ASYMPTOTIC EFFICIENCY OF CERTA】N TEST
BASED ON THE METHOD OFηRANKINGS
BY
RYuzo KANNO
1. intrOduction. 1£t us consider the randomized 1)lock layout where m(≧3)treat三
ments∠1,...,Am are applied to m difirerent plots of. a. block, and the experiment i&
replicated,in n(≧2)blockS. Here we esp㏄ia皿y suppose that thesβ m t gatments・−4re clas−
sified into two groups G1={A1,...,Ak}and G2={∠4元+1,...,ノ{m}.which have a、
difirgrence betwoen the two and no−d遜erence withhl each group, and we wish tρtest the:
difference betWeen the two groups. The author studied this problem in・【3】by the method、
of〃rank口1gs which is based on the set ofπ血trablock rank m−tuples ri=(ri1,...,r伽),
’=1,...,n, where々denotes、 the rank of Xi,(challce variable associated witll the・
observation of theノーth treatment of the i−th block)among Xi1,...,】陥m,ノ=1,...,〃z,
i=1,...,〃,and showed, under the null hypothesis that there is no−difference betWeen.
the two groups, the test statistic:
k(m−k)(互…_互…)・
(1’1)
@ f,、[ξ、(嵩・・)・+、融艮…)・]・
w・͡一
メ筋互…一÷ξ、輌・互…一▲皇、・R・…d・・一・pP・・x・−
mately as F−distribution With 1,η3−2d.f. However, the ef五ciency of the test was not:
discussed in I3].
Tlie purpose of this paper is to study the・ asymptotic relative eMciency(A.R.E.)of the・
Fr−test with respect to the correspondng ANOVA .タニtest which is derived by using the
original observations血stead of lank in(1.1), fbr the sequenoe of translation−type alter・・
natives:
(1.2) Kn:1㌃ゴ(x)=
疏(・+h)旬一1・…・k
瓦←一☆・吉)輌一凪…・沈
where Fii(x)=P{為≦x}denotes 4 contiluous c.d.f. of Xtf, and Fi,s may be arbitrarily
* Received June 20,1975
57
づ8
R.KANNO
イ1itferent丘om each other and fUrther瓦, i=1,...,〃, is absolutely continuous distri−
Ib・…n・haVi・g・…血・・u・d・n…y血・・…n綱一毒民ω・a…fy…∫:..∫…鋤
く◎o.
The nu1ゴh6や6伍esis. that there is n6−differert¢e betweeri the伽09t6ups(蓋and G2 is
’τ=0,namely, it is repreSented as
イ1.3) 疏:、厩1ω=...=・Fim(x)=・Fi(x). .
For large〃the hypothesis、Kn is near、日b. so that this type of limit process provides a way
っfstudying the e径ect of small translations on the Fr−test.
We shaI1 start our discussion assurning neither the normality ofthe Fi’s nor the additivity
・of the block effects nor the homoscedasticity of the errors. At first we consider to find the
乏5ymptotie distributions of the、房一and房test under、Kn tespectively and next using the
fesults we deriVe the general fbnnula fbr the A.R.E. of the Fr−test with resp㏄t to the
一房test. As shoWn in later section, undet Kn, the asy血ptotic diStributionS of the statistics
君a血d戸afe nonCentral chi−Squale distributions with sa血e d.f., bUt unequal noncentrality
parametefs, and the A.R.E. is e卯al to the A.RE. of the F亘edfnan’s x蓼 test with respect
to the corresponding ANOVA test. At last寵shall also consider t6.compate the’ power
of the・Fi−test with tespect to the Fried血aむfs x? test. in eonclusi皿, it may be fb加d that
fbt testing the hyp6thes▲s・召b against」(n, the・Fr−te§t is血ote e伍cient thari the z Z test.
The asymptot量C relative e缶cienCy of other rank tests fbr the translation type alternatjve
・has been studied by Andrews【1】, Elteren−Noether【2】, Mehra−safarigi[4], sen[6】, and
.others.
2.Asy叫totic Distribution①f the Statistic、Fr皿der Thms{ation Alternatiyes. Assum一
丘1g the translati皿aiteniativeS Kn in(1.2), we conSider to fi ld the asymptotic disnibution
・of the statistic Fr described in(L1). We shall use the same notations as Puri and Sen[5】
thfoughout thiS papet. N6w let, f()r i=1,...,Zi,
(2・1) .ρま1膓==1⊃{rii==r}・ ρ鰍を#=」P{rif=r’rih=s}・ノ≠ノち r≠ぷ=1・・ … m
孤de・pecia皿yρ擁=動・鴎ρ}鍵=δ・・ク,沙,ノ, h,・,・=1,…・鵬吐e⇒・andδ・8
are the]臼OneCker ddtas.’
Then, fro面he dei㎞tion in(2、1);螺and蟻)are㏄p㏄sented by the飴Uowing飴ms
。。チrl in‘r. ..,
(2.2)
〈2・3) P:ii’g)
・1!輪一?,∫一..,u, Fi・ω(・)
一曇∫:..∫1..冨踊・ω(・)8喜1
dFiゴω『dFih(y),
−[1一Fit(の(x)】dFiゴ(x),
t=・1
s
[Fi。a)O’)一砺(1)(x)]H[(1−[Fiw(1)(ン)]・
Is 1
・where Sj extends over a皿possible(m−1)−tuples:(ぷ(1),....,δ(r_1);’(1),...,t(m_r))of divid一
晦幼二i醐㏄t§(1づ..弓ノー1うノ+1,.、.、.m)・ihtO・tw6・9tOUp§C・血ta面ing・一1,m−r・bjeCtS
Tespectively, and similarly Sjh extends over all possible(〃1−2)−tuples:(故1),◆◆.,〃(r−1):γ(1,,
...,ッ(s_r_1):w(1ハ,...,w(m−8})of dividing〃1−20Ujects(1,...,.ノー1,ノ十1,...,h
−1,h十1,...,〃1)into tlur㏄groups containing r−1,ぷ一アーl」〃1− s objects respectively.
ASYMPTOTIC EFFICIENCY OF CERTAIN TEST
Further, under Kn,(2.2)is represented as follow§:
forノ=1,...,k
5i鈴
(2・・) ・;C一菖(rEτLI,)(−gk)∫...[Fi←+瓢一1一㌔[1−Fi(・+吉)]元
・[Fi←一卿竺濠)]㌔[1一長←−m三・・司r砺←噺
and fbrノ=:k十1,.・・,〃2
(2・・酷一記一f−,)(m−k−1 ξ)∫1..[Fi←+吉)] ∼
卜Fi←+iE;) ]k−r+i+4[民←一栩三・・芳)]‘・
[1一疏←一☆・h)] 鵡←一…昂ン5)
wh・・eifa<・,・h・nw・1・・(;)一・・
Also, we find that irnder Kn(2.3)is equal to
(2・・)螺一(,−1)!(。睾;,(◎!∫:∫:酌一・)8−「−1(1−7)m’sd・d・’一
十〇(1). 、
Thus, we obtain after si皿plifip.a. tiQ. ns the fQllowing resUlt・ 、
LEMMA 2.1. under Kn,励o泌吻’
(2・7) ρ{鍔」==
(2.8)’ρ{儂」召)=
励θrεβ凱一2
0η4ωηVεητ∫卿1砂垣β8生.悟_2
Now, we de血1e
(2.9) Tn.ゴ=
and Iet
(2.10)
∼π=(6n,ゴh);σn,in=COV(Tn
Then,丘om the Lemma 2.1. we can show that tmder」Kn, as〃→oo
(・,11) fi(・一 )一・[1
i−m・吉蹴_・一β昇…・)+・(吉)fo・鳳…・4一
吉+浩・tt(β£ 一・一β』一・)+・(七)f・・ノーk
十1,...,m
輪1
│1)+・(1)如〃j.tk・r≠鳳…s・.
イこ2)∫1..蹴・͡ω]rξω婦(・)一軸
≒β援一、.m=一、=o・for i・=1・・…π・
⊥邑噺=1,e..,m;Tn =:(烏,、,_,T・.m)’
”f司
、,。・== On.1,...,刷….・・一・E(T・.・)一÷語嶋
み)一÷鰭ξ1・嶋一・・・・・・・・・…一
,
ml−・・−ILit. 皇∫1..fi・(叶一・(1)・
’
’
㌔
dO ・ R. KANNO
(2・12) ←・一(m十112){・m一吉ω㌔D一蕊
司where lm=.(1,...,1)’,
・一(−L二}・☆∵・・自:^’
k −
Im:identity mat血of order m.
Using the resU lt・iin Puri and Sen C5】,ρ.270), we know that under Kn the distribution
つ・Va(Tn−mま1輪)・・nv・・g・・t…・−1・di・励・・…輔・h−・…一・
踊∫1..f‘・(x)dU] nd面・・・n一血畷1)1・m一吉⇒・…m・h・d・f・−
nition of 2㌦,」R=(R1,...,Rm), is equal to nTn. Therefbre we have immediately the
’fb皿owing]Lemma.−
LEMMA 2・2・ むitder」Kn・功θdistribu’ionげReonツθ㎎・es to a〃ormal硫〃功〃tio〃ハr(θ,
・V),1〃here
一吻・θ一ヂ1)lm+m・v・i’・巨菖∫二捌弗
t
卿・』ぬ・γ一 (m十112)1・m一吉脇}・
Using the Sverdmp’s Theorem fbr the Hlniting distributioll of a cont血uous function of
エandom variab1卵(s{ePuri and $en【5],、ρ.21), we can obta血the fb皿oWing result from
オhe tc…:2.2.
−2…磁橘一一…͡’・め〃’】再三た)(R(1) R()/
VPt(m十112)一・卿・⊇・・㎜1鋤輌・励一
・・一ノ伽+1袈儒一、)・隠∫二九・α司
⊇・’W・・疏鵬〃(mf紛ぽ…一艮…)・怜;1)一・一一 ・・1
・chi−square distrib吻πw吻η∫’邸励磁η・〃centrality parameter:
〈2・13)
@(;+li511Z#k,m..2−k)・・隠∫二f‘・(州三一^
雲us血glPe Law°f晦N輌題t°theiY g卵e㎜sh°wt』t皿de「K・・ as
側一ぺ』砺≒}しξ㈹一艮…)・+,叢、(Ri一艮…)・]一竿;1)」L…
Thus,丘om(1.1),(2.14), and bβmma 2.3, we obtam the fbllow血g resUlt.
THEOREM 2.1. Undqr」(n, as n→?g, the distributionげthe P}¢onワerges to a noncen’ral
4乃助拠砲伽’i・〃wi吻〃£’d・f.・and.n・〃C・力滅砂クarczm・t・’」。(Fi)噸〃・鋤(2.13).
ASYMPrO皿C EFFIC正NCY OF CERTA困TEST
、10 1
玩三、[k . _Σ(x・i二x(グ=1)・+,鷲、ぽ・一童…)ゴ
where
x・・一菖鞠・矛・1・一÷を・孤・呈…一沈三議み
.NoW,.consider the situation where the nomlality of the Fi’s or the additivity of the
bl・・k緬㏄・・…h・h・m・・㏄d…id・y…h・緬rs皿・…h・1d…pP・・e・h・・為+f,,
ノー1・……跡☆・☆,ノーk+・,…,一血・・P・・d・nt・・d・d・n…ally
di・t・ib・t・d・㏄・・d血g t・a・c・d・晒,・u中th・t∫国4+δdFi(x)<∞, f・…m・δ>01 ahd
司1ゴ=1,...,n, that is, it assures the existence of the丘rst fbur moments of the轟’s.
Under these conditions, we consider to fUid the asymptotic distribution of the statistic
m(3.1).At血st, applying the Iow of large numbers t.o the average, we⑳show that
〈…)姻と、)L£【X・・一矛…)・+、。£、(x・d x()2]二÷皇巧・−e−…
whereσ∼denotes the va亘ance of Xid. NeXt, us血g the dassical central 1imit theo】rem
(under the Liapounoff condition;e.g. see Pud and Sen[5], p.23), we ca血obta丘1 that
as〃→◎◎,
・(3・・) V”(初蒜勾ぽ…一又…)/隠馬・]き
・・−g・・t・an・㎜1d・・垣・・…nw・也一一」識・/巨剖4and皿…a・・一
・…一・・ek(m一幻ぽω一王・nm)・
B砺・]・・一・g・・t・a・・nCe翻←are
distribution with unit d.f. and the noncentrality parameter:
⑭ 蕊・・鵬馬・]
Hence, from(3.1),(3.2)and(3.3), it is fbund that㎝∂εr品舵吻’加化タフ㎞α汐〃lptoti一
侃1砂・…鋤解励棚・τr鋤功晒㎝w乃励』
@磁ψ・・…e・・t・ali力a・am・ter
∠n(ダ)defined b7(3.∋
Thus, we may obtain the A.R.E., following the same arguments as Elteren and Noether
I2]・that is;s五1ce both test statistics Fr andタ}have, under the alternative hypothesis」丘,
asymptotica皿y non㏄entral chi−square・distdbutions with the same num6et Of d.f., the
A.R.E. of the Frtest withエesp㏄t to the字t{苓t(we denote it by ARE(F}toプ))is given
4)ythe ratio of the辻respective:noncentraHty paranleters.且owever, the computation.of
the A.R.E tequ廿es the existen◎e of血n 4%(君)=∠(Fr)a耳d 1㎞ An(ダ)=A(ア). When it
n−°60 n→oo
3.Comparison with1 the Startdard V叡riance Ratio Test. In this s㏄tio耳we cOnsider to
compare the Fr−test with the parametric ANOVA字test based on the variance ratio
criterion . ・ ・ . . ・
k(m−k)(i…_矛…)2−’
m
〈3.1) y=
62 R.KANNO
hQlds, th.e A.R,E, may bg gbta註1ed aS∠ARE(1膓tO夕っ…=A(LFIf)μ(tタ》).廿we here assUme
the existence of
(…) 農鵠馬・一・・㌍÷皇∫:..f’・繊一∫:..f2(肱
then it fbllows that
(3.6)
12玲m2
∠(F∂=
伽十1)伽一k)
⑰一浩・・/・2・
・・
m∫:。.∫・(・)dx]2・
Thus, we have the fb皿gW祖g result.
TliEoREM 3.1. Under the a∬umptionぷげ(3.5). the・A.R.E. of the Fr−teぷt with respect to
the・f一彪ぷ’far teぷting・th, e hypothesis Ho.against Kn is
(3.7) ARE(1 lr庭∼テ)一袈σi[∫:..∫・幽]2・
This is cqual to the A,R,耳. qf the Fr;e4m釦【1 test with respcct to the coπespond血9 para−
metric ANOyA test based on thC vaτiange ratio, fbr tCStinS thC hyPothesis疏against
Kn’・⑭一み(一掛
血the dassical ANOVA Model, bl㏄k e飴cts are additive and the errorS are homo−
scedastic, sg th4‡(3.5)is always satis6ed, Furthe劃nore when∫てκ)is nρ皿a1, it becomes
顛t
(…) ∫:..f・(鋤一,嶽・
Hen㏄, in the dassical ANov.A Mρde1,(3.7)b㏄gmesメ理僻’oグ)=3m/n(m十1)・
4. ComparisOn w汕the.1㌔ie輌,§X多Test. The Friedman’s statistic is
(4・1) ・2=渤(±≒1)菖(R・一〃(m芸’〕12−R’凪
吐・・e・一
@(12
刻¥1){九一㌃脇}胆d疋…under・Kn・・皿・・・…a・…皿・1伽
tribution 2V(θ, V)as shown m Lemna 2.2〔
Now, it can be fb皿d that
(4.2)
(i)
(li)
(lii)θ’A.e =
(4V)2.:=Aγ:i4e皿Pρtent
〃(4γ)=m−1
12km2
(m十1)(m一
k)・・
、茎∫°°
Thus, under Kn, the distribution ofκ2 converges†o a nop㏄ntral iStribUtion
withηm−1d.f. and noncentra五ty paranleter
(・,3) 螺)㍉+1蒜一k)・・隠∫:..f’・(・畔
This is equal to the noncentrality parameter of the、Fr’s distributiol1. Namely, under
Kn, b6th distributions of the Fr andκ多are nonCCntral chi−square distribution with the
ノ・(・)dU]乳
ASYMPrOTIC EFFICIENCY OF CERTAIN TEST
63
same noncelltrality parameter. However, one has〃nit d.f.,the otherhas m−1 d.f. Hence we
can soe that for testing the hypothesis Ho. against瓦1, theルtest is generany more e伍cient
than the Friedman,s X多test.
Acknowledgement The author wishes to express his deepest appreciation to Prof. Y?
Tumu]ra fbr his guidan㏄.
【1】
[21
[3工
凹
【5]
[61
REFERENCES
ANDREWS,・F. C.(1954):Asymptotic behavior of some rank tests fbr an…ltysiS of varian◎e.
ノ1ηπ.1し(ath.・Sta’白’.,25.724−736.
ELTEREN, PH、 VAN and NO正ITHER, G. E(1959):The asymptotic e伍ciency of the Z多・
test for a balai;ced incomplete block design. Biometrika,46.475−477.
KANNO, R.(1’974):Arank sum test血the analysis of variance. ZRσハ4athematics,10.
55−64.
MEHRA,.K. L. and§ARANGI,」.(1967):Asymptotic eMciency of certqin, rank tests永)r.
comparative ex画ments. Ann. Math. Stati t”38・9《FIO7・
PURI, M. L. and SEN, P. K.(1971):Nonpara栩etric methods in multivariate analysis. New
York:John Wiley and Sons.
SEN, P. K.(1967):A note on the asymptotic ethciency of Friedman’s X2−test.助〃昭鰍砺
54.677−678.