Upload
lee-shen-chen
View
215
Download
1
Embed Size (px)
Citation preview
Contents lists available at ScienceDirect
Journal of Statistical Planning and Inference
Journal of Statistical Planning and Inference 141 (2011) 2670–2681
0378-37
doi:10.1
� Cor
E-m
journal homepage: www.elsevier.com/locate/jspi
Empirical Bayes testing for equivalence
Lee-Shen Chen a,�, Ming-Chung Yang b,c
a Applied Statistics and Information Science, Ming Chuan University, Taoyuan, Taiwanb Graduate Institute of Statistics, National Central University, Chung-Li, Taiwanc Department of Banking and Finance, Kainan University, Taoyuan, Taiwan
a r t i c l e i n f o
Article history:
Received 30 August 2010
Received in revised form
22 January 2011
Accepted 11 February 2011Available online 18 February 2011
Keywords:
Asymptotically optimal
Empirical Bayes test
Equivalence
Regret
58/$ - see front matter & 2011 Elsevier B.V. A
016/j.jspi.2011.02.019
responding author.
ail address: [email protected] (L.-S. Ch
a b s t r a c t
For a fixed point y0 and a positive value c0, this paper studies the problem of testing the
hypotheses H0 : jy�y0jrc0 against H1 : jy�y0j4c0 for the normal mean parameter yusing the empirical Bayes approach. With the accumulated past data, a monotone
empirical Bayes test is constructed by mimicking the behavior of a monotone Bayes
test. Such an empirical Bayes test is shown to be asymptotically optimal and its regret
converges to zero at a rate ðlnnÞ2:5=n where n is the number of past data available, when
the current testing problem is considered. A simulation study is also given, and the
results show that the proposed empirical Bayes procedure has good performance for
small to moderately large sample sizes. Our proposed method can be applied for testing
close to a control problem or testing the therapeutic equivalence of one standard
treatment compared to another in clinical trials.
& 2011 Elsevier B.V. All rights reserved.
1. Introduction
Since Robbins (1956, 1964), the empirical Bayes approach to statistical decision problems has generated considerableinterests among the researchers, and such a procedure has been extensively studied in the literature. For the recentdevelopment of empirical Bayes procedures, see Pensky (2002, 2003), Gupta and Li (2006), Li and Gupta (2002), and Liang(2002, 2006a) and the references cited in these papers. Although there are many works on empirical Bayes procedures,two-tailed tests, which have great applications to practical problems, have unfortunately drawn a little attention inempirical Bayes literatures. We mention a few research work in this direction. Wei (1991) and Liang (1999) have studiedempirical Bayes two-tailed tests for one-parameter discrete exponential family, respectively. Wei (1989), Singh and Wei(2000), and Liang (2006b) have, respectively, studied empirical Bayes two-tailed tests for one-parameter continuousexponential families. Liang (2006c) has studied a simultaneous inspection of variable equivalence for finitepopulations. Chen (2009) has studied on empirical Bayes two-tail tests for double exponential distributions. It is surprisingthat normal distribution is an important distribution model but there is no empirical Bayes work done regarding the two-tailed tests for the normal mean.
Consider a normal distribution with mean y and variance s2, where both y and s2 are unknown. Let y0 be a fixed pointand c0 be a positive value. One is interested in testing the hypotheses H0 : jy�y0jrc0 versus H1 : jy�y0j4c0. Let i, i=0,1,denote an action deciding in favor of Hi. For the parameter y and action i, the loss is defined to be
lðy,iÞ ¼ ð1�iÞ½ðy�y0Þ2�c2
0 �Iðjy�y0j�c0Þþ i½c20�ðy�y0Þ
2�Iðc0�jy�y0jÞ, ð1:1Þ
ll rights reserved.
en).
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2671
where I(x)=1 if x40, and 0 otherwise. It is assumed that the parameter y is a realization of a random variable Y having anunknown nondegenerate prior distribution G over the parameter space O¼ ð�1,1Þ and s2 is a fixed, unknown parameter.
For making a decision, a sample of size mðmZ2Þ random variables X1, . . . ,Xm is taken from a Nðy,s2Þ distribution. Let
Y ¼ ð1=mÞPm
j ¼ 1 Xj and W ¼Pm
i ¼ 1ðXi�YÞ2. Thus, Y follows a Nðy,t2Þ distribution with probability density f ðyjy,t2Þ ¼
1=ðffiffiffiffiffiffi2pp
tÞexpð�ðy�yÞ2=ð2t2ÞÞ, where t2 ¼ s2=m; W=s2 has a w2ðm�1Þ distribution. Obviously, Y and W are independent andY is a sufficient statistic for the parameter y. Since we are interested in Bayes and empirical Bayes tests regarding theparameter y, it suffices to consider tests based on Y.
Let Y be the sample space of Y. A test d is defined to be a mapping on Y into the interval [0, 1], such that for each y in Y.dðyÞ ¼ Pfaccepting H1jY ¼ yg is the probability of accepting H1 when Y=y is observed. Since y0 is known and Y�y0 has aNðy�y0,t2Þ distribution, without loss of generality, it is assumed that y0 ¼ 0.
Let RðG,dÞ denote the Bayes risk of a test d when G is the true prior distribution. Assume thatRy2 dGðyÞ �m2ðGÞo1. Let
fGðyÞ ¼
Zf ðyjy,t2Þ dGðyÞ : the marginal density of Y ,
HGðyÞ ¼
Zðc2
0�y2Þf ðyjy,t2Þ dGðyÞ ¼ c2
0 fGðyÞ�aGðyÞ,
aGðyÞ ¼
Zy2f ðyjy,t2Þ dGðyÞ,jGðyÞ ¼HGðyÞ=fGðyÞ,
cG ¼
Zðy2�c2
0ÞIðy2�c2
0Þ dGðyÞ: ð1:2Þ
By Fubini’s theorem, a straightforward computation yields that
RðG,dÞ ¼ZYdðyÞHGðyÞ dyþcG ¼
ZYdðyÞfGðyÞjGðyÞ dyþcG: ð1:3Þ
A Bayes test dG, which minimizes the Bayes risks among a class of tests, can be obtained from (1.3) as follows:
dGðyÞ ¼ 1 if HGðyÞo0, and 0 otherwise
¼ 1 if jGðyÞo0, and 0 otherwise: ð1:4Þ
The minimum Bayes risk of this testing problem is
RðG,dGÞ ¼
ZYdGðyÞHGðyÞ dyþcG: ð1:5Þ
Since the prior distribution G is unknown, it is not possible to implement the Bayes test dG for the underlying testingproblem. When a sequence of past data is available, we shall study this testing problem using the empirical Bayesapproach of Robbins (1956, 1964).
The rest of the paper is arranged as follows. In Section 2, we construct an empirical Bayes test d�n. Then, the asymptoticoptimality of d�n is studied in Section 3. It is shown that the regret of d�n converges to zero at a rate ðlnnÞ2:5=n where n is thenumber of past data available when the current testing problem is considered. The detailed proof to derive theconvergence rate of the proposed test can be found in Appendix. In Section 4, a Monte Carlo simulation study is givento illustrate the performance of regret for small to moderately large sample sizes of the proposed test. We also use thepercentage of relative regret reduction as a measure of the performance of the proposed empirical Bayes test for anysample size. The results of the simulation study demonstrate that our proposed empirical Bayes test has goodperformance. Our proposed method can be applied for testing close to a control problem or testing the therapeuticequivalence of one standard treatment compared to another in clinical trials.
2. Construction of an empirical Bayes test
2.1. Empirical Bayes framework
In the empirical Bayes framework, let Xi1, . . . ,Xim be a sample of size m arising from a Nðyi,s2Þ distribution at stage i,where yi is a realization of a random variable Yi. Also, let Yi and Wi be the corresponding (Y,W) based on ðXi1, . . . ,XimÞ. It isassumed that ðYi,Wi,YiÞ, i¼ 1,2, . . . are iid copies of ðY ,W ,YÞ, where ðYi,WiÞ, i¼ 1,2, . . . are observable, but Yi,i¼ 1,2, . . . arenot observable. Let YðnÞ ¼ ðY1, . . . ,YnÞ and WðnÞ ¼ ðW1, . . . ,WnÞ. At stage n+1, ðYi,WiÞ,i¼ 1, . . . ,n, denote the n past data, andðYnþ1,Wnþ1Þ � ðY ,WÞ is the present random observation. Let ynþ1 be a realized value of the present random variable Ynþ1.We are interested in testing H0 : jynþ1jrc0 versus H1 : jynþ1j4c0 using the error loss lðynþ1,iÞ of (1.1) with y0 ¼ 0.
An empirical Bayes test dn is a probability function defined on Y=y and (Y(n),W(n)) such that dnðy,YðnÞ,WðnÞÞ � dnðyÞ isthe probability of accepting H1 : jynþ1j4c0. Let RðG,dnjYðnÞ,WðnÞÞ denote the Bayes risk of an empirical Bayes test dn
conditioning on ðYðnÞ,WðnÞÞ, and RðG,dnÞ � EnRðG,dnjYðnÞ,WðnÞÞ the overall Bayes risk of dn, where the expectation En istaken with respect to (Y(n),W(n)). Since RðG,dGÞ is the minimum Bayes risk, which could be achieved if the prior
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–26812672
distribution G were known and the Bayes test dG had been applied. Thus, RðG,dnjYðnÞ,WðnÞÞ�RðG,dGÞZ0 for all ðYðnÞ,WðnÞÞand for all n. Hence, RðG,dnÞ�RðG,dGÞZ0 for all n. The nonnegative difference RðG,dnÞ�RðG,dGÞ, called as the regret of theempirical Bayes test dn, is often used as a measure of performance of the empirical Bayes test dn; for example, see Johnsand Van Ryzin (1972) and van Houwelingen (1976). An empirical Bayes test dn is said to be asymptotically optimal,relative to the prior distribution G, at a rate en if RðG,dnÞ�RðG,dGÞ ¼OðenÞ, where feng is a sequence of positive, decreasingnumbers such that limn-1en ¼ 0.
2.2. An alternative form of the Bayes test
Throughout the paper, it is assumed that the prior distribution G satisfies the following assumption:
Assumption G. [G1] G is nondegenerate and symmetric about the point y0 ¼ 0.
½G2� jGð0Þ404 limy-1
jGðyÞ:
½G3� m2ðGÞ �
Zy2 dGðyÞo1:
Under Assumption [G1], fG(y), HG(y) and jGðyÞ ¼HGðyÞ=fGðyÞ are even functions. Also, jGðyÞ is strictly decreasing in jyj. SincejGðyÞ is a continuous function, by Assumption [G2], there exists a point aG,0oaGo1, such that
jGðyÞ40 for jyjoaG, jGðaGÞ ¼ 0, jGðyÞo0 for jyj4aG: ð2:1Þ
Combining (1.4) and (2.1), the Bayes test dG can be expressed as
dGðyÞ ¼ 1 if jyj4aG, and 0 if jyjraG: ð2:2Þ
aG is thus called a critical point of the Bayes test dG. From (2.2), dG can be viewed as a monotone test of jyj.
2.3. Empirical Bayes estimation
Since an empirical Bayes test often mimics the behavior of a Bayes test, in order to construct an empirical Bayes test,from (1.4) and (2.2), we need to estimate the function HG(y) and the critical point aG. For this purpose, we give arepresentation of HG(y) on which an empirical Bayes estimator Hn(y) for HG(y) is based.
First note that aGðyÞ ¼ t4f ð2ÞG ðyÞþ2t2yf ð1ÞG ðyÞþðt2þy2ÞfGðyÞ, where f ðiÞG ðyÞ denotes the i-th derivative of fG(y) with respectto y and f ð0ÞG ðyÞ ¼ fGðyÞ. Thus,
HGðyÞ ¼ ðc20�t
2�y2ÞfGðyÞ�2t2yf ð1ÞG ðyÞ�t4f ð2ÞG ðyÞ: ð2:3Þ
Hence, we need to estimate f ðlÞG ðyÞ,l¼ 0,1 and 2.We let cfG
ðtÞ, cNðtÞ and cGðtÞ denote the characteristic functions associated with the probability density fG, a Nð0,t2Þ
distribution and the prior distribution G, respectively. Thus, cNðtÞ ¼ expð�t2t2=2Þ and cfGðtÞ ¼cNðtÞcGðtÞ. By the inversion
formula,
f ðlÞG ðyÞ ¼1
2p
Zð�itÞlexpð�ityÞcfG
ðtÞ dt, l¼ 0,1,2: ð2:4Þ
Consider the kernel function KðxÞ ¼ sinx=ðpxÞ. The Fourier transform of K is cK ðtÞ ¼ I½�1,1�ðtÞ. Also,K ð1ÞðxÞ ¼ ðxcosx�sinxÞ=ðpx2Þ, K ð2ÞðxÞ ¼ ð2sinx�2xcosx�x2sinxÞ=ðpx3Þ. Thus, jK ðlÞðxÞjr1 for l=0,1,2, and
RjK ðlÞðxÞj2
dxrk0o1 for some finite k040. We note that the types of estimators of (2.4) have been used in Liang (2006a)and Chen (2009) for constructing empirical Bayes tests.
Estimation of s2 and t2: Note that W1, . . . ,Wn are iid, Wi follows a s2w2ðm�1Þ distribution. Define s2n ¼WðnÞ=ðnðm�1ÞÞ,
WðnÞ ¼W1þ � � � þWn and t2n ¼ s2
n=m. Thus, nðm�1Þt2n=t2 ¼ nðm�1Þs2
n=s2 � w2ðnðm�1ÞÞ, E½t2n� ¼ t2, Varðt2
nÞ ¼ 2t4=ðnðm�1ÞÞand E½t4
n� ¼ t4½1þ2=ðnðmþ1ÞÞ�.Estimation of f ðlÞG ðyÞ,l¼ 0,1,2: Let b¼ 1=
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi4mlnnp
. Define cnðtÞ ¼ ð1=nÞPn
j ¼ 1 expðitYjÞ. Note that cnðtÞ is the empiricalcharacteristic function based on Y(n) and is an unbiased and consistent estimator of cfG
ðtÞ. By mimicking the form of fG(l)
(y)given in (2.4), define
f ðlÞn ðyÞ ¼1
2p
Zð�itÞlexpð�ityÞcnðtÞcK ðbsntÞ dt: ð2:5Þ
Note that fn(l)
(y) can also be expressed as
f ðlÞn ðyÞ ¼1
nðbsnÞlþ1
Xn
j ¼ 1
K ðlÞy�Yj
bsn
� �: ð2:6Þ
Estimation of HGðyÞ: Mimicking the form (2.3), we estimate HG(y) by Hn(y), where
HnðyÞ ¼ ðc20�t
2n�y2ÞfnðyÞ�2t2
nyf ð1Þn ðyÞ�t4nf ð2Þn ðyÞ: ð2:7Þ
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2673
Estimation of aG: Define cn ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffilnn=ð4m2Þ4
p. Since cn-1 as n-1, thus, for sufficiently large n, 0raGocn. Since jGðyÞ
and HG(y) have the same sign and jGðyÞZ0 for 0ryraG, and jGðyÞo0 for y4aG, the critical point aG can be expressed as:aG ¼
R cn
0 I½jGðyÞ� dy¼R cn
0 I½HGðyÞ� dy. Thus, we estimate aG by ann, where
a�n ¼
Z cn
0I½HnðyÞ� dy: ð2:8Þ
We have
a�n�aG ¼�
Z aG
0I½�HnðyÞ� dyþ
Z cn
aG
I½HnðyÞ� dy: ð2:9Þ
2.4. Monotone empirical Bayes test d�n
By mimicking the behavior of the Bayes test dG given in (2.2), we propose an empirical Bayes test d�n as follows:
d�nðyÞ ¼ 1 if jyj4a�n, and 0 otherwise: ð2:10Þ
The Bayes risk of d�n is
RðG,d�nÞ ¼ZY
En½d�
nðyÞ�HGðyÞ dyþcG: ð2:11Þ
From (2.10), we see that d�n is a monotone test of jyj.
3. Asymptotic optimality
The empirical Bayes test d�n in (2.10) possesses the following asymptotic optimality.
Theorem 3.1. Suppose the prior distribution G satisfies Assumption G. Then, the empirical Bayes test d�n is asymptotically
optimal, and
RðG,d�nÞ�RðG,dGÞ ¼OððlnnÞ2:5=nÞ:
Proof. Since jGðyÞ is continuous and strictly decreasing in jyj and jGðaGÞ ¼ 0, we let aG1ðnÞoaGoaG2
ðnÞ be points such that
jGðaG1ðnÞÞ ¼ 1=
ffiffiffiffiffiffiffiffinb5
pand jGðaG2
ðnÞÞ ¼ �1=ffiffiffiffiffiffiffiffinb5
p. Note that aG1
ðnÞ-aG and aG2ðnÞ-aG as n-1. Let a1=aG/2 and a2=aG+1.
We consider the case where n is sufficiently large so that a1oaG1ðnÞoaGoaG2
ðnÞoa2ocn. From (1.4)–(1.5)
and (2.10)–(2.11), the regret of the empirical Bayes test d�n can be written as
0rRðG,d�nÞ�RðG,dGÞ ¼ 2En
Z aG
a�n
HGðyÞ dy
" #¼ 2En Ið1�ja�n�aGjÞ
Z aG
a�n
HGðyÞ dy
( )þ2En Iðja�n�aGj�1Þ
Z aG
a�n
HGðyÞ dy
( ):
ð3:1Þ
Since HG(aG)=0, by Taylor series expansion, for some yn between ann
and aG,
En Ið1�ja�n�aGjÞ
Z aG
a�n
HGðyÞ dy
( )¼ En Ið1�ja�n�aGjÞH
ð1ÞG ðy
�Þða�n�aGÞ2=2
n orHð1Þ
�
G ðaGÞEnða�n�aGÞ
2, ð3:2Þ
where Hð1Þ�
G ¼maxfj12Hð1ÞðyÞj; jy�aGjr1g.
Also, by Markov inequality
En Iðja�n�aGj�1Þ
Z aG
a�n
HGðyÞ dy
( )rEnða
�n�aGÞ
2Z 1
0jHGðyÞj dy¼mHG
Enða�n�aGÞ
2, ð3:3Þ
where mHG¼R1
0 jHGðyÞj dyo1 by Lemma A4 in the Appendix. Therefore,
0rRðG,d�nÞ�RðG,dGÞrc0Enða�n�aGÞ
2, ð3:4Þ
where c0 ¼Hð1Þ�
G ðaGÞþmHG. From (2.9)
a�n�aG ¼�
Z a1
0I½�HnðyÞ� dy�
Z aG1ðnÞ
a1
I½�HnðyÞ� dy�
Z aG
aG1ðnÞ
I½�HnðyÞ� dy
þ
Z aG2ðnÞ
aG
I½HnðyÞ� dyþ
Z a2
aG2ðnÞ
I½HnðyÞ� dyþ
Z cn
a2
I½HnðyÞ� dy�X6
i ¼ 1
AiðnÞ:
Hence,
Enða�n�aGÞ
2r6X6
i ¼ 1
En½A2i ðnÞ�: ð3:5Þ
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–26812674
Therefore, it suffices to investigate the asymptotic behaviors of En[Ai2(n)] for each i=1,y,6. The asymptotic properties of
Ai2(n) have been evaluated in Appendix. From the Appendix, we have
En½A21ðnÞ�ra2
G½4g1ðnÞþ3g2ðnÞþg3ðnÞþg4ðnÞþg5ðnÞ�, ð3:6Þ
En½A22ðnÞ�r aG½4g1ðnÞþ3g2ðnÞ�j4
Gða1Þþr2
1ðnÞ
ðnb5Þ2þ
64
c41½nðm�1Þ�2
( )nb5
½jð1Þ� �2, ð3:7Þ
En½A23ðnÞ�r
1
nb5½jð1Þ� �2
, ð3:8Þ
En½A24ðnÞ�r
1
nb5½jð1Þ� �2
, ð3:9Þ
En½A25ðnÞ�r a2½4g1ðnÞþ3g2ðnÞ�j4
Gða2Þþr2
1ðnÞ
ðnb5Þ2þ
64
c41½nðm�1Þ�2
( )nb5
½jð1Þ� �2, ð3:10Þ
En½A26ðnÞ�rc2
n ½4g1ðnÞþ3g2ðnÞþg3ðnÞþg4ðnÞþg5ðnÞ�: ð3:11Þ
Here c1,jð1Þ� are positive constants, r1(n), giðnÞ,i¼ 1,2,3,4,5 are functions of n defined in the Appendix. From the Appendix,
we obtain giðnÞ ¼ oðn�2Þ,i¼ 1,2,3,4,5 and r1(n) is a bounded function of n. From (3.1)–(3.11), we conclude that
RðG,d�nÞ�RðG,dGÞ ¼O1
nb5
� �¼O
ðlnnÞ2:5
n
!: &
Table 1
The regrets of empirical Bayes test procedure d�n for a2 ¼ 2 and 4 with various values of s2 and c20 ¼ 0:1þs2=3.
n a2 ¼ 2 4
s2 =0.1 0.2 0.1 0.2
R RR R RR R RR R RR
11 0.002607 (0.000190) 44.6694 0.002385 (0.000183) 19.1452 0.001853 (0.000162) 43.7302 0.001961 (0.000166) 21.4542
21 0.001657 (0.000141) 28.3941 0.001598 (0.000139) 12.8256 0.001475 (0.000137) 34.8216 0.001291 (0.000130) 14.1246
31 0.001360 (0.000126) 23.3008 0.001297 (0.000124) 10.4133 0.000879 (0.000101) 20.7361 0.001066 (0.000110) 11.6635
41 0.000930 (0.000099) 15.9375 0.000962 (0.000100) 7.7243 0.000664 (0.000082) 15.6756 0.000608 (0.000081) 6.6522
51 0.000722 (0.000087) 12.3671 0.000624 (0.000078) 5.0061 0.000661 (0.000086) 15.6065 0.000664 (0.000087) 7.2680
61 0.000592 (0.000073) 10.1417 0.000681 (0.000080) 5.4695 0.000513 (0.000070) 12.1059 0.000631 (0.000081) 6.9059
71 0.000520 (0.000068) 8.9157 0.000625 (0.000078) 5.0173 0.000438 (0.000065) 10.3392 0.000459 (0.000069) 5.0215
81 0.000522 (0.000069) 8.9456 0.000458 (0.000065) 3.6725 0.000552 (0.000074) 13.0237 0.000397 (0.000060) 4.3401
91 0.000335 (0.000056) 5.7359 0.000242 (0.000046) 1.9438 0.000243 (0.000045) 5.7366 0.000304 (0.000052) 3.3253
101 0.000440 (0.000060) 7.5406 0.000381 (0.000057) 3.0601 0.000368 (0.000056) 8.6854 0.000225 (0.000040) 2.4642
111 0.000373 (0.000052) 6.3965 0.000288 (0.000046) 2.3086 0.000308 (0.000048) 7.2650 0.000248 (0.000042) 2.7109
121 0.000282 (0.000045) 4.8390 0.000197 (0.000040) 1.5774 0.000264 (0.000046) 6.2368 0.000266 (0.000046) 2.9117
131 0.000227 (0.000038) 3.8970 0.000139 (0.000028) 1.1167 0.000286 (0.000047) 6.7609 0.000212 (0.000041) 2.3159
141 0.000235 (0.000038) 4.0302 0.000205 (0.000038) 1.6493 0.000271 (0.000045) 6.4067 0.000146 (0.000031) 1.5958
151 0.000205 (0.000036) 3.5061 0.000115 (0.000027) 0.9194 0.000195 (0.000038) 4.5933 0.000141 (0.000031) 1.5472
161 0.000143 (0.000028) 2.4548 0.000170 (0.000031) 1.3644 0.000176 (0.000034) 4.1536 0.000168 (0.000034) 1.8432
171 0.000156 (0.000030) 2.6746 0.000172 (0.000032) 1.3818 0.000159 (0.000030) 3.7642 0.000148 (0.000033) 1.6220
181 0.000148 (0.000026) 2.5362 0.000111 (0.000022) 0.8900 0.000126 (0.000026) 2.9665 0.000178 (0.000035) 1.9513
191 0.000109 (0.000023) 1.8596 0.000149 (0.000030) 1.1947 0.000102 (0.000027) 2.4055 0.000083 (0.000021) 0.9075
201 0.000085 (0.000020) 1.4620 0.000077 (0.000017) 0.6205 0.000076 (0.000019) 1.7984 0.000119 (0.000030) 1.3004
211 0.000092 (0.000020) 1.5837 0.000143 (0.000028) 1.1500 0.000183 (0.000033) 4.3211 0.000112 (0.000028) 1.2225
221 0.000086 (0.000020) 1.4777 0.000114 (0.000026) 0.9180 0.000097 (0.000024) 2.2936 0.000110 (0.000024) 1.2007
231 0.000094 (0.000022) 1.6140 0.000060 (0.000016) 0.4850 0.000060 (0.000018) 1.4169 0.000067 (0.000017) 0.7377
241 0.000060 (0.000015) 1.0342 0.000130 (0.000027) 1.0442 0.000113 (0.000025) 2.6704 0.000081 (0.000023) 0.8858
251 0.000073 (0.000017) 1.2553 0.000114 (0.000026) 0.9147 0.000060 (0.000015) 1.4070 0.000086 (0.000022) 0.9431
261 0.000074 (0.000016) 1.2762 0.000082 (0.000019) 0.6566 0.000069 (0.000020) 1.6245 0.000083 (0.000021) 0.9076
271 0.000079 (0.000017) 1.3496 0.000082 (0.000020) 0.6600 0.000022 (0.000010) 0.5166 0.000046 (0.000015) 0.5024
281 0.000086 (0.000020) 1.4815 0.000074 (0.000019) 0.5915 0.000059 (0.000018) 1.4029 0.000077 (0.000020) 0.8410
291 0.000047 (0.000015) 0.8060 0.000118 (0.000026) 0.9509 0.000053 (0.000016) 1.2617 0.000069 (0.000018) 0.7571
301 0.000050 (0.000017) 0.8496 0.000120 (0.000026) 0.9615 0.000059 (0.000018) 1.3852 0.000077 (0.000022) 0.8411
Note: Estimated standard deviations of regrets are in parentheses.
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2675
4. Simulation study
In this section, we design a Monte Carlo simulation to investigate the performance of small to moderately large samplesizes for the proposed empirical Bayes test procedure with the conjugate normal prior. For the (n+1)-st component testingproblem, let X(n) denote the historical data, and Xnþ1 ¼ ðXnþ1,1, . . . ,Xnþ1,mÞ denote the present random observations. Fori=1,y,n+1, (Yi,Wi) is defined, based on (Xi1,y,Xim), as Yi ¼
Pmj ¼ 1 Xij=m and Wi ¼
Pmj ¼ 1ðXij�YiÞ
2. From (2.8), (2.10) and(3.1), we define In � I½a�no jYnþ1joaG�þ I½aGo jYnþ1joa�n� and SRðnÞ � jjGðYnþ1Þj � In=2. Then,
E½SRðnÞ� ¼ E
Z aG
a�n
jGðyÞfGðyÞ dy
" #¼ RðG,d�nÞ�RðG,dGÞ,
where dG is the Bayes test based only on the present observation. Obviously, SR(n) is an unbiased estimator of the regretRðG,d�nÞ�RðG,dGÞ.
Let N and m be fixed positive integers, and let N1=N+1. The simulation scheme is described as follows.
(1)
Tabl
The
n
Note
Generate Yi,i¼ 1, . . . ,N1 from the normal distribution Nð0,a2Þ where a240 is the given value.
(2) At the stage i, given Yi ¼ yi, generate a sample of size m, Xi1, . . . ,Xim from the normal distribution Nðyi,s2Þ where s2 isthe given positive value. Let xi1,y,xim denote the observations at the stage i, and yi ¼Pm
j ¼ 1 xij=m and wi ¼Pmj ¼ 1ðxij�yiÞ
2, i=1,y,N1. Here, (y1,y,yn) and (w1,y,wn) denote the past data, and let (yn + 1,wn + 1) denote the presentobservation. Note that from (2.3), one can derive jGðyÞ ¼HGðyÞ=fGðyÞ ¼ c2
0�t2�y2 þðt4y2þ2t2a2y2þt4a2Þ=ðt2þa2Þ2,
and aG is the root of jGðyÞ. Next, do the following computation for each n=10 (10) 300.
(3) Compute Hn(y) in (2.7). (4) Compute annin (2.8).
(5)
Compute in ¼ I½a�no jynþ1joaG�þ I½aGo jynþ1joa�n� and SRðnÞ ¼ jjGðynþ1Þj � in=2.e 2
regrets of empirical Bayes test procedure d�n for a2 ¼ 6 and 8 with various values of s2 and c20 ¼ 0:1þs2=3.
a2 ¼ 6 8
s2 =0.1 0.2 0.1 0.2
R RR R RR R RR R RR
11 0.001388 (0.000143) 39.0594 0.001442 (0.000141) 19.0655 0.001554 (0.000157) 51.9532 0.001332 (0.000139) 20.5441
21 0.001035 (0.000113) 29.1281 0.001165 (0.000127) 15.4013 0.000809 (0.000101) 27.0421 0.000878 (0.000108) 13.5438
31 0.000758 (0.000094) 21.3189 0.000741 (0.000094) 9.7946 0.000730 (0.000095) 24.4152 0.000633 (0.000087) 9.7659
41 0.000636 (0.000085) 17.8794 0.000650 (0.000085) 8.5878 0.000598 (0.000084) 19.9840 0.000719 (0.000093) 11.0869
51 0.000622 (0.000087) 17.4989 0.000687 (0.000091) 9.0872 0.000571 (0.000082) 19.0725 0.000612 (0.000085) 9.4474
61 0.000527 (0.000075) 14.8180 0.000488 (0.000069) 6.4515 0.000368 (0.000059) 12.3021 0.000491 (0.000071) 7.5807
71 0.000517 (0.000074) 14.5381 0.000545 (0.000074) 7.2064 0.000418 (0.000062) 13.9648 0.000381 (0.000061) 5.8750
81 0.000453 (0.000069) 12.7481 0.000557 (0.000077) 7.3631 0.000492 (0.000070) 16.4297 0.000499 (0.000069) 7.7007
91 0.000335 (0.000057) 9.4255 0.000276 (0.000050) 3.6544 0.000301 (0.000052) 10.0459 0.000313 (0.000055) 4.8279
101 0.000380 (0.000058) 10.7027 0.000333 (0.000054) 4.4030 0.000405 (0.000062) 13.5497 0.000302 (0.000053) 4.6583
111 0.000366 (0.000056) 10.2973 0.000199(0.000040) 2.6250 0.000351 (0.000055) 11.7346 0.000254 (0.000047) 3.9172
121 0.000268 (0.000047) 7.5325 0.000251 (0.000045) 3.3158 0.000210 (0.000042) 7.0087 0.000246 (0.000045) 3.7880
131 0.000235 (0.000041) 6.5991 0.000250 (0.000046) 3.3073 0.000235 (0.000042) 7.8501 0.000180 (0.000038) 2.7789
141 0.000249 (0.000047) 7.0135 0.000190 (0.000040) 2.5165 0.000204 (0.000040) 6.8336 0.000206 (0.000042) 3.1828
151 0.000293 (0.000051) 8.2389 0.000130 (0.000031) 1.7185 0.000226 (0.000042) 7.5696 0.000175 (0.000037) 2.7009
161 0.000219 (0.000040) 6.1700 0.000138 (0.000031) 1.8218 0.000160 (0.000034) 5.3560 0.000175 (0.000036) 2.6990
171 0.000176 (0.000034) 4.9501 0.000152 (0.000033) 2.0072 0.000168 (0.000035) 5.6021 0.000147 (0.000036) 2.2719
181 0.000142 (0.000030) 3.9904 0.000136 (0.000031) 1.7917 0.000131 (0.000031) 4.3670 0.000124 (0.000028) 1.9171
191 0.000131 (0.000030) 3.6969 0.000124 (0.000029) 1.6427 0.000148 (0.000034) 4.9620 0.000134 (0.000031) 2.0725
201 0.000108 (0.000029) 3.0475 0.000105 (0.000026) 1.3875 0.000109 (0.000028) 3.6306 0.000101 (0.000027) 1.5590
211 0.000109 (0.000026) 3.0743 0.000115 (0.000027) 1.5168 0.000061 (0.000020) 2.0390 0.000119 (0.000030) 1.8296
221 0.000119 (0.000028) 3.3426 0.000120 (0.000027) 1.5929 0.000087 (0.000025) 2.9235 0.000080 (0.000025) 1.2306
231 0.000066 (0.000020) 1.8625 0.000105 (0.000030) 1.3827 0.000083 (0.000021) 2.7735 0.000056 (0.000020) 0.8708
241 0.000077 (0.000025) 2.1561 0.000065 (0.000018) 0.8632 0.000122 (0.000029) 4.0864 0.000051 (0.000019) 0.7844
251 0.000050 (0.000014) 1.4056 0.000081 (0.000022) 1.0758 0.000076 (0.000021) 2.5369 0.000070 (0.000020) 1.0729
261 0.000051 (0.000017) 1.4468 0.000079 (0.000022) 1.0408 0.000083 (0.000021) 2.7593 0.000056 (0.000017) 0.8661
271 0.000060 (0.000016) 1.6846 0.000055 (0.000018) 0.7270 0.000127 (0.000030) 4.2323 0.000070 (0.000021) 1.0751
281 0.000035 (0.000012) 0.9725 0.000069 (0.000020) 0.9133 0.000085 (0.000024) 2.8466 0.000070 (0.000022) 1.0786
291 0.000099 (0.000023) 2.7938 0.000072 (0.000018) 0.9541 0.000065 (0.000018) 2.1706 0.000052 (0.000019) 0.8014
301 0.000081 (0.000020) 2.2864 0.000082 (0.000023) 1.0811 0.000057 (0.000018) 1.9169 0.000100 (0.000023) 1.5496
: Estimated standard deviations of regrets are in parentheses.
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–26812676
(6)
Repeat step (1) through step (5) 2000 times. Then take its average denoted by Dnþ1, which is used as estimator of thedifference RðG,d�nÞ�RðG,dGÞ.The simulation results of the proposed empirical Bayes test procedure for the given m=3, s2 ¼ 0:1,0:2 andc2
0 ¼ 0:1þs2=m, are shown in Table 1 with a2 ¼ 2,4 and in Table 2 with a2 ¼ 6,8, respectively. The estimated standarddeviations of regrets are also reported in Tables 1 and 2. We find that the regret(R) of empirical Bayes test decreasesoscillately as n increases for both tables. Next, we consider the percentage relative regret(RR) of the empirical Bayes testprocedure d�n, which is defined by RRðd�nÞ ¼ 100½RðG,d�nÞ�RðG,dGÞ�=RðG,dGÞ, to evaluate the proposed testing procedure forsmall to moderately large sample sizes. Here the minimum Bayes risk RðG,dGÞ can be obtained from (1.5), and we interpretRR as the percentage regret reduction of d�n in terms of the minimum Bayes risk. The smaller the value of RR is, the betterthe performance of d�n is. These RR’s are reported in Tables 1 and 2. We find that these RR’s decrease oscillately as n
increases. For the small sample size n=11 at different values of a in Table 1, we observe that the RR’s are around 44 percentwhen s2 ¼ 0:1 and around 20 percent when s2 ¼ 0:2. Further, we find d�n always has better performance with largervariance under the same sample size. This phenomenon also occurs in Table 2. For s2 ¼ 0:2 in both tables, the RR’s arearound 20 percent when n=11 and around 10 percent when n=31. Thus, the risk of d�n is relatively close to the minimumBayes risk even for the sample size n=31. Overall, the proposed empirical Bayes test procedure has good performance forsmall to moderately large sample sizes in the simulation study.
Acknowledgments
We would like to thank two referees for the careful review and valuable suggestions which lead to the improvement of thepresentation of the paper. This research was supported by the project NSC 98-2118-M-130-001 of the National Science Councilof ROC.
Appendix A
In order to investigate the asymptotic behavior of En[A2i (n)], we have to study certain properties pertaining to the
estimator Hn(y). From (2.6)–(2.7), Hn(y) can be expressed as
HnðyÞ ¼1
n
Xn
j ¼ 1
Vðy,Yj,tn,bÞ, ðA:1Þ
where
Vðy,Yj,tn,bÞ ¼ðc2
0�y2�t2nÞ
bsnK
y�Yj
bsn
� ��
2yt2n
ðbsnÞ2
K ð1Þy�Yj
bsn
� ��
t4n
ðbsnÞ3
K ð2Þy�Yj
bsn
� �:
Conditioning on tn,Vðy,Yj,tn,bÞ are iid. Since jKðlÞðyÞjr1 for all y and l=0,1,2,
jVðy,Yj,tn,bÞjrðbsnÞ
2ðc2
0þy2þt2nÞþ2b
ffiffiffiffiffimpjyjt3
nþt4n
ðbsnÞ3
¼pðy,tn,bÞ
2ðbsnÞ3
,
where pðy,tn,bÞ ¼ 2fðbsnÞ2½c2
0þy2þt2n�þ2b
ffiffiffiffiffimpjyjt3
nþt4ng. Thus,
jVðy,Yj,tn,bÞ�En½Vðy,Yj,tn,bÞjtn�jrpðy,tn,bÞ
ðbtnÞ3: ðA:2Þ
Let
HnðyjtÞ ¼ ðc20�y2�t2ÞfnðyÞ�2yt2f ð1Þn ðyÞ�t
4f ð2Þn ðyÞ,
p1ðy,tnÞ ¼ ½fnðyÞ�fGðyÞ�þ2y½f ð1Þn ðyÞ�f ð1ÞG ðyÞ�þ½t2nþt
2�½f ð2Þn ðyÞ�f ð2ÞG ðyÞ�,
p2ðy,tnÞ ¼ fGðyÞþ2yf ð1ÞG ðyÞþ½t2nþt
2�f ð2ÞG ðyÞ: ðA:3Þ
Thus,
HnðyÞ�HGðyÞ ¼ fHnðyÞ�En½HnðyjtÞjtn�gþEn½HnðyjtÞ�HGðyÞjtn��ðt2n�t
2ÞEn½p1ðy,tnÞjtn��ðt2n�t
2Þp2ðy,tnÞ: ðA:4Þ
The following Lemmas A.1 and A.2 can be obtained through straightforward computation and the details are omitted.
Lemma A1.
(a)
jEn½fnðyÞ�fGðyÞjtn�jrbsn=ðpt2Þexpð�2s2lnn=s2nÞ.(b)
jEn½fð1Þn ðyÞ�f ð1ÞG ðyÞjtn�jrexpð�2s2lnn=s2nÞ=ðpt2Þ.
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2677
(c)
(a)
(b)
(c)
(d)
(a)
(b)
jEn½fð2Þn ðyÞ�f ð2ÞG ðyÞjtn�jr
1
pt2exp
�2s2lnn
s2n
� �1
bsnþ
bsn
t2
� �:
(d)
jEn½HnðyjtÞ�HGðyÞjtn�jrB1ðy,sn,bÞ, whereB1ðy,sn,bÞ ¼ exp�2s2lnn
s2n
� �1
pt2bsnðc
20þy2þt2
nÞþ2jyjt2nþt
4n
� �:
(e)
jEn½p1ðy,tnÞjtn�jrB2ðy,sn,bÞ, whereB2ðy,sn,bÞ ¼1
pt2exp
�2s2lnn
s2n
� �bsnþ2jyjþðt2
nþt2Þ
1
bsnþ
bsn
t2
� �� �:
(f)
jp2ðy,tnÞjr1
t þ2jyj½jyjþm1ðGÞ�
t3þðt2
nþt2Þ
t2þ2y2þ2m2ðGÞ
t5
� �� B3ðy,snÞ,
where miðGÞ ¼Rjyji dGðyÞ, i¼ 1,2.
Lemma A2.
Var Ky�Yj
bsn
� �tn
� �rk0bsn=t:
Var K ð1Þy�Yj
bsn
� �tn
� �rk0bsn=t:
Var K ð2Þy�Yj
bsn
� �tn
� �rk0bsn=t:
Var Vðy,Yj,tn,bjtnÞ �
rQ ðy,sn,bÞ
ðbsnÞ5
:
where Q ðy,sn,bÞ ¼ ½3ðc20þy2þt2
nÞ2k0ðbsnÞ
4þ12y2t4
nk0ðbsnÞ2þ3t8
nk0�=t.
Lemma A3. Let gG ¼R
expð�y2=ð2t2ÞÞdGðyÞ. Suppose that G is symmetric about the point 0. Then,
(a)
fGðyÞZgGf ðyj0,t2Þ for all y. ffiffiffiffiffiffip ffiffiffiffiffiffiffiffip (b) For sufficiently large n, fGðyÞZgG=ð 2ptÞexpð� lnn=ð8s2ÞÞ � qðnÞ for jyjrcn.Proof. Since G is symmetric about 0, thus, fGðyÞ=f ðyj0,t2Þ ¼R
expðyy=t2Þexpð�y2=ð2t2ÞÞ dGðyÞ is strictly increasing in y for
yZ0. So, fGðyÞ=f ðyj0,t2ÞZ fGð0Þ=f ð0j0,t2Þ ¼R
expð�y2=ð2t2ÞÞ dGðyÞ ¼ gG. Hence, fGðyÞZgGf ðyj0,t2Þ. Part (b) follows directly
from the definition of the point cn and the function f ðyj0,t2Þ. &
Lemma A4. mHG¼R1
0 jHGðyÞj dyo1, if m2ðGÞ ¼Ry2 dGðyÞo1.
Proof.Z 10jHGðyÞj dy¼
Z 10
Zðc2
0�y2Þf ðyjy,t2Þ dGðyÞ
dyrc2
0
ZY
Z 10
f ðyjy,t2Þ dy dGðyÞþZYy2Z 1
0f ðyjy,t2Þ dy dGðyÞ
rc20þ
Zy2 dGðyÞ ¼ c2
0þm2ðGÞo1: &
The following lemma is from Liang (1997).
Lemma A5.
Pt2
n
t2�14c
� �rexp �
nðm�1Þ
2½c�lnð1þcÞ�
� for c40:
Pt2
n
t2�1oc
� � ¼ 0 if cr�1,
rexp �nðm�1Þ
2½c�lnð1þcÞ�
� for �1oco0:
8><>:
Let jð1Þ� ¼ inff½�jð1ÞG ðyÞ�; a1ryra2g. Since jð1ÞG ðyÞo0 and is continuous for all y40, thus, jð1Þ� 40.
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–26812678
Lemma A6.
(a)
(b)
0oaG�aG1ðnÞr1
jð1Þ�ffiffiffiffiffiffiffiffinb5
p :
0oaG2ðnÞ�aGr1
jð1Þ�ffiffiffiffiffiffiffiffinb5
p :
Proof. We provide proof for part (a) only. Note that jGðyÞ is decreasing in y for y40 and jGðaGÞ ¼ 0. By the definition of
aG1(n) and mean-value theorem, �1=ffiffiffiffiffiffiffiffinb5
p¼jGðaGÞ�jGðaG1ðnÞÞ ¼ ðaG�aG1ðnÞÞjð1ÞG ða
�Þ for some immediate value an
between aG1(n) and aG. Thus, 0oaG�aG1ðnÞ ¼�1=ðjð1ÞG ða�Þ
ffiffiffiffiffiffiffiffinb5
pÞr1=ðjð1Þ�
ffiffiffiffiffiffiffiffinb5
pÞ. &
Lemma A7. For aG2ðnÞryrcn,
jHGðyÞjZrðnÞ �f�ffiffiffiffiffiffiffiffinb5
p ,gGffiffiffiffiffiffi2pp
texp �
ffiffiffiffiffiffiffiffilnnp
8s2
!jjGða2Þj
!Z
1ffiffiffinp for sufficiently large n:
where f� ¼minffGðyÞj0ryra2g.
Proof. Note that for aG2ðnÞryrcn,jGðyÞo0, and is strictly decreasing in y. Thus as aG2ðnÞryra2, jHGðyÞj ¼ fGðyÞjjGðyÞj
Z f�jjðaG2ðnÞj ¼ f�=ffiffiffiffiffiffiffiffinb5
p. For a2ryrcn, by Lemma A3, jHGðyÞjZ fGðyÞjjGða2ÞjZgG=ð
ffiffiffiffiffiffi2pp
tÞexpð�ffiffiffiffiffiffiffiffilnnp
=ð8s2ÞÞjjGða2Þj.
Hence, jHGðyÞjZrðnÞZ1=ffiffiffinp
for sufficiently large n. &
For aG2ðnÞryrcn by (A.4),
PfHnðyÞ40g ¼ PfHnðyÞ�HGðyÞ4�HGðyÞgrP HnðyÞ�En½HnðyÞjtn�4�14HGðyÞ
� �þP En½HnðyjtÞ�HGðyÞjtn�4�1
4HGðyÞ� �
þP ðt2n�t
2ÞEn½p1ðy,tnÞjtn�o14HGðyÞ
� �þP ðt2
n�t2Þp2ðy,tnÞo1
4HGðyÞ� �
� Iðy,nÞþ IIðy,nÞþ IIIðy,nÞþ IVðy,nÞ, ðA:5Þ
Iðy,nÞrPfs2n42s2gþP s2
no12s
2� �
þP HnðyÞ�En½HnðyÞjtn�4�14 HGðyÞ,
12s
2rs2nr2s2
� �� I1ðy,nÞþ I2ðy,nÞþ I3ðy,nÞ:
By Lemma A5, we have
I1ðy,nÞ ¼ Ps2
n
s2�141
� rexp �
nðm�1Þ
2ð1�ln2Þ
� � g1ðnÞ, ðA:6Þ
I2ðy,nÞ ¼ Ps2
n
s2�1o�
1
2
� rexp �
nðm�1Þ
2ðln2�1=2Þ
� � g2ðnÞ, ðA:7Þ
and
I3ðy,nÞ ¼ En P HnðyÞ�En½HnðyÞjtn�4�14HGðyÞjtn
� �I 1
2s2rs2
nr2s2 �� �
:
By (A.2), Lemma A2(d) and Bernstein inequality,
P HnðyÞ�En½HnðyÞjtn�4�14HGðyÞjtn
� �
rexp �n½HGðyÞ=4�2=2
VarðVðy,Yi,tn,bÞjtnÞþpðy,tn,bÞ
3ðbsnÞ3�jHGðyÞj
4
8>>><>>>:
9>>>=>>>;rexp �
3nðbsnÞ5f 2
G ðyÞj2GðyÞ
8½12Q ðy,sn,bÞþb2s2npðy,tn,bÞfGðyÞjjGðyÞj�
( ):
Note that both Q ðy,sn,bÞ and pðy,tn,bÞ are increasing in sn. Therefore,
I3ðy,nÞrexp�3nb5s5f 2
G ðyÞj2GðyÞ
32ffiffiffi2p½12Q ðy,
ffiffiffi2p
s,bÞþ2b2s2pðy,ffiffiffi2p
t,bÞfGðyÞjjGðyÞj�
( )� I31ðy,nÞ: ðA:8Þ
Combining the preceding results yields
Iðy,nÞrg1ðnÞþg2ðnÞþ I31ðy,nÞ: ðA:9Þ
In (A.5), we have
IIðy,nÞ ¼ P En½HnðyjtÞ�HGðyÞjtn�4�14HGðyÞ
� �rPfs2
n42s2gþP En½HnðyjtÞ�HGðyÞjtn�4�14HGðyÞ,s2
n r2s2� �
� II1ðy,nÞþ II2ðy,nÞ,
ðA:10Þ
where II1ðy,nÞrg1ðnÞ.
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2679
For aG2ðnÞryrcn, by Lemma A7, j 14 HGðyÞjZ1=ffiffiffinp
for sufficiently large n. On the event that s2n r2s2, by Lemma A1(d),
jEn½HnðyjtÞ�HGðyÞjtn�jr1
npt2½ffiffiffi2p
bsðc20þc2
nþ2t2Þþ4cns2þ4s4�o1ffiffiffinp :
Thus, II2ðy,nÞ ¼ 0 for sufficiently large n.Combining the preceding results leads to
IIðy,nÞrg1ðnÞ for sufficiently large n: ðA:11Þ
From Lemma A1(e),
IIIðy,nÞrP jt2n�t
2jB2ðy,sn,bÞ4�14HGðyÞ
� �rPft2
n42t2gþPft2no
12t
2g
þP ðt2n�t
2ÞB2ðy,sn,bÞ4�14HGðyÞ,t2rt2
nr2t2� �
þP ðt2n�t
2ÞB2ðy,sn,bÞo14 HGðyÞ,
12t
2rt2nrt2
� �� III1ðy,nÞþ III2ðy,nÞþ III3ðy,nÞþ III4ðy,nÞ,
where III1ðy,nÞrg1ðnÞ,III2ðy,nÞrg2ðnÞ, and by a discussion analogous to the term II2(y,n), we have, for sufficiently large n,III3(y,n)=0 and III4(y,n)=0. Therefore,
IIIðy,nÞrg1ðnÞþg2ðnÞ for sufficiently large n: ðA:12Þ
From Lemma A1(f),
IVðy,nÞrP jt2n�t
2jB3ðy,snÞ4�14HGðyÞ
� �rPft2
n42t2gþPft2no
12t
2gþP ðt2n�t
2ÞB3ðy,snÞ4�14HGðyÞ,t2rt2
nr2t2� �
þP ðt2n�t
2ÞB3ðy,snÞo14 HGðyÞ,
12t
2rt2nrt2
� �� IV1ðy,nÞþ IV2ðy,nÞþ IV3ðy,nÞþ IV4ðy,nÞ,
where IV1ðy,nÞrg1ðnÞ and IV2ðy,nÞrg2ðnÞ. Since B3ðy,snÞ is increasing in jyj and sn, thus,
IV3ðy,nÞ ¼ P t2n�t
24�HGðyÞ
4B3ðy,snÞ,t2rt2
nr2t2
� rP
t2n
t2�14
�HGðyÞ
4t2B3ðy,ffiffiffi2p
sÞ,0r
t2n
t2�1r1
( ),
which is equal to 0 if �HGðyÞ=ð4t2B3ðy,ffiffiffi2p
sÞÞ41. Without loss of generality, we may assume that0o�HGðyÞ=ð4t2B3ðy,
ffiffiffi2p
sÞÞr1. Thus, by the preceding inequality and Lemma A5,
IV3ðy,nÞrexp �nðm�1Þ
2
�HGðyÞ
4t2B3ðy,ffiffiffi2p
sÞln �1
HGðyÞ
4t2B3ðy,ffiffiffi2p sÞ
!" #( )� IV31ðy,nÞ,
IV4ðy,nÞrPt2
n
t2�1o
HGðyÞ
4t2B3ðy,ffiffiffi2p
sÞ,�1
2ot2
n
t2�1o0
( ),
which is equal to 0, if HGðyÞ=ð4t2B3ðy,ffiffiffi2p
sÞÞo� 12. Without loss of generality, we may assume that
� 12 rHGðyÞ=ð4t2B3ðy,
ffiffiffi2p
sÞÞo0. Hence, by Lemma A5 and the preceding inequality,
IV4ðy,nÞrexp �nðm�1Þ
2
HGðyÞ
4t2B3ðy,ffiffiffi2p
sÞ�ln 1þ
HGðyÞ
4t2B3ðy,ffiffiffi2p
sÞ
!" #( )� IV41ðy,nÞ:
Combining the preceding results, we obtain
IV4ðy,nÞrg1ðnÞþg2ðnÞþ IV31ðy,nÞþ IV41ðy,nÞ: ðA:13Þ
Now, combining the results of (A.5), (A.9)–(A.14), we obtain for sufficiently large n,
PfHnðyÞ40gr4g1ðnÞþ3g2ðnÞþ I31ðy,nÞþ IV31ðy,nÞþ IV41ðy,nÞ: ðA:14Þ
Proof of (3.11). A6ðnÞ ¼R cn
a2I½HnðyÞ� dyrcn. By using (A.14), we obtain
En½A26ðnÞ�rcn
Z cn
a2
PfHnðyÞ40g dyrc2n ½4g1ðnÞþ3g2ðnÞ�þcn
Z cn
a2
½I31ðy,nÞþ IV31ðy,nÞþ IV41ðy,nÞ� dy: ðA:15Þ
For a2ryrcn, j2GðyÞZj2
Gða2ÞZj2Gða1,a2Þ, where jGða1,a2Þ ¼minfjjGðaiÞj; i¼ 1,2g, f 2
G ðyÞZg2G=ð2pt2Þexpð�
ffiffiffiffiffiffiffiffilnnp
=ð4s2ÞÞ,
Q ðy,ffiffiffi2p
s,bÞrQ ðcn,ffiffiffi2p
s,bÞ, pðy,ffiffiffi2p
s,bÞrpðcn,ffiffiffi2p
s,bÞ. Substituting these inequalities into I31ðy,nÞ, We obtain
I31ðy,nÞrexp �
3nb5s5g2G=ð2pt2Þexp �
ffiffiffiffiffiffiffiffilnnp
4s2
!j2
Gða1,a2Þ
32ffiffiffi2p½12Q ðcn,
ffiffiffi2p
s,bÞþ2b2s2pðcn,ffiffiffi2p
t,bÞjGða1,a2Þ�
8>>>><>>>>:
9>>>>=>>>>;� g3ðnÞ:
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–26812680
Note that pðcn,ffiffiffi2p
t,bÞ-8t4 and Q ðcn,ffiffiffi2p
s,bÞ-48k0t7 as n-1. Thus, g3ðnÞ ¼ oðn�2Þ. Therefore, we have
cn
Z cn
an
I31ðy,nÞ dyrg3ðnÞc2n ¼ oðn�1Þ: ðA:16Þ
From Lemma A7 �HGðyÞZqðnÞjjGða2Þj for a2ryrcn. Also, B3ðy,ffiffiffi2p
sÞ is increasing in jyj. Thus, on a2ryrcn, �HGðyÞ=
ð4t2B3ðy,ffiffiffi2p
sÞÞZ�qðnÞjGða2Þ=ð4t2B3ðcn,ffiffiffi2p
sÞÞ. Note that qðnÞ=B3ðcn,ffiffiffi2p
sÞ � qðnÞ=c2n � ð1=
ffiffiffiffiffiffiffilnnp
Þexpð�ffiffiffiffiffiffiffilnnp
=ð8s2ÞÞZn�1=4 for
sufficiently large n. For t40, expð�nðm�1Þ=2½t�lnð1þtÞ�Þ is decreasing in t. Thus,
IV31ðy,nÞrexp �nðm�1Þ
2
�qðnÞjGða2Þ
4t2B3ðcn,ffiffiffi2p
s�ln 1�
qðnÞjGða2Þ
4t2B3ðcn,ffiffiffi2p
sÞ
!" #( )� g4ðnÞ,
where g4ðnÞ ¼ oðn�2Þ. Hence,
cn
Z cn
an
IV31ðy,nÞ dyrc2ng4ðnÞ ¼ oðn�1Þ: ðA:17Þ
Also note that expð�nðm�1Þ=2½t�lnð1þtÞ�Þ is increasing in t for �1oto0. Thus,
IV41ðy,nÞrexp �nðm�1Þ
2
qðnÞjGða2Þ
4t2B3ðcn,ffiffiffi2p
sÞ�ln 1þ
qðnÞjGða2Þ
4t2B3ðcn,ffiffiffi2p
sÞ
!" #( )� g5ðnÞ,
where g5ðnÞ ¼ oðn�2Þ. Thus,
cn
Z cn
an
IV41ðy,nÞ dyrc2ng5ðnÞ ¼ oðn�1Þ: ðA:18Þ
Combining (A.15)–(A.18) yields En½A26ðnÞ�rc2
n ½4g1ðnÞþ3g2ðnÞþg3ðnÞþg4ðnÞþg5ðnÞ�. &
Proof of (3.10). By Holder inequality,
A5ðnÞrZ a2
aG2ðnÞ
I½HnðyÞ�
j3GðyÞj
ð1ÞG ðyÞ
dy
" #1=2
�
Z a2
aG2ðnÞI½HnðyÞ�j3
GðyÞjð1ÞG ðyÞ dy
� �1=2
,
where
Z a2
aG2ðnÞ
I½HnðyÞ�
j3GðyÞj
ð1ÞG ðyÞ
dyrZ a2
aG2ðnÞ
jð1ÞG ðyÞ
j3GðyÞ½j
ð1ÞG ðyÞ�
2dyr
1
½jð1Þ�G �2
Z a2
aG2ðnÞ
jð1ÞG ðyÞ
j3GðyÞ
dyr1
½jð1Þ�G �2j2
GðaG2ðnÞÞ¼
nb5
½jð1Þ�G �2:
Hence,
En½A25ðnÞ�r
nb5
½jð1Þ�G �2
Z a2
aG2ðnÞPfHnðyÞ40gj3
GðyÞjð1ÞG ðyÞ dy, ðA:19Þ
where by (A.14),Z a2
aG2ðnÞPfHnðyÞ40gj3
GðyÞ djGðyÞra2½4g1ðnÞþ3g2ðnÞ�j4Gða2Þþ
Z a2
aG2ðnÞ½I31ðy,nÞþ IV31ðy,nÞþ IV41ðy,nÞ�j3
GðyÞ djGðyÞ: ðA:20Þ
Let j�Gða2Þ ¼maxðjjGða2Þj,jGð0ÞÞ,
r1ðnÞ ¼32
ffiffiffi2p½12Q ða2,
ffiffiffi2p
s,bÞþ2b2s2pða2,ffiffiffi2p
t,bÞj�Gða2Þ=t�3s5f 2
�
:
Note that r1ðnÞ-384ffiffiffi2p
k0t7=ðs5f 2� Þ as n-1 and pða2,
ffiffiffi2p
t,bÞ-8t8 as n-1. Hence, on aG2ðnÞryra2,
I31ðyÞrexp �3nb5s5f 2
� j2GðyÞ
32ffiffiffi2p½12Q ða2,
ffiffiffi2p
s,bÞþ2b2s2pða2,ffiffiffi2p
t,bÞj�Gða2Þ�
( )¼ expf�nb5j2
GðyÞ=r1ðnÞg:
Thus,
Z a2
aG2ðnÞI31ðy,nÞj3
GðyÞ djGðyÞrZ a2
aG2ðnÞexpf�nb5j2
GðyÞ=r1ðnÞgj3GðyÞ djGðyÞr
Z j2Gða2Þ
j2GðaG2ðnÞÞ
expð�nb5t=r1ðnÞÞt dtrr1ðnÞ
nb5
� �2
:
ðA:21Þ
L.-S. Chen, M.-C. Yang / Journal of Statistical Planning and Inference 141 (2011) 2670–2681 2681
For aG2ðnÞryra2, �HGðyÞ=ð4t2B3ðy,ffiffiffi2p
sÞÞZ�f�jGðyÞ=ð4t2B3ða2,ffiffiffi2p
sÞÞ ��c1jGðyÞ where c1 ¼�f�=ð4t2B3ða2,ffiffiffi2p
sÞÞ. Since
expð�nðm�1Þ=2½t�lnð1þtÞ�Þ is decreasing in t for t40,
IV31ðy,nÞrexp �nðm�1Þ
2½�c1jGðyÞ�ln½1�c1jGðyÞ��
� :
Also, for 0oto1, t�lnð1þtÞZt2=4. Thus,Z a2
aG2ðnÞIV31ðy,nÞj3
GðyÞ djGðyÞrZ a2
aG2ðnÞexp �
nðm�1Þ
2½�c1jGðyÞ�ln½1�c1jGðyÞ��
� j3
GðyÞ djGðyÞ
¼1
c41
Z �c1jGða2Þ
�c1jGðaG2ðnÞÞexp �
nðm�1Þ
2½t�lnð1þtÞ�
� t3 dtr
1
2c41
Z �c1jGða2Þ
�c1jGðaG2ðnÞÞexp �
nðm�1Þt2
8
� t2 dtr
32
c41½nðm�1Þ�2
:
ðA:22Þ
Also, expf�nðm�1Þ=2½t�lnð1þtÞ�g is increasing in t for �1oto0, and t�lnð1þtÞZt2=4 for �1oto0. Thus, on
aG2ðnÞryra2.
IV41ðy,nÞrexp �nðm�1Þ
2½c1jGðyÞ�ln½1þc1jGðyÞ��
� :
Hence,Z a2
aG2ðnÞIV41ðy,nÞj3
GðyÞ djGðyÞrZ a2
aG2ðnÞexp �
nðm�1Þ
2½c1jGðyÞ�ln½1þc1jGðyÞ��
� j3
GðyÞ djGðyÞ
¼1
c41
Z c1jGða2Þ
c1jGðaG2ðnÞÞexp �
nðm�1Þ
2½t�lnð1þtÞ�
� t3 dtr
1
c41
Z c1jGða2Þ
c1jGðaG2ðnÞÞexp �
nðm�1Þt2
2
� t3 dtr
32
c41 ½nðm�1Þ�2
: ðA:23Þ
Combining (A.19)–(A.20) yields
En½A25ðnÞ�r a2½4g1ðnÞþ3g2ðnÞ�j4
Gða2Þþr1ðnÞ
ðnb5Þ2þ
64
c41 ½nðm�1Þ�2
( )nb5
½jð1Þ�G �2¼O
1
nb5
� �: &
Proof of (3.9).
A4ðnÞ ¼
Z aG2ðnÞ
aG
I½HnðyÞ� dyraG2ðnÞ�aG ¼1
jð1Þ�G
ffiffiffiffiffiffiffiffinb5
p :
Thus,
En½A24ðnÞ�r
1
½jð1Þ�G �2nb5
: &
Proof of (3.6), (3.7), and (3.8). The proofs of (3.6), (3.7), and (3.8) are similar to that of (3.11), (3.10), and (3.9),respectively. The details are thus omitted. &
References
Chen, L.-S., 2009. On empirical Bayes two-tail tests for double exponential distributions. J. Nonparametr. Statist. 21, 1037–1049.Gupta, S.S., Li, J., 2006. Optimal rate of convergence of monotone empirical Bayes tests for a normal mean. J. Statist. Plann. Inference 136, 2352–2366.Johns, M.V., Van Ryzin, J., 1972. Convergence rates for empirical Bayes two-action problem. II. Continuous case. Ann. Math. Statist. 43, 934–947.Li, J., Gupta, S.S., 2002. Empirical Bayes tests based on kernel sequence estimation. Statist. Sinica 12, 1061–1072.Liang, T., 1999. Empirical Bayes two-tail tests in a discrete exponential family. Statist. Decisions 17, 157–184.Liang, T., 1997. Simultaneously selecting normal populations close to a control. J. Statist. Plann. Inference 61, 297–316.Liang, T., 2002. An improved empirical Bayes test for positive exponential families. Statist. Neerlandica 56, 346–361.Liang, T., 2006a. Empirical Bayes testing for a normal mean: variance unknown case. J. Statist. Plann. Inference 136, 1376–1393.Liang, T., 2006b. On empirical Bayes two-tail tests in a positive exponential family. J. Nonparametr. Statist. 18, 449–464.Liang, T., 2006c. Simultaneous inspection of variable equivalence for finite populations. J. Statist. Plann. Inference 136, 2112–2128.Pensky, M., 2002. A new approach to empirical Bayes estimation with errors in variables. Statist. Decisions 20, 225–240.Pensky, M., 2003. Rate of convergence of empirical Bayes tests for a normal mean. J. Statist. Plann. Inference 111, 181–196.Robbins, H., 1956. An empirical Bayes approach to statistics. In: Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability,
vol. 1, University of California Press, Berkeley, CA, pp. 157–163.Robbins, H., 1964. An empirical Bayes approach to statistical decision problems. Ann. Math. Statist. 35, 1–20.Singh, R.S., Wei, L., 2000. Nonparametric empirical Bayes procedures, asymptotic optimality and rates of convergence for two-tail tests in exponential
family. J. Nonparametr. Statist. 12, 475–501.van Houwelingen, J.C., 1976. Monotone empirical Bayes test for continuous one-parameter exponential family. Ann. Statist. 4, 981–989.Wei, L., 1989. Empirical Bayes two-sided test problem for continuous one-parameter exponential families. System Sci. Math. Sci. 2, 369–384.Wei, L., 1991. Empirical Bayes two-sided test problem about discrete one-parameter exponential families. Chinese J. Appl. Probab. Statist. 7, 299–310.