Large sample property of the bayes factor in a spline semiparametric regression model
-
Upload
alexander-decker -
Category
Technology
-
view
52 -
download
0
description
Transcript of Large sample property of the bayes factor in a spline semiparametric regression model
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
91
Large Sample Property of The Bayes Factor in a Spline
Semiparametric Regression Model
Ameera Jaber Mohaisen and Ammar Muslim Abdulhussain
Mathematics Department College of Education for Pure Science AL-Basrah University-Iraq
E-mail: [email protected]
Abstract
In this paper, we consider semiparametric regression model where the mean function of this model has two part,
the first is the parametric part is assumed to be linear function of p-dimensional covariates and nonparametric
( second part ) is assumed to be a smooth penalized spline. By using a convenient connection between penalized
splines and mixed models, we can representation semiparametric regression model as mixed model. In this
model, we investigate the large sample property of the Bayes factor for testing the polynomial component of
spline model against the fully spline semiparametric alternative model. Under some conditions on the prior and
design matrix, we identify the analytic form of the Bayes factor and show that the Bayes factor is consistent.
Keywords: Mixed Models, Semiparametric Regression Model, Penalized Spline, Bayesian Model, , Marginal
Distribution, Prior Distribution, Posterior Distribution, Bayes Factor, Consistent.
1. Introduction
In many applications in different fields, we need to use one of a collection of models for correlated data
structures, for example, multivariate observations, clustered data, repeated measurements, longitudinal data and
spatially correlated data. Often random effects are used to describe the correlation structure in clustered data,
repeated measurements and longitudinal data. Models with both fixed and random effects are called mixed
models. The general form of a linear mixed model for the ith
subject (i = 1,…, n) is given as follows (see
[10,13,15]), �� = ��� + ∑ ���� � + �� , ��~��0, �� , ��~�(0, ��) (1)
where the vector �� has length �� , �� and � are, respectively, a �� × � design matrix and a �� × �� design
matrix of fixed and random effects. β is a p-vector of fixed effects and u ! are the q -vectors of random effects.
The variance matrix � is a �� × �� matrix and �� is a �� × �� matrix.
We assume that the random effects { �� ; % = 1, … , ( ; ) = 1, … , *} and the set of error terms {��, … , �,} are
independent. In matrix notation (see [13,15]), � = �� + � + � (2)
Here � = (��, … . , �,). has length � = ∑ ��,� � , � = (��. , … , �,.). is a � × � design matrix of fixed effects, Z
is a � × � block diagonal design matrix of random effects, q = ∑ q!/! � , u = (u�0, … , u/0)0 is a q-vector of
random effects, � = 1%23(��, … , �,) is a � × � matrix and � = 1%23(��, … , ��) is a � × � block diagonal
matrix. In this paper, we consider semiparametric regression model (see [1,6,8,9,10,13,15]), for which the mean
function has two part, the parametric ( first part ) is assumed to be linear function of p-dimensional covariates
and nonparametric ( second part) is assumed to be a smooth penalized spline. By using a convenient connection
between penalized splines and mixed models, we can representation semiparametric regression model as mixed
model. In this model we investigate large sample properties of the Bayes factor for testing the pure polynomial
component of spline null model whose mean function consists of only the polynomial component against the
fully spline semiparametric alternative model whose mean function comprises both the pure polynomial and the
component spline basis functions. The asymptotic properties of the Bayes factor in nonparametric or
semiparametric models have been studied mainly in nonparametric density estimation problems related to
goodness of fit testing. These theoretical results include (see [4,7,12,14]). Compared to the previous approaches
to density estimations problems for goodness of fit testing and model selection, little work has been done on
nonparametric regression problems. Choi and et al 2009 studied the semiparametric additive regression models
as the encompassing model with algebraic smoothing and obtained the closed form of the Bayes factor and
studied the asymptotic behavior of the Bayes factor based on the closed form ( see [3] ).
In this paper, we obtain the closed form and studied the asymptotic behavior of the Bayes factor in spline
semiparametric regression model and we proved that the Bayes factor converges to infinity under the pure
polynomial model and the Bayes factor converges to zero almost surely under the spline semiparametric
regression alternative and we show that the Bayes factor is consistent.
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
92
2. Semiparametric and Penalized Spline
Consider the model: 4� = ∑ �5� + �(567�,�6 8 ) + �� , % = 1,2, … , ( (3)
Where 4� , … , 4, response variables and the unobserved errors are �� , … , �, are known to be i.i.d. normal with
mean 0 and covariance :;<= with :;< known.
The mean function of the regression model in (3) has two parts. The parametric ( first part ) is assumed to be
linear function of p-dimensional covariates 5� and nonparametric (second part) �(567�,�) is function defined on
some index set > ⊂ ��.
The model (3) can be expressed as a smooth penalized spline with q degree, then it's become as: 4� = ∑ �5� + ∑ �67567�,�@ �6 8 + ∑ {�A(567�,� − CADA � )7@ + �� (4)
Where C�, … , CD are inner knots 2 < C� < , , , < CD < F.
By using a convenient connection between penalized splines and mixed models. Model (4) is rewritten as
follows (see [13,15])
� = �� + � + � (5)
where
� = G4�⋮4,I , � = JKKKKL �8⋮�6�67�⋮�67@MN
NNNO , � = G��⋮�DI , = P(567�,� − C�)7@ ⋯ (567�,� − CD)7@⋮ ⋱ ⋮(567�,, − C�)7@ ⋯ (567�,, − CD)7@
S
� = JKKKL1 5�� … 56�1 5�< … 56<
567�,� … 567�,�@567�,< … 567�,<@
⋮ ⋮ ⋱ ⋮1 5�, … 56, ⋮ ⋱ ⋮567�,, … 567�,,@ MN
NNO
We assume ( < C + � + � + 1, ( + T = C + � + � + 1, (T is integer number and greater than 1 ), and we
assume that the design matrix is orthogonal in the following way: U,. U, = (=, ∀( ≥ 1 (6)
where:
U, = ( � , �) and � be the ( − (� + � + 1) × ( − (� + � + 1) matrix, and let = � + X
Based on the above setup for regression model (5) and the assumption of the orthogonal design matrix (6), we
consider a Bayesian model selection problem in a spline semiparametric regression problem. Specifically, we
would like to choose between a Bayesian spline semiparametric model and its pure polynomial counterpart by
the criterion of the Bayes factor for two hypotheses, Y8: � = �� + � versus Y�: � = �� + � + � (7) As for the prior of �, 2(1 � under Y�, we assume � and u are independent and � ~ �(0, ∑8) , ∑8 = :[<=67@7� (8) � ~ �(0, ∑�) , ∑� = :\<=D (9)
3. Bayes factor and marginal distribution
The response 4� in (4) follows the normal distribution with mean ]� and variance :;< , where ]� = ∑ �5� +6 8)=1��)5�+1,%)+C=1^{�C(5�+1,%−^C)+�, for, %=1,2,…,(. Thus, given covariates and ](=(]1, …, ]()>, the
n-dimensional response vector �, = (4� , … , 4,). follows the n-dimensional normal distribution with mean ],
and covariance matrix :;<=, where =, is the ( × ( identity matrix. Also from the prior distributions specified in
(8) and (9), we can deduce the joint distribution of ], is the multivariate normal distribution (U_�) with mean
zero and ( × ( covariance matrix �∑8�. + ∑�. .
In summary, we have the following:- �,|], ~ U_�,(],, :;<=,) , ], ~ U_�,(0,, �∑8�. + ∑�.) (10)
Result 1:
Suppose the distribution of ], and �, are given by (10). Then the posterior distribution of P(],|�,)is the
multivariate normal with the following mean and variance. a(],|�, , :;<) = ∑<(∑< + :;<=, )b��,
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
93
c2*(],|�, , :;<) = (∑<b� + (:;<=, )b�)b� = ∑<(∑< + :;<=, )b�:;<=,
Where ∑< = �∑8�. + ∑�. . Furthermore, marginally �, follows
�,~ U_�(0, :;<=, + ∑<) (11)
Note the (%, ))th element of ∑<, ∑<(%, )) = ∑ :[<67@7�A � 5�A5A + ∑ :\<d�AdADA � , %, ) = 1, … , ( (12)
where d�A = (567�,� − CA)7@
Applications of the previous result yield that, the marginal distribution of �, under Y8 and Y� are respectively �,(0, _8) and �,(0, _�), where _8 = :;<=, + �∑8�. _� = :;<=, + �∑8�. + ∑�. Hence, the Bayes factor for testing problem (7) is given by:
e8� = �(�|Y8)�(�|Y�) = 1f2gdet (_8) exp { − 12 �,._8b��,}1f2gdet (_�) exp { − 12 �,._�b��,}
= fmno (pq)fmno (pr) exp{− �< �,.( _8b�−_�b�)�,} (13)
4. Asymptotic behavior of Bayes factor
The Bayes factor is said to be consistent if ( see[3,14]) s%�,→u e8� = ∞ , 2. T. w8u under the Y8, (14) s%�,→u e8� = 0 , 2. T. w�u under the Y�, (15)
We investigate the consistency of the Bayes factor, by establishing (14) and (15), under the spline
semiparametric regression model described in the section 2. By truncating the regression up to the first ( terms.
We get 4� = ∑ �5� + ∑ �67567�,�@ �6 8 + ∑ {�A(567�,� − CA,b(67@7�)A � )7@ } + �� Now define _�,, as the covariance matrix of the marginal distribution of �, under Y� as:- _�,, = :;<=, + �∑8�. + �∑�,��.
= :;<=, + U,x,U,. (16)
Where ∑�,� is diagonal matrix with diagonal elements :\< from size ( − (� + � + 1) × ( − (� + � + 1) and
x, = y∑8 0��0�� ∑�,�z
where 0��, 0�� are the matrices for zeros element with size (� + � + 1) × ( − (� + � + 1) and ( − (� + � +1) × (� + � + 1), respectively.
Then, we can rewrite (16) by using x, as: _�,, = U, { |}~, =, + x, � U,. (17)
Let _8 be the covariance matrix of the marginal distribution of �, under Y8, then _8 can be represented as: _8 = :;<=, + �∑8�. + �0���.
= :;<=, + U,�,U,. where: �, = y ∑8 0��0�� 0��z
where 0�� is the matrix for zeros element with size ( − (� + � + 1) × ( − (� + � + 1), then we can rewrite _8
as the following: _8 = U, { |}~, =, + �, � U,. (18)
Then we can use the two matrices above (_8 , _�,,) for matrix inversions and calculating determinants as shown
in following lemma.
Lemma1: Let _�,, and _8 be defined as in (17) and (18) respectively, and suppose U, satisfied (6). Then, the
following hold.
i) U,U,. = U,. U, = (=,
ii) det(_8) = (67@7�(|}~, + :[<)67@7� :;<(,b(67@7�)) iii) det�_�,,� = (, (|}~, + :[<)67@7�(|}~, + :\<),b(67@7�) iv) _8b� = �,~ U, { |}~, =, + �, �b� U,.
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
94
v) _�,,b� = �,~ U, { |}~, =, + x, �b� U,. Proof:-
i) By multiplying U, and U,. to equation (6) from right and left side, we have:- U,U,.U,U,. = U,(=,U,.
= U,(U,.
= (U,U,.
Multiplying (U,U,.)b� to the above equation from left side, we get:- U,U,.U,U,.(U,U,.)b� = (U,U,.(U,U,.)b� U,U,. = (=,
ii) _8 = U, { |}~, =, + �, � U,., and by using (6), we get
= �� , ��JKKKKKKL(|}~, + :[<) 0 ⋯ 0 0
0 (|}~, + :[<) ⋯ 0 0⋮00
⋮00⋱ ⋮ ⋮⋯ |}~, 0⋯ 0 |}~, MN
NNNNNO �� , ��.
= JKKKKKL((|}~, + :[<) 0 ⋯ 0 0
0 ((|}~, + :[<) ⋯ 0 0⋮00⋮00
⋱ ⋮ ⋮⋯ :;< 0⋯ 0 :;<MNNNNNO
Then
det(_8) = (67@7� (|}~, + :[<)67@7�:;<(,b(67@7�))
iii) _�,, = U, { |}~, =, + x, � U,.
= �� , ��JKKKKKKL(|}~, + :[<) 0 ⋯ 0 0
0 (|}~, + :[<) ⋯ 0 0⋮00
⋮00⋱ ⋮ ⋮⋯ |}~, + :\< 0⋯ 0 |}~, + :\<MN
NNNNNO �� , ��.
= JKKKKKKL((|}~, + :[<) 0 ⋯ 0 0
0 ((|}~, + :[<) ⋯ 0 0⋮00
⋮00⋱ ⋮ ⋮⋯ ((|}~, + :\<) 0⋯ 0 ((|}~, + :\<)MN
NNNNNO
Then
det�_�,,� = (, (:;<( + :[<)67@7�(:;<( + :\<),b(67@7�) iv) _8b� = {U, { |}~, =, + �, � U,. }b�
= U,.b� { |}~, =, + �, �b� U,b�
= ���q, (=, U,.b� { |}~, =, + �, �b� U,b�(=, ���q
,
= �,~ U,U,. U,.b� { |}~, =, + �, �b� U,b�U,U,. by lemma 1.i
= �,~ U, { |}~, =, + �, �b� U,.
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
95
v) _�b� = {U, { |}~, =, + x, � U,. }b�
= U,.b� { |}~, =, + x, �b� U,b�
= ���q, (=, U,.b� { |}~, =, + x, �b� U,b�(=, ���q
,
= �,~ U,U,. U,.b� { |}~, =, + x, �b� U,b�U,U,.
= �,~ U, { |}~, =, + x, �b� U,.
Now let e�8� = fmno (pq,�)fmno (pr) exp{− �< �,.( _8b�−_�,,b�)�,} (19)
Which an approximation of e8� with _� replaced by _�,, in (13)
Thus,
s�3 e�8� = 12 s�3 det (_�,,)det (_8) − 12 �,.( _8b�−_�,,b�)�,
2 log e�8� = log det (_�,,)det (_8) − �,.( _8b�−_�,,b�)�,
By results of lemma1 we have:-
2 log e�8� = log (, (:;<( + :[<)67@7�(:;<( + :\<),b(67@7�) (67@7� (:;<( + :[<)67@7�:;<
(,b(67@7�)) − 1(< ��,.U, �� :;<( =, + �, �b� − � :;<( =, + x, �b�� U,.�,�
= log ,��(����q) (�}~� 7|�~)��(����q) |}~(��(����q)) − �,~ {�,.U, y0(67@7�)×(67@7�) 0��0�� �� z U,.�,}
= log( |}~7,|�~ |}~ ),b(67@7�) − �,.�,�,}
where �, = 1(< U, y0(67@7�)×(67@7�) 0��0�� �� z U,.
and �� = x%23 ( ,~|�~(|}~7,|�~)|}~ )
To establish the consistency of the Bayes factor we focus on e�8� , first and the remaining terms will be
considered later. Without loss of generality we assume :;< = 1 in the remainder of this paper.
Now let �,,� = ∑ ,|�~(�7,|�~),� � = ,~|�~(�7,|�~) (20) �,,< = ∑ log(1 + (:\<),� � = ( log(1 + (:\<) (21) �,,� = ∑ ( ,|�~�7,|�~)<,� � = (�( |�~�7,|�~)< (22)
Lemma2: let �, = (4� , … , 4,). where 4� ′T are independent normal random variable with mean ∑ �5� +6 8)=1���+)5�+1,%) and variance :�2=1, then there exist a positive constant �1 such that
���,q { ∑ log(1 + (:\<) − ,b(67@7�)� � �,.�,�,� > ��, with probability tending to 1.
This implies that, under Y8.
log e�8� ,→u���� ∞, in probability. Proof
Let �, = ( ��, �<, … , �,) are independent standard normal random variables, Note that �, �, � + �� and �,.�,�, = (�, + ��).�, (�, + ��)
= �,.�,�, + (��).�,(��) + �,.�,�� + (��).�,�,.
By the orthogonality of the design matrix
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
96
(��).�, = (��). �,~ U, y0(67@7�)×(67@7�) 0��0�� �� z U,. = 0.
Then �,.�,�, = �,.�,�,.
Using the expectation formula of the quadratic form given in (12), we obtain
0 < a(�,.�,�,) = �*(�,) = � (:\<(1 + (:\<),
� �= �,,�
Note that �,< = 1( U, y0(67@7�)×(67@7�) 0��0�� �� z U,.U, y0(67@7�)×(67@7�) 0��0�� �� z U,. = �,¡ U, y0(67@7�)×(67@7�) 0��0�� ��< z U,.
and �*(�,<) = �,~ ∑ ( ,~|�~�7,|�~)<,� � = ∑ ( ,|�~�7,|�~)<,� � = �,,�.
Similarly, using the variance formula of the quadratic form of the multivariate normal variables, we have c2*(�,.�,�,) = 2�*(�,<) = 2 ∑ ( ,|�~�7,|�~)<,� � = 2�,,�.
Let ¢, = ��,~b��,q<��,q . Then,
w* { ���,q { ∑ log(1 + (:\<)< − ,b(67@7�)� � �,.�,�,� ≤ ¢, }
= w* � { �,.�,�,}�,,� ≥ �,,<�,,� − ¢, � = w* � { �,.�,�,}�,,� − a(�,.�,�,)�,,� ≥ �,,<�,,� − ¢, − a(�,.�,�,)�,,� � = w* � 1�,,� {(�,.�,�,) − a(�,.�,�,)} ≥ �,,< − a(�,.�,�,)�,,� − ¢, � ≤ ���,q~ ¤¥��¦�§¨�¦��
(©�,~�©�,q~©�,q )~ = ª��,¡(��,~b��,q)~ ,→u���� 0,
where the last statement holds from the Chebyshev inequality.
Then �,,< − �,,� ≥ ¢��,,� for some ¢� > 0 and let ¢, ≥ ¢�/2.
Hence, we conclude that there exists a positive constant �� = ¢�/2 such that ���,q { ∑ log(1 + (:\<)< − ,b(67@7�)� � �,.�,�,� > ��, with probability tending to 1.
Consequently, it follows that 2 log e�8� = ∑ log(1 + (:\<)< − ,b(67@7�)� � �,.�,�, ,→u���� ∞ in probability under Y8 log e�8� ,→u���� ∞
Lemma3
Suppose we have the penalized spline smoothing, then 1�,,� �*� _�,,b� − _�b� � ,→u���� 0
Proof:
Since _� − _�,, is the covariance matrix of ]∗(5) = ∑ dA,7XA ,b(67@)7� �A = X� it is a positive definite matrix.
Thus _�,,b� − _�b� is also a positive definite matrix. Thus, �*� _�,,b� � ≥ �*( _�b�), also from the matrix analogue
of the Cauchy-Schwarz inequality, we have
�*�X. _�,,b� � ≤ �*(X. X)�*�_�,,b<� = ®X® ¯_�,,b�¯ where
¯_�,,b�¯ = { �*�_�,,b<� }�< = � 1(< { � °:;<( + :[<±b<67@7�� �
+ � °:;<( + :\<±b<,b(67@7�)� �
}� �<
Note that
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
97
∑ ²|}~, + :[<³b<67@7�� � = 67@7�´�}~� 7|µ~¶~ = 67@7�
·�}~�� �µ~� ¸~ = ,~(67@7�)(|}~7, |µ~)~ ≤ 67@7�|µ¹ = º(1)
And
� °:;<( + :\<±b<,b(67@7�)� �
= ( − (� + � + 1)´:;<( + :\<¶< ≤ (
´:;<( + :\<¶< = (�(:;< + (:\<)< ≤ (�((:\<)< = º(() Thus, it follows that ¯_�,,b�¯ = »º ²�,³¼� <⁄ = º(()b� <⁄
Since sup¿∈�8,��®X® < 1 from the distances between truncated power functions, it follows that ®X®< = ∑ XA< (5�),� � ≤ ( , where XA< (5�) = (567�,� − CA)7@ is element in X
Thus, ���,q �*� _�,,b� − _�b� � = ���,q �*� _�,,b�( _� − _�,,) _�b� �
= ���,q ∑ :\<(A.,7XA ,b(67@7�) _�,,b�(A. _�b�).)
≤ |�~��,q ¯_�,,b�¯< ∑ ®A®<,7XA ,b(67@7�) ≤ ,|�~, ��,q ∑ 1,7XA ,b(67@7�) ≤ D |�~ ��,q = D|�~�~��~(q����~ )
= D(�7,|�~),~ Thus 1�,,� �*� _�,,b� − _�b� � ≤ ^(1 + (:\<)(< ,→u���� 0
Theorem(1)
Consider the testing problem in (7), suppose that the true model is Y8 ( pure polynomial model ) and let w8,
denote the true distribution of the whole data, with p.d.f w8(4|�8) = ∅{( 4 − �.�8)/ :;}, where ∅(∙) is the
standard normal density, then, the Bayes factor is consistent under the null hypothesis Y8: lim,→u e8� = ∞ in w8, probability.
Proof log e8� = 12 log det (_�)det (_8) − 12 �,.( _8b�−_�b�)�, = �< log mno (pq)mno (pq,�) + �< log mno (pq,�)mno (pr) − �< �,.( _8b�−_�,,b�)�, − �< �,.(_�,,b�−_�b�)�, = loge�8� + �< log mno (pq)mno (pq,�) − �< �,.(_�,,b�−_�b�)�,
= loge�8� + Å� + Å< where Å� = �< log mno (pq)mno (pq,�) , Å< = − �< �,.(_�,,b�−_�b�)�, Since _� − _�,, is a positive definite matrix, 1Æ� (_�) ≥ 1Æ� (_�,,), thus Å� ≥ 0 a( �,.�_�,,b�−_�b�)�, � = :;< �*� _�,,b� − _�b�� + (��).� _�,,b� − _�b�� ��, since _�,,b� − _�b�, _�,,b� and _�b� are all nonnegative definite matrices, it follows that 0 ≤ (��).(_�,,b� − _�b�) �� ≤ (��). _�,,b� ��. By orthogonality of the design matrix (6), we have �.U, = U,.� = ( � =67@7� , 0,b(67@7�)�. Thus similarly to the proof of lemma (3), we get:
(��). _�,,b� �� = �,~ (��).U, { |}~, =, + x, �b� U,. ��
= �,~ �.( � =67@7� , 0,b(67@7�)� { |}~, =, + x, �b� ( � =67@7� , 0,b(67@7�)� �
= �. � =67@7� , 0,b(67@7�)� { |}~, =, + x, �b� � =67@7� , 0,b(67@7�)� �
= ∑ ��<67@7�� � { |}~, + :[< �b� = º(1). Let Ç > 0, by the Markov inequality when r = 1 and lemma (3), we have:
�* { 1�,,� �,.�_�,,b�−_�b�)�, > Ç È ≤ 1�,,� { a( �,.�_�,,b�−_�b�)�, �Ç }
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
98
≤ ���,q É��pq,��qbpq�q �Ê + ���,q (Ë[)§ pq,��q Ë[Ê ,→u���� 0.
Note that _�,,b� − _�b� is nonnegative definite and thus, �,.(_�,,b�−_�b�)�, is nonnegative random variable.
Therefore, Å< = ���,,��. The above results with lemma (2) imply there exists a constant �� such that 1�,,� log e8� ≥ 1�,,� log e�8� + �(1)
> �� + �(1). That is, e8� ≥ exp( �,,�{ �� + �(1)}) ,→u���� ∞ in w8, probability.
This completes the proof.
Now we consider asymptotic behavior of the Bayes factor under Y� and assuming that the true model for �, is
in Y�.
Lemma(4):
Let �, = (4� , … , 4,) where 4� ′T are independent normal random variables with mean ∑ �5� +6 8)=1���+)5�+1,%)+C=1(−(�+�+1)d%C�C and variance :�2=1 and let Ì(= a �( > �( a �( . Then, lim,→u �Í� �,.�,�, = 1, in w�u probability.
Proof:
From the moment formula of the quadratic form of normal random variables, the expectation and variance of
quadratic form �,.�,�, are given by: a(�,.�,�, ) = �* ( �, ) + a( �, ). �, a( �, ) = ∑ ,|�~(�7,|�~),b(67@7�)� � + a( �, ). �, a( �, ) = �,,� + Ì, , and c2*(�,.�,�, ) = 2�* � �,< � + 4a( �, ). �,< a( �, ). Note Ì, = a( �, ). �, a( �, )
= � �. , �,b(67@7�). �U,. �,~ U, y0(67@7�)×(67@7�) 0��0�� �� z U,. U,� �. , �,b(67@7�). �.
= � �. , �,b(67@7�). � (. �,~ y0(67@7�)×(67@7�) 0��0�� �� z ( � �. , �,b(67@7�). �.
= � �. , �,b(67@7�). � y0(67@7�)×(67@7�) 0��0�� �� z � �. , �,b(67@7�). �.
= �Ab(67@7�). �� �,b(67@7�) = ∑ ,~|�~\Ï~�7,|�~,b(67@7�)� � ,
since T��� �� < ∞ , we have Ì, ≤ T��� ��< ( ∑ ,|�~�7,|�~,� � = º�( �,,� �.
On the other hand, consider a fixed positive integer � ≥ 1 such that �Ð< > 0 and a sufficiently large ( with ,|�~�7,|�~ ≥ �<. For such ( and �.
Ì, = � (<:\<��<1 + (:\<,b(67@7�)
� �≥ (<:\<1 + (:\< �Ð<
= ( ,|�~�7,|�~ �Ð< ≥ ( \Ñ~< .
Further, similarly to the previous calculation, we have
a(�,).�,<a(�,) = ��. , �,b(67@7�). �U,. 1(� U, y0(67@7�)×(67@7�) 0��0�� ��< z U,.U,��. , �,b(67@7�). �. = 1( �,b(67@7�). ��,,b(67@7�)< �,b(67@7�) = �, ∑ ,¹|�¹\Ï~(�7,|�~)~,b(67@7�)� � .
As in case of a( �, ). �, a( �, )
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
99
a( �, ).�,< a( �, ) ≤ ∑ ,¡|�¹\Ï~��7,|�~�~,� � ≤ ( T��� ��< ∑ ,~|�¹��7,|�~�~,� � = º( (�,,�).
Consider a fixed positive integer � ≥ 1 such that �Ð< > 0 and a sufficiently large n with ,|�~�7,|�~ ≥ �< . Again, for
such n and N, a( �, ). �,< a( �, ) = ∑ ,¡|�¹\Ï~��7,|�~�~ ,b(67@7�)� � ≥ ( �Ð< ² ,|�~�7,|�~³< ≥ \Ñ~ (.
By combining value of �,,� and the previous result, we have c2*(�,.�,�, ) ≤ º�( �,,� �.
Furthermore, note that a( �, ). �, a( �, ) ≥ a( �, ). �,< a( �, ). Let Ç > 0, by the Chebshev inequality. �* » Ò �Í� �,.�,�, − Ó�¦�§¨�¦��Í� Ò > Ç ¼ ≤ ¤¥��¦�§¨�¦��Í�~Ê~ = º( (b< ) ,→u���� 0.
Since lim,→u Ó�¦�§¨�¦��Í� = lim,→u ²��,qÍ� + Í�Í� ³ = 1 lim,→u �Í� �,.�,�, = 1 , in probability
This completes the proof.
Lemma(5):
Suppose we have the penalized spline smoothing, then
lim,→u1Ì, log 1Æ�(_�)1Æ�(_�,,) = 0
Proof:
From a property of the determinant of the positive definite matrix, we have 1Æ�(_�) ≤ ∏ �:;< + ∑ :[<67@7�A � 5�A< + ∑ :\<d�A<,7XA � �,� � = (, ∏ ´ |}~, + ∑ |µ~¿ÏÕ~,67@7�A � + ∑ |�~ÖÏÕ~,,7XA � ¶,� � ,
where 5�A , d�A is the k th element of 5� , d� respectively.
From lemma (1), we have: det�_�,,� = (, (|}~, + :\<),b(67@7�) (|}~, + :[<)67@7�.
Thus
log 1Æ�(_�)1Æ�(_�,,) = log(1Æ�(_�)) − log(1Æ��_�,,�)
≤ ( log ( + ∑ log ´|}~, + ∑ |µ~¿ÏÕ~,67@7�A � + ∑ |�~ÖÏÕ~,,A � ¶,� � − { ( log ( + (� + � + 1) log ²|}~, +:�2+(−�+�+1:�2(+:�2=º1−º1=º(1).
Therefore, it follows that �Í� log �×É(pq)�×É(pq,�) ,→u���� 0
Theorem(2):
Consider the testing problem in (7), suppose that the true model is in Y� ( the penalized spline semiparametric
model ) and let w�, denote the true distribution of the whole data, with p.d.f. ��(4|��, ��) = ∅{(4 −(�.�� + ��))/:;}, where ∅(∙) is the standard normal density. Then the Bayes factor is consistent under the
alternative hypothesis Y�: lim,→u e8� = 0 in w�, probability.
Proof: log e8� = 12 log det (_�)det (_8) − 12 �,.( _8b�−_�b�)�, = �< log mno (pq)mno (pq,�) + �< log mno (pq,�)mno (pr) − �< �,.( _8b�−_�,,b�)�, − �< �,.(_�,,b�−_�b�)�, = loge�8� − �< �,.( _8b�−_�,,b�)�, − �< �,.(_�,,b�−_�b�)�,
= loge�8� + Å� + Å< ,
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
100
where
loge�8� = 12 log det (_�,,)det (_8) − 12 �,.( _8b�−_�,,b�)�,
Å� = �< log mno (pq)mno (pq,�) , Å< = − �< �,.(_�,,b�−_�b�)�, . First, by lemma (1),
mno (pq,�)mno (pr) = ,� (�}~� 7|µ~)����q(�}~� 7|�~)��(����q)
,����q(�}~� 7|µ~)����q |}~(��(����q)) = ,��(����q)´�}~� 7|�~¶��(����q)|}~(��(����q)) = ²|}~7,|�~|}~ ³,b(67@7�)
=²1 + ,|�~|}~ ³,b(67@7�) = º(() ⟹ log ²1 + ,|�~|}~ ³,b(67@7�) = º(().
And from lemma (4) Ì, ≤ º((<). Then �Í� log mno (pq,�)mno (pr) ≤ º((b�) ,→u���� 0.
Therefore, 1Ì, loge�8� = 12 ° 1Ì log det�_�,,�det(_8) − 1Ì �,.( _8b�−_�,,b�)�,±
Since _8b� − _�,,b� = �,, then �Í� �,.( _8b�−_�,,b�)�, = �Í� �,.�,�,, from lemma (4), lim,→u �Í� �,.�,�, = 1. Then 1Ì, loge�8� ,→u���� = 12 ° 1Ì log det�_�,,�det(_8) − 1Ì �,.( _8b�−_�,,b�)�,± ,→u���� = 12 (0 − 1) = −12
∴ �Í� loge�8� ,→u���� b�< , in the w�, probability.
And from lemma (5), �Í� Å� ,→u���� 0, finally, note that Å< is nonpositive random variable since _�,,b� − _�b� is a nonnegative definite
matrix.
Therefore, by combining the above results, there exist �< > 0 such that lim,→u sup �Í� log e8� ≤ bÚ~< , with probability tending to 1.
Hence, e8� ,→u���� 0, in w�, probability.
This completes the proof.
5.Conclusion
1- The Bayes factor for testing problem Y8: � = �� + � versus Y�: � = �� + � + � is given by: e8� = fmno (pq)fmno (pr) exp{− �< �,.( _8b�−_�b�)�,}. 2- If �, = (4� , … , 4,). where 4� ′T are independent normal random variable with mean ∑ �5� +6 8 ∑ �67567�,�@ � and variance :;< = 1, then there exist a positive constant �� such that ���,q { ∑ log(1 + (:\<) − ,b(67@7�)� � �,.�,�,� > ��, with probability tending to 1.
This implies that, under Y8. log e�8� ,→u���� ∞, in probability.
3- If the true model is Y8 ( the pure polynomial model ), then, the Bayes factor is consistent under the null
hypothesis Y8. This implies that, lim,→u e8� = ∞ in w8, probability.
4- If �, = (4� , … , 4,) where 4� ′T are independent normal random variables with mean ∑ �5� +6 8)=1���+)5�+1,%)+C=1(−(�+�+1)d%C�C and variance :�2=1 and if Ì(=a�(>�(a(�(). Then, lim,→u �Í� �,.�,�, = 1, in w�u probability.
5- If the true model is in Y� ( the penalized spline semiparametric model ) ,then the Bayes factor is
consistent under the alternative hypothesis Y�. This implies that, lim,→u e8� = 0 in w�, probability.
6.References
[1] Angers, J.-F. and Delampady, M. (2001). Bayesian nonparametric regression using wavelets. Sankhy2Û Ser. B
63, 287–308.
[2] Berger, J. O. (1985). Statistical Decision Theory and Bayesian Analysis. Springer, New York.
Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.11, 2013
101
[3] Choi, T., Lee, J. and Roy, A. (2008). A note on the Bayes factor in a semiparametric regression model,
Journal of Multivariate Analysis. 1316-1327.
[4] G.Gosal, J. Lember, A. Van der Vaart, Nonparametric Bayesian model selection and averaging, Electron.J.
Stat. 2 (2008) 63-89.
[5] Ghosh, J. K., Delampady, M. and Samanta, T. (2006). Introduction to Bayesian Analysis: Theory and
Methods. Springer, New York.
[6] Ghosh, J. K. and Ramamoorthi, R. V. (2003). Bayesian Nonparametric. Springer, New York.
[7] I. Verdinelli, L. Wasserman, Bayesian goodness –of-fit testing using infinite- dimentional exponential
families, Ann. Stat.26 (4) (1998)1215-1241.
[8] L. H. Zhao, Bayesian aspects of some nonparametric problems., Ann.Statist. 28 (2000) 532-552.
[9] P. J. Lenk, Bayesian inference for semiparametric regression using a Fourier representation, J. R, Stat. Soc.
Ser. B Stat. Methodol. 61 (4) 1999) 863-879.
[10] Natio, k. (2002) " Semiparametric regression with multiplicative adjustment '' Communications in statistics,
Theory and methods 31, 2289-2309.
[11] Pelenis, J (2012) Bayesian Semiparametric Regression. Institute for Advanced Studies, Vienna 1-39.
[12] R.McVinish, J. Rousseau,K. Mengerson, Bayesian goodness –of-fit testing with mixtures of triangular
distributions, Scand. J, statist. (2008), in press (doi:10.1111/j.1467-9469.2008.00620.x).
[13] Ruppert, D, Wand, M. P. and Carroll, R. J.(2003) Semiparametric Regression. Cambridge University Press.
[14] S.C. Dass, lee, A note on the consistency of Bayes factors for testing point null versus non-parametric
alternative, J.Statist. 26 (4) (1998) 1215-1241.
[15] Tarmaratram, K. (2011) Robust Estimation and Model Selection in Semiparametric Regression Models,
proefschrift voorgedragen tot het behalen van de grad van Doctor.
This academic article was published by The International Institute for Science,
Technology and Education (IISTE). The IISTE is a pioneer in the Open Access
Publishing service based in the U.S. and Europe. The aim of the institute is
Accelerating Global Knowledge Sharing.
More information about the publisher can be found in the IISTE’s homepage:
http://www.iiste.org
CALL FOR JOURNAL PAPERS
The IISTE is currently hosting more than 30 peer-reviewed academic journals and
collaborating with academic institutions around the world. There’s no deadline for
submission. Prospective authors of IISTE journals can find the submission
instruction on the following page: http://www.iiste.org/journals/ The IISTE
editorial team promises to the review and publish all the qualified submissions in a
fast manner. All the journals articles are available online to the readers all over the
world without financial, legal, or technical barriers other than those inseparable from
gaining access to the internet itself. Printed version of the journals is also available
upon request of readers and authors.
MORE RESOURCES
Book publication information: http://www.iiste.org/book/
Recent conferences: http://www.iiste.org/conference/
IISTE Knowledge Sharing Partners
EBSCO, Index Copernicus, Ulrich's Periodicals Directory, JournalTOCS, PKP Open
Archives Harvester, Bielefeld Academic Search Engine, Elektronische
Zeitschriftenbibliothek EZB, Open J-Gate, OCLC WorldCat, Universe Digtial
Library , NewJour, Google Scholar