The Integrated Conditional Moment Test

Herman J. Bierens
Abstract

In this lecture I will review the integrated conditional 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-trivial n local power. I will focus on the mathematical foundations of the ICM approach, in particular the consistency proof and the derivation of the asymptotic null distribution.

1 The Fourier transform of a Borel measurable function

Let g(x) be a Borel measurable real function on Rk . The Fourier transform of g(x) relative to a probability measure X (.) on the Borel k sets in RZ is defined by 0 () = exp (i. x) g(x)dX (x), i = 1, R provided that |g(x)|dX (x) < . LEMMA 1: Let g1(x) and g2(x) be Borel meaR k surableR functions on R satisfying |g1(x)|dX (x) < , |g2(x)|dX (x) < , with Fourier transforms 1 () , 2 () , respectively, relative to a probability measure (.) on the Borel sets in Rk . Then g1(x) = g2(x) a.s. X (.), i.e., k B0 = x R : g1(x) g2(x) = 0 X (B0) = 1, if and only if 1 () 2 () .

Proof : Suppose 1 () 2 () and X (B0) < 1. Let r1(x) = max (0, g1(x) g2(x)) , r2(x) = max (0, g1(x) + g2(x)) Then g1(x) g2(x) = r1(x) r2(x) and Z

If c = 0 then r1(x) = r2(x) = 0 a.s. (.), hence g1(x) = g2(x) a.s. (.). Therefore, assume that c > 0. Then we can define the probability measures Z 1 m (B) = rm(x)dX (x), m = 1, 2, c B with corresponding characteristic functions

= 1 () 2 () 0 Substituting = 0 yields Z Z r1(x)dX (x) =

exp (i. 0x) r1(x)dX (x) Z exp (i. 0x) r2(x)dX (x)

r2(x)dX (x) = c 0.

exp (i. 0y) dm (y) Z 1 = exp (i. 0x) rm(x)dX (x) c for m = 1, 2. But I have just established that R R 0 exp (i. x) r1(x)dX (x) exp (i. 0x) r2(x)dX (x), hence 1 () 2 () , which by the uniqueness of characteristic functions implies that 1 (B) = 2 (B) for all Borel sets B Rk . It is now an easy (ECON 501) exercise to verify that the latter implies r1(x) = r2(x) a.s. X (.), hence g1(x) = g2(x) a.s. X (.). Corollary: m () =
LEMMA 2: Let U be a random variable satisfying E [|U|] < , and let X Rk be a random vector If P [E(U|X) = 0] < 1 then there . exists a Rk such that E [U exp (i. 0X)] 6= 0. Question: Where to look for such a ?

LEMMA 3: If X is bounded then under the conditions of Lemma 2, for each > 0 there exists a satisfying kk < such that E [U exp (i. 0X)] 6= 0. Proof : Let X R. Then"
X im m X m

E [U exp (i.X)] = E U

Xi E [U.X m] = m! m=0 Since E [U exp (i.X)] 6= 0 for some we must have that E [U.X m] 6= 0 for some integer m 0. Let m0 be the smallest m for which E [U.X m] 6= 0. Then dm0 E [U exp (i.X)] = im0 E [U.X m0 ] 6= 0 m0 (d) =0 which implies that E [U exp (i.X)] 6= 0 for 6= 0 arbitrarily close to zero.

m=0 m m

m!

LEMMA 4: Under the conditions of Lemma 3, the set S0 = Rk : E [U. exp (i. 0X)] = 0 has Lebesgue measure zero and is nowhere dense. Proof : Let k = 1 and 0 S0. Define U0 = U exp (i.0X) . Then P (E[U0|X] = 0) < 1, hence for an arbitrarily small > 0 there exists a (, 0) (0, ) such that E [U exp (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 is countable and is nowhere dense. In the general case k 1, S0 has Lebesgue measure zero and is nowhere dense. More generally, we have:

where |s| < and at most a finite number of ss are zero. Then under the conditions of Lemma 3, the set k 0 S0 = R : E [U.w ( X)] = 0 has Lebesgue measure zero and is nowhere dense. For example, let w(u) = cos(u) + sin(u), or w(u) = exp(u). The condition that the random vector X is bounded can be get rid of by replacing X with (X), where is a Borel measurable bounded one-to-one mapping, because the -algebra generated by X is then the same as the -algebra generated by (X), hence conditioning on (X) is equivalent to conditioning on X .

LEMMA 5: Let w(u) be a real or complex valued function of the type X w(u) = (s/s!) us
s=0

THEOREM 1: Let U be a random variable satisfying E [|U |] < and let X Rk be a random vector Denote . k 0 S = R : E [U.w ( (X))] = 0 , where w(.) is defined in Lemma 5, and (.) is a Borel measurable bounded one-to-one mapping. 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 = Rk .

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, Theorem 1 suggests to test the correctness of the functional specification of this model on the basis of following ICM statistic:

b where Uj = Yj g(Xj , b with b the NLLS es), timator of 0, is a bounded one-to-one mapping, and w(.) is a weight function satisfying the conditions of Theorem 1. More formally, the null hypothesis to be tested is that H0: There exists a 0 such that P [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,

In this expression, () is an absolutely continuous (w.r.t. Lebesgue measure) probability measure with compact support Rk , and n 1 Xb z () = b Uj w ( 0(Xj )) . n j=1

|b()|2 d () z

Under the null hypothesis and standard regularity conditions, n X g(Xj , ) 1 b 0 = A1 n Uj 0 n

j=1 =0

+op(1)

where

n X g(Xj , ) g(Xj , ) 0 1 A = p lim 0 0 n n j=1

=0

Hence, by the uniform law of large numbers, n 1 Xb z () = b Uj w ( 0(Xj )) n j=1

n

1 X = Uj j () + op(1), n j=1
10

1 X Uj w ( 0(Xj )) = n j=1 n 1 X g(Xj , b g(Xj , 0) w ( 0(Xj )) ) n j=1

n

say, where j () = w ( 0(Xj ))

b ()0 A1
with

n X g(Xj , ) 1 b () = p lim 0 n n j=1

g(Xj , ) 0
=0

=0

w ( 0(Xj )) ,

and op(1) is uniform in .

11

THEOREM 2: Under the null hypothesis and some regularity conditions (one of these conditions is that is compact ), n 1 Xb Uj w ( 0(Xj )) z() on , z () = b n j=1 where z() is a zero-mean Gaussian process on , with covariance function (1, 2) = E [z(1)z(2)] . Hence by the continuous Z mapping theorem, Z

|b()|2 d () d z

|z()|2 d () .

provided that () is absolutely continuous w.r .t. Lebesgue measure and its support has positive Lebesgue measure.

Under the alternative that the null is false, z ()/ n p () uniformly on , where () 6= b 0 except on a set with zero Lebesgue measure, so thatZ Z (1/n) |b()|2 d () p |()|2 d () > 0, z

12

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)] n 1X 2 = lim E Uj j (1) j (2) . n n j=1 This covariance function is symmetric and positive semidefinite, in the following sense: Z Z

(1)(1, 2)(2)d (1) d (2) 0

for all Lebesgue integrable functions () on . Such functions have non-negative eigenvalues and corresponding orthonormal eigenfunctions:

13

THEOREM 3: The functional eigenvalue probR lem (1, 2)(2)d (2) = .(1) a.e. on has a countable number of solutions Z

(1, 2)j (2)d (2) = j .j (1)a.e. on ,

j = 1, 2, ..... where X j 0, j < , Z

j=1

j1 ()j2 ()d ()

= 0 if j1 6= j2 = 1 if j1 = j2

Moreover, THEOREM 4 (Mercers Theorem): The covariance function (1, 2) can be written as P (1, 2) = j=1 j j (1)j (2).

14

The sequence {t()} is an orthonormal t=1 basis for a Hilbert space H () of Lebesgue integrable functions on , with inner product Z

hf, gi =

f () g () d () .

t = hf, ti , t = 1, 2, 3, ..... It can be shown that z() is a random element of H () , hence X z() = ztt(), where t=1 Z zt = z()t()d () , t = 1, 2, 3, ....

so that every function f in H () can be written as X X f () = tt(), t2 < , where

t=1 t=1

Consequently,

15

Z = =

|z()|2 d () =
XX t1 =1 t2=1 XX t1 =1 t2=1

zt1 zt2

Z X
t=1

ztt()

!2

d ()

t1 ()t2 ()d ()
X t=1 2 zt

zt1 zt2 I (t1 = t2) =

The sequence zt is a zero-mean Gaussian process, with variance function "Z 2# 2 z()t()d () E zt = E Z Z = E [z(1)z(2)] t(1)t(2)d (1) d (2) Z Z X j j (1)j (2) t(1)t(2) =
j=1

d (1) d (2) where the latter equality follows from Mercers theorem.

16

Thus by the orthonormality of the eigenfunctions, 2 X Z 2 j ()t()d () E zt = j

= 0 if t1 6= t2. Hence, denoting t = zt/ t if t > 0, we have:

THEOREM 5: |z()| d () = t=1 t2, t where the t are i.i.d. N (0, 1) and the t are s s the eigenvalues of the covariance function .

Moreover, by a similar argument it follows that X E [zt1 zt2 ] = j I (j = t1) I (j = t2)

j=1

j=1 X j=1

j I (j = t) = t.

17

4 Critical values
The problem is that the eigenvalues t are casedependent: They depend on the distribution of the regressors, the functional form of the NLLS model, and the conditional variance of the erR rors. Therefore, the distribution of |z()|2 d () cannot be tabulated. A possible way to get around this problem is to bootstrap this distribution. However, a convenient way to get around this problem is to derive upper bounds of the critical values, as follows. Without loss of generality we may assume that the ts are positive and arranged in decreasing order. Moreover, it follows from Mercers theorem that

(, )d () = =

X j=1 X j=1

j j

j ()2d ()

18

THEOREM 6: Denoting pt = t/ we have R X |z()|2 d () R pt2 = t (, )d () t=1

j=1 j ,

say,

p1 p2 ....., pt =1 t=1 t=1 m 1 X 2 sup = T, m i=1 i m1

sup P

pt2 t

so that asymptotic critical values can be derived from the latter distribution. The actual test statistic of the ICM test is therefore R |b()|2 d() z b , TICM = R b )d() (,

b where (1, 2) is a consistent estimator of (2, 2), uniformly on .

19

5 Local power of the ICM test

Consider the local alternative hypothesis h (Xj ) L H1 : E[Yj |Xj ] = g(Xj , 0) + a.s., n where h (Xj ) is not constant: P [h (Xj ) = E (h (Xj ))] < 1. L Then under H1 , z () z() + () on , b where z() is the same zero-mean Gaussian process on as before, and () R a deterministic mean is function satisfying 0 < ()2d () < . Similar to the case under the null hypothesis, we can write X z() + () = ztt()
t=1

where now

zt = t t + t R with t i.i.d. N (0, 1) and t = ()t()d () .

20

Z Hence, |b()|2 d () z

2 X p = t t + t t=1

|z() + ()|2 d ()

THEOREM 7: The ICM test has nontrivial n-local power in the sense that for every K > , 0, " # 2 X p P t t + t K

<P

t=1 "

X t=1

t2 K t

21

Proof : Let 2 p 2 X p t t + t n n + n Cn =
t=1

and suppose that n 6= 0. Then # " 2 X p t t + t K P

t=1

2 p = P n n + n K Cn and Cn K h p p p = P K Cn n n + n K Cn and C K h p n p p < P K Cn n n K Cn and Cn K 2 = P nn + Cn K and Cn K 2 = P nn + 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 induction.
22

6 Bibliography
Bierens, H. J. (1982): Consistent Model Specification 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 Conditional Moment Test of Functional Form, Econometrica 58, 1443-1458. Bierens, H. J. and W. Ploberger (1997): Asymptotic Theory of Integrated Conditional Moment Tests, Econometrica 65, 1129-1151. De Jong, R. M. (1996): On the Bierens Test Under Data Dependence, Journal of Econometrics 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