Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Large-Sample Based Inference in the Linear Model
Walter Sosa-Escudero
Econ 507. Econometric Analysis. Spring 2009
February 26, 2009
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Motivation
In the classical linear model with finite n, we had to introduce
distributional assumptions to perform basic hypothesis tests
and confidence intervals.
A really important advantage of the large sample theory is to
provide a framework for inference, without distributional
assumptions.
We will derive simple test statistics to evaluate linear
hypothesis in the linear model.
We will discuss a simple strategy to handle the non-linear case.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Our starting point will be the asympototic normality result:
p
n n 0 N (0, 1 S1 )
x
where x = E(xi x0i ) and S = V (xi ui ).
Note that n1 Xi0 Xi is a consistent estimator for x .
Recall that our assumptions for asymptotic analysis allow for
conditional heteroskedasticity. Alternative consistent
estimators for S depend on what we are willing to assume on
this.
1
We will use the notation AV (n ) 1
x Sx and
d (n ) =
1 S
1 .
AV
x
x
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Road Map
Road Map
1
Derive general test statistics assuming there is available a
p
consistent estimator S S
Derive consistent estimators for S with and without
conditional heteroskedasticity.
Obtain particular cases of the previous tests for the
conditional homoskedastic case.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Test for Linear Hypothesis under general S
Hypothesis about single coefficients:
H0 : j = j0 vs. HA : j 6= j0
From the asymptotic normality, using Slutzkys theorem, it easy to
check that when H0 : j = j0 holds:
n j j0 d
tj = q
N (0, 1)
d (j )
AV
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Multiple Linear Hypothesis
H0 : R r = 0, R is a q K matrix with (R) = q, and r <q
Result: under all the assumptions for asymptotic analysis, and
p
S S, when H0 : R r = 0 holds:
h
i1
d
d (n ) R0
W = n (Rn r)0 R AV
(Rn r) 2 (q)
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Proof: by asympototic normality, under H0
d
n(Rn r) N 0, R AV(n ) R0
or
h
i1/2
d
n(Rn r) N (0, Ik )
R AV(n ) R0
By Slutzkys theorem and linearity:
h
i1/2
d
c n ) R0
R AV(
n(Rn r) N (0, Ik )
The desired result follows by forming the quadratic form.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Estimation of S
A general estimator
S = E(xi ui ui x0i ) = E(u2i xi x0i )
We will need an additional assumption
Assumption 6 (fourth moments): E (xik xij )2 exists and is finite
for all k, j = 1, 2, . . . , K.
Result
1X 2 2 p
Sw
e i xi xi S
n
i=1
where ei are the OLS residuals.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Proof:
ei = yi x0i = yi x0i x0i ( ) = ui x0i ( )
e2i = u2i 2ui x0i ( ) + ( )0 xi x0i ( )
Replacing
n
1X 2 0
e i xi xi =
n
i=1
i=1
i=1
1X 2 0 2X
ui xi xi
ui x0i ( )xi x0i +
n
n
n
1X
( )0 xi x0i ( )xi x0i
+
n
i=1
First note that
1X 2 0 p
ui xi xi E(u2i xi x0i )
n
i=1
by Kolmogorovs LLN, since we assumed finite second moments
(expectaction exists) and iid.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
We will show the other two terms converge to zero
I) A =
2
n
Pn
i=1 ui
x0i ( )xi x0i
"K
#
n
X
2X
ui
xik (k k ) xi x0i
A =
n
i=1
k=1
Pn
K
0
X
u
x
x
x
i
i
ik
i
i=1
= 2
(k k )
n
k=1
Note k k 0, by consistency. So if we can show
we are done.
Walter Sosa-Escudero
h i
< ,
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
1 Pn
n
i=1 ui xik
xi x0i is a K K matrix with typical (h, j) element:
Pn
i=1 ui xik xih xij
n
By the Cauchy-Schwartz inequality:
1/2
1/2
E|xik xih xij ui | E |xik xih |2
E |xij ui |2
Both factors in the RHS are < , by our fourth moments
assumption 5 and by assumption 3. Hence, we can use the LLN:
n
1X
p
ui xik xi x0i E(ui xik xi x0i ) < ,
n
i=1
so by the product rule and continuity, A 0.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
II) B =
1
n
Pn
)0 xi x0 ( ) xi x0
i
i
i=1 (
Using the same trick as before:
#" K
"K
#
n
X
1X X
xik (k k )
xik0 (k0 k0 ) xi x0i
B=
n
0
i=1
k=1
k =1
Now we have a sum of K 2 matrices. The (h, j) element of the
k, k 0 summand will be
n
1X
(k k )(k0 k0 )
xik xik0 xih xij
n
i=1
Using again the Cauchy Schwartz inequality and the finite fourth
moments assumption: E|xik xik0 xih xij | <
p
And again, by consistency and LLN, B 0. q.e.d.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Then, using Sw as an estimator for S and noting
n
1X 2 2
1
Sw =
ei xi xi = (X 0 BX)
n
n
i=1
with B diag(e21 , . . . , e2n ),
c w (n ) =
1
1
AV
x Sw x
= n(X 0 X)1 n1 (X 0 BX)n(X 0 X)1
= n(X 0 X)1 (X 0 BX)(X 0 X)1
This is Whites heteroskedasticity consistent estimator for the
asymptotic variance of n . Remember that in the derivation of all
result we never ruled out the possibility of conditional
heteroskedasticity, then its consistency does not depend on it.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Returns-to-scale revisited:
. reg ltc lq lpl lpf lpk
-----------------------------------------------------------------------------ltc |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------lq |
.7209135
.0174337
41.35
0.000
.6864462
.7553808
lpl |
.4559645
.299802
1.52
0.131
-.1367602
1.048689
lpf |
.4258137
.1003218
4.24
0.000
.2274721
.6241554
lpk | -.2151476
.3398295
-0.63
0.528
-.8870089
.4567136
_cons | -3.566513
1.779383
-2.00
0.047
-7.084448
-.0485779
-----------------------------------------------------------------------------. reg ltc lq lpl lpf lpk, robust
-----------------------------------------------------------------------------|
Robust
ltc |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------lq |
.7209135
.0325376
22.16
0.000
.656585
.785242
lpl |
.4559645
.260326
1.75
0.082
-.0587139
.9706429
lpf |
.4258137
.0740741
5.75
0.000
.2793653
.5722622
lpk | -.2151476
.3233711
-0.67
0.507
-.8544698
.4241745
_cons | -3.566513
1.718304
-2.08
0.040
-6.963693
-.1693331
------------------------------------------------------------------------------
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Variance estimation under conditional homoskedasticity
If we further assume E(u2i |xi ) = 02 , then using LIE:
S = E(u2i xi x0i ) = 02 E(xi x0i ) = 02 x
So the asymptotic variance of n reduces to
1
2 1
1
2 1
1
x Sx = 0 x x x = x
So a consistent estimator for AV(n ) can be
c h = n s2 (X 0 X)1
AV
which is n times the finite sample estimator for the variance of n
in the classical linear model.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Tests under conditional homoskedasticity
Under conditional homoskedasticity AV(n ) = 2 1
x and
2
0
1
c
AVh = n s (X X) provides a consistent estimator:
a) Single linear hypothesis: Our test for H0 : j = j0
n j j0
tj = q
d (j )
AV
reduces to:
j j0
tj = p
s2 ajj
where ajj is the jth element of the main diagonal of X 0 X, so we
get our old t test. Note that we do not get the t distribution in
finite samples, but instead we get the N (0, 1) distribution
asymptotically.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Multiple linear hypothesis
Our W statistic
h
i1
d (n ) R0
W = n (Rn r)0 R AV
(Rn r)
simplifies to:
i1
h
d (n ) R0
(Rn r)
= n (Rn r)0 R AV
1
= (Rn r)0 R s2 (X 0 X)1 R0
(Rn r)
= qF
This provides an approximation of our old F statistic for the
asymptotic case.
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Summary of results
We derived general test for single coefficient (t) and multiple
(W ) hypothesis, in our asympotic framework that does not
assume normality and allows for conditional heteroskedasticity.
c w estimator provides a general consistent strategy
Whites AV
for estimating the AVAR of n , robust to the presence of
conditional heteroskedasticity.
Under the conditional homoskedasticity assumption, our the
general t test simplifies to the old one, but this is an
asympotic result.
W simplifies to nF .
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
The Delta Method
Suppose we want to perform inference about a non-linear function
of , say a().
Example 1
yi = 0 + 1 south + zi0 2 + ui
yi is log-wages, south is a dummy indicating if the person lives in the
southern region zi is a vector of control variables.
The percent difference between south/not south is given by
e1 1
and for small values of 1 is very similar to 1 . Suppose we are interested
exactly. A consistent estimator is e1 1. A natural problem is how to
construct a confidence interval for .
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Example 2: consider now
yi = 1 exper + 2 exper2 + zi0 + ui
where exper is work experience in years. The level of experience
that maximizes expected wages is:
1
22
and a consistent estimate is provided by
=
1
22
How can we construct an estimate for the standard deviation of a
confidence interval for ?
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Result (Delta Method): suppose xn is a sequence of random
vector of dimension K such that
p
d
xn and n(xn ) Z
and a(x) : <K <r is a function with continuous derivatives
A()
a()
0
(note A() is an r K matrix).
Then:
d
n a(xn ) a() A()Z
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Proof: Take a first-order mean value expansion of a(xn ) around :
a(xn ) = a() + A(yn ) (xn )
where the mean value yn is a vector between xn and .
From this, get
n a(xn ) a() = A(yn ) (xn )
p
Now yn (why?) so A(yn ) A() by continuous
differentiability.
Then, by the hypothesis of the theorem and Slutzkys Theorem
d
n a(xn ) a() A()Z
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
As a simple corollary note that if
then
d
n(n ) N (0, AV(n ))
d
n a(n ) a() N 0, A()AV(n )A()0
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Example 1 (Blackburn and Neumark, 1992, also in Wooldridge,
2002)
yi = 0 + 1 south + zi0 2 + ui
a(1 ) = e1 1
with
A(1 ) = e1
So, according to the delta-method
h i2
c e1 1 = e1 AV(1 )
AV
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
. reg lwage exper tenure married black south urban educ
-----------------------------------------------------------------------------lwage |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------exper |
.014043
.0031852
4.41
0.000
.007792
.020294
tenure |
.0117473
.002453
4.79
0.000
.0069333
.0165613
married |
.1994171
.0390502
5.11
0.000
.1227801
.276054
black | -.1883499
.0376666
-5.00
0.000
-.2622717
-.1144281
south | -.0909036
.0262485
-3.46
0.001
-.142417
-.0393903
urban |
.1839121
.0269583
6.82
0.000
.1310056
.2368185
educ |
.0654307
.0062504
10.47
0.000
.0531642
.0776973
_cons |
5.395497
.113225
47.65
0.000
5.17329
5.617704
-----------------------------------------------------------------------------. nlcom exp(_b[south])-1
lwage |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------_nl_1 | -.0868943
.0239677
-3.63
0.000
-.1339315
-.0398571
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Example 2:
yi = 1 exper + 2 exper2 + zi0 + ui ,
a(1 , 2 ) = 1 /(22 )
. regress lwage edup edusi edus eduui eduu exper exper2 if muest==1
-----------------------------------------------------------------------------lwage |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------edup |
.2104513
.0629835
3.34
0.001
.0869123
.3339903
edusi |
.4148728
.0678469
6.11
0.000
.2817946
.547951
edus |
.7587112
.0695764
10.90
0.000
.6222406
.8951817
eduui |
1.018209
.077569
13.13
0.000
.866061
1.170356
eduu |
1.560496
.0769774
20.27
0.000
1.409509
1.711483
exper |
.0283668
.0071065
3.99
0.000
.0144279
.0423058
exper2 | -.0002502
.0001509
-1.66
0.098
-.0005462
.0000458
_cons |
.2130178
.0934142
2.28
0.023
.0297906
.3962449
-----------------------------------------------------------------------------. nlcom -_b[exper]/(2*_b[exper2])
_nl_1:
-_b[exper]/(2*_b[exper2])
-----------------------------------------------------------------------------lwage |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------_nl_1 |
56.6962
20.7191
2.74
0.006
16.05673
97.33567
------------------------------------------------------------------------------
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model
Tests for General S
Estimation of S
Tests under conditional homoskedasticity
The Delta Method and Non-Linear Hypothesis
Summary of variance results
Starting point:
p
n n 0 N (0, AV (n ))
1
0
2
0
AV (n ) = 1
x Sx , x = E(xi xi ), S = V (xi ui ) = E(ui xi xi ).
0
1
x = X X. S is any consistent estimator of S.
d (n ) =
1
1
AV
x S x
Tests
H0 : j = j0 , use tj
h
i1
d (n ) R0
H0 : R r = 0, use W = n (Rn r)0 R AV
(Rn r)
Allowing for conditional heteroskedasticity
P
2
2
0
1
Sw = n1 n
i=1 ei xi xi = n (X BX)
c w (n ) =
1
1
AV
x Sw x
Under homoskedasticity
1
2 1
S = E(u2i xi x0i ) = 02 E(xi x0i ) = 02 x , so AV (n ) = 1
x Sx = 0 x
d h (n ) = n s2 (X 0 X)1
AV
Walter Sosa-Escudero
Large-Sample Based Inference in the Linear Model