-II right censoring scheme.
Because the computational complexity increases with n, Hofmann et al. (2005)
adopted an asymptotic approach for the determination of optimal progressive censor-
ing schemes and presented several interesting theoretical and computational results,
including some discussion on the optimality of the two extremal censoring schemes.

11 Progressively censored reliability sampling plans

Let Y denote the quality of a product with some lifetime distribution F (·) and L
denote the one-sided lower specification of the limit for the quality, meaning that
the fraction of defectives is p = P (Y ≤ L) = Pr(X = ln Y ≤ L ), where L = ln L.
The design of acceptance sampling plans requires agreement between the producer
and the consumer, and lots for which p ≤ pα are presumed to be acceptable and
the consumer accepts these lots with probability at least 1 − α, while lots for which
p > pβ are presumed to be unacceptable and the consumer rejects these lots with
probability at least 1 − β; here, α and β are the producer’s risk and consumer’s
risk, respectively. For pertinent details, one may refer to Schilling (1982). Then, the
Progressive censoring methodology: an appraisal 245

well-known Lieberman–Resnikoff procedure stipulates to accept the lot under test if


μ̂ − σ̂ > L , where L is the specified lower limit for the quality of the product,
and (μ̂, σ̂ ) are some optimal estimates of the parameters (μ, σ ), such as the MLEs.
The construction of a reliability sampling plan is then to determine the acceptance
constant  and the minimum sample size n for two chosen points (pα , 1 − α) and
(pβ , 1 − β); see Balasooriya (2001).
Balasooriya et al. (2000) were the first to discuss the construction of such a reli-
ability sampling plan for the Weibull distribution (or the extreme value distribution
on the log-scale), based on progressively Type-II right censored samples. Similar
work was carried out by Balasooriya and Balakrishnan (2000) for the log-normal
distribution (or the normal distribution on the log-scale). These authors used ap-
proximate maximum likelihood estimators of μ and σ as they are explicit in form,
while the MLEs can be determined only numerically in this case (see Sect. 9.2).
Ng et al. (2004) discussed the implementation of their optimal progressive censor-
ing schemes (see Sect. 10) in this context and they used the MLEs μ̂ and σ̂ in the
Weibull plan rather than the approximate MLEs. They showed that the use of ap-
proximate MLEs, for the sample sizes determined, result in larger values of α (pro-
ducer’s risk) and β (consumer’s risk) than the pre-fixed ones, while those based on
the MLEs attain values closer to the pre-fixed values of α and β. It is important to
mention here that these authors also demonstrated that, under the same degree of cen-
soring and the same specifications, the use of optimal progressive censoring reduces
the required sample size n significantly. Balasooriya and Saw (1998) discussed the
construction of a reliability sampling plan for the two-parameter exponential distrib-
ution.

Open Problem 10 It might be of interest to consider here a general family of distri-


butions, which includes the above mentioned lifetime distributions as special cases,
and then determine the optimal progressive censoring schemes and the reliability
sampling plans based on them. This will then provide a flexible and practical ap-
proach to this issue!

12 Goodness-of-fit tests

In Sects. 9–11, several parametric inferential methods, based on progressively Type-


II right censored samples, have been developed by assuming specific distributions
for the observed data. For example, the progressively Type-II right censored times to
breakdown of insulating fluids in Table 1 were assumed in Example 1 to be from a
scaled exponential distribution, while the log-times to breakdown were assumed in
Example 2 to be from an extreme value distribution. Naturally, it will be of interest
to test the validity of this distributional assumption for the observed data.
To this end, Balakrishnan et al. (2002) proposed the test statistic
m−1
i=1 (m − i)zi
T=  (73)
(m − 1) mi=1 zi
246 N. Balakrishnan

1 m (R ,...,R )
to test that the observed progressively Type-II right censored sample (x1:m:n ,...,
(R1 ,...,Rm )
xm:m:n ) is from a scaled exponential distribution, with small and large values
of T leading to the rejection of exponentiality. The zi ’s in (73) are the normalized
spacings defined earlier in (25). These authors have shown that the null distribution
of T is exactly the same as the average of m − 1 i.i.d. Uniform(0,1) random variables
and, therefore, tends to normal N ( 12 , 12(m−1)
1
), when m is large. While Sanjel and
Balakrishnan (2007) discussed a Laguerre polynomial approximation, Balakrishnan
and Lin (2003) derived the exact null distribution of T . It should be mentioned that
the test statistic T in (73) was suggested by Tiku (1980) for complete and usual
Type-II censored samples and its power was examined empirically by Balakrishnan
(1983).

Example 6 Let us consider the progressively Type-II right censored times to break-
down in Table 1. For these data, we find T = 0.4325 and the approximate p-value
(based√ on the above mentioned normal approximation) to be p  2Φ((0.4325 −
0.5)/ 1/84) = 0.5362. We, therefore, fail to reject the null hypothesis that the ob-
served progressively Type-II right censored sample is from a scaled exponential dis-
tribution.

Balakrishnan et al. (2004) subsequently generalized this test and proposed the
statistic
m−1
∗ (m − i)z∗
T = i=2 m i∗ (74)
(m − 2) i=2 zi

1 m(R ,...,R )
to test that the observed progressively Type-II right censored sample (x1:m:n ,...,
(R1 ,...,Rm )
xm:m:n ) is from a specified location-scale family of distributions F (·), with small
and large values of T ∗ leading to rejection. The zi∗ ’s in (74) are the normalized spac-
ings defined as
(R ,...,Rm ) (R ,...,R )
1
xi:m:n − xi−1:m:n
1 m
zi∗ = (R ,...,Rm ) (R ,...,R )
, for i = 2, . . . , m,
1
μi:m:n − μi−1:m:n
1 m

where μ(R 1 ,...,Rm )


i:m:n denotes the mean of the ith progressively censored order statistic
from the standard distribution (with μ = 0 and σ = 1). These authors have shown
that the null distribution of T ∗ in (74) can be approximated by a normal distribution,
and also examined approximate and simulated power of this test in case of normal
and extreme value distributions for different alternative models.

Example 7 Let us consider the progressively Type-II right censored log-times to


breakdown in Table 1. For these data, note that an extreme value distribution was
assumed in Example 2. To test this assumption, we calculate from (74) that T ∗ =
0.5396; also, Balakrishnan et al. (2004) computed the approximate p-value (based
on the normal approximation) to be p  2{1 − Φ( 0.5396−0.5
√ )} = 0.7341, thus provid-
0.013597
ing strong evidence that the observed progressively Type-II right censored sample is
Progressive censoring methodology: an appraisal 247

from the extreme value distribution or the Weibull distribution in the original time
scale.

Open Problem 11 Many different types of goodness-of-fit tests have been studied
in the literature for complete and ordinary censored data, and these include the well-
known Shapiro–Wilk test, Anderson–Darling test, Kolmogorov–Smirnov test, tests
based on spacings, tests based on information measures, etc.; see, for example, the au-
thoritative book on this topic by D’Agostino and Stephens (1986). It will be of great
interest, of course, to generalize these test procedures to the case of progressively
Type-II censored data. One should also mention the recent work of Marohn (2002)
who exploited a characterization result to develop goodness-of-fit tests for a gener-
alized Pareto distribution and of Balakrishnan et al. (2007) who utilized Kullback–
Leibler information, based on progressively Type-II right censored samples, to de-
velop a test of exponentiality.

13 Nonparametric inference

In this section, we will summarize the developments on nonparametric inferential


methods, based on progressively Type-II right censored samples.

13.1 Confidence intervals for quantiles


(R ,...,R )
1 m (R ,...,R )
1 m
Let Y1:m:n , . . . , Ym:m:n be a progressively Type-II right censored sample
from a continuous cdf F (y), and ξp be the pth quantile of F (y), i.e., F (ξp ) = p,
for 0 < p < 1. Guilbaud (2001) then constructed nonparametric confidence intervals
for the quantile ξp , based on this progressively Type-II right censored sample. To
this end, since the probability integral transformation U = F (Y ), being monotone
(R1 ,...,Rm )
increasing, transforms the progressively censored order statistic Yi:m:n into the
(R1 ,...,Rm )
uniform progressively censored order statistic Ui:m:n for i = 1, . . . , m, we have
for 1 ≤ r < s ≤ m
(R1 ,...,Rm )
P (Yr:m:n ≤ ξp ≤ Ys:m:n
(R1 ,...,Rm )
)
= P (Ur:m:n
(R1 ,...,Rm )
≤ p) − P (Us:m:n
(R1 ,...,Rm )
≤ p)
s r
ai,s ai,r
= cs−1 (1 − p)γi − cr−1 (1 − p)γi , 0 < p < 1, (75)
γi γi
i=1 i=1

upon using the expression in (11). The integers r and s need to be determined so
(R1 ,...,Rm )
that the confidence level on the RHS of (75) is at least 1 − α in order for [Yr:m:n ,
(R1 ,...,Rm )
Ys:m:n ] to form a 100(1 − α)% confidence interval for the quantile ξp . Of course,
Guilbaud (2001) used his representation in (14) for the computation of this confidence
level.

Example 8 Let us consider the progressively Type-II right censored times to break-
down in Table 1. With these data, suppose we wish to construct a near 90% confidence
248 N. Balakrishnan

interval for the quantile ξ0.2 without making any assumption on the underlying distri-
bution F (y). Then, since the choice of r = 1, s = 6, and p = 0.2 results in the RHS
of (75) being 0.897, a near 90% confidence interval for the quantile ξ0.2 is simply
[0.19, 4.85], as presented by Guilbaud (2001).

Balakrishnan and Han (2007) discussed the determination of optimal r and s for
this confidence interval, by using the expected interval mass as an optimality crite-
rion, and they illustrated this method with intervals for the population median ξ0.5 .
Specifically, by noting that [see (23)]

(R ,...,Rm ) ! (R ,...,Rm ) !
1
E F Yi:m:n = E Ui:m:n
1


m
j + Rm−j +1 + · · · + Rm
=1−
j + 1 + Rm−j +1 + · · · + Rm
j =m−i+1

= ei (say),

Balakrishnan and Han (2007) considered the expected interval mass es − er of the in-
(R1 ,...,Rm ) (R1 ,...,Rm )
terval [Yr:m:n , Ys:m:n ] and discussed the determination of r and s that min-
imizes the expected interval mass of the 100(1 − α)% confidence interval for the
quantile ξp in (75).

Example 9 Corresponding to n = 19, m = 8, R = (0, 0, 3, 0, 3, 0, 0, 5), p = 0.2, and


(R) (R)
the confidence interval [Y1:8:19 , Y6:8:19 ], presented in Example 8, we have the corre-
sponding confidence level to be 0.8970 and the expected interval mass to be 0.3024.
If the goal had been to construct a 90% confidence interval for the quantile ξ0.2 , then
Balakrishnan and Han (2007) determined the optimal progressive censoring scheme
(R∗ ) (R∗ )
to be R∗ = (3, 0, 0, 0, 1, 0, 0, 7) and the confidence interval to be [Y1:8:19 , Y6:8:19 ]
with the corresponding confidence level being 0.9004 and the expected interval mass
being 0.3023. Note that the progressive censoring scheme in Table 1 chosen by
Viveros and Balakrishnan (1994) still provides a confidence interval for ξ0.20 that
is almost as efficient as the optimal one, if the minor difference between the nominal
and actual confidence levels is not of concern.

Example 10 Let us consider once again the progressively Type-II censored times to
breakdown in Table 1. With these data, suppose we wish to construct a confidence in-
terval for the population median ξ0.5 . As mentioned by Balakrishnan and Han (2007),
(R) (R)
the widest nonparametric confidence interval [Y1:8:19 , Y6:8:19 ] has a confidence level
of only 0.4929, while the widest confidence interval from the conventional Type-II
right censoring scheme (0 ∗ 7, 11) has a confidence level of only 0.1796. So, if the
goal was to construct a 90% confidence interval for the quantile ξ0.5 , the optimal
progressive censoring scheme (that minimizes the expected interval mass) would be
(R∗ ) (R∗ )
R∗ = (9, 0 ∗ 6, 2), and the confidence interval would be [Y3:8:19 , Y8:8:19 ] with the
corresponding confidence level being 0.9009.
Progressive censoring methodology: an appraisal 249

13.2 Tolerance and prediction intervals

Guilbaud (2004), by continuing along the lines of Sect. 13.1, discussed the construc-
(R1 ,...,Rm )
tion of nonparametric tolerance and prediction intervals. The interval [Yr:m:n ,
(R1 ,...,Rm )
Ys:m:n ] is said to be a 100(1 − α)% γ -tolerance interval if it contains at least γ
proportion of the population distribution with probability of at least 1 − α. Hence, the
integers r and s (for 1 ≤ r < s ≤ m) need to be determined so that
 Ys:m:n 

P f (y)dy ≥ γ = P F (Ys:m:n ) − F (Yr:m:n ) ≥ γ
Yr:m:n
= P (Us:m:n − Ur:m:n ≥ γ ) ≥ 1 − α. (76)

Guilbaud (2004) also discussed the construction of prediction intervals of the form
(R1 ,...,Rm ) (R1 ,...,Rm )
[Yr:m:n , Ys:m:n ] that will contain at least k order statistics of a future inde-
pendent sample of size  from the same continuous cdf F (y).

13.3 Counting and quantile processes, hazard rate estimation and homogeneity test

Bordes (2004) and Alvarez-Andrade and Bordes (2004) studied the counting and
quantile processes, associated with progressively Type-II right censored samples, and
their properties. Balakrishnan and Bordes (2004) applied these results to discuss the
nonparametric hazard rate estimation, while Alvarez-Andrade et al. (2007) used this
hazard rate estimate for testing the homogeneity of two or k distribution functions or,
equivalently, the equality of the corresponding hazard rate functions.

13.4 Exceedances, precedences and precedence tests


(R ,...,R )
1 m (R ,...,R )
1 m
Suppose (Y1:m:n 2
, . . . , Ym:m:n2 ) is a progressively Type-II right censored sam-
ple from a continuous cdf FY (y) and X1:n1 , . . . , Xn1 :n1 are the usual order statistics
from another independent sample from the same cdf. Then, Guilbaud (2001) and
Bairamov and Eryilmaz (2006) derived the probability distributions of exceedances,
which are simply the numbers of Xj :n1 ’s that exceed a chosen progressively censored
(R1 ,...,Rm )
order statistic Yi:m:n 2
. Ng and Balakrishnan (2005), motivated by a two-sample
hypothesis testing problem, discussed the distribution of precedences, which are the
(R1 ,...,Rm )
numbers of Xj :n1 ’s that precede Yi:m:n 2
. Clearly, the distributions of precedences
and exceedances are quite closely related. Specifically, by denoting the number of
Xj :n1 ’s less than Y1:m:n2 by M1 , and by Mi the number of Xj :n1 ’s between Yi−1:m:n2
and Yi:m:n2 , for i = 2, . . . , m, Ng and Balakrishnan (2005) derived the joint pmf of
(M1 , . . . , Mm ). They then used these Mi ’s to propose weighted precedence, weighted
maximal precedence, and maximal Wilcoxon rank-sum precedence test statistics for
testing the null hypothesis that FX (x) = FY (x) against the stochastically ordered al-
ternative FX (x) > FY (x); see also Balakrishnan and Ng (2006). These authors have
derived the exact null distributions of these test statistics as well as their power func-
tions under the Lehmann alternative FY (x) = (FX (x))γ , for γ > 1. By considering
progressively Type-II right censored samples from both populations FX (·) and FY (·),
Balakrishnan et al. (2007) have generalized these distributional results as well as the
precedence test.
250 N. Balakrishnan

Table 4 Robustness m
comparison of progressive Censoring scheme P (Am ) j =1 P (Aj )
censoring schemes
(0,0,7) 0.05944 0.16780
(0,1,6) 0.06001 0.16837
(0,2,5) 0.06079 0.16915
(0,3,4) 0.06192 0.17028
(0,4,3) 0.06369 0.17205
(0,5,2) 0.06687 0.17523
(0,6,1) 0.07430 0.18266
(0,7,0) 0.11146 0.21981

14 Progressive censoring under heterogeneity and robustness

All the inferential results discussed so far are under the assumption that the progres-
sively Type-II right censored sample arises from n iid variables. Earlier in Sect. 4.5,
Balakrishnan–Cramer representation of the joint density of the progressively Type-II
censored sample, arising from n independent and nonidentically distributed random
variables, was presented; see (18) and (19). By using these expressions, Balakrishnan
and Cramer (2007) established some properties of progressively Type-II right cen-
sored order statistics from independent and nonidentical variables, and especially in
the case when the underlying variables are exponentially distributed. They also used
these results to examine the effect of an outlier in the underlying sample and the ro-
bust estimation of the exponential mean, when an outlier is possibly present in the
progressively Type-II right censored sample.
With Xi ∼ Exp(λ), i ∈ {1, . . . , n} \ {}, and X ∼ Exp(μ)(μ ≥ λ), with X corre-
sponding to the outlier in the sample, and Aj denoting the event that the outlier X is
(R1 ,...,Rm )
the j th failure time Xj :m:n , Balakrishnan and Cramer (2007) showed that

(λ/μ)  j
γi
P (Aj ) = , (77)
n − 1 + (λ/μ) γi − 1 + (λ/μ)
i=2

where γi = m j =i (Rj + 1), i = 1, . . . , m. From (77), it can be easily verified that
P (Am ) is minimal for the right censoring scheme (i.e., R1 = · · · = Rm−1 = 0 and
Rm = n−m) and is maximal for the left censoring scheme (i.e., R1 = n−m and R2 =
· · · = Rm = 0). Thus, on the basis of P (Am ), the conventional Type-II censoring is the
most robust censoring scheme, while the left censoring scheme is the least robust one.
Another criterion one could employ in determining a robust censoring scheme is the
probability that the outlier is one of the observations, i.e., m ). Therefore,
j =1 P (Aj
as pointed out by Balakrishnan and Cramer (2007), either P (Am ) or m j =1 P (Aj )
could be used to choose between different progressive censoring schemes in terms of
robustness. For example, with n = 10, m = 3, and λ = μ/2, they reported the results
in Table 4.

Open Problem 12 The permanent expressions for densities of the usual order sta-
tistics from independent and nonidentically distributed variables have been used to
Progressive censoring methodology: an appraisal 251

establish many theoretical properties as well as applications to robustness studies;


see Balakrishnan (2007), for example. The question that arises now is to what extent
the permanent representation in (19) can be utilized to generalize the theoretical re-
sults on progressively Type-II right censored order statistics to the nonidentical case,
and to develop robust inference for other distributions in the presence of outliers!

15 Progressive Type-I right censoring

As mentioned already in Sect. 1.2, though progressive Type-I right censoring is


quite practical and useful while designing a life-testing experiment, it poses diffi-
culties in developing exact inference as well as in studying the theoretical proper-
ties of failure times arising from such a censoring scheme. For this reason, most
of the work on inference based on progressively Type-I right censored data deal
with maximum likelihood estimation and numerical methods for their determina-
tion; see, for example, Cohen (1963, 1966, 1975, 1976, 1991), Ringer and Sprinkle
(1972), Wingo (1973, 1993), Cohen and Norgaard (1977), Nelson (1982), Gibbons
and Vance (1983), Cohen and Whitten (1988), Balakrishnan and Cohen (1991), and
Wong (1993). Gajjar and Khatri (1969) considered the progressive Type-I right cen-
soring situation in which at each censoring time Ti the population parameters change,
and discussed the corresponding inference for log-normal and logistic distributions.
Progressive Type-I censoring model is also suitable for the case wherein patients
randomly withdraw from a study before its termination; see Sampford (1952) and
London (1988). Chatterjee and Sen (1973) and Majumdar and Sen (1978) discussed
nonparametric tests under progressive Type-I censoring, while Sinha and Sen (1982)
considered clinical trials with staggered entry times and random withdrawals under
the model of progressive Type-I censoring.

Open Problem 13 The Fisher information in the usual Type-I and Type-II censored
samples and their comparison have been discussed extensively in the literature. It
will, of course, be of interest to make a similar comparison between progressively
Type-I and Type-II censored samples!

16 Some other developments

I conclude this article by briefly describing in this section some other recent develop-
ments relating to progressive censoring.

16.1 Progressive censoring with random removals

In the progressive Type-II right censoring scheme, instead of fixing the progressive
censoring scheme (R1 , . . . , Rm ), one could make these removal numbers Ri ’s to be
random and ascribe a probability distribution for it. Yuen and Tse (1996), Tse and Xi-
ang (2003) and Tse and Yang (2003) have considered such a model with the Weibull
distribution and discussed inferential methods for the parameters of the underlying
252 N. Balakrishnan

model. For example, the binomial removal model would stipulate that at the time of
d
the first failure, a random number R1 = Binomial(n − 1, p) of surviving units will be
withdrawn from the test; at the time of the next failure, given R1 , a random number
d
R2 = Binomial(n − 2 − R1 , p) of surviving units will be withdrawn from the test,
and so on. As these authors have assumed this random removal process to be indepen-
dent of the failure times, the likelihood function is decomposable and the likelihood
estimation becomes simple.

Open Problem 14 A more realistic and interesting problem will be to allow for the
random removal process to be dependent on the failure times, thus allowing for larger
number of removals with a higher probability at larger failure times, and discuss
inferential methods under such a model.

16.2 Progressive hybrid censoring

Hybrid censoring schemes that combine the usual Type-I and Type-II censoring sce-
narios have been discussed extensively in the literature. In a similar vein, Childs et
al. (2007) recently introduced two progressive hybrid censoring schemes. Suppose
(R1 ,...,Rm ) (R1 ,...,Rm )
X1:m:n , . . . , Xm:m:n are the progressively Type-II right censored order sta-
tistics corresponding to n units on a life-test with the progressive censoring scheme
(R1 , . . . , Rm ). Then, the Type-I progressive hybrid censoring arises if the termination
time of the life-test is chosen to be T1∗ = min{Xm:m:n
(R1 ,...,Rm )
, T }, where T is a pre-fixed
time. Similarly, the Type-II progressive hybrid censoring corresponds to the situation
when the termination time of the life-test is chosen to be T2∗ = max{Xm:m:n
(R1 ,...,Rm )
, T }.
Note that it is guaranteed that the life-test would not last beyond time T in the former
case, while the latter case guarantees that at least m complete failures will be ob-
served, and that is why they are referred to as Type-I and Type-II progressive hybrid
censoring schemes, respectively. Childs et al. (2007) have then discussed exact con-
ditional inferential procedures for the exponential mean under these two generalized
progressive hybrid censoring schemes.
Specifically, in the case of Type-I progressive hybrid censoring scheme, these au-
thors have shown that the MLE of the exponential mean θ is given by
⎧  

⎪ 1
D

⎪ ∗
(Ri + 1)Xi:m:n + RD+1 T , if D = 1, 2, . . . , m − 1,

⎨D
i=1
θ̂ =

⎪ m


1

⎩m (Ri + 1)Xi:m:n , if D = m,
i=1


where D denotes the number of failures up to time T and Ri∗ = m k=i (Rk + 1). The
MLE of θ does not exist in this case if the number of failures up to the termination
time T1∗ is zero.
Progressive censoring methodology: an appraisal 253

Similarly, in the case of Type-II progressive hybrid censoring scheme, Childs et


al. (2007) have shown that the MLE of the exponential mean θ is given by
⎧ m

⎪ 1

⎪ (Ri + 1)Xi:m:n , if D = 0, 1, . . . , m − 1,

⎨m i=1
θ̂ =  

⎪ m D m−1


1
(Ri + 1)Xi:m:n + Xi:n + RD T , if D = m, · · · , n −

⎩D Ri ,
i=1 i=m+1 i=1

where RD = n − D − R1 − · · · − Rm−1 and Rm = 0 if D ≥ m.


They then derived the exact distribution of θ̂ in these two cases and discussed
the construction of exact confidence intervals for θ , and compared these confidence
intervals with those obtained by asymptotic and bootstrap methods.
Kundu and Joarder (2006) considered the Type-I progressive hybrid censoring
scheme under exponential distribution and compared the performance of the confi-
dence intervals based on the asymptotic distribution of the MLE, likelihood ratio test,
bootstrap method, and the Bayesian credible interval obtained with a gamma prior for
the parameter θ .

16.3 Competing risks under progressive censoring

Suppose n identical units are placed on a life-test and that every unit can fail due to
any one of k competing causes C1 , . . . , Ck , with possibly different lifetime distribu-
tions (it is common to assume them to belong to the same family, but with different
parameters). Suppose now a progressive Type-II right censoring is adopted in this
(R1 ,...,Rm )
life-test and that the observed data are of the form (Xi:m:n , δi ), where δi = j if
the ith failure is due to cause Cj for j = 1, . . . , k. Under this general set-up, with
exponential distribution as the underlying lifetime model, Kundu et al. (2004) have
discussed exact conditional inference methods based on the maximum likelihood es-
timates of the parameters θ1 , . . . , θk .

16.4 Step-stress test under progressive censoring

Consider a step-stress experiment in which N1 = n units are placed on test at an


initial stress level of s1 , run until time τ at which point the stress is changed to s2
and c1 surviving units are randomly withdrawn from the test. The test is continued
on N2 = n − n1 − c1 units until time 2τ , where n1 is the number of units that failed
at stress s1 , when the stress is changed to s3 and c2 surviving units are randomly
withdrawn from the test, and so on. Finally, at time kτ , all the surviving
 units are
withdrawn from the test and the experiment is terminated. Since n = ki=1 (ni + ci ),
the number of surviving units at time kτ will be

k k−1
ck = n − ni − ci = N k − n k .
i=1 i=1
254 N. Balakrishnan

In such a progressively Type-I right censored step-stress experiment, under the as-
sumption of a cumulative exposure model with exponential lifetimes, Gouno et al.
(2004), Han et al. (2006), and Balakrishnan and Han (2007) have all discussed the
maximum likelihood estimation of the parameters, their efficiencies, and the determi-
nation of optimal time duration τ that minimizes the trace or the determinant of the
variance-covariance matrix of the MLEs. Xie et al. (2007) have studied the same step-
stress problem in the case of progressive Type-II censoring, and they also examined
the optimal choice of the progressive censoring scheme in this context.

Acknowledgements I express my sincere thanks to Professors Maria Angeles Gil and Leandro Pardo,
Editors of TEST, for inviting me to prepare this Discussion Paper and also for their constant support
and encouragement through out the course of preparation and completion of this article. I am grateful to
the Natural Sciences and Engineering Research Council of Canada for funding this research. My special
thanks go to Ms. Debbie Iscoe for her expert assistance in typesetting the entire manuscript.

References

Aggarwala R (2001a) Progressive interval censoring: some mathematical results with applications to in-
ference. Commun Stat Theory Methods 30:1921–1935
Aggarwala R (2001b) Progressive censoring: a review. In: Balakrishnan N, Rao CR (eds) Advances in
reliability. Handbook of statistics, vol 20. North-Holland, Amsterdam, pp 373–429
Aggarwala R, Balakrishnan N (1996) Recurrence relations for single and product moments of progressive
Type-II right censored order statistics from exponential and truncated exponential distributions. Ann
Inst Stat Math 48:757–771
Aggarwala R, Balakrishnan N (1998) Some properties of progressive censored order statistics from arbi-
trary and uniform distributions with applications to inference and simulation. J Stat Plann Inference
70:35–49
Aggarwala R, Balakrishnan N (2002) Maximum likelihood estimation of the Laplace parameters based
on progressive Type-II censored samples. In: Balakrishnan N (ed) Advances on methodological and
applied aspects of probability and statistics. Taylor & Francis, Philadelphia, pp 159–167
Alvarez-Andrade S, Bordes L (2004) Empirical quantile process under Type-II progressive censoring. Stat
Probab Lett 68:111–123
Alvarez-Andrade S, Balakrishnan N, Bordes L (2007) Homogeneity tests based on several progressively
Type-II censored samples. J Multivar Anal 98(6):1195–1213
Arnold BC (1983) Pareto distributions. International Co-operative Publishing House, Fairland
Arnold BC, Balakrishnan N (1989) Relations, bounds and approximations for order statistics. Lecture
notes in statistics, vol 53. Springer, New York
Arnold BC, Balakrishnan N, Nagaraja HN (1992) A first course in order statistics. Wiley, New York
Bairamov I, Eryilmaz S (2006) Spacings, exceedances and concomitants in progressive type II censoring
scheme. J Stat Plann Inference 136:527–536
Balakrishnan N (1983) Empirical power study of a multi-sample test of exponentiality based on spacings.
J Stat Comput Simul 18:265–271
Balakrishnan N (1990) Improving the Hartley–David–Gumbel bound for the mean of extreme order sta-
tistics. Stat Probab Lett 9:291–294
Balakrishnan N (ed) (1992) Handbook of the logistic distribution. Dekker, New York
Balakrishnan N (1993) A simple application of binomial-negative binomial relationship in the derivation
of sharp bounds for moments of order statistics based on greatest convex minorants. Stat Probab Lett
18:301–305
Balakrishnan N (2007) Permanents, order statistics, outliers, and robustness. Rev Mat Complut 20:7–107
Balakrishnan N, Aggarwala R (2000) Progressive censoring: theory, methods, and applications.
Birkhäuser, Boston
Balakrishnan N, Asgharzadeh A (2005) Inference for the scaled half-logistic distribution based on pro-
gressively Type-II censored samples. Commun Stat Theory Methods 34:73–87
Progressive censoring methodology: an appraisal 255

Balakrishnan N, Bordes L (2004) Non-parametric hazard rate estimation under progressive Type-II censor-
ing. In: Balakrishnan N, Rao CR (eds) Advances in survival analysis. Handbook of statistics, vol 23.
North-Holland, Amsterdam, pp 227–249
Balakrishnan N, Brain C, Mi J (2002) Stochastic order and MLE of the mean of the exponential distribu-
tion. Methodol Comput Appl Probab 4:83–93
Balakrishnan N, Burkschat M, Cramer E, Hofmann G (2007) Can progressive censoring be better than
right censoring? (submitted)
Balakrishnan N, Childs A, Chandrasekar B (2002) An efficient computational method for moments of
order statistics under progressive censoring. Stat Probab Lett 60:359–365
Balakrishnan N, Cohen AC (1991) Order statistics and inference: estimation methods. Academic, Boston
Balakrishnan N, Cramer E (2007) Progressive censoring from heterogeneous distributions with applica-
tions to robustness. Ann Inst Stat Math (to appear)
Balakrishnan N, Cramer E, Kamps U (2001) Bounds for means and variances of progressive Type II
censored order statistics. Stat Probab Lett 54:301–315
Balakrishnan N, Cramer E, Kamps U (2005) Relation for joint densities of progressively censored order
statistics. Statistics 39:529–536
Balakrishnan N, Cramer E, Kamps U, Schenk N (2001) Progressive Type II censored order statistics from
exponential distributions. Statistics 35:537–556
Balakrishnan N, Dembińska A (2007a) Progressively Type-II right censored order statistics from discrete
distributions. J Stat Plann Inference (to appear)
Balakrishnan N, Dembińska A (2007b) Ordered random variables from discontinuous distributions (sub-
mitted)
Balakrishnan N, Habibi Rad A, Arghami NR (2007) Testing exponentiality based on Kullback–Leibler
information with progressively Type-II censored data. IEEE Trans Reliab 56:301–307
Balakrishnan N, Han D (2007) Optimal progressive Type-II censoring schemes for non-parametric confi-
dence intervals of quantiles. Commun Stat Simul Comput (to appear)
Balakrishnan N, Hossain A (2007) Inference for the Type-II generalized logistic distribution under pro-
gressive Type-II censoring. J Stat Comput Simul (to appear)
Balakrishnan N, Joshi PC (1981) Moments of order statistics from doubly truncated power function distri-
bution. Aligarh J Statist 1:98–105
Balakrishnan N, Joshi PC (1982) Moments of order statistics from doubly truncated Pareto distribution. J
Indian Stat Assoc 20:109–117
Balakrishnan N, Kannan N (2001) Point and interval estimation for parameters of the logistic distribution
based on progressively Type-II censored samples. In: Balakrishnan N, Rao CR (eds) Advances in
reliability. Handbook of statistics, vol 20. North-Holland, Amsterdam, pp 431–456
Balakrishnan N, Kannan N (2007) Recursive algorithm for single and product moments of progressive
Type-II right censored order statistics from logistic distribution (submitted)
Balakrishnan N, Kannan N, Lin C-T, Ng HKT (2003) Point and interval estimation for Gaussian distribu-
tion, based on progressively Type-II censored samples. IEEE Trans Reliab 52:90–95
Balakrishnan N, Kannan N, Lin C-T, Wu SJS (2004) Inference for the extreme value distribution under
progressive Type-II censoring. J Stat Comput Simul 74:25–45
Balakrishnan N, Kateri M (2007) On the maximum likelihood estimation of parameters of Weibull distri-
bution based on complete and censored data (submitted)
Balakrishnan N, Kim J-A (2004) EM algorithm and optimal censoring schemes for progressively Type-II
censored bivariate normal data. In: Balakrishnan N, Kannan N, Nagaraja HN (eds) Advances in rank-
ing and selection, multiple comparisons, and reliability: methodology and applications. Birkhäuser,
Boston, pp 21–45
Balakrishnan N, Kim J-A (2005) Point and interval estimation for bivariate normal distribution based on
progressively Type-II censored data. Commun Stat Theory Methods 34:1297–1347
Balakrishnan N, Lin C-T (2002) Exact linear inference and prediction for exponential distributions based
on general progressively Type-II censored samples. J Stat Comput Simul 72:677–686
Balakrishnan N, Lin C-T (2003) On the distribution of a test for exponentiality based on progressively
Type-II right censored spacings. J Stat Comput Simul 73:277–283
Balakrishnan N, Malov SV (2005) Some characterizations of exponential distribution based on progres-
sively censored order statistics. In: Balakrishnan N, Bairamov IG, Gebizlioglu OL (eds) Advances
on models, characterizations and applications. Chapman & Hall, New York, pp 97–109
Balakrishnan N, Mi J (2003) Existence and uniqueness of the MLEs for normal distribution based on
general progressively Type-II censored samples. Stat Probab Lett 64:407–414
256 N. Balakrishnan

Balakrishnan N, Ng HKT (2006) Precedence-type tests and applications. Wiley, Hoboken


Balakrishnan N, Ng HKT, Kannan N (2002) A test of exponentiality based on spacings for progressively
Type-II censored data. In: Huber-Carol C, Balakrishnan N, Nikulin MS, Mesbah M (eds) Goodness-
of-fit tests and model validity. Birkhäuser, Boston, pp 89–111
Balakrishnan N, Ng HKT, Kannan N (2004) Goodness-of-fit tests based on spacings for progressively
Type-II censored data from a general location-scale distribution. IEEE Trans Reliab 53:349–356
Balakrishnan N, Rao CR (1997) Large-sample approximations to the best linear unbiased estimation and
best linear unbiased prediction based on progressively censored samples and some applications. In:
Panchapakesan S, Balakrishnan N (eds) Advances in statistical decision theory and applications.
Birkhäuser, Boston, pp 431–444
Balakrishnan N, Sandhu RA (1995) A simple simulational algorithm for generating progressive Type-II
censored samples. Am Stat 49:229–230
Balakrishnan N, Sandhu RA (1996) Best linear unbiased and maximum likelihood estimation for expo-
nential distributions under general progressive Type-II censored samples. Sankhyā Ser B 58:1–9
Balakrishnan N, Sultan KS (1998) Recurrence relations and identities for moments of order statistics. In:
Balakrishnan N, Rao CR (eds) Order statistics: theory and methods. Handbook of statistics, vol 16.
North-Holland, Amsterdam, pp 149–228
Balakrishnan N, Tripathi RC, Kannan N (2007) On the joint distribution of placement statistics under
progressive censoring and applications to precedence test. J Stat Plann Inference (to appear)
Balasooriya U (2001) Progressively censored variables-sampling plans for life testing. In: Balakrishnan
N, Rao CR (eds) Advances in reliability. Handbook of statistics, vol 20. North-Holland, Amsterdam,
pp 457–467
Balasooriya U, Balakrishnan N (2000) Reliability sampling plans for the lognormal distribution, based on
progressively censored samples. IEEE Trans Reliab 49:199–203
Balasooriya U, Saw SLC (1998) Reliability sampling plans for the two-parameter exponential distribution
under progressive censoring. J Appl Stat 25:707–714
Balasooriya U, Saw SLC, Gadag VG (2000) Progressively censored reliability sampling plans for the
Weibull distribution. Technometrics 42:160–168
Basak I, Balakrishnan N (2003) Robust estimation under progressive censoring. Comput Stat Data Anal
44:349–376
Basak I, Basak P, Balakrishnan N (2006) On some predictors of times to failure of censored items in
progressively censored samples. Comput Stat Data Anal 50:1313–1337
Bhattacharya B (2007) Testing for ordered failure rates under general progressive censoring. J Stat Plann
Inference 137:1775–1786
Bordes L (2004) Non-parametric estimation under progressive censoring. J Stat Plann Inference 119:179–
189
Brascamp HJ, Lieb EH (1975) Some inequalities for Gaussian measures and the long-range order of the
one-dimensional plasma. In: Arthurs AM (ed) Functional integration and its applications. Clarendon
Press, Oxford, pp 1–14
Burkschat M, Cramer E, Kamps U (2006) On optimal schemes in progressive censoring. Stat Probab Lett
76:1032–1036
Cacciari M, Montanari GC (1987) A method to estimate the Weibull parameters for progressively censored
tests. IEEE Trans Reliab 36:87–93
Cacoullos T, Papathanasiou V (1985) On upper bounds for the variance of functions of random variables.
Stat Probab Lett 3:175–184
Chandrasekar B, Balakrishnan N (2002) On a multiparameter version of Tukey’s linear sensitivity measure
and its properties. Ann Inst Stat Math 54:796–805
Chandrasekar B, Leo Alexander T, Balakrishnan N (2002) Equivariant estimation for parameters of expo-
nential distributions based on Type-II progressively censored samples. Commun Stat Theory Methods
31:1675–1686
Chatterjee SK, Sen PK (1973) Nonparametric testing under progressive censoring. Calcutta Stat Assoc
Bull 22:13–50
Childs A, Balakrishnan N (2000) Conditional inference procedures for the Laplace distribution when the
observed samples are progressively censored. Metrika 52:253–265
Childs A, Chandrasekar B, Balakrishnan N (2007) Exact likelihood inference for an exponential parameter
under progressive hybrid censoring schemes. In: Vonta F, Nikulin M, Limnios N, Huber-Carol C (eds)
Statistical models and methods for biomedical and technical systems. Birkhäuser, Boston, pp 323–
334
Progressive censoring methodology: an appraisal 257

Cohen AC (1963) Progressively censored samples in life testing. Technometrics 5:327–329


Cohen AC (1966) Life testing and early failure. Technometrics 8:539–549
Cohen AC (1975) Multi-censored sampling in the three parameter Weibull distribution. Technometrics
17:347–351
Cohen AC (1976) Progressively censored sampling in the three parameter log-normal distribution. Tech-
nometrics 18:99–103
Cohen AC (1991) Truncated and censored samples: theory and applications. Dekker, New York
Cohen AC, Norgaard NJ (1977) Progressively censored sampling in the three parameter gamma distribu-
tion. Technometrics 19:333–340
Cohen AC, Whitten BJ (1988) Parameter estimation in reliability and life span models. Dekker, New York
Cramer E (2004) Logconcavity and unimodality of progressively censored order statistics. Stat Probab
Lett 68:83–90
Cramer E, Kamps U, Rychlik T (2004) Unimodality of uniform generalized order statistics, with applica-
tions to mean bounds. Ann Inst Stat Math 56:183–192
D’Agostino RB, Stephens MA (eds) (1986) Goodness-of-fit techniques. Dekker, New York
David HA, Nagaraja HN (2003) Order statistics, 3rd edn. Wiley, Hoboken
Dharmadhikari S, Joag-Dev K (1988) Unimodality, convexity, and applications. Academic, Boston
Epstein B, Sobel M (1953) Life testing. J Amer Stat Assoc 48:486–502
Epstein B, Sobel M (1954) Some theorems relevant to life testing from an exponential distribution. Ann
Math Stat 25:373–381
Fischer T, Balakrishnan N, Cramer E (2007) Mixture representation for order statistics from INID pro-
gressive censoring and its applications (submitted)
Fisher RA (1934) Two new properties of mathematical likelihood. Proc Roy Soc Ser A 144:285–307
Gajjar AV, Khatri CG (1969) Progressively censored samples from log-normal and logistic distributions.
Technometrics 11:793–803
Gertsbakh I, Kagan AM (1999) Characterization of the Weibull distribution by properties of the Fisher
information under Type-I censoring. Stat Probab Lett 42:99–105
Gibbons DI, Vance LC (1983) Estimators for the 2-parameter Weibull distribution with progressively cen-
sored samples. IEEE Trans Reliab 32:95–99
Gouno E, Sen A, Balakrishnan N (2004) Optimal step-stress test under progressive Type-I censoring. IEEE
Trans Reliab 53:388–393
Guilbaud O (2001) Exact non-parametric confidence intervals for quantiles with progressive Type-II cen-
soring. Scand J Statist 28:699–713
Guilbaud O (2004) Exact non-parametric confidence, prediction and tolerance intervals with progressive
Type-II censoring. Scand J Statist 31:265–281
Han D, Balakrishnan N, Sen A, Gouno E (2006) Corrections on “Optimal step-stress test under progressive
Type-I censoring”. IEEE Trans Reliab 55:613–614
Herd RG (1956) Estimation of the parameters of a population from a multi-censored sample. Ph.D. Thesis,
Iowa State College, Ames, Iowa
Hofmann G, Balakrishnan N, Ahmadi J (2005) Characterization of hazard function by Fisher information
in minima and upper record values. Stat Probab Lett 72:51–57
Hofmann G, Cramer E, Balakrishnan N, Kunert G (2005) An asymptotic approach to progressive censor-
ing. J Stat Plann Inference 130:207–227
Ibragimov IA (1956) On the composition of unimodal distributions. Teor Veroyatnost Primenen 1:283–
288
Johnson NL, Kotz S, Balakrishnan N (1994) Continuous univariate distributions, 2nd edn, vol 1. Wiley,
New York
Joshi PC (1978) Recurrence relations between moments of order statistics from exponential and truncated
exponential distributions. Sankhyā Ser B 39:362–371
Joshi PC (1979) A note on the moments of order statistics from doubly truncated exponential distribution.
Ann Inst Stat Math 31:321–324
Joshi PC (1982) A note on the mixed moments of order statistics from exponential and truncated exponen-
tial distributions. J Stat Plann Inference 6:13–16
Kamps U (1995a) A concept of generalized order statistics. J Stat Plann Inference 48:1–23
Kamps U (1995b) A concept of generalized order statistics. Teubner, Stuttgart
Kamps U, Cramer E (2001) On distributions of generalized order statistics. Statistics 35:269–280
Keseling C (1999) Characterizations of probability distributions by generalized order statistics (in Ger-
man). Ph.D. Thesis, Aachen University of Technology, Aachen, Germany
258 N. Balakrishnan

Keseling C, Kamps U (2003) A theorem of Rossberg for generalized order statistics. Sankhyā 65:259–270
Kundu D, Joarder A (2006) Analysis of Type-II progressively hybrid censored data. Comput Stat Data
Anal 50:2509–2528
Kundu D, Kannan N, Balakrishnan N (2004) Analysis of progressively censored competing risks data.
In: Balakrishnan N, Rao CR (eds) Advances in survival analysis. Handbook of statistics, vol 23.
North-Holland, Amsterdam, pp 331–348
Lin C-T, Wu SJS, Balakrishnan N (2004) Interval estimation of parameters of log-gamma distribution
based on progressively censored data. Commun Stat Theory Methods 33:2595–2626
Lin C-T, Wu SJS, Balakrishnan N (2006a) Monte Carlo methods for Bayesian inference on the linear
hazard rate distribution. Commun Stat Theory Methods 35:575–590
Lin C-T, Wu SJS, Balakrishnan N (2006b) Inference for log-gamma distribution based on progressively
Type-II censored data. Commun Stat Theory Methods 35:1271–1292
London D (1988) Survival models. Actex Publications, Connecticut
Louis TA (1982) Finding the observed information matrix when using the EM algorithm. J Roy Stat Soc
Ser B 44:226–233
Majumdar H, Sen PK (1978) Nonparametric tests for multiple regression under progressive censoring. J
Multivar Anal 8:73–95
Malmquist S (1950) On a property of order statistics from a rectangular distribution. Skand Aktuarietidskr
33:214–222
Mann NR (1969) Exact three-order-statistic confidence bounds on reliable life for a Weibull model with
progressive censoring. J Am Stat Assoc 64:306–315
Mann NR (1971) Best linear invariant estimation for Weibull parameters under progressive censoring.
Technometrics 13:521–533
Marohn F (2002) A characterization of generalized Pareto distributions by progressive censoring schemes
and goodness-of-fit tests. Commun Stat Theory Methods 31:1055–1065
Montanari GC, Cacciari M (1988) Progressively-censored aging tests on XLPE-insulated cable models.
IEEE Trans Electr Insulation 23:365–372
Moriguti S (1953) A modification of Schwarz’s inequality with applications to distributions. Ann Math
Stat 24:107–113
Nagaraja HN (1982) On the non-Markovian structure of discrete order statistics. J Stat Plann Inference
7:29–33
Nagaraja HN (1986) A note on conditional Markov property of discrete order statistics. J Stat Plann Infer-
ence 13:37–43
Nagaraja HN (1992) Order statistics from discrete distributions (with discussion). Statistics 23:189–216
Nagaraja HN (1994) Tukey’s linear sensitivity and order statistics. Ann Inst Stat Math 46:757–768
Nelson W (1982) Applied life data analysis. Wiley, New York
Nevzorov VB (2001) Records: mathematical theory. Translations of mathematical monographs, vol 194.
American Mathematical Society, Providence
Ng HKT, Balakrishnan N (2005) Weighted precedence and maximal precedence tests and an extension to
progressive censoring. J Stat Plann Inference 135:197–221
Ng HKT, Chan PS, Balakrishnan N (2002) Estimation of parameters from progressively censored data
using the EM algorithm. Comput Stat Data Anal 39:371–386
Ng HKT, Chan PS, Balakrishnan N (2004) Optimal progressive censoring plans for the Weibull distribu-
tion. Technometrics 46:470–481
Rényi A (1953) On the theory of order statistics. Acta Math Acad Sci Hung 4:191–231
Rezaei M (2007) Analysis of competing risks data for lognormal model under progressive Type-II censor-
ing. Commun Stat Theory Methods (to appear)
Ringer LJ, Sprinkle EE (1972) Estimation of the parameters of the Weibull distribution from multicensored
samples. IEEE Trans Reliab 21:6–51
Robinson JA (1983) Bootstrap confidence intervals in location-scale models with progressive censoring.
Technometrics 25:179–187
Rüschendorf L (1985) Two remarks on order statistics. J Stat Plann Inference 11:71–74
Sampford MR (1952) The estimation of response-time distributions, Part II. Biometrics 9:307–369
Sanjel D, Balakrishnan N (2007) A Laguerre polynomial approximation for a goodness-of-fit test for
exponential distribution based on progressively censored data. J Stat Comput Simul (to appear)
Schilling EG (1982) Acceptance sampling in quality control. Dekker, New York
Shah BK (1966) On the bivariate moments of order statistics from a logistic distribution. Ann Math Stat
37:1002–1010
Progressive censoring methodology: an appraisal 259

Shah BK (1970) Note on moments of a logistic order statistics. Ann Math Stat 41:2151–2152
Sherif A, Tan P (1978) On structural predictive distribution with Type-II progressively censored Weibull
data. Stat Hefte 19:247–255
Sinha AN, Sen PK (1982) Tests based on empirical processes for progressive censoring schemes with
staggered entry and random withdrawal. Sankhyā Ser B 44:1–18
Sukhatme PV (1937) Tests of significance for samples of the χ 2 population with two degrees of freedom.
Ann Eugen 8:52–56
Tanner MA (1993) Tools for statistical inference: observed data and data augmentation methods, 2nd edn.
Springer, New York
Thomas DR, Wilson WM (1972) Linear order statistic estimation for the two-parameter Weibull and ex-
treme value distributions from Type-II progressively censored samples. Technometrics 14:679–691
Tiku ML (1980) Goodness of fit statistics based on the spacings of complete or censored samples. Aust J
Stat 22:260–275
Tran TTH (2006) Discrete generalized order statistics. Ph.D. Thesis, University of Oldenburg, Oldenburg,
Germany
Tse S-K, Xiang L (2003) Interval estimation for Weibull-distributed life data under Type II progressive
censoring with random removals. J Biopharm Stat 13:1–16
Tse S-K, Yang C (2003) Reliability sampling plans for the Weibull distribution under Type II progressive
censoring with binomial removals. J Appl Stat 30:709–718
Viveros R, Balakrishnan N (1994) Interval estimation of parameters of life from progressively censored
data. Technometrics 36:84–91
Wingo DR (1973) Solution of the three-parameter Weibull equations by constrained modified quasilin-
earization (progressively censored samples). IEEE Trans Reliab 22:96–102
Wingo DR (1993) Maximum likelihood methods for fitting the Burr Type XII distribution to multiply
(progressively) censored life test data. Metrika 40:203–210
Wong JY (1993) Simultaneously estimating the three Weibull parameters from progressively censored
samples. Microelectron Reliab 3:2217–2224
Xie Q, Balakrishnan N, Han D (2007) Exact inference and optimal censoring scheme for a simple step-
stress model under progressive Type-II censoring (submitted)
Yuen H-K, Tse S-K (1996) Parameters estimation for Weibull distributed lifetimes under progressive cen-
soring with random removals. J Stat Comput Simul 55:57–71
Zheng G (2001) A characterization of the factorization of hazard function by the Fisher information under
Type-II censoring with application to the Weibull family. Stat Probab Lett 52:249–253
Zheng G, Gastwirth JL (2003) Fisher information in ordered randomly censored data with applications to
characterization problems. Stat Sin 13:507–517
Zheng G, Park S (2004) On the Fisher information in multiply censored and progressively censored data.
Commun Stat Theory Methods 33:1821–1835

You might also like