REVSTAT – Statistical Journal

Volume 0, Number 0, Month 0000, 000-000

Pitman’s measure of closeness for weighted random

variables

Authors: Mosayeb Ahmadi

– Department of Mathematics, University of Science and Technology of Mazandaran,
Behshahr, Iran ([email protected], [email protected])

G. R. Mohtashami Borzadaran ∗

– Department of Statistics, Ferdowsi University of Mashhad,

Mashhad, Iran ([email protected])
Mostafa Razmkhah
– Department of Statistics, Ferdowsi University of Mashhad,
Mashhad, Iran (razmkhah− [email protected])

Received: Month 0000 Revised: Month 0000 Accepted: Month 0000

Abstract:
• Statistical inference based on weighted random variables is developed in the sense of
Pitman’s measure of closeness. Some general formulas are presented to compute the
Pitman closeness of two weighted random variables to the parameter of interest. A new
general weighted model is also introduced and some properties are investigated. Also,
the concept of Pitman’s measure of closeness is used for measuring the nearness of
some weighted random variables with respect to each other. The results are illustrated
using some real data sets. Eventually, some conclusions are stated.

Key-Words:

• Weighted distribution; Exponential distribution; Ordered data; Population quantile;

Skew distributions.

AMS Subject Classification:

• 62G30, 62G05.

∗
Corresponding author
2 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

1. INTRODUCTION

There are situations in which the observed data would not be modeled
by a proper stochastic model. In these cases, some general models leading to
weighted distributions may be used. The concept of weighted distributions has
been introduced by Fisher (1934) and Rao (1965) in connection with model-
ing statistical data in situation that the standard distribution was not appro-
priate for their purposes. A formal definition of a weighted distribution for
random variable X with density (mass) function f (x) and non-negative weight
w(x) is obtained by g(x) = w(x)f (x)
E(w(X)) , where E(w(X)) > 0. The correspond-
ing weighted random variable is denoted by Xw in this paper. For more de-
tails about the statistical applications of weighted distributions, we refer the
readers to Patil and Rao (1977, 1978) and Rao (1985). From Gupta and Keat-
ing (1986) to Bartoszewicz (2009) and the references therein, the idea of the
weighted distributions is developed to various applications and properties. Saghir
et al. (2017) carried out a brief review of weighted distributions, and inves-
tigated the implications of the differing weight models as well as characteriza-
tions of these distributions based on a simple relationship between two trun-
cated moments. In recent years, this concept has been applied in many areas
of statistics such as biomedical, human study, wildlife populations, electronic,
etc. For example, Jiang (2020) used the weight function distribution to an-
alyze measurement errors of an electromagnetic flowmeter. It is of great im-
portance to note that the models of ordered data are special cases of weighted
distributions, for example the ith order statistic in sample of size n, denoted by
Xi:n 1 ≤ i ≤ n, from a population with probability density function (pdf) f (·)
and cumulative distribution function (cdf) F (·) is a weighted random variable
with weight function w(x) = (F (x))i−1 (F̄ (x))n−i . The nth upper k-record and
the nth lower
[ ]n−1are also k−1
k-record weighted random variables with weight functions
w(x) = − log F̄ (x) (F̄ (x)) and w(x) = [− log F (x)]n−1 F (x)k−1 , respec-
tively. For more details and literature on order statistics and record values, see for
example, Arnold et al. (1998, 2008) and David and Nagaraja (2003). The skew
distributions are also another kind of weighted distributions. Recently, Gómez-
Déniz et al. (2021) investigated the properties and applications of a new family
of skew distributions. Azzalini (2022) has done an overview on the progeny of
the skew-normal family.

It is known that the theory of estimation is a fundamental discipline dealing

with the specification of probabilistic model in terms of observed data. There are
many measures to evaluate the performance of an estimator like mean squared
error (MSE), mean absolute deviation, etc. The Pitman’s measure of closeness
(PMC) introduced by Pitman (1937) is another criterion which has been used by
several authors to compare the performance of the estimators. Let us first recall
the formal definition:

Definition 1.1. Let θ̂1 and θ̂2 be two estimators of a common parameter
PMC for weighted random variables 3

θ. The Pitman’s measure of closeness of θ̂1 relative to θ̂2 is denoted by π(θ̂1 , θ̂2 |θ)
and defined as

(1.1) π(θ̂1 , θ̂2 |θ) = Pr(|θ̂1 − θ| < |θ̂2 − θ|), ∀ θ ∈ Θ,

where Θ is the parameter space. If Pr(θ̂1 = θ̂2 ) = 0 and π(θ̂1 , θ̂2 |θ) ≥ 1/2 for all
θ ∈ Θ, with strict inequality holding for at least one θ, then θ̂1 is said to be a
Pitman closer estimator than θ̂2 with respect to θ.

It is obvious that π(θ̂1 , θ̂2 |θ) determines the relative frequency that the
estimator θ̂1 is closer than θ̂2 to the true but unknown value of the parameter
θ. Note that the absolute difference in (1.1) may be replaced by any other loss
function, in this case the associated probability is called generalized PMC. Several
developments in estimation theory via PMC arguments have been written by
Efron (1975). Also, Rao (1980, 1981) concluded noticeable comments about the
necessity of inference based on PMC. Keating (1985) derived various aspects of
PMC. Keating and Mason (1985) provided many practical examples in which
the PMC may be more useful than MSE. Sen (1986) presented the condition on
variance for an estimator being closer than another. Ghosh and Sen (1989) have
shown under certain conditions a median unbiased estimator of parameter is the
Pitman closest within a certain class of estimators. Some characterization results
in statistical inference and decision theory with examples have been explained by
Lee (1990) in view of the PMC for location and scale families. Nayak (1990) used
the PMC to find best estimators of a location or scale parameter. Many relevant
references can be found for example in the monograph by Keating et al. (1993).
A useful discussion to the question “Is Pitman closeness a reasonable criterion?”
with comments and rejoinder about this measure can be found in Robert et al.
(1993). Kourouklis (1996) used the PMC to get an improved estimation.

It is of great importance to note that two various statistics in estimating a

common parameter would be considered as two weighted random variables. This
motivated us to study the PMC in view of the weighted distributions. Toward
this end, the PMC of weighted random variables with respect to any unknown
parameter is first investigated. A new general weighted model is also introduced
and some properties are investigated. Then, the concept of PMC is used to
measure the nearness of some weighted random variables with respect to each
other.

The rest of the paper is organized as follows: In Section 2, some general

results are given to compute the PMC of the weighted random variables to the
parameter of interest. In Section 3, a new general weighted model is introduced
and some numerical results and conclusions are presented. A real data set is used
in this section to illustrate the proposed procedure. In Section 4, the PMC of two
weighted random variables with respect to another one is investigated. Moreover,
the results are applied to another real data set. Finally, some conclusions are
presented in Section 5.
4 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

2. GENERAL RESULTS

Let X be a random variable with pdf f (·) and cdf F (·) which is defined on
finite interval [a, b]. Assume Xw1 and Xw2 are two independent weighted random
variables of X with weight functions w1 (·) and w2 (·), respectively. Again, it
is pointed out that these random variables may be considered as two various
estimators for a common parameter θ. Based on (1.1), the PMC of Xw1 and Xw2
with respect to θ is given by

(2.1) π(Xw1 , Xw2 |θ) = P (|Xw1 − θ| < |Xw2 − θ|).

A general expression for the probability in (2.1) can be written as follows:

π(Xw1 , Xw2 |θ) = P (Xw1 < Xw2 , Xw1 + Xw2 > 2θ)
+P (Xw1 > Xw2 , Xw1 + Xw2 < 2θ)
∫ θ ∫ b ∫ b∫ b
= g1 (x)g2 (y)dydx + g1 (x)g2 (y)dydx
max{a,2θ−b} 2θ−x θ x
∫ θ∫ x ∫ min{b,2θ−a} ∫ 2θ−x
+ g1 (x)g2 (y)dydx + g1 (x)g2 (y)dydx
a a θ a
∫ θ ∫ b
= Ḡ2 (2θ − x)g1 (x)dx + Ḡ2 (x)g1 (x)dx
max{a,2θ−b} θ
∫ θ ∫ min{b,2θ−a}
(2.2) + G2 (x)g1 (x)dx + G2 (2θ − x)g1 (x)dx,
a θ

where for i = 1, 2, the functions gi (x) and Gi (x) are respectively the pdf and cdf
Bi (x)
of Xwi . It can be shown that Ḡi (x) = 1 − Gi (x) = E(w i (X))
, where
∫ ∞
Bi (x) = F̄ (x)E(wi (X)|X > x) = wi (t)f (t)dt.
x

Therefore, we have
∫ θ
B2 (2θ − x) w1 (x)f (x)
π(Xw1 , Xw2 |θ) = dx
max{a,2θ−b} E(w2 (X) E(w1 (X))
∫ b
B2 (x)w1 (x)f (x)
+ dx
θ E(w2 (X)) E(w1 (X))
∫ θ( )
B2 (x) w1 (x)f (x)
+ 1− dx
a E(w2 (X)) E(w1 (X))
∫ min{b,2θ−a} ( )
B2 (2θ − x) w1 (x)f (x)
(2.3) + 1− dx.
θ E(w2 (X)) E(w1 (X))
PMC for weighted random variables 5

Table 1. Some special cases of φi (u) related to PMC.

References φ1 (u) φ2 (u) Descriptions

Ahmadi and ui−1 (1 − u)m−i+1 uj−1 (1 − u)n−j+1 Order statistics in
Raqab (2013) two samples
Raqab and [− log(1 − u)]i [− log(1 − u)]j Record values from
Ahmadi (2012) two sequences
∑i R
−1
∑j S
−1
l (i)(1 − u) l (j)(1 − u)
Volterman et al. aR γl aS γl Two-sample
l=1 l=1
(2013) progressive type-II
censoring
Ahmadi and [− log(1 − u)]m ui−1 (1 − u)n−i+1 Record values
Mohtashami and order statistics
Botzadaran (2014)

( )
In special case of wi (x) = φi F (x) (i = 1, 2), by transforming u = F (x), we get
{∫ F (θ)
1
π(Xw1 , Xw2 |θ) = B2 (2θ − F −1 (u))φ1 (u)du
γ1 γ2 F (max{a,2θ−b})
∫ 1 }
−1
+ B2 (F (u))φ1 (u)du
F (θ)
{∫ F (θ)
1
− B2 (F −1 (u))φ1 (u)du
γ1 γ2 0
∫ F (min{b,2θ−a}) }
+ B2 (2θ − F −1 (u))φ1 (u)du
F (θ)
{∫ F (θ) ∫ F (min{b,2θ−a}) }
1
(2.4) + φ1 (u)du + φ1 (u)du ,
γ1 0 F (θ)

where γi = E[φi (U )] (i = 1, 2), such that U is a Uniform(0, 1) random variable.

Table 1 shows some special cases of φi (u) (i = 1, 2), which have been previously
studied by several authors.

The definition of the weighted random variables may be extended to a

random sample of size n. Let X1 , X2 , . . . , Xn be independent and identically
distributed (iid) random variables with cdf F (·) and pdf f (·), then a weighted
version of this sample can be defined by (X1∗ , X2∗ , ..., Xn∗ ), which have the joint
pdf

w(x∗1 , x∗2 , ...x∗n ) ∏ n

(2.5) hX1∗ ,X2∗ ,...,Xn∗ (x∗1 , x∗2 , ...x∗n ) = ∗ ∗ ∗
f (x∗i ).
E[w(X1 , X2 , ..., Xn )]
i=1
∏n
Notice that the joint pdf in (2.5) is a weighted version of i=1 f (x∗i ). Now, for
the special case of n = 2, this joint distribution can be expressed as
w(x, y)
(2.6) hX1∗ ,X2∗ (x, y) = f (x)f (y),
E[w(X1∗ , X2∗ )]
where include the joint distribution of two order statistics in a finite random
sample or two k-records in a sequence of iid random variables. By assuming
6 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

Table 2. Some special cases of φ(u, v) related to PMC.

Refrences φ(u, v) Descriptions

Balakrishnan et al. (2009b) um−1 (v − u)i−m−1 (1 − v)n−i Sample median and
the ith order statistic
Balakrishnan et al. (2009a) ui−1 (v − u)j−i−1 (1 − v)n−i Two order statistics
in one sample
[ ]j−i−1
[− log(1−u)]i
Ahmadi and 1−u
− log( 1−u
1−v
) Two upper records
Balakrishnan (2009) in one sequence
Ahmadi and [− log(u + 1 − v)]i−1 The ith lower and upper
Balakrishnan (2010b) records in one sequence
[ ]j−i−1
(1−v)k−1
Ahmadi and [− log(1 − u)]i − log( 1−u
1−v
) 1−u
Two k-records
Balakrishnan (2010a) in one sequence

w(x, y) = φ(u, v), where u = F (x) and v = F (y), the PMC of X1∗ and X2∗ with
respect to parameter θ can be determined as follows
{ ∫ F (θ) ∫ 1
1
π(X1∗ , X2∗ |θ) = φ(u, v)dvdu
A F (max(a,2θ−b)) F (2θ−F −1 (u))
∫ 1 ∫ 1 ∫ F (θ) ∫ u
+ φ(u, v)dvdu + φ(u, v)dvdu
F (θ) u 0 0
∫ F (min(b,2θ−a)) ∫ F (2θ−F −1 (u)) }
(2.7) + φ(u, v)dvdu ,
F (θ) 0

where A = E[φ(F (X1∗ ), F (X2∗ ))]. Some special cases of the weighted model in
(2.6) have been previously studied in view of PMC by some authors which are
summarized in Table 2.

Remark 2.1. It is worthwhile to note that not only for the special cases
presented in Tables 1 and 2, but also the PMC of the general versions of weighted
random variables with respect to the parameter of interest can be derived by the
use of (2.4) and (2.7).

3. A GENERAL WEIGHTED MODEL

In Section 2, a weighted random variable whose weight function was a func-

tion of the baseline cdf was considered in some special cases. Here, we introduce
a more general model contains all previous ones. Let X be a continuous random
variable with pdf f (·) and cdf F (·). For positive real constants α, β, γ and δ, one
may consider a general weighted random variable with the following pdf
1
(3.1) gF (x; λ) = [F (x)]α−1 [F̄ (x)]γ−1 [− log F (x)]β−1 [− log F̄ (x)]δ−1 f (x),
M (λ)
PMC for weighted random variables 7

where λ = (α, γ, β, δ) and M (λ) is the normalizing constant which is given by

M (λ) = M (α, γ, β, δ)
∫ 1
(3.2) = uα−1 (1 − u)γ−1 (− log u)β−1 [− log(1 − u)]δ−1 du.
0

The model defined in (3.1) includes many famous families of distributions, such
as distribution of order statistics (see David and Nagaraja, 2003), distributions
of upper and lower k-records (see, Arnold et al, 1998), Jones model (see Jones,
2004) and proportional (reversed) hazard rate model. This model can be also
considered as the pdf of a weighted k-record statistic and also weighted order
statistics. In the next two subsection we consider two famous member of the
proposed model.

3.1. Results based on records

Let us recall the sequences of upper k-record times, Tn,k , and upper k-
record values, Rn,k , which are defined as follow: T1,k = k with probability one,
R1,k = X1:k and for n ≥ 2

Tn,k = min{j : j > Tn−1,k , Xj > XTn−1,k −k+1:Tn−1,k },

and the nth upper k-record value is defined by Rn,k = XTn,k −k+1:Tn,k , for n ≥ 1.
For k = 1, the ordinary records are recovered. Then, the pdf of Rn,k is given by

kn [ ]n−1
fn,k (x) = − log F̄ (x) (F̄ (x))k−1 f (x),
Γ(n)
∫∞
where Γ(n) = 0 xn−1 e−x dx stands for the complete gamma function; see Arnold
et al. (1998) for more details. It is obvious that by taking γ = k and δ = n, the
pdf in (3.1) can be interpreted as the pdf of a weighted the nth upper k-record
w . Similarly, the weighted lower k-records may
statistic, which is denoted by Rn,k
be defined. The first question arises here is that whether k-records or the weighted
version ones are closer to the parameter of interest. By using (2.4) and performing
some algebraic calculations, it can be shown that the PMC of Rn,k w and R
n,k with
8 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

respect to θ is

1 ∑ ∑ (α − 1)
∞ α−1
w
πα,β (Rn,k , Rn,k |θ) = (−1)r Cj (β − 1)
M (α, β, k, n) r
j=β−1 r=0
{ ( )
Γ n, −(r + j + k) log F̄ (max(a, 2θ − b))
×
(r + j + k)n
∞ ( ( )
∑ k i Γ(n) − 2Γ n, −(r + j + 2k) log F̄ (θ)
+
i! (n + j + 2k)n+i
i=n
∫ − log F̄ (min(b,2θ−a))
+ ψ(y; i, j, r)dy
− log F̄ (θ)
∫ − log F̄ (θ) )}
(3.3) − ψ(y; i, j, r)dy ,
− log F̄ (max(a,2θ−b))
∫∞
where Γ(n, a) = a xn−1 e−x dx stands for the incomplete gamma function, the
function M (α, β, k, n) is as defined in (3.2), Cj (n) is the coefficient of wj in the
∑
expansion of ( ∞ wi n
i=1 i ) , and
( )i ( )k
ψ(y; i, j, r) = − log F̄ (2θ − F −1 (1 − e−y )) F̄ (2θ − F −1 (1 − e−y ))
×y n−1 e−(r+j+k)y .

Remark 3.1. It is clear that for other different versions of weight func-
tions, we can express the PMC via similar arguments in (3.3), for which they can
be interpreted as applications and specifications of several forms.

In the rest of this section, the population quantile is considered as the

parameter of interest θ. We recall that the population quantile ξp of order
p (0 < p < 1) of the cdf F (·) is defined by ξp = inf{x : F (x) ≥ p}. Balakr-
ishnan et al. (2009b, 2010) determined the closest order statistic in a random
sample of size n to a specific population quantile and specially studied the PMC
of sample median to population median. Ahmadi and Balakrishnan (2009) ex-
amined the PMC of record statistics to the population quantiles of location-scale
family of distributions. Moreover, Ahmadi and Balakrishnan (2010a) investi-
gated the PMC of k-records and Ahmadi and Balakrishnan (2010b) obtained
the PMC of current records for location-scale families. Similar work was done
by Razmkhah and Ahmadi (2013) regarding the current k-records. The PMC
of upper (lower) records in two independent sequences of iid continuous ran-
dom variables to population quantiles was studied by Raqab and Ahmadi (2012).
Similarly, the PMC of order statistics in a two-sample problem was investigated
by Ahmadi and Raqab (2013). Moreover, a comparison study for order statis-
tics and records was performed with some remarks by Ahmadi and Mohtashami
Borzadaran (2014). Davies (2021) studied some PMC results for type-I hybrid
censored data from exponential distribution. Morabbi and Razmkhah (2020)
used the PMC to get the quantile estimation based on modified ranked set sam-
pling schemes. The weighted versions of the aforementioned papers, for example,
PMC for weighted random variables 9

w,U
Table 3. Values of πα,β (Rn,k U
, Rn,k |ξp ) for standard exponential distribution.
p
n k β α 0.10 0.25 0.5 0.60 0.75 0.80 0.90 0.95 0.99
3 3 1 2 0.4036 0.4044 0.4439 0.4899 0.5654 0.5819 0.5955 0.5964 0.5964
3 0.3310 0.3315 0.3805 0.4540 0.5989 0.6347 0.6667 0.6689 0.6690
4 0.2754 0.2757 0.3213 0.4077 0.6099 0.6664 0.7204 0.7244 0.7246
5 0.2321 0.2322 0.2702 0.3591 0.6056 0.6827 0.7613 0.7676 0.7679
2 1 0.6789 0.6745 0.5730 0.4839 0.3626 0.3397 0.3222 0.3212 0.3211
2 0.5748 0.5759 0.5727 0.5437 0.4675 0.4463 0.4267 0.4253 0.4252
3 0.4888 0.4902 0.5363 0.5613 0.5467 0.5319 0.5130 0.5113 0.5112
4 0.4184 0.4194 0.4838 0.5497 0.6030 0.5985 0.5838 0.5816 0.5815
5 0.3609 0.3614 0.4273 0.5201 0.6400 0.6491 0.6414 0.6392 0.6391
4 3 1 2 0.4291 0.4292 0.4371 0.4560 0.5166 0.5392 0.5674 0.5706 0.5709
3 0.3720 0.3720 0.3812 0.4088 0.5149 0.5596 0.6200 0.6275 0.6280
4 0.3254 0.3254 0.3335 0.3633 0.5012 0.5663 0.6609 0.6737 0.6746
5 0.2869 0.2870 0.2933 0.3219 0.4800 0.5631 0.6928 0.7116 0.7130
2 1 0.6884 0.6881 0.6554 0.5937 0.4321 0.3790 0.3183 0.3120 0.3116
2 0.6123 0.6123 0.6099 0.5914 0.4988 0.4559 0.3959 0.3883 0.3877
3 0.5461 0.5462 0.5566 0.5678 0.5440 0.5163 0.4632 0.4546 0.4539
4 0.4890 0.4890 0.5032 0.5325 0.5711 0.5618 0.5211 0.5118 0.5110
5 0.4397 0.4397 0.4535 0.4921 0.5835 0.5943 0.5708 0.5613 0.5603
4 4 1 2 0.4167 0.4169 0.4456 0.4889 0.5612 0.5745 0.5830 0.5833 0.5833
3 0.3498 0.3500 0.3871 0.4577 0.5992 0.6291 0.6495 0.6501 0.6502
4 0.2959 0.2960 0.3318 0.4166 0.6189 0.6674 0.7028 0.7040 0.7041
5 0.2522 0.2523 0.2830 0.3722 0.6246 0.6927 0.7456 0.7477 0.7477
2 1 0.6601 0.6589 0.5828 0.4942 0.3715 0.3518 0.3402 0.3399 0.3399
2 0.5715 0.5718 0.5667 0.5370 0.4599 0.4417 0.4289 0.4285 0.4285
3 0.4949 0.4953 0.5269 0.5485 0.5310 0.5179 0.5056 0.5051 0.5050
4 0.4297 0.4299 0.4763 0.5371 0.5855 0.5809 0.5710 0.5703 0.5703
5 0.3743 0.3744 0.4232 0.5103 0.6250 0.6317 0.6264 0.6257 0.6257
3 1 0.7803 0.7753 0.5810 0.4165 0.2483 0.2289 0.2198 0.2197 0.2197
2 0.6988 0.6983 0.6137 0.4973 0.3368 0.3136 0.3015 0.3012 0.3012
3 0.6226 0.6231 0.6117 0.5510 0.4180 0.3926 0.3778 0.3774 0.3774
4 0.5537 0.5542 0.5856 0.5789 0.4892 0.4639 0.4468 0.4463 0.4463

weighted order statistics and weighted k-records may be of great interest which
are discussed in this paper.

From (3.3), it is seen that the PMC is not distribution-free. So, the baseline
distribution has to be specified. Therefore, for as an example, the standard
exponential distribution has been considered and the PMC investigated for the
population quantiles. With this in mind, let {Xi , i ≥ 1} be a sequence of iid
random variables from standard exponential distribution. The PMC of Rn,k w and

Rn,k with respect to the pth quantile of this distribution, can be simplified as
follows

1 ∑∞ α−1∑ (α − 1)
w
πα,β (Rn,k , Rn,k |ξp ) = (−1)r Cj (β − 1)
M (α, β, k, n) r
j=β−1 r=0
{ ∑ ∞ (
Γ(n) k i Γ(n) − 2Γ(n, −(r + j + 2k) log q)
× +
(r + j + k)n i! (r + j + 2k)n+i
i=n
∫ − log q2 ∫ − log q )}
(3.4) +q 2k
h(y; i, j, r)dy − q 2k
h(y; i, j, r)dy ,
− log q 0

where ( )i
h(y; i, j, r) = − log(q 2 ey ) y n−1 e−(r+j)y .

n,k |ξp ) have been computed

w ,R
Using (3.4), the numerical values of πα,β (Rn,k
for n = 3, 4, k = 3, 4 and some selected values of α, β and p. The results are
presented in Table 3.
10 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

The bold numbers in Table 3 are referring to the fact that the weighted
k-records are Pitman closer than the usual k-records to the specific population
quantile. From this table, it is intuitively observed that for given α and β,
there exist real values α0 and p0 ∈ (0, 1) such that for α ≤ α0 and p ≤ p0 ,
U,w U |ξ ) > 0.5. Similarly for α > α , there exist a real value p ∈ (0, 1),
π(Rn,k , Rn,k p 0 ∗
U,w
so that for p ≥ p∗ , π(Rn,k U |ξ ) > 0.5.
, Rn,k p

3.2. Results based on order statistics

We recall that the pdf of the rth order statistic from an iid sample of size
n from cdf F (·) and pdf f (·) is given by
( )
n
(3.5) fr:n (x) = k F r−1 (x)F̄ n−r (x)f (x).
r
Let us denote the weighted order statistic by Xr:n w . Here, the PMC of X w and
r:n
Xr:n with respect to the pth quantile ξp is studied when the baseline distribution
is standard exponential distribution. From (2.4) and (3.5), we find
{∑ r−1 ∑∞ ( )
1 j r−1
πβ,δ (Xr:n , Xr:n |ξp ) =
w
(−1) Ck (β − 1)
M (r, n − r + 1, β, δ) j
j=0 k=1
(
Γ(δ) 1
× −
(j + n − r + k + 1) M (r, n − r + 1, 1, 1)
r+i ( ( )(r+i)
∑∑
n−r
(−1)i+l n−r l (1 − p)
2l
× i
r+i
i=0 l=0
( ) ))
Γ(δ) − Γ δ, −(j + n − r + k − l + 1) ln(1 − p)
×
(j + n − r + k − l + 1)δ
∞ (
( )(2r+i−1)
1 ∑ 2r+i−1
n−r ∑ ∑ (−1)i+j n−r
i j
+
M (r, n − r + 1, 1, 1) r+i
i=0 j=0 k=1
( ) )}
Γ(δ) − 2Γ δ, −(j + n − r + k + 1) ln(1 − p)
(3.6) ×Ck (β − 1) .
(j + n − r + k + 1)δ
For extreme order statistics X1:n and Xn:n we have
∑∞ {
1 Γ(δ)
w
πβ,δ (X1:n , X1:n |ξp ) = Ck (β − 1)
M (1, n, β, δ) (n + k)
k=1
( )( )
∑∑
n−1 i+1
(−1)i+j n−1i
i+1
j
+n
i+1
i=0 j=0
( ( )
Γ(δ) − 2Γ δ, −(n + k + j) ln(1 − p)
×
(n + k + j)δ
( ) )}
Γ(δ) − Γ δ, −(n + k − j) ln(1 − p)
− (1 − p)2j
.
(n + k − j)δ
PMC for weighted random variables 11

w
Table 4. Values of πβ,δ (Xr:n , Xr:n |ξp ) for standard exponential distribution.

p
n r β δ 0.10 0.25 0.5 0.75 0.90
5 1 2 3 0.2326 0.5067 0.7858 0.8018 0.8019
3 2 0.5243 0.5585 0.5299 0.5276 0.5275
3 2 3 0.3628 0.3658 0.4716 0.6292 0.6372
3 2 0.6371 0.6360 0.5349 0.3758 0.3629
4 2 3 0.4301 0.4302 0.4447 0.5515 0.5751
3 2 0.7064 0.7063 0.6731 0.4180 0.3005
8 1 2 3 0.2361 0.6905 0.8162 0.8169 0.8169
3 2 0.5307 0.5699 0.5636 0.5635 0.5635
7 2 3 0.4833 0.4834 0.4839 0.5180 0.5341
3 2 0.7338 0.7337 0.7313 0.5615 0.2974
10 2 2 3 0.2741 0.5501 0.7414 0.7429 0.7434
3 2 0.5131 0.5259 0.5170 0.5024 0.5000
5 2 3 0.3758 0.3777 0.5190 0.6237 0.6241
3 2 0.5966 0.5958 0.4858 0.4038 0.4033
7 2 3 0.4216 0.4307 0.4415 0.5573 0.5693
3 2 0.6758 0.6547 0.6320 0.3869 0.3454

and
∑∞ { n−1
∑ ( )
1 j n−1 Γ(δ)
w
πβ,δ (Xn:n , Xn:n |ξp ) = Ck (β − 1) (−1)
M (n, 1, β, δ) j (j + k + 1)
k=1 j=0
( ) ( )
∑
2n−1
j 2n − 1 Γ(δ) − 2Γ δ, −(j + k + 1) ln(1 − p)
+ (−1)
j (j + k + 1)δ
j=0
n (
( )(n)
∑ n−1
n−r ∑∑ (−1)j+l n−1
j l (1 − p)
2l
−n
n+i
i=0 j=0 l=0
( ) )}
Γ(δ) − Γ δ, −(j + k − l + 1) ln(1 − p)
× ,
(j + k − l + 1)δ

r:n |ξp ) have been

w ,X
respectively. Using (3.6), the numerical values of πβ,δ (Xr:n
computed for some selected values of n,r, β, δ and p. The results are presented
in Table 4. The bold numbers in this table are referring to the fact that the
weighted order statistics are Pitman closer than the usual order statistics to the
specific population quantile.

From Table 4 we observe that, if β ≤ δ (or β > δ), there exist p0 ∈ (0, 1)
such that for p ≥ p0 (or p ≤ p0 ), we get πβ,δ (Xn:n
w ,X
n:n |ξp ) ≥ 0.5.

4. PITMAN CLOSENESS AND WEIGHTED RANDOM VARI-

ABLES

Note that in the field of PMC, there is no necessity to restrict our attention
to parameters and their estimators. There are situations in which the problem of
closeness of some random variables with respect to each others may be of great
12 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

importance. For instance, if Z is a random variable and X and Y are two other
random variables, then for an appropriate measure of distance d(·, ·), we may
define GP N (X, Y, Z) = P (d(X, Z) < d(Y, Z)) to determine the closer random
variable between X and Y to Z; see for example, Mendes and Merkle (2005).
Here, we focus on the PMC by using the distance d(X, Z) = |X − Z|.

Now, consider the situation in which the cdf of a distribution varies in some
steps when the time proceeds. Let us show the corresponding random variable
to the baseline (initial) distribution by X. Then, the transformed distribution at
the ith temporal step can be characterized by a weighted random variable, such
as Xwi (i ≥ 1) with pdf gi (·). So, the PMC of any two transformed distribution
to the baseline distribution is discussed in what follows.

Let Xw1 , Xw2 and Xw3 be independent weighted random variables which
are considered to describe the transformations on a baseline distribution by pdf
f (·) and cdf F (·). Therefore, the pdf of Xwi is given by

wi (x)f (x)
(4.1) gi (x) = , E(wi (X)) > 0, i = 1, 2, 3.
E(wi (X))
Here, we generally focus our attention to compute the probability of closeness of
Xw1 and Xw2 with respect to Xw3 . Note that in the special case of w3 (x) = 1, in
fact the closeness of two transformed distributions with respect to the baseline
distribution is studied. Using (1.1), we get

π(Xw1 , Xw2 |Xw3 ) = P (|Xw1 − Xw3 | < |Xw2 − Xw3 |)

= P (−|Xw2 − Xw3 | < Xw1 − Xw3 < |Xw2 − Xw3 |)
= P (Xw2 < Xw1 < 2Xw3 − Xw2 , Xw2 < Xw3 )
+P (2Xw3 − Xw2 < Xw1 < Xw2 , Xw2 > Xw3 ).

Assuming the random variables are defined on [a, b], we have

∫ b ∫ z ∫ 2z−y
π(Xw1 , Xw2 |Xw3 ) = g1 (x)g2 (y)g3 (z)dxdydz
a a y
∫ b∫ b∫ y
+ g1 (x)g2 (y)g3 (z)dxdydz
a z 2z−y
{∫ b∫ z
1
= A1 (y)F̄ (y)w2 (y)f (y)w3 (z)f (z)dydz
Λ a a
∫ b∫ z
− A1 (2z − y)F̄ (2z − y)w2 (y)f (y)w3 (z)f (z)dydz
a a
∫ b∫ b
+ A1 (2z − y)F̄ (2z − y)w2 (y)f (y)w3 (z)f (z)dydz
a z
∫ b∫ b }
(4.2) − A1 (y)F̄ (y)w2 (y)f (y)w3 (z)f (z)dydz ,
a z

where
Λ = E(w1 (X))E(w2 (X))E(w3 (X))
PMC for weighted random variables 13

and ∫ b
1
A1 (x) = E(w1 (X)|X > x) = w1 (t)f (t)dt.
F̄ (x) x

Let us now suppose that wi (x) = φi (F (x)), for i = 1, 2, 3, then by using

transformations u = F (y) and v = F (z), the PMC in (4.2) can be rewritten as
follows:
{∫ 1∫ v
1
π(Xw1 , Xw2 |Xw3 ) = [C1 (u, v) − C2 (u, v)]dudv
Λ 0 0
∫ 1∫ 1 }
− [C1 (u, v) − C2 (u, v)]dudv ,
0 v

where
C1 (u, v) = A1 (F −1 (u))(1 − u)φ2 (u)φ3 (v)
and

C2 (u, v) = A1 (2F −1 (v) − F −1 (u))F̄ (2F −1 (v) − F −1 (u))φ2 (u)φ3 (v).

By selecting various choices of wi (x), specially those of presented in the

previous sections, the above probability of closeness can be derived via tedious
calculations.
It is worth to mention that formula (4.2) can also be used to assess the relationship
among different statistics. Suppose we have three independent sample from the
same distribution. It is known that in the context of nonparametric X[np]:n used
as an estimator for ξp . Denote the the sample mean of a standard exponential
population by X̄, and consider the sample quantiles of two independent standard
exponential populations by Xr:n and Xs:n , respectively. Then by using (4.2) we
have
( )( ) ∑ r−1 ∑
s−1 ( )( )
n n i+j r − 1 s−1
π(Xr:n , Xs:n |X) = n rsn
(−1)
r s i j
i=1 j=1
{ (
(n − r + i + j)−1 1 1
× −
(2n − r − s + i + j + 2) n n (3n − r − s + i + j + 3)n
)
1 1
− +
(3n − 2r + 2i + 2)n (4n − 3r − s + 3i + j + 4)n
}
(3n − r − s + i + j + 2)−n
+ .
(r − s − i + j)(2n − r − s + i + j + 2)

5. APPLICATIONS

To illustrate the performance of the proposed procedure in Sections 4 and

5, we use two real data sets in the following examples.
14 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

Example 1 (Telephone calls). Table 5 contains the data concerning the times
(in minutes) between 48 consecutive telephone calls to a company’s switchboard
which is presented by Castillo et al. (2005). They assumed that the data come
from the exponential distribution.

Table 5. Times (in minutes) between 48 consecutive calls.

1.34 0.14 0.33 1.68 1.86 1.31 0.83 0.33 2.20 0.62 3.20 1.38
0.96 0.28 0.44 0.59 0.25 0.51 1.61 1.85 0.47 0.41 1.46 0.09
2.18 0.07 0.02 0.64 0.28 0.68 1.07 3.25 0.59 2.39 0.27 0.34
2.18 0.41 1.08 0.57 0.35 0.69 0.25 0.57 1.90 0.56 0.09 0.28

By using the definition of k-records, as presented in subsection 3.1, from

this data set, the second upper 2-records (values of R2,2 ) have been extracted
which are 0.33, 1.86, 1.38, 0.44, 0.51, 1.85, 0.07, 0.68, 2.39, 0.34, 0.57 and 0.57.
Note that these data are indeed the second largest observations in the partial
samples. Moreover, after observing each data point, the procedure of collecting
the next second upper 2-record has been restarted. Moreover, the initial sample
maxima to attain the second upper 2-records are observed as

1.34, 2.2, 3.2, 0.96, 0.59, 2.18, 0.64, 1.07, 3.25, 2.18, 1.08, 0.69.

The null hypothesis that the above data are coming from a weighted dis-
tribution with pdf (3.1) and the parameters λ = (α, β, 2, 2) (distribution of the
second upper 2-record), is checked by using the Kolmogorov-Smirnov (K-S) dis-
tances between the empirical distribution functions and the fitted distribution
function. The observed MLEs of α and β are 3.5 and 0.8, respectively. Also, the
observed value of K-S statistic is 0.2158 and the associated p-value is 0.6311. So,
based on these observations, the weighted distribution is adequate for the data
regarding the sample maxima to attain the second upper 2-records. That is, one
may accept that these data are some observed values of R2,2 w .

w , R |ξ ) have been numerically ob-

Using (3.4), the values of π3.5,0.8 (R2,2 2,2 p
tained and presented in Table 6 for some choices of p. It is observed that for
upper quantiles (p ≥ 0.75), R2,2
w is Pitman closer to ξ than R .
p 2,2

w
Table 6. Values of π3.5,0.8 (R2,2 , R2,2 |ξp ) for standard exponential distribution.
p 0.10 0.25 0.5 0.60 0.75 0.80 0.90 0.95 0.99
w
π3.5,0.8 (R2,2 , R2,2 |ξp ) 0.2378 0.2389 0.2935 0.3727 0.5616 0.6341 0.7366 0.7581 0.7621

Example 2 (Air conditioning system). In this example we use the data set
which consist of the intervals between failures (in hours) of the air conditioning
system in three Boeing 720 jet aircrafts. The data are reported in Table 7.
See Proschan (1963) for a detailed description of the data set. He tested and
accepted the hypothesis that the successive failure times were iid exponential for
each plane, but with different failure rates.
PMC for weighted random variables 15

Table 7. Intervals between failures of the air conditioning system in three Boeing 720
jet aircraft.

Plane 7909 90,10,60,186,61,49,14,24,56,20,79,84,44,59,29,118,25,156,310,76,26,44,23,62,130,208,70,101,208

Plane 7912 23,261,87,7,120,14,62,47,225,71,246,21,42,20,5,12,120,11,3,14,71,11,14,11,16,90,1,16,52,95
Plane 7913 97,51,11,4,141,18,142,68,77,80,1,16,106,206,82,54,31,216,46,111,39,63,18,191,18,163,24

Let us denote the intervals between failures of the air conditioning system
of Planes 7913, 7912 and 7909 by X, Xw1 and Xw2 , respectively. As well as
assumed by Proschan (1963), the observed values of X come from exponential
distribution. Moreover, it can be deduced that Xw1 and Xw2 are two different
weighted versions of X. More precisely, according to the general weighted pdf
presented in (3.1) with baseline exponential distribution, the null hypotheses that
the associated data for Planes 7912 and 7909 are coming from

gF (x; 1, 1.75, 1, 1) = 1.75[F̄ (x)]0.75 f (x)

and
1
gF (x; 1, 1, 1, 1.15) = [− log F̄ (x)]0.15 f (x),
Γ(1.15)
respectively, are checked by using the K-S distances. The observed p-values are
0.6455 and 0.9123, respectively. So, based on these observations, the mentioned
weighted versions of the exponential distribution are adequate for the data. By
using (4.2) and doing some algebraic calculations and numerical computations,
we get π(Xw1 , Xw2 |X) = 0.0452. That is, the intervals between failures of the air
conditioning system of Plane 7909 are Pitman closer than those of Plane 7912
with respect to Plane 7913.

6. CONCLUSION

In this paper, the weighted random variables were considered and their
closeness to a common parameter was investigated in the sense of PMC. Some
general results were derived and a new general weighted model was introduced
which subsumes most of the previous works as special cases. Some numerical
results and conclusions were presented for exponential distribution in details. It
was seen that the weighted upper k-records are Pitman closer than the usual
ones to certain population quantiles. To illustrate the proposed procedure, a real
data set was used. Furthermore, the concept of PMC was applied for measuring
the nearness of some weighted random variables with respect to each other. This
procedure was also explained via application to a real data set.
16 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

ACKNOWLEDGMENTS

The authors would to thank the Associate Editor and the reviewer for the
useful comments and suggestions.

REFERENCES

[1] Ahmadi, J. and Balakrishnan, N. (2009). Pitman closeness of record values to

population quantiles, Statistics and Probability Letters, 79, 2037–2044.
[2] Ahmadi, J. and Balakrishnan, N. (2010a). On Pitman’s measure of closeness of
k-records, Journal of Statistical Computation and Simulation, 81, 497–509.
[3] Ahmadi, J. and Balakrishnan, N. (2010b). Pitman closeness of current records for
location-scale families, Statistics and Probability Letters, 80, 1577–1583.
[4] Ahmadi, J. and Raqab, M. Z. (2013). Comparison of order statistics in two-sample
problem in the sense of Pitman closeness, Statistics, 47, 729–743.
[5] Ahmadi, Mosayeb and Mohtashami Borzadaran, G. R. (2014). On the nearness
of records to population quantiles with respect to order statistics in the sense of
Pitman closeness, Communications in Statistics–Theory and Methods, 43, 4357-
4370.
[6] Arnold, B. C., Balakrishnan, N. and Nagaraja, H. N. (2008). A First Course in
Order Statistics, SIAM.
[7] Arnold, B. C., Balakrishnan, N. and Nagaraja, H. N. (1998). Records, John Wiley
& Sons, New York.
[8] Azzalini, A. (2022). An overview on the progeny of the skew-normal family-
A personal perspective, Journal of Multivariate Analysis, 188, 104851,
https://doi.org/10.1016/j.jmva.2021.104851.
[9] Balakrishnan, N., Davies, K. and Keating, J. P. (2009a). Pitman closeness of order
statistics to population quantiles, Communications in Statistics – Simulation and
Computation, 38, 802–820.
[10] Balakrishnan, N., Davies, K., Keating, J. P. and Mason, R. L. (2010) Simultane-
ous closeness among order statistics to population quantiles, Journal of Statistical
Planning and Inference, 140, 2408–2415.
[11] Balakrishnan, N., Iliopoulos, G., Keating, J. P. and Mason, R. L. (2009b). Pit-
man closeness of sample median to population median, Statistics and Probability
Letters, 79, 1759–1766.
[12] Bartoszewicz, J. (2009). On a representation of weighted distributions. Statistics
and Probability Letters, 79, 1690–1694.
[13] Castillo, F., Hadi, A. S., Balakrishnan, N. and Sarabia, J. M. (2005). Extreme
Value and Related Models with Applications in Engineering and Science, John
Wiley & Sons, Inc.
PMC for weighted random variables 17

[14] David, H. A. and Nagaraja, H. N. (2003). Order Statistics, Third edition, John
Wiley & Sons, Hoboken, New Jersey.
[15] Davies K. F. (2021). Pitman closeness results for Type-I hybrid censored data
from exponential distribution, Journal of Statistical Computation and Simulation,
91, 58-80.
[16] Efron, B. (1975). Biased versus unbiased estimation. Advances in Mathematics,
16, 259–277.
[17] Fisher, R. A. (1934). The effect of methods of ascertainment upon the estimation
of frequencies. Annals of Eugenics, 6, 13–25.
[18] Ghosh, M. and Sen, P. K. (1989). Median unbiasedness and Pitman closeness.
Journal of the American Statistical Association, 84, 1089-1091.
[19] Gómez-Déniz, E., Arnold, B. C., Sarabia, J. M. and Gómez, H. W. (2021). Prop-
erties and applications of a new family of skew distributions. Mathematics, 9, P.
87.
[20] Gupta, R. C. and Keating, J. P. (1986). Relations for the reliability measures
under length biased sampling. Scandinavian Journal of Statistics, 13, 49-56.
[21] Jiang, Y. (2020). Study on weight function distribution of hybrid
gas-liquid two-phase flow electromagnetic flowmeter. Sensors, 20, 1431,
https://doi.org/10.3390/s20051431.
[22] Jones, M. C. (2004). Families of distributions arising from distributions of order
statistics (with discussion). Test, 13, 1–43.
[23] Keating, J.P. (1985). More on Rao’s phenomenon, Sankhya, Series A, 47, 18–21.
[24] Keating, J. P. and Mason, R. L. (1985). Practical relevance of an alternative
criterion in estimation, American Statistician, 39, 203–205.
[25] Keating, J. P., Mason, R. L. and Sen, P. K. (1993). Pitman’s Measure of Close-
ness: A Comparison of Statistical Estimators, Society for Industrial and Applied
Mathematics, Philadelphia.
[26] Kourouklis, S. (1996). Improved estimation under Pitman’s measure of closeness,
Annuals of Statistics 48, 509–518.
[27] Lee, C. M. S. (1990). On the characterization of Pitman’s measure of nearness.
Statistics and Probability Letters, 8 , 41–46.
[28] Mendes, B. and Merkle, M. (2005). Some remarks regarding Pitman closeness,
Univ. Beograd Publ. Elektrotehn. Fak. Ser. Mat., 16, 1–11.
[29] Morabbi, H. and Razmkhah, M. (2020). Quantile estimation based on modified
ranked set sampling schemes using Pitman closeness, Communications in Statistics
- Simulation and Computation, DOI:10.1080/03610918.2020.1811329.
[30] Nayak, T. K. (1990). Estimation of location and scale parameters using general-
ized Pitman nearness criterion, Journal of Statistical Planning and Inference, 24,
259–268.
[31] Patil, G. P. and Rao, C. R. (1977). Weighted distributions: a survey of their
application. In P. R. Krishnaiah (Ed.), Applications of Statistics, North Holland
Publishing Company, 383–405.
[32] Patil, G. P. and Rao, C. R. (1978). Weighted distributions and size-biased sam-
pling with applications to wildlife populations and human families. Biometrics, 34,
179–180.
18 Mosayeb Ahmadi, G. R. Mohtashami Borzadaran and Mostafa Razmkhah

[33] Pitman, E. J. G. (1937). The closest estimates of statistical parameters, Proceed-

ings of The Cambridge Philosophical Society, 33, 212–222.
[34] Proschan, F. (1963). Theoretical explanation of observed decreasing failure rate.
Technometrics, 5, 375–383.
[35] Rao, C. R. (1965). Weighted distributions arising out of methods of ascertain-
ment. In classical and contagious discrete distributions, G. P. Patil (Eds), Calcuta:
Pergamon Press and Statistical Publishing Society, 320-332.
[36] Rao, C. R. (1980). Discussion of Minimum chi-square, not maximum likelihood
by J. Berkson, Annals of Statistics, 8, 482–485.
[37] Rao, C. R. (1981). Some comments on the minimum mean square error as a
criterion of estimation. In statistics and related topics, eds. M. Csorgo, D. Dawson,
J. Rao, and A. Saleh, North-Holland, Statistics and Related Topics, Amsterdam:
North Holland, 123-143.
[38] Rao, C.R. (1985). Weighted distributions arising out of methods of ascertainment,
in A Celebration of Statistics, A.C. Atkinson and S.E. Fienberg, eds, Springer-
Verlag, New York, Chapter 24, pp. 543-569.
[39] Raqab, M. Z. and Ahmadi, J. (2012). Pitman closeness of record values from two
sequences to population quantiles, Journal of Statistical Planning and Inference,
142, 855–862.
[40] Razmkhah, M. and Ahmadi, J. (2011). Nonparametric prediction intervals for fu-
ture order statistics in a proportional hazard model, Communications in Statistics–
Theory and Methods, 40, 1807-1820.
[41] Razmkhah, M., Ahmadi, J., Khatib, B. (2008). Nonparametric confidence in-
tervals and tolerance limits based on minima and maxima, Communications in
Statistics–Theory and Methods, 37, 1523-1542.
[42] Razmkhah, M. and Ahmadi, J. (2013). Pitman closeness of current k-records to
population quantiles, Statistics & Probability Letters, 83, 148–156.
[43] Robert, C. P., Hwang, J. T. G. and Strawderman, W. E. (1993). Is Pitman
closeness a reasonable criterion?, Journal of the American Statistical Association,
88, 52–76.
[44] Saghir, A., Hamedani, G. G., Tazeem, S. and Khadim, A. (2017). Weighted dis-
tributions: a brief review, perspective and characterizations, International Journal
of Statistics and Probability, 6, 109-131.
[45] Sen, P. K. (1986). Are BAN estimators the Pitman closest ones too?, Sankhya
Series A, 48, 51–58.
[46] Sen, P. K., Kubokawa, T. and Saleh, A. K. M. E. (1989). The Stein paradox in
the sense of the Pitman measure of closeness, Annals of Statistics, 17, 1375–1386.
[47] Volterman, William, Davies, Katherine F. and Balakrishnan, N. (2013). Simulta-
neous Pitman closeness of progressively type-II right censored order statistics to
population quantiles, Statistics, 47, 439–452.