of 23
7/30/2019 The Integrated
1/23
The IntegratedConditional Moment Test
Herman J. Bierens
AbstractIn this lecture I will review the integrated con-
ditional moment (ICM) test for functional form
of a conditional expectation model. This is a
consistent test: the ICM test has asymptotic
power 1 against all deviations from the null
hypothesis. Moreover, this test has non-trivialn local power.I will focus on the mathematical foundations
of the ICM approach, in particular the consis-
tency proof and the derivation of the asymp-
totic null distribution.
1
7/30/2019 The Integrated
2/23
1 The Fourier transform of
a Borel measurable function
Let g(x) be a Borel measurable real functionon Rk. The Fourier transform of g(x) relativeto a probability measure X(.) on the Borelsets in Rk is defined by
() = Zexp(i.0x) g(x)dX(x), i = 1,provided that
R|g(x)|dX(x) < .
LEMMA 1: Let g1(x) and g2(x) be Borel mea-surable functions on Rk satisfying R|g1(x)|dX(x)< , R|g2(x)|dX(x) < , with Fouriertransforms 1 () , 2 () , respectively, rela-tive to a probability measure (.) on the Borelsets in Rk. Then g1(x) = g2(x) a.s. X(.),i.e.,
B0 = x Rk : g1(x) g2(x) = 0X(B0) = 1, if and only if 1 () 2 () .2
7/30/2019 The Integrated
3/23
Proof: Suppose1 ()
2 () and X(B0) 0. Then we can
define the probability measures
m (B) =1
c ZBrm(x)dX(x), m = 1, 2,
with corresponding characteristic functions
3
7/30/2019 The Integrated
4/23
m () = Zexp(i.0y) dm (y)=
1
c
Zexp(i.0x) rm(x)dX(x)
for m = 1, 2. But I have just established that
Rexp(i.0x) r1(x)dX(x)
Rexp(i.0x) r2(x)dX(x)
hence 1 ()
2 () , which by the unique-
ness of characteristic functions implies that1 (B)= 2 (B) for all Borel sets B Rk. It is nowan easy (ECON 501) exercise to verify that the
latter implies r1(x) = r2(x) a.s. X(.), henceg1(x) = g2(x) a.s. X(.).
Corollary:
LEMMA 2: Let Ube a random variable satis-fying E[|U|] < , and let X Rk be a ran-dom vector. If P [E(U|X) = 0] < 1 then thereexists a Rk such that E[Uexp(i.0X)] 6=0.
Question: Where to look for such a ?
4
7/30/2019 The Integrated
5/23
LEMMA 3: If X is bounded then under the
conditions of Lemma 2, for each > 0 thereexists a satisfying kk < such thatE[Uexp(i.0X)] 6= 0.
Proof: Let X R. Then
E[Uexp(i.X)] = E"U Xm=0
immXm
m! #=
Xm=0
imm
m!E[U.Xm]
Since E[Uexp(i.X)] 6= 0 for some wemust have that E[U.Xm] 6= 0 for some integer
m 0. Let m0 be the smallest m for whichE[U.Xm] 6= 0. Thendm0E[Uexp(i.X)]
(d)m0
=0
= im0E[U.Xm0] 6= 0
which implies that E[Uexp(i.X)] 6= 0 for6= 0 arbitrarily close to zero.
5
7/30/2019 The Integrated
6/23
LEMMA 4: Under the conditions of Lemma 3,
the set S0 = Rk : E[U. exp(i.0X)] = 0has Lebesgue measure zero and is nowhere dense.
Proof: Let k = 1 and 0 S0. Define U0 =Uexp(i.0X) . Then P(E[U0|X] = 0) < 1,hence for an arbitrarily small > 0 there exists
a (, 0) (0, ) such thatE[Uexp(i.0X)exp(i.X)] 6= 0.
By continuity it follows now that for each 0 S0 there exists an > 0 such that
/ S0 for all (0 , 0) (0, 0 + ) .Consequently, in the case k = 1, the set S0 iscountable and is nowhere dense. In the general
case k 1, S0 has Lebesgue measure zero andis nowhere dense.
More generally, we have:
6
7/30/2019 The Integrated
7/23
LEMMA 5: Let w(u) be a real or complexvalued function of the type
w(u) =X
s=0
(s/s!) us
where |s| < and at most a finite numberof ss are zero. Then under the conditions of
Lemma 3, the setS0 = Rk : E[U.w (0X)] = 0
has Lebesgue measure zero and is nowhere dense.
For example, let w(u) = cos(u) + sin(u), orw(u) = exp(u).
The condition that the random vector X isbounded can be get rid of by replacing X with(X), where is a Borel measurable boundedone-to-one mapping, because the-algebra gen-
erated by X is then the same as the -algebragenerated by(X), hence conditioning on(X)
is equivalent to conditioning on X.
7
7/30/2019 The Integrated
8/23
THEOREM 1: Let U be a random variable
satisfying E[|U|] < and let X Rk be arandom vector. Denote
S = Rk : E[U.w (0(X))] = 0
,
where w(.) is defined in Lemma 5, and (.) isa Borel measurable bounded one-to-one map-
ping. If P [E(U|X) = 0] < 1 then S has
Lebesgue measure zero and is nowhere dense,
whereas if P [E(U|X) = 0] = 1 then S =R
k.
2 The ICM test
Given a random sample (Yj, Xj), j = 1, . . ,n,Xj Rk, and a conditional expectation model
E(Yj|Xj) = g(Xj, 0), 0 ,where Rm is the parameter space, The-orem 1 suggests to test the correctness of the
functional specification of this model on thebasis of following ICM statistic:
8
7/30/2019 The Integrated
9/23
Z|bz()|2 d ()In this expression, () is an absolutely con-tinuous (w.r.t. Lebesgue measure) probability
measure with compact support Rk, and
bz() =
1n
n
Xj=1bUjw (
0(Xj)) .
where bUj = Yjg(Xj,b), withb the NLLS es-timator of0, is a bounded one-to-one map-ping, and w(.) is a weight function satisfyingthe conditions of Theorem 1.
More formally, the null hypothesis to be tested
is thatH0: There exists a 0 such thatP [E(yj|xj) = g(xj, 0)] = 1,
and the alternative hypothesis is that H0 is false:H1: For all ,
P [E(yj|xj) = g(xj, )] < 1,
9
7/30/2019 The Integrated
10/23
Under the null hypothesis and standard reg-
ularity conditions,
nb 0 = A1 1
n
nXj=1
g(Xj, )
0
=0
Uj
+op(1)
where
A = p limn 1n
nXj=1
g(Xj, )0
g(Xj, )0
0=0
Hence, by the uniform law of large numbers,
bz() =
1n
nXj=1
bUjw (
0(Xj))
= 1n
nXj=1
Ujw (0(Xj))
1n
nXj=1
g(Xj,b) g(Xj, 0)w (0(Xj))
=
1
n
n
Xj=1 Ujj () + op(1),10
7/30/2019 The Integrated
11/23
say, where
j () = w (0(Xj))
b ()0 A1 g(Xj, )0
=0
with
b () = p limn
1
n
n
Xj=1g(Xj, )
0 =0 w (0(Xj)) ,
and op(1) is uniform in .
11
7/30/2019 The Integrated
12/23
THEOREM 2: Under the null hypothesis and
some regularity conditions (one of these con-ditions is that is compact),
bz() = 1n
nXj=1
bUjw (0(Xj)) z() on ,where z() is a zero-mean Gaussian process
on
, with covariance function(1, 2) = E[z(1)z(2)] .
Hence by the continuous mapping theorem,Z|bz()|2 d () d Z|z()|2 d () .
Under the alternative that the null is false,
bz()/n p () uniformly on , where () 6=0 except on a set with zero Lebesgue measure,so that
(1/n)
Z|bz()|2 d () p Z|()|2 d () > 0,
provided that () is absolutely continuous w.r.t.Lebesgue measure and its support has pos-
itive Lebesgue measure.
12
7/30/2019 The Integrated
13/23
3 The null distribution of
the ICM test
If we choose the weight function w real-valued,for example w(u) = cos(u)+sin(u), then z()is a real-valued zero-mean Gaussian process
on , with real-valued covariance function
(1, 2) = E[z(1)z(2)]
= limn
1
n
nXj=1
E
U2j j (1)j (2)
.
This covariance function is symmetric and pos-
itive semidefinite, in the following sense:
Z Z(1)(1, 2)(2)d (1) d (2) 0for all Lebesgue integrable functions () on. Such functions have non-negative eigenval-ues and corresponding orthonormal eigenfunc-
tions:
13
7/30/2019 The Integrated
14/23
THEOREM 3: The functional eigenvalue prob-
lem R(1, 2)(2)d (2) = .(1) a.e. on has a countable number of solutionsZ
(1, 2)j(2)d (2) = j.j(1)a.e. on ,
j = 1, 2, .....
where
j 0, Xj=1
j < ,Z
j1()j2()d ()
= 0 if j1 6= j2= 1 if j1 = j2
Moreover,
THEOREM 4 (Mercers Theorem): The co-
variance function (1, 2) can be written as(1, 2) =
Pj=1 jj(1)j(2).
14
7/30/2019 The Integrated
15/23
The sequence {t()}t=1 is an orthonormal
basis for a Hilbert space H () of Lebesgue in-tegrable functions on , with inner product
hf, gi =
Zf() g () d () .
so that every function finH () can be writtenas
f() =X
t=1
tt(),X
t=1
2t < , where
t = hf,ti , t = 1, 2, 3, .....
It can be shown that z() is a random elementofH () , hence
z() =
Xt=1
ztt(), where
zt =
Z
z()t()d () , t = 1, 2, 3,....
Consequently,
15
7/30/2019 The Integrated
16/23
Z|z()|2 d () = Z Xt=1
ztt()!2 d ()=
Xt1=1
Xt2=1
zt1zt2
Zt1()t2()d ()
=
Xt1=1
Xt2=1 zt1zt2I(t1 = t2) =
Xt=1 z2tThe sequence zt is a zero-mean Gaussian process,with variance function
E
z2t
= E
"Zz()t()d ()
2#
= Z ZE[z(1)z(2)]t(1)t(2)d (1) d (2)=
Z Z Xj=1
jj(1)j(2)
t(1)t(2)
d (1) d (2)
where the latter equality follows from Mercers
theorem.
16
7/30/2019 The Integrated
17/23
Thus by the orthonormality of the eigenfunc-
tions,
E
z2t
=X
j=1
j
Zj()t()d ()
2=
Xj=1jI(j = t) = t.
Moreover, by a similar argument it follows that
E[zt1zt2] =
Xj=1
jI(j = t1) I(j = t2)
= 0 ift1 6= t2.
Hence, denoting t = zt/t ift > 0, we
have:
THEOREM 5:R|z()|2 d () =
Pt=1 t
2
t ,where the ts are i.i.d. N(0, 1) and the ts arethe eigenvalues of the covariance function .
17
7/30/2019 The Integrated
18/23
4 Critical values
The problem is that the eigenvaluest are case-
dependent: They depend on the distribution of
the regressors, the functional form of the NLLS
model, and the conditional variance of the er-
rors. Therefore, the distribution of
R|z()|2 d ()
cannot be tabulated. A possible way to get aroundthis problem is to bootstrap this distribution.
However, a convenient way to get around this
problem is to derive upper bounds of the criti-
cal values, as follows.
Without loss of generality we may assume
that the ts are positive and arranged in de-creasing order. Moreover, it follows from Mer-
cers theorem thatZ(, )d () =
Xj=1
j
Zj()
2d ()
=X
j=1
j
18
7/30/2019 The Integrated
19/23
THEOREM 6: Denoting pt = t/Pj=1 j,we haveR|z()|2 d ()R(, )d ()
=X
t=1
pt2
t
supp1p2.....,
Pt=1
pt=1
Xt=1pt
2
t
= supm1
1m
mXi=1
2i = T ,
say,
so that asymptotic critical values can be de-
rived from the latter distribution. The actual
test statistic of the ICM test is thereforebTICM = R|bz()|2 d()Rb(, )d() ,where
b(1, 2) is a consistent estimator of(2, 2),
uniformly on .
19
7/30/2019 The Integrated
20/23
5 Local power of the ICM
test
Consider the local alternative hypothesis
HL1
: E[Yj|Xj] = g(Xj, 0) +h (Xj)
na.s.,
where h (Xj) is not constant:
P [h (Xj) = E(h (Xj))] < 1.Then under HL
1,bz() z() + () on ,
where z() is the same zero-mean Gaussian processon as before, and() is a deterministic meanfunction satisfying 0 < R()2d () < .Similar to the case under the null hypothesis,we can write
z() + () =X
t=1
ztt()
where now
zt =
tpt + twith t i.i.d. N(0, 1) andt = R()t()d () .20
7/30/2019 The Integrated
21/23
Hence,Z|bz()|2 d () d Z|z() + ()|2 d ()=
Xt=1
tpt + t
2
THEOREM 7: The ICM test has nontrivialn-local power, in the sense that for every K >
0,
P
" Xt=1
tpt + t
2 K
#
< P" Xt=1
t2t K#
21
7/30/2019 The Integrated
22/23
Proof:
Let
Cn =X
t=1
tpt + t
2npn + n
2and suppose that n 6= 0. Then
P"
Xt=1 tpt + t2
K#
= P
npn + n
2 K Cn and Cn K
= Php
K Cn npn + n
pK Cn
and Cn K
< PhpK Cn npn pK Cnand Cn K
= P
2nn + Cn K and Cn K
= P
2nn + Cn K
where the inequality is due to the symmetry
and unimodality of the N(0, n) distribution.The result of Theorem 7 follows now by in-duction.
22
7/30/2019 The Integrated
23/23
6 Bibliography
Bierens, H. J. (1982): Consistent Model Spec-
ification Tests, Journal of Econometrics 20,
105-134.
Bierens, H. J. (1984): Model Specification
Testing of Time Series Regressions, Journal
of Econometrics 26, 323-353.Bierens, H. J. (1990): A Consistent Condi-
tional Moment Test of Functional Form,Econo-
metrica 58, 1443-1458.
Bierens, H. J. and W. Ploberger (1997): As-
ymptotic Theory of Integrated Conditional Mo-
ment Tests, Econometrica 65, 1129-1151.De Jong, R. M. (1996): On the Bierens Test
Under Data Dependence, Journal of Econo-
metrics 72, 1-32.
Stinchcombe, M. B., and H. White (1998):
Consistent Specification Testing with Nuisance
Parameters Present Only Under the Alternative,
Econometric Theory 14, 295-325.
23