Beruflich Dokumente
Kultur Dokumente
mer O
ztu rk
The Ohio State University, 1465 Mt. Vernon Avenue,
Marion, OH 43302
E-mail: omer@stat.ohio-state.edu
ABSTRACT
This paper introduces a nonparametric test of symmetry for
ranked-set samples to test the asymmetry of the underlying
distribution. The test statistic is constructed from the Crame r-
von Mises distance function which measures the distance
between two probability models. The null distribution of the
test statistic is established by constructing symmetric boot-
strap samples from a given ranked-set sample. It is shown
that the type I error probabilities are stable across all practical
symmetric distributions and the test has high power for asym-
metric distributions.
Key Words: Bootstrap distribution; Cramer-von Mises dis-
tance; Design; HodgesLehmann estimator; Median; Mode;
Unequal allocation
Mathematics Subject Classication Codes (1991): 62G30;
62H10
2117
Copyright & 2001 by Marcel Dekker, Inc. www.dekker.com
COMMUN. STATIST.THEORY METH., 30(10), 21172133 (2001)
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
1. INTRODUCTION
Appropriate location and scale measures for a distribution depend
on its general structure. Either the mean or median can be used to dene
the location of an underlying symmetric distribution, but it is not clear
how to dene the appropriate location measure when the underlying
distribution is asymmetric. Similarly, the standard deviation makes sense
as a measure of spread (or scale) if the underlying distribution is symmetric,
but in an asymmetric setting, the researcher has to reect more intensively
about what to use as a measure of spread for the underlying distribution.
Therefore, to select an appropriate statistical analysis, it is necessary to
check if the data set is coming from a symmetric distribution. There are
several available testing procedures to check the symmetry of the underlying
distribution for a simple random sample (Antille, Kersting and Zucchini (1),
Boos (2), Koziol (3), Schuster and Barker (4), Randles, Fligner, Policello
and Wolfe (5)). The nite sample null distributions of all these testing
procedures, with the exception of Schuster and Barker (4) test, depend on
the underlying probability models. Schuster and Barker use the bootstrap
distribution to get around this problem.
In many scientic investigations, where the actual measurement of
the observations is dicult relative to a complete or partial ranking of all
units in a set, ranked-set sampling provides improved statistical inference.
The data collection procedure in a ranked-set sampling involves selecting m
2
units at random from an innite population F. These units are divided
into m sets, each having m units. The units are then judgment ranked by
some criterion that does not require actual measurements. The ith judgment
order statistic, X
i1
, from the ith set is quantied for i 1, . . . , m.
The observations, X
i1
, i 1, . . . , m, constitute a cycle and the process can
be repeated for n cycles to have n m quantied observations, X
i j
,
i 1, . . . , m; j 1, . . . , n. If the judgment ranking is perfect, the ith
judgment order statistic is the same as the ith order statistic for each of the
n cycles. In this case we use the usual order statistic notation, X
i j
, to denote
the ith order statistic from the jth cycle, i 1, . . . , m, j 1, . . . , n. Unless
stated otherwise, we assume perfect judgment ranking throughout the paper.
Note that the ranked-set sample described above is balanced, namely
each of the m order statistics is quantied an equal number of times. Recent
research has shown that unequal allocation in a ranked-set sample improves
the eciency of a balanced ranked-set sample (O
zt uurk (6); O
zt uurk
and Wolfe (7,8), Chen (9,10), Kaur, Patil and Taillie (11,12)). In the
same spirit but in a dierent context, Kvam and Samaniego (13) and
Stokes (14) used unequal weighting in L-estimators. Thus, we consider
unequal allocation to cover several optimal allocation procedures.
2118 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
For a xed t 2 f1, . . . , mg, let D fd
1
, . . . , d
t
g be a set of integers
that determines the judgment ranks of the quantied observations. For
example, if D f1, 5g and the set size m is 5, we quantify two extremes
in each cycle. This denition is general enough to cover all practical alloca-
tion procedures. For example, O
zt uurk and
Wolfe (7,8). In a recent work, O
t
i1
n
j1
oX
d
i
j
y
be the Stokes and Sager (18) cdf estimator of F, where oa 1, 0 as
a , > 0.
Denition 1. The design D is called symmetric if d
i
d
m1i
m1, for
i 1, . . . , t.
We note that F
D, n
y is not necessarily an unbiased estimator for F
if the design D does not correspond to the standard ranked-set sample
design D SRSS f1, . . . , mg. On the other hand, it is an unbiased esti-
mator for
t
i1
F
d
i
sr
y,t, which is a symmetric distribution if the under-
lying probability model F and the design D are symmetric. Thus, even
though F
D, n
t is a biased estimator for Ft, it is still possible to use
F
D, n
t to devise a test of symmetry as long as D is a symmetric design.
Therefore, we restrict our attention to the symmetric designs D in the
remainder of this paper. Let
W
_
1
1
fF
D, n
y F
D, n
2 y 1g
2
dy.
A functional form of W can be written as
W
_
1
1
1
t
t
j1
F
d
j
y F
m1d
j
2 y 1
_ _
2
dy
2120 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
where F
d
i
y
F
m1d
j
t
i1
n
j1
t
k1
n
r1
jX
d
i
j
X
d
k
r
j
1
t
2
n
2
t
i1
n
j1
t
k1
n
r1
jX
d
i
j
X
d
k
r
2j. 1
Note that if the underlying distribution F is symmetrical about and the
design D is symmetric, then
t
i1
t
k1
jX
d
i
j
X
d
k
r
2j
d
t
i1
t
k1
jX
d
i
j
X
d
k
r
j,
where
d
stands for the equality in distribution. We propose to use
Boos (2) test statistic to asses the symmetry of the underlying distribution
F, namely,
T
n
^
n 1
t
i1
n
j1
t
k1
n
r1
jX
d
i
j
X
d
k
r
2
^
j
t
i1
n
j1
t
k1
n
r1
jX
d
i
j
X
d
k
r
j
_ _
, 2
where
^
is a suitable estimator for the point of symmetry. We note that
the denumerator of the ratio in Eq. (2) is the nt 1,nt times Ginis
mean dierence which estimates the dispersion of the underlying distribu-
tion. Thus, W
^
is standardized with respect to the dispersion of F.
If
^
is a consistent estimator for the point of symmetry , under the
null hypothesis the second term in the curly brackets in Eq. (2) is,
in distribution, equal to one. Thus, we expect that T
n
^
will be close to
zero under the symmetry assumption and we reject the null hypothesis
for large values of T
n
^
. The problem with this approach is that the null
distribution of the test statistic T
n
^
, as in the simple random sampling
Boos (2), depends on the underlying probability model. Boos has shown
that the null limiting distribution of T
n
^
in a simple random sample is
an innite sum of the weighted chi-square random variables. On the other
hand, the null distribution of T
n
^
appears to be stable across many
NONPARAMETRIC TEST OF SYMMETRY VERSUS ASYMMETRY 2121
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
practical unimodal symmetric distributions if it is evaluated at the logistic
model. The limiting null distribution of T
n
^
in a ranked-set sample will
also be a function of an innite sum of independent chi-square random
variables. On the other hand, it will be much more complicated than
its simple random sample counterparts since it involves order statistics.
Thus, it will not have much use in practical applications. In order to est-
ablish the null distribution of T
n
^
, in Section 3, we study the bootstrap null
distribution.
3. BOOTSTRAP DISTRIBUTION OF T
n
^
hh
A close investigation of Eq. (1) shows that W is a convex function
of . Thus, there exists a unique minimizer. Let
^
HL
arg min
2R
W.
O
HL
is the median of all Walsh averages X
d
i
j
X
d
k
r
,2, i, k 1, . . . , t; j, r 1, . . . , n. For the standard ranked-set sample
D SRSS the asymptotic variance of
n
p
^
HL
is
o
2
1
6mm1f
_
f
2
y dyg
2
.
For a general design D, the variance of
n
p
^
HL
is given in O
zt uurk (7).
We construct our test statistic by replacing
^
with
^
HL
in Eq. (2) and
reject the null hypothesis H
0
in favor of H
1
for large values of T
n
T
n
HL
.
Our test rejects the null hypothesis of symmetry if and only if
T
n
HL
T
n
F
D, n
, X
d
1
1
, . . . , X
d
t
n
,
^
HL
! t
,
where
P
F
fT
n
HL
! t
jH
0
is trueg ot
, F
is the probability of type I error for the proposed test at t
and F.
Apparently, the type I error of the test is a function of the unknown under-
lying distribution F. Thus, the test is not distribution free. We approximate
ot
HL
.
2122 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
The bootstrap is a nonparametric estimation procedure for the distri-
bution of some functional, say M MF, X
1
, . . . , X
n
. An algorithm for the
bootstrap distribution of M is given as follows:
For b 1 to B
I Sample X
b
1
, . . . , X
b
n
with replacement from the empirical CDF F
n
based on the original sample X
1
, . . . , X
n
.
II Compute M
b
MF
n
, X
b
1
, . . . , X
b
n
.
Then the bootstrap estimate of the distribution of M is the empirical dis-
tribution of M
i
, i 1, . . . , B.
It is clear that the bootstrap distribution of T
n
HL
must be esta-
blished by utilizing the symmetry restriction. For a nite sample size n,
F
D, n
is not symmetric even if the ranked-set sample, X
d
i
j
, 1, . . . , t;
j 1, . . . , n, is generated from a symmetric distribution F. Thus, we
cannot sample our bootstrap sample from F
D, n
since the bootstrap estimate
ot
, F
D, n
will not estimate the type I error probabilities correctly. In order
to get a good approximation for ot
, F, F
D, n
must be replaced with a
suitable symmetric distribution. In the following theorem, we show that a
symmetric distribution exists for any symmetric design D and sample size n.
Theorem 1. Let X
d
i
j
, i 1, . . . , t; j 1, . . . , n, be a ranked-set sample
from a symmetric distribution F 2 O. Then G
D, n
y,
^
HL
fF
D, n
y
F
D, n
2
^
HL
y 1g,2 is the closest symmetric distribution in O to
F
D, n
y, where the closeness is measured with respect to the Cramer-von
Mises distance function; i.e,
W
^
HL
_
1
1
fF
D, n
y G
D, n
y,
^
HL
g
2
dy
min
_
1
1
F
D, n
y G y
2
dy; G 2 O
_ _
.
Theorem 1 allows us to select a bootstrap sample from a symmetric
distribution which is the closest distribution to F
D, n
in Crame r-von Mises
distance. In order to construct a bootstrap sample under the null hypothesis,
we rst estimate with
^
HL
. Then, we sample nt observations with replace-
ment from 2tn observations, X
d
i
j
^
HL
, j 1, . . . , n; i 1, . . . , t. This
bootstrap sample will be a realization from the distribution G
D, n
y,
^
HL
b
HL
be the value of
T
n
HL
obtained from the bth symmetric bootstrap sample. Then the
empirical cdf of T
b
n
, b 1, . . . , B, estimates the distribution of T
n
HL
.
Let
TT
n
T
n
F
D, n
, X
d
1
1
, . . . , X
d
t
n
,
^
HL
be the value of T
n
HL
obtained
from the original sample. Then,
o
B
TT
n
B
b1
oT
b
n
T T
n
,B
is the p-value of the test, which can be used to reject the null hypothesis
for its small values.
As suggested in Schuster and Barker (4), for bootstrap samples we
recommend using a nearby continuous symmetric distribution instead of a
discrete G
D, n
. Let
G
D, n
y
1
nt
nt
i1
K
y Y
i
a
_ _
,
where Y
1
, . . . , Y
nt
is a symmetric bootstrap sample from G
D, n
, K is the
cdf of the uniform distribution over 1, 1, and positive constant a is
the smoothing parameter. A good choice of a can be considered from
the optimal smoothing parameter which makes the derivative of G
D, n
y
a good estimate of the derivative of F
D
y
t
i1
F
d
i
y,t. From
Lemma 4A of Parzen (20), the optimal smoothing parameter of this density
estimation is
a
opt
j
2,5
2
f
_
k
2
y dyg
1,5
f
_
f
002
D
y dyg
1,5
n
1,5
, 3
where k y is the kernel density, j
2
is the second moment of the
kernel density, and f
00
D
y is the second derivative of
t
i1
f
d
i
y,t.
Unfortunately, a
opt
is not distribution free since it is a function of the under-
lying distribution F. For a simple random sample, Boos (2) has shown that
the null distribution of T
n
HL
does not vary much across all practical
unimodal distributions if it is evaluated at the logistic distribution. Thus,
we expect and show in a simulation study that the smoothing parameter
evaluated at the logistic distribution will perform suciently well across all
practical distributions. Let H y be the logistic distribution with mean zero
and scale 1.
2124 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
Lemma 1. Let F y H y ,o be the logistic distribution with mean
and scale o. For a given design D, uniform kernel density over 1, 1 and a
perfect ranking, the optimal smoothing parameter for f
D
y is
a
opt
1.35o
3,5
c
1,5
D
n
1,5
,
where
c
D
_
1
t
t
i1
m
m1
d
i
1
_ _
u
d
i
1
1 u
md
i
A
d
i
u
_ _
2
u1 u du,
A
d
i
u d
2
i
m12d
i
1u m2m1u
2
.
For an easy reference, Table 1 presents the values of c
D
for selected practical
designs and set sizes.
In order to estimate o, we recommend using interquartile range (IQR)
since it will be less aected by the extreme observations. For the logistic
distribution, the estimate of o can be expressed as
S
IQR
t
D
,
where t
D
is given in Table 1 and suggested for the unbiasedness of the
estimator at the logistic distribution and ranked-set sample design D.
We note that one can use some other symmetric kernel densities in
G
D
,
n
y such as logistic or normal. On the other hand, the selection of the
uniform kernel density over 1, 1 makes the sampling easy since G
D
,
n
y
can be expressed as the convolution of G
D, n
y with the cdf of the uniform
Table 1. The Coecient c
D
and t
D
for Selected Set Sizes and Designs at the
Logistic Distribution Under the Perfect Judgment Ranking
m D t
D
c
D
m D t
D
c
D
2 {1, 2} 2.197 0.024 4 {1, 2, 3, 4} 2.197 0.024
3 {1, 3} 2.748 0.009 5 {1, 5} 3.812 0.065
3 {2} 1.449 0.166 5 {2, 4} 1.705 0.069
3 {1, 2, 3} 2.197 0.024 5 {2, 3, 4} 1.489 0.143
4 {1, 4} 3.334 0.033 5 {3} 1.156 0.494
4 {2, 3} 1.449 0.166 5 {1, 2, 3, 4, 5} 2.197 0.024
NONPARAMETRIC TEST OF SYMMETRY VERSUS ASYMMETRY 2125
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
distribution over a, a. Thus, we can sample a symmetric bootstrap sample
from G
D, n
y by observing that Z Y aU is a realization from G
D, n
y,
where Y and U are independent random variables, Y has cdf G
D, n
y, and U
has uniform cdf over 1, 1.
4. SIMULATION RESULTS
We must address two basic questions about the proposed testing
procedure: (1) How good is the type I probability approximation across
all practical distributions? and (2) How powerful is the test to detect
asymmetry?
Table 2 contains the type I error probability estimates of the proposed
test. The simulation study is done with IMSL random generator for several
symmetric distributions. The distributions include normal, double exponen-
tial, Cauchy, logistic, uniform, beta(0.9, 0.9) and t
i
, i 2, 3, where t
i
is the
Students t-distribution with the degrees of freedom i. We considered sample
sizes n 10, 20, 30 and several symmetric designs. In the estimation of the
type I error probabilities, we generated 100 dierent samples for each
sample size, design and distribution combinations. For each one of these
100 samples, we generated 5000 symmetric bootstrap samples from G
D, n
y
and calculated the test statistic. Based on these 5000 T
b
n
, b 1, . . . , B, we
estimated the critical value for a ve-percent test, i.e. t
i
: f
B
b1
oT
b
n
! t
i
,
Bg 0.05, i 1, . . . , 100. The bootstrap estimate of the critical value of a
ve-percent test is taken as the average of these 100 t
i
, i 1, . . . , 100,
tt
100
i1
t
i
,100.
In order to see how good
tt
tt
10000
i1
of
TT
i
n
HL
tt
g
10 000
,
where
TT
i
n
HL
is the test statistic evaluated at an independent sample.
Note that if
tt
tt
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
Table 2. Estimates of the Type I Error Probabilities for Several Symmetric
Distributions: Logistic(L), Normal(N), Double Exponential (DE), Cauchy(C),
Uniform(U), Students T-distribution with Degrees of Freedom i (t
i
), Beta
Distributions with Parameters 0.9 and 0.9 (B)
m D n L N DE C U t
2
t
3
B
2 {1, 2} 10 0.053 0.045 0.069 0.040 0.076 0.055 0.056 0.078
20 0.049 0.050 0.057 0.039 0.079 0.055 0.062 0.100
30 0.051 0.048 0.063 0.038 0.089 0.058 0.048 0.101
3 {1, 3} 10 0.036 0.031 0.052 0.041 0.048 0.059 0.049 0.055
20 0.040 0.031 0.059 0.041 0.052 0.058 0.051 0.058
30 0.038 0.032 0.057 0.033 0.061 0.052 0.057 0.068
3 {2} 10 0.045 0.050 0.063 0.069 0.052 0.052 0.055 0.058
20 0.048 0.049 0.067 0.071 0.059 0.054 0.051 0.057
30 0.051 0.049 0.054 0.047 0.064 0.049 0.055 0.065
3 {1, 2, 3} 10 0.052 0.047 0.063 0.037 0.063 0.065 0.058 0.071
20 0.052 0.051 0.057 0.033 0.071 0.061 0.056 0.080
30 0.052 0.049 0.064 0.034 0.073 0.059 0.057 0.083
4 {1, 4} 10 0.015 0.008 0.028 0.031 0.021 0.043 0.030 0.025
20 0.016 0.008 0.032 0.031 0.022 0.047 0.035 0.029
30 0.016 0.007 0.035 0.032 0.026 0.047 0.030 0.026
4 {2, 3} 10 0.048 0.048 0.057 0.061 0.054 0.056 0.054 0.059
20 0.052 0.051 0.060 0.052 0.055 0.058 0.055 0.057
30 0.048 0.049 0.057 0.050 0.060 0.054 0.052 0.065
4 {1, 2, 3, 4} 10 0.049 0.044 0.066 0.038 0.057 0.053 0.057 0.059
20 0.051 0.047 0.060 0.029 0.064 0.054 0.059 0.070
30 0.050 0.042 0.059 0.033 0.062 0.052 0.051 0.074
5 {1, 5} 10 0.004 0.000 0.014 0.026 0.007 0.030 0.019 0.009
20 0.003 0.000 0.014 0.030 0.008 0.036 0.023 0.008
30 0.003 0.000 0.016 0.028 0.007 0.034 0.022 0.007
5 {2, 4} 10 0.042 0.043 0.055 0.056 0.046 0.054 0.051 0.048
20 0.044 0.043 0.051 0.056 0.048 0.050 0.044 0.053
30 0.049 0.043 0.055 0.050 0.057 0.054 0.049 0.057
5 {3} 10 0.051 0.050 0.060 0.053 0.046 0.055 0.054 0.049
20 0.050 0.047 0.059 0.053 0.051 0.049 0.049 0.052
30 0.046 0.051 0.056 0.060 0.058 0.051 0.051 0.056
5 {2, 3, 4} 10 0.047 0.047 0.056 0.059 0.051 0.053 0.057 0.054
20 0.047 0.047 0.061 0.059 0.055 0.057 0.052 0.060
30 0.052 0.047 0.055 0.044 0.056 0.055 0.053 0.065
5 {1, 2, 3, 4, 5} 10 0.044 0.041 0.055 0.027 0.047 0.050 0.054 0.054
20 0.042 0.040 0.050 0.025 0.048 0.046 0.051 0.057
30 0.044 0.040 0.048 0.026 0.051 0.048 0.049 0.055
NONPARAMETRIC TEST OF SYMMETRY VERSUS ASYMMETRY 2127
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
Table 3. Empirical Power of the Bootstrap Test of Symmetry for Several
Asymmetric Distributions: Chi Square Distribution with Degrees of Freedom i
(Chi-i), Lognormal Distribution (LN), Exponential Distribution (E), Weibull
Distribution with Shape Parameter 0.5 and Scale Parameter 3 (W(0.5, 3))
m D n Chi-1 Chi-2 LN E W(0.5, 3)
2 {1, 2} 10 0.891 0.646 0.750 0.653 0.871
20 0.998 0.944 0.975 0.950 0.998
30 1.000 0.994 0.999 0.995 1.000
3 {1, 3} 10 0.954 0.754 0.813 0.776 0.941
20 1.000 0.984 0.995 0.985 1.000
30 1.000 1.000 1.000 1.000 1.000
3 {2} 10 0.333 0.167 0.214 0.169 0.346
20 0.702 0.360 0.493 0.389 0.745
30 0.887 0.537 0.690 0.580 0.951
3 {1, 2, 3} 10 0.987 0.857 0.913 0.870 0.984
20 1.000 0.996 0.999 0.997 1.000
30 1.000 1.000 1.000 1.000 1.000
4 {1, 4} 10 0.964 0.745 0.857 0.797 0.945
20 1.000 0.993 0.998 0.996 1.000
30 1.000 1.000 1.000 1.000 1.000
4 {2, 3} 10 0.727 0.367 0.485 0.387 0.790
20 0.971 0.725 0.847 0.744 0.993
30 0.999 0.907 0.969 0.913 1.000
4 {1, 2, 3, 4} 10 0.999 0.963 0.982 0.962 0.999
20 1.000 1.000 1.000 1.000 1.000
30 1.000 1.000 1.000 1.000 1.000
5 {1, 5} 10 0.957 0.711 0.859 0.785 0.945
20 1.000 0.989 0.999 0.996 1.000
30 1.000 1.000 1.000 1.000 1.000
5 {2, 4} 10 0.824 0.464 0.595 0.481 0.877
20 0.993 0.828 0.920 0.839 0.999
30 1.000 0.957 0.987 0.962 1.000
5 {3} 10 0.262 0.125 0.156 0.129 0.269
20 0.578 0.270 0.367 0.274 0.702
30 0.802 0.417 0.548 0.427 0.894
5 {2, 3, 4} 10 0.906 0.599 0.741 0.614 0.960
20 0.999 0.914 0.975 0.919 1.000
30 1.000 0.988 0.998 0.990 1.000
5 {1, 2, 3, 4, 5} 10 1.000 0.991 0.997 0.991 1.000
20 1.000 1.000 1.000 1.000 1.000
30 1.000 1.000 1.000 1.000 1.000
2128 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
sample underestimate the size of the test. For example, design D f1, 5g
with m 5 underestimates the type I error probabilities.
For asymmetric distributions, the critical value
tt
D, n
y. Thus, o
s
tt
tt
HL
has good power for standard ranked-set samples (D f1, . . . , mg)
and the designs that quantify extreme order statistics in each cycle. For
example, the designs D f1, 4g and D f1, 5g have higher powers than
the designs D f2, 3g and D f2, 4g when m 4 and m 5, respectively.
There is a loss of power for the central designs that quantify only the middle
order statistics as expected. This can be explained from the fact that asym-
metry for unimodal distribution can be expressed in the form of a right-
skewness or left-skewness. Thus, the central designs (such as D f3g with
m 5) ignore the largest order statistic which contributes to the skewness
(asymmetry) and result a loss in the power of the test.
5. CONCLUDING REMARKS
This paper establishes a test of symmetry for ranked-set samples to
assess the asymmetry of the underlying distribution. The null distribution of
the test is constructed by using the symmetric bootstrap samples. It appears
in Table 2 that the null distribution is stable across all practical distributions
and the sample sizes as small as n 10. A simulation study indicates that
the test has high power for asymmetric distributions.
APPENDIX
Proof Theorem 1: Let F
. For the
proof of the theorem it is sucient to show that
min
1--1
min
F
2O
HF
D, n
, F
W
^
HL
.
NONPARAMETRIC TEST OF SYMMETRY VERSUS ASYMMETRY 2129
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
We now consider
HF
D, n
, F
_
fF
D, n
y F
yg
2
dy
_
fF
D, n
y G
D, n
, yg
2
dy
2
_
fF
D, n
y G
D, n
, ygfG
D, n
, y F
yg dy
_
fG
D, n
, y F
yg
2
dy. 4
Note that both F
. and G
D, n
, . are in O. Let
U, y F
D, n
y G
D, n
, y
F
D, n
y F
D, n
2 y 1
2
.
It is clear that U, y is symmetric around and U, y U, 2 y. The
cross-product term in Eq. (4) can be expressed as
_
fF
D, n
y G
D, n
, ygfG
D, n
, y F
yg dy
_
1
1
U, yG
D, n
, y dy
_
1
1
U, yF
y dy.
We now show that this cross-product term is equal to zero. We start with the
rst integral
_
1
1
U, yG
D, n
, y dy
_
1
U, yG
D, n
, y dy
_
1
U, yG
D, n
, y dy
_
1
U, 2 zG
D, n
, 2 z dz
_
1
U, yG
D, n
, y dy
_
1
U, zf1 G
D, n
, zg dz
_
1
U, yG
D, n
, y dy
_
1
U, z dz. 5
2130 O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
Since F
2 O, it follows that
_
1
1
U, yF
y dy
_
1
U, y dy.
By combining Eqs (5) and (6), we observe that the cross-product term is
zero. Therefore, we have shown that
min
F
2O
HF
D, n
, F
W.
The remaining part of the proof of the theorem follows from Lemma 1 of
O
zt uurk (17).
Proof of Lemma 1: For the uniform kernel over 1, 1, j
2
1,3 and
_
k
2
ydy 1,2. Without loss of generality we assume 0. Then, we
need to evaluate the second derivative of
f
D
y
1
t
t
i1
h
d
i
y,o.
Using the fact that, for logistic distribution, f y F yf1 F yg, we can
express h
d
i
y,o as
h
d
i
y,o m
m1
d
i
1
_ _
H
d
i
1
y,of1 H y,og
md
i
h y,o
m
m1
d
i
1
_ _
H
d
i
y,of1 H y,og
md
i
1
.
After some algebraic manipulations, we have
h
00
d
i
y,o
m
o
2
m1
d
i
1
_ _
H
d
i
1
y,of1 H y,og
md
i
A
d
i
fH y,ogh y,o,
where
A
a
u a
2
m12a 1u m2m1u
2
.
NONPARAMETRIC TEST OF SYMMETRY VERSUS ASYMMETRY 2131
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
Note that
C
D
_
f
002
D
y dy
_ _
1
t
t
i1
h
00
d
i
y,o
_
2
dy
_ _
m
o
2
t
t
i1
m
o
2
m1
d
i
1
_ _
H
d
i
1
y,o.
f1 H y,og
md
i
A
d
i
fH y,ogh
1,2
y,o
_
2
h y,o dy.
Let H y,o u. Then we have
C
D
1
o
3
_ _
m
t
t
i1
m1
d
i
1
_ _
u
d
i
1
1 u
md
i
A
d
i
uu
1,2
1 u
1,2
_
2
du
c
D
o
3
,
which completes the proof.
ACKNOWLEDGMENT
The author thanks to the anonymous referee for his/her helpful com-
ments and suggestions.
REFERENCES
1. Antille, A.; Kersting, G.; Zucchini, W. Testing symmetry, J. Amer.
Statist. Assoc., 1982, 77, 639646.
2. Boos, D. A test of asymmetry associated with the HodgesLehmann
estimator, J. Amer. Statist. Assoc., 1982, 77, 647651.
3. Koziol, J.A. Test for symmetry about unknown value based on
the empirical distribution function, Commun. Statist., Theory and
Methods, 1983, 12(24), 28232844.
4. Schuster, F.; Barker, R.C. Using the bootstrap in testing symmetry
versus asymmetry, Commun. Statist., Simulation, 1987, 16(1), 6984.
5. Randles, R.H.; Fligner, M.A.; Policello, G.E.; Wolfe, D.A. An
asymptotically distribution-free test for symmetry versus asymmetry,
J. Amer. Statist. Assoc., 1980, 75, 168172.
6. O
ztu rk, O
ZTU
RK
D
o
w
n
l
o
a
d
e
d
b
y
[
P
a
l
e
s
t
i
n
i
a
n
P
o
l
y
t
e
c
h
n
i
c
U
n
i
v
e
r
s
i
t
y
]
a
t
0
5
:
2
5
2
2
S
e
p
t
e
m
b
e
r
2
0
1
2
ORDER REPRINTS
7. O
ztu rk, O
ztu rk, O
ztu rk, O