Statlstms & Probabdlty Letters 7 (1989) 179-185
North-Holland
December 1988
TESTING FOR DISPERSIVE ORDERING
Ibrahim A. AHMAD and Subhash C. KOCHAR *
Dwtswn of Stattsttcs, Nothern Ilhnols Unwerslly, DeKalb, IL 60115, USA
Recewed December 1987
Revased April 1988
dlsp
Abstract An asymptotmally dxstnbutmn-free test is proposed and studied for testing the null hypothesis H 0 F = G versus
dlsp
H 1 F < G, that is G-l(fl)-G-l(a)>/F l(fl)_ F-l(a) for 0 ~< a < fl ~<1 The proposed test is based on an estimator of
]fZ(x) dz - fg2(x) dx, where f(g) is the probability density function corresponding to the distribution function F(G) The
cases of two independent samples as well as that of pmred samples are discussed in detads
A M S 1980 SubJect Classtficatton: Primary 62G05, 62E20, Secondary. 62G10
Keywords nonparametnc, asymptoUcally distribution-free, density estimate, kernel, window size
1. Introduction
Let X and Y be two random variables having
absolutely continuous distribution functions
(d.f.'s) F and G, respectively, with F -a and G -1
as their left continuous inverses.
Definition 1.1. G is said to be more dispersed than
dlsp
F, written F < G, if
G-l(fl) - G-l(a) >_. F-l(fl) - F-l(a)
forO~<a<fl~<l, (1.1)
i.e., if any two quantiles of G are at least as far
apart as the corresponding quantiles of F.
The above definition is equix;alent to saying
that G-1F(x)-x is nondecreasing in x. Doksum
(1969) calls this ordering as tail ordering.
Let f and g denote the probability density
functions of F and G, respectively. The failure
(hazard) rate of F is defined as
rF(X )=f(x)/[1-F(x)], F(x)<I.
* On leave from Panjab Umverslty, Chandlgarh, INDIA
0167-7152/88/$3 50 © 1988, Elsevier Science Pubhshers B V (North-Holland)
Differentiating G-1F(x)- x, we find that (1.1) is
equivalent to
g[G-l(u)]<~f[F-a(u)] for 0~< u~< 1. (1.2)
Equivalently,
ro[G-a(u)] <~rF[F-t(u)] for0~<u~< 1. (1.3)
The above several equivalent versions of disper-
sive ordering have been discussed by many authors
including Doksum (1969), Bickel and Lehmann
(1979), Oja (1981), Lewis and Thompson (1981),
and Shaked (1982). For positive random variables,
which are mainly applicable in reliability theory,
Deshpande and Kochar (1983), Bartoszewicz
(1986), and Ahmad, Alzaid, Bartoszewicz, and
Kochar (1986), have discussed relations between
dispersive ordering, star ordering, superadditive
ordering, and failure rate ordering.
Note that if F(x)= H((x- 01)/,11) and G(x)
dlsp
=H((x-02)/'02), then F < G if and only if
1"/1 < "02.
It is possible sometimes to compare two distri-
butions F and G not necessarily belonging to the
same location-scale family. Doksum (1969) has
shown that for distributions symmetric about 0, if
F is ordered with respect to G in the sehse of
179