The Uniform Weak Law of Large Numbers and the Consistency of M-Estimators of Cross-Section and Time Series Models

Transcription

1 The Uniform Weak Law of Large Numbers and the Consistency of M-Estimators of Cross-Section and Time Series Models Herman J. Bierens Pennsylvania State University September 16, The uniform weak law of large numbers In econometrics we often have to deal with sample means of random functions. A random function is a function that is a random variable for each fixed value of its argument. In crosssection econometrics random functions usually take the form of a function g(z,2) of a random vector Z and a non-random vector 2. For example, consider a Logit model: P[Y j ' y X j ] ' y % (1 & y)exp(&α & βt X j ) 1 % exp(&α & β T X j ), y ' 0,1, where Y j 0 {0,1} is the dependent variable and X j 0ú k is a vector of explanatory variables. Denoting Z j ' (Y j,x T j ) T, and given a random sample {Z 1,Z 2,...,Z n }, the log-likelihood function involved takes the form ' n j'1 g(z j,θ), where g(z j,θ) ' ln Y j % (1 & Y j )exp(&α & β T X j ) & ln 1 % exp(&α & β T X j ) ' Y j (α % β T X j ) & ln 1 % exp(α % β T X j ),whereθ ' (α,β T ) T. (1) For such functions we can extend the weak law of large numbers for i.i.d. random variables to a Uniform Weak Law of Large Numbers (UWLLN): Theorem 1: Let Z j, j = 1,..,n, be a random sample from a k-variate distribution. Let g(z,2) be a Borel measurable function on Ζ Θ, where Ζ dú k is a Borel set such that P[ 0 Ζ] ' 1, and Θ is a compact subset of ú m, such that for each z 0 Ζ, g(z,θ) is a continuous function on Θ. Furthermore, let 1

2 E[ *g(z j,θ)*] <4. (2) Then plim n64 *(1/n)' n j'1 g(z j,θ) & E[g(Z 1,θ)]* ' 0. Note that subsets of Euclidean spaces are compact if and only if they are closed and bounded. See, for example, Bierens (2004), Appendix II, Theorem II.2. The original proof of the stronger result *(1/n)' n j'1 g(z j,θ) & E[g(Z 1,θ)]* 60a.s., was given in the seminal paper of Jennrich (1969). This proof is explained in detail in Bierens (2004, Appendix to Chapter 6). The condition that the random vectors Z j are i.i.d. can be relaxed, because the result in Theorem 1 also holds for strictly stationary time series processes with a vanishing memory: Definition 1: A (vector) time series process X t 0ú k is strictly stationary if for arbitrary integers m 1 < m 2 <...< m n the joint distribution of X T t&m 1,...,X T does not depend on the time index t. t&m n T Definition 2: A (vector) time series process X t 0ú k has a vanishing memory if all the sets in the remote F-algebra ö &4 ' _ t σ {X t&j } 4 j'0 have either probability zero or one. Note that if the X t s are independent then by Kolmogorov s zero-one law the time series X t has a vanishing memory. It has been shown in Bierens (2004, Theorem 7.4) that Theorem 2: If X t 0ú k is a strictly stationary time series process with vanishing memory, and E[ X t ] < 4, then plim n64 (1/n)' n t'1 X t ' E[X 1 ]. I will use this result to prove the following more general version of Theorem 1. To be able to generalize the UWLLN to the time series case where the random functions involved depend on 2

3 the entire past of the time series rather than on a finite dimensional vector of variables, I will reformulate and prove Theorem 1 under slightly different moment conditions. Theorem 3: Let 0ú k be a strictly stationary vector time series process with a vanishing memory, 1 defined on a common probability space {Ω,ö,P}. Let g(z,2) be a Borel measurable real function on Ζ Θ 0, where Ζ dú k is a Borel set such that P[ 0 Ζ] ' 1, and Θ 0 is an open subset of ú m, such that for each z 0 Ζ, g(z,θ) is a continuous function on Θ 0. Furthermore, let 1 be a compact subset of Θ 0. Finally, assume that for each θ ( 0 Θ there exists an arbitrary small * > 0, possibly depending on θ (, such that E sup θ&θ( # δ g(z 1,θ) < 4, E inf θ&θ ( # δ g(z 1,θ) > &4. (3) Then plim n64 *(1/n)' n j'1 g(z j,θ) & E[g(Z 1,θ)]* ' 0. Proof: Observe from condition (3) that for each θ 0 Θ, E[g(Z 1,θ)] is well-defined. Actually, due to the compactness of 1, (3) implies (2) [Exercise: Why?], so that the latter is a weaker condition than (3). Moreover, it follows from condition (3), the continuity of g(z,θ) in 2, and the dominated convergence theorem, that lim δ90 E sup θ&θ( # δ g(z 1,θ) & inf θ&θ ( # δ g(z 1,θ) ' 0, (4) pointwise in θ ( 0 Θ. Therefore, for an arbitrary g > 0 and each θ ( 0 Θ we can choose a positive number δ(θ (,g) such that, with we have N(θ ( g) ' {θ 0 Θ 0 : θ & θ ( < δ(θ (,g)}, (5) 0 # E sup θ 0 N(θ( g) g(z 1,θ) & inf θ 0 N(θ ( g) g(z 1,θ) < g. (6) Next, observe that the sets (5) are open, so that is an open covering of Θ. ^θ( 0Θ N(θ ( g) Then by the compactness of Θ there exists a finite sub-covering of Θ: 1 Which includes the case that the s are i.i.d. 3

4 Θ d ^ K i'1 N(θ i g), (7) where K and the vectors θ i 0 Θ depend on g. Using the easy inequality for each θ i 0 Θ, sup x *f(x)* # sup x f(x) % inf x f(x), it is not hard to verify that sup θ0n(θi g) *(1/n)'n t'1 g(,θ) & E[g(Z 1,θ)* # 2 (1/n)' n t'1 sup θ0n(θ i g) g(,θ) & E[sup θ0n(θ i g) g(z 1,θ)] % 2 (1/n)' n t'1 inf θ0n(θ i g) g(,θ) & E[inf θ0n(θ i g) g(z 1,θ)] (8) % 2 E[sup θ0n(θi g) g(z 1,θ)] & E[inf θ0n(θ i g) g(z 1,θ)]. It follows from Theorem 2 that the first two terms at the right-hand side of (8) converge in probability to zero, and from (6) that the last term is less than 2.g. Hence, *(1/n)' n t'1 g(,θ) & E[g(Z 1,θ)* # max 1#i#K sup θ0n(θi g) *(1/n)'n t'1 g(,θ) & E[g(Z 1,θ)* (9) # R n (g) % 2.g, where plim n64 R n (g) ' 0. Theorem 3 follows now straightforwardly from (9). Q.E.D. In time series econometrics there are quite a few cases where we need a UWLLN for functions g(.,θ) depending on &j for all j $ 0. In that case g(.,θ) takes a more general form as a random function: Definition 3: Let {S,ö,P} be the probability space. A random function f(2) on a subset 1 of a Euclidean space is a mapping f(ω,θ): Ω Θ 6úsuch that for each Borel set B in ú and each 2 0 1, {ω 0 Ω: f(ω,θ) 0 B} 0ö. 4

5 Definition 4: A random function f(θ) on a subset Θ of a Euclidean space is almost surely continuous on Θ if there exists a set A with probability one such that for each continuous in θ 0 Θ. ω 0 A, f(ω,θ) is For example, let 0ú MA(1) process]: be a stationary Gaussian moving average process or order 1 [alias an ' U t & α 0 U t&1, α 0 <1,U t - i.i.d. N(0,σ 2 0 ). (10) Then backwards substitution of U t ' α 0 U t&1 % yields U t ' ' 4 j'0 αj 0 &j, hence '&' 4 j'1 αj 0 &1 % U t (11) Thus, denoting ö t ' σ(u t,u t&1,u t&2,...), the distribution of conditional on ö t&1 is normal with conditional expectation &' 4 j'1 αj 0 &1 and conditional variance σ 2 0. If the s were observable for all t # n, a version of the log-likelihood would take the form ' n j'1 g t (θ), where g t (θ) '& 1 2σ 2 '4 j'0 αj 2 1 &j & 2 ln(σ2 ) & ln 2π, θ ' (α,σ 2 ) T, (12) is a random function. In that case we need to reformulate Theorem 3 as follows. Theorem 4: Let ö t ' σ(u t,u t&1,u t&2,...), where U t is a time series process with vanishing memory. Let g t (θ) be a sequence of a.s. continuous random function on an open subset Θ 0 of a Euclidean space, and let Θbe a compact subset of Θ 0. If for each θ ( 0 Θ there exists an arbitrarily small * > 0 such that (a) g t (θ ( ), sup θ&θ( #δ g t (θ) and inf θ&θ are measurable and strictly stationary, ( #δ g t (θ) ö t (b) E[sup θ&θ( #δ g 1 (θ)] < 4, E[inf θ&θ ( #δ g 1 (θ)] > &4, then plim n64 (1/n)' n t'1 g t (θ) & E[g 1 (θ)] ' 0. 5

6 2. Consistency of M-estimators Theorems 3 and 4 are important tools for proving consistency of parameter estimators. A large class of estimators are obtained by maximizing or minimizing an objective function of the form (1/n)' n t'1 g t (θ), for example maximum likelihood estimators or nonlinear least squares estimators. These estimators are called M-estimators (where the M indicates that the estimator is obtained by Maximizing or Minimizing a Mean of random functions). Suppose that the conditions of Theorem 4 are satisfied, and that the parameter vector of interest is θ 0 ' argmax θ0θ E[g 1 (θ)]. (13) Note that "argmax" is a short-hand notation for the argument for which the function involved is maximal. Then it seems a natural choice to use ˆθ ' argmax θ0θ (1/n)' n t'1 g t (θ) (14) as an estimator of θ 0. Indeed, under some mild conditions the estimator involved is consistent: Theorem 5: (Consistency of M-estimators) Let ˆθ = argmax θ0θ ˆQ(θ) and θ 0 = argmax θ0θ Q(θ), where ˆQ(θ) = (1/n)' n t'1 g t (θ) and Q(θ) ' E[ ˆQ(θ)] ' E[g 1 (θ)]. If 2 0 is unique then under the conditions of Theorem 4, plim n64ˆθ ' θ 0. Proof: Since a continuous function on a compact set takes its maximum value in this set [see, for example, Bierens (2004, Appendix II )], it follows that ˆθ 0 Θ and θ 0 0 Θ. Moreover, by the same result it follows from the continuity of every g > 0 for which the set {θ 0 Θ: θ&θ 0 $ g} is non-empty, Q(θ) and the uniqueness of 2 0 that for Q(θ 0 )>, 2θ&θ0 2$g Q(θ) (15) [Exercise: Why?] Now by the definition of 2 0, 6

7 0 # Q(θ 0 ) & Q(ˆθ) ' Q(θ 0 ) & ˆQ(θ 0 ) % ˆQ(θ 0 ) & Q(ˆθ) # Q(θ 0 ) & ˆQ(θ 0 ) % ˆQ(ˆθ) & Q(ˆθ) # 2. * ˆQ(θ) & Q(θ)*, (16) and it follows from Theorem 4 that the right-hand side of (16) converges in probability to zero. Thus: plim n64 Q(ˆθ) ' Q(θ 0 ). (17) Moreover, (15) implies that for arbitrary g > 0 there exists a * > 0 such that Q(θ 0 ) & Q(ˆθ) $ δ if 2ˆθ & θ 0 2$g, hence P 2ˆθ & θ 0 2 > g # P Q(θ 0 ) & Q(ˆθ) $ δ. (18) Combining (17) and (18), the theorem under review follows. Q.E.D. It is easy to verify that Theorem 5 carries over to the "argmin" case. References Bierens, H. J. (2004): Introduction to the Mathematical and Statistical Foundations of Econometrics, Cambridge University Press, Cambridge, U.K. Jennrich, R. I. (1969): Asymptotic Properties of Non-Linear Least Squares Estimators, Annals of Mathematical Statistics 40,