Biometrika (1971), 58, 1, p.

213 213
Printed in Great Britain

A combination of Wilcoxon's and Ansari-Bradley's statistics

B Y YVES LEPAGE
Universite de Montreal

SUMMARY
This paper deals with a nonparametric two-sample test for location and dispersion. I t
proposes a statistic T and derives some properties for T: a recursion formula for the exact

Downloaded from [Link] at Penn State University (Paterno Lib) on September 16, 2016
distribution, the asymptotic distribution and the consistency.

1. INTRODUCTION
Two samples, the X-sample and the F-sample of m and n independent observations from
populations with continuous distribution functions, F(x) and G(y) respectively, are con-
sidered. The problem is to test the hypothesis G(x) = F(x) versus alternatives of the form
G(x) = F{ax + b) with a * 1, or b 4= 0 (a > 0).
The proposed T statistic is a certain function of the Wilcoxon and Ansari-Bradley
statistics. The Wilcoxon statistic is well known and an important property is that the two-
sided test based on this statistic is consistent for testing the hypothesis F(x) = G(x) versus
alternatives of the form G(x) = F(x + b) (b #= 0); see Mann & Whitney (1947) and van Dant-
zig (1951). The Ansari-Bradley statistic is denned as follows: in the combined samples, the
observations less than or equal to the median are replaced by their ranks in increasing
order and those larger than the median are replaced by their ranks in decreasing order; the
statistic is the sum of these ranks for the X-sample. The two-sided test based on this statistic
is consistent for testing the hypothesis F(x) = G(x) versus alternatives of the form

G(x-d) = F{a(x-d)} (a>0,a=M),

where 6 is the common median; see Ansari & Bradley (1960).

In § 2 a description of the test is given and § 3 contains a recursion formula for the exact
distribution of T under the hypothesis. In § 4, by the use of the Chernoff-Savage asymptotic
normality theorem (Chernoff & Savage, 1958), the asymptotic distribution of T, under the
hypothesis, is found. Finally, in §5 the conditions for the consistency of the test based on T
are given.

2. DESCRIPTION OF THE TEST

Let Xlt..., Xm and Yv ..., Yn be two independent random samples from populations with
continuous distribution functions, F(x) and G(y) = F(ay + b)(a > 0), respectively. Let
Ti = 1 when the ith smallest of the N = m + n observations is an X and V{ = 0 otherwise.
Finally, let Ho denote the hypothesis F = G.
Let Tx be the Wilcoxon statistic denned by

Tx= S*Ff (2-1)

214 YVES LEPAGE

and let T2 be the Ansari-Bradley statistic defined by

(2-2)
Let
<r\ = var

(N is odd), (2-3)
{N i s e v e n )
'
2
~ {N is odd).

Downloaded from [Link] at Penn State University (Paterno Lib) on September 16, 2016
A test for Ho versus the alternative a ==
| 1 or & 4= 0 can be based on the pair of statistics
(21, T2). For N> 2, define 2 2
r = (hzM + (hzJh) (2.4)
\ o-i / \ o-2 /
and reject i70 when ^ i s large. This is the test that will be considered here. However, other
possibilities are to base a test on
Zi-/*i Tt-/H | or +
max ^2

but these will not be considered in this paper.

3. EXACT DISTRIBUTION UNDER HO

Let Pmn(A) represent the probability, under Ho, of an event A coming from samples with
m i ' s and n F's, let M be such that N = 2ilf or N = 2M + 1 depending on the parity of N
and finally let k be an integer function define as
\ix(x + 2) (seven),
U (a odd).
The following theorems will give a recursion formula for the random couple (21, T2)
and thus for T.
THEOREM 1. For all n,
(3-1)
For all m,
Pm,oiTi = km(m+l),Tt = h(m)} = 1. (3-2)
For
i <*•*> (i = l,...,M),
(

(3-3)
For

...,N),
(3-4)
Proof. The results (3-1), (3-2) and (3-3) are trivial. For (3-4), it is sufficient to remark that
P«.iVi = h,T2 = t2) = Plim{7i = ^ ( ^ + 1 ) - < 1 J 2 ' 2 = k(N)-t2}. (3-5)
 A combination of Wilcoxon's and Ansari-Bradley's statistics 215
THEOREM 2. For m and n^2,we have under Ho

.n(Ti = h,T2 = y = | ^n(2i = h-2m-n+l,T2 = t2-m)

7/7 77
i^-u-iffi = t1-m,T2 = t2-m)
(3-6)

T1-)Pm-i,n-i(T1 = t1-2m-n+l,T2 = «2-m).J

Downloaded from [Link] at Penn State University (Paterno Lib) on September 16, 2016
Proof. Let

Thus, we have
Pm Tx = *lf T2 = ta\At). (3-7)
It is obvious that
-l)}, Pm>n(A3) = n(n - 1)},
P m>n (^ 2 ) = P m > n (^ 4 ) = mnl{N(N-1)}.
Also, since each of the N\j(m\n\) configurations of the m i ' s and the nY's is equally
probable,
Pm,n(Ti = h,T2 = t2\Ax) = P ^ , ^ ! = t1-2m-n+l,T2 = <2-m), (3-9)
^,,n(^i = h, T2 = <2|^L2) = Pm^n-^ = t.-m, T2 = t2-m), (3-10)
^ . n ^ i = h,T2 = t a |4,) = P ^ ^ ^ = tx-m,T2 = t 2 - m ) , (3-11)
1
[Link].^ = J . I 1 = <2M4) = P ^ n - i f Z k = « x - 2 m - » + 1 , 2 , = <2 -»»). (3-12)
Then, by substituting (3-8) to (3-12) in (3-7), we obtain exactly the result (3-6).

4. ASYMPTOTIC DISTRIBUTION UNDER HO

The term asymptotic distribution is used in the sense of convergence in law when m -*• oo
and n ->• oo with the ratio m/N constant.
THEOREM 3. Under Ho, the statistics Tt and T2 are uncorrelated for all m and n.
Proof. I t is sufficient to show that for all m and n,
EIT&IH,) = ElTjHJEiTJlHo). (4-1)
Since, for all r
, E(VrV8\H0) = m(m- - 1)} (r # «),

we have
216 YVES LEPAGE

which gives after evaluation,

(N is even),
{Niaodd).

The equality (4-3) being also the product of E^H^) and E(T2\H0), by (2-3); thus (4-1) is
satisfied.
THEOREM 4. The asymptotic distribution ofT, under Ho, is x2 with 2 degrees offreedom.
Proof. Let TltN = Txj{mN) and T2N = T2\(mN). It can be easily shown that for all real
numbers <xx and a 2 the statistic cc1 Tr N + a2 T2 N satisfies the asymptotic normality theorem
of Chernoff & Savage (1958). Thus, for all oct and a 2 with (alt a2) * (0,0)

Downloaded from [Link] at Penn State University (Paterno Lib) on September 16, 2016
is asymptotically normal with zero mean and unit variance.
Let Vx = (Tx —fi^jo-x and V2 = (T2-/i2)ja2. Then, by substitution in (4-4) and since
a\ — 4<j|, we have an asymptotic standard normal distribution for

f0ra11 a
i'a* ^ («i. a 2)*(0,0). (4-5)

Thus, by (4-5) and Theorem 3, we can conclude that (T^, V2) is asymptotically bivariate
normal with mean vector zero and identity covariance matrix. Hence, the proof is com-
pleted since T = V\ + V\.

5. CONSISTENCY OF THE TEST

In this section we will find a class of alternatives H such that the test for Ho, of level a,
with critical region T 5s T(a\m,n) is consistent for H.
LEMMA 1. A test for Ho versus a general class of alternatives H based on T, is consistent if the
two-sided test based on Tx or. on T2 is consistent for H.
Proof. From the inequalities
1 ^ P{T > T(a; m,n)\H} > P^-^flo-l > T(oc; m,n)\H}, (5-1)
1 $5 P{T > T[a; m,n)\H) > P{{T2-fi^al > T(oc; m,n)\H}, (5-2)
the result is immediate.
THEOREM 5. Let G(z) = F(ax + b) for all x(a > 0). Then a test based on T for H0:a = 1,
6 = 0 versus H:a + lorb^0is consistent.
Proof, (i) K P ( Z > Y) 4= | , Tx is consistent (van Dantzig, 1951) and the proof follows from
Lemma 1.
(ii) If P{X > Y) = £, obviously a + 1 and, by van Eeden (1964), it is sufficient to show
that «„,
Z> = £ - \AF(x) + {l-A)F(ax+b)-%\dF{x)*O, (5-3)
J -00

where A = m/N, to verify the consistency of T2 and by Lemma 1 to complete the proof.
Let x' = 6/(1 — a) and suppose a < 1. The proof for a > 1 is analogous. Then, for any
x < x', we have F(x) ^ F(ax + b) and, for any x > x', we have F(x) > F(ax + 6).
 A combination of Wilcozon's and Ansari-Bradley's statistics 217
Let [xo,x'o] = {x\AF(x) + (l-A)F(ax + b) = |}. Thus, by introducing x0 in the relation
(5-3), we obtain
D = \F*(xo)-F(xo) + 2(l-A)\X' F(ax + b)dF(x) + l (5-4)
J -oo

Suppose first that x0 < x', then

D>{F(xo)-$*. (5-5)
If F(x0) 4= \, the result is immediate. Therefore, let F(x0) = £. Then it is sufficient to show
that
r*. ex,
F(ax + b)dF(x) > F(x)dF(x). (5-6)
J —00 J — 00

If i^(x) < _F(<zx + 6) for all x < x0, the relation (5-6) is satisfied. If not, since F(x) is a continu-

Downloaded from [Link] at Penn State University (Paterno Lib) on September 16, 2016
ous distribution function, there exists an interval [xv x2] with xt < x2 < x0 where F(x)
is nonconstant and F(x) < F(ax + b), and thus the relation (5-6) is also satisfied. The case
x0 ^ x' is analogous.

The author wishes to give particular thanks to Professor Constance van Eeden of the
Universite de Montreal for her constructive suggestions during the preparation of this
paper.

REFERENCES
ANSABI, A. R. & BRADLEY, R. A. (1960). Rank-sum tests for dispersions. Ann. Math. Statist. 31, 1174-
89.
CHERNOFF, H. & SAVAGE, I. R. (1958). Asymptotic normality and efficiency of certain non-parametric
test statistics. Ann. Math. Statist. 29, 972-94.
MANN, H. B. & WHITNEY, D. R. (1947). On a test of whether one of two random variables is stochas-
tically larger than the other. Ann. Math. Statist. 18, 50-60.
VAN DANTZIG, D. (1951). On the consistency and the power of Wilcoxon's two sample test. Indag.
Math. 13, 1-8.
VAN EEDEN, C. (1964). Note on the consistency of some distribution-free tests for dispersion. J. Am.
Statist. Ass. 59, 104-19.

[Received March 1970. Revised September 1970]

Some key words: Nonparametric rank-sum two-sample test for location and dispersion; Distribution-
free test for location and dispersion; Wilcoxon test; Ansari-Bradley test; Combination of tests.