Professional Documents
Culture Documents
Bruce E. Hansen
University of Wisconsin
www.ssc.wisc.edu/~bhansen
1
This manuscript may be printed and reproduced for individual or instructional use, but may not be
printed for commercial purposes.
Contents
2 Regression Models 23
2.1 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Bias and Variance of OLS estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Multicollinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Forecast Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 NonLinearity in Regressors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.6 NonLinear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.7 Normal Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.8 Least Absolute Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Model Selection 39
3.1 Omitted Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Irrelevant Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Testing for Omitted NonLinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
i
3.5 log(Y ) versus Y as Dependent Variable . . . . . . . . . . . . . . . . . . . . . . . . 46
6 Empirical Likelihood 63
6.1 Non-Parametric Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2 Asymptotic Distribution of EL Estimator . . . . . . . . . . . . . . . . . . . . . . . 65
6.3 Overidentifying Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.5 Numerical Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.1 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5.2 Inner Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.5.3 Outer Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7 Endogeneity 71
7.1 Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.2 Reduced Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.3 Identi cation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.5 Special Cases: IV and 2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.6 Bekker Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.7 Identi cation Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
ii
8 The Bootstrap 81
8.1 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
8.2 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
8.3 The Empirical Distribution Function . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.4 De nition of the Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.5 Bootstrap Estimation of Bias and Variance . . . . . . . . . . . . . . . . . . . . . . 89
8.6 Percentile Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
8.7 Percentile-t Equal-Tailed Interval . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.8 Symmetric Percentile-t Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
8.9 Asymptotic Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8.10 One-Sided Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.11 Symmetric Two-Sided Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.12 Percentile Con dence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
8.13 Bootstrap Methods for Regression Models . . . . . . . . . . . . . . . . . . . . . . . 99
8.14 Bootstrap GMM Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
iii
11 Limited Dependent Variables 121
11.1 Binary Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11.2 Count Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
11.3 Censored Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
11.4 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
13 Nonparametrics 132
13.1 Kernel Density Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
13.2 Asymptotic MSE for Kernel Estimates . . . . . . . . . . . . . . . . . . . . . . . . . 134
iv
17.5 Asymptotic Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
v
Chapter 1
1.1 Framework
An econometrician has observational data
1
It is unimportant whether the observations yi and xi may come from continuous or discrete
distributions. For example, many regressors in econometric practice are binary, taking on only the
values 0 and 1, and are typically called dummy variables.
A linear regression model for yi given xi takes the form
where 1 through k are parameters and ei is the error. The parameter vector is written as
0 1
1
B C
B 2 C
= B . C:
@ .. A
k
The model is incomplete without a description of the error ei . It should be mean zero, a nite
variance, and be uncorrelated with the regressors. We state the needed conditions here.
Assumption 1.1.1
1. E(ei ) = 0
2. E (xi ei ) = 0
3. 2 = Ee2i < 1
4. Ex0i xi < 1
Assumptions 1.1.1.3 and 1.1.1.4 are made to guarantee that all variables in the model have a
nite variance. This is necessary to ensure that E (xi ei ) is well de ned. Indeed by the Cauchy-
Schwarz inequality,
E jxi ei j E jxi j2 E jei j2 < 1
under these assumptions.
We can use Assumption 1.1.1 to derive a moment representation for the parameter vector :
Take equation (1.2) and pre-multiply by xi
xi yi = xi x0i + xi ei :
2
Now take expectations:
E (xi yi ) = E xi x0i + E (xi ei )
= E xi x0i
where the second equality is Assumption 1.1.1.2. Since E (xi x0i ) is invertible by Assumption 1.1.1.5,
we can solve for :
1
= E xi x0i E (xi yi ) : (1.3)
Thus the parameter is an explicit function of population second moments of (yi ; xi ):
In fact, this derivation shows that if is de ned by (1.3), then Assumption 1.1.1.2 must hold
true by construction. In this sense, Assumption 1.1.1.2 is very weak. However, it is important
to not misinterpret this statement. In many economic models, the parameter may be de ned
within the model, rather than by construction as in (1.3). In this case (1.3) may not hold. These
structural models require alternative estimation methods, and are discussed in Chapter 5.
To emphasize this distinction, we may describe the model of this section as a linear projection
model rather than a linear regression model. This is an accurate label as the equation (1.3) shows
that ei is explicitly de ned as a projection error. However, conventional econometric practice
labels (1.2) as a linear regression model, so we will adhere to this convention.
1.2 Estimation
Equation (1.3) writes the regression parameter as an explicit function of population moments
E (xi yi ) and E (xi x0i ) : Their moment estimators are the sample moments
n
^ (xi yi ) = 1X
E xi yi
n
i=1
n
^ xi x0i 1X
E = xi x0i :
n
i=1
It follows that the moment estimator of is (1.3) with the population moments replaced by the
sample moments:
n
! 1 n
1X 1X
= xi x0i xi yi
n n
i=1 i=1
n
! 1 n
X X
= xi x0i xi yi : (1.4)
i=1 i=1
Another way to derive ^ is as follows. Observe that Assumptions 1.1.1.2 can be written in the
parametric form
E xi yi x0i = 0: (1.5)
3
The function E (xi (yi x0i )) can be estimated by
n
^ xi yi 1X
E x0i = xi yi x0i
n
i=1
The Ordinary Least Squares (OLS) estimator is the value of which minimizes Sn ( ): Observe
that we can write the latter as
n
X n
X n
X
Sn ( ) = yi2 2 0
xi yi + 0
xi x0i
i=1 i=1 i=1
Vector calculus (see section 1.1.13) gives the rst-order conditions for minimization:
@
0 = Sn ( ^ )
@
Xn n
X
= 2 xi yi + 2 xi x0i ^
i=1 i=1
whose solution is (1.4). Following convention, we will call ^ the OLS estimator of :
As a by-product of OLS estimation, we de ne the predicted value
y^i = x0i ^
e^i = yi y^i
= yi x0i ^ :
4
Note that yi = y^i + e^i :It is important to understand the distinction between the error ei and the
residual e^i : The error is unobservable, while the residual is a by-product of estimation. These two
variables are frequently mislabeled, which can cause confusion.
Equation (1.6) implies that
n
1X
xi e^i = 0:
n
i=1
Thus the sample correlation between the regressors and the residual is zero. Furthermore, since
xi (typically) contains a constant, one implication is that
n
1X
e^i = 0:
n
i=1
Thus the residuals have a sample mean of zero. These are algebraic results, and hold true for all
linear regression estimates.
The error variance 2 is also a parameter of interest. A method of moments estimator for it is
the sample average
n
1X 2
^2 = e^i :
n
i=1
1.3 E ciency
Is the OLS estimator e cient, in the sense of achieving the smallest possible mean-squared error
among feasible estimators? The answer was a rmatively provided by Chamberlain (1987).
Suppose that the joint distribution of (yi ; xi ) is discrete. That is, for nite r;
P yi = j; xi = j = j; j = 1; :::; r
5
for some constant vectors j ; j ; and j : Assume that the j and j are known, but the j are
unknown. (We know the values yi and xi can take, but we don't know the probabilities.)
In this discrete setting, the moment condition (1.5) can be rewritten as
r
X
0
j j j j = 0: (1.7)
j=1
for j = 1; :::; r; where 1 ( ) is the indicator function. That is, ^ j is the percentage of the observations
which fall in each category. The MLE ^ mle for is then the function of (^ 1 ; :::; ^ r ) which satis es
the analog of (1.7) with the i replaced by the ^ i :
r
X
0^
^j j j j mle = 0:
j=1
But this is the same expression as (1.6), which means that ^ mle = ^ ols : In other words, if the
data have a discrete distribution, the maximum likelihood estimator is simply the OLS estimator.
Since this is a regular parametric model the MLE is asymptotically e cient, and thus so is the
OLS estimator.
Chamberlain (1987) extends this argument to the case of continuously-distributed data: He
observes that the above argument holds for all multinomial distributions, and any continuous
distribution can be arbitrarily well approximated by a multinomial distribution. He proves that
generically the OLS estimator is asymptotically e cient for the class of regression models satisfying
Assumption 1.1.1.
6
1.4 Model in Matrix Notation
For some purposes, including computation, it is convenient to write the model and statistics in
matrix notation. We de ne
0 1 0 1 0 1
y1 x01 e1
B y2 C B x02 C B e2 C
B C B C B C
Y = B . C; X=B .. C; e=B .. C:
@ .. A @ . A @ . A
yn x0n en
y1 = x01 + e1
y2 = x02 + e2
..
.
yn = x0n + en :
or equivalently
Y = X + e:
Sample sums can also be written in matrix notation. For example
n
X
xi x0i = X 0 X
i=1
Xn
xi yi = X 0 Y:
i=1
Thus the estimator (1.4), residual vector, and sample error variance can be written
^ = 1
X 0X X 0Y
e^ = Y X^
^2 = n 1 0
e^ e^
Then
1
Y^ = X ^ = X X 0 X X 0Y = P Y
7
and
e^ = Y X^ = Y P Y = (In P ) Y = M Y: (1.8)
Another way of writing this is
Y = (P + M ) Y = P Y + M Y = Y^ + e^:
Y^ 0 e^ = (P Y )0 (M Y ) = Y 0 P M Y = 0:
Y = X1 1 + X2 2 + e: (1.9)
0 0 0
Observe that the OLS estimator of = ( 1; 2) can be obtained by regression of Y on X = [X1
X2 ]: OLS estimation can be written as
Y = X1 ^ 1 + X2 ^ 2 + e^: (1.10)
M1 = In X1 X10 X1 1 X10
M2 = In X2 X20 X2 1 X20 :
8
Thus
^ 1
1 X10 X1 X10 X2 X10 Y
^ =
2 X20 X1 X20 X2 X20 Y
(X10 M2 X1 ) 1 (X10 M2 X1 ) 1 X10 X2 (X20 X2 ) 1
X10 Y
=
(X20 X2 ) 1 X20 X1 (X10 M2 X1 ) 1
(X20 M1 X2 ) 1 X20 Y
1
(X10 M2 X1 ) (X10 M2 Y )
= 1
(X20 M1 X2 ) (X20 M1 Y )
0 1 1
~0X
X ~ ~ 0 Y~1
X
1 1 1
= @ 1
A (1.12)
~0X
X ~ ~ 0
X2 Y2~
2 2
where
~ 1 = M2 X 1
X
Y~1 = M2 Y
~ 2 = M1 X 2
X
Y~2 = M1 Y
The variables X~ 1 and Y~1 are least-squares residuals from the regression of X1 and Y; respectively,
on the matrix X2 only. Similarly, the variables X ~ 2 and Y~2 are least-squares residuals from the
regression of X2 and Y on the matrix X1 only.
Formula (1.12) shows that the subvector ^ 1 of the OLS estimator ^ can be calculated by the
OLS regression of Y~1 on X ~ 1 ; and similarly ^ 2 can be calculated by the OLS regression of Y~2 on
~
X2 : This technique is called residual regression.
Furthermore, recalling the de nition M = I X (X 0 X) 1 X 0 ; observe that X20 M = 0 and hence
M1 M = I X2 X20 X2 1 X20 M = M
Theorem 1.5.1 (Frisch-Waugh-Lovell). In the model (1.9), the OLS estimator of 1 and the OLS
residuals e^ may be equivalently computed by either the OLS regression (1.10) or via the following
algorithm:
9
~1;
2. Regress X1 on X2 ; obtain residuals X
~ 1 ; obtain OLS estimates ^ 1 and residuals e^:
3. Regress Y~ on X
In some contexts, the FWL theorem can be used to speed computation, but in most cases there
is little computational advantage to using the two-step algorithm. Rather, the theorem's primary
use is theoretical.
A common application of the FWL theorem, which you may have seen in an introductory
econometrics course, is the demeaning formula for regression. Partition X = [X1 X2 ] where
X1 = is a vector of ones, and X2 is the vector of observed regressors. In this case,
0 1 0
M1 = I :
Observe that
~ 2 = M1 X 2 = X 2 0 1 0
X X2
= X2 X2
and
1 0
Y~ = M1 Y = Y 0
Y
= Y Y;
which are \demeaned". The FWL theorem says that ^ 2 is the OLS estimate from a regression of
Y~ on X
~ 2 ; or yi y on x2i x2 :
n
! 1 n
!
X X
^ =
2 (x2i x2 ) (x2i x2 )0 (x2i x2 ) (yi y) :
i=1 i=1
Thus the OLS estimator for the slope coe cients is a regression with demeaned data.
1.6 Consistency
The OLS estimator ^ is a statistic, and thus has a statistical distribution. In general, this distribu-
tion is unknown. Asymptotic (large sample) methods approximate sampling distributions based
on the limiting experiment that the sample size n tends to in nity. A preliminary step in this
approach is the demonstration that estimators are consistent { that they converge in probability
to the true parameters as the sample size gets large.
10
The following decomposition is quite useful.
n
! 1 n
X X
^ = xi x0 xi yi
i
i=1 i=1
n
! 1 n
X X
= xi x0i xi x0i + ei
i=1 i=1
n
! 1 n
! n
! 1 n
X X X X
= xi x0i xi x0i + xi x0i xi ei
i=1 i=1 i=1 i=1
n
! 1 n
X X
= + xi x0i xi ei : (1.13)
i=1 i=1
This shows that after centering, the distribution of ^ is determined by the joint distribution of
(xi ; ei ) only.
We can now deduce the consistency of ^ : First, Assumption 1.1.1 and the WLLN (Section 17.2)
imply that
n
1X
xi x0i !p E xi x0i = Q (1.14)
n
i=1
and
n
1X
xi ei !p E (xi ei ) = 0: (1.15)
n
i=1
n
! 1 n
X X
^ = + xi x0i xi ei
i=1 i=1
n
! 1 n
!
1X 1X
= + xi x0i xi ei
n n
i=1 i=1
n n
!
1X 0 1
X
= +g xi xi ; xi ei
n n
i=1 i=1
where g(A; b) = A 1 b is a continuous function of A and b; at all values of the arguments such that
A 1 exist. Now by (1.14) and (1.15),
n n
!
1X 1 X
xi x0i ; xi ei !p (Q; 0) :
n n
i=1 i=1
11
Assumption 1.1.1.5 implies that Q 1 exists and thus g( ; ) is continuous at (Q; 0): Hence by the
continuous mapping theorem (CMT) (Section 17.5),
n n
!
1X 1 X
g xi x0i ; xi ei !p g (Q; 0) = Q 1 0 = 0
n n
i=1 i=1
so !
n n
^= 1X 1X
+g xi x0i ; xi ei !p +0=0
n n
i=1 i=1
In Section 1.2 we also de ned the sample error variance ^ 2 : We now demonstrate its consistency
for 2 : Using (1.8),
n^ 2 = e0 M M e = e0 M e = e0 e e0 P e: (1.16)
An application of the WLLN yields
n
1X 2
ei !p Ee2i = 2
n
i=1
so ^ 2 is consistent for 2:
Assumption 1.7.1 In addition to Assumption 1.1.1, Ee4i < 1 and E jxi j4 < 1:
Now de ne
= E xi x0i e2i :
Assumption (1.7.1) guarantees that the elements of are nite. To see this,by the Cauchy-Schwarz
inequality and Assumption 1.7.1,
12
Thus xi ei is iid with mean zero and has covariance matrix . By the central limit theorem (Section
17.4),
n
1 X
p xi ei !d N (0; ) (1.19)
n
i=1
Then using (1.13), (1.14), and (1.19),
n
! 1 n
!
p 1X 1 X
n ^ = xi x0i p xi ei
n n
i=1 i=1
1
!d Q N (0; )
1 1
= N 0; Q Q :
where V = Q 1 Q 1:
p
As V is the variance of the asymptotic distribution of n ^ ; V is often referred to as
the asymptotic covariance matrix of ^ : The form V = Q 1 Q 1 is called a sandwich form.
It may be insightful to examine a special case where and V simplify:
Homoskedastic Projection Error: Cov(xi x0i ; e2i ) = 0
Condition (1.7) holds, for example, when xi and ei are independent, but this is not a necessary
condition. We should not expect it to generically hold, but when it does the asymptotic variance
formulas simplify. If (1.7) is true, then
= E xi x0i E e2i = Q 2
(1.20)
1 1 1 2 0
V = Q Q =Q V (1.21)
V^ 0 = Q
^ 1 2
^ (1.22)
where
X n
^= 1
Q
1
xi x0i = X 0 X
n n
i=1
13
^ !p Q and ^ !p 2 (see (1.14) and (1.17)) it
is the method of moments estimator for Q: Since Q
is clear that V^ !p V :
0 0
X n
^= 1 xi x0i e^2i
n
i=1
Hence
14
n
^ 1X
= xi x0i e^2i
n
i=1
n n n
1X 2X 0 1X 0
= xi x0i e2i xi x0i ^ xi ei + xi x0i ^ xi x0i ^ : (1.23)
n n n
i=1 i=1 i=1
We now examine the each sum on the right-hand-side of (1.23) in turn. First, (1.18) and the
WLLN show that
n
1X
xi x0i e2i !p E xi x0i e2i = :
n
i=1
so by the WLLN
n
1X
jxi j3 jei j !p E jxi j3 jei j ;
n
i=1
n n
!
1X 0 1X 3
xi x0i ^ xi ei ^ jxi j jei j !p 0:
n n
i=1 i=1
Theorem 1.8.1 As n ! 1; ^ !p :
15
p p
When is scalar, and V^ is an estimator of the variance of n ^ ; we set s( ^ ) = n 1=2 V^
: When is a vector, we focus on individual elements of one-at-a-time, vis., j ; j = 1; :::; k: Thus
q
s( ^ j ) = n 1=2
V^jj :
Generically, standard errors are not unique, as there may be more than one estimator of the
variance of the estimator. It is therefore important to understand what formula and method is
used by an author when studying their work. It is also important to understand that a particular
standard error may be relevant under one set of model assumptions, but not under another set of
assumptions, just as any other estimator.
From a computational standpoint, the standard method to calculate the standard errors is to
rst calculate n 1 V^ , then take the diagonal elements, and then the square roots.
= h( ):
We will assume from now on that h( ) is continuously di erentiable at the true value of :
The estimate of is
^ = h( ^ ):
h( ^ ) ' h( ) + H 0 ^ :
where
@
H = h( ) k q:
@
Thus
p p
n ^ = n h( ^ ) h( )
p
' H0 n ^
!d H 0 N (0; V )
= N (0; V ): (1.24)
where
V = H0 V H :
16
If V^ is the estimated covariance matrix for ^ ; then the natural estimate for the variance of ^ is
^ 0 V^ H
V^ = H ^
where
^ = @ h( ^ ):
H
@
In many cases, the function h( ) is linear:
h( ) = R0
^ = R; so V^ = R0 V^ R:
for some k q matrix R: In this case, H = R and H
For example, if R is a \selector matrix"
I
R=
0
I
V^ = I 0 V^ = V^11 ;
0
1.10 t tests
Let = h( ) : Rk ! R be any parameter of interest, ^ its estimate and s(^) its asymptotic
standard error. Consider the studentized statistic
^
tn ( ) = (1.25)
s(^)
It is easy to calculate that this statistic has the asymptotic distribution
^
tn ( ) =
s(^)
p ^
n
= q
V^
N (0; V )
!d p = N (0; 1)
V
17
the standard normal. This distribution is known. Since this distribution does not depend on the
parameters, we say that tn ( ) is asymptotically pivotal. In special cases (such as the normal
regression model, see Section 2.7), the statistic tn has an exact t distribution, and is therefore
exactly free of unknowns. In this case, we say that tn is an exactly pivotal statistic. In general,
however, pivotal statistics are unavailable and so we must rely on asymptotically pivotal statistics.
A simple null and composite hypothesis takes the form
H0 : = 0
H1 : 6= 0
where 0 is some pre-speci ed value, and = h( ) is some function of the parameter vector. (For
example, could be a single element of ):
The standard test for H0 against H1 is the t-statistic (or studentized statistic)
^ 0
tn = tn ( 0 ) = :
s(^)
Under H0 ; tn !d N (0; 1): Let z =2 is the upper =2 quantile of the standard normal distribution.
That is, if Z N (0; 1); then P (Z > z =2 ) = =2 and P (jZj > z =2 ) = : For example, z:025 = 1:96
and z:05 = 1:645: A test of asymptotic signi cance rejects H0 if jtn j > z =2 : Otherwise the test
does not reject, or \accepts" H0 : This is because
pn = p(tn ):
If the p-value pn is small (close to zero) then the evidence against H0 is strong. In a sense, p-
values and hypothesis tests are equivalent since pn < if and only if jtn j > z =2 , thus an equivalent
statement of a Neyman-Pearson test is to reject at the % level if and only if pn < : The p-value
is more general, however, in that the reader is allowed to pick the level of signi cance ( ), in
contrast to Neyman-Pearson rejection/acceptance reporting, where the researcher picks the level.
Another helpful observation is that the p-value function has simply made a unit-free transfor-
mation of the test statistic. That is, under H0 ; pn !d U [0; 1]; so the \unusualness" of the test
18
statistic can be compared to the easy-to-understand uniform distribution, regardless of the com-
plication of the distribution of the original test statistic. To see this fact, note that the asymptotic
distribution of jtn j is F (x) = 1 p(x): Thus
P (1 pn u) = P (1 p(tn ) u)
= P (F (tn ) u)
1
= P jtn j F (u)
1
! F F (u) = u;
establishing that 1 pn !d U [0; 1]; from which it follows that pn !d U [0; 1]:
It may be helpful to note that in the GAUSS language, the function p(t) may be computed by
the expression p = 2 cdf nc(t):
19
1.12 Wald Tests
Sometimes = h( ) is a q 1 vector, and it is desired to test the joint restrictions simultaneously.
In this case the t-statistic approach does not work. We have the null and alternative
H0 : = 0
H1 : 6= 0:
where
H^ = @ h( ^ ):
@
The Wald statistic for H0 against H1 is
0
Wn = n ^ 0 V^ 1 ^ 0
0 1
= n h( ^ ) 0
^ 0 V^ H
H ^ h( ^ ) 0 :
When h is a linear function of ; h( ) = R0 ; then the Wald statistic takes the form
0 1
Wn = n R 0 ^ 0 R0 V^ R R0 ^ 0 :
p
The delta method (1.24) showed that n ^ !d Z N (0; V ); and Theorem 1.8.1 showed
that V^ !p V: Furthermore, H ( ) is a continuous function of ; so by the continuous mapping
theorem, H ( ^ ) !p H : Thus V^ = H^ 0 V^ H
^ !p H 0 V H = V > 0 if H has full rank q: Hence
0
Wn = n ^ 0 V^ 1 ^ 0 !d Z 0 V 1
Z= 2
q;
20
1.13 F Tests
Take the linear model
Y = X1 1 + X2 2 +e
where X1 is n k1 and X2 is n k2 and k = k1 + k2 : The null hypothesis is
H0 : 2 = 0:
0
In this case, = 2; and there are q = k2 restrictions. Also h( ) = R0 is linear with R =
I
a selector matrix. We know that the Wald statistic takes the form
0
Wn = n^ V^ 1^
0 1
= n ^ 2 R0 V^ R ^ :
2
1
What we will show in this section is that if V^ is replaced with V^ 0 = ^ 2 n 1 X 0 X ; the covariance
matrix estimator valid under homoskedasticity, then the Wald statistic can be written in the form
~2 ^2
Wn = n (1.27)
^2
where
1 0 1
~2 =
e~ e~; e~ = Y X1 ~ 1 ; ~ = X 0 X1
1 1 X10 Y
n
are from OLS of Y on X1 ; and
1 0 1
^2 = e^ e^; e^ = Y X ^; ^ = X 0X X 0Y
n
are from OLS of Y on X = (X1 ; X2 ):
The elegant feature about (1.27) is that it is directly computable from the standard output
from two simple OLS regressions, as the sum of square errors is a typical output from statistical
packages. This statistic is typically reported as an \F-statistic" which is de ned as
n k Wn ~ 2 ^ 2 =k2
F = = :
n k2 ^ 2 =(n k)
1
While it should be emphasized that equality (1.27) only holds if V^ 0 = ^ 2 n 1 X 0 X ; still this
formula often nds good use in reading applied papers.
We now derive expression (1.27). First, note that using (1.11),
! 1
1 0X 0X 1
X 1 1 X 1 2
R0 V^ 0 R = n 1 ^ 2 R0 R = ^ 2 n 1 X20 M1 X2 ;
X20 X1 X20 X2
21
where M1 = I X1 (X10 X1 ) 1X 0 :
1 Thus
0 1
Wn = n ^ 2 R0 V^ 0 R ^
2
^ 0 (X 0 M1 X2 ) ^
2 2 2
= 2 :
^
To simplify this expression further, note that if we regress Y on X1 alone, the residual is
e~ = M1 Y: Now consider the residual regression of e~ on X ~ 2 = M1 X2 : By the FWL theorem,
~ ^ ~ 0
e~ = X2 2 + e^ and X2 e^ = 0: Thus
0
e~0 e~ = ~ 2 ^ 2 + e^
X ~ 2 ^ 2 + e^
X
0
= ^2X ~ 2 ^ 2 + e^0 e^
~ 20 X
0
= ^ X 0 M1 X2 ^ + e^0 e^;
2 2 2
or alternatively,
^ 0 X 0 M1 X2 ^ = e~0 e~ e^0 e^:
2 2 2
Also, since
^2 = n 1 0
e^ e^
we conclude that
e~0 e~ e^0 e^ ~2 ^2
Wn = n =n ;
e^0 e^ ^2
as claimed.
In many statistical packages, when an OLS regression is reported, an \F statistic" is reported.
This is
~ 2y ^ 2 =(k 1)
F = 2 :
^ =(n k)
where
1
(y y)0 (y y)
~ 2y =
n
is the sample variance of yi ; equivalently the residual variance from an intercept-only model. This
special F statistic is testing the hypothesis that all slope coe cients (other than the intercept) are
zero. This was a popular statistic in the early days of econometric reporting, when sample sizes
were very small and researchers wanted to know if there was \any explanatory power" to their
regression. This is rarely an issue today, as sample sizes are typically su ciently large that this F
statistic is highly \signi cant". Certainly, there are special cases where this F statistic is useful,
but these cases are atypical.
22
Chapter 2
Regression Models
2.1 Regression
In regression, we want to nd the central tendency of the conditional distribution of yi given xi :
A standard measure of central tendency is the mean. The conditional analog is the conditional
mean m(x) = E (yi j xi = x). In general, m(x) can take any form.
The regression error ei is de ned to be the di erence between yi and its conditional mean:
ei = yi m(xi ):
yi = m(xi ) + ei : (2.1)
It is worth emphasizing that no assumptions have been used to develop (2.1), other than that
(yi ; xi ) have a joint distribution and E jyi j < 1:
1. E (ei j xi ) = 0:
2. E(ei ) = 0:
4. E(xi ei ) = 0:
Proof:
23
1. By the de nition of ei and the linearity of conditional expectations,
2. By the law of iterated expectations (Theorem 16.7) and the rst result,
E(ei ) = E (E (ei j xi ))
= E(0)
= 0:
Equation (2.1) plus Proposition 2.1.1.1 are often stated jointly as the regression framework:
yi = m(xi ) + ei (2.2)
E (ei j xi ) = 0:
It is important to understand that this is a framework, not a model, because no restrictions have
been placed on the joint distribution of the data. These equations hold true by de nition. A
regression model imposes further restrictions on the joint distribution; most typically, restrictions
on the permissible class of regression functions m(x):
The most common choice is the linear regression model. It speci es that m(x) is a linear
function of x :
yi = x0i + ei
E (ei j xi ) = 0
Since this is a linear equation is a special case of the general conditional conditional mean equation,
this is a substantive restriction which may or may not be true in a speci c application. The fact
24
that Proposition 2.1.1.4 is the same as Assumption 1.1.1.2 means that the linear regression model
is a special case of the least-squares projection model of Chapter 1. Another way of saying this
is that the conditional mean assumption that E (ei j xi ) = 0 is stronger than the uncorrelated
assumption E (xi ei ) = 0:
It is also useful to de ne the conditional variance of yi given xi = x:
V ar (yi j xi = x) = E e2i j xi = x = 2
(x):
Generally, this is a function of x: Just as the conditional mean function may take any form, so
may the conditional variance function (other than the restriction that it is non-negative). Given
the random variable xi , the conditional variance is 2i = 2 (xi ): In the general case where 2 (x)
is not necessarily a constant function, so 2i may di erent across i; we say that the error ei is
heteroskedastic.
When 2 (x) is a constant, so that
E e2i j xi = 2
(2.3)
yi = x0i + ei
E (ei j xi ) = 0
E e2i j xi = 2
is called the homoskedastic linear regression model. In this case, by the law of iterated
expectations
E xi x0i e2i = E xi x0i E e2i j xi = Q 2
which is (1.20). Thus the homoskedastic linear regression model is a special case of the homoskedas-
tic projection error model.
25
observations
2 ! 3
n 1 n
X X
E ^ jX = E4 xi x0i xi ei j X 5
i=1 i=1
n
! 1 n
X X
= xi x0i xi E (ei j X)
i=1 i=1
n
! 1 n
X X
= xi x0i xi E (ei j xi )
i=1 i=1
= 0
V ar(Y ) = E (Y EY ) (Y EY )0
= EY Y 0 (EY ) (EY )0 :
and we nd ! ! !
n 1 n n 1
X X X
V ar ^ j X = xi x0i xi x0i 2i xi x0i :
i=1 i=1 i=1
In the special case of the linear homoskedastic regression model, 2 = 2 and the covariance
i
matrix simpli es to
n
! 1
X
V ar ^ j X = 0
xi xi 2
:
i=1
2
Recall the method of moments estimator ^ for 2 : We now calculate its nite sample bias
in the context of the homoskedastic linear regression model. Using (1.16) and linear algebra
manipulations
26
E n^ 2 j X = E e0 M e j X
= E tr e0 M e j X
= E tr M ee0 j X
= tr E M ee0 j X
= tr M E ee0 j X
2
= tr M
2
= (n k);
the nal equality by (15.4). We have found that under these assumptions
(n k)
E ^2 = 2
n
so ^ 2 is biased towards zero. Since the bias is proportional to 2; it is common to de ne the
bias-corrected estimator
n
1 X 2
s2 = e^i
n k
i=1
so that Es2 = 2is unbiased. It is important to remember, however, that this estimator is only
unbiased in the special case of the homoskedastic linear regression model. It is not unbiased in the
absence of homoskedasticity, or in the projection model.
2.3 Multicollinearity
If rank(X 0 X) < k; then ^ is not de ned. This can be called strict multicollinearity. This
happens when the columns of X are linearly dependent, i.e., there is some such that X = 0:
Most commonly, this arises when sets of regressors are included which are identically related.
For example, if X includes both the logs of two prices and the log of the relative prices log(p1 );
log(p2 ) and log(p1 =p2 ): When this happens, the applied researcher quickly discovers the error as
the statistical software will be unable to construct (X 0 X) 1 : Since the error is discovered quickly,
this is rarely a problem for applied econometric practice.
The more relevant issue is near multicollinearity, which is often called \multicollinearity"
for brevity. This is the situation when the X 0 X matrix is near singular, when the columns of X are
close to linearly dependent. This de nition is not precise, because we have not said what it means
for a matrix to be \near singular". This is one di culty with the de nition and interpretation of
multicollinearity.
One implication of near singularity of matrices is that the numerical reliability of the calcula-
tions is reduced. It is possible that the reported calculations will be in error due to oating-point
calculation di culties.
27
More relevantly in practice, an implication of near multicollinearity is that estimation precision
of individual coe cients will be poor. We can see this most simply in a model with two regressors
and no intercept:
yi = x1i 1 + x2i 2 + ei ;
where ei is independent of x1i and x2i ; Ee2i = 1 and
The correlation indexes collinearity, since as approaches 1 the matrix becomes singular. We can
see the e ect of collinearity on precision by examining the asymptotic variance of either coe cient
2 1
estimate, which is 1 : As approaches 1, the variance rises quickly to in nity. Thus the
more \collinear" are the regressors, the worse the precision of the individual coe cient estimates.
Basically, what is happening is that when the regressors are highly dependent, it is statistically
di cult to disentangle the impact of 1 from that of 2 : The precision of individual estimates are
reduced.
Is there a simple solution? Basically, No. Fortunately, multicollinearity does not lead to errors
in inference. The asymptotic distribution is still valid. Regression estimates are asymptotically
normal, and estimated standard errors are consistent for the asymptotic variance. So reported
con dence intervals are not inherently misleading. They will be large, correctly indicating the
inherent uncertainty about the true parameter value
m(x) = E (yi j xi = x) = x0 :
In some cases, we want to estimate m(x) at a particular point x: Notice that this is a (linear)
function 0
pof : Letting h( ) = x and = h( ); we see that m(x) ^ = ^ = x0 ^ and H = x; so
s(^) = n 1 x0 V^ x: Thus an asymptotic 95% con dence interval for m(x) is
h p i
x0 ^ 2 n ^x
1 x0 V :
It is interesting to observe that if this is viewed as a function of x; the width of the con dence set
is dependent on x:
28
For a given value of xi = x; we may want to forecast (guess) yi out-of-sample. A reasonable
guess is the conditional mean m(x); and indeed this is the mean-square-minimizing decision rule.
Thus a point forecast is m(x)
^ = x0 ^ ; the estimated conditional mean, as discussed above. We
would also like a measure of uncertainty for the forecast.
The forecast error is e^i = yi m(x)^ = ei x0 ^ : As the out-of-sample error ei is
independent of the in-sample estimate ^ ; this has variance
0
e2i = E e2i j xi = x + x0 E ^
E^ ^ x
2 1 0
= (x) + n x V x:
but this turns out to be incorrect. In general, the validity of an asymptotic con dence interval is
based on the asymptotic normality of the studentized ratio. In the present case, this would require
the asymptotic normality of the ratio
ei x0 ^
p :
^2 + n ^x
1 x0 V
But no such asymptotic approximation can be made. The only special exception is the case where
ei has the exact distribution N (0; 2 ); which is generally invalid.
To get an accurate forecast interval, we need to estimate the conditional distribution of ei given
xi = x; which is a much more di cult h task. pGiven the di culty,
i most applied forecasters focus on
the simple and unjusti ed interval x ^ 2 ^ + n 1 x0 V^ x :
0 2
Letting =( 0; 1 ; :::; k) and zi = (1; xi ; x2i ; :::; xki ); this is the linear regression yi = zi0 + ei .
29
Now suppose that x 2 R2 : A simple quadratic approximation is
2 2
yi = 0 + 1 x1i + 2 x2i + 3 x1i + 4 x2i + 5 x1i x2i + ei :
Non-linear approximations can also be made using alternative basis functions, such as Fourier
series (sins and cosines), splines, neural nets, or wavelets.
Since these non-linear models are linear in the parameters, they can be estimated by OLS, and
inference is convention. However, the model is non-linear so interpretation must take this into
account. For example, in the cubic model given above, the slope with respect to xi1 is
@ 2
E (yi j xi ) = 1 +2 3 x1i +3 5 x1i + 7 x2i ;
@x1i
which is a function of x1i and x2i ; making reporting of the \slope" di cult. In many applications,
it will be important to report the slopes for di erent values of the regressors, carefully chosen to
illustrate the point of interest. In other applications, an average slope may be su cient. There
are two obvious candidates: the derivative evaluated at the sample averages
@ 2
E (yi j xi ) jxi =x = 1 +2 3 x1 +3 5 x1 + 7 x2
@x1i
and the average derivative
n n
1X @ 1X 2
E (yi j xi ) = 1 + 2 3 x1 + 3 5 x1i + 7 x2 :
n @x1i n
i=1 i=1
30
functions include
x
m(x; ) = 1 + 2
1 + 3x
m(x; ) = 1 + 2x
3
m(x; ) = 1 + 2 exp( 3 x)
0
m(x; ) = G(x ); G known
x2 5
m(x; ) = 1 + 2 x1 +( 3 + 4 x1 )
6
m(x; ) = 1+ 2x + 4 (x 3 ) 1 (x > 3)
m(x; ) = ( 1 + 2 x1 ) 1 (x2 < 3) +( 4 + 5 x1 ) 1 (x2 > 3)
In the rst ve examples, m(x; ) is (generically) di erentiable in the parameters : In the nal
two examples, m is not di erentiable with respect to 3 ; which alters some of the analysis. When
it exists, let
@
m (x; ) = m(x; ):
@
Nonlinear regression is frequently adopted because the functional form m(x; ) is suggested
by an economic model. In other cases, it is adopted as a exible approximation to an unknown
regression function.
The least squares estimator ^ minimizes the sum-of-squared-errors
n
X
Sn ( ) = (yi m(xi ; ))2 :
i=1
When the regression function is nonlinear, we call this the nonlinear least squares (NLLS)
estimator. The NLLS residuals are e^i = yi m(xi ; ^):
One motivation for the choice of NLLS as the estimation method is that the parameter is
the solution to the population problem min E (yi m(xi ; ))2
Since sum-of-squared-errors function Sn ( ) is not quadratic, ^ must be found by numerical
methods. See Appendix E. When m(x; ) is di erentiable, then the FOC for minimization are
n
X
0= m (xi ; ^)^
ei : (2.4)
i=1
Theorem 2.6.1 If the model is identi ed and m(x; ) is di erentiable with respect to ,
p
n ^ 0 !d N (0; V )
1 1
V = E m i m0 i E m i m0 i e2i E m i m0 i
where m i = m (xi ; 0 ):
31
Sketch of Proof. First, it must be shown that ^ !p 0 : This can be done using arguments
for optimization estimators, but we won't cover that argument here. Since ^ !p 0 ; ^ is close to
0 for n large, so the minimization of Sn ( ) only needs to be examined for close to 0 : Let
yi0 = ei + m0 i 0 :
For close to the true value 0; by a rst-order Taylor series approximation,
m(xi ; ) ' m(xi ; 0) + m0 i ( 0) :
Thus
yi m(xi ; ) ' (ei + m(xi ; 0 )) m(xi ; 0) + m0 i ( 0)
0
= ei m i( 0)
= yi0 m 0
i :
Hence the sum of squared errors function is
Xn n
X
2 2
Sn ( ) = (yi m(xi ; )) ' yi0 m0 i
i=1 i=1
and the right-hand-side is the SSE function for a linear regression of yi0 on m i : Thus the NLLS
estimator ^ has the same asymptotic distribution as the (infeasible) OLS regression of yi0 on m i ;
which is that stated in the theorem.
Based on Theorem 2.6.1, an estimate of the asymptotic variance V is
n
! 1 n
! n
! 1
1 X 1 X 1 X
V^ = m 0
^ i m 0 2
^ i e^i
^ im m
^ im 0
^ i
n n n
i=1 i=1 i=1
32
2.7 Normal Regression Model
The normal regression model adds the additional assumption that the error ei is independent of xi
and has distribution N (0; 2 ): This is a parametric model, where likelihood methods can be used
for estimation, testing, and distribution theory.
The log-likelihood function for the normal regression model is
n
!
X 1 1 2
2 0
Ln ( ; ) = log exp yi xi
(2 2 )1=2 2 2
i=1
n
n 2 1 X 2
= log 2 2
yi x0i
2 2
i=1
33
The spectral decomposition of M yields
In k 0
M =H H0
0 0
(n k) s2 1
2
= 2
e^0 e^
1
= 2
e0 M e
1 In k 0
= 2
e0 H H 0e
0 0
In k 0
= u0 u
0 0
2
n k;
Theorem 2.7.1 If ei is independent of xi and distributed N (0; 2 ); and standard errors are cal-
culated using the homoskedastic formula (1.22) then
^ N 0; 2 (X 0 X) 1
(n k)s2 2
2 n k;
^
j j
tn k
s( ^ j )
In Chapter 1 we showed that in large samples, ^ and t are approximately normally distributed.
In contrast, Theorem 2.7.1 shows that under the strong assumption of normality, ^ has an exact
normal distribution and t has an exact t distribution. As inference (con dence intervals) are based
on the t-ratio, the notable distinction is between the N (0; 1) and tn k distributions. The critical
34
values are quite close if n k 30; so as a practical matter it does not matter which distribution
is used. (Unless the sample size is unreasonably small.)
Now let us partition = ( 1 ; 2 ) and consider tests of the linear restriction
H0 : 2 =0
H1 : 2 6= 0
In the context of parametric models, a good testing procedure is based on the likelihood ra-
tio statistic, which is twice the di erence in the log-likelihood function evaluated under the null
and alternative hypotheses. The estimator under the alternative is the unrestricted estimator
( ^ 1 ; ^ 2 ; ^ 2 ) discussed above. The log-likelihood at these estimates is
n 1 0
Ln ( ^ 1 ; ^ 2 ; ^ 2 ) = log 2 ^ 2 e^ e^
2 2^ 2
n n n
= log ^ 2 log (2 ) :
2 2 2
The MLE of the model under the null hypothesis is ( ~ 1 ; 0; ~ 2 ) where ~ 1 is the OLS estimate from
a regression of yi on x1i only, with residual variance ~ 2 : The log-likelihood of this model is
n n n
Ln ( ~ 1 ; 0; ~ 2 ) = log ~ 2 log (2 ) :
2 2 2
The LR statistic for H0 is
LR = 2 Ln ( ^ 1 ; ^ 2 ; ^ 2 ) Ln ( ~ 1 ; 0; ~ 2 )
= n log ~ 2 log ^ 2
~2
= n log :
^2
By a rst-order Taylor series approximation
~2 ~2
LR = n log 1 + 1 'n 1 = Wn :
^2 ^2
the F statistic.
35
Let Y be a continuous random variable with median 0 = M ed(Y ). De ne the sign function
1 if u 0
sgn (u) =
1 if u < 0
P (Y 0) = P (Y > 0) = :5
E sgn (Y 0) =0
0 = min E jY j
Given a random sample fy1 ; ::; yn g from this distribution, these three de nitions motivate three
estimators of : The rst suggests
P taking the 50 quantile. The second suggests nding Pthe solution
to the moment equation n1 ni=1 sgn (yi ) ; and the rst suggests minimizing n1 ni=1 jyi j:
These distinctions are illusory, however, as these estimators are indeed identical.
Now let's consider the conditional median of Y given a random variable X: Let m(x) =
M ed (Y j X = x) denote the conditional median of Y given X = x; and let M ed (Y j X) = m(X)
be this function evaluated at the random variable X: The linear median regression model takes
the form
yi = x0i + ei
M ed (ei j xi ) = 0
In this model, the linear function M ed (yi j xi = x) = x0 is the conditional median function, and
the substantive assumption here is that the median function is linear in x:
Conditional analogs of the facts about the median are
E (sgn (ei ) j xi ) = 0
be the average of absolute deviations. The least absolute deviations (LAD) estimator of
minimizes this function
^ = argmin Ln ( )
36
Equivalently, it is a solution to the moment condition
n
1X
xi sgn yi x0i ^ = 0: (2.5)
n
i=1
Let f (e j x) denote the conditional density of ei given xi = x: The LAD estimator has the
asymptotic distribution
p
Theorem 2.8.1 n ^ 0 !d N (0; V );
1 1 1
V = E xi x0i f (0 j xi ) Exi x0i E xi x0i f (0 j xi )
4
The variance of the asymptotic distribution inversely depends on f (0 j x) ; the conditional
density of the error at its median. When f (0 j x) is large, then there are many innovations near
to the median, and this improves estimation of the median. In the special case where the error is
independent of xi ; then f (0 j x) = f (0) and the asymptotic variance simpli es
1
(Exi x0i )
V = (2.6)
4f (0)2
This simpli cation is similar to the simpli cation of the asymptotic covariance of the OLS estimator
under homoskedasticity.
Computation of standard error for LAD estimates typically is based on equation (2.6). The
main di culty is the estimation of f (0); the height of the error density at its median. This can
be done with kernel estimation techniques. See Chapter 13. The proof of Theorem 2.8.1 is a bit
advanced, but we provide it here for completeness.
^
1
Pn : Since sgn (a) = 1 2 1 (a 0) ;0 (2.5) is equivalent to g n ( ) = 0; where g n ( ) =
Proof
n i=1 gi ( ) and gi ( ) = xi (1 2 1 (yi xi )) : Let g( ) = Egi ( ). We need three preliminary
result. First, by the central limit theorem
n
X
p 1=2
n (g n ( 0) g( 0 )) = n gi ( 0) !d N (0; Exi x0i )
i=1
since Egi ( 0 )gi ( 0 )0 = Exi x0i : Second using the law of iterated expectations and the chain rule of
di erentiation,
@ @
g( ) = Exi 1 2 1 yi x0i
@ 0 @ 0
@
= 2 0 E xi E 1 ei x0i x0i 0 j xi
@
" Z 0 #
xi x0i 0
@
= 2 0 E xi f (e j xi )
@ 1
37
so
@
g( 0) = 2E xi x0i f (0 j xi ) :
@ 0
Third, by a Taylor series expansion and the fact g( 0) =0
@
g( ^ ) ' g( 0)
^
0 :
@ 0
Together
1
p @ p
n ^ 0 ' g( 0) ng( ^ )
@ 0
1p
= 2E xi x0i f (0 j xi ) n g( ^ ) gn( ^ )
1 1p
' E xi x0i f (0 j xi ) n (g n ( 0) g( 0 ))
2
1 1
! d E xi x0i f (0 j xi ) N (0; Exi x0i )
2
= N (0; V ):
38
Chapter 3
Model Selection
and
g2 (x1 ; x2 ) = E (yi j x1i = x1 ; x2i = x2 ) :
Both of these functions exist and are well de ned. Given data, either can be estimated. Thus, if
the function g1 (x1 ) is estimated by regression of yi on x1i only, there is no bias.
However, the function g1 may not be of interest. Rather, the function g2 may be of interest.
Thus if g1 is estimated, when the true relationship of interest is g2 ; then there will be estimation
bias. That is, what may be of interest is the e ect of x1i on the conditional mean of yi , holding
x2i constant, namely
@ @
g2 (x1 ; x2 ) 6= g1 (x1 ):
@x1 @x1
In this sense, omission of x2i from the regression can induce bias.
Another way to see this is by focusing on the linear regression model. Suppose that
yi = x01i 1 + x02i 2 + ei
E (ei j x1i ; x2i ) = 0:
Then
39
Thus a regression of yi on x1i does not yield the coe cient 1; unless E (x2i j x1i ) = 0 or 2 = 0:
Furthermore, suppose E (x2i j x1i ) = x1i : Then
E (yi j x1i ) = x01i 1 + ( x1i )0 2
= x01i 1 + 0
2 :
So a regression of yi on x1i will consistently estimate 1 + 0 2 : 1 cannot be uncovered from this
regression, unless = 0 or 2 = 0; and thus the regression is \biased", if the parameter 1 is of
interest.
Notice that the omitted variable bias problem disappears if = 0 (so x1i and x2i are uncor-
related) or if 2 = 0 (so x2i does not enter the joint regression). The rst can be assessed by
examining the correlation between x1i and x2i ; but the second can only be assessed by computing
the joint regression. Therefore the standard advice is when in doubt, to always estimate the more
general model, since it is by construction free of the omitted variables problem.
say, and
1 1
lim nV ar( ^ 1 j X) = Ex1i x01i Ex1i x02i Ex2i x02i Ex2i x01i 2
= (Q11 Q121 ) 2
;
n!1
say. If Ex1i x02i = 0 (so the variables are uncorrelated) then these two variance matrices equal, and
the two estimators have equal asymptotic e ciency.
40
Proposition 3.2.1 If Ex1i x02i = 0; ^ 1 and ~ 1 are both consistent and have equal asymptotic
variances.
so
1
Q111 < (Q11 Q121 ) ;
meaning that ~ 1 has a lower asymptotic variance matrix than ^ 1 : The inclusion of irrelevant
variables reduces e ciency if these variables are correlated with the relevant variables.
Y = X1 1 + X2 2 + e;
M1 : Y = X1 1 + e; E (e j X1 ; X2 ) = 0
M2 : Y = X1 1 + X2 2 + e; E (e j X1 ; X2 ) = 0:
41
One useful property is consistency, that it selects the true model with probability one if the sample
is su ciently large. A model selection procedure is consistent if
c = M1 j M1
P M ! 1
c = M2 j M2
P M ! 1
1 but P M c = M2 j M2 could vary dramatically, depending on the sample size, etc. An-
other problem is that if is held xed, then this model selection procedure is inconsistent, as
c
P M = M1 j M1 ! 1 < 1:
Adjusted R-squared
Since R2 is not a useful model selection rule, as it always \prefers" the larger model, Theil
proposed an adjusted coe cient of determination
2 e0 e^) = (n
(^ k)
R = 1
^ 2y
s2
= 1 :
^ 2y
2
At one time, it was popular to pick between models based on R : This rule is to select M1 if
2 2 2
R1 > R2 ; else select M2 : Since R is a monotonically decreasing function of s2 ; this rule is the
same as selecting the model with the smaller s2 ; or equivalently, the smaller log(s2 ): It is helpful
42
to observe that
n
log(s2 ) = log ^ 2
n k
k
= log ^ 2 + log 1 +
n k
k
' log ^ 2 +
n k
k
' log ^ 2 + ;
n
2
(the rst approximation is log(1 + x) ' x for small x): Thus selecting based on R is the same
as selecting based on log ^ 2 + nk ; which is a particular choice of penalized likelihood criteria. It
turns out that model selection based on any criterion of the form
k
log ^ 2 + c ; c > 0; (3.1)
n
is inconsistent, as the rule tends to over t. Indeed, since under M1 ;
c = M1 j M1 2 2
P M = P R1 > R2 j M1
' P n log(s21 ) < n log(s22 ) j M1
' P n log ^ 21 + ck1 < n log ^ 22 + c (k1 + k2 ) j M1
= P (LRn < ck2 j M1 )
2
! P k2 < ck2 < 1:
43
Since the AIC criterion (3.3) takes the form (3.1), it is an inconsistent model selection criterion,
and tends to over t.
Schwarz Criterion
While many modi cations of the AIC have been proposed, the most popular appears to be one
proposed by Schwarz, based on Bayesian arguments. His criterion, known as the BIC, is
k
BIC = log ^ 2 + log(n) : (3.4)
n
Since log(n) > 2 (if n > 8); the BIC places a larger penalty than the AIC on the number of
estimated parameters and is more parsimonious.
In contrast to the other methods studied above, BIC model selection is consistent. Indeed,
since (3.2) holds under M1 ;
LRn
!p 0;
log(n)
so
c = M1 j M1
P M = P (BIC1 < BIC2 j M1 )
= P (LRn < log(n)k2 j M1 )
LRn
= P < k2 j M1
log(n)
! P (0 < k2 ) = 1:
c = M2 j M2 LRn
P M = P > k2 j M2
log(n)
! 1:
and the question is which subset of the coe cients are non-zero (equivalently, which regressors
enter the regression).
There are two leading cases: ordered regressors and unordered.
44
In the ordered case, the models are
M1 : 1 6= 0; 2 = 3 = = K =0
M2 : 1 6= 0; 2 6= 0; 3 = = K =0
..
.
MK : 1 6= 0; 2 6= 0; : : : ; K 6= 0:
which are nested. The AIC selection criteria estimates the K models by OLS, stores the residual
variance ^ 2 for each model, and then selects the model with the lowest AIC (3.3). Similarly for
the BIC, selecting based on (3.4).
In the unordered case, a model consists of any possible subset of the regressors fx1i ; :::; xKi g;
and the AIC or BIC in principle can be implemented by estimating all possible subset models.
However, there are 2K such models, which can be a very large number. For example, 210 = 1024;
and 220 = 1; 048; 576: In the latter case, a full-blown implementation of the BIC selection criterion
would seem computationally prohibitive.
yi = x0i + ei
which is estimated by OLS, yielding predicted values y^i = x0i ^ : Now let
0 2 1
y^i
B .. C
zi = @ . A
y^im
by OLS, and form the Wald statistic Wn for = 0: It is easy (although somewhat tedious) to
show that under the null hypothesis, Wn !d 2m 1 : Thus the null is rejected at the % level if Wn
exceeds the upper % tail critical value of the 2m 1 distribution.
45
To implement the test, m must be selected in advance. Typically, small values such as m = 2,
3, or 4 seem to work best.
The RESET test appears to work well as a test of functional form against a wide range of
smooth alternatives. It is particularly powerful at detecting single-index models of the form
yi = G(x0i ) + ei
where G( ) is a smooth \link" function. To see why this is the case, note that (3.5) may be written
as
2 3 m
yi = x0i ~ + x0i ^ ~ 1 + x0i ^ ~ 2 + x0i ^ ~ m 1 + e~i
46
Chapter 4
yi = x0i + ei
E (ei j xi ) = 0
imposes the condition of zero conditional mean. This is stronger than the orthogonality condition
E (xi ei ) = 0: This stronger condition can be exploited to improve estimation e ciency. Recall the
de nition of the conditional variance 2i = E e2i j xi :
The Generalized Least Squares (GLS) estimator of is
~ = X 0D 1 1
X X 0D 1
Y (4.1)
where 0 1
2 0 0
1
B 0 2 0 C
2 2 B 2 C
D = diagf 1 ; :::; ng =B .. .. .. .. C:
@ . . . . A
0 0 2
n
The GLS estimator is sometimes called the Aitken estimator.
Since Y = X + e; then
~= 1
+ X 0D 1
X X 0D 1
e :
1 1
V ar( ~ j X) = X 0 D 1
X X 0D 1
DD 1
X X 0D 1
X
1
= X 0D 1
X :
47
The class of unbiased linear estimators take the form
~ = A(X)0 Y; A(X)0 X = Ik
L
where A(X), n k, is a function only of X: This is called linear because it is a linear function
of Y; even though it is nonlinear in X: OLS is the case A(X) = X(X 0 X) 1 and GLS is the case
1
A(X) = D 1 X X 0 D 1 X : Observe that
E ~ L j X = A(X)0 X =
The \best" estimator within this class is the one with the smallest variance.
C 0 X = A0 X A 0X
= Ik Ik = 0
and
1
C 0 DA = C 0 DD 1
X X 0D 1
X
1
= C 0X X 0D 1
X = 0:
Then
A0 DA = (C + A )0 D (C + A )
= C 0 DC + C 0 DA + A 0 DC + A 0 DA
= C 0 DC + A 0 DA A DA 0 :
The Gauss-Markov theorem tells us that the OLS estimator is ine cient in linear regression
models, and that within the class of linear estimators, GLS is e cient. However, the restriction
to linear estimators is unsatisfactory, as the theorem leaves open the possibility that a non-linear
estimator could have lower mean squared error than the GLS estimator.
48
Chamberlain (1987) established a more powerful and general result. He showed that in the
regression model, no regular consistent estimator can have a lower asymptotic variance than the
GLS estimator. This establishes that the GLS estimator is asymptotically e cient. The proof of
his theorem is quite deep and we cannot cover it here.
In Section 1.3 we claimed that OLS is asymptotically e cient in the class of models with
E (xi ei ) = 0: Now we have shown that OLS is ine cient if E (ei j xi ) = 0: The gain of e ciency
(through use of GLS) comes through the exploitation of the stronger conditional mean assumption,
which has the cost of reduced robustness. If E (ei j xi ) 6= 0 then the GLS estimator will be
inconsistent for the projection coe cient ; but OLS will be consistent.
49
4.3 Estimation of Skedastic Regression
Suppose ei (and thus i) were observed. Then we could estimate by OLS:
1
^ = Z 0Z Z 0 !p
and p
n (^ ) !d N (0; V )
where
1 1
V = E zi zi0 E zi zi0 2
i E zi zi0 : (4.4)
While ei is not observed, we have the OLS residual e^i = yi x0i ^ = ei x0i ( ^ ): Thus
^ i = e^2i e2i
= 2ei x0i ^ + (^ )0 xi x0i ( ^ )
= i;
Let
1
~ = Z 0Z Z0^ (4.5)
be from OLS regression of ^i on zi : Then
p p 1 1
n (~ ) = n (^ )+ n Z 0Z n 1=2
Z0
!d N (0; V ) (4.6)
Thus the fact that i is replaced with ^i is asymptotically irrelevant. We may call (4.5) the
skedastic regression, as it is estimating the conditional variance of the regression of yi on xi :
We have shown that is consistently estimated by a simple procedure, and hence we can
2
estimate i = zi by ~ i = zi0 ~ : We now discuss how to use these results to test hypotheses on ;
2 0
H0 : 1 =0
50
in the regression (4.2). We may therefore test this hypothesis by the estimation (4.5) and con-
structing a Wald statistic.
This hypothesis does not imply that i is independent of xi : Typically, however, we impose
the stronger hypothesis and test the hypothesis that ei is independent of xi ; in which case i is
independent of xi and the asymptotic variance (4.4) for ~ simpli es to
1
V = E zi zi0 E 2
i : (4.7)
Hence the standard test of H0 is a classic F (or Wald) test for exclusion of all regressors from the
skedastic regression (4.5). The asymptotic distribution (4.6) and the asymptotic variance (4.7)
under independence show that this test has an asymptotic chi-square distribution.
Most tests for heteroskedasticity take this basic form. The main di erences between popular
\tests" is which transformations of xi enter zi : Motivated by the form of the asymptotic variance
of the OLS estimator ^ ; White (1980) proposed that the test for heteroskedasticity be based on
setting zi to equal all non-redundant elements of xi ; its squares, and all cross-products. Breusch-
Pagan (1979) proposed what might appear to be a distinct test, but the only di erence is that
they allowed for general choice of zi ; and used an assumption of normality to use the simpli cation
(4.3) for their test. If this simpli cation is replaced by the standard formula (under independence
of the error), the two tests coincide.
and
1
~ = X 0D
~ 1
X ~
X 0D 1
Y:
51
that there is a need to bound the estimated variances away from zero. A trimming rule might
make sense:
2 2 2
i = max[~ i ; ]
for some 2 > 0:
It is possible to show that if the skedastic regression is correctly speci ed, then FGLS is
asymptotically equivalent to GLS, but the proof of this is tricky in our notational structure. We
just state the result without proof.
Theorem 4.5.1 If the skedastic regression is correctly speci ed,
p
n ~ GLS ~ F GLS !p 0;
and thus p
n ~ F GLS !d N (0; V );
where
2 1
V = E i xi x0i :
which is consistent for V as n ! 1: This estimator V~ 0 is appropriate when the skedastic regression
(4.2) is correctly speci ed.
It may be the case that 0 zi is only an approximation to the true conditional variance 2i =
E(e2i j xi ). In this case we interpret 0 zi as a linear projection of e2i on zi : ~ should perhaps be
called a quasi-FGLS estimator of : Its asymptotic variance is not that given in Theorem 4.5.1.
Instead,
1 1 2 1 1
0
V = E zi xi x0i E 0
zi 2 0
i xi xi E 0
zi xi x0i :
V takes a sandwich form, similar to the covariance matrix of the OLS estimator. Unless 2i = 0 zi ,
V~ 0 is inconsistent for V .
An appropriate solution is to use a White-type estimator in place of V~ 0 : This may be written
as
n
! 1 n
! n
! 1
1 X 1 X 1 X
V~ = 2 0
~ i xi xi 4 2 0
~ i e^i xi xi 2
~ i xi xi 0
n n n
i=1 i=1 i=1
1 1
~
= n XD 0 1
X ~
XD 0 1 ^D
D ~ 1
X ~
XD 0 1
X
52
where D ^ = diagf^e21 ; :::; e^2n g: This is an estimator which is robust to misspeci cation of the condi-
tional variance, and was proposed by Cragg (Journal of Econometrics, 1992).
53
Chapter 5
yi = x0i + ei
= x01i 1 + x02i 2 + ei
E (xi ei ) = 0
yi = x01i 1 + ei
E (xi ei ) = 0:
In this case, how should 1 be estimated? One method is OLS regression of yi on x1i alone. This
method, however, is not necessarily e cient, as their are ` restrictions in E (xi ei ) = 0, while 1 is
of dimension k < `. This situation is called overidenti ed. There are ` k = r more moment
restrictions than free parameters. We call r the number of overidentifying restrictions.
This is a special case of a more general class of moment condition models. Let g(y; z; x; ) be
an ` 1 function of a k 1 parameter with ` k such that
Eg(yi ; zi ; xi ; 0) =0 (5.1)
where 0 is the true value of : In our previous example, g(y; x; ) = x(y x01 ): In econometrics,
this class of models are called moment condition models. In the statistics literature, these are
known as estimating equations.
54
As an important special case we will devote special attention to linear moment condition
models, which can be written as
yi = zi0 + ei
E (xi ei ) = 0:
where the dimensions of zi and xi are k 1 and ` 1 , with ` k: If k = ` the model is just
identi ed, otherwise it is overidenti ed. The variables zi may be components and functions of
xi ; but this is not required. This model falls in the class (5.1) by setting
g(y; z; x; 0) = x(y z0 ) (5.2)
The method of moments estimator for is de ned as the parameter value which sets g n ( ) = 0;
but this is generally not possible when ` > k: The idea of the generalized method of moments
(GMM) is to de ne an estimator which sets g n ( ) \close" to zero.
For some ` ` weight matrix Wn > 0; let
Jn ( ) = n g n ( )0 Wn g n ( ):
This is a non-negative measure of the \length" of the vector g n ( ): For example, if Wn = I; then,
Jn ( ) = n g n ( )0 g n ( ) = n jg n ( )j2 ; the square of the Euclidean length. The GMM estimator
minimizes Jn ( ):
De nition 5.2.1 ^ GM M = argmin Jn ( ):
55
Proposition 5.2.1
^ 1
GM M = Z 0 XWn X 0 Z Z 0 XWn X 0 Y:
While the estimator depends on Wn ; the dependence is only up to scale, for if Wn is replaced
by cWn for some c > 0; ^ GM M does not change.
1 1
V = Q0 W Q Q0 W W Q Q0 W Q :
In general, GMM estimators are asymptotically normal with \sandwich form" asymptotic
variances.
The optimal weight matrix W0 is one which minimizes V: This turns out to be W0 = 1 : The
^ = Z 0X 1 1
X 0Z Z 0X 1
X 0 Y:
Thus we have
p 1
Theorem 5.3.2 For the e cient GMM estimator, n ^ !d N 0; Q0 1Q :
This estimator is e cient only in the sense that it is the best (asymptotically) in the class of
GMM estimators with this set of moment conditions.
W0 = 1 is not known in practice, but it can be estimated consistently. For any W ! W ;
n p 0
^
we still call the e cient GMM estimator, as it has the same asymptotic distribution.
56
We have described the estimator ^ as \e cient GMM" if the optimal (variance minimizing)
weight matrix is selected. This is a weak concept of optimality, as we are only considering alter-
native weight matrices Wn : However, it turns out that the GMM estimator is semiparametrically
e cient, as shown by Gary Chamberlain (1987).
If it is known that E (gi ( )) = 0; and this is all that is known, this is a semi-parametric
problem, as the distribution of the data is unknown. Chamberlain showed that in this context,
no semiparametric estimator (one which is consistent globally for the class of models considered)
1
can have a smaller asymptotic variance than G0 1 G : Since the GMM estimator has this
asymptotic variance, it is semiparametrically e cient.
This results shows that in the linear model, no estimator has greater asymptotic e ciency than
the e cient linear GMM estimator. No estimator can do better (in this rst-order asymptotic
sense), without imposing additional assumptions.
Then Wn !p 1
= W0 ; and GMM using Wn as the weight matrix is asymptotically e cient.
A common alternative choice is to set
n
! 1
1X 0
Wn = g^i g^i
n
i=1
which uses the uncentered moment conditions. Since Egi = 0; these two estimators are asymptot-
ically equivalent under the hypothesis of correct speci cation. However, Alastair Hall (2000) has
shown that the uncentered estimator is a poor choice. When constructing hypothesis tests, under
the alternative hypothesis the moment conditions are violated, i.e. Egi 6= 0; so the uncentered
estimator will contain an undesirable bias term and the power of the test will be adversely a ected.
A simple solution is to use the centered moment conditions to construct the weight matrix, as in
(5.4) above.
57
Here is a simple way to compute the e cient GMM estimator: First, set Wn = (X 0 X) 1 ,
estimate ^ using this weight matrix, and construct the residual e^i = yi zi0 ^ : Then set g^i = xi e^i ;
and let g^ be the associated n ` matrix: Then the e cient GMM estimator is
1 1 1
^ = Z 0 X g^0 g^ ng n g 0n X 0Z Z 0 X g^0 g^ ng n g 0n X 0 Y:
In most cases, when we say \GMM", we actually mean \e cient GMM". There is little point in
using an ine cient GMM estimator as it is easy to compute.
An estimator of the asymptotic variance of ^ can be seen from the above formula. Set
1 1
V^ = n Z 0 X g^0 g^ ng n g 0n X 0Z :
Asymptotic standard errors are given by the square roots of the diagonal elements of V^ :
There is an important alternative to the two-step GMM estimator just described. Instead, we
can let the weight matrix be considered as a function of : The criterion function is then
n
! 1
0 1X
J( ) = n g n ( ) gi ( )gi ( )0 g n ( ):
n
i=1
where
gi ( ) = gi ( ) gn( )
The ^ which minimizes this function is called the continuously-updated GMM estimator,
and was introduced by L. Hansen, Heaton and Yaron (1996).
The estimator appears to have some better properties than traditional GMM, but can be
numerically tricky to obtain in some cases. This is a current area of research in econometrics.
58
and !
n 1
1X 0
Wn = g^i g^i g n g 0n ;
n
i=1
with g^i = gi ( ~ ) constructed using a preliminary consistent estimator ~ , perhaps obtained by rst
setting Wn = I: Since the GMM estimator depends upon the rst-stage estimator, often the weight
matrix Wn is updated, and then ^ recomputed. This estimator can be iterated if needed.
p 1
Theorem 5.5.1 Under general regularity conditions, n ^ !d N 0; G0 1G ; where
1
= (E (gi gi0 )) 1 and G = E @@ 0 gi ( ): The variance of ^ may be estimated by G
^0 ^ ^
1G where
^ =n 1 P ^=n 1 P @ ^
ig ^i g^i 0 and G i @ 0 gi ( ):
The general theory of GMM estimation and testing was exposited by L. Hansen (1982).
Eg(wi ; ) = 0
holds. Thus the model { the overidentifying restrictions { are testable.
For example, take the linear model yi = 01 x1i + 02 x2i + ei with E (x1i ei ) = 0 and E (x2i ei ) = 0:
It is possible that 2 = 0; so that the linear equation may be written as yi = 01 x1i + ei : However,
it is possible that 2 6= 0; and in this case it would be impossible to nd a value of 1 so that
both E (x1i (yi x01i 1 )) = 0 and E (x2i (yi x01i 1 )) = 0 hold simultaneously. In this sense an
exclusion restriction can be seen as an overidentifying restriction.
Note that g n !p Egi ; and thus g n can be used to assess whether or not the hypothesis that
Egi = 0 is true or not. The criterion function at the parameter estimates is
J = n g 0n Wn g n
1
= n2 g 0n g^0 g^ ng n g 0n gn:
Theorem 5.6.1 (Sargan-Hansen). Under the hypothesis of correct speci cation, and if the weight
matrix is asymptotically e cient,
J = J( ^ ) !d 2` k :
59
The proof of the theorem is left as an exercise. This result was established by Sargan (1958)
for a specialized case, and by L. Hansen (1982) for the general case.
The degrees of freedom of the asymptotic distribution are the number of overidentifying re-
strictions. If the statistic J exceeds the chi-square critical value, we can reject the model. Based on
this information alone, it is unclear what is wrong, but it is typically cause for concern. The GMM
overidenti cation test is a very useful by-product of the GMM methodology, and it is advisable to
report the statistic J whenever GMM is the estimation method.
When over-identi ed models are estimated by GMM, it is customary to report the J statistic
as a general test of model adequacy.
J( ) = n g n ( )0 Wn g n ( )
H0 : h( ) = 0:
The two minimizing criterion functions are J( ^ ) and J( ~ ): The GMM distance statistic is the
di erence
D = J( ~ ) J( ^ ):
Proposition 5.7.1 If the same weight matrix Wn is used for both null and alternative,
1. D 0
2. D !d 2
r
60
If h is non-linear, the Wald statistic can work quite poorly. In contrast, current evidence
suggests that the D statistic appears to have quite good sampling properties, and is the preferred
test statistic.
Newey and West (1987) suggested to use the same weight matrix Wn for both null and alter-
native, as this ensures that D 0. This reasoning is not compelling, however, and some current
research suggests that this restriction is not necessary for good performance of the test.
This test shares the useful feature of LR tests in that it is a natural by-product of the compu-
tation of alternative models.
61
and
2
i = E ei ( )2 j xi :
Then the \optimal instrument" is
2
Ai = i Ri
so the optimal moment is
gi ( ) = Ai ei ( ):
Setting gi ( ) to be this choice (which is k 1; so is just-identi ed) yields the best GMM estimator
possible.
In practice, Ai is unknown, but its form does help us think about construction of optimal
instruments.
In the linear model ei ( ) = yi zi0 ; note that
Ri = E (zi j xi )
and
2
i = E e2i j xi ;
so
2
Ai = i E (zi j xi ) :
In the case of linear regression, zi = xi ; so Ai = i 2 xi : Hence e cient GMM is GLS, as we
discussed earlier in the course.
In the case of endogenous variables, note that the e cient instrument Ai involves the estimation
of the conditional mean of zi given xi : In other words, to get the best instrument for zi ; we need the
best conditional mean model for zi given xi , not just an arbitrary linear projection. The e cient
instrument is also inversely proportional to the conditional variance of ei : This is the same as the
GLS estimator; namely that improved e ciency can be obtained if the observations are weighted
inversely to the conditional variance of the errors.
62
Chapter 6
Empirical Likelihood
where is a Lagrange multiplier. The n rst order conditions are 0 = pi 1 : Combined with
1
the constraint (6.1) we nd that the MLE is pi = n yielding the log-likelihood n log(n):
Now consider the case of an overidenti ed model with moment condition
Egi ( 0) =0
63
where g is ` 1 and is k 1 and for simplicity we write gi ( ) = g(yi ; zi ; xi ; ): The multinomial
distribution which places probability pi at each observation (yi ; xi ; zi ) will satisfy this condition if
and only if
X n
pi gi ( ) = 0 (6.3)
i=1
The empirical likelihood estimator is the value of which maximizes the multinomial log-
likelihood (6.2) subject to the restrictions (6.1) and (6.3).
The Lagrangian for this maximization problem is
n n
! n
X X X
0
Ln ( ; p1 ; :::; pn ; ; ) = ln(pi ) pi 1 n pi gi ( )
i=1 i=1 i=1
where and are Lagrange multipliers. The rst-order-conditions of Ln with respect to pi , ;
and are
1
= + n 0 gi ( )
pi
n
X
pi = 1
i=1
n
X
pi gi ( ) = 0:
i=1
Multiplying the rst equation by pi , summing over i; and using the second and third equations,
we nd = n and
1
pi = :
n 1 + 0 gi ( )
Substituting into Ln we nd
n
X
0
Rn ( ; ) = n ln (n) ln 1 + gi ( ) : (6.4)
i=1
This minimization problem is the dual of the constrained maximization problem. The solution
(when it exists) is well de ned since Rn ( ; ) is a convex function of : The solution cannot be
obtained explicitly, but must be obtained numerically (see section 6.5). This yields the (pro le)
empirical log-likelihood function for .
Ln ( ) = Rn ( ; ( ))
n
X
= n ln (n) ln 1 + ( )0 gi ( )
i=1
64
The EL estimate ^ is the value which maximizes Ln ( ); or equivalently minimizes its negative
^ = argmin [ Ln ( )] (6.6)
and
1
V = G0 1
G (6.9)
1
V = G G0 1
G G0 (6.10)
For example, in the linear model, Gi (; ) = xi zi0 ; G = E (xi zi0 ), and = E xi x0i e2i :
The asymptotic variance V for ^ is the same as for e cient GMM. Thus the EL estimator is
asymptotically e cient.
65
Proof. ( ^ ; ^ ) jointly solve
@
n
X gi ^
0 = Rn ( ; ) = 0
(6.11)
@ 1 + ^ gi ^
i=1
0
@
n
X Gi ; ^
0 = Rn ( ; ) = 0
: (6.12)
@ 1 + ^ gi ^ i=1
P P P
Let Gn = n1 ni=1 Gi ( 0 ) ; g n = n1 ni=1 gi ( 0 ) and n = n1 ni=1i g ( 0 ) gi ( 0
0) :
Expanding (6.12) around = 0 and = 0 = 0 yields
0 ' G0n ^ 0 : (6.13)
Expanding (6.11) around = 0 and = 0 = 0 yields
0' gn Gn ^ 0 + n
^ (6.14)
Premultiplying by G0n n
1 and using (6.13) yields
0 ' G0n n
1
gn G0n n
1
Gn ^ 0 + G0n n
1
n
^
= G0n n
1
gn G0n n
1
Gn ^ 0
66
6.3 Overidentifying Restrictions
In a parametric likelihood context, tests are based on the di erence in the log likelihood functions.
The same statistic can be constructed for empirical likelihood. Twice the di erence between the
unrestricted empirical likelihood n log (n) and the maximized empirical likelihood for the model
(6.7) is
Xn
0
LRn = 2 ln 1 + ^ gi ^ : (6.17)
i=1
The EL overidenti cation test is similar to the GMM overidenti cation test. They are asymp-
totically rst-order equivalent, and have the same interpretation. The overidenti cation test is a
very useful by-product of EL estimation, and it is advisable to report the statistic LRn whenever
EL is the estimation method.
Proof. First, by a Taylor expansion, (6.15), and (6.16),
n
1 X p
p g wi ; ^ ' n g n + Gn ^ 0
n
i=1
1 p
' I Gn G0n n
1
Gn G0n n
1
ng n
p
' n n^:
6.4 Testing
Let the maintained model be
Egi ( ) = 0 (6.18)
67
where g is ` 1 and is k 1: By \maintained" we mean that the overidentfying restrictions
contained in (6.18) are assumed to hold and are not being challenged (at least for the test discussed
in this section). The hypothesis of interest is
h( ) = 0:
where h : Rk ! Ra : The restricted EL estimator and likelihood are the values which solve
~ = argmax Ln ( )
h( )=0
L~n = Ln ( ~ ) = max Ln ( ):
h( )=0
http://www.ssc.wisc.edu/~bhansen/progs/elike.prc
6.5.1 Derivatives
The numerical calculations depend on derivatives of the dual likelihood function (6.4). De ne
gi ( )
gi ( ; ) =
1 + 0 gi ( )
Gi ( )0
Gi ( ; ) =
1 + 0 gi ( )
68
The rst derivatives of (6.4) are
n
X
@
R = Rn ( ; ) = gi ( ; )
@
i=1
n
X
@
R = Rn ( ; ) = Gi ( ; ) :
@
i=1
A quadratic function can be t exactly through these three points. The value of which minimizes
this quadratic is
^ = R2 + 3R0 4R1 :
4R2 + 4R0 8R1
yielding the steplength to be plugged into (6.19)..
A complication is that must be constrained so that 0 pi 1 which holds if
0
n 1+ gi ( ) 1 (6.20)
for all i: If (6.20) fails, the stepsize needs to be decreased.
69
6.5.3 Outer Loop
The outer loop is the minimization (6.6). This can be done by the modi ed Newton method
described in the previous section. The gradient for (6.6) is
@ @ 0
L = Ln ( ) = Rn ( ; ) = R + R =R
@ @
since R ( ; ) = 0 at = ( ); where
@ 1
= ( )= R R ;
@ 0
the second equality following from the implicit function theorem applied to R ( ; ( )) = 0:
The Hessian for (6.6) is
@
L = Ln ( )
@ @ 0
@ 0
= R ( ; ( )) + R ( ; ( ))
@ 0
= R ( ; ( )) + R0 + 0
R + 0
R
0 1
= R R R R :
It is not guaranteed that L > 0: If not, the eigenvalues of L should be adjusted so that all are
positive. The Newton iteration rule is
1
j+1 = j L L
70
Chapter 7
Endogeneity
We say that there is endogeneity in the linear model y = zi0 + ei if is the parameter of interest
and E(zi ei ) 6= 0: This cannot happen if is de ned by linear projection, so requires a structural
interpretation. The coe cient must have meaning separately from the de nition of a conditional
mean or linear projection.
Example: Measurement error in the regressor. Suppose that (yi ; xi ) are joint random
variables, E(yi j xi ) = xi 0 is linear, is the parameter of interest, and xi is not observed. Instead
we observe xi = xi + ui where ui is an k 1 measurement error, independent of yi and xi : Then
yi = xi 0 + ei
= (xi ui )0 + ei
= x0i + vi
where
vi = ei u0i :
The problem is that
^ !p 1
= E xi x0i E ui u0i 6= :
71
e1i
Assume that ei = is iid, Eei = 0; 1 + 2 = 1 and Eei e0i = I2 (the latter for simplicity).
e2i
The question is, if we regress qi on pi ; what happens?
It is helpful to solve for qi and pi in terms of the errors. In matrix notation,
1 1 qi e1i
=
1 2 pi e2i
so
1
qi 1 1 e1i
=
pi 1 2 e2i
2 1 e1i
=
1 1 e2i
2 e1i + 1 e2i
= :
(e1i e2i )
qi = pi + " i
E (pi "i ) = 0
where
E (pi qi ) 2 1
= =
E p2i 2
Hence if it is estimated by OLS, ^ !p ; which does not equal either 1 or 2: This is called
simultaneous equations bias.
Y = Z + e: (7.2)
Any solution to the problem of endogeneity requires additional information which we call
instruments.
De nition 7.1.1 The ` 1 random vector xi is an instrumental variable for (7.1) if E (xi ei ) = 0:
72
In a typical set-up, some regressors in zi will be uncorrelated with ei (for example, at least the
intercept). Thus we make the partition
z1i k1
zi = (7.3)
z2i k2
where E(z1i ei ) = 0 yet E(z2i ei ) 6= 0: We call z1i exogenous and z2i endogenous. By the above
de nition, z1i is an instrumental variable for (7.1); so should be included in xi : So we have the
partition
z1i k1
xi = (7.4)
x2i `2
where z1i = x1i are the included exogenous variables, and x2i are the excluded exogenous
variables. That is x2i are variables which could be included in the equation for yi (in the sense
that they are uncorrelated with ei ) yet can be excluded, as they would have true zero coe cients
in the equation.
The model is just-identi ed if ` = k (i.e., if `2 = k2 ) and over-identi ed if ` > k (i.e., if
`2 > k2 ):
We have noted that any solution to the problem of endogeneity requires instruments. This
does not mean that valid instruments actually exist.
u i = zi x0i
as the projection error. Then the reduced form linear relationship between zi and xi is
0
zi = xi + ui : (7.5)
Z =X +u (7.6)
where u is n k:
By construction,
E(xi u0i ) = 0;
73
so (7.5) is a projection and can be estimated by OLS:
Z = X^ + u
^
^ = X X 1 X 0Z :
0
Y = (X + u) +e
= X + v; (7.7)
where
= (7.8)
and
v = u + e:
Observe that
E (xi vi ) = E xi u0i + E (xi ei ) = 0:
Thus (7.7) is a projection equation and may be estimated by OLS. This is
Y = X ^ + v^;
^ = X 0X 1 X 0Y
The equation (7.7) is the reduced form for Y: (7.6) and (7.7) together are the reduced form
equations for the system
Y = X +v
Z = X + u:
rank( ) = k: (7.9)
Assume that (7.9) holds. If ` = k; then = 1 : If ` > k; then for any W > 0; =
0 1 0
( W ) W :
If (7.9) is not satis ed, then cannot be recovered from ( ; ): Note that a necessary (although
not su cient) condition for (7.9) is ` k:
74
Since X and Z have the common variables X1 ; we can rewrite some of the expressions. Using
(7.3) and (7.4) to make the matrix partitions X = [X1 ; X2 ] and Z = [X1 ; Z2 ]; we can partition
as
11 12
=
21 22
I 12
=
0 22
Z1 = X1
Z2 = X1 12 + X2 22 + u2 : (7.10)
is identi ed if rank( ) = k; which is true if and only if rank( 22 ) = k2 (by the upper-diagonal
structure of ): Thus the key to identi cation of the model rests on the `2 k2 matrix 22 in (7.10).
7.4 Estimation
The model can be written as
yi = zi0 + ei
E (xi ei ) = 0
or
Eg (wi ; ) = 0
g (wi ; ) = xi yi zi0 :
This a moment condition model. Appropriate estimators include GMM and EL. The estimators
and distribution theory developed in those Chapter 8 and 9 directly apply. Recall that the GMM
estimator, for given weight matrix Wn ; is
^ = Z 0 XWn X 0 Z 1
Z 0 XWn X 0 Y:
75
This estimator is often called the instrumental variables estimator (IV) of ; where X is used
as an instrument for Z: Observe that the weight matrix Wn has disappeared. In the just-identi ed
case, the weight matrix places no role. This is also the MME estimator of ; and the EL estimator.
Another interpretation stems from the fact that since = 1 ; we can construct the Indirect
1 1 1
= X 0X X 0Z X 0X X 0Y
1 1
= X 0Z X 0X X 0X X 0Y
1
= X 0Z X 0Y :
Recall that the optimal weight matrix is an estimate of the inverse of = E xi x0i e2i : In the
special case that E e2i j xi = 2 (homoskedasticity), then = E (xi x0i ) 2 / E (xi x0i ) suggesting
the weight matrix Wn = (X 0 X) 1 : Using this choice, the GMM estimator equals
1 1 1
^ = Z 0X X 0X X 0Z Z 0X X 0X X 0Y
2SLS
This is called the two-stage-least squares (2SLS) estimator. It was originally proposed by Theil
(1953) and Basmann (1957), and is the classic estimator for linear equations with instruments.
Under the homoskedasticity assumption, the 2SLS estimator is e cient GMM, but otherwise it is
ine cient.
It is useful to observe that writing
1
PX = X X 0 X X 0;
1
Z^ = PX Z = X X 0 X X 0 Z;
then
^ = 1
Z 0 PX Z Z 0 PX Y
= Z^ 0 Z^ Z^ 0 Y:
76
^ Recall, Z = [Z1 ; Z2 ] and X = [Z1 ; X2 ]: Then
It is useful to scrutinize the projection Z:
h i
Z^ = Z^1 ; Z^2
= [PX Z1 ; PX Z2 ]
= [Z1 ; PX Z2 ]
h i
= Z1 ; Z^2 ;
since Z1 lies in the span of X: Thus in the second stage, we regress Y on Z1 and Z^2 : So only the
endogenous variables Z2 are replaced by their tted values:
Z^2 = X1 ^ 12 + X2 ^ 22 :
77
which is zero only when S21 = 0 (when Z is exogenous).
We now derive a similar result for the 2SLS estimator.
^ 1
2SLS = Z 0 PX Z Z 0 PX Y :
where
l
= :
n
Using (7.12) and this result,
1 1 1
E Z 0 PX e = E 0
X 0e + E u0 PX e = S21 ;
n n n
and
1 1
E Z 0 PX Z = 0
E xi x0i + 0
E (xi ui ) + E ui x0i + E u 0 PX u
n n
0
= Q + S22 :
Together
1
1 0 1 0
E ^ 2SLS E Z PX Z E Z PX e
n n
0 1
= Q + S22 S21 : (7.14)
In general this is non-zero, except when S21 = 0 (when Z is exogenous). It is also close to zero
when = 0. Bekker (1994) pointed out that it also has the reverse implication { that when
= l=n is large, the bias in the 2SLS estimator will be large. Indeed as ! 1; the expression in
(7.14) approaches that in (7.13), indicating that the bias in 2SLS approaches that of OLS as the
number of instruments increases.
Bekker (1994) showed further that under the alternative asymptotic approximation that is
xed as n ! 1 (so that the number of instruments goes to in nity proportionately with sample
size) then the expression in (7.14) is the probability limit of ^ 2SLS
78
7.7 Identi cation Failure
Recall the reduced form equation
Z2 = X1 12 + X2 22 + u2 :
The parameter fails to be identi ed if 22 has de cient rank. The consequences of identi cation
failure for inference are quite severe.
Take the simplest case where k = l = 1 (so there is no X1 ): Then the model may be written as
yi = zi + ei
zi = xi + ui
and 22 = = E (xi zi ) =Ex2i : We see that is identi ed if and only if 22 = 6= 0; which occurs
when E (zi xi ) 6= 0. Thus identi cation hinges on the existence of correlation between the excluded
exogenous variable and the included endogenous variable.
Suppose this condition fails, so E (zi xi ) = 0: Then by the CLT
n
1 X
p xi ei !d N1 N 0; E x2i e2i (7.15)
n
i=1
n n
1 X 1 X
p xi zi = p xi ui !d N2 N 0; E x2i u2i (7.16)
n n
i=1 i=1
therefore Pn
p1
n i=1 xi ei N1
^ = Pn !d Cauchy;
p1 N2
n i=1 xi zi
since the ratio of two normals is Cauchy. This is particularly nasty, as the Cauchy distribution
does not have a nite mean. This result carries over to more general settings, and was examined
by Phillips (1989) and Choi and Phillips (1992).
Suppose that identi cation does not complete fail, but is weak. This occurs when 22 is full
rank, but small. This can be handled in an asymptotic analysis by modeling it as local-to-zero,
viz
1=2
22 = n C;
where C is a full rank matrix. The n 1=2 is picked because it provides just the right balancing to
allow a rich distribution theory.
To see the consequences, once again take the simple case k = l = 1: Here, the instrument xi is
weak for zi if
= n 1=2 c:
79
Then (7.15) is una ected, but (7.16) instead takes the form
n n n
1 X 1 X 2 1 X
p xi zi = p xi + p xi ui
n n n
i=1 i=1 i=1
Xn n
X
1 1
= x2i c +p xi ui
n n
i=1 i=1
!d Qc + N2
therefore
^ N1
!d :
Qc + N2
As in the case of complete identi cation failure, we nd that ^ is inconsistent for and the
asymptotic distribution of ^ is non-normal. In addition, standard test statistics have non-standard
distributions, meaning that inferences about parameters of interest can be misleading.
The distribution theory for this model was developed by Staiger and Stock (1997) and extended
to nonlinear GMM estimation by Stock and Wright (2000). Further results on testing were obtained
by Wang and Zivot (1998).
The bottom line is that it is highly desirable to avoid identi cation failure. Once again, the
equation to focus on is the reduced form
Z2 = X1 12 + X2 22 + u2
80
Chapter 8
The Bootstrap
Gn (x; F ) = P (Tn x j F)
Given the structure of the problem, Gn is a function only of F; the distribution of (yi ; xi ) : In
general, Gn (x; F ) depends on F; meaning that G changes as F changes.
Ideally, inference on would be based on Gn (x; F0 ); the true value of the sampling distribution.
This is generally impossible for two reasons. First, the function Gn (x; F ) is unknown. Second, F0
is unknown.
The idea of Monte Carlo simulation is to solve the rst problem through numerical simulation.
The idea is that for any given F; the distribution function Gn (x; F ) can be calculated numerically
through simulation. Since F is unknown, this does not solve the problem of inference. Instead,
the method is typically used to assess the adequacy of statistical methods in practical settings.
The name Monte Carlo derives from the famous Mediterranean gambling resort, where games
of chance are played.
The method of Monte Carlo is quite simple to describe. The researcher chooses F0 (the dis-
tribution of the data) and the sample size n. A \true" value of 0 is implied by this choice, or
equivalently the value 0 is selected directly by the researcher.
Then the following experiment is conducted
81
n independent random vectors yi ; xi ; i = 1; :::; n; are drawn from the distribution F using
the computer's random number generator.
For step 1, most computer packages have built-in procedures for generating U [0; 1] and N (0; 1)
random numbers, and from these most random variables can be constructed. (For example, a
chi-square can be generated by sums of squares of normals.)
For step 2, it is important that the statistic be evaluated at the \true" value of corresponding
to the choice of F:
The above experiment creates one random draw from the distribution Gn (x; F ): This is one
observation from an unknown distribution. Clearly, from one observation very little can be said.
So the researcher repeats the experiment B times, where B is a large number. Typically, we set
B = 1000 or B = 5000; and we will discuss this choice later.
Notationally, let the b0 th experiment result in the draw Tnb ; b = 1; :::; B: These results are
stored. They constitute a random sample of size B from the distribution of Gn (x; F ) = P (Tnb x) =
P (Tn x j F ) :
From a random sample, we can estimate any feature of interest using (typically) a method of
moments estimator. For example:
Suppose we are interested in the bias, mean-squared error (MSE), or variance of the distribution
of ^ : We then set Tn = ^ ; run the above experiment, and calculate
B
\^) = 1 X
Bias( Tnb
B
b=1
XB
1
M\
SE(^) = (Tnb )2
B
b=1
2
V\
ar(^) = M\
SE(^) \^)
Bias(
Suppose we are interested in the Type I error associated with an asymptotic 5% two-sided
t-test. We would then set Tn = ^ =s(^) and calculate
B
^ 1 X
P = 1 (Tnb 1:96) ; (8.1)
B
b=1
the percentage of the simulated t-ratios which exceed the asymptotic 5% critical value.
Suppose we are interested in the 5% and 95% quantile of ^ or ^ =s(^): We then set Tn to
either choice, and compute the 10% and 90% sample quantiles of the sample fTnb g: The % sample
quantile is a number q such that % of the sample are less than q : A simple way to compute
82
sample quantiles is to sort the sample fTnb g from low to high. Then q is the N 'th number in
this ordered sequence, where N = (B + 1) : It is therefore convenient to pick B so that N is an
integer. For example, if we set B = 999; then the 5% sample quantile is 50'th sorted value and
the 95% sample quantile is the 950'th sorted value.
The typical purpose of a monte carlo simulation is to investigate the performance of a statistical
procedure (estimator or test) in realistic settings. Generally, the performance will depend on n
and F: In many cases, an estimator or test may perform wonderfully for some values, and poorly
for others. It is therefore useful to conduct a variety of experiments, for a selection of choices of n
and F:
As discussed above, the researcher must select the number of experiments, B: Often this is
called the number of replications. Quite simply, a larger B results in more precise estimates of
the features of interest of Gn ; but requires more computational time. In practice, therefore, the
choice of B is often guided by the computational demands of the statistical procedure. However,
it should be recognized that the results of a monte carlo experiment are all estimates computed
from a random sample of size B; and therefore it is straightforward to calculate standard errors
for any quantity of interest. If the standard error is too large to make a reliable inference, then B
will have to be increased.
In particular, it is simple to make inferences about rejection probabilities from statistical tests,
such as the percentage estimate reported in (8.1). The random variable 1 (Tnb 1:96) is iid
Bernoulli, equalling 1 with probability P = E1 (Tnb 1:96) : The average (8.1) is therefore an
p
unbiased estimator of P with standard error s P^ = P (1 P ) =B. As P is unknown, this may
r
be approximated using the estimated value s P = P^ 1 P^ =B or using a hypothesized value.
^
p
For example, if we are assessing an asymptotic 5% test, then we can set s P^ = (:05) (:95) =B '
p
:22= B: Hence the standard errors for B = 100; 1000, and 5000, are, respectively, s P^ = :022;
:007; and .003.
8.2 An Example
Here we illustrate Monte Carlo simulation to investigate estimation and tests on a nonlinear
function of regression parameters.
Our model is an iid sample fyi ; x1i ; x2i : 1 i ng from the linear Gaussian regression
yi = 0 + x1i 1 + x2i 2 + ei
x1i
N (0; I2 )
x2i
2
ei N 0;
We set = 3; 0 = 0; 1 = 1; 2 = :5; and n = 300: In this example, the distribution of the data,
83
F; is determined by the above choices. We don't have to be more explicit about F; as the above
equations are su cient to specify the joint distribution.
The parameter of interest is the ratio of the regression slopes
1
= :
2
The goal is to estimate , construct con dence intervals for , and test the hypothesis that = 0:
Note that 0 = 2:
We estimate the parameters of the model by OLS, and then estimate by
^
^= 1
:
^
2
84
Inference is typically based on the t-ratio tn = (^ ^
0 )=s( ): The asymptotic approximation
to tn is N (0; 1): Using the same simulated samples, we estimated the exact density of tn ; which is
displayed along with the asymptotic distribution in the bottom section of Figure 1.
Once again we nd that the divergence between the exact and asymptotic distributions to be
dramatic. The exact distribution is highly skewed and non-normal. The accuracy of hypothesis
testing and con dence interval coverage depends on tail probabilities. We calculate the exact prob-
abilities P (tn > 1:645) = 0:00; P (tn < 1:645) = :115 and P (jtn j > 1:96) = :084; meaning that
both one-tailed and two-tailed tests and con dence intervals based on asymptotic approximations
have signi cant Type I error. (The nominal, or asymptotic, Type I error for each of these tests is
.050).
This example shows that asymptotic approximations may be quite poor, even in very simple
regression models with reasonably large samples.
85
8.3 The Empirical Distribution Function
The observations are wi ; i; :::; n; e.g. wi = (yi ; xi ) with CDF F (w) = P (wi w) and true value
F0 : Note that F0 (w) = P (wi w) = E1 (wi w) ; where 1( ) is the indicator function, so F0 (w)
can be expressed as a population moment. The MME is the corresponding sample moment:
n
1X
Fn (w) = 1 (wi w) :
n
i=1
To see the e ect of sample size on the EDF, in Figure 2, I have plotted the EDF and true
CDF for three random samples of size n = 25; 50, and 100. The random draws are from the
N (0; 1) distribution. For n = 25; the EDF is only a crude approximation to the CDF, but the
approximation appears to improve for the large n. Yet even for n = 100; there is a signi cant
divergence between the EDF and CDF around w = :6 in this sample. In general, as the sample
size gets larger, the EDF step function gets uniformly close to the true CDF.
86
8.4 De nition of the Bootstrap
Let Tn = Tn (w1 ; :::; wn ; ) be a statistic of interest with CDF Gn (x; F ) = P (Tn x j F ): For exact
inference on , we desire Gn (x; F0 ): This is impossible since F0 is unknown.
In a seminal contribution, Efron (1979) observed that since F0 can be estimated by Fn ;
Gn (x; F0 ) can be estimated by plugging Fn into Gn (x; F ) to get the estimator
87
pretends that the distribution of wi is Fn rather than F0 ; but it also pretends that the true value
of the parameter2 is the sample estimate ^; rather than 0 :
The EDF is a valid discrete probability distribution which puts probability mass 1=n at each
point wi = (yi ; xi ), i = 1; :::; n: Notationally, it will be helpful to think of a random variable wi
with the distribution Fn : That is,
P (wi x) = Fn (x):
We can easily calculate the moments of wi :
Z
Eh(wi ) = h(w)dFn (w)
n
X
= h(wi )P (w = wi )
i=1
n
1X
= h(wi );
n
i=1
P (Tn x) = Gn (x):
We call wi and Tn the bootstrap distribution of the data and statistic. Tn is the correct analog of
Tn when the true CDF is Fn ; as the data wi are sampled from the CDF Fn and the parameter ^n
is the true value under Fn :
Since the EDF Fn is a multinomial (with n support points), in principle the distribution of Tn
could be calculated by direct methods. Since there are 2n possible samples fw1 ; :::; wn g; however,
such a calculation is computationally infeasible unless n is very small. The popular alternative is
to use Monte Carlo simulation to approximate the distribution. The algorithm is identical to our
discussion of Monte Carlo simulation, with the following points of clari cation:
The sample size n used for the simulation is the same as the sample size
The random vectors wi are drawn randomly from the distribution function Fn
The generation of random vectors from Fn depends on the form of Fn : When Fn equals the
EDF, this is particularly simple. Recall that Fn is a discrete probability distribution putting
probability mass 1=n at each sample point wi : Thus a random draw from Fn is just a random
draw from the sample fw1 ; :::; wn g: For a bootstrap sample we need n independent random draws
2
More precisely, the bootstrap pretends that the true value of is the value consistent with Fn ; the estimate of
F used to construct the bootstrap. In most cases, and the ones we consider, this value of is the sample estimate ^:
88
from Fn : This requires that we make n independent random draws from the sample fw1 ; :::; wn g
with replacement. In consequence, a bootstrap sample fw1 ; :::; wn g will necessarily have some
ties and multiple values, which is generally not a problem.
A theory for the determination of the number of bootstrap replications B has been developed
by Andrews and Buchinsky (2000).
Let Tn ( ) = ^ : Then
n = E(Tn ( 0 )):
The bootstrap counterparts are ^ = ^(w1 ; :::; wn ) and Tn = ^ n = ^ ^: The bootstrap
estimate of n is
n = E(Tn ):
If this is calculated by the simulation described in the previous subsection, the estimate of n
is
B
1 X
^n = Tnb
B
b=1
XB
1 ^ ^
= b
B
b=1
= ^ ^:
= 2^ ^ :
Note, in particular, that the biased-corrected estimator is not ^ : Intuitively, the bootstrap makes
the following experiment. Suppose that ^ is the truth. Then what is the average value of ^
calculated from such samples? The answer is ^ : If this is lower than ^; this suggests that the
estimator is downward-biased, so a biased-corrected estimator of should be larger than ^; and
89
the best guess is the di erence between ^ and ^ : Similarly if ^ is higher than ^; then the estimator
is upward-biased and the biased-corrected estimator should be lower than ^.
Let Tn = ^: The variance of ^ is
Vn = E(Tn ETn )2 :
A bootstrap
q standard error for ^ is the square root of the bootstrap estimate of variance,
s( ^ ) = V^n :
While this standard error may be calculated and reported, it is not clear if it is useful. The
primary use of asymptotic standard errors is to construct asymptotic con dence intervals, which
are based on the asymptotic normal approximation to the t-ratio. However, the use of the boot-
strap presumes that such asymptotic approximations might be poor, in which case the normal
approximation is suspected. It appears superior to calculate bootstrap con dence intervals, and
we turn to this next.
90
The interval C1 is a popular bootstrap con dence interval often used in empirical practice.
This is because it is easy to compute, simple to motivate, was popularized by Efron early in the
history of the bootstrap, and also has the feature that it is translation invariant. That is, if we
de ne = f ( ) as the parameter of interest for a monotonic function f; then percentile method
applied to this problem will produce the con dence interval [f (qn ( =2)); f (qn (1 =2))]; which
is a naturally good property.
However, as we show now, C1 is in a deep sense very poorly motivated.
It will be useful if we introduce an alternative de nition C1 . Let Tn ( ) = ^ and let qn ( )
be the quantile function of its distribution. (These are the original quantiles, with subtracted.)
Then C1 can alternatively be written as
C1 = [^ + qn ( =2); ^ + q (1
n =2)]:
P 0 2 C10 = P ^ + qn ( =2) 0
^ + qn (1 =2)
= P qn (1 =2) ^ 0 qn ( =2)
= Gn ( qn ( =2); F0 ) Gn ( qn (1 =2); F0 )
which generally is not 1 ! There is one important exception. If ^ 0 has a symmetric distribution,
then Gn ( x; F0 ) = 1 Gn (x; F0 ); so
and this idealized con dence interval is accurate. Therefore, C10 and C1 are designed for the case
that ^ has a symmetric distribution about 0 :
When ^ does not have a symmetric distribution, C1 may perform quite poorly.
However, by the translation invariance argument presented above, it also follows that if there
exists some monotonic transformation f ( ) such that f (^) is symmetrically distributed about f ( 0 );
then the idealized percentile bootstrap method will be accurate.
Based on these arguments, many argue that the percentile interval should not be used unless
the sampling distribution is close to unbiased and symmetric.
The problems with the percentile method can be circumvented by an alternative method.
91
Let Tn ( ) = ^ . Then
1 = P (qn ( =2) Tn ( 0 ) qn (1 =2))
= P ^ qn (1 =2) 0
^ qn ( =2) ;
P (Tn ( 0 ) < c) = :
Thus c = qn ( ): Since this is unknown, a bootstrap test replaces qn ( ) with the bootstrap estimate
qn ( ); and the test rejects if Tn ( 0 ) < qn ( ):
Similarly, if the alternative is H1 : > 0 ; the bootstrap test rejects if Tn ( 0 ) > qn (1 ):
Computationally, these critical values can be estimated from a bootstrap simulation by sorting
the bootstrap t-statistics Tn = ^ ^ =s(^ ): Note, and this is important, that the bootstrap test
statistic is centered at the estimate ^; and the standard error s(^ ) is calculated on the bootstrap
sample. These t-statistics are sorted to nd the estimated quantiles q^n ( ) and/or q^n (1 ):
Let Tn ( ) = ^ =s(^). Then
= P ^ s(^)qn (1 =2) 0
^ s(^)qn ( =2) ;
92
so an exact (1 )% con dence interval for 0 would be
This is often called a percentile-t con dence interval. It is equal-tailed or central since the proba-
bility that 0 is below the left endpoint approximately equals the probability that 0 is above the
right endpoint, each =2:
Computationally, this is based on the critical values from the one-sided hypothesis tests, dis-
cussed above.
P (jTn ( 0 )j > c) = :
Note that
which is a symmetric distribution function. The ideal critical value c = qn ( ) solves the equation
Gn (qn ( )) = 1 :
Gn (qn ( )) = Gn (qn ( )) Gn ( qn ( )) = 1 :
93
where qn ( ) is the bootstrap critical value for a two-sided hypothesis test. C4 is called the sym-
metric percentile-t interval. It is designed to work well since
P( 0 2 C4 ) = P ^ s(^)qn ( ) 0
^ + s(^)q ( )
n
= P (jTn ( 0 )j < qn ( ))
' P (jTn ( 0 )j < qn ( ))
= 1 :
or some other asymptotically chi-square statistic. Thus here Tn ( ) = Wn ( ): The ideal test rejects
if Wn qn ( ); where qn ( ) is the (1 )% quantile of the distribution of Wn : The bootstrap test
rejects if Wn qn ( ); where qn ( ) is the (1 )% quantile of the distribution of
0
Wn = n ^ ^ V^ 1 ^ ^ :
Computationally, the critical value qn ( ) is found as the quantile from simulated values of Wn :
Note in the simulation that the Wald statistic is a quadratic form in ^ ^ ; not ^ 0 :
[This is a typical mistake made by practitioners.]
94
De nition 8.9.2 an = O(1) if jan j is uniformly bounded.
We can interpret Theorem 8.9.1 as follows. First, Gn (x; F ) converges to the normal limit at
rate n1=2 : To a second order of approximation,
x 1=2
Gn (x; F ) +n g1 (x; F ):
v
Since the derivative of g1 is odd, the density function is skewed. To a third order of approximation,
x 1=2 1
Gn (x; F ) +n g1 (x; F ) + n g2 (x; F )
v
which adds a symmetric non-normal component to the approximate density (for example, adding
leptokurtosis).
95
so the order of the error is O(n 1=2 ):
A bootstrap test is based on Gn (x); which from Theorem 8.9.1 has the expansion
1 1
Gn (x) = Gn (x; Fn ) = (x) + g1 (x; Fn ) + O(n ):
n1=2
Because (x) appears in both expansions, the di erence between the bootstrap distribution and
the true distribution is
1 1
Gn (x) Gn (x; F0 ) = (g1 (x; Fn ) g1 (x; F0 )) + O(n ):
n1=2
p
Since Fn converges to F at rate n; and g1 is continuous with respect to F; the di erence
p
(g1 (x; Fn ) g1 (x; F0 )) converges to 0 at rate n: Heuristically,
@
g1 (x; Fn ) g1 (x; F0 ) g1 (x; F0 ) (Fn F0 )
@F
= O(n 1=2 );
@
The \derivative" @F g1 (x; F ) is only heuristic, as F is a function. We conclude that
1
Gn (x) Gn (x; F0 ) = O(n );
or
1
P (Tn x) = P (Tn x) + O(n );
which is an improved rate of convergence over the asymptotic test (which converged at rate
O(n 1=2 )). This rate can be used to show that one-tailed bootstrap inference based on the t-
ratio achieves a so-called asymptotic re nement { the Type I error of the test converges at a faster
rate than an analogous asymptotic test.
P (jXj x) = P ( x X x)
= P (X x) P (X x)
= H(x) H( x):
96
Similarly, if Tn has exact distribution Gn (x; F ); then jTn j has the distribution function
Gn (x; F ) = Gn (x; F ) Gn ( x; F ):
A two-sided hypothesis test rejects H0 for large values of jTn j : Since Tn !d Z; then jTn j !d
jZj : Thus asymptotic critical values are taken from the distribution, and exact critical
values are taken from the Gn (x; F0 ) distribution. From Theorem 8.9.1, we can calculate that
Gn (x; F ) = Gn (x; F ) Gn ( x; F )
1 1
= (x) + 1=2 g1 (x; F ) + g2 (x; F )
n n
1 1 3=2
( x) + 1=2 g1 ( x; F ) + g2 ( x; F ) + O(n )
n n
2
= (x) + g2 (x; F ) + O(n 3=2 ); (8.5)
n
where the simpli cations are because g1 is even and g2 is odd. Hence the di erence between the
asymptotic distribution and the exact distribution is
2 3=2 1
(x) Gn (x; F0 ) = g2 (x; F0 ) + O(n ) = O(n ):
n
The order of the error is O(n 1 ):
Interestingly, the asymptotic two-sided test has a better coverage rate than the asymptotic
one-sided test. This is because the rst term in the asymptotic expansion, g1 ; is an even function,
meaning that the errors in the two directions exactly cancel out.
Applying (8.5) to the bootstrap distribution, we nd
2 3=2
Gn (x) = Gn (x; Fn ) = (x) + g2 (x; Fn ) + O(n ):
n
Thus the di erence between the bootstrap and exact distributions is
2
Gn (x) Gn (x; F0 ) = (g2 (x; Fn ) g2 (x; F0 )) + O(n 3=2 )
n
= O(n 3=2 );
p
the last equality because Fn converges to F0 at rate n; and g2 is continuous in F: Another way
of writing this is
P (jTn j < x) = P (jTn j < x) + O(n 3=2 )
so the error from using the bootstrap distribution (relative to the true unknown distribution) is
O(n 3=2 ): This is in contrast to the use of the asymptotic distribution, whose error is O(n 1 ):
Thus a two-sided bootstrap test also achieves an asymptotic re nement, similar to a one-sided
test.
97
A reader might get confused between the two simultaneous e ects. Two-sided tests have better
rates of convergence than the one-sided tests, and bootstrap tests have better rates of convergence
than asymptotic tests.
The analysis shows that there may be a trade-o between one-sided and two-sided tests. Two-
sided tests will have more accurate size (Reported Type I error), but one-sided tests might have
more power against alternatives of interest. Con dence intervals based on the bootstrap can be
asymmetric if based on one-sided tests (equal-tailed intervals) and can therefore be more informa-
tive and have smaller length than symmetric intervals. Therefore, the choice between symmetric
and equal-tailed con dence intervals is unclear, and needs to be determined on a case-by-case
basis.
98
8.13 Bootstrap Methods for Regression Models
The bootstrap methods we have discussed have set Gn (x) = Gn (x; Fn ); where Fn is the EDF.
Any other consistent estimate of F0 may be used to de ne a feasible bootstrap estimator. The
advantage of the EDF is that it is fully nonparametric, it imposes no conditions, and works in
nearly any context. But since it is fully nonparametric, it may be ine cient in contexts where more
is known about F: We discuss some bootstrap methods appropriate for the case of a regression
model where
yi = x0i + ei
E (ei j xi ) = 0:
The non-parametric bootstrap distribution resamples the observations (yi ; xi ) from the EDF,
which implies
yi = xi 0 ^ + ei
E (xi ei ) = 0
but generally
E (ei j xi ) 6= 0:
The the bootstrap distribution does not impose the regression assumption, and is thus an ine cient
estimator of the true distribution (when in fact the regression assumption is true.)
One approach to this problem is to impose the very strong assumption that the error ei is
independent of the regressor xi : The advantage is that in this case it is straightforward to con-
struct bootstrap distributions. The disadvantage is that the bootstrap distribution may be a poor
approximation when the error is not independent of the regressors.
To impose independence, it is su cient to sample the xi and ei independently, and then create
yi = xi 0 ^ + ei : There are di erent ways to impose independence. A non-parametric method is
to sample the bootstrap errors ei randomly from the OLS residuals f^ e1 ; :::; e^n g: A parametric
method is to generate the bootstrap errors ei from a parametric distribution, such as the normal
ei N (0; ^ 2 ):
For the regressors xi , a nonparametric method is to sample the xi randomly from the EDF
or sample values fx1 ; :::; xn g: A parametric method is to sample xi from an estimated parametric
distribution. A third approach sets xi = xi : This is equivalent to treating the regressors as xed
in repeated samples. If this is done, then all inferential statements are made conditionally on the
observed values of the regressors, which is a valid statistical approach. It does not really matter,
however, whether or not the xi are really \ xed" or random.
The methods discussed above are unattractive for most applications in econometrics because
they impose the stringent assumption that xi and ei are independent. Typically what is desirable
is to impose only the regression condition E (ei j xi ) = 0: Unfortunately this is a harder problem.
99
One proposal which imposes the regression condition without independence is the Wild Boot-
strap. The idea is to construct a conditional distribution for ei so that
E (ei j xi ) = 0
E ei 2 j xi = e^2i
E ei 3 j xi = e^3i :
A conditional distribution with these features will preserve the main important features of the
data. This can be achieved using a two-point distribution of the form
p ! ! p
1+ 5 5 1
P ei = e^i = p
2 2 5
p ! ! p
1 5 5+1
P ei = e^i = p
2 2 5
E gi ^ = g n ( ^ ) 6= 0:
This means that wi do not satisfy the same moment conditions as the population distribution.
A correction suggested by Hall and Horowitz (1996) can solve the problem. Given the bootstrap
sample (Y ; X ; Z ); de ne the bootstrap GMM criterion
0
J ( )=n gn( ) gn( ^ ) Wn g n ( ) gn( ^ )
where g n ( ^ ) is from the in-sample data, not from the bootstrap data.
100
Let ^ minimize J ( ); and de ne all statistics and tests accordingly. In the linear model, this
implies that the bootstrap estimator is
^ = Z 0X W X 0Z 1
n Z 0 X Wn X 0 Y X 0 e^ :
101
Chapter 9
A time series yt is a process observed in sequence over time, t = 1; :::; T . To indicate the dependence
on time, we adopt new notation, and use the subscript t to denote the individual observation, and
T to denote the number of observations.
Because of the sequential nature of time series, we expect that Yt and Yt 1 are not independent,
so classical assumptions are not valid.
We can separate time series into two categories: univariate (yt 2 R is scalar); and multivariate
(yt 2 Rm is vector-valued). The primary model for univariate time series is autoregressions (ARs).
The primary model for multivariate time series is vector autoregressions (VARs).
E(Yt ) =
is independent of t; and
Cov (Yt ; Yt k) = (k)
is independent of t for all k:
De nition 9.1.2 fYt g is strictly stationary if the joint distribution of (Yt ; :::; Yt k) is independent
of t for all k:
102
The following two theorems are essential to the analysis of stationary time series. There proofs
are rather di cult, however.
Theorem 9.1.1 If Yt is strictly stationary and ergodic and Xt = f (Yt ; Yt 1 ; :::) is a random
variable, then Xt is strictly stationary and ergodic.
Theorem 9.1.2 (Ergodic Theorem). If Xt is strictly stationary and ergodic and E jXt j < 1; then
as T ! 1;
T
1X
Xt !p E(Xt ):
T
t=1
Theorem 9.1.3 If Yt is strictly stationary and ergodic and EYt2 < 1; then as T ! 1;
1. ^ !p E(Yt );
2. ^ (k) !p (k);
3. ^(k) !p (k):
Proof. Part (1) is a direct consequence of the Ergodic theorem. For Part (2), note that
T
1X
^ (k) = (Yt ^ ) (Yt k ^)
T
t=1
T T T
1X 1X 1X
= Yt Yt k Yt ^ Yt k^ + ^2:
T T T
t=1 t=1 t=1
103
By Theorem 9.1.1 above, the sequence Yt Yt k is strictly stationary and ergodic, and it has a nite
mean by the assumption that EYt2 < 1: Thus an application of the Ergodic Theorem yields
T
1X
Yt Yt k !p E(Yt Yt k ):
T
t=1
Thus
2 2 2 2
^ (k) !p E(Yt Yt k) + = E(Yt Yt k) = (k):
Part (3) follows by the continuous mapping theorem: ^(k) = ^ (k)=^ (0) !p (k)= (0) = (k):
9.2 Autoregressions
In time-series, the series f:::; Y1 ; Y2 ; :::; YT ; :::g are jointly random. We consider the conditional
expectation
E (Yt j It 1 )
where It 1 = fYt 1 ; Yt 2 ; :::g is the past history of the series.
An autoregressive (AR) model speci es that only a nite number of past lags matter:
E (Yt j It 1) = E (Yt j Yt 1 ; :::; Yt k ) :
A linear AR model (the most common type used in practice) speci es linearity:
E (Yt j It 1) = + 1 Yt 1 + 2 Yt 1 + + k Yt k :
Letting
et = Yt E (Yt j It 1) ;
then we have the autoregressive model
Yt = + 1 Yt 1 + 2 Yt 1 + + k Yt k + et
E (et j It 1) = 0:
The last property de nes a special time-series process.
De nition 9.2.1 et is a martingale di erence sequence (MDS) if E (et j It 1) = 0:
Regression errors are naturally a MDS. Some time-series processes may be a MDS as a conse-
quence of optimizing behavior. For example, some versions of the life-cycle hypothesis imply that
either changes in consumption, or consumption growth rates, should be a MDS. Most asset pricing
models imply that asset returns should be the sum of a constant plus a MDS.
The MDS property for the regression error plays the same role in a time-series regression as
does the conditional mean-zero property for the regression error in a cross-section regression. In
fact, it is even more important in the time-series context, as it is di cult to derive distribution
theories without this property.
A useful property of a MDS is that et is uncorrelated with any function of the lagged information
It 1 : Thus for k > 0; E(Yt k et ) = 0:
104
9.3 Stationarity of AR(1) Process
A mean-zero AR(1) is
Yt = Yt 1 + et :
Assume that et is iid, E(et ) = 0 and Ee2t = 2 < 1:
By back-substitution, we nd
2
Yt = et + et 1 + et 2 + :::
X1
k
= et k:
k=0
Loosely speaking, this series converges if the sequence ke gets small as k ! 1: This occurs
t k
when j j < 1:
1
X 2
2k
V ar(Yt ) = V ar (et k) = 2
:
1
k=0
If the equation for Yt has an intercept, the above results are unchanged, except that the mean
of Yt can be computed from the relationship
EYt = + EYt 1;
Yt Yt 1 + et
105
or
(1 L) Yt 1 = et :
The operator (L) = (1 L) is a polynomial in the operator L: We say that the root of the
polynomial is 1= ; since (z) = 0 when z = 1= : We call (L) the autoregressive polynomial of Yt .
From Theorem 9.3.1, an AR(1) is stationary i j j < 1: Note that an equivalent way to say
this is that an AR(1) is stationary i the root of the autoregressive polynomial is larger than one
(in absolute value).
where the 1 ; :::; k are the complex roots of (z); which satisfy ( j ) = 0:
We know that an AR(1) is stationary i the absolute value of the root of its autoregressive
polynomial is larger than one. For an AR(k), the requirement is that all roots are larger than one.
Let j j denote the modulus of a complex number :
Theorem 9.5.1 The AR(k) is strictly stationary and ergodic if and only if j j j > 1 for all j:
One way of stating this is that \All roots lie outside the unit circle."
If one of the roots equals 1, we say that (L); and hence Yt ; \has a unit root". This is a special
case of non-stationarity, and is of great interest in applied time series.
106
9.6 Estimation
Let
0
xt = 1 Yt 1 Yt 2 Yt k
0
= 1 2 k :
Theorem 9.6.1 If the AR(k) process Yt is strictly stationary and ergodic and EYt2 < 1; then
^ !p as T ! 1:
Proof. The vector xt is strictly stationary and ergodic, and by Theorem 9.1.1, so is xt x0t : Thus
by the Ergodic Theorem,
T
1X
xt x0t !p E xt x0t = Q:
T
t=1
Combined with (9.1) and the continuous mapping theorem, we see that
T
! 1 T
!
^= 1X 1X
+ xt x0t xt et !p Q 1
0 = 0:
T T
t=1 t=1
107
9.7 Asymptotic Distribution
Theorem 9.7.1 MDS CLT. If ut is a strictly stationary and ergodic MDS and E(ut u0t ) = < 1;
then as T ! 1;
T
1 X
p ut !d N (0; ) :
T t=1
where
= E(xt x0t e2t ):
Theorem 9.7.2 If the AR(k) process Yt is strictly stationary and ergodic and EYt4 < 1; then as
T ! 1;
p
T ^ !d N 0; Q 1 Q 1 :
This is identical in form to the asymptotic distribution of OLS in cross-section regression. The
implication is that asymptotic inference is the same. In particular, the asymptotic covariance
matrix is estimated just as in the cross-section case.
Yt = ^ + ^1 Yt 1 + ^2 Yt 2 + + ^k Yt k + et :
108
This construction imposes homoskedasticity on the errors ei ; which may be di erent than the
properties of the actual ei : It also presumes that the AR(k) structure is the truth.
4. This allows for arbitrary stationary serial correlation, heteroskedasticity, and for model-
misspeci cation.
5. The results may be sensitive to the block length, and the way that the data are partitioned
into blocks.
Yt = 0 + 1t + St (9.2)
St = 1 St 1 + 2 St 2 + + k St l + et ; (9.3)
or
Yt = 0 + 1t + 1 Yt 1 + 2 Yt 1 + + k Yt k + et : (9.4)
There are two essentially equivalent ways to estimate the autoregressive parameters ( 1 ; :::; k ):
You can estimate (9.2)-(9.3) sequentially by OLS. That is, rst estimate (9.2), get the residual
S^t ; and then perform regression (9.3) replacing St with S^t : This procedure is sometimes called
Detrending.
The reason why these two procedures are (essentially) the same is the Frisch-Waugh-Lovell
theorem.
Seasonal E ects
Include dummy variables for each season. This presumes that \seasonality" does not change
over the sample.
109
Use \seasonally adjusted" data. The seasonal factor is typically estimated by a two-sided
weighted average of the data for that season in neighboring years. Thus the seasonally
adjusted data is a \ ltered" series. This is a exible approach which can extract a wide
range of seasonal factors. The seasonal adjustment, however, also alters the time-series
correlations of the data.
First apply a seasonal di erencing operator. If s is the number of seasons (typically s = 4
or s = 12);
s Yt = Yt Yt s ;
or the season-to-season change. The series s Yt is clearly free of seasonality. But the long-run
trend is also eliminated, and perhaps this was of relevance.
or
Yt = (1 ) + ( + ) Yt 1 Yt 2 + et = AR(2):
Thus under H0 ; Yt is an AR(1), and under H1 it is an AR(2). H0 may be expressed as the
restriction that the coe cient on Yt 2 is zero.
An appropriate test of H0 against H1 is therefore a Wald test that the coe cient on Yt 2 is
zero. (A simple exclusion test).
In general, if the null hypothesis is that Yt is an AR(k), and the alternative is that the error
is an AR(m), this is the same as saying that under the alternative Yt is an AR(k+m), and this
is equivalent to the restriction that the coe cients on Yt k 1 ; :::; Yt k m are jointly zero. An
appropriate test is the Wald test of this restriction.
110
9.11 Model Selection
What is the appropriate choice of k in practice? This is a problem of model selection.
One approach to model selection is to choose k based on a Wald tests.
Another is to minimize the AIC or BIC information criterion, e.g.
2k
AIC(k) = log ^ 2 (k) + ;
T
where ^ 2 (k) is the estimated residual variance from an AR(k)
One ambiguity in de ning the AIC criterion is that the sample available for estimation changes
as k changes. (If you increase k; you need more initial conditions.) This can induce strange
behavior in the AIC. The best remedy is to x a upper value k; and then reserve the rst k as
initial conditions, and then estimate the models AR(1), AR(2), ..., AR(k) on this (uni ed) sample.
(L)Yt = + et
k
(L) = 1 1L kL :
1 + 2 + + k = 1:
In this case, Yt is non-stationary. The ergodic theorem and MDS CLT do not apply, and test
statistics are asymptotically non-normal.
A helpful way to write the equation is the so-called Dickey-Fuller reparameterization:
Yt = + 0 Yt 1 + 1 Yt 1 + + k 1 Yt (k 1) + et : (9.7)
These models are equivalent linear transformations of one another. The DF parameterization
is convenient because the parameter 0 summarizes the information about the unit root, since
(1) = 0 : To see this, observe that the lag polynomial for the Yt computed from (9.7) is
(1 L) 0L 1 (L L2 ) k 1 (L
k 1
Lk )
But this must equal (L); as the models are equivalent. Thus
(1) = (1 1) 0 (1 1) (1 1) = 0:
H0 : 0 = 0:
111
Note that the model is stationary if 0 < 0: So the natural alternative is
H1 : 0 < 0:
Under H0 ; the model for Yt is
Yt = + 1 Yt 1 + + k 1 Yt (k 1) + et ;
which is an AR(k-1) in the rst-di erence Yt : Thus if Yt has a (single) unit root, then Yt is a
stationary AR process. Because of this property, we say that if Yt is non-stationary but d Yt is
stationary, then Yt is \integrated of order d"; or I(d): Thus a time series with unit root is I(1):
Since 0 is the parameter of a linear regression, the natural test statistic is the t-statistic for
H0 from OLS estimation of (9.7). Indeed, this is the most popular unit root test, and is called the
Augmented Dickey-Fuller (ADF) test for a unit root.
It would seem natural to assess the signi cance of the ADF statistic using the normal table.
However, under H0 ; Yt is non-stationary, so conventional normal asymptotics are invalid. An
alternative asymptotic framework has been developed to deal with non-stationary data. We do
not have the time to develop this theory in detail, but simply assert the main results.
^0
ADF = ! DFt :
s(^ 0 )
The limit distributions DF and DFt are non-normal. They are skewed to the left, and have
negative means.
The rst result states that ^ 0 converges to its true value (of zero) at rate T; rather than the
conventional rate of T 1=2 : This is called a \super-consistent" rate of convergence.
The second result states that the t-statistic for ^ 0 converges to a limit distribution which is
non-normal, but does not depend on the parameters : This distribution has been extensively
tabulated, and may be used for testing the hypothesis H0 : Note: The standard error s(^ 0 ) is the
conventional (\homoskedastic") standard error. But the theorem does not require an assumption
of homoskedasticity. Thus the Dickey-Fuller test is robust to heteroskedasticity.
Since the alternative hypothesis is one-sided, the ADF test rejects H0 in favor of H1 when
ADF < c; where c is the critical value from the ADF table. If the test rejects H0 ; this means that
the evidence points to Yt being stationary. If the test does not reject H0 ; a common conclusion is
that the data suggests that Yt is non-stationary. This is not really a correct conclusion, however.
All we can say is that there is insu cient evidence to conclude whether the data are stationary or
not.
We have described the test for the setting of with an intercept. Another popular setting includes
as well a linear time trend. This model is
Yt = 1 + 2t + 0 Yt 1 + 1 Yt 1 + + k 1 Yt (k 1) + et : (9.8)
112
This is natural when the alternative hypothesis is that the series is stationary about a linear time
trend. If the series has a linear trend (e.g. GDP, Stock Prices), then the series itself is non-
stationary, but it may be stationary around the linear time trend. In this context, it is a silly
waste of time to t an AR model to the level of the series without a time trend, as the AR model
cannot conceivably describe this data. The natural solution is to include a time trend in the tted
OLS equation. When conducting the ADF test, this means that it is computed as the t-ratio for
0 from OLS estimation of (9.8).
If a time trend is included, the test procedure is the same, but di erent critical values are
required. The ADF test has a di erent distribution when the time trend has been included, and
a di erent table should be consulted.
Most texts include as well the critical values for the extreme polar case where the intercept has
been omitted from the model. These are included for completeness (from a pedagogical perspective)
but have no relevance for empirical practice where intercepts are always included.
113
Chapter 10
A multivariate time series Yt is a vector process m 1. Let It 1 = (Yt 1 ; Yt 2 ; :::) be all lagged
information at time t: The typical goal is to nd the conditional expectation E (Yt j It 1 ) : Note
that since Yt is a vector, this conditional expectation is also a vector.
A linear VAR speci es that this conditional mean is linear in the arguments:
E (Yt j Yt 1 ; :::; Yt k ) = A0 + A1 Yt 1 + A2 Yt 2 + Ak Yt k:
114
and the m (mk + 1) matrix
A= A0 A1 A2 Ak ;
then
Yt = Axt + et :
The VAR model is a system of m equations. One way to write this is to let a0j be the jth row
of A. Then the VAR system can be written as the equations
10.2 Estimation
Consider the moment conditions
E (xt ejt ) = 0;
j = 1; :::; m: These are implied by the VAR model, either as a regression, or as a linear projection.
The GMM estimator corresponding to these moment conditions is equation-by-equation OLS
^j = (X 0 X)
a 1
X 0 Yj :
^ we nd
And if we stack these to create the estimate A;
0 1
Y10
B Y0 C
B 2 C
A^ = B . C X(X 0 X) 1
@ .. A
0
Ym+1
= Y 0 X(X 0 X) 1
;
where
Y = Y1 Y2 Ym
the T m matrix of the stacked yt0 :
This (system) estimator is known as the SUR (Seemingly Unrelated Regressions) estimator,
and was originally derived by Zellner (1962)
115
10.3 Restricted VARs
The unrestricted VAR is a system of m equations, each with the same set of regressors. A restricted
VAR imposes restrictions on the system. For example, some regressors may be excluded from some
of the equations. Restrictions may be imposed on individual equations, or across equations. The
GMM framework gives a convenient method to impose such restrictions on estimation.
Yjt = a0j xt + et ;
and xt consists of lagged values of Yjt and the other Ylt0 s: In this case, it is convenient to re-de ne
the variables. Let yt = Yjt ; and Zt be the other variables. Let et = ejt and = aj : Then the single
equation takes the form
yt = x0t + et ; (10.1)
and h i
0
xt = 1 Yt 1 Yt k Zt0 1 Zt0 k :
This is just a conventional regression, with time series data.
yt = x0t + et
et = et 1 + ut (10.2)
E (ut j It 1) = 0:
H0 : =0 H1 : 6= 0:
Take the equation yt = x0t + et ; and subtract o the equation once lagged multiplied by ; to get
yt yt 1 = x0t + et x0t 1 + et 1
= x0t xt 1 + et et 1;
or
yt = y t 1 + x0t + x0t 1 + ut ; (10.3)
116
which is a valid regression model.
So testing H0 versus H1 is equivalent to testing for the signi cance of adding (yt 1 ; xt 1 ) to
the regression. This can be done by a Wald test. We see that an appropriate, general, and simple
way to test for omitted serial correlation is to test the signi cance of extra lagged values of the
dependent variable and regressors.
You may have heard of the Durbin-Watson test for omitted serial correlation, which once was
very popular, and is still routinely reported by conventional regression packages. The DW test is
appropriate only when regression yt = x0t + et is not dynamic (has no lagged values on the RHS),
and et is iid N (0; 1): Otherwise it is invalid.
Another interesting fact is that (10.2) is a special case of (10.3), under the restriction = :
This restriction, which is called a common factor restriction, may be tested if desired. If valid,
the model (10.2) may be estimated by iterated GLS. (A simple version of this estimator is called
Cochrane-Orcutt.) Since the common factor restriction appears arbitrary, and is typically rejected
empirically, direct estimation of (10.2) is uncommon in recent applications.
where p is the number of parameters in the model, and e^t (k) is the OLS residual vector from the
model with k lags. The log determinant is the criterion from the multivariate normal likelihood.
The information set I1t is generated only by the history of Yt ; and the information set I2t is
generated by both Yt and Zt : The latter has more information.
117
We say that Zt does not Granger-cause Yt if
That is, conditional on information in lagged Yt ; lagged Zt does not help to forecast Yt : If this
condition does not hold, then we say that Zt Granger-causes Yt :
The reason why we call this \Granger Causality" rather than \causality" is because this is not
a physical or structure de nition of causality. If Zt is some sort of forecast of the future, such as a
futures price, then Zt may help to forecast Yt even though it does not \cause" Yt : This de nition
of causality was developed by Granger (1969) and Sims (1972).
In a linear VAR, the equation for Yt is
Yt = + 1 Yt 1 + + k Yt k + Zt0 1 1 + + Zt0 k k + et :
H0 : 1 = 2 = = k = 0:
10.8 Cointegration
The idea of cointegration is due to Granger (1981), and was articulated in detail by Engle and
Granger (1987).
If the series Yt is not cointegrated, then r = 0: If r = m; then Yt is I(0): For 0 < r < m; Yt is
I(1) and cointegrated.
In some cases, it may be believed that is known a priori. Often, = (1 1)0 : For example,
if Yt is a pair of interest rates, then = (1 0
1) speci es that the spread (the di erence in
returns) is stationary. If Y = (log(Consumption) log(Income))0 ; then = (1 1)0 speci es
that log(Consumption=Income) is stationary.
In other cases, may not be known.
If Yt is cointegrated with a single cointegrating vector (r = 1); then it turns out that can
be consistently estimated by an OLS regression of one component of Yt on the others. Thus Yt =
118
(Y1t ; Y2t ) and = ( 1 2 ) and normalize 1 = 1: Then ^ 2 = (Y20 Y2 ) 1 Y2 Y1 !p 2 : Furthermore
this estimation is super-consistent: T ( ^ 2 2 ) !d Limit; as rst shown by Stock (1987). This
is not, in general, a good method to estimate ; but it is useful in the construction of alternative
estimators and tests.
We are often interested in testing the hypothesis of no cointegration:
H0 : r = 0
H1 : r > 0:
A(L)Yt = et
A(L) = I A1 L A2 L2 Ak Lk
or alternatively as
Yt = Yt 1 + D(L) Yt 1 + et
where
= A(1)
= I + A1 + A2 + + Ak :
Thus cointegration imposes a restriction upon the parameters of a VAR. The restricted model
can be written as
0
Yt = Yt 1 + D(L) Yt 1 + et
Yt = zt 1 + D(L) Yt 1 + et :
119
If is known, this can be estimated by OLS of Yt on zt 1 and the lags of Yt :
If is unknown, then estimation is done by \reduced rank regression", which is least-squares
subject to the stated restriction. Equivalently, this is the MLE of the restricted parameters under
the assumption that et is iid N (0; ):
One di culty is that is not identi ed without normalization. When r = 1; we typically just
normalize one element to equal unity. When r > 1; this does not work, and di erent authors have
adopted di erent identi cation schemes.
In the context of a cointegrated VAR estimated by reduced rank regression, it is simple to
test for cointegration by testing the rank of : These tests are constructed as likelihood ratio
(LR) tests. As they were discovered by Johansen (1988, 1991, 1995), they are typically called
the \Johansen Max and Trace" tests. Their asymptotic distributions are non-standard, and are
similar to the Dickey-Fuller distributions.
120
Chapter 11
A \limited dependent variable" Y is one which takes a \limited" set of values. The most common
cases are
Binary: Y = f0; 1g
Censored: Y = fx : x 0g
The traditional approach to the estimation of limited dependent variable (LDV) models is
parametric maximum likelihood. A parametric model is constructed, allowing the construction of
the likelihood function. A more modern approach is semi-parametric, eliminating the dependence
on a parametric distributional assumption. We will discuss only the rst (parametric) approach,
due to time constraints. They still constitute the majority of LDV applications. If, however, you
were to write a thesis involving LDV estimation, you would be advised to consider employing a
semi-parametric estimation approach.
For the parametric approach, estimation is by MLE. A major practical issue is construction of
the likelihood function.
P (Yi = 1 j xi ) = x0i :
121
As P (Yi = 1 j xi ) = E (Yi j xi ) ; this yields the regression: Yi = x0i + ei which can be estimated by
OLS. However, the linear probability model does not impose the restriction that 0 P (Yi j xi ) 1:
Even so estimation of a linear probability model is a useful starting point for subsequent analysis.
The standard alternative is to use a function of the form
P (Yi = 1 j xi ) = F x0i
where F ( ) is a known CDF, typically assumed to be symmetric about zero, so that F (z) =
1 F ( z): The two standard choices for F are
Logistic: F (u) = (1 + e u) 1 :
If F is logistic, we call this the logit model, and if F is normal, we call this the probit model.
This model is identical to the latent variable model
Yi = x0i + ei
ei F()
1 if Yi > 0
Yi = :
0 otherwise
For then
f (y) = py (1 p)1 y
; y = 0; 1:
In the Binary choice model, Yi is conditionally Bernoulli with P (Yi = 1 j xi ) = pi = F (x0i ) : Thus
the conditional density is
f (yi j xi ) = pyi i (1 pi )1 yi
yi
= F x0i (1 F x0i )1 yi
:
122
Hence the log-likelihood function is
n
X
ln ( ) = log f (yi j xi )
i=1
Xn
yi
= log F x0i (1 F x0i )1 yi
i=1
n
X
= yi log F x0i + (1 yi ) log(1 F x0i )
i=1
X X
= log F x0i + log(1 F x0i ):
yi =1 yi =0
The MLE ^ is the value of which maximizes ln ( ): Standard errors and test statistics are
computed by asymptotic approximations. Details of such calculations are left to more advanced
courses.
The conditional density is the Poisson with parameter i: The functional form for i has been
picked to ensure that i > 0.
The log-likelihood function is
n
X n
X
ln ( ) = log f (yi j xi ) = exp(x0i ) + yi x0i log(yi !) :
i=1 i=1
V ar (Yi j xi ) = i = exp(x0i );
so the model imposes the restriction that the conditional mean and variance of Yi are the same.
This may be considered restrictive. A generalization is the negative binomial.
123
11.3 Censored Data
The idea of \censoring" is that some data above or below a threshold are mis-reported at the
threshold. Thus the model is that there is some latent process yi with unbounded support, but
we observe only
yi if yi 0
yi = : (11.1)
0 if yi < 0
(This is written for the case of the threshold being zero, any known value can substitute.) The
observed data yi therefore come from a mixed continuous/discrete distribution.
Censored models are typically applied when the data set has a meaningful proportion (say 5%
or higher) of data at the boundary of the sample support. The censoring process may be explicit
in data collection, or it may be a by-product of economic constraints.
An example of a data collection censoring is top-coding of income. In surveys, incomes above
a threshold are typically reported at the threshold.
The rst censored regression model was developed by Tobin (1958) to explain consumption of
durable goods. Tobin observed that for many households, the consumption level (purchases) in a
particular period was zero. He proposed the latent variable model
yi = x0i + ei
2
ei iid N (0; )
with the observed variable yi generated by the censoring equation (11.1). This model (now called
the Tobit) speci es that the latent (or ideal) value of consumption may be negative (the household
would prefer to sell than buy). All that is reported is that the household purchased zero units of
the good.
The naive approach to estimate is to regress yi on xi . This does not work because regression
estimates E (Yi j xi ) ; not E (Yi j xi ) = x0i ; and the latter is of interest. Thus OLS will be biased
for the parameter of interest :
[Note: it is still possible to estimate E (Yi j xi ) by LS techniques. The Tobit framework pos-
tulates that this is not inherently interesting, that the parameter of is de ned by an alternative
statistical structure.]
Consistent estimation will be achieved by the MLE. To construct the likelihood, observe that
the probability of being censored is
x0i
= :
124
The conditional distribution function above zero is Gaussian:
Z y
1 z x0i
P (yi = y j xi ) = dz; y > 0:
0
yi = x0i + e1i
Ti = 1 zi0 + e0i > 0
where 1 ( ) is the indicator function. The dependent variable yi is observed if (and only if) Ti = 1:
Else it is unobserved.
For example, yi could be a wage, which can be observed only if a person is employed. The
equation for Ti is an equation specifying the probability that the person is employed.
The model is often completed by specifying that the errors are jointly normal
e0i 1
N 0; 2 :
e1i
125
It is presumed that we observe fxi ; zi ; Ti g for all observations.
Under the normality assumption,
e1i = e0i + vi ;
where vi is independent of e0i N (0; 1): A useful fact about the standard normal distribution is
that
(x)
E (e0i j e0i > x) = (x) = ;
(x)
and the function (x) is called the inverse Mills ratio.
The naive estimator of is OLS regression of yi on xi for those observations for which yi is
available. The problem is that this is equivalent to conditioning on the event fTi = 1g: However,
The OLS standard errors will be incorrect, as this is a two-step estimator. They can be
corrected using a more complicated formula. Or, alternatively, by viewing the Probit/OLS
estimation equations as a large joint GMM problem.
The Heckit estimator is frequently used to deal with problems of sample selection. However,
the estimator is built on the assumption of normality, and the estimator can be quite sensitive
to this assumption. Some modern econometric research is exploring how to relax the normality
assumption.
126
The estimator can also work quite poorly if (zi0 ^ ) does not have much in-sample variation.
This can happen if the Probit equation does not \explain" much about the selection choice. An-
other potential problem is that if zi = xi ; then (zi0 ^ ) can be highly collinear with xi ; so the second
step OLS estimator will not be able to precisely estimate : Based this observation, it is typically
recommended to nd a valid exclusion restriction: a variable should be in zi which is not in xi :
If this is valid, it will ensure that (zi0 ^ ) is not collinear with xi ; and hence improve the second
stage estimator's precision.
127
Chapter 12
Panel Data
A panel is a set of observations on individuals, collected over time. An observation is the pair
fyit ; xit g; where the i subscript denotes the individual, and the t subscript denotes time. A panel
may be balanced :
fyit ; xit g : t = 1; :::; T ; i = 1; :::; n;
or unbalanced :
fyit ; xit g : For i = 1; :::; n; t = ti ; :::; ti :
E (xit ui ) = 0 (12.1)
If this condition fails, then OLS is inconsistent. (12.1) fails if the individual-speci c unobserved
e ect ui is correlated with the observed explanatory variables xit : This is often believed to be
plausible if ui is an omitted variable.
If (12.1) is true, however, OLS can be improved upon via a GLS technique. In either event,
OLS appears a poor estimation choice.
Condition (12.1) is called the random e ects hypothesis. It is a strong assumption, and most
applied researchers try to avoid its use.
128
12.2 Fixed E ects
This is the most common technique for estimation of non-dynamic linear panel regressions.
The motivation is to allow ui to be arbitrary, and have arbitrary correlated with xi : The goal
is to eliminate ui from the estimator, and thus achieve invariance.
There are several derivations of the estimator.
First, let 8
< 1 if i = j
dij = ;
:
0 else
and 0 1
di1
B C
di = @ ... A ;
din
an n 1 dummy vector with a \1" in the i0 th place. Let
0 1
u1
B C
u = @ ... A :
un
If xit contains an intercept, it will be collinear with di ; so the intercept is typically omitted
from xit :
Any regressor in xit which is constant over time for all individuals (e.g., their gender) will
be collinear with di ; so will have to be omitted.
There are n + k regression parameters, which is quite large as typically n is very large.
129
Computationally, you do not want to actually implement conventional OLS estimation, as the
parameter space is too large. OLS estimation of proceeds by the FWL theorem. Stacking the
observations together:
Y = X + Du + e;
then by the FWL theorem,
^ = 1
X 0 (1 PD ) X X 0 (1 PD ) Y
1
= X 0X X 0Y ;
where
Y = Y D(D0 D) 1
D0 Y
X = X D(D0 D) 1
D0 X:
Since the regression of yit on di is a regression onto individual-speci c dummies, the predicted
value from these regressions is the individual speci c mean y i ; and the residual is the demean
value
yit = yit y i :
The xed e ects estimator ^ is OLS of yit on xit , the dependent variable and regressors in deviation-
from-mean form.
Another derivation of the estimator is to take the equation
and then take individual-speci c means by taking the average for the i0 th individual:
ti ti ti
1 X 1 X 0 1 X
yit = x + ui + eit
Ti t=t Ti t=t it Ti t=t
i i i
or
y i = x0i + ui + ei :
Subtracting, we nd
yit = xit0 + eit ;
which is free of the individual-e ect ui :
130
This is a model suitable for studying dynamic behavior of individual agents.
Unfortunately, the xed e ects estimator is inconsistent, at least if T is held nite as n ! 1:
This is because the sample mean of yit 1 is correlated with that of eit :
The standard approach to estimate a dynamic panel is to combine rst-di erencing with IV or
GMM. Taking rst-di erences of (12.3) eliminates the individual-speci c e ect:
131
Chapter 13
Nonparametrics
132
be the kernel K rescaled by the bandwidth h: The kernel density estimator of f (x) is
n
1X
f^(x) = Kh (Xi x) :
n
i=1
This estimator is the average of a set of weights. If a large number of the observations Xi are near
x; then the weights are relatively large and f^(x) is larger. Conversely, if only a few Xi are near x;
then the weights are small and f^(x) is small. The bandwidth h controls the meaning of \near".
Interestingly, f^(x) is a valid density. That is, f^(x) 0 for all x; and
Z Z n n Z n Z
1 1
1X 1X 1
1X 1
f^(x)dx = Kh (Xi x) dx = Kh (Xi x) dx = K (u) du = 1
1 1 n n 1 n 1
i=1 i=1 i=1
the sample mean of the Xi ; where the second-to-last equality used the change-of-variables u =
(Xi x)=h which has Jacobian h:
The second moment of the estimated density is
Z n Z
1
1X 1
x f^(x)dx =
2
x2 Kh (Xi x) dx
1 n 1
i=1
Xn Z 1
1
= (Xi + uh)2 K (u) du
n 1
i=1
n n Z n Z
1X 2 2X 1
1X 2 1
= Xi + Xi h K(u)du + h u2 K (u) du
n n 1 n 1
i=1 i=1 i=1
n
1X 2
= X i + h2 2
K
n
i=1
133
where Z 1
2
K = x2 K (x) dx
1
is the variance of the kernel. It follows that the variance of the density f^(x) is
Z Z n n
!2
1 1 2
1X 2 1X
x f^(x)dx
2
xf^(x)dx = X i + h2 2
K Xi
1 1 n n
i=1 i=1
2
= ^ + h2 2K
The second equality uses the change-of variables u = (z x)=h: The last expression shows that
the expected value is an average of f (z) locally about x:
This integral (typically) is not analytically solvable, so we approximate it using a second order
Taylor expansion of f (x + hu) in the argument hu about hu = 0; which is valid as h ! 0: Thus
1
f (x + hu) ' f (x) + f 0 (x)hu + f 00 (x)h2 u2
2
and therefore
Z 1
1
EKh (X x) ' K (u) f (x) + f 0 (x)hu + f 00 (x)h2 u2 du
1 2
Z 1 Z 1 Z 1
1
= f (x) K (u) du + f 0 (x)h K (u) udu + f 00 (x)h2 K (u) u2 du
1 1 2 1
1
= f (x) + f 00 (x)h2 2K :
2
The bias of f^(x) is then
n
1X 1
Bias(x) = E f^(x) f (x) = EKh (Xi x) f (x) = f 00 (x)h2 2
K:
n 2
i=1
We see that the bias of f^(x) at x depends on the second derivative f 00 (x): The sharper the derivative,
the greater the bias. Intuitively, the estimator f^(x) smooths data local to Xi = x; so is estimating
134
a smoothed version of f (x): The bias results from this smoothing, and is larger the greater the
curvature in f (x):
We now examine the variance of f^(x): Since it is an average of iid random variables, using
rst-order Taylor approximations and the fact that n 1 is of smaller order than (nh) 1
1
V ar (x) = V ar (Kh (Xi x))
n
1 1
= EKh (Xi x)2 (EKh (Xi x))2
n n
Z 1
1 z x 2 1
' 2
K f (z)dz f (x)2
nh h n
Z 11
1
= K (u)2 f (x + hu) du
nh 1
Z
f (x) 1
' K (u)2 du
nh 1
f (x) R(K)
= :
nh
R1
where R(K) = 1 K (x)2 dx is called the roughness of K:
Together, the asymptotic mean-squared error (AMSE) for xed x is the sum of the approximate
squared bias and approximate variance
1 f (x) R(K)
AM SEh (x) = f 00 (x)2 h4 4
K + :
4 nh
A global measure of precision is the asymptotic mean integrated squared error (AMISE)
Z
h4 4K R(f 00 ) R(K)
AM ISEh = AM SEh (x)dx = + : (13.1)
4 nh
R
where R(f 00 ) = (f 00 (x))2 dx is the roughness of f 00 : Notice that the rst term (the squared bias)
is increasing in h and the second term (the variance) is decreasing in nh: Thus for the AMISE to
decline with n; we need h ! 0 but nh ! 1: That is, h must tend to zero, but at a slower rate
than n 1 :
Equation (13.1) is an asymptotic approximation to the MSE. We de ne the asymptotically
optimal bandwidth h0 as the value which minimizes this approximate MSE. That is,
h0 = argmin AM ISEh
h
d R(K)
AM ISEh = h3 4 00
K R(f ) =0
dh nh2
135
yielding
1=5
R(K) 1=2
h0 = 4 R(f 00 ) n : (13.2)
K
This solution takes the form h0 = cn 1=5 where c is a function of K and f; but not of n: We
thus say that the optimal bandwidth is of order O(n 1=5 ): Note that this h declines to zero, but
at a very slow rate.
In practice, how should the bandwidth be selected? This is a di cult problem, and there is a
large and continuing literature on the subject. The asymptotically optimal choice given in (13.2)
depends on R(K); 2K ; and R(f 00 ): The rst two are determined by the kernel function. Their
values for the three functions introduced in the previous section are given here.
R1 R1
K 2
K = 2
1x K (x) dx R(K) = 1 K (x)2 dx
p
Gaussian 1 1/(2 )
Epanechnikov 1=5 1=5
Biweight 1=7 5=7
An obvious di culty is that R(f 00 ) is unknown. A classic simple solution proposed by Silverman
(1986)has come to be known as the reference bandwidth or Silverman's Rule-of-Thumb.
It uses formula (13.2) but replaces R(f 00 ) with ^ 5 R( 00 ); where is the N (0; 1) distribution
and ^ 2 is an estimate of 2 = V ar(X): This choice for h gives an optimal rule when f (x) is
normal, and gives a nearly optimal rule when f (x) is close to normal. The downside is that if the
density is very far from normal, the rule-of-thumb h can be quite ine cient. We can calculate that
p
R( 00 ) = 3= (8 ) : Together with the above table, we nd the reference rules for the three kernel
functions introduced earlier.
Gaussian Kernel: hrule = 1:06n 1=5
Epanechnikov Kernel: hrule = 2:34n 1=5
Biweight (Quartic) Kernel: hrule = 2:78n 1=5
Unless you delve more deeply into kernel estimation methods the rule-of-thumb bandwidth is
a good practical bandwidth choice, perhaps adjusted by visual inspection of the resulting estimate
f^(x): There are other approaches, but implementation can be delicate. I now discuss some of these
choices. The plug-in approach is to estimate R(f 00 ) in a rst step, and then plug this estimate into
the formula (13.2). This is more treacherous than may rst appear, as the optimal h for estimation
of the roughness R(f 00 ) is quite di erent than the optimal h for estimation of f (x): However, there
are modern versions of this estimator work well, in particular the iterative method of Sheather
and Jones (1991). Another popular choice for selection of h is cross-validation. This works by
constructing an estimate of the MISE using leave-one-out estimators. There are some desirable
properties of cross-validation bandwidths, but they are also known to converge very slowly to
the optimal values. They are also quite ill-behaved when the data has some discretization (as is
common in economics), in which case the cross-validation rule can sometimes select very small
136
bandwidths leading to dramatically undersmoothed estimates. Fortunately there are remedies,
which are known as smoothed cross-validation which is a close cousin of the bootstrap.
137
Chapter 14
n n!
=
r r! (n r)!
A number x is the limit of the sequence fxn g if for every " > 0; there is some N < 1 such
that jxn xj < " for all n N: The sequence is said to converge if it has a limit.
The limit superior (limsup) and limit inferior (liminf ) of the sequence fxn g are
138
A function f : R ! R is continuous at x if limx!a f (x) = f (a): A function f (x) is uniformly
continuous if for every " > 0 there exists a > 0 such that jf (x) f (y)j < " for every x and y
in its domain with jx yj < :
Taylor Series in one variable
N
X xn xN +1 (N +1)
f (a + x) = f (n) (a) + f (a + x); 0 1
n! N!
n=0
@ 1 @2
f (a + x) = f (a) + x0 f (a) + x0 f (a + x)x; 0 1
@a 2 @a@a0
139
Chapter 15
15.1 Terminology
A scalar a is a single number.
A vector a is a k 1 list of numbers, typically arranged in a column. We write this as
0 1
a1
B a2 C
B C
a=B . C
@ . A.
ak
140
where 2 3
a1i
6 a2i 7
6 7
ai = 6 .. 7
4 . 5
aki
are column vectors and
0
j = aj1 aj2 ajr
are row vectors.
The transpose of a matrix, denoted B = A0 ; is obtained by ipping the matrix on its diagonal.
2 3
a11 a21 ak1
6 a12 a22 a 7
6 k2 7
B = A0 = 6 . . .. 7
4 .. .. . 5
a1r a2r akr
Thus bij = aji for all i and j: Note that if A is k r, then A0 is r k: If a is a k 1 vector, then
a0 is a 1 k row vector.
A matrix is square if k = r: A square matrix is symmetric if A = A0 ; which implies aij = aji :
A square matrix is diagonal if the only non-zero elements appear on the diagonal, so that aij = 0
if i 6= j: A square matrix is upper (lower) diagonal if all elements below (above) the diagonal
equal zero.
A partitioned matrix takes the form
2 3
A11 A12 A1r
6 A21 A22 A2r 7
6 7
A=6 . .. .. 7
4 .. . . 5
Ak1 Ak2 Akr
Note that a0 b = b0 a:
141
If A is k r and B is r s; then we de ne their product AB by writing A as a set of row
vectors and B as a set of column vectors (each of length r). Then
2 0 3
a1
6 a0 7
6 2 7
AB = 6 . 7 b1 b2 br
4 .. 5
a0k
2 3
a01 b1 a01 b2 a01 br
6 a02 b1 a02 b2 a02 br 7
6 7
= 6 .. .. .. 7
4 . . . 5
0 0
ak b1 ak b2 a0k br
When the number of columns of A equals the number of rows of B; we say that A and B; or the
product AB; are conformable, and this is the only case where this product is de ned.
An alternative way to write the matrix product is to use matrix partitions. For example,
and
2 3
B1
6 B2 7
6 7
AB = A1 A2 Ar 6 .. 7
4 . 5
Br
= A1 B1 + A2 B2 + + Ar Br
Xr
= Ar Br
j=1
142
An important diagonal matrix is the identity matrix, which has ones on the diagonal. A
k k identity matrix is denoted as
2 3
1 0 0
6 0 1 0 7
6 7
Ik = 6 . . .. 7
4 .. .. . 5
0 0 1
tr (cA) = c tr (A)
tr A0 = tr (A)
tr (A + B) = tr (A) + tr (B)
tr (Ik ) = k
tr (AB) = tr (BA)
143
A k k matrix A has full rank, or is nonsingular, if there is no c 6= 0 such that Ac = 0:
In this case there exists a unique matrix B such that AB = BA = Ik : This matrix is called the
inverse of A and is denoted by A 1 : Some properties include
1 1
AA = A A = Ik
1 0 0 1
A = A
1 1 1
(AC) = C A
1 1 1 1 1 1
(A + C) = A A +C C
1 1 1 1 1 1
A (A + C) = A A +C A
Also, if A is an orthogonal matrix, then A 1 = A:
The following fact about inverting partitioned matrices is sometimes useful
1 1 1 BD 1
A B E E
= 1 CE 1 1 (15.1)
C D D F
where
1 1 1 1 1 1 1
E = A +A BF CA = A BD C
1 1 1 1 1 1 1
F = D +D CE BD = D CA B
Even if a matrix A does not possess an inverse, we can still de ne a generalized inverse A
as a matrix which satis es
AA A = A: (15.2)
The matrix A is not necessarily unique. The Moore-Penrose generalized inverse A satis es
(15.2) plus the following three conditions
A AA = A
AA is symmetric
A A is symmetric
For any matrix A; the Moore-Penrose generalized inverse A exists and is unique.
The determinant is de ned for square matrices.
If A is 2 2; then its determinant is det A = a11 a22 a12 a21 :
For a general k k matrix A = [aij ] ; we can de ne the determinant as follows. Let =
(j1 ; :::; jk ) denote a permutation of (1; :::; k) : There are k! such permutations. There is a unique
count of the number of inversions of the indices of such permutations (relative to the natural order
(1; :::; k)); and let " = +1 if this count is even and " = 1 if the count is odd. Then
X
det A = " a1j1 a2j2 akjk
144
det A = det A0
det ( A) = k det A
A B 1C
det = (det D) det A BD if det D 6= 0
C D
det A 6= 0 if and only if A is nonsingular.
Qk
If A is triangular (upper or lower), then det A = i=1 aii
15.4 Eigenvalues
The characteristic equation of a square matrix A is
det (A Ik ) = 0:
The left side is a polynomial of degree k in so has exactly k roots, which are not necessarily
distinct and may be real or complex. They are called the latent roots or characteristic roots
or eigenvalues of A. If i is an eigenvalue of A; then A i Ik is singular so there exists a non-zero
vector hi such that
(A i Ik ) hi = 0
If A has distinct characteristic roots, there exists a nonsingular matrix P such that A =
P 1 P and P AP 1 = .
If A is symmetric, then A = H H 0 and H 0 AH = ; and the characteristic roots are all real.
145
1 1 1 1
The characteristic roots of A are 1 ; 2 ; ..., k :
A = AA = H H 0 H H 0 = H 2
H 0:
By the uniqueness of the characteristic roots, we deduce that 2 = and 2i = i for i = 1; :::; k:
Hence they must equal either 0 or 1.
It follows that if A is symmetric and idempotent, then tr(A) = rank(A):
Let X be an n k matrix; k < n: Two projection matrices are
1
P = X X 0X X0
M = In P
1
= In X X 0X X 0:
They are called projection matrices due to the property that for any matrix Z which can be written
as Z = X for some matrix ; (we say that Z lies in the range space of X) then
1
P Z = P X = X X 0X X 0X = X = Z
and
M Z = (In P)Z = Z PZ = Z Z = 0:
146
As an important example of this property, partition the matrix X into two matrices X1 and
X2 ; so that
X = [X1 X2 ] :
Then P X1 = X1 and M X1 = 0:
P and M are symmetric:
1 0
P0 = X X 0X X0
0 1 0
= X0 X 0X (X)0
0 1
= X X 0X X0
0 1
= X (X)0 X 0 X0
= P
and
M 0 = (In P )0 = In0 P 0 = In P = M:
The projection matrices P and M are idempotent:
1 1
PP = X X 0X X0 X X 0X X0
1 1
= X X 0X X 0X X 0X X0
1
= X X 0X X 0 = P;
and
MM = (In P ) (In P)
= In In P In In P + P P
= In P P +P
= In P = M:
Furthermore,
M +P = In P + P = In
MP = (In P )P = P PP = P P = 0:
tr P = k (15.3)
tr M = n k: (15.4)
147
Indeed,
1
tr P = tr X X 0 X X0
1
= tr X 0X X 0X
= tr (Ik )
= k;
and
tr M = tr (In P ) = tr (In ) tr (P ) = n k:
From this, we deduce that the ranks of P and M are k and n = k, respectively. Since M is
symmetric and idempotent, its spectral decomposition takes the form
In k 0
M =H H0 (15.5)
0 0
with H 0 H = In .
Let B be any matrix. The Kronecker product of A and B; denoted A B; is the matrix
2 3
a11 B a12 B a1n B
6 a21 B a22 B a2n B 7
6 7
A B=6 .. .. .. 7:
4 . . . 5
am1 B am2 B amn B
Some important properties are now summarized. These results hold for matrices for which all
matrix multiplications are conformable.
(A + B) C=A C +B C
(A B) (C D) = AC BD
A (B C) = (A B) C
148
(A B)0 = A0 B0
tr (A B) = tr (A) tr (B)
If A is m m and B is n n; det(A B) = (det (A))n (det (B))m
1 1 1
(A B) =A B
If A > 0 and B > 0 then A B>0
vec (ABC) = (C 0 A) vec (B)
tr (ABCD) = vec (D0 )0 (C 0 A) vec (B)
and
@ @ @
g(x) = @x1 g(x) @xk g(x) :
@x0
Some properties are now summarized.
@ @
@x (a0 x) = @x (x0 a) = a
@
@x0 (Ax) = A
@
@x (x0 Ax) = (A + A0 ) x
@2
@x@x0 (x0 Ax) = (A + A0 )
A = [aij ] be m n and g(A) : Rmn ! R: We de ne
@ @
g(A) = g(A)
@A @aij
Some properties are now summarized.
@
@A (x0 Ax) = xx0
@
ln (A) = A 1 0
@A
@
@A tr (AB) = B 0
@ 1B 1 BA 1
@A tr A = A
149
Chapter 16
Appendix C: Probability
16.1 Foundations
The set S of all possible outcomes of an experiment is called the sample space for the experiment.
Take the simple example of tossing a coin. There are two outcomes, heads and tails, so we
can write S = fH; T g: If two coins are tossed in sequence, we can write the four outcomes as
S = fHH; HT; T H; T T g:
An event A is any collection of possible outcomes of an experiment. An event is a subset of S;
including S itself and the null set ;: Continuing the two coin example, one event is A = fHH; HT g;
the event that the rst coin is heads. We say that A and B are disjoint or mutually exclusive
if A \ B = ;: For example, the sets fHH; HT g and fT Hg are disjoint. Furthermore, if the sets
A1 ; A2 ; ::: are pairwise disjoint and [1
i=1 Ai = S; then the collection A1 ; A2 ; ::: is called a partition
of S:
The following are elementary set operations:
Union: A [ B = fx : x 2 A or x 2 Bg:
Intersection: A \ B = fx : x 2 A and x 2 Bg:
Complement: Ac = fx : x 2 = Ag:
The following are useful properties of set operations.
Communtatitivity: A [ B = B [ A; A \ B = B \ A:
Associativity: A [ (B [ C) = (A [ B) [ C; A \ (B \ C) = (A \ B) \ C:
Distributive Laws: A\(B [ C) = (A \ B)[(A \ C) ; A[(B \ C) = (A [ B)\(A [ C) :
c c c c
DeMorgan's Laws: (A [ B) = A \ B ; (A \ B) = Ac [ B c :
A probability function assigns probabilities (numbers between 0 and 1) to events A in S:
This is straightforward when S is countable; when S is uncountable we must be somewhat more
careful:A set B is called a sigma algebra (or Borel eld) if ; 2 B , A 2 B implies Ac 2 B, and
A1 ; A2 ; ::: 2 B implies [1i=1 Ai 2 B. A simple example is f;; Sg which is known as the trivial sigma
algebra. For any sample space S; let B be the smallest sigma algebra which contains all of the
open sets in S: When S is countable, B is simply the collection of all subsets of S; including ; and
150
S: When S is the real line, then B is the collection of all open and closed intervals. We call B the
sigma algebra associated with S: We only de ne probabilities for events contained in B.
We now can give the axiomatic de nition of probability. Given S and B, a probability function
P satis es P (S)
P1= 1; P (A) 0 for all A 2 B, and if A1 ; A2 ; ::: 2 B are pairwise disjoint, then
1
P ([i=1 Ai ) = i=1 P (Ai ):
Some important properties of the probability function include the following
P (;) = 0
P (A) 1
P (Ac ) = 1 P (A)
P (B \ Ac ) = P (B) P (A \ B)
P (A [ B) = P (A) + P (B) P (A \ B)
For some elementary probability models, it is useful to have simple rules to count the number
of objects in a set.
When counting the number of objects in a set, there are two important distinctions. Counting
may be with replacement or without replacement. Counting may be ordered or unordered.
For example, consider a lottery where you pick six numbers from the set 1, 2, ..., 49. This selection is
without replacement if you are not allowed to select the same number twice, and is with replacement
if this is allowed. Counting is ordered or not depending on whether the sequential order of the
numbers is relevant to winning the lottery. Depending on these two distinctions, we have four
expressions for the number of objects (possible arrangements) of size r from n objects.
Without With
Replacement Replacement
n!
Ordered (n r)! nr
n n+r 1
Unordered r r
In the lottery example, if counting is unordered and without replacement, the number of
potential combinations is 49
6 = 13; 983; 816.
If P (B) > 0 the conditional probability of the event A given the event B is
P (A \ B)
P (A j B) = :
P (B)
151
For any B; the conditional probability function is a valid probability function where S has been
replaced by B: Rearranging the de nition, we can write
P (A \ B) = P (A j B) P (B)
which is often quite useful. We can say that the occurrence of B has no information about the
likelihood of event A when P (A j B) = P (A); in which case we nd
We say that the events A and B are statistically independent when (16.1) holds. Furthermore,
we say that the collection of events A1 ; :::; Ak are mutually independent when for any subset
fAi : i 2 Ig; !
\ Y
P Ai = P (Ai ) :
i2I i2I
Theorem 2 (Bayes' Rule). For any set B and any partition A1 ; A2 ; ::: of the sample space, then
for each i = 1; 2; :::
P (B j Ai ) P (Ai )
P (Ai j B) = P1
j=1 P (B j Aj ) P (Aj )
2. F (x) is nondecreasing in x
3. F (x) is right-continuous
We say that the random variable X is discrete if F (x) is a step function. In the latter case,
the range of X consists of a countable set of real numbers 1 ; :::; r : The probability function for
X takes the form
P (X = j ) = j ; j = 1; :::; r (16.3)
152
P
where 0 j 1 and rj=1 j = 1.
We say that the random variable X is continuous if F (x) is continuous in x: In this case P (X =
) = 0 for all 2 R so the representation (16.3) is unavailable. Instead, we represent the relative
probabilities by the probability density function (PDF)
d
f (x) = F (x)
dx
so that Z x
F (x) = f (u)du
1
and Z b
P (a X b) = f (x)dx:
a
These expressions only make sense if F (x) is di erentiable. While there are examples of continuous
random variables which do not possess a PDF, these cases are unusualRand are typically ignored.
1
A function f (x) is a PDF if and only if f (x) 0 for all x 2 R and 1 f (x)dx:
16.3 Expectation
For any measurable real function g; we de ne the mean or expectation Eg(X) as follows. If X
is discrete,
Xr
Eg(X) = g( j ) j ;
j=1
and if X is continuous Z 1
Eg(X) = g(x)f (x)dx:
1
The latter is well de ned and nite if
Z 1
jg(x)j f (x)dx < 1: (16.4)
1
153
Since E (a + bX) = a + bEX; we say that expectation is a linear operator.
For m > 0; we de ne the m0 th moment of X as EX m and the m0 th central moment as
E (X EX)m :
Two special
p moments are the mean = EX and variance 2 = E (X )2 = EX 2 2 : We
call = 2 the standard deviation of X: We can also write 2 = V ar(X). For example, this
M ( ) = E exp ( X) :
The MGF does not necessarily exist. However, when it does and E jXjm < 1 then
dm
M( ) = E (X m )
d m =0
C( ) = E exp (i X) :
p
where i = 1 is the imaginary unit. The CF always exists, and when E jXjm < 1
dm
C( ) = im E (X m ) :
d m =0
P (X = x) = px (1 p)1 x
; x = 0; 1; 0 p 1
EX = p
V ar(X) = p(1 p)
Binomial
n x
P (X = x) = p (1 p)n x
; x = 0; 1; :::; n; 0 p 1
x
EX = np
V ar(X) = np(1 p)
154
Geometric
P (X = x) = p(1 p)x 1
; x = 1; 2; :::; 0 p 1
1
EX =
p
1 p
V ar(X) =
p2
Multinomial
n!
P (X1 = x1 ; X2 = x2 ; :::; Xm = xm ) = px1 p x2 pxmm ;
x1 !x2 ! xm ! 1 2
x1 + + xm = n;
p1 + + pm = 1
EX =
V ar(X) =
Negative Binomial
r+x 1
P (X = x) = p(1 p)x 1
; x = 1; 2; :::; 0 p 1
x
EX =
V ar(X) =
Poisson
x
exp ( )
P (X = x) = ; x = 0; 1; 2; :::; >0
x!
EX =
V ar(X) =
We now list some important continuous distributions.
Beta
( + )
f (x) = x 1 (1 x) 1 ; 0 x 1; > 0; >0
( ) ( )
=
+
V ar(X) =
( + + 1) ( + )2
Cauchy
1
f (x) = ; 1<x<1
(1 + x2 )
EX = 1
V ar(X) = 1
155
ExponentiaI
1 x
f (x) = exp ; 0 x < 1; >0
EX =
2
V ar(X) =
Logistic
exp ( x)
f (x) = ; 1 < x < 1;
(1 + exp ( x))2
EX = 0
V ar(X) =
Lognormal
!
1 (ln x )2
f (x) = p exp 2
; 0 x < 1; >0
2 x 2
2
EX = exp + =2
2 2
V ar(X) = exp 2 + 2 exp 2 +
Pareto
f (x) = +1
; x < 1; > 0; >0
x
EX = ; >1
1
2
V ar(X) = ; >2
( 1)2 ( 2)
Uniform
1
f (x) = ; a x b
b a
a+b
EX =
2
(b a)2
V ar(X) =
12
Weibull
1 x
f (x) = x exp ; 0 x < 1; > 0; >0
1= 1
EX = 1+
2= 2 2 1
V ar(X) = 1+ 1+
156
16.5 Multivariate Random Variables
A pair of bivariate random variables (X; Y ) is a function from the sample space into R2 : The joint
CDF of (X; Y ) is
F (x; y) = P (X x; Y y) :
If F is continuous, the joint probability density function is
@2
f (x; y) = F (x; y):
@x@y
For a Borel measurable set A 2 R2 ;
Z Z
P ((X < Y ) 2 A) = f (x; y)dxdy
A
For any measurable function g(x; y);
Z 1 Z 1
Eg(X; Y ) = g(x; y)f (x; y)dxdy:
1 1
The marginal distribution of X is
FX (x) = P (X x)
= lim F (x; y)
y!1
Z x Z 1
= f (x; y)dydx
1 1
so the marginal density of X is
Z 1
d
fX (x) = FX (x) = f (x; y)dy:
dx 1
Similarly, the marginal density of Y is
Z 1
fY (y) = f (x; y)dx:
1
The random variables X and Y are de ned to be independent if f (x; y) = fX (x)fY (y):
Furthermore, X and Y are independent if and only if there exist functions g(x) and h(y) such that
f (x; y) = g(x)h(y):
If X and Y are independent, then
Z Z
E (g(X)h(Y )) = g(x)h(y)f (y; x)dydx
Z Z
= g(x)h(y)fY (y)fX (x)dydx
Z Z
= g(x)fX (x)dx h(y)fY (y)dy
= Eg (X) Eh (Y ) : (16.5)
157
if the expectations exist. For example, if X and Y are independent then
E(XY ) = EXEY:
MZ ( ) = E exp ( (X + Y ))
= E (exp ( X) exp ( Y ))
0 0
= E exp X E exp Y
= MX ( )MY ( ): (16.6)
V ar (X + Y ) = V ar(X) + V ar(Y );
F (x) = P (X x)
@k
f (x) = F (x):
@x1 @xk
158
where the symbol dx denotes dx1 dxk : In particular, we have the k 1 multivariate mean
= EX
= E (X ) (X )0
= EXX 0 0
The conditional mean m(x) is a function, meaning that when X equals x; then the expected value
of Y is m(x):
159
Similarly, we de ne the conditional variance of Y given X = x as
2
(x) = V ar (Y j X = x)
= E (Y m(x))2 j X = x
= E Y2 jX =x m(x)2 :
Evaluated at x = X; the conditional mean m(X) and conditional variance 2 (x) are random
variables, functions of X: We write this as E(Y j X) = m(X) and V ar (Y j X) = 2 (X): For
example, if E (Y j X = x) = + x; then E (Y j X) = + X; a transformation of X:
The following are important facts about conditional expectations.
Simple Law of Iterated Expectations:
E (E (Y j X)) = E (Y ) (16.7)
Proof :
E (E (Y j X)) = E (m(X))
Z 1
= m(x)fX (x)dx
1
Z 1 Z 1
= yfY jX (y j x) fX (x)dydx
1 1
Z 1Z 1
= yf (y; x) dydx
1 1
= E(Y ):
Law of Iterated Expectations:
E (E (Y j X; Z) j X) = E (Y j X) (16.8)
Conditioning Theorem. For any function g(x);
E (g(X)Y j X) = g (X) E (Y j X) (16.9)
Proof : Let
h(x) = E (g(X)Y j X = x)
Z 1
= g(x)yfY jX (y j x) dy
1
Z 1
= g(x) yfY jX (y j x) dy
1
= g(x)m(x)
where m(x) = E (Y j X = x) : Thus h(X) = g(X)m(X), which is the same as E (g(X)Y j X) =
g (X) E (Y j X) :
160
16.7 Transformations
Suppose that X 2 Rk with continuous distribution function FX (x) and density fX (x): Let Y =
g(X) where g(x) : Rk ! Rk is one-to-one, di erentiable, and invertible. Let h(y) denote the
inverse of g(x). The Jacobian is
@
J(y) = det h(y) :
@y 0
Consider the univariate case k = 1: If g(x) is an increasing function, then g(X) Y if and
only if X h(Y ); so the distribution function of Y is
FY (y) = P (g(X) y)
= P (X h(Y ))
= FX (h(Y ))
so the density of Y is
d d
fY (y) = FY (y) = fX (h(Y )) h(y):
dy dy
If g(x) is a decreasing function, then g(X) Y if and only if X h(Y ); so
FY (y) = P (g(X) y)
= 1 P (X h(Y ))
= 1 FX (h(Y ))
This is known as the change-of-variables formula. This same formula (16.10) holds for k > 1;
but its justi cation requires deeper results from analysis.
As one example, take the case X U [0; 1] and Y = ln(X). Here, g(x) = ln(x) and
h(y) = exp( y) so the Jacobian is J(y) = exp(y): As the range of X is [0; 1]; that for Y is [0,1):
Since fX (x) = 1 for 0 x 1 (16.10) shows that
an exponential density.
161
16.8 Normal and Related Distributions
The standard normal density is
1 x2
(x) = p exp ; 1 < x < 1:
2 2
This density has all moments nite. Since it is symmetric about zero all odd moments are zero. By
iterated integration by parts, we can also show that EX 2 = 1 and EX 4 = 3: It is conventional to
write X N (0; 1); and to denote the standard normal density function by (x) and its distribution
function by (x): The latter has no closed-form solution.
If Z is standard normal and X = + Z; then using the change-of-variables formula, X has
density !
1 (x )2
f (x) = p exp ; 1 < x < 1:
2 2 2
which is the univariate normal density. The mean and variance of the distribution are and
2 ; and it is conventional to write X N ( ; 2 ).
k
For x 2 R ; the multivariate normal density is
1 (x )0 1 (x )
f (x) = k=2 1=2
exp ; x 2 Rk :
(2 ) det ( ) 2
The mean and covariance matrix of the distribution are and ; and it is conventional to write
X N ( ; ).
It useful to observe that the MGF and CF of the multivariate normal are exp 0 + 0 =2
and exp i 0 0
=2 ; respectively.
k
If X 2 R is multivariate normal and the elements of X are mutually uncorrelated, then
= diagf 2j g is a diagonal matrix. In this case the density function can be written as
!!
2 2 2 2
1 (x1 1) = 1 + + (xk k) = k
f (x) = exp
(2 )k=2 1 k
2
k 2
!
Y 1 xj j
= 1=2
exp 2
j=1 (2 ) j 2 j
which is the product of marginal univariate normal densities. This shows that if X is multivariate
normal with uncorrelated elements, then they are mutually independent.
Another useful fact is that if X N ( ; ) and Y = a + BX with B an invertible matrix, then
by the change-of-variables formula, the density of Y is
!
0 1
1 (y Y ) (y Y )
f (y) = exp Y
; x 2 Rk :
(2 )k=2 det ( Y )1=2 2
162
where Y = a+B and Y = B B 0 ; where we used the fact that det (B B 0 )1=2 = det ( )1=2 det (B) :
This shows that linear transformations of normals are also normal.
Let X N (0; Ir ) and set Q = X 0 X: We show at the end of this section that the Q has density
1
f (y) = r y r=2 1
exp ( y=2) ; y 0: (16.11)
2 2r=2
and is known as the chi-square density with r degrees of freedom, denoted 2. Its mean and
r
variance are = r and 2 = 2r: A useful result is:
Z
tr = p :
Q=r
Z 1
1
E exp (tQ) = r y r=2 1
exp (ty) exp ( y=2) dy
0 2 2r=2
r=2
= (1 2t) (16.13)
R1
where the second equality uses the fact that 0 y a 1 exp ( by) dy = b a (a); which can be found
by applying change-of-variables to the gamma function. For Z N (0; 1) the distribution of Z 2 is
p
P Z2 y = 2P (0 Z y)
Z py
1 x2
= 2 p exp dx
0 2 2
Z y
1 s
= 1 1=2
s 1=2 exp ds
0 2 2
2
1 p
using the change{of-variables s = x2 and the fact 2 = : Thus the density of Z 2 is (16.11)
with r = 1: From (16.13), we see that the MGF of Z 2 is (1 2t) 1=2 . Since we can write Q =
163
Pr
X 0X = 2
j=1 Zj where the Zj are independent N (0; 1), (16.6) can be used to show that the MGF
r=2
of Q is (1 2t) ; which we showed in (16.13) is the MGF of the density (16.11).
Proof of (16.8.1). The fact that A > 0 means that we can write A = CC 0 where C is
non-singular. Then A 1 = C 10 C 1 and
1 1 10 1
C Z N (0; C AC ) = N (0; C CC 0 C 10
) = N (0; Iq ):
Thus
0
Z 0A 1
Z = Z 0C 10
C 1
Z= C 1
Z C 1
Z 2
q:
Proof of (16.12). Using the simple law of iterated expectations, tr has distribution function
!
Z
F (x) = P p x
Q=r
( r )
Q
= E Z x
r
" r !#
Q
= E P Z x jQ
r
r !
Q
= E x
r
164
of the distribution F: The space is the set of permissible value for : In this setting the method
of maximum likelihood is the appropriate technique for estimation and inference on :
If the distribution F is continuous then the density of Yi can be written as f (y; ) and the joint
density of a random sample Y~ = (Y1 ; :::; Yn ) is
n
Y
fn Y~ ; = f (Yi ; ) :
i=1
The likelihood of the sample is this joint density evaluated at the observed sample values, viewed
as a function of . The log-likelihood function is its natural log
n
X
Ln ( ) = ln f (Yi ; ) :
i=1
If the distribution F is discrete, the likelihood and log-likelihood are constructed by setting
f (y; ) = P (Y = y; ) :
De ne the Hessian
@2
H= E ln f (Yi ; 0 ) (16.14)
@ @ 0
and the outer product matrix
@ @ 0
=E ln f (Yi ; 0) ln f (Yi ; 0) : (16.15)
@ @
Two important features of the likelihood are
Theorem 16.9.1
@
E ln f (Yi ; ) =0 (16.16)
@ = 0
H= I0 (16.17)
The matrix I0 is called the information, and the equality (16.17) is often called the infor-
mation matrix equality.
The Cramer-Rao Theorem gives a lower bound for estimation. However, the restriction to
unbiased estimators means that the theorem has little direct relevance for nite sample e ciency.
The maximum likelihood estimator or MLE ^ is the parameter value which maximizes
the likelihood (equivalently, which maximizes the log-likelihood). We can write this as
^ = argmax Ln ( ):
2
165
In some simple cases, we can nd an explicit expression for ^ as a function of the data, but these
cases are rare. More typically, the MLE ^ must be found by numerical methods.
Why do we believe that the MLE ^ is estimating the parameter ? Observe that when stan-
dardized, the log-likelihood is a sample average
n
1 1X
Ln ( ) = ln f (Yi ; ) !p E ln f (Yi ; ) L( ):
n n
i=1
As the MLE ^ maximizes the left-hand-side, we can see that it is an estimator of the maximizer
of the right-hand-side. The rst-order condition for the latter problem is
@ @
0= L( ) = E ln f (Yi ; )
@ @
which holds at = 0by (16.16). In fact, under conventional regularity conditions, ^ is consistent
for this value, ^ !p 0 as n ! 1:
p
Theorem 16.9.3 Under regularity conditions, n ^ 0 !d N 0; I0 1 .
Thus in large samples, the approximate variance of the MLE is (nI0 ) 1 which is the Cramer-
Rao lower bound. Thus in large samples the MLE has approximately the best possible variance.
Therefore the MLE is called asymptotically e cient.
Typically, to estimate the asymptotic variance of the MLE we use an estimate based on the
Hessian formula (16.14)
X n
^ = 1 @2 ^
H 0 ln f Yi ; (16.18)
n @ @
i=1
We then set I^0 1 = H ^ 1 : Asymptotic standard errors for ^ are then the square roots of the diagonal
elements of n I^0 :
1 1
Sometimes a parametric density function f (y; ) is used to approximate the true unknown
density f (y); but it is not literally believed that the model f (y; ) is necessarily the true density.
In this case, we refer to Ln (^) as a quasi-likelihood and the its maximizer ^ as a quasi-mle or
QMLE.
In this case there is not a \true" value of the parameter : Instead we de ne the pseudo-true
value 0 as the maximizer of
Z
E ln f (Yi ; ) = f (y) ln f (y; ) dy
166
the Kullback-Leibler information distance between the true density f (y) and the parametric density
f (y; ): Thus the QMLE 0 is the value which makes the parametric density \closest" to the true
value according to this measure of distance. The QMLE is consistent for the pseudo-true value,
but has a di erent covariance matrix than in the pure MLE case, since the information matrix
equality (16.17) does not hold. A minor adjustment to Theorem (16.9.3) yields the asymptotic
distribution of the QMLE:
p
n ^ 0 !d N (0; V ) ; V =H 1
H 1
Asymptotic standard errors (sometimes called qmle standard errors) are then the square roots of
the diagonal elements of n 1 V^ :
Proof of Theorem 16.9.1. To see (16.16);
Z
@ @
E ln f (Yi ; ) = ln f (y; ) f (y; 0 ) dy
@ = 0 @ = 0
Z
@ f (y; 0 )
= f (y; ) dy
@ f (y; ) = 0
Z
@
= f (y; ) dy
@ = 0
@
= 1 = 0:
@ = 0
167
Proof of Cramer-Rao Lower Bound.
n
X
@ @
S= ln fn Y~ ; 0 = ln f (Yi ; 0)
@ @
i=1
which by Theorem (16.9.1) has mean zero and variance nH: Write the estimator ~ = ~(Y~ ) as a
function of the data. Since ~ is unbiased for any ;
Z
~
= E = ~(~ y )f (~
y ; ) d~
y
Proof of Theorem 16.9.3 Taking the rst-order condition for maximization of Ln ( ), and
making a rst-order Taylor series expansion,
@
0 = Ln ( )
@ =^
n
X @
= ln f Yi ; ^
@
i=1
n
X Xn
@ @2 ^
' ln f (Yi ; 0) + 0 ln f (Yi ; n) 0 ;
@ @ @
i=1 i=1
where n lies on a line segment joining ^ and 0 : (Technically, the speci c value of n varies by
row in this expansion.) Rewriting this equation, we nd
n
! 1 n !
X @ 2 X @
^ 0 = ln f (Yi ; n ) ln f (Yi ; 0 ) :
i=1
@ @ 0 i=1
@
@
Since @ ln f (Yi ; 0) is mean-zero with covariance matrix ; an application of the CLT yields
n
1 X @
p ln f (Yi ; 0) !d N (0; ) :
n @
i=1
168
The analysis of the sample Hessian is somewhat more complicated due to the presence of n : Let
2
H( ) = @ @@ 0 ln f (Yi ; ) : If it is continuous in ; then since n !p 0 we nd H( n ) !p H and
so
n n
1 X @2 1X @2
0 ln f (Yi ; n) = 0 ln f (Yi ; n) H( n) + H( n)
n @ @ n @ @
i=1 i=1
! pH
169
Chapter 17
17.1 Inequalities
Triangle inequality.
jX + Y j jXj + jY j :
C r inequality. 8
< jXjr + jY jr : 0<r 1
r
jX + Y j :
: 1 (jXjr
2r + jY jr ) r 1
Proofs of the following statements are at the end of this section.
Jensen's Inequality. If g( ) : R ! R is convex, then
g(E(X)) E(g(X)):
E jXY j kXkp kY kq :
Cauchy-Schwarz Inequality.
E jXY j kXk2 kY k2
170
kX + Y kp kXkp + kY kp :
1
P (g(X) > ) Eg(X):
Proof of Jensen's Inequality. Let a + bx be the tangent line to g(x) at x = EX: Since g(x)
is convex, tangent lines lie below it. So for all x; g(x) a + bx yet g(EX) = a + bEX since the
curve is tangent at EX: Applying expectations, Eg(X) a + bEX = g(EX); as stated.
Proof of Lr Inequality. Let Y = jXjr and g(x) = xp=r ; which is convex. By Jensen's inequality,
g(EY ) Eg(Y ); so
(E jXjr )p=r E (jXjr )p=r = E jXjp
Raised to the 1=p power is the inequality.
Proof of Holder's Inequality. By renormalization, without loss of generality we can assume
E jXjp = 1 and E jY jq = 1, so that we need to show E jXY j 1: By the theorem of geometric
means, for A > 0 and B > 0
1 1
A1=p B 1=q A+ B (17.1)
p q
so h i 1 1 1 1
E jXY j = E (jXjp )1=p (jY jq )1=q E jXjp + jY jq = + = 1
p q p q
as needed.
We now show (17.1). De ne a random variable T which takes the value ln A with probability
1=p; and the value ln B with probability 1=q: The exponential function is convex, so Jensen's
inequality yields
1 1
A1=p B 1=q = exp ln A + ln B
p q
= exp(E(T ))
E (exp(T ))
1 1
= A+ B
p p
171
For p > 1; de ne its conjugate q = p=(p 1) (so 1=p + 1=q = 1): By the triangle inequality,
Holder's inequality,
E jX + Y jp = E jX + Y jp 1
jX + Y j
= E jX + Y jp 1
jXj + E jX + Y jp 1
jY j
= jX + Y jp 1
kXkp + jX + Y jp 1
kY kp
q q
= (E jX + Y jp )1 1=p
kXkp + kY kp
where the nal inequality holds since (p 1)q = p: Multiplying both sides by (E jX + Y jp )1=p 1
Proof of Markov's Inequality. Set Y = g(X); and let f g denote the indicator function: For
simplicity suppose that Y has density f (y): Then
Z Z Z Z 1
P (Y > ) = EfY > g = fy > gf (y)dy = f (y)dy yf (y)dy yf (y)dy = E(Y )
fy> g fy> g 1
the second-to-last inequality using the region of integration fy > g: Hence P (Y > ) 1 E(Y )
and we are done.
Proof: Without loss of generality, we can set E(X) = 0 (by recentering Xi on its expectation).
We need to show that for all > 0 and > 0 there is some N < 1 so that for all n N;
P Xn > : Fix and : Set " = =3: Pick C < 1 large enough so that
172
(where 1 ( ) is the indicator function) which is possible since E jXj < 1: De ne the random vectors
E Zn E jZi j
E jXi j 1 (jXi j > C) + jE (Xi 1 (jXi j > C))j
2E jXi j 1 (jXi j > C)
2": (17.3)
By Jensen's inequality, the fact that the Wi are iid and mean zero, and the bound jWi j 2C;
2 2
E Wn EW n
EWi2
=
n
4C 2
n
"2 (17.4)
the equality by the de nition of ": We have shown that for any > 0 and > 0 then for all
n 36C 2 = 2 2 ; P X n > ; as needed.
P lim Zn = Z = 1:
n!1
173
Almost sure convergence is stronger than convergence in probability. For example, consider
the sequence of discrete random variables Zn with
1
P (Zn = 1) = n
1
P (Zn = 0) = 1 n
You can see that Zn !p Z yet Zn does not converge almost surely:
Strong Law of Large Numbers (SLLN). If fX1 ; :::; Xn g are iid with E jXj < 1 and
EX = then as n ! 1
n
1X
Xn = Xi !a:s: :
n
i=1
To show the SLLN, we need some intermediate results. The rst is a maximal form of Markov's
inequality.
n
1 X
P max Si > 2 EXi2 : (17.5)
1 i n
i=1
Proof. Since
n
X i
X
E (Sn j X1 ; :::; Xi ) = E (Xj j X1 ; :::; Xi ) = X j = Si
j=1 j=1
174
n o
2 2
Let Ii 1 = Si2 > ; maxj<i Sj2 which are disjoint events. Then
2 2
P max jSi j > = P max Si2 > 2
1 i n 1 i n
n
X
2
= P (Ii 1)
i=1
n
X
2 2 2
= P Ii 1 Si >
i=1
n
X
2
E Ii 1 Si
i=1
n
X
E Ii 1E Sn2 j X1 ; :::; Xi 1
i=1
Xn
2
= E Ii 1 Sn
i=1
E Sn2
Xn
= EXi2 :
i=1
The fourth line is Markov's inequality, the fth is (17.6), the sixth uses the conditioning theorem
and the law of iterated expectations.
The next two are summation results. P
Toeplitz Lemma. Suppose xn ! x with jxn xj B and for aj 0; nj=1 aj ! 1 as
n ! 1: Then as n ! 1 Pn
aj xj
Pj=1
n !x
j=1 aj
P Fix
Proof: " > 0: P
Find N1 < 1 such that jxn xj " for all n N1 : Then nd N2 < 1 such
that N 1
j=1 aj B" N2
j=1 aj : Then for all for n > N2 ;
Pn PN1 Pn
aj xj j=1 aj (xj x) j=N1 +1 aj (xj x)
Pj=1
n x = Pn + Pn 2":
j=1 aj j=1 aj j=1 aj
Pn
Kronecker Lemma. If bn is increasing with bn ! 1; and Sn = j=1 xj ! x, then as n ! 1
n
1 X
bj xj ! 0:
bn
j=1
175
Proof: Set aj = bj bj 1 0: Then
n
X n
X n
X
bj xj = bj S j bj S j 1
j=1 j=1 j=1
Xn Xn n
X
= bj S j bj 1 Sj 1 aj Sj 1
j=1 j=1 j=1
n
X
= bn Sn aj Sj 1:
j=1
Hence as n ! 1 Pn
n
1 X j=1 aj Sj 1
bj xj = Sn P n !x x=0
bn j=1 aj
j=1
almost surely.
176
17.4 Convergence in Distribution
Let Zn be a random variable with distribution Fn (x) = P (Zn x) : We say that Zn converges
in distribution to Z as n ! 1, denoted Zn !d Z; where Z has distribution F (x) = P (Z x) ;
if for all x at which F (x) is continuous, Fn (x) ! F (x) as n ! 1:
Central Limit Theorem (CLT). If Xi 2 Rk is iid and E jXi j2 < 1; then as n ! 1
n
p 1 X
n Xn =p (Xi ) !d N (0; V ) :
n
i=1
where = EX and V = E (X ) (X )0 :
Proof: Without loss of generality, it is su cient to consider the case = 0 and V = Ik : For
2 Rk ; let C( ) = E exp i 0 X denote the characteristic function of X and set c( ) = ln C( ).
Then observe
@
C( ) = iE X exp i 0 X
@
@2
C( ) = i2 E XX 0 exp i 0 X
@ @ 0
so when evaluated at =0
C(0) = 1
@
C(0) = iE (X) = 0
@
@2
C(0) = E XX 0 = Ik :
@ @ 0
Furthermore,
@ @
c ( ) = c( ) = C( ) 1 C( )
@ @
@2 1 @
2
2 @ @
c ( ) = c( ) = C( ) C( ) C( ) C( ) 0 C( )
@ @ 0 @ @ 0 @ @
so when evaluated at =0
c(0) = 0
c (0) = 0
c (0) = Ik :
177
where lies on the line segment joining 0 and :
p p
We now compute Cn ( ) = E exp i 0 nX n the characteristic function of nX n : By the
properties of the exponential function, the independence of the Xi ; the de nition of c( ) and
(17.7)
0 1
n 0
X
1
ln Cn ( ) = log E exp @i p Xj A
n
j=1
n
Y 1
= log E exp i p 0 Xj
n
j=1
n
Y 1
= log E exp i p 0 Xj
n
i=1
= nc p
n
1 0
= c ( n)
2
p
where n ! 0 lies on the line segment joining 0 and = n: Since c ( n) ! c (0) = Ik ; we see
that as n ! 1;
1 0
Cn ( ) ! exp
2
the characteristic function of the N (0; Ik ) distribution. This is su cient to establish the theorem.
178
Proof : By a vector Taylor series expansion, for each element of g;
gj ( n) = gj ( 0 ) + gj ( jn ) ( n 0)
where nj lies on the line segment between n and 0 and therefore converges in probability to 0:
It follows that ajn = gj ( jn ) gj !p 0: Stacking across elements of g; we nd
p p
n (g ( n ) g( 0 )) = (g + an ) n ( n 0 ) !d g N (0; ) = N 0; g g0
179
Chapter 18
180
| 1) + 2k(b a)=G2 : k = 0; :::; Gg: Let k^ = argmin0 k G Q( 0 (k)): The estimate of ^ is
(^
0 ^
(k): The advantage of the two-step method over a one-stepp grid search is that the number of
function evaluations has been reduced from (b a)=" to 2 (b a)=" which can be substantial.
The disadvantage is that if the function Q( ) is irregular, the rst-step grid may not bracket ^
which thus would be missed.
Q( + j ") Q( )
gj ( ) ' :
"
Similarly,
Q( + j" + k ")
Q( + k ") Q( + j ") + Q( )
gjk ( ) '
"2
In many cases, numerical derivatives can work well but can be computationally costly relative to
analytic derivatives. In some cases, however, numerical derivatives can be quite unstable.
Most gradient methods are a variant of Newton's method which is based on a quadratic
approximation. By a Taylor's expansion for close to ^
0 = g(^) ' g( ) + H( ) ^
which implies
^= H( ) 1
g( ):
This suggests the iteration rule
^i+1 = i H( i ) 1
g( i ):
where
181
One problem with Newton's method is that it will send the iterations in the wrong direction if
H( i ) is not positive de nite. One modi cation to prevent this possibility is quadratic hill-climbing
which sets
^i+1 = i (H( i ) + i Im ) 1 g( i ):
where i is set just above the smallest eigenvalue of H( i ) if H( ) is not positive de nite.
Another productive modi cation is to add a scalar steplength i : In this case the iteration
rule takes the form
i+1 = i Di gi i (18.2)
where gi = g( i ) and Di = H( i ) 1 for Newton's method and Di = (H( i ) + i Im ) 1 for quadratic
hill-climbing.
Allowing the steplength to be a free parameter allows for a line search, a one-dimensional
optimization. To pick i write the criterion function as a function of
Q( ) = Q( i + Di gi )
a one-dimensional optimization problem. There are two common methods to perform a line search.
A quadratic approximation evaluates the rst and second derivatives of Q( ) with respect to
; and picks i as the value minimizing this approximation. The half-step method considers the
sequence = 1; 1/2, 1/4, 1/8, ... . Each value in the sequence is considered and the criterion
Q( i + Di gi ) evaluated. If the criterion has improved over Q( i ), use this value, otherwise move
to the next element in the sequence.
Newton's method does not perform well if Q( ) is irregular, and it can be quite computationally
costly if H( ) is not analytically available. These problems have motivated alternative choices for
the weight matrix Di : These methods are called Quasi-Newton methods. Two popular methods
are do to Davidson-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS).
Let
gi = gi gi 1
i = i i 1
For any of the gradient methods, the iterations continue until the sequence has converged in
some sense. This can be de ned by examining whether j i i 1 j ; jQ ( i ) Q ( i 1 )j or jg( i )j
has become small.
182
18.3 Derivative-Free Methods
All gradient methods can be quite poor in locating the global minimum when Q( ) has several
local minima. Furthermore, the methods are not well de ned when Q( ) is non-di erentiable. In
these cases, alternative optimization methods are required. One example is the simplex method
of Nelder-Mead (1965).
A more recent innovation is the method of simulated annealing (SA). For a review see
Go e, Ferrier, and Rodgers (1994). The SA method is a sophisticated random search. Like the
gradient methods, it relies on an iterative sequence. At each iteration, a random variable is drawn
and added to the current value of the parameter. If the resulting criterion is decreased, this new
value is accepted. If the criterion is increased, it may still be accepted depending on the extent of
the increase and another randomization. The latter property is needed to keep the algorithm from
selecting a local minimum. As the iterations continue, the variance of the random innovations
is shrunk. The SA algorithm stops when a large number of iterations is unable to improve the
criterion. The SA method has been found to be successful at locating global minima. The downside
is that it can take considerable computer time to execute.
183
Bibliography
[1] Aitken, A.C. (1935): \On least squares and linear combinations of observations," Proceedings
of the Royal Statistical Society, 55, 42-48.
[2] Akaike, H. (1973): \Information theory and an extension of the maximum likelihood prin-
ciple." In B. Petroc and F. Csake, eds., Second International Symposium on Information
Theory.
[3] Anderson, T.W. and H. Rubin (1949): \Estimation of the parameters of a single equation in
a complete system of stochastic equations," The Annals of Mathematical Statistics, 20, 46-63.
[4] Andrews, D.W.K. (1988): \Laws of large numbers for dependent non-identically distributed
random variables,' Econometric Theory, 4, 458-467.
[5] Andrews, D.W.K. (1993), \Tests for parameter instability and structural change with un-
known change point," Econometrica, 61, 821-8516.
[6] Andrews, D.W.K. and M. Buchinsky: (2000): \A three-step method for choosing the number
of bootstrap replications," Econometrica, 68, 23-51.
[7] Andrews, D.W.K. and W. Ploberger (1994): \Optimal tests when a nuisance parameter is
present only under the alternative," Econometrica, 62, 1383-1414.
[8] Basmann, R. L. (1957): \A generalized classical method of linear estimation of coe cients in
a structural equation," Econometrica, 25, 77-83.
[9] Bekker, P.A. (1994): \Alternative approximations to the distributions of instrumental variable
estimators, Econometrica, 62, 657-681.
184
[13] Breusch, T.S. and A.R. Pagan (1979): \The Lagrange multiplier test and its application to
model speci cation in econometrics," Review of Economic Studies, 47, 239-253.
[14] Brown, B.W. and W.K. Newey (2002): \GMM, e cient bootstrapping, and improved infer-
ence ," Journal of Business and Economic Statistics.
[15] Carlstein, E. (1986): \The use of subseries methods for estimating the variance of a general
statistic from a stationary time series," Annals of Statistics, 14, 1171-1179.
[16] Chamberlain, G. (1987): \Asymptotic e ciency in estimation with conditional moment re-
strictions," Journal of Econometrics, 34, 305-334.
[17] Choi, I. and P.C.B. Phillips (1992): \Asymptotic and nite sample distribution theory for
IV estimators and tests in partially identi ed structural equations," Journal of Econometrics,
51, 113-150.
[18] Chow, G.C. (1960): \Tests of equality between sets of coe cients in two linear regressions,"
Econometrica, 28, 591-603.
[19] Davidson, J. (1994): Stochastic Limit Theory: An Introduction for Econometricians. Oxford:
Oxford University Press.
[20] Davison, A.C. and D.V. Hinkley (1997): Bootstrap Methods and their Application. Cambridge
University Press.
[21] Dickey, D.A. and W.A. Fuller (1979): \Distribution of the estimators for autoregressive time
series with a unit root," Journal of the American Statistical Association, 74, 427-431.
[22] Donald S.G. and W.K. Newey (2001): \Choosing the number of instruments," Econometrica,
69, 1161-1191.
[23] Dufour, J.M. (1997): \Some impossibility theorems in econometrics with applications to
structural and dynamic models," Econometrica, 65, 1365-1387.
[24] Efron, B. (1979): \Bootstrap methods: Another look at the jackknife," Annals of Statistics,
7, 1-26.
[25] Efron, B. and R.J. Tibshirani (1993): An Introduction to the Bootstrap, New York: Chapman-
Hall.
[26] Eicker, F. (1963): \Asymptotic normality and consistency of the least squares estimators for
families of linear regressions," Annals of Mathematical Statistics, 34, 447-456.
[27] Engle, R.F. and C.W.J. Granger (1987): \Co-integration and error correction: Representa-
tion, estimation and testing," Econometrica, 55, 251-276.
185
[28] Frisch, R. and F. Waugh (1933): \Partial time regressions as compared with individual
trends," Econometrica, 1, 387-401.
[29] Gallant, A.F. and D.W. Nychka (1987): \Seminonparametric maximum likelihood estima-
tion," Econometrica, 55, 363-390.
[30] Gallant, A.R. and H. White (1988): A Uni ed Theory of Estimation and Inference for Non-
linear Dynamic Models. New York: Basil Blackwell.
[31] Go e, W.L., G.D. Ferrier and J. Rogers (1994): \Global optimization of statistical functions
with simulated annealing," Journal of Econometrics, 60, 65-99.
[32] Gauss, K.F. (1809): \Theoria motus corporum coelestium," in Werke, Vol. VII, 240-254.
[33] Granger, C.W.J. (1969): \Investigating causal relations by econometric models and cross-
spectral methods," Econometrica, 37, 424-438.
[34] Granger, C.W.J. (1981): \Some properties of time series data and their use in econometric
speci cation," Journal of Econometrics, 16, 121-130.
[35] Granger, C.W.J. and T. Ter•asvirta (1993): Modelling Nonlinear Economic Relationships,
Oxford University Press, Oxford.
[36] Hall, A. R. (2000): \Covariance matrix estimation and the power of the overidentifying re-
strictions test," Econometrica, 68, 1517-1527,
[37] Hall, P. (1992): The Bootstrap and Edgeworth Expansion, New York: Springer-Verlag.
[38] Hall, P. (1994): \Methodology and theory for the bootstrap," Handbook of Econometrics, Vol.
IV, eds. R.F. Engle and D.L. McFadden. New York: Elsevier Science.
[39] Hall, P. and J.L. Horowitz (1996): \Bootstrap critical values for tests based on Generalized-
Method-of-Moments estimation," Econometrica, 64, 891-916.
[41] Hansen, B.E. (1992): \E cient estimation and testing of cointegrating vectors in the presence
of deterministic trends," Journal of Econometrics, 53, 87-121.
[42] Hansen, B.E. (1996): \Inference when a nuisance parameter is not identi ed under the null
hypothesis," Econometrica, 64, 413-430.
[43] Hansen, L.P. (1982): \Large sample properties of generalized method of moments estimators,
Econometrica, 50, 1029-1054.
186
[44] Hansen, L.P., J. Heaton, and A. Yaron (1996): \Finite sample properties of some alternative
GMM estimators," Journal of Business and Economic Statistics, 14, 262-280.
[45] Hausman, J.A. (1978): \Speci cation tests in econometrics," Econometrica, 46, 1251-1271.
[46] Heckman, J. (1979): \Sample selection bias as a speci cation error," Econometrica, 47, 153-
161.
[47] Imbens, G.W. (1997): \One step estimators for over-identi ed generalized method of moments
models," Review of Economic Studies, 64, 359-383.
[48] Imbens, G.W., R.H. Spady and P. Johnson (1998): \Information theoretic approaches to
inference in moment condition models," Econometrica, 66, 333-357.
[49] Jarque, C.M. and A.K. Bera (1980): \E cient tests for normality, homoskedasticity and serial
independence of regression residuals, Economic Letters, 6, 255-259.
[50] Johansen, S. (1988): \Statistical analysis of cointegrating vectors," Journal of Economic
Dynamics and Control, 12, 231-254.
[51] Johansen, S. (1991): \Estimation and hypothesis testing of cointegration vectors in the pres-
ence of linear trend," Econometrica, 59, 1551-1580.
[52] Johansen, S. (1995): Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Mod-
els, Oxford University Press.
[53] Johansen, S. and K. Juselius (1992): \Testing structural hypotheses in a multivariate cointe-
gration analysis of the PPP and the UIP for the UK," Journal of Econometrics, 53, 211-244.
[54] Kitamura, Y. (2001): \Asymptotic optimality and empirical likelihood for testing moment
restrictions," Econometrica, 69, 1661-1672.
[55] Kitamura, Y. and M. Stutzer (1997): \An information-theoretic alternative to generalized
method of moments," Econometrica, 65, 861-874..
[56] Kunsch, H.R. (1989): \The jackknife and the bootstrap for general stationary observations,"
Annals of Statistics, 17, 1217-1241.
[57] Kwiatkowski, D., P.C.B. Phillips, P. Schmidt, and Y. Shin (1992): \Testing the null hypothesis
of stationarity against the alternative of a unit root: How sure are we that economic time
series have a unit root?" Journal of Econometrics, 54, 159-178.
[58] Lafontaine, F. and K.J. White (1986): \Obtaining any Wald statistic you want," Economics
Letters, 21, 35-40.
[59] Lovell, M.C. (1963): \Seasonal adjustment of economic time series," Journal of the American
Statistical Association, 58, 993-1010.
187
[60] MacKinnon, J.G. (1990): \Critical values for cointegration," in Engle, R.F. and C.W. Granger
(eds.) Long-Run Economic Relationships: Readings in Cointegration, Oxford, Oxford Univer-
sity Press.
[61] MacKinnon, J.G. and H. White (1985): \Some heteroskedasticity-consistent covariance matrix
estimators with improved nite sample properties," Journal of Econometrics, 29, 305-325.
[62] Magnus, J. R., and H. Neudecker (1988): Matrix Di erential Calculus with Applications in
Statistics and Econometrics, New York: John Wiley and Sons.
[63] Muirhead, R.J. (1982): Aspects of Multivariate Statistical Theory. New York: Wiley.
[64] Nelder, J. and R. Mead (1965): \A simplex method for function minimization," Computer
Journal, 7, 308-313.
[65] Newey, W.K. and K.D. West (1987): \Hypothesis testing with e cient method of moments
estimation," International Economic Review, 28, 777-787.
[66] Owen, Art B. (1988): \Empirical likelihood ratio con dence intervals for a single functional,"
Biometrika, 75, 237-249.
[67] Owen, Art B. (2001): Empirical Likelihood. New York: Chapman & Hall.
[68] Phillips, P.C.B. (1989): \Partially identi ed econometric models," Econometric Theory, 5,
181-240.
[69] Phillips, P.C.B. and S. Ouliaris (1990): \Asymptotic properties of residual based tests for
cointegration," Econometrica, 58, 165-193.
[70] Politis, D.N. and J.P. Romano (1996): \The stationary bootstrap," Journal of the American
Statistical Association, 89, 1303-1313.
[71] Potscher, B.M. (1991): \E ects of model selection on inference," Econometric Theory, 7,
163-185.
[72] Qin, J. and J. Lawless (1994): \Empirical likelihood and general estimating equations," The
Annals of Statistics, 22, 300-325.
[73] Ramsey, J. B. (1969): \Tests for speci cation errors in classical linear least-squares regression
analysis," Journal of the Royal Statistical Society, Series B, 31, 350-371.
[74] Rudin, W. (1987): Real and Complex Analysis, 3rd edition. New York: McGraw-Hill.
[75] Said, S.E. and D.A. Dickey (1984): \Testing for unit roots in autoregressive-moving average
models of unknown order," Biometrika, 71, 599-608.
[76] Shao, J. and D. Tu (1995): The Jackknife and Bootstrap. NY: Springer.
188
[77] Sargan, J.D. (1958): \The estimation of economic relationships using instrumental variables,"
Econometrica, 2 6, 393-415.
[78] Sheather, S.J. and M.C. Jones (1991): \A reliable data-based bandwidth selection method
for kernel density estimation, Journal of the Royal Statistical Society, Series B, 53, 683-690.
[79] Shin, Y. (1994): \A residual-based test of the null of cointegration against the alternative of
no cointegration," Econometric Theory, 10, 91-115.
[80] Silverman, B.W. (1986): Density Estimation for Statistics and Data Analysis. London: Chap-
man and Hall.
[81] Sims, C.A. (1972): \Money, income and causality," American Economic Review, 62, 540-552.
[82] Sims, C.A. (1980): \Macroeconomics and reality," Econometrica, 48, 1-48.
[83] Staiger, D. and J.H. Stock (1997): \Instrumental variables regression with weak instruments,"
Econometrica, 65, 557-586.
[84] Stock, J.H. (1987): \Asymptotic properties of least squares estimators of cointegrating vec-
tors," Econometrica, 55, 1035-1056.
[85] Stock, J.H. (1991): \Con dence intervals for the largest autoregressive root in U.S. macroe-
conomic time series," Journal of Monetary Economics, 28, 435-460.
[86] Stock, J.H. and J.H. Wright (2000): \GMM with weak identi cation," Econometrica, 68,
1055-1096.
[87] Theil, H. (1953): \Repeated least squares applied to complete equation systems," The Hague,
Central Planning Bureau, mimeo.
[88] Tobin, J. (1958): \Estimation of relationships for limited dependent variables," Econometrica,
2 6, 24-36.
[89] Wald, A. (1943): \Tests of statistical hypotheses concerning several parameters when the
number of observations is large," Transactions of the American Mathematical Society, 54,
426-482.
[90] Wang, J. and E. Zivot (1998): \Inference on structural parameters in instrumental variables
regression with weak instruments," Econometrica, 66, 1389-1404.
[91] White, H. (1980): \A heteroskedasticity-consistent covariance matrix estimator and a direct
test for heteroskedasticity," Econometrica, 48, 817-838.
[92] White, H. (1984): Asymptotic Theory for Econometricians, Academic Press.
[93] Zellner, A. (1962): \An e cient method of estimating seemingly unrelated regressions, and
tests for aggregation bias," Journal of the American Statistical Association, 57, 348-368.
189