Upload
roger-asencios-nunez
View
217
Download
0
Embed Size (px)
DESCRIPTION
Econometria
Citation preview
Large Sample TheoryLarge Sample Results for
Consistency of S2
Large Sample Results for the Linear Model
Walter Sosa-Escudero
Econ 507. Econometric Analysis. Spring 2009
February 19, 2009
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Why Large Sample Theory?
= (X X)1X Y , then E() = . Easy, mostly due tolinearity.
What about = g(X,Y ). In many cases, impossible toderive finite sample properties without being too specificabout g(.).
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Consider H0 : = 0 vs. HA : 6= 0. We relied on checking = 0, based on the distribution of . We had to assume unormal and, through linearity, we got the distribution of .
It is equivalent to think about H0 : = 0 vs. HA : 6= 0for 6= 0 and check whether = 0. Why? If we choose cleverly, the distribution of is much easier to get than thatof . For example, if we set =
n, when n, ' N .
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Example: X (, 2), H0 : = 0, S2 = (n 1)1
(Xi X)2
X normal:
Z =X 0S2/n
a' tn1
Reject if |z| > c, with Pr(|Z| > c) = .
X not normal:
Z =X 0S2/n
' N(0, 1)
Reject if |z| > c, with Pr(|Z| > c) = .If X is normal, we use the t distribution. The result holds for anysample size.
If X is not normal, we use the normal distribution. The result holdsasymptotically.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
What do we mean by valid asymptotically?
How relevant is the previous result in practice, when n isnecesarily finite
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
X is normal
We should use the t distributionto compute c.
What if we use the normal?
The approximation works fine.For n = 30 differences arenegligible.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
X is not-normal
Here we do not haveinformation tocompute c.
What if we use thenormal?
Uniform case: theapproximation workswell!
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Large sample approximations work well in many cases, evenfar from infinity.
In most cases it is much easier to derive the asymptoticapproximations instead of the finite sample result.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Basic Concepts
Convergence in probability:A sequence {zn} of random scalars converges in probability to anon-random constant a if for every > 0:
limnPr [|zn a| > ] = 0
We use the notation znp a, or plim zn = a
It extendes naturally to sequences of random vectors ormatrices, requiring element-by-element convergence.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Almost sure convergence:A sequence {zn} of random scalars converges almost surely to aconstant a if
Pr(
limn zn =
)= 1
We use the notation zna.s. a, or plim zn = a
Almost sure convergence implies convergence in probability.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Convergence in Distribution A sequence {zn} of random scalarswith distribution cumulative distribution Fn converges indistribution to a random scalar z with distribution F
limnFn = F
at every continuity point of F .
F is the limiting distribution of {zn}.Notation: zn
d z.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Law of Large NumbersLet zn be a secuence of random scalars, and construct a newsequence
zn =n
i=1 zin
Kolmogorovs Strong LLN: {zn} and i.i.d. secuence of randomscalars with E(zi) = (exists and is finite). Then zn
a.s. .
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Central Limit Theorem
Lindeberg-Levy CLT: {zn} and i.i.d. secuence of random scalarswith E(zi) = and V (zi) = 2 (both exist and are finite). Then
nzn
d N(0, 1)
Careful, the theorem does not refer to zn but to atransformation involving n.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Useful Results
Continuity: {zn} a sequence of random vectors, z a randomvector, a a vector of constants. Let g(.) be a vector valuedcontinuous function that does not depend on n
znp a g(zn) p g(a), and zn a.s. a g(zn) a.s. g(a)
znd z g(zn) d g(z)
Product rule: znp 0 and xn d x, then znxn p 0
Slutzkys Theorem:
xnd x and yn p , then xn + yn d x+ .
xnd x, An p A, then Anxn d Ax, provided An and xn are
conformable.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Cramer Wold Device (restricted): xn a sequence of K 1random vectors. If for any vector
Large Sample TheoryLarge Sample Results for
Consistency of S2
Estimators as sequences of random variables
An estimator n is any function of the sample. The sequence{n}n refers to the sequence formed by increasing the sample sizeprogressively.
Consistency: n is consistent for 0 if n 0 (pis weak anda.s. strong consistency.
Asymptotic normality: n is asymptotically normal ifn(n 0) d N(0,).
is the asymptotic variance. Careful with this!Such estimators are usually called
n-consistent.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Large Sample Properties of OLS estimators
Aside: some new notation
Let us write our linear model
yi = 1x1i + 2x2i + + KxKi + ui, i = 1, . . . , n
as: yi = xi + ui
where xi is an K 1 vector (x1i, x2i, . . . , xKi).
Note that xixi is a K K matrix. It is easy to check
X X =n
i=1 xixi, X
Y =n
i=1 xiyi
n = (n
i=1 xixi)1 (
ni=1 xiyi)
n refers to the sequence of OLS estimators formed by increasingthe sample size.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
The Model
Assumptions for asymptotic analysis
1 Linearity: yi = xi0 + ui i = 1, . . . , n.2 Random sample: {xi, ui} is a jointly i.i.d. process.3 Predeterminedness: E(xikui) = 0 for all i and k = 1, . . . ,K.4 Rank condition: x E(xixi) finite positive definite.5 V (xiui) = S finite positive definite.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
The results
Consistency: np 0.
Asymptotic normality:n(n 0
)p N(0,1x S1x )
Plan: first prove results, then we discuss the assumptions.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Consistency
n = 0 + (X X)1X u
= 0 +(X Xn
)1(X un
)
We will show:
n1X u p 0(XXn
)1does not explode
This argument will appear several times in this course!
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
The hth element of(Xun
)is
1n
ni=1
xhiui =1n
ni=1
zi, zi xhiui
with E(zi) = 0, so by Kolmogorovs LLN
1n
ni=1
zi =n
i=1 xhiuin
p E(xhiui) = 0
hence (X un
)p 0
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
The (h, j) element of XXn is
ni=1 xhixji
n . Since E(xhixji) = x,hj ,which exists and is finte, by the LLN (element-by-element)
X Xn
p x
Since x is invertible and by continuity:(X Xn
)1p 1x
Summarizing:
n = 0 +(XXn
)1 (Xun
)p 1x p 0
Hence, by the product rule: np 0
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Asymptotic normalityThe starting point is now:
n(n 0
)=(X Xn
)1(X un
)We have already shown:(
X Xn
)1p 1x
Now the goal is to find the asymptotic distribution of(Xun
)and
use our continuity results.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
(X un
)=nX un
is a vector of K VAs. By the Cramer-Wold Device, we will findthe distribution of:
n c(X un
)for every vector c
Large Sample TheoryLarge Sample Results for
Consistency of S2
It is easy to check (do it as exersise) that the assumptions imply:
E(zi) = 0V (zi) = cSc
Large Sample TheoryLarge Sample Results for
Consistency of S2
Then: n(n 0
)=
(XXn
)1 (Xun
)p 1x d N(0, S)
By Slutzkys theorem and by the linearity property of multivariatenormal variables:
n(n 0
)p N(0,1x S1x )
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
A useful simplification
If we further assume E(u2i |xi) = 20, then using LIE:
S = V (xiui) = E(xiuiuixi) = 20E(xix
i) =
20 x
So the asymptotic variance of n reduces to
1x S1x =
201x x
1x =
201x
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Comments on the assumptions
1 Predeterminedness vs. Mean independence: E(xikui) = 0 isan orthogonality assumption. If there is a constant in themodel, E(ui) = 0 and this implies Cov(xi, ui) = 0.E(ui|xik) = 0 is a stronger assumption, since it implies thatui is uncorrelated to any (measurable) function of xi.
2 Predeterminedness vs. Strict Exogeneity: our old assumptionwas E(u|X) = 0, which implies E(uiXi) = 0. We arerequiring a bit less than we did to show unbiasedness (careful).
Example: Yt = 1 + 2Yt1 + ut, t = 1, 2, ... E(utYt1) = 0(consistency), but E(u|Y1) 6= 0 (why?). The OLS estimator willbe consistent but biased!
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
3 Rank condition as no multicollinearity in the limit: thisguarantees the invertibility of n1 X X in the limit.
4 Rank condition requires variability in X The rank conditionimplies
1nX X p E(xixi), finite pd.
Consider the following model:
yi = 0 + 11i
+ ui, i = 1, . . . , n
Here the explanatory variable provides less information asn.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
5 We are not ruling out conditional heteroskedasticity!: the iidassumption implies E(u2i ) is constant (unconditionalhomoskedasticity), but the error term can be conditionallyheteroskedastic.
6 The intercept. Note that
Cov(Xk, ui) = E(Xkiui) E(Xki)E(ui)
If there is a constant in the model, E(X1iui) = E(ui) = 0.Then, this joint with predeterminedness implies nocontemporaneous correlation between explanatory variablesand the error term.
7 Do not confuse consistency with unbiasedness.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Extensions
The iid case: No fixed regressors, heteroskedasticity,dependences.
Heterogeneity: moment restrictions (Markov, Lyapunov CLT).
Dependences: martingale difference, ergodic theorem, etc.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for
Consistency of S2
Consistency of S2
Result: Under Assumptions 1-4, s2p E(u2i ), provided E(u2i )
exists and is finite.
Proof:
s2 =ee
nK =n
nKeen
=n
nKuMun
NowuMu = u(I X(X X)1X )u
Ill let you finish the proof in the homework.
Walter Sosa-Escudero Large Sample Results for the Linear Model
Large Sample TheoryLarge Sample Results for Consistency of S2