UNCORRECTED PROOF - University at Buffalo

Journal Code:
S
J
O
S
Article ID
1
2
0
7
9
Dispatch 30.01.14
No. of Pages: 18
CE: Patrimonio, Cheryl F.
ME:
Scandinavian Journal of Statistics
doi: 10.1111/sjos.12079
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics. Published by Wiley Publishing Ltd.
ALBERT VEXLER
PR
OO
F
Computing Critical Values of Exact Tests
by Incorporating Monte Carlo Simulations
Combined with Statistical Tables
Department of Biostatistics, The University at Buffalo, State University of New York
YOUNG MIN KIM
Department of Biostatistics, The University at Buffalo, State University of New York
JIHNHEE YU
Department of Biostatistics, The University at Buffalo, State University of New York
NICOLE A. LAZAR
Department of Statistics, University of Georgia
ED
ALAN D. HUTSON
Department of Biostatistics, The University at Buffalo, State University of New York
CO
R
RE
CT
ABSTRACT. Various exact tests for statistical inference are available for powerful and accurate decision rules provided that corresponding critical values are tabulated or evaluated via
Monte Carlo methods. This article introduces a novel hybrid method for computing p-values of
exact tests by combining Monte Carlo simulations and statistical tables generated a priori. To
use the data from Monte Carlo generations and tabulated critical values jointly, we employ kernel density estimation within Bayesian-type procedures. The p-values are linked to the posterior
means of quantiles. In this framework, we present relevant information from the Monte Carlo
experiments via likelihood-type functions, whereas tabulated critical values are used to reflect
prior distributions. The local maximum likelihood technique is employed to compute functional
forms of prior distributions from statistical tables. Empirical likelihood functions are proposed
to replace parametric likelihood functions within the structure of the posterior mean calculations
to provide a Bayesian-type procedure with a distribution-free set of assumptions. We derive the
asymptotic properties of the proposed nonparametric posterior means of quantiles process. Using
the theoretical propositions, we calculate the minimum number of needed Monte Carlo resamples
for desired level of accuracy on the basis of distances between actual data characteristics (e.g.
sample sizes) and characteristics of data used to present corresponding critical values in a table.
The proposed approach makes practical applications of exact tests simple and rapid. Implementations of the proposed technique are easily carried out via the recently developed STATA and R
statistical packages.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
Key words: Bayesian point estimation, empirical likelihood, exact tests, local maximum
likelihood, posterior expectation, quantile estimation
1. Introduction
Exact tests are well known to be simple, efficient and reliable statistical tools in a variety
of applications. The statistical literature has extensively addressed many parametric, semiparametric and nonparametric exact tests that have finite sample type I error control, for
example, exact tests for quantiles (Serfling, 2002, p. 102), Fisher’s exact test, the exact F -test,
the Shapiro–Wilk test of normality, the Wilcoxon rank-sum test, Hall and Welsh’s test for
normality (Hall & Welsh, 1983) and the exact test for comparing multivariate distributions
2
Scand J Statist
CO
R
RE
CT
ED
PR
OO
F
proposed by Rosenbaum (2005), to name just a few. Many of these tests are now found in standard statistical software packages, for example, SPSS. Exact tests are not only useful in the small
sample size setting but also invaluable in the case of rare events in the large sample setting; for
example, see Mudholkar & Hutson (1997) for the case of contingency tables and the inaccuracy
of asymptotic methods when there is a large number of zero count cells. A major disadvantage is
that, in general, exact tests are computationally intensive and may not be feasible for moderate
to large sample sizes. In these instances, Monte Carlo (MC) methods may provide accurate and
computationally feasible approximations. Commonly, tables of corresponding critical points
are required for the exact tests’ implementation. Because all possible critical values cannot be
tabulated, two different approaches can be employed in practice to calculate the exact p-value.
One method is related to interpolation/extrapolation of the relevant critical points on the basis
of tabulated values. The use of tables with corresponding critical values is a standard method
applied in various statistical software routines. Employing a table-based methodology for use
within the testing algorithm is quick and efficient. However, the interpolation/ extrapolation
method using table-based methods becomes much less reliable when real data characteristics
(e.g. sample sizes) differ from those used to tabulate the critical values. Alternatively, as a second
approach, one can conduct MC simulations to approximate the exact p-values. The approach
based on MC simulations is simple and accurate but oftentimes computationally intensive.
In this article, we propose and examine a novel Bayesian hybrid method to compute p-values
of exact tests on the basis of both statistical tables of related critical values and MC simulations.
The key feature of the proposed method is that if actual data characteristics (e.g. sample sizes)
are close to those used in the table of critical values, then relatively few MC simulations are
needed to improve the accuracy of p-value calculations. When this is not the case, more MC
simulations are necessary.
Consider the following simple illustration. Table 1 represents a part of ‘percentiles of the
2 distribution’ (Kutner et al., 2005) at 29, 30, 40 and 50 degrees of freedom. Suppose that
we are interested in the percentiles of the 2 distribution with 35 degrees of freedom and level
of significance 0.05. Table 1 does not provide an exact percentile in this case. Interpolation
between the 30-degree-of-freedom and 40-degree-of-freedom entries or an MC procedure based
on randomly generated 2 values can accurately approximate the exact (missing and unknown)
percentile. In another scenario, suppose we are interested in the case with 31 degrees of freedom
and 0.05 level of significance. In this case, we could conduct an MC study with fewer replications than required in the previous case (35 degrees of freedom) because the desired percentile is
nearer to the tabulated value at 30 degrees of freedom. The proposed Bayesian hybrid method
calculates the percentile of interest by incorporating the information from Table 1 as a prior
distribution to improve the accuracy and reduce the computational cost of the simulations.
In this article, we consider a Bayesian solution to the problem through the incorporation of
information from MC simulations and tabulated critical values. Hence we consider the tables to
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
T125
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
Table 1. Percentiles of the 2 distribution (Kutner et al., 2005) where d.f. is degrees
of freedom
Level of significance
d.f.
0.025
0.050
0.100
29
30
40
50
14.26
14.95
22.16
37.48
16.05
16.79
24.43
40.48
17.71
18.49
26.51
43.19
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
Scand J Statist
CO
R
RE
CT
ED
PR
OO
F
represent prior information, whereas the likelihood part of the analysis is based on the output
of the relevant MC simulations. Therefore, the two key issues that need to be addressed for
the proposed hybrid method are as follows: (i) constructing the functional forms of the prior
distribution based on of tabulated critical values; and (ii) developing nonparametric likelihoodtype functions from information extracted from MC simulations.
In this work, we apply the local maximum likelihood (LML) technique (Tibshirani & Hastie,
1987; Fan & Gijbels, 1995, 1996; Fan et al., 1995; Fan et al., 1998) to obtain functional forms
of prior distributions from tables of critical values. The LML methodology is usually applied
to construct functional forms of dependency between variables. For example, in regression,
the aim of the LML approach is to explore the association between dependent and independent variables when the regression functions are data driven instead of being limited to a
certain presumed form (Fan & Gijbels, 1995). The flexibility and relative simplicity of the LML
methodology makes it a practical solution to the first of our problems.
As for the second problem, that is, construction of a likelihood to summarize the data based
on MC generations of exact-test-statistic values, we propose the nonparametric technique of
empirical likelihood (EL; e.g. Owen, 1988, 1990, 2001; Lazar & Mykland, 1998; Yu et al.,
2010; Vexler & Gurevich, 2010; Vexler & Yu, 2011). It is well known that EL is a powerful
nonparametric technique of statistical inference. Lazar (2003) further demonstrated that the EL
functions can be used to construct distribution-free Bayesian posterior probabilities. Following
the results of Lazar (2003), we develop the nonparametric posterior expectations of quantiles
via EL in our proposed procedure for computing the p-values of exact tests.
Evaluation of the exact p-value in our context requires statistical inference for quantiles. We
achieve this inference using the EL technique based on kernel densities. As shown by Chen &
Hall (1993), the use of kernel densities to smooth EL significantly improves the performance of
the EL ratio tests for quantiles, in terms of relative accuracy. Zhou & Jing (2003) proposed an
alternative smoothed EL approach where the likelihood ratio has an explicit form based on Mestimators. Yu et al. (2011) also showed that the use of kernel densities significantly improves
two-sample EL ratio tests for medians.
In order to implement our approach, we need to calculate the posterior expectation given the
a priori information (tables of critical values) and the observed data (MC simulations). In the
case of the classical Bayesian analysis, there are some instances where the integrals related to
the posterior expectations may be evaluated analytically. More often than not, however, these
calculations are intractable, and numerical methods are used to obtain a solution. Alternatively, this article provides an asymptotic expression for the proposed nonparametric posterior
expectations, which utilizes the asymptotic Gaussian form of the posterior distribution of
quantiles based on maximum likelihood estimation. Note that the proposed nonparametric
posterior expectations are based on integrated EL functions. The EL functions themselves do
not have closed analytical forms. Hence, we are required to use numerical methods. That is,
computation of the posterior expectation in this instance is not trivial. In a similar manner
to classical Bayesian inference (e.g. DasGupta, 2008; DiCiccio et al., 1997; Kass & Raftery,
1995), the asymptotic results developed in this article have good utility relative to the posterior expectation calculation, thus making our methodology tractable. The asymptotic approach
also provides a method for calculating the required number of samples needed to generate specific MC simulations in order to obtain accurate critical values for the proposed exact tests.
The criterion for computing the minimum number of MC repetitions is based on the prior
information (tables of pre-populated critical values).
This article is organized as follows. In Section 2, we describe the problem more fully and
define our formal notation. We then define the nonparametric posterior expectation of the
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
3
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
4
Scand J Statist
CT
ED
PR
OO
F
critical values based on EL with kernel functions. The asymptotic properties of the posterior expectation are shown, and the theoretical properties of the estimated critical values are
evaluated asymptotically. Section 2 also presents the method to create functional forms of the
prior distributions using the LML method based on statistical tables. Finally, we provide the
procedure to calculate critical values of exact tests using our new method. Concluding remarks
are presented in Section 3.
Section S1 of the Supporting information demonstrates and examines the applicability of the
proposed approach in practice. Note also that Tanajian et al. (2013) and Shepherd et al. (2013)
have employed the proposed method in the STATA and R software packages, respectively, to
apply new statistical testing strategies in practice. The STATA package provides nonparametric
tests for symmetry of data distributions as well as comparing K-sample distributions. Recognizing the fact that recent statistical software packages do not sufficiently address K-sample
nonparametric comparisons of data distributions, Tanajian et al. (2013) proposed a new STATA
command VX_DBEL to execute exact tests on the basis of K-samples. To calculate p-values
of the exact tests, the STATA command uses the following methods: (i) a classical technique
based on MC p-value evaluations; (ii) an interpolation technique based on tabulated critical
values; and (iii) a new hybrid technique that combines (i) and (ii). The second and third cutting
edge methods are based on the methodology presented in this paper. The R package provides
a function dbELnorm to be used for joint assessment of normality of K-independent samples
with varying means and standard deviations. The function provides the test statistic and associated p-values. The p-values can be calculated by incorporating MC simulations combined
with statistical tables in a manner similar to the technique considered in this paper.
2. Statements and methods
CO
R
RE
In this section, we formally state the problem related to the procedure development in order
to compute critical values of exact tests using MC simulations and statistical tables. We begin
with the following example. Suppose, for simplicity, that we have a two sample exact test
1˛
, where n; m are the respective sample sizes and ˛ is the level of
with critical values qn;m
1˛
significance. Let the values of qn;m
be tabulated for n 2 N , m 2 M and ˛ 2 A, where N
and M are sets of integer numbers and A is a set of real numbers from 0 to 1. That is, we have
® 1˛
¯
the table defined as qn;m
I n 2 N; m 2 M; ˛ 2 A . We are interested in obtaining a value of
®
¯
0
0
1˛
qn1˛
… qn;m
I n 2 N; m 2 M; ˛ 2 A . There are two ways to evaluate qn1˛
: (i) a technique
0 ;m0
0 ;m0
® 1˛
¯
based on interpolation/extrapolation using the table values qn;m I n 2 N; m 2 M; ˛ 2 A ; and
(ii) an approach based on a MC study for generating values of the test statistic. In this paper,
® 1˛
¯
0
we use the table qn;m
I n 2 N; m 2 M; ˛ 2 A to construct a prior distribution for qn1˛
that
0 ;m0
will be combined through Bayes rule with a nonparametric likelihood function based on MC
0
generated values of the test statistic. The value of qn1˛
, the .1 ˛0 / 100% quantile, is then
0 ;m0
estimated from the posterior distribution that results. The next section presents the nonpara0
metric Bayesian approach to estimate qn1˛
using the posterior expectation of quantiles based
0 ;m0
on the smoothed EL method. Section 2.2 introduces the LML technique to derive a prior distribution of quantiles based on related statistical tables. In Section 2.3, the final procedure of
the hybrid method is provided to be applied to calculate critical values of exact tests efficiently.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
2.1. Bayesian empirical likelihood evaluation of quantiles
Smoothed empirical likelihood for quantiles. In this section, we denote the distribution-free
posterior expectation of quantiles based on independent identically distributed observations
X1 ; : : : ; Xt . Let Xj .j D 1; : : : ; t / denote a realization of the test statistic value at the j -th MC
iteration when the MC simulations provide in total t generated values of the test statistic. The
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
Scand J Statist
PR
OO
F
distribution function, F; of X is unknown. We focus on the evaluation of the .1 ˛/ 100%
quantile, Pr X < q01˛ D 1 ˛, ˛ 2 .0; 1/. To provide nonparametric statistical inference
regarding q01˛ , we apply the nonparametric likelihood approach proposed by Chen & Hall
(1993). Towards this end, denote K./ to be a smoothed version of the indicator function F0
defined by F0 .u/ D 1, for u 0, 0 otherwise. Let k./ be a -th order kernel density function, commonly used in nonparametric density estimation (e.g. Silverman, 1986; Härdle, 1990).
Then, the function K./ can be represented in the form
Z
k.u/du;
(1)
K.x/ D
u<x
where k./ satisfies
Z
8
ˆ
< 1 if j D 0;
j
u k.u/du D 0 if 1 j 1;
ˆ
: if j D ;
CT
ED
for any integer 2 and a constant ¤ 0. We use the notation Kh ./ D K.= h/ and
kh ./ D h1 k.= h/ as the kernel distribution and kernel density representations, respectively,
with bandwidth h satisfying the following conditions: h ! 0 and th ! 1 as t ! 1.
Write the smoothed EL function for quantiles as
ˇ
2
3
ˇ
t
t
t
Y
X
X
ˇ
®
¯
EL.q/ D sup 4
pj ˇˇ pj 0;
pj D 1 and
pj Kh q Xj .1 ˛/ D 05:
ˇ
j D1
j D1
j D1
(2)
RE
The method of Lagrange multipliers then can be used to find the values of p1 ; : : : ; pt in (2)
yielding the result:
¯1
®
pj D t C Kh q Xj .1 ˛/
;
(3)
where , the Lagrange multiplier, is a root of
Kh .q Xj /
D 1 ˛:
t C ¹Kh .q Xj / .1 ˛/º
CO
R
t
X
j D1
(4)
Chen & Hall (1993) showed that this smoothed EL for quantiles improves the coverage
accuracy of confidence intervals from order t 1=2 to order t 1 .
Bayesian approach based on the smoothed empirical likelihood for quantiles. We assume that,
in addition to the sample, X1 ; : : : ; Xt , we have prior information regarding the quantile q01˛
in the form of a density function denoted as .q/. Without loss of generality and for simplicity
of notation, we also assume that the form of the density function of X1 is known to be f .xjq/.
Classical Bayesian methodology yields the posterior probability of the quantile q01˛ as
Qt
j D1 f Xj jq .q/
PP .q/ D R 1 Qt
:
(5)
j D1 f .Xj jq/ .q/ dq
1
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
5
Thus, the posterior expectation of a given quantile is
R1
Z1
E.qjX1 ; : : : ; Xt / D
qPp .q/ dq D
1
Qt
j D1 f .Xj jq/ .q/dq
1 q
:
R 1 Qt
j D1 f .Xj jq/ .q/dq
1
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
(6)
6
Scand J Statist
In this work, we replace in (5) with a nonparametric counterpart. Lazar (2003) proposed and
validated the Bayesian EL method as an alternative to the classical parametric setting. This
suggests the nonparametric form of (5)
EL.q/ .q/
PNP .q/ D R X
.t/
X.1/ EL .q/ .q/ dq
(7)
F
;
R X.t/
X
Z .t/
1˛
qO NP
D
X
qPNP .q/ dq D R X.1/
.t/
X.1/
X.1/
qEL.q/ .q/ dq
:
EL.q/ .q/ dq
PR
OO
where X.1/ X.2/ : : : X.t / are the order statistics based on X1 ; : : : ; Xt and the function
EL.q/ is defined at (2), the kernel density-smoothed EL. It follows that the nonparametric
posterior expectation of the (1-˛/th quantile is given by
(8)
RE
CT
ED
Note that if, for example, for some q, we have maxj D1;:::;t ¹Kh .q Xj / .1 ˛/º < 0, then
P
there are no p1 ; : : : ; pt summing to 1 for which tj D1 pj ¹Kh .q Xj / .1 ˛/º D 0. If this
happens, we define EL.q/ D 0 (Owen, 2001), and hence, by convention, and without loss of
generality, we propose the support of the integrals in (7) and (8) to be restricted to ŒX.1/ ; X.t / .
To investigate the asymptotic properties of the nonparametric posterior expectation of
quantiles at (8), we assume that the data distribution function, F ./, is three times differentiable in a neighbourhood of q01˛ and that the applied kernel function is twice
differentiable in a neighbourhood of q01˛ . Let us also define ´.r/ .q/ D d r ´.q/=dq r for
P
a smooth function ´./ and r D 3; 4; : : :, and kNh .q/ D t 1 tj D1 kh .q Xj /, kNh0 .q/ D
P
P
.t h2 /1 tj D1 k 0 .h1 .q Xj // and KN s;h .q/ D t 1 tj D1 ¹Kh .q Xj / .1 ˛/ºs for s D
1; 2; 3. Then we have the following:
Proposition 1. Let X1 ; : : : ; Xt be a random sample from a density
function f ./. Suppose that
q
2 1
2
the prior density is defined as .q/ D exp .q 0 / .20 /
202 . Then, the asymptotic
1˛
qO NP
D
CO
R
approximation to the nonparametric posterior expectation of the (1-˛/-th quantile is given as
1˛
qO M
C
1C
0
1˛
02 Bt2 .q
OM
/t
1
1˛
02 Bt2 .q
OM
/t
1˛
where qO M
satisfies 1 ˛ D t 1
1˛ KN 2;h qO
.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
C op
1
as t ! 1;
t
1˛
1˛ ® 1˛ ¯2 .
2
q
O
and
B
q
O
D
kNh qO M
K
X
h
j
t
j D1
M
M
Pt
M
By proposition 1, one can show that the asymptotic distribution of the estimator (8) is
p
t
Vt1=2
1˛
qO NP
q01˛
! p 1=2 t Vt
0 q01˛ d
1
1˛ 1C 2 2 1˛ ! N.0; 1/ as t ! 1;
0 Bt qO M
02 Bt2 qO M
t
t
1˛ where Vt D tVar qO M
.
The next proposition provides a method to control the accuracy of the proposed pvalues’ evaluations. To formulate the following result, we assume that F .x/ is an absolutely
continuous cumulative distribution function with corresponding density function f .x/ and
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
Scand J Statist
F 00 .x/ D f 0 .x/. We also assume that F .5/ q01˛ D f .4/ q01˛ exists. The application of
the Taylor theorem then implies
(9)
F
1
2 1˛
1˛
1˛
D ˛ qO NP
q01˛ f q01˛ q01˛ f 0 q01˛
qO NP
1 F qO NP
2
3
1 1˛
1˛
00
1˛
qO
q0
f
q0
C ::::
6 NP
PR
OO
1˛
2
1˛
Q
q01˛ f q01˛ C 0:5 qO NP
q01˛
In this case, controlling values of .t/
D qO NP
Q
f 0 q01˛ , we can monitor the accuracy of the p-value evaluations.
Here,
ˇ
ˇ .t/ may be thought
ˇ
ˇ
Q
of as a risk function. For example, the choice of t such that E .t / < (e.g. D 0:001/
provides the presumed accuracy of the q01˛ -estimation. This leads to the following:
Proposition 2. The asymptotic representation of the expectation of the estimated probability at
(9) is
ED
°
±
1˛
Q / C o.t 1 /;
E 1 F qO NP
D ˛ E .t
where
CO
R
RE
CT
´
d2 f 0 q01˛ h2
f q01˛
d4 f .3/ q01˛ h4
Q
1˛ 1˛ E .t/
D 2f q0
24f q0
1 C 2 f˛.1˛/
2 q 1˛ t
. 0 /
0
μ
.1 C 2˛/ f q01˛
.d2 /2 f 0 q01˛ f 00 q01˛ h4
C
C
t
4f 2 q01˛
´
1˛ μ
1˛
1˛
0
f q0
˛.1 ˛/f q0
˛.1 ˛/ 0 q0
1˛ C
2f 3 q0
02 f 2 q01˛ t
t
1 C 2 f˛.1˛/
2 q 1˛ t
. 0 /
0
´
μ
f 0 q01˛
˛ .1 ˛/
;
2
f 2 q01˛ t
2 1 C 2 f˛.1˛/
2 q 1˛ t
. 0 /
0
and d D
R
u k.u/du; D 2; 4:
We outline the proofs of propositions 1 and 2 in the Appendix.
Remark 1. Proposition 1 illustrates that the asymptotic result has a remainder term of
1˛
1˛
order o.t 1 /. By virtue of proposition 1, we can easily show that qO NP
D qO M
C
2 2 1˛ 1
1˛
1
0 qO M
0 Bt qO M
t
C op .t /. This indicates that we may asymptotically detect
effects of the prior on the estimated quantile (8) with relative accuracy, because the term
dependent on the prior information vanishes as t 1 . One can further show that if the kernel
method is not used, the EL function could still be represented through the empirical distribution functions, for example, EL.q/ t.1 ˛ Ft .q//2 =.2˛.1 ˛//, where Ft .q/ D
P
t 1 tj D1 I.Xj q/ for an indicator function, I./. In this case, because of lack of smoothness of the estimator, the technique used to prove proposition 1 cannot be used because Ft
is a step function. Instead, we can apply the classic Bahadur asymptotic results (e.g. Serfling,
2002, p. 93) ; these provide remainder terms calculated to be of order t 3=4 . This is not enough
evidence to assess the asymptotic impact of the prior information in the evaluation of the
distribution-free posterior expectation.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
7
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
8
Scand J Statist
Remark 2. If X is discrete or has defined discontinuities, we can reformulate the propositions
below using the smoothing transformation given by Zj D Xj C "j , where "j N.0; 1/,
j D 1; 2; : : : ; t, and ! 0 as t ! 1.
PR
OO
F
Remark 3. We suggest to choose a bandwidth of h D 0:2t 1=6 . In the context of the quantile
evaluations, this choice of bandwidth provides reasonable performance for various underlying
distributions in our extensive MC study. In addition, the choice of bandwidth satisfies the
asymptotic conditions on required in Chen & Hall (1993) (see also Yu et al., 2011, for details).
2.2. Derivation of the functional form of prior information via the local maximum likelihood
estimation based on statistical tables
® 1˛
¯
In this section, we use the table qn;m
I n 2 N; m 2 M; ˛ 2 A to define a prior distribution on
0
qn1˛
. Consider the model
0 ;m0
1˛
qn;m
D .˛; n; m/ C "1˛
n;m ;
"1˛
n;m N.0; 1/;
(10)
rn X
rm
r˛ X
X
iD0 j D0 kD0
RE
.˛; n; m/ Š
CT
ED
where ./ is an unknown function corresponding to location and > 0 is an unknown scale
parameter. We estimate the function in order to provide the prior location .˛0 ; n0 ; m0 /
0
for qn1˛
, the parameter of interest. In order to estimate .˛0 ; n0 ; m0 / and , we propose
0 ;m0
directly applying the LML methodology (e.g. Fan et al., 1998). The LML methodology is a
nonparametric maximum likelihood technique based on a polynomial-type approximation of
a parameterized function of interest, where the parameterized function is often unknown. The
point of interest in our procedure is the point .˛0 ; n0 ; m0 / for which the corresponding quan® 1˛
¯
tile is not tabulated in the table of critical values qn;m
I n 2 N; m 2 M; ˛ 2 A . In the LML
approach, we approximate the parametric function via Taylor expansion as
ˇijk .˛ ˛0 /i .n n0 /j .m m0 /k ;
ˇijk D ˇikj :
(11)
CO
R
The restriction ˇijk D ˇikj corresponds to a wide class of two sample hypothesis tests, including the applications considered in Section S1 of the Supporting information. This condition
can be avoided in general evaluations. Equation (11) allows an approximation to the location
parameter around the point of interest .˛0 ; n0 ; m0 /. The ultimate parameter of interest is
ˇ000 D .˛0 ; n0 ; m0 / in the polynomial approximation, while incorporating all available data
points into the parameter estimation. Let 0 be the standard deviation corresponding to (the
estimator of) ˇ000 . ˇijk and are estimated through maximizing the log likelihood given by
L.“; ˛0 ; n0 ; m0 / D log
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
p
1
2 2
2 2
X
k.˛ ˛0 ; n n0 ; m m0 /
n2N;m2M;˛2A
X
0
@q 1˛ n;m
n2N;m2M;˛2A
r˛ X
rn X
rm
X
12
i
j
kA
ˇijk .˛ ˛0 / .n n0 / .m m0 /
i D0 j D0 kD0
k.˛ ˛0 ; n n0 ; m m0 /;
(12)
where “ indicates the vector composed of ˇijk and k./ is a joint kernel function. We write
k˛0 ;n0 ;m0 D k.˛ ˛0 ; n n0 ; m m0 / in order to ease the notational burden of this
formulation. The joint kernel function k˛0 ;n0 ;m0 defines the contribution of tabulated data
points to the log-likelihood (12), thus yielding a local kernel weighted log-likelihood function.
The kernel function centred around the values of interest .˛0 ; n0 ; m0 / provides a larger weight
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
Scand J Statist
for the data points closer to the values of interest relative to those points more distant. Without
compromising this concept, we define the joint kernel function as
k˛0 ;n0 ;m0 D kh1 .˛ ˛0 /kh2 .n n0 /kh3 .m m0 /;
1
2
PR
OO
F
where khi ./ D k.= hi /= hi is a univariate density function with khi .x/ D khi .x/ for all real
x and bandwidth hi ; hi > 0. Typically, the ranges of tabulated n and m are similar so that we
let h2 D h3 . The values of r˛ ; rn ; rm and the bandwidth are selected on the basis of the bias
and variance estimates of the nonparametric estimator (Fan et al., 1998). As building blocks
of the bias and variance estimation, we note that @L.“; ˛; n; m/[email protected]“ is a vector whose elements
corresponding to ˇi 0 j 0 k 0 are given as
9
28
r˛ X
rn X
rm
=
<
X
1˛
i
j
k
4 q
ˇijk .˛ ˛0 / .n n0 / .m m0 /
;
: n;m
n2N;m2M;˛2A
i D0 j D0 kD0
3
X
i0
k0 5
j0
k˛0 ;n0 ;m0 ;
ED
.˛ ˛0 / .n n0 / .m m0 /
(13)
ı
and @2 L.“I ˛; n; m/ @“2 is a matrix whose elements are given as
CT
@2 L.“; ˛; n; m/
;
@ˇi 0 j 0 k 0 @ˇi j k with diagonal elements corresponding to ˇi 0 j 0 k 0 , 2
0
(14)
P
0
.˛ ˛0 /2i .n n0 /2j
0
n2N;m2M;˛2A
n2N;m2M;˛2A
RE
.m m0 /2k k˛0 ;n0 ;m0 and off-diagonal elements corresponding to ˇi 0 j 0 k 0 and ˇi j k ,
P
0
0
0
2
.˛ ˛0 /i Ci .n n0 /j Cj .m m0 /k Ck k˛0 ;n0 ;m0 . Analogous to
´
CO
R
the univariate approach of Fan et al. (1998), the bias for the vector of estimators “O can be
estimated by
@2 L.“; ˛; n; m/
μ1
@“2
ˇ
ˇ
@L.“; ˛; n; m/ ˇˇ
ˇ
@“
ˇ
(15)
O
“D“
given values of r˛ , rn and rm . If we set .˛; n; m/ D .˛0 ; n0 ; m0 /, then on the basis of the
results in (13) and (14), the bias defined at (15) becomes 0 for each component. Thus the bias
is not of concern when we obtain the likelihood function estimates at the value of interest,
namely .˛0 ; n0 ; m0 /. This result follows immediately because the Taylor expansion is carried
out around the point .˛0 ; n0 ; m0 / so that the first term of (11) always corresponds to the true
value of ˇ000 regardless of the order of the expansion. Hence there is no bias involved in the
estimation of ˇ000 .
Analogous to Fan et al. (1998), the variance estimators “O are obtained in the form of
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
9
´
O ˛; n; m/ D
var.“;
@2 L.“; ˛; n; m/
@“2
μ1
²
@L .“; ˛; n; m/
var
@“
³´
@2 L .“; ˛; n; m/
@“2
μ1 ˇˇ
ˇ
ˇ
ˇ
ˇ
O
“D“
(16)
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
10
Scand J Statist
In (16), because elements of the vector, var ¹@L.“; ˛; n; m/[email protected]“º corresponding to ˇi 0 j 0 k 0 , have
the form
8
9
r˛ X
rn X
rm
<
=
X
X
1
2
1˛
i
j
k
k˛0 ;n0 ;m0 var qn;m ˇijk .˛ ˛0 / .n n0 / .m m0 /
2
:
;
iD0 j D0 kD0
n2N;m2M;˛2A
.˛ ˛0 /
2j 0
.n n0 /
2k 0
.m m0 /
;
F
2i 0
PR
OO
(17)
the only term not to be 0 is the element for ˇ000 , when evaluated at .˛; n; m/ D .˛0 ; n0 ; m0 /.
Thus, on the basis of (14), (15) and (16), (17) reduces to
ˇ
ˇ
® 1˛
¯ˇ
P 2
k˛0 ;n0 ;m0 var qn;m ˇ000 ˇˇ
n2N;m2M;˛2A
ˇ
var.ˇO000 ; ˛0 ; n0 ; m0 / D
:
(18)
ˇ
!2
ˇ
P
ˇ
ˇ
k˛0 ;n0 ;m0
ˇ
n2N;m2M;˛2A
O
“D“
ED
0
Thus, we have a prior for q01˛ D qn1˛
N 0 ; 02 , where 0 D ˇ000 .˛0 ; n0 ; m0 /
0 ;m0
and 2 D var ˇO000 . The values of r˛ , rn and rm are obtained to minimize 2 D var ˇO000 .
0
0
CT
We employ the optimal bandwidth hi on the basis of minimum variance (Simonoff, 1998, p.
105). In practice, the values of r˛ , rn and rm also can be assumed to be fixed, say, to be 2 or 3.
2.3. The procedure to calculate critical values of exact tests incorporating Monte Carlo
simulations and statistical tables
CO
R
RE
In this section, we provide the algorithm for executing the proposed method in practice. The
procedure is based on the following steps:
(a) Obtain the prior distribution with the parameters 0 ; 02 using the LML method on
the basis of tabulated critical values. This prior has the form defined in Section 2.2.
(b) Generate a learning sample (e.g. defining t D 200/ of the test statistic values under the
corresponding null hypothesis using MC simulations.
(c) Using the learning sample, estimate f q01˛ f 0 q01˛ f 00 q01˛ and f .3/ q01˛ to
Q
present E .t
ˇ / of ˇproposition 2 as a function of t.
Q /ˇ with a presumed threshold (e.g. D 0:001/ to compute an appropri(d) Compare ˇE .t
ˇ
ˇ
Q 0 /ˇ < .
ate value of t0 when ˇE .t
(e) Run the MC simulations to obtain t0 values of the exact test statistic.
(f) Use (8) or its asymptotic form from proposition 1 to obtain the estimator of the quantile
of interest, evaluating the critical values.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
Remark 4. The proposed estimator given at (8) or its asymptotic form from proposition 1 can
be used to directly evaluate the critical values given a fixed number of MC simulations (e.g.
t0 D 10; 000/. In this case, proposition 2 provides the method from which we can measure the
accuracy of the given estimation procedure.
Remark 5. Classical hypothesis testing methodology offers to preselect the significance level of
the test by defining the corresponding critical value of the test statistic. The test statistic is then
compared with the critical value with the possible decisions of rejecting the null hypothesis H0
or not rejecting H0 . An alternative approach to hypothesis testing is based on computing the
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
Scand J Statist
0
(19)
PR
OO
1p
T0 D @qO M
10
11
1
0
A @1 C
A ;
C
1p
1p
t
t
02 Bt2 qO M
02 Bt2 qO M
F
p-value (e.g. Gibbons & Chakraborti, 2011). Generally speaking, the p-value is a value of a
uniformly distributed random variable that can measure the smallest significance level at which
the null hypothesis would be rejected conditionally on the observed data. The procedure we
have developed can be easily modified for evaluating p-values. For example, using proposition
1, we can define p as a root of
CT
ED
where T0 represents a value of the test statistic computed conditional on the observed data.
1p
That is, qO M
can
be computed
as a function of T0 , and then by virtue of the definition1 p D
P
t
1p
Xj , we obtain a value of p that gives an approximation to the true
t 1 j D1 Kh qO M
underlying p-value. Similarly, but slightly more complex, one can use (8) to obtain a pvalue approximation. Several other procedures to examine the p-value, on the basis of the
proposed method, are possible. Algorithms employed to develop the software packages of
Tanajian et al. (2013) and Shepherd et al. (2013), for example, consist of the following steps: (1):
(a), (b) and (c) as mentioned earlier with q01˛ D T0 ; (2) compute the value of p using (19); and
Q /j with a presumed threshold . If jE .t
Q /j , we approx(3) taking ˛ D p, compare jE .t
Q
imate the p-value using p. If jE .t/j > , then we generate more test statistic values under
the corresponding null hypothesis using MC simulations, thus obtaining t D t C d , where, for
example, d D 200, and go to Step (1.c).
3. Concluding remarks
CO
R
RE
In this paper, we defined and examined the nonparametric posterior expectation of quantiles within the general setting of hypothesis testing. The goal of the proposed approach
was to develop a general framework for evaluating the p-values of a variety of exact tests,
incorporating the relevant MC simulations and statistical tables in a Bayesian framework as
observed and prior information, respectively. We used the smoothed EL function (Chen &
Hall, 1993) to encapsulate information from the MC experiments, whereas tabulated critical
values were used to compute prior distributions. The new method improves the schemes used to
calculate critical values of exact tests. We showed two asymptotic propositions to conduct exact
tests using asymptotic critical points and provided the rule to select the appropriate number of
MC simulations needed for a desired degree of accuracy. The asymptotic result we developed is
useful in practice for evaluating critical values of interest. The MC simulation studies demonstrated that the proposed method yields very accurate estimators for the critical values of exact
tests. It should be noted that the proposed method can also be applied to power calculations
(Cohen, 1988). The proposed method is also very flexible in terms of the variety of applications.
For example, Tanajian et al. (2013) employed the main idea of this paper to develop STATA statistical software packages for testing symmetry and for K-sample comparisons. In a subsequent
paper, we plan to address the use of the bootstrap and asymptotic Wilks-type propositions
to achieve accurate estimation of critical values. Further studies are needed to evaluate the
approach in other contexts. We hope that this paper will stimulate future theoretical and applied
research on this topic.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
11
Appendix: The proofs of propositions 1 and 2
R
1˛
To prove proposition 1, we calculate the integral EL.q/.q/dq in the neighbourhood of qO M
1˛
for large t , where qO M
is a maximum of the smoothed EL function, EL.q/. The approach
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
12
A. Vexler et al.
Scand J Statist
UN
CO
R
RE
CT
ED
PR
OO
F
for evaluating the integral is based on the following stages. The Taylor theorem can be applied
01
to represent approximately the function L./ at (A.1), which is the first derivative of EL.q/
02
with respect to the Lagrange multiplier . The obtained Taylor approximation ofL./ at (A.5)
03
can be used in the evaluation of on the basis of the observed data points. This form can
04
yield values of the integrals of interest. Thus, we begin with a lemma about the existence of the
05
maximum value of the smoothed EL function in the convex hull of ¹X1 ; : : : ; Xt º. The lemma
06
shows that the behaviour of the EL is similar to that of parametric likelihoods with respect to
07
increasing/decreasing of the likelihoods and their maximum occurrences.
08
09
1˛
Lemma 1. The smooth EL function, EL.q/, defined at (2) monotonically decreases if q > qO M
,
10
1˛
1˛
and monotonically increases if q < qO M , where the maximum of EL.q/ occurs at q D qO M that
11
1˛
P
satisfies 1 ˛ D t 1 tj D1 Kh qO M
Xj .
12
13
Proof. By virtue of (2) and (3), the derivative of the log EL.q/ is
14
15
t
Y
@
@
1
16
log EL.q/ D
log
@q
@q
t
C
¹K
.q
Xj / .1 ˛/º
h
17
j D1
18
@
t
¹Kh .q Xj / .1 ˛/º C kh .q Xj /
X
@q
19
D
t C ¹Kh .q Xj / .1 ˛/º
20
j D1
21
t
X
kh .q Xj /
22
D
;
t C ¹Kh .q Xj / .1 ˛/º
23
j D1
24
this follows from the fact that is a root of
25
t
X
Kh .q Xj / .1 ˛/
26
D 0:
27
t C ¹Kh .q Xj / .1 ˛/º
j D1
28
Thus, the sign of @ log EL.q/[email protected] depends on the sign of .q/. In order to obtain the sign of
29
.q/, we define the function
30
31
t
X
Kh .q Xj / .1 ˛/
32
L./ :
(A.1)
t C ¹Kh .q Xj / .1 ˛/º
j D1
33
34
1˛
Note that @L./[email protected] < 0, and hence L./ is decreasing with respect to . Because qO M
1˛
Pt
35
1˛
1
satisfies t
O M Xj .1 ˛/ D 0, at the point q D qO M is 0, that is,
j D1 Kh q
1˛ 36
qO M
D 0. In this case, pj , j D 1; 2; : : : ; t , in (8) are just t 1 , and the constraint,
37
Pt
j D1 pj ¹Kh .q Xj / .1 ˛/º D 0, is automatically satisfied.
38
Consider the two situations:
39
40
1˛
i) If q < qO M
. If .q/ D 0, then
41
t
t
±
42
1X
1 X°
1˛
L..q/ D 0/ D
¹Kh .q Xj / .1 ˛/º Kh qO M
Xj .1 ˛/ D 0;
43
t
t
j D1
j D1
44
because Kh is a smooth function belonging to .0; 1/.
45
1˛
We are interested in 0 such that L.0 / D 0. So, we have 0 < 0, for q < qO M
.
46
The details of this case are depicted in Fig. A.1, taking the function, L./, into account
47 FA.1
which decreases when increases. Therefore, because @ log EL.q/[email protected] > 0, log EL.q/
48
1˛
monotonically increases for q < qO M
.
49
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
Computing critical values of exact tests
PR
OO
CT
ED
1˛
Fig. A.1. The function L./ is plotted against for the case of q < qO M
to evaluation of the sign of 0
such that L.0 / D 0.
RE
1˛
Fig. A.2. The function L./ is plotted against for the case of q > qO M
to evaluation of the sign of 0
such that L.0 / D 0.
1˛
ii) If q > qO M
. Likewise, we have
t
t
±
1 X
1 X°
1˛
¹Kh .q Xj / .1 ˛/º Kh qO M
Xj .1 ˛/ D 0:
t
t
CO
R
L¹.q/º D
j D1
j D1
1˛
We have 0 > 0, for q > qO M
, shown in Fig. A.2. L./ decreases as decreases.
1˛
In a similar manner to the case, q < qO M
, because @ log EL.q/[email protected] < 0, log EL.q/
1˛
monotonically decreases for q > qO M
. The proof of lemma 1 is complete.
Pt
To simplify the proof of proposition 1, let us define kNh .q/ D
j D1 kh .q Xj /=t and
®
¯s
Pt
KN s;h .q/ D j D1 Kh q Xj .1 ˛/ =t for s D 1; 2; 3.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
13
F
Scand J Statist
Proof of proposition 1. To prove proposition 1, we first note that
p
1˛
q
OM
't =
X
Z .t/
Z
EL.q/.q/dq D
Z
EL.q/.q/dq C
X.1/
1˛
q
OM
't =
X
Z .t/
1˛
q
OM
C't
EL.q/.q/dq;
p
=
t
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
t
EL.q/.q/dq
p
X.1/
C
p
1˛
q
OM
C't =
t
t
(A.2)
FA.2
14
Scand J Statist
given 't D t ˇ , 0 < ˇ < 1=2. By lemma 1, we have
p
1˛
q
OM
't =
.p i Z1
h
1˛
expŒlog EL.q/.q/dq exp log EL qO M
't
t
.q/dq
t
X.1/
1
h
.p i
1˛
D exp log EL qO M
't
t :
(A.3)
F
Z
L./ D KN 1;h .q/ PR
OO
In order to show that the upper bound at (A.3) vanishes to 0, we consider the function L./
defined in (A.1). This function can be rewritten as
t
X
¹K.h1 .q Xj // .1 ˛/º2
:
2
t
1 C t ¹K.h1 .q Xj // .1 ˛/º
j D1
(A.4)
Define 0 D t =gt with gt D t . Then, defining 2 .0; 1=2/, one can show that
p
t L.0 / D
p
t KN 1;h .q/C
t
t X
gt
p
j D1
tX
gt
t
j D1
¹K.h1 .q Xj // .1 ˛/º2
1C
1
gt
¹K.h1 .q Xj // .1 ˛/º
ED
p
p
p
t L.0 / D t KN 1;h .q/ ¹K.h1 .q Xj // .1 ˛/º2
1
1
gt
¹K.h1 .q Xj // .1 ˛/º
! 1 and
! 1; as t ! 1;
RE
CT
p
because t KN 1;h .q/ D Op .1/ (Chen & Hall, 1993). Thus, we obtain the order D O.t=gt /,
where is the solution to L./ D 0. In this case, because KN 4;h .q/ D Op .1/ by Chen &
Hall (1993), the Taylor’s expansion of the function L./ with respect to around 0 provides
the approximation
!
2
3
:
(A.5)
L./ D KN 1;h .q/ KN 2;h .q/ C 2 KN 3;h .q/ C Op
t
t
t3
D
CO
R
Solving (A.5) for gives
t KN 1;h .q/
C Op
KN 2;h .q/
t
gt2
(A.6)
;
which gives gt D o.h1 /; that is, if h D o.t 1=6 /, then 2 .0; 1=6/. Because KN 3;h .q/ D Op .1/
by Chen & Hall (1993), the Taylor expansion of the function log EL.q/ when is considered
around 0, and (A.6) implies that
!
!
2
3
t¹KN 1;h .q/º2
3
log EL.q/ D KN 1;h .q/C KN 2;h .q/COp
D
: (A.7)
CO
p
2t
t2
t2
2KN 2;h .q/
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
1˛
1˛
By the Taylor expansion of the function (A.7) around qO M
at q D qO M
't
1˛ Op .1/ and KN 2;h qO M
D Op .1/ (Chen & Hall, 1993), we obtain
!
® 1˛ ıp ¯2
.p 't
t
t kNh qO M
't3
1˛
1˛ log EL qO M 't
C Op 1=2
t D
t
2KN 2;h qO M
!
3
't
D C 't2 C Op 1=2
! 1 as t ! 1;
t
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
ıp
1˛ t , kNh qO M
D
(A.8)
for some constant C > 0. Therefore, by virtue of (A.3) and (A.6), we conclude that
p
1˛
q
OM
't =
Z
t
EL.q/.q/dq D op
1
:
t
(A.9)
X.1/
EL.q/.q/dq D op
1˛
q
OM
C't
p
= t
1
:
t
PR
OO
X
Z .t/
F
Similarly, we also obtain
(A.10)
Equations (A.9) and (A.10) imply that (A.2) can be rewritten as
1˛
q
OM
C't =
X
Z .t/
Z
EL.q/.q/dq D
X.1/
p
t
EL.q/.q/dq C op
1˛
q
OM
't
p
= t
1
:
t
(A.11)
ED
Through the result at (A.7), the right hand side of (A.11) can be written as
p
t
p
1˛
¯2#
qO M ZC't = t " ® N
t K1;2 .q/
1
1
EL.q/.q/dq C op
exp D
.q/dq C op
:
N 2;h .q/
t
t
2
K
p
p
1˛
1˛
q
OM
't = t
q
OM
't = t
(A.12)
Z
CT
1˛
q
OM
C't =
Hence, we approximate (A.12) as the function
1
"
®
¯2 #
t KN 1;2 .q/
1
exp :
.q/dq C op
t
2KN 2;h .q/
RE
Z1
CO
R
®
¯ q
By the assumption of .q/ D exp .q 0 /2 =202 = 202 and the Taylor expansion of the
q
1˛
function KN 1:h .q/= KN 2;h .q/ with respect to q around qO M
, we have
"
.q 0 /2
t¹KN 1;2 .q/º2
exp 202
2KN 2;h .q/
#
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
15
Computing critical values of exact tests
Scand J Statist
3
° ±2
1˛
N
2
/
.q
1
2
7
6 t kh qO M
0
1˛
1˛ q qO M
D exp 4
5 C op
2
t
20
2KN 2;h qO M
2
2
61
D exp 4
2
´
1˛ 1
C 2
tBt2 qO M
0
tBt2
1˛
qO M
μ8
<
1˛
qO M
1˛ 1
q tBt2 qO M
C 2
:
0
0
C 2
0
! 92
=
;
3
7
5 C op
!1
1
;
t
1˛ ® 1˛ ¯2 .
1˛ where Bt2 qO M
D kNh qO M
KN 2;h qO M
. Thus, the asymptotic estimator of the
nonparametric posterior expectation of the quantiles is
1˛
qO NP
D
1˛
qO M
C
1C
0
1˛
02 Bt2 .q
OM
/t
1
1˛
02 Bt2 .q
OM
/t
C op
1
:
t
This completes the proof.
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
(A.13)
16
Scand J Statist
1˛
To prove proposition 2, we start with the fact that qO M
satisfies the relationship 1 ˛ D
1˛
Pt
t
q
O
.
It
follows
that
the
Taylor
expansion
of the function around q01˛
K
X
j
j D1 h
M
provides the approximation
2
KN 1;h q01˛
kNh0 q01˛ 1˛
1
1˛
1˛
qO M D q0
:
(A.14)
1˛ 1˛ qO M q01˛ C op
N
N
t
kh q0
2kh q0
Note that we have covariance expression
D
1
cov
t
1
k
k
q01˛ X1
h
f .q01˛ /.1C2˛/
t
Co
.1 ˛/
PR
OO
cov kNh q01˛ ; KN 1;h q01˛ D
F
1
;K
1
t
q01˛ X1
h
:
:
(A.15)
1˛ 3
1˛
Proof of proposition 2. Because
qO M
q0
D
oP .t 1 / from
®
1˛ ¯3
2ı D Op h t and (9), we obtain the following equation as
KN 1;h q
0
1˛
qO NP
qO 1˛ q01˛
D˛
f
NP
2
1
1
Q
C op
˛ .t/
C op
:
t
t
1˛
qO NP
q01˛
q01˛
ED
1F
2
(A.14)
and
f 0 q01˛
(A.16)
CT
Equations (A.14) and (A.16) jointly imply that
2
1
1˛
.0 q01˛ /
.0 q01˛ /
1˛
f 0 .q 1˛ /
qM q01˛ C 2 B 2 qO 1˛
qM
q01˛ C 2 B 2 qO 1˛
t
.
/
0 t
M
B
0 t . M / t C
1˛
Q
.t/
[email protected]
/
:
f
.q
A
2
1 C 2 B 2 1qO 1˛ t
0 t . M /
2 1 C 2 B 2 1qO 1˛ t
0 t . M /
1˛
Because Bt q0
be rewritten as
0
B
Q
.t/
[email protected]
Df
(A.17)
ı
1˛
q01˛ ˛.1 a/ C op .1/ and qO M
D q01˛ C op .1/, (A.17) can
2
CO
R
2
RE
0
1˛
qM
q01˛ C
1C
˛.1˛/.0 q01˛ /
02 f 2 .q01˛ / t
˛.1˛/
02 f 2 .q01˛ / t
1
1˛
1˛ 2
f 0 q 1˛
1
C 1˛ qM q0
:
Af q
2 C op
t
˛.1˛/
2 1 C 2 f 2 q 1˛ t
. 0 /
0
(A.18)
Q
We take the expectation of the function .t/
in (A.18) using (A.14). Then, we obtain
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
0
1˛
q01˛ C
E qM
1
1˛
1˛ 2
f 0 q 1˛
1
C 1˛ E qM q0
Co
f
q
A
2
˛.1˛/
t
1 C 2 f 2 q 1˛ t
˛.1˛/
. 0 /
0
2 1 C 2 f 2 q 1˛ t
. 0 /
0
0 1
N 1;h .q 1˛ /
K
˛.1˛/f 0 .q01˛ /
˛.1˛/.0 q01˛ /
0
E
C
1˛
1˛
2 2
1˛
3
N
B
kh .q0
2f .q0
0 f .q0
/
/t
/t C
C
DB
@
A
1 C 2 f˛.1˛/
2 .q 1˛ / t
B
Q
E .t/
[email protected]
˛.1˛/.0 q01˛ /
02 f 2 .q01˛ / t
0
˛.1˛/f 0 .q 1˛ /
f 2 .q01˛ / t
f .q 1˛ / 2 1C
˛.1˛/
02 f 2 .q01˛ / t
0
2 C o
1
:
t
(A.19)
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
F2
Computing critical values of exact tests
Scand J Statist
ı To compute E KN 1;h q01˛ kNh q01˛ , we use the Taylor expansion in two variables of
f .x; y/ D y=x given as
y
Ey
Ey
1
.x Ex/ C
.y Ey/
x
Ex .Ex/2
.Ex/2
3Ey
1
C
.x Ex/2 .x Ex/.y Ey/:
.Ex/3
.Ex/2
(A.20)
F
f .x; y/ D
PR
OO
Equations (A.15) and (A.20) imply that
!
KN 1;h q01˛
2 f 0 q 1˛ h2
4 f .3/ q01˛ h4
1˛ E
D k 01˛ C k
1˛ 2f q0
24f q0
kNh q0
2 2 0 1˛ 00 1˛ 4
f q01˛ .1 C 2˛/
f q0
f q
h
1
1˛ 0
Co
:
k
t
t
4f 2 q0
The proof of proposition 2 is complete.
ED
Acknowledgements
CT
This research is supported by the NIH grant 1R03DE020851 - 01A1 (the National Institute of
Dental and Craniofacial Research). The authors are grateful to the editor and the referee for
suggestions that led to a substantial improvement in this paper.
References
CO
R
RE
Chen, S. X. & Hall, P. (1993). Smoothed empirical likelihood confidence intervals for quantiles. Ann. Stat.
21, 1166–1181.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences, Lawrence Erlbaumn Associates, New
Jersey.
DasGupta, A. (2008). Asymptotic theory of statistics and probability, Springer, New York.
DiCiccio, T. J., Kass, R. E., Raftery, A. & Wasserman, L. (1997). Computing Bayes factors by combining
simulation and asymptotic approximations. J. Am. Stat. Assoc. 92, 903–915.
Fan, J., Farmen, M. & Gijbels, I. (1998). Local maximum likelihood estimation and inference. J. R. Stat.
Soc. Ser. B 60, 591–608.
Fan, J. & Gijbels, I. (1995). Data-driven bandwidth selection in local polynomial fitting: variable bandwidth
and spatial adaptation. J. R. Stat. Soc. Ser. B 57, 371–394.
Fan, J. & Gijbels, I. (1996). Local polynomial modelling and its applications, Chapman and Hall, London.
Fan, J., Heckman, N. E. & Wand, M. P. (1995). Local polynomial kernel regression for generalized linear
models and quasi-likelihood functions. J. Am. Stat. Assoc. 90, 141–150.
Gibbons, J. D. & Chakraborti, S. (2011). Nonparametric statistical inference, Chapman and Hall, London.
Hall, P. & Welsh, A. H. (1983). A test for normality based on the empirical characteristic function.
Biometrika 70, 485–489.
Härdle, W. (1990). Applied nonparametric regression, Cambridge Univ. Press.
Kass, R. E. & Raftery, A. E. (1995). Bayes factors. J. Am. Stat. Assoc. 90, 773–795.
Kutner, M. H., Nachtsheim, C. J., Neter, J. & Li, W. (2005). Applied linear statistical methods, (5th edition).,
McGraw Hill, New York.
Lazar, N. A. & Mykland, P. A. (1998). An evaluation of the power and conditionality properties of
empirical likelihood. Biometrika 85, 523–534.
Lazar, N. A. (2003). Bayesian empirical likelihood. Biometrika 90, 319–326.
Mudholkar, G. S. & Hutson, A. D. (1997). Continuity corrected approximations for and “exact” inference
with Pearson’s X 2 . J. Stat. Plann. Inference 59, 61–78.
Owen, A. B. (1988). Empirical likelihood ratio confidence intervals for a single functional. Biometrika 75,
237–249.
UN
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
17
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.
AQ1
18
Scand J Statist
CT
Received December 2012, in final form January 2014
ED
PR
OO
F
Owen, A. B. (1990). Empirical likelihood ratio confidence regions. Ann. Stat. 19, 237–249.
Owen, A. B. (2001). Empirical likelihood, Chapman and Hall/CRC, New York.
Rosenbaum, P. R. (2005). An exact distribution-free test comparing two multivariate distributions based
on adjacency. J. R. Stat. Soc. Ser. B 67, 515–530.
Serfling, R. J. (2002). Approximation theorems of mathematical statistics, Wiley, New York.
Shepherd, L. A., Tsai, W. -M., Vexler, A. & Miecznikowski, J. C. (2013). dbEmpLikeNorm: test
for joint assessment of normality, R package. Available at: http://cran.r-project.org/web/packages/
dbEmpLikeNorm/index.html.
Silverman, B. J. (1986). Density estimation for statistics and data analysis, Chapman and Hall, London.
Simonoff, J. S. (1998). Smoothing methods in statistics, Springer-Verlag, New York.
Tanajian, H., Vexler. A. & Hutson, A. D. (2013). Novel and efficient density based empirical likelihood procedures for symmetry and K-sample comparisons: STATA package. Available at: http://sphhp.buffalo.
edu/biostatistics/research-and-facilities/software/stata.html.
Tibshirani, R. & Hastie, T. J. (1987). Local likelihood estimation. J. Am. Stat. Assoc. 82, 559–567.
Vexler, A. & Gurevich, G. (2010). Empirical likelihood ratios applied to goodness-of-fit tests based on
sample entropy. Comput. Stat. Data Anal. 54, 531–545.
Vexler, A. & Yu, J. (2011). Two-sample density-based empirical likelihood tests for incomplete data in
applications to a pneumonia study. Biometrical J. 53, 628–651.
Yu, J., Vexler, A. & Tian, L. (2010). Analyzing incomplete data subject to a threshold using empirical
likelihood methods: an application to a pneumonia risk study in an ICU setting. Biometrics 66, 123–130.
Yu, J., Vexler, A., Kim, S. & Hutson, A. D. (2011). Two-sample empirical likelihood ratio tests for medians
application to biomarker evaluations. Can. J. Stat. 39, 671–689.
Zhou, W. & Jing, B. -Y. (2003). Adjusted empirical likelihood method for quantiles. Ann. Inst. Stat. Math.
55, 689–703.
Supporting information
RE
Albert Vexler, Department of Biostatistics, The University at Buffalo, State University of New York , NY
14214, USA.
E-mail: [email protected]
CO
R
Additional supporting information may be found in the online version of this article at the
publisher’s website.
UN
01
02
03
04
05
06 AQ2
07
08
09
10 AQ3
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
A. Vexler et al.
© 2014 Board of the Foundation of the Scandinavian Journal of Statistics.