6.1. THE DMRL ALTERNATIVE 53
where
… J
内 )ds (6.1.8)and
C
J u a zu
s
可EEEEEBEEBd
11
a︐b S ︐
dt
v
u
FEE‑‑EEEEa﹄∞
ρ t ' ' ' o
2 一 一 ψ
σ
From this corollary the asymptotic variance
σ 3 "
under the null hypothesis予'1
depends on the unknown parametersμand G(t). By the similar method as given in the previous sections we can construct a consistent estimator
3 3 1 イ
[v一品
(t一 ) } ]
2 S!(t‑)de(t)by the theory of counting processes.
Next we consider the asymptotic behavior of the test statistic P2(α
, s )
defiI凶 inthe equation (6.1.6). This result can be proved by Theorem 4.1.1 of Section 4.1 and stated as follows.
COROLLARY 6.1.3. Suppose that the df 's F(t) and G(t) satis今 theconditions
l
T Hh~(t)dC(t)
<∞
(6.1.9)and
η1/2 h1(T)→
o
in probab出tyasπ→∞, (6.1.10) whereん(t):=
1
00 S( u)d匂Then the sequence of the rv's
ポ/2
九 {
(o:,s)‑告 )
converges in distribution as n→∞ to a normal rv with zero mean and variance
1I F‑
‑ ' t w ‑
︐ ︐
••. ︑ ‑
C
一
︐α一角4
‑
︑
E︐ E
︑︑.. 一
︐ ︐ ‑
a︐
b E
︐ ︐
•. ︑ ‑
今 必
‑
L
‑
hN
一4F
54 CHAPTER 6. TESTS FOR DMRL AND HNBUE
where
内:=
1
00 g{S(t)}dt,
ん 川 州
(例tt吋 ) 寸 ∞
S(州珂仰削州包吋併)け}and the function g(t) is defined in the equation (6.1.2).
COROLLARY 6.1.4. Suppose that under the null hypothesis討。 thedf G(t) satis五esthe conditions (6.1.9) and (6.1.10) of Corollary 6.1.3. Then we have as n‑→00
η巾 九α(
,
β)→4N(O, d
,β),
where
σ3β:=f 州
t)}dC(t)Because of the dependency of the asymptotic variance of the statistic 九(α
,
s) under the null hypothesis on the unknowns μand G( t),
we may consider an estimator8 3 β
:=1
Tl { S n い‑州
t)The consistency can be proved by the same technique as used in the proof of Lemma 2.4 of Kumazawa (1987a).
Now we compare the e罰caciesof the test statistics P1 (仇)and P2(α
,
s) forthe alternatives (i)‑(iv) presented in Section 4.1 under the proportional censoring model with the censori泊ngpa町ramete白Eλλ.Fr旨omCorollarie白s6.1.2 and 6.1.4 we may take μ 1 and 0 < λ < 1. And the asymptotic variances of the suit旬ably ωE口rmalize吋dversions of P1 (
ψd
and P2(α,
β) under冗oare given byand
σ~ :=
I
{v ‑<p(t)}2t一λdtJO
σ2ー α ? + α 2 + α 3 +
一 一 ‑
2・ 1‑λI 2β+3‑λI 2α+2β+7ー λI
+~α1α~+
̲2α1αs+2α2αs s+2ー λ α+s+4一入『 α+2β+5‑λ'6.1. THE DMRL ALTERNATIVE 55 Table 6.1
Efficacies
0 /
the DMRL‑test statisticsωhen the
A1ternative
Statistic (i)
。
i} (iii) (iv)λ = 1/10
九(0
,
0) 0.5638 0.6609 0.0469 0.1131 九(0・ ,
5) 0.6575 0.6334 0.0540 0.1435 九(0,
1) 0.7284 0.5997 0.0585 0.1701 P2(ム
0) 0.5942 0.6608 0.0491 0.1222 P2(.5,
.5) 0.6840 0.6293 0.0557 0.1526 P2(.5,
1) 0.7509 0.5931 0.0597 0.1787 九(1,
0) 0.6211 0.6611 0.0510 0.1306 P2(1,
.5) 0.7076 0.6256 0.0571 0.1608 P2(1,
1) 0.7711 0.5874 0.0606 0.1866λ = 3/4
九(0
,
0) 0.0534 0.0625 0.0044 0.0107 九(0,
.5) 0.0857 0.0826 0.0070 0.0187 P2(0,
1) 0.1182 0.0973 0.0095 0.0276 P2(.5,
0) 0.0594 0.0660 0.0049 0.0122 九(.5,
.5) 0.0934 0.0859 0.0076 0.0208 P2(.5,
1) 0.1269 0.1002 0.0100 0.0302 九(1,
0) 0.0647 0.0689 0.0053 0.0136 九(1,
.5) 0.1003 0.0887 0.0081 0.0228 P2( l ,
1) 0.1348 0.1026 0.0106 0.0326respectively
,
whereψ(t),
v and αj'S are given in the equations (6.1.7),
(6.1.8) and (6.1.3)ー(6.1.5),
respectively. Here ior the statistic P1(ψ1) we consider the weight iunction ψ1 (t) to be oi iorm ta. 50 the resulting statistic P1 (ψ1) is equivalent to the statistic九α(,
α).56 CHAPTER 6. TESTS FOR DMRL AND HNBUE
Then the efficacies of the九(a
,
s)‑test are given as follows: for the alter‑ native (i)f
1n(s + 2) ln(α+β+4) ) 2 I ( f ,n¥2 ̲̲2 eff{P2(α,
β ) } = { ‑ } / { ( α + 2 y.lσi } ,
l (
β + 1)(β+ 2) (β+ 2)(α+β+ 3)J
for (ii)
eff{
p : 仇
β)}={ l / ( / l
β +'
,2,¥)"),2(α1+s+4)σ, l / ,
,. 2~2 J
for (iu)
eff{P2(α
,
s)} = ~竺 +JL+α~~L
2 .s
十o
α+β+ 5J
and for (iv)
f ̲
,a2ln(s + 2)α31n(α+β+4)) 2 I 2 eff{乃(α,
β)}= { α 1 + + } z L‑‑.o. ' s+
1 α +s+3J
Some numerical evaluations of the above expressions with some values of αand βyield the entries of Table 6.1. Here we recommend the test based on the P2(1
,
l)‑statistic for this testing problem.6.!. THE HNBUE ALTERNATIVE 57
6.2. The HNBUE A1ternative For testing the null hypothesis
π 。 :
F(t)=
1 ‑exp( ‑tjμ) for t三o
(μunspecl五ed)versus the alternative
冗6 : F(t) is HNBUE
,
but not exponential,
under the random censorship model
,
Bergman and Klefsjo (1985) proposed the class of the test statistics Q1(k) and Q2(k) with k integerと2based on the property that if F(t) is HNBUE then for k=
2,
3, ・ ・ ・ ,
1
0 0 S1c(t)dt三 守
and
f い
1c(t)}dt三 同with l/1c :=
2 : 7 = 1 j .
In Bergman and Klefsjる(1985) a modi泊五制edKa句,子plar吋在島似el estimator was used to de. n
ne the statistics Qdk) and Q2(k) and the asymptotic normality of the suitably normalized versions of the statistics was derived under some strong conditions. Then these statistics can be represented asand
r
S!(t)dtQ1(k) : = . JU ‑;:.
r‑n
f
‑ :
{1 ‑F;(t)}dt
Q2(k) : = ...u l ‑ .... μπ
(6.2.1)
(6.2.2) by using the Kaplan‑Meier estimator Fn(t) de
. n
I凶 inthe equation (4.1.1) of Section 4.1. Here we reject the null hypothesis 冗oin favor of 冗6for large values of Q1(k) and reject冗。 forsmall values of Q2(k). It is seen that the test statistic N2(ψ) withψ(t) = ta and αnonnegative integer,
introduced in Section 5.2 fortesting against the NBUE alternatives
,
is asymptotically equivalent to Q1(α+2).58 CHAPTER 6. TESTS FOR DMRL AND HNBUE
Kumazawa (1989a) introduced a measure of exponentiality against the HN‑
BUE life df's given by
d.1 :
= 1
わ∞ 刊世州刈(0 0
=付∞伽
t)e‑tdt‑1 0 0 κ イ仰
)d仇with nonnegative weight functionゆ(t).Note that d.1 = 0 when Fε冗oand d.1 > 0 when F E 冗6・Ifwe select the weight function
ψ
(t) := ta with α > 0,
the above measure d.1 becomes
+1n/̲¥
J ; ∞ 包
as(u)ぬ1=μa+lr(α)‑~一一 α and a class of the test statistics
f T r g n (
包)du Q3(α). = . " . . ;
U,,‑‑̲内‑1‑1μ九. (6.2.3) for α > 0 may be constructed by using the Kaplan‑Meier estimator Fn(t). This statistic in the uncensored case closely relates to the class of the statistics Ta introdl悶 din Kimball (1947) and may be considered部 anatural extension of Ta for the censored observations. Some properties of Ta are discussed in Lee et al. (1980) in detail. Under the uncensored model Singh and Kochar (1986) considered the above testing problem by using the weight functionゆ(t):=
exp( ‑tjμF)jμF in the measure d.1 and discussed some properties ofthe resulting test statistic.
Here Theorem 9.4 oi Dharmadhikari and Joag‑dev (1988) states that F(t) is an HNBUE liie df ii
,
and only ii,
f
∞f ∞
g(t)ほ p(‑tjμF)dt1
g(t)dF(t)三 3Jo μF
for all nondecreasing
,
convex function g(t) on [0, ∞ ) .
Hence the measure d.1may be also derived from this characterization oi the HNBUE life distributions. Note that we reject
π 。
infavor of冗6ior small values oi Q3(α).Now the asymptotic distribution of the test statistic Q1(k) can be derived from Theorem 4.1.1 oi Section 4.1.
6.2. THE HNBUE ALTERNATIVE 59
COROLLARY 6.2.1. Suppose that for五xedinteger kさ2the df's F(t) and G(t) satisfy the conditions
1
T Hh~(t)dC(t)
<∞
(6.2.4)and
π1/2h1(T)→
o
in probability asη →∞, (6.2.5) whereん(t):=
1
00 S(u)d包Then we have as n→ ∞
ポ ヤ
where
σ;:=j ; H {
μ2h1(t)‑fFh(t)}2dC(t) μ企μ2 : =
1
00s
1&( t)dtand
ん(t):
= イ 内 )
COROLLARY 6.2.2. Suppose that for fixed integer k
2 :
2 and the null hypothesis F(t),
the censoring df G(t) satisfies the conditions (6.2.4) and (6.2.5) ofCorollary 6.2.1. Then we have as n→ ∞1・同
舟4噌Aσ
AU
/a
・ ︑
N J帥↓ ︑EEt﹀ ︐
EEJ1
一
h︑ . ︐
F' m
︐ ︐
I︑
唱A
n E v
rB
﹄t
EK
司z u
︐ ︐ ︐ ︐
'i
制
品
where
σ2L‑r
{S(t) ‑kS"(t)PdC(t)1," .‑ k2
Again we can prove the asymptotic normality of the suitably normalized version of the statistic Q2(k) from Theorem 4.1.1 of Section 4.1.
60 CHAPTER 6. TESTS FOR DMRL AND HNBUE
COROLLARY 6.2.3. Suppose that fOI五xedintegeI k三2the df F( t) and the censoIing df G(t) satisfy the conditions (6.2.4) and (6.2.5) of COIollaIY 6.2.1. Then we have as n
→∞
wheIe
and
ポ ヤ
σ;:=j;H{μ2h1(t)‑fFh(t)}2dC(t) μF
μ2 :
= 1
00 {1 ー戸内内内(ω
附ttり川川)汀}h1(
川州(
ω
収吟t
):= 1
00 S(u)d包ん(t):
=
k1
00 Fc‑ I
1州 包
COROLLARY 6.2.4. Suppose that fOI fixed integeI k三2and the null hypothesis F(t)
,
the censoIing df G(t) satisfies the conditions (6.2.4) and (6.2.5) ofCoIollary 6.2.1. Then we have as n→∞
η1/2{Q2(k) ‑V
I c }
→d N(O, イ I c ) ,
wheIe
σ ; h イ∞山
From Corollaries 6.2.2 and 6.2.4 the asymptotic variances of the suitably normalized versions OfQl(k) and Q2(k) under冗。hypothesisare found to depend on the unknowns μand G(t)
,
and may be estimated byづJ;:{丸
(t‑)‑kS!(t‑)Pde(t)v 1,1c k2
and
勾 I c: = 1
T ν{ふ 日 勾
(tー)ー 1}2de例respectively. The consistency of these estimators can be proved by the same
6.!!. THE HNBUE ALTERNATIVE 61
method as given in Section 4.1.
Next we consider the asymptotic distribution of the test statistic Q3(α) de五nedin the equation (6.2.3).
THEOREM 6.2.5 (Kumazawa (1989a)). Suppose that for五xedconstant α > 0 the df's F(t) and G(t) satis今 theconditions
f 同
(t)dC(t)<∞and
η1/2h;(T)→
o
in probabi1ity as n→ ∞ for i = 1 and 2,
whereん(t):=
1
00 S( u)d包and
ん(t):=
1
00 uCt SThen we have as n→ ∞
η1/2
~
Q3(α)‑ζ:1~→d
N(O, ぺ ) ,
t μF J where J.L2 :=ん(0)and
σ2 .̲
h ; H {
μ2(α+
1)h1(t) ‑μFh2(t)PdC(t)PROOF: We have
肌 π1/2
~
Q3(α) ‑~:1 ~
L μF . J
f T Z π
(s )dh2( s) π1/2(向+1μ~+1 )η1/2h2
(T).‑.,...,‑1‑1 μ 2 ......"'‑l‑1 ↓ . ‑ . , . . . L 1
μ九・ μ九'μF μ九'
(6.2.6)
(6.2.7)
with Zn(t)
=
η1/2{Fn(t) ‑F(t)}jS(S). It is seen from Corollary 3.3 of Serfiing (1980) and the fact thatん
isa consistent estimator ofμF that the second term62 CHAPTER 6. TESTS FOR DMRL AND HNBUE
of the right hand side is asymptotically equivalent to μ2(α+ 1)
: J z .
叫(s)dhds)μ
予
+2 from the conditions (6.2.6) and (6.2.7)Now Lemma 3.1.2 of Section 3.1 together with the Cramむ‑Wold Device implies that the random vector
( l
T Zn(t)dh1(t), l
T Zn(t附 ) )
converges in distribution as n→ ∞ to
(
l
TH却 附
with the limiting process Z(t) of み(t).Hence the limiting rv of Wn asπ→ ∞ can be expressed as
f ; H
Z(s)dh2(S) JL2(α + 1)J ; H
Z(s)dhl(S) μa+l μF ..a+2J ; H {
μ2(α+
1)h1(t) ‑μFh2( t)}dZ( t) μF a+2The desired result follows from the Fubini's Theorem and some calculations. I COROLLARY 6.2.6. Suppose that for五xedconstant α > 0 and the null hypoth‑
esis F(t)
,
the censori時 dfG(t) satisfies the conditions (6.2.6) and (6.2.7) of Theorem 6.2.5. Then we have as n→ ∞η1/2{Q3(α) ‑r(α+1)}→d N(O
, イ
a)' where~,a
:=1
00 { r(α+仰 )̲J
1LO O戸
t)小
C(包)Because of the dependency of the asymptotic varianceσ3,a UIIdeE冗oon the unknown parametersμand G(t)
,
we may use a consistent estimatorrT[n ,
,n¥n L ¥: 1
ftaEn(t)dt ) 2: = l 刊
(α+2)ぬい‑)‑
A O M }dC(包)Jo ..l μ九回 J
6.~. THE HNBUE ALTERNATIVE 63
by the theory of counting processes.
We shall consider the e侃caciesof the tests based on the statistics Q1(k)
,
Q2(k) and Q3(α) against the alternatives (i)‑(iv) listed in Section 4.1 under the proportional censoring model with否(t)= Sλ(t). Then Corollaries 6.2.2
,
6.2.4 and 6.2.6 imply that μ = 1 and 0 < 入 <1. And the asymptotic variances of their suitably normalized versions under冗oare given byand
宅 1 2 1
σ1,11 :
=
k2(1ー λ)一夜 E て可+
2k‑λ‑1' 仏 ・ ー ヤ 旬一 ‑
2
, " . ‑
~ i‑1一入σ d
必心 ふ
2ふ
hν,,:= 戸 = 斗= (件h削k似!川白
(ο1 一 λ刈げ)i+1'respectively
,
whereand
αi,"
乞
ω C7n,II,
l+7n=i
川 :=
2 :
dl,,,d7n,,,,l+7n=i
ω : = ( 川 合 )
ffoorr i i=
= 21,
3, . . . ,
k,
for i
=
0for i = 1
,
2, . . . ,
k.Here we assume for the statistic Q3(α) that α= k is positive integer.
As stated in the beginning of this section the Q1 (k )‑statistic is equivalent to the N2(ψ)‑statistic with ψ(t)
=
t"‑2,
we do not give the expressions for theefficacies of the test based on Q1(k). Then some calculations show that for the
64 CHAPTER 6. TESTS FOR DMRL AND HNBUE
Table 6.2
Efficαcies 01 the HNB UE‑test dαtutics ωhen the censoring parameterλ=
ム
and2 Statutic /りQ1(2) 1.2474 Q1(3) 1.3732 Q2(2) 1.2474 Q2(3) 1.1289 Q3(1) 0.7217 Q3(2) 0.3889
Q1(2) 0.1863 Q1(3) 0.3497 Q2(2) 0.1863 Q2(3) 0.1386 Q3(1) 0.01 Q3(2) 0.0008
alter口rna
eff州 仰
f パ ω { ω
偽仰附Q仏ω2
似バ μ
州(刈h作)例仲 ザ州f
汽 刷 { for (ii)A1ternative
。
i) (iiり0.6490 0.0721 0.5056 0.0711 0.6490 0.0721 0.6932 0.0691 0.7217 0.0451 0.5600 0.0243
0.0969 0.0107 0.1287 0.0181 0.0969 0.0107 0.0851 0.0084 0.01 0.0006 0.0011 0.0001
(iv
0.3874 0.4 776 0.3874 0.3287 0.1804 0.0847
0.0578 0.1216 0.0578 0.0403 0.0025 0.0002
6.~. THE HNBUE ALTERNATIVE
for (i益)
ザ引州
f
パ 馴 {eザ引州
f
汀ff{バ ω
胤 刷{
仙ω
Q仏似3(め
k刈 ) リ } ト = { ド 「 灯 刷 r t ( ケ : γ
什叫γ + γ)
2幻 リ
}2jσ ; h '吋引州仰
パ ω f 刷 ω {
偽ω
仏Q似バ
2( k 刈 刈 ) リ } ト = { 巳 ; z 右 川 ( k ぺ γ ; 7 ~ 1 1 う ) ( い 川
山i++刊叫1
り)ι い ; し 九 υ れ山
i+什J
2幻 ψ ) }
e
ザ引州f 汀 爪 f パ m
胤 刷ω {
偽仏ω3(Q似仏
k)ト } ト = ゴ = r ド巾
2代2(伴k+ベ 与 1 + 持 ) ' / σ ;
h,
and for (iv)
ザf{Q2(kト (νh-hZ(叫ん~ ) 1 思 3 ) 2 / 心
eff{Q3(k)}
= r
2(k+
1){ν10+1 ‑k ̲1}2jσ;,h,
respectively
,
whereγdenotes the Euler's constant.65
Table 6.2 shows the e罰caciesof the tests based on the statistics Ql(2), Q1(3)
,
Q2(2),
Q2(3),
Q3(2) and Q3(2) against the alternatives (i)‑(iv) and some values of the censoring parameterλThe poorness of the performance of the test based on the Ta‑statistic,
equivalent to the Q3(α)‑statistic in the uncensored case,
was pointed out in Lee et al. (1980) and it seems that the Q3(α)‑statistic in the censored case inherits the characteristic of the Ta‑statistic. Here we recom‑ mend the Q1 (3)‑test for testing exponentiality against the HNBUE alternatives under the censored model based on the concept of the Pitman asymptotic rela‑ tive efficiency.Acknowledgments
I wish to thank Professor S. Shirahata of Osaka U niversity for his advice and constructive criticism. My thanks also go to Professor M. Okamoto of Otemon University
,
Professor N. Inagaki and Professor K. Ishu of Osaka University from whom I acquired statistical backgrounds and helpful discussions.References
Aalen
,
O. O. (1978). Nonparametric inference for a family of counting processes. Ann. Stαtist. 6,
701‑726.Andersen
,
P. K.,
Borgan, o . ,
Gill,
R. D. and Keiding,
N. (1982). Linear non‑ parametric tests for comparison of counting processes,
with applications to censored survival data. lnter. Statist. Rev. 50,
219‑258.Barlow
,
R. E.,
Bartholomew,
D. J.,
Bremr悶 ,
J. M. and Bru此 ,
H. D. (1972). Stαtisticallnference under Order Restrictions. John Wiley,
New York. Barlow,
R. E. and Campo,
R. (1975). Total time on test processes and applica‑tions to failure data analysis. In Reliαbility and ,Pαult Tree Analysis
,
R. E. Barlow,
J. Fussell and N. D. Singpurwalla,
eds.,
SIAM,
Philadelphia,
451‑481.
Barlow
,
R. E. and Doksu凧 K.(1972). Isotonic tests for convex orderings. Proc. Sizth Berkely Sympo. 1,
293‑323.Barlow
,
R. E. and Proschan,
F. (1969). A note on tests for monotone if仏白a.ilu rate based on incomplete data. Ann. Math. Statist. 40,
595‑600.Barlow
,
R. E. and Proschan,
F. (1975). Stαtistical Theo叩 ofReliabilityαnd Life Testing: Probability Models. Holt,
Rinehart and Winston,
New York. Bergman,
B. and Kle色UO,
B. (1985). Testing against HNBUE with randomlycensored data. In Contributions to Prob. Statist. in Honour of G. Blom
,
J. Lanke and G. Lindgrem
,
eds.,
Lund,
49‑72.Bergman, B. and Klefsj,る B. (1989). A family of test statistics for detecting monotone mean residuallife. J. Stαtist. Plan. lnfer. 21
,
161‑178.Billingsley
,
P. (1968). Converget問 。IfProbability Measures. John Wiley,
NewREFERENCES 69
York.
Chen
,
Y. Y.,
Hollander,
M. and LaIゆerg,
N. A. (1983a). Testing whether new is better than used with randomly censored data. Ann. Statist. 11,
267‑ 274.Chen
,
Y. Y.,
Hollander,
M. and La時berg,
N. A. (1983b). Tests for monotone mean residuallife,
using randomly censored data. Biometrics 39,
119‑127. De Souza Borges,
W.,
Proschan,
F. and Rodrig町 s,
J. (1984). A simple test for new better than used in expectation. Commun. Stαtist.‑Theory Meth. A13,
3217‑3223.Deshpande
,
J. V. (1983). A class of tests for expOI悶 ltialityagainst increasing failure rate average alternative. Biometrika 70,
514‑518.Dharmadhikari
,
S. and Joag‑dev,
K. (1988). Unimodαlity,
Convezity,
αnd Appli‑ cations. Academic Press,
New York.Fernholz
,
L. T. (1983). Von Mises Calculus for Statistical Functionals. Springer‑Verlag,
New York.G出, R. D. (1980). Censori吋 andStochastic Integral. Mathematical Centre Tracts 124
,
Mathematisch Centrum,
Amsterdam.Gill
,
R. D. (1983). Large sample behaviours of the product‑limit estimator on the whole line. Ann. Statist. 11,
49‑58.Hollander
,
M. and Proschan,
F. (1972). Testing whether new is better than used. Ann. Math. Statist. 43,
1136‑1146.Hollander
,
M. and Proschan,
F. (1975). Tests for the mean residual life. Biometrika 62,
585‑593.Hollander
,
M. and Proscl則 l,
F. (1984). Nonparametric concepts and methods in reliability. In Handbook of Statistics,
Vol. ./. P. K. Krishnaiah and P. K. Sen. eds.,
North‑Holland,
New York,
613‑655.Hollander
,
M.,
Park,
D. H. and Proschan,
F. (1985). Testing whether new is better than used of a speci:fied age,
with randomly censored data. Canad. J. Statist. 13,
45‑52.70 REFERENCES
Hollander
,
M.,
Park,
D. H. and Proschan,
F. (1986). A class oflife distributions for aging. J. Amer. Statut. A.uoc. 81,
91‑95.Joe
,
H. and Proschan,
F. (1983). Tests for properties of the percentile residual life function. Commun. Statiot.‑Theory Meth. A12,
1087‑1119.Joe
,
H. and Proschan,
F. (1984). Percentile residualli長 functions.Oper. Reo. 32,
668‑678.Kaplan
,
E. L. and Meier,
P. (1958). Nonparametric estimation from incomplete observations. J. Amer. Statut. Auoc. 53,
457‑481.Kimball
,
B. F. (1947). Some basic theorems for developing tests of五tfor the case of nonparametric probability distribution functions. Ann. Math. Statut. 18,
540‑548.Kle色jo
,
B. (1981). HNBUE survival under some shock models. Scαnd. J. Stati8t. 8,
39‑47.Kle色j,る B. (1982a). On aging properties and total time on test transforrr比
Scand. J. Stαtiot. 9, 37‑41.
Klefsj,る B. (1982b). The HNBUE and HNWUE classes of life distributions. Naval Reo. Logut. Quart. 29
,
331‑344.Kle色uo
,
B. (1983). Some tests against aging based on the total time on test transform. Commun. Statiot.‑Theory Meth. A12,
907‑927.Ko叫 H.L. (1978a). A class of tests for testing new is better than used.' Canad. J. Statiot. 6
,
249‑271.Koul
,
H. L. (1978b). Testing for new better than used in expectation. Commun.Statut.‑Theory Meth. A7
,
685‑701.Ko叫 H.L. and Susarla
,
V. (1980). Testing for new better than used in expec‑ tation with incomplete data. J. Amer. Stαtiot. Aoooc. 75,
952‑956. Kumazawa,
Y. (1983). A class oftest statistics for testing whether new is betterthan used. Commun. Statiot.‑Theory Meth. A12
,
311‑321.Kumazawa
,
Y. (1984). Tests for increasing failure rate average with randomly censored data. Working Paper No. 11,
Shiga Univ.REFERENCES 71
K umazawa
,
Y. (1986a). Tests for new better than used in expectation with randomly censored data. Seq. Anal. 5,
85‑92.Kumazawa
,
Y. (1986b). Large sample behaviors of statistics under random censorship. Working Paper No. 15,
Shiga Univ.Kumazawa
,
Y. (1986c). On testing NBU for random censors}吟 .Economic Rev. Shiga Univ. 240,
79‑85.Kumazawa
,
Y. (1986d). Testing whether new better than used in expedation for random censorship. Seq. Anal. 5,
339‑346.Kumazawa
,
Y. (1986e). A class of tests for new better than used in expedation with incomplete data. Seq. Anal. 5,
347‑361.Kumazawa
,
Y. (1987a). On testing whether new is better than used using ran‑ domly censored data. Ann. Statiot. 15,
420‑426.Kumazawa
,
Y. (1987b). A note on an estimator oflife expectancy with random censorship. Biometrika 47,
655‑658.Kumazawa
,
Y. (1988a). Tests for NBUP‑αand NBU‑to under random censor‑ ship. Economic Rev. Shiga Univ. 253・254,
169‑182.Kumazawa
,
Y. (1988b). Testing exponentiality based on total time on test transform under random censorship. Working Paper No. 19,
Shiga Univ. Kumazawa,
Y. (1989a). Testing expor悶 ttialityfor HNBUE under random cen‑sorship. Economic Rev. Shiga Univ. 260・261.(in press).
Kumazawa
,
Y. (1989b). Testing against NBUE under random censorship. Eco‑ nomic Rev. Shiga Univ. 257,
95‑102.Larゆerg
,
N. A.,
Leon,
R. V. and Proschan,
F. (1980). Characterization ofnon‑parametric classes of life distributions. Ann. Prob. 8
,
1163‑1170.Lee
,
S. C. S.,
Locke,
C. and Spurrier,
J. D. (1980). On a class of tests of exponentiality. Technometric$ 22,
547‑554.Marshall
,
M. and Proschan,
F. (1972). Classes of distributions applicable in replacement with renewal theory implications. Proc. Sizth Berkely Sympo・1