Formulary Eco II

Formulas and Tables
for
Econometrics II
April 27, 2016

c Prof. Dr. Kai Carstensen
Institute for Statistics and Econometrics, CAU Kiel
Formulas and Tables for Econometrics II
Part A. Cross Section Econometrics

1 Preliminaries
1.1 Useful Theorems
Convergence in probability: A sequence of random variables {xN : N = 1, 2, . . .} converges
p
in probability to the constant a, xN a, if for all > 0,
Pr [|xN a| > ] 0 as N .
Uniform convergence: Let w be a random vector taking values in W RM , let be a

subset of RP , and let q : W R be a real-valued function. Assume that (a) is compact;
(b) for each , q(, ) is Borel measurable on W; (c) for each w W, q(w, ) is continuous
on ; and (d) |q(w, )| b(w) for all , where b is a nonnegative function on W such
that E[b(w)] < . Then uniform convergence in probability holds:

N
1 X p
max N q(wi , ) E[q(w, )] 0.

i=1
Mean value theorem (scalar version): Suppose that O is an open interval and f : O R
is a differentiable function. Then for any [a, b] O, there exists a c (a, b) such that
f (b) f (a) = f 0 (c) (b a).
Mean value theorem (vector version): Suppose G is an open convex subset of RP and
f : G R is a differentiable function of P arguments with gradient vector x f (x). Then for
any vectors a, b G, there exists a vector c between a and b such that
f (b) f (a) = x f (c) (b a).
1.2 Vector differentiation

Gradient and Hessian: Let f () be a twice continuously differentiable scalar function that
depends on the (P 1) vector = (1 , . . . , P )0 . The gradient is

f () f () f ()
f () = = ,...,
0 1 P
2
and the Hessian is

2 f (w,) 2 f (w,) 2 f (w,)

...
1 1
2 f (w,)
1 2
2 f (w,)
1 P
2 f (w,)
2 f (w, ) ...
= [ f ()]0 =
2 2 2 P

.
2 1
0 .. .. ...
. .
2 f (w,) 2 f (w,) 2 f (w,)
P 1 P 2
... P P
Example 1: Consider the K 1 vector function g() = f (), where f () is a L 1 vector

function of the P 1 vector and is a K L matrix of constants. Then
g() f ()
0 = = [ f ()] = f ().
0
In the simple case f () = ,
g()
0 = = .
0
Example 2: Consider the scalar quadratic function g() = f ()0 f (), where f () is a L 1
vector function of the P 1 vector and is a matrix of constants. Then
g() f ()0 f ()
0 = 0 = [f ()0 f ()] = f ()0 ( + 0 ) [ f ()] .

In the simple case that f () = and is symmetric,
g() 0
= = [ 0 ] = 2 0 .
0 0
1.3 Delta Method

Suppose the P -dimensional random variable has asymptotic normal distribution
d
N 1/2 ( o ) Normal(0, V).
Now let c : RQ , Q P , be a continuously differentiable function and assume that is in

the interior of the parameter space. Define the Q P Jacobian C() = c(). Then
d
c( o )]
N 1/2 [c() Normal(0, C()VC()0 ).
3
2 M-estimation
2.1 The M-estimator
M-estimator: Consider the population parameter o that minimizes the population function
min E [q(w, )] .

It is estimated by the M-estimator that minimizes the sample function

N
X
min N 1 q(wi , ).

i=1
Score: Suppose the objective function is once continuously differentiable with respect to .
Then the score is
0
0 q(w, ) q(w, )
s(w, ) = q(w, ) = ,...,
1 p
First order condition:

N
X
= 0.
s(wi , )
i=1
Hessian: Suppose the objective function is twice continuously differentiable with respect to
. Then the Hessian is

(1)
s (w, )
2 q(w, ) ..
H(w, ) = = s(w, ) =

0 .
(P )
s (w, )
2.2 Asymptotic Properties

Consistency: Suppose the M-estimator is identified and uniform convergence of the sample
average of q(wi , ) holds. Then the M-estimator is consistent,
p
o .
4
Asymptotic normality: Suppose the M-estimator is identified and some regularity condi-
tions are satisfied. Then the M-estimator is asymptotically normal,
d
N 1/2 ( o ) Normal(0, A1 1
o Bo Ao ),
where
Ao = E[H(w, o )]
and
Bo = E[s(w, o )s(w, o )0 ] = Var[s(w, o )].
2.3 Estimators of the Asymptotic Variance

Estimator of Ao : Either use
N
X N
X
= N 1
A = N 1
H(wi , ) i
H
i=1 i=1
or
N
X N
X
= N 1
A = N 1
A(xi , ) i,
A
i=1 i=1
where
A(xi , o ) = E[H(wi , o )|x].
Estimator of Bo :
N
X N
X
= N 1
B
s(wi , )s(w 0
i , ) = N
1
sis0i .
i=1 i=1
Estimator of Vo :

V [ N ( o )] = A
= Avar[ 1 B
A 1 .
Take the square roots of the elements on the main

Asymptotic standard errors for :
diagonal of
= Avar(
V [ ) = V/N
1 B
=A A 1 /N.

5
2.4 Generalized information matrix equality

The generalized information matrix equality means
E[s(w, o )s(w, o )0 ] = o2 E[H(w, o )], o2 > 0.
2.5 Nonlinear Least Squares

Assumptions. NLS.1: For some o , E(y|x) = m(x, o ). NLS.2: E {[m(x, o ) m(x, )]2 } >
0, for all , 6= o . NLS.3: Var(y|x) = Var(u|x) = o2 .
NLS estimator: Suppose E(y|x) = m(x, o ) and thus y = m(x, o ) + u, where E(u|x) = 0.
The NLS estimator minimizes
N
X
N 1 [yi m(xi , )]2 /2.
i=1
In terms of the M-estimator, the objective function specializes to
q(w, ) = [y m(x, )]2 /2.
NLS score: Suppose m(x, o ) is once continuously differentiable with respect to . Then the
score is
m(x, )
s(w, ) = 0 q(w, ) = u .

NLS Hessian: Suppose m(x, o ) is twice continuously differentiable with respect to . Then
the Hessian is
2 m(x, ) m(x, ) m(x, )
H(w, ) = u + .
0 0
NLS variance matrix: The asymptotic variance matrix is
Vo = Ao1 Bo A1
o ,
where the matrix Ao becomes

m(x, ) m(x, )
Ao = E[H(w, o )] = E
0
and the matrix Bo becomes

0 2 m(x, o ) m(x, o )
Bo = E[s(w, o )s(w, o ) ] = E u .
0
6
Homoscedasticity: Under homoscedasticity, i.e., under Assumption NLS.3, Var(y|x) =

Var(u|x) = o2 , the generalized information matrix equality holds. Hence,
Vo = o2 A1
o .
2.6 Inference
Wald test of linear hypotheses: To test the linear hypotheses Q H0 : R = r against
H1 : R 6= r, the Wald statistic is (distribution under H0 )
h i0 h i1 h i
a
WN R r R(V/N )R0 R r 2 . Q
Wald test of nonlinear hypotheses: To test the Q nonlinear hypotheses H0 : c() = 0

against H1 : c() 6= 0, the Wald statistic is (distribution under H0 )
h i1
0 C()(
V/N
0 a
WN c() )C() c() 2 . Q
2.7 Optimization Methods

Newton-Raphson method: To find the coefficient vector that is the root of the FOC
N
X
= 0,
s(wi , )
i=1
the Newton-Raphson method uses the iteration

" N #1 " N #
X X
{g+1} = {g} H(wi , {g} ) s(wi , {g} ) .
i=1 i=1
Generalized Gauss-Newton method: The generalized Gauss-Newton method uses the

iteration
" N #1 " N #
X X
{g+1} = {g} r A(xi , {g} ) s(wi , {g} ) .
i=1 i=1
BHHH method: The Berndt, Hall, Hall, and Hausman (BHHH) algorithm uses the iteration
" N #1 " N #
X X
{g+1} = {g} r s(wi , {g} )s(wi , {g} )0 s(wi , {g} ) .
i=1 i=1
7
3 Conditional Maximum Likelihood Estimation

3.1 The conditional maximum likelihood estimator
Conditional log likelihood for observation i:
ì () `(yi , xi , ) = log f (yi |xi ; ),
where f (yi |xi ; o ), o , is the conditional density for the random vector yi given the random
vector xi .
Sample log likelihood function:

N
X N
X
L() = ì () = log f (yi |xi ; )
i=1 i=1
Conditional maximum likelihood estimator: the CMLE of o is the vector that solves
N
X N
X
1 1 1
max N L() = max N ì () = max N log f (yi |xi ; ).

i=1 i=1
Score: Suppose the conditional log likelihood function is once continuously differentiable with
respect to . Then the score is
0
0 ì ì
si () = ì () = (), . . . , () .
1 P

N
X
= 0.
si ()
i=1
Hessian: Suppose the conditional log likelihood function is twice continuously differentiable
with respect to . Then the Hessian is
2 ` () 2 ` () 2 ì ()

i i
. . .
2 `1i 1
()
1 2
2 ì ())
1 P
2 ì ()
2 ì ()
2 1 2 2
. . . 2 P

Hi () = 0 = si () = . . .
.. .. ..

2 ì () 2 ì () 2 ì ()
P 1 P 2
... P P
Expectation of the Hessian:
Ao = E[Hi ( o )]
8
Fisher information matrix:
Bo = Var[si ( o )] = E[si ( o )si ( o )0 ].
Conditional information matrix equality (CIME): Under fairly general conditions, in

the maximum likelihood context the conditional information matrix equality holds
E[Hi ( o )|xi ] = E[si ( o )si ( o )0 |xi ]
Unconditional information matrix equality (UIME): By the law of iterated expecta-

tions,
Ao = E[Hi ( o )] = E[si ( o )si ( o )0 ] = Bo .

Consistency: Suppose conditions equivalent to those for the M-estimator are satisfied. Then
the CMLE is consistent,
p
o .
Asymptotic normality: Suppose conditions equivalent to those for the M-estimator are
satisfied. Then the CMLE is asymptotically normal,
d
N 1/2 ( o ) Normal(0, Vo ),
where Vo = A1
o
3.3 Estimators of the Asymptotic Variance

(a) Direct estimate:
" N
#1
X
=A
V 1 = N 1
Hi () .
i=1
(b) Using the conditional expectation:

" N
#1
X
=A
V 1 = N 1
A(xi , ) ,
i=1
where A(xi , o ) E[H(yi , xi , o )|xi )].
9
(c) Outer product of the score:

" N
#1
X
=A
V 1 = N 1 i ()
si ()s 0 .
i=1

diagonal of
= Avar(
V [ ) = V/N
1 /N.
=A

3.4 Inference
Wald test of linear hypotheses: To test the linear hypotheses Q H0 : R = r against
H1 : R 6= r, the Wald statistic is (distribution under H0 )
h i0 h i1 h i
a
WN R r R(V/N )R0 R r 2Q .
Wald test of nonlinear hypotheses: To test the Q nonlinear hypotheses H0 : c() = 0

against H1 : c() 6= 0, the Wald statistic is (distribution under H0 )
h i1
0 0 a
WN c() C()(V/N )C() c() 2Q .
Likelihood ratio (LR) test: To test the Q nonlinear hypotheses H0 : c() = 0 against
H1 : c() 6= 0, the LR statistic is (distribution under H0 )
L()]
a
LR 2[L() 2 , Q
where is the restricted estimator (estimated under H0 ) and is the unrestricted estimator
(estimated under H1 ).
Lagrange multiplier (LM) or score test: To test the Q nonlinear hypotheses H0 : c() =
0 against H1 : c() 6= 0, the LM statistic is (distribution under H0 )
N
!0 N
!
a
X X
LM N 1/2 si A 1 N 1/2 si 2 , Q
i=1 i=1
is the P 1 score evaluated at the restricted estimate and A

where si = si () is an estimator
of Ao . One of the following estimators can be used:
N
X N
X N
X
= N 1
A
Hi () or = N 1
A
A(xi , ) or = N 1
A sis0i .
i=1 i=1 i=1
10
4 Generalized Method of Moments Estimation

4.1 The Generalized Method of Moments Estimator
Moment restrictions: Let {w RM : i = 1, 2, . . .} denote a set of independent, identically
distributed random vectors, where some feature of the distribution of wi is indexed by the
P 1 parameter vector . It is assumed that for some function g(wi , ) RL , the parameter
o RP satisfies
E [g (wi , o )] = 0.
Generalized method of moments (GMM) estimator: The GMM estimator minimizes

" N
#0 " N
#
X X
min QN () = min N 1 N 1
g (wi , ) g (wi , ) ,

i=1 i=1
is an L L symmetric, positive semidefinite weighting matrix.

where

" N #0 " N #
X X

g(wi , ) 0.
g(wi , )
i=1 i=1
Expected gradient of the moment condition:
Go = E [ g (wi , o )] .
Variance of the moment condition:
o = E [g(wi , o )g(wi , o )0 ] = Var [g(wi , o )] .

Consistency: Suppose conditions similar to those for the M-estimator are satisfied and
p

o , where o is an LL positive definite matrix. Then the GMM estimator is consistent,
p
o .
11
Asymptotic normality: Suppose conditions equivalent to those for the M-estimator are
p

satisfied, o , where o is an L L positive definite matrix, and Go has rank P . Then
the GMM estimator is asymptotically normal,
d
N 1/2 ( o ) Normal(0, Vo ),
where
Vo = Ao1 Bo A1
o
with
Ao G0o o Go
and
Bo G0o o o o Go .
4.3 Estimation of the Variance

Estimator of o :
N
X
N 1
gi ()
gi () 0
i=1
Estimator of Go :
N
X
N 1
G
gi ().
i=1
Estimator of Ao :
=G
A 0
G
Estimator of Bo :
=G
B 0

G
Estimator of Vo :
=A
V 1 B
A 1
12

diagonal of
= Avar(
V [ ) = V/N
1 B
=A A 1 /N.

4.4 Efficient GMM estimation

p
opt
Optimal weighting matrix: The optimal weighting matrix is chosen such that 1
o ,
e.g.,
1 .
opt =

Efficient GMM estimator: The asymptotically efficient GMM estimator solves

" N
#0 " N
#
X 1 X
min N 1
gi () N 1
gi () .

i=1 i=1
Asymptotic distribution of the efficient GMM estimator:

d
N o Normal (0, Vo ) ,
where
1
Vo = G0o 1

o Go .
Asymptotic standard errors for the efficient GMM estimator: Take the square roots
of the elements on the main diagonal of
i1
0 1
h
[
V = Avar() = V/N = G G /N.
4.5 Inference
Test of the validity of the moment conditions: Hansens J statistic is (distribution
under H0 )
" N
#0 " N #
X 1 X a
= N N 1
J = N QN ()
gi ()
N 1 gi () 2LP ,
i=1 i=1
where L is the number of moment conditions and P is the number of parameters.
13
GMM distance statistic: To test the Q nonlinear hypotheses H0 : c() = 0 against H1 :

c() 6= 0, the GMM distance statistic is (distribution under H0 )
(" N #0 " N # " N #0 " N #)
X X X 1 X d

gi () 1
gi ()
gi ()
gi () /N 2 , Q
i=1 i=1 i=1 i=1
where is the restricted estimator (estimated under H0 ), is the unrestricted estimator (esti-
mated under H1 ), and is obtained from an initial unrestricted estimator.
5 Binomial Choice Models

5.1 Model setup
Latent variable representation: The observable variable yi takes the values 0 and 1 ac-
cording to
(
1 if yi > 0
yi =
0 if yi 0,
where yi is a continuous latent variable that is determined by
yi = xi + ei .
Distribution of ei : The error ei is assumed to be distributed according to the twice contin-

uously differentiable distribution function (cdf) G() that has symmetric first derivative (pdf)
g(). Moreover, E(ei ) = 0 (inclusion of an intercept in the latent model).
Conditional probability that yi = 1:

p(xi ) = Pr(yi = 1|xi ) = G(xi ).
Conditional expectation:
E(y|x) = G(x).
5.2 Conditional Maximum Likelihood Estimation

Log likelihood function: for observation i
ì () = log f (yi |xi ; ) = yi log[G(xi )] + (1 yi ) log[1 G(xi )].
and for the full sample of size N
N
X N
X
L() = ì () = {yi log[G(xi )] + (1 yi ) log[1 G(xi )]} .
i=1 i=1
14
Score:

yi 1 yi
si () = g(xi )x0i
G(xi ) 1 G(xi )
or, defining ui yi E(yi |xi ) = yi G(xi o ),
g(xi )
si () = x0 ui .
G(xi )[1 G(xi )] i
Hessian:

yi gi (1 yi )gi 0 yi 1 yi 0
Hi () = + gi xi xi + g (xi )x0i xi ,
G2i [1 Gi ]2 Gi 1 Gi
where Gi G(xi ) and gi g(xi ).
Conditional expectation of the Hessian:

g(xi o )2
A(xi , o ) = E[Hi ( o )|x] = x0 x i .
G(xi o )[1 G(xi o )] i
Estimator of the asymptotic variance:

" N
#1
X g(x i

) 2
= N 1
V x0i xi .
G(x i

)[1 G(x i

)]
i=1
Asymptotic standard errors: Take the square roots of the elements on the main diagonal
of
= Avar(
V [ ) = V/N.

5.3 Probit
Probit model: standard normal distribution for ei ,
Z x
Pr(y = 1|x) = (x) = (t)dt

where () and () are the cdf and pdf, respectively, of the standard normal distribution.
FOC:
N N
X
=
X (xi ) =0
si () x0i [yi (xi )]
(xi )]
(xi )[1
i=1 i=1
15
of
" N #1
X (x i
2
)
Avar(
[ ) = x0i xi .

i=1 (xi )[1 (xi )]
5.4 Logit
Logit model: logistic distribution for ei ,
exp(x)
Pr(y = 1|x) = (x) = ,
1 + exp(x)
where () is the cdf of a standard logistic distribution with pdf
exp(z)
(z) = = (z)[1 (z)].
[1 + exp(z)]2
FOC:
N
X N
X
=
si () = 0.
x0i [yi (xi )]
i=1 i=1
of
" N #1
X
Avar(
[ ) = 0 xi
(xi )x .
i
i=1
5.5 Partial Effects

Partial effect of a continuous variable:
E(yi |xi )
= g(xi )k .
xi,k
Partial effect of a discrete variable: The partial effect of a dummy variable, xi,P , i.e., the
effect a change in xi,P from 0 to 1 has on Pr(yi = 1|xi ), is
Pi = Pr(yi = 1|xi,P = 1) Pr(yi = 1|xi,P = 0).
where the xi,1 , . . . , xi,P 1 are as observed. The probabilities are computed as follows:
Pi = G([xi,1 , . . . , xi,P 1 , 1]) G([xi,1 , . . . , xi,P 1 , 0]).
16
Partial effect of the average (PEA): For a continuous explanatory variable xk , this is in
population
P EA = g (E[xi ]) k
which is estimated as

P[EA = g x k .
Average partial effect (APE): For a continuous explanatory variable xk , this is in popu-
lation
AP E = E [g(xi )] k
which is estimated as
N
X
AP
[ E = N 1 g(xi )k .
i=1
PEA and APE for discrete variables: For a discrete explanatory variable xP , one com-
putes
P[
EA = G([ G([
x1 , . . . , xP 1 , 1])
x1 , . . . , xP 1 , 0])
N h
X i
AP
[ E = N 1 G([xi,1 , . . . , xi,P 1 , 0])
G([xi,1 , . . . , xi,P 1 , 1]) .
i=1
17
Part B. Econometrics for Stationary Time

Series Processes
6 Stationary Time Series Regression
6.1 Properties of Stochastic Time-Series Processes
Autocovariance of order k:
Cov(yt , ytk ) = E{[yt E(yt )][ytk E(ytk )]}
Autocorrelation of order k:
E{[yt E(yt )][ytk E(ytk )]}
Corr(yt , ytk ) = p p .
Var(yt ) Var(ytk )
Cross covariance of order k:

Cov(xt , ytk ) = E{[xt E(xt )][ytk E(ytk )]}.
Cross correlation of order k:

E{[xt E(xt )][ytk E(ytk )]}
Corr(xt , ytk ) = p p .
Var(xt ) Var(ytk )
Weak stationarity: A time series process {yt } is called weakly stationary if the first and
second moments are time-invariant and finite, i.e.,
E(yt ) = < t
Var(yt ) = 2 < t
Cov(yt , ytk ) = k < t
Strong stationarity: A time series process {yt } is called strongly stationary if the joint
probability distribution of any set of k observations in the sequence [yt , yt+1 , . . . , yt+k1 ] is the
same regardless of the origin, t, in the time scale.
Ergodicity: A strongly stationary time-series process, {yt }, is ergodic if for any two bounded
functions that map vectors in the a and b dimensional real vector spaces to real scalars, f :
Ra R1 and g : Rb R1 ,
lim |E[f (yt , . . . , yt+a1 )g(yt+k , . . . , yt+k+b1 )]|
k
= |E[f (yt , . . . , yt+a1 )]| |E[g(yt+k , . . . , yt+k+b1 )]|
18
Ergodic Theorems for scalar processes: If {yt } is a time-series process that is strongly
stationary and ergodic and E[yt ] = exists and is a finite constant, then
T
a.s.
X
1
yT = T yt .
t=1
If in addition E[yt2 ] = m exists and is finite then

T
a.s.
X
1
T yt2 m.
t=1
Ergodic Theorems for vector processes: If {yt } is a K 1 vector of strongly stationary

and ergodic processes with existing and finite moments E[yt ] = and E[yt yt0 ] = M, then
T
a.s.
X
1
T yt
t=1
and
T
a.s.
X
1
T yt yt0 M.
t=1
Martingale sequence: A vector sequence zt is a martingale sequence if E[zt |zt1 , zt2 , . . .] =

zt1 .
Martingale difference sequence: A vector sequence zt is a martingale difference sequence

if E[zt |zt1 , zt2 , . . .] = 0.
Stationary linear process: A stationary linear process with mean zero is defined as
ut = 0 t + 1 t1 + 2 t2 + = (L)t ,
where t is iid white noise with mean zero and variance 2 > 0, if

X
(C1) j|j | <
j=0
(C2) (1) = 0 + 1 + 2 + =
6 0
Then ut has mean E(ut ) = 0, variance Var(ut ) = 2 (02 + 12 + 22 + ) and long-run variance

2 lim Var( T uT ) = 2 [(1)]2 > 0.
T
19
6.2 Asymptotic Properties of Time Series Regressions

Population model and OLS assuptions: The population model with K regressors is
y t = xt + u t , t = 1, . . . , T,
where {[x, u]} is a jointly stationary and ergodic process with finite first and second moments
and the usual OLS assumptions E(x0t ut ) = 0 and rank[E(x0t xt )] = K hold.
Consistency: The OLS estimator of the population model is consistent,

p
.
Martingale Difference Central Limit Theorem: If zt is a vector valued stationary and

ergodic martingale difference sequence, with E[zt z0t ] = , where is a finite positive definite
matrix, then
T
X d
T zT = T 1/2 zt Normal(0, ).
t=1
Asymptotic distribution of the OLS estimator when ut is white noise: By the mar-
tingale difference CLT,
T
d
X
1/2
T x0t ut Normal(0, B),
t=1
where B E(u2t x0t xt ). From this follows

d
T ( ) = Normal(0, V),
where V = A1 BA1 and A E(x0t xt ).
Estimation of the variance of the OLS estimator when ut is white noise: Using the
OLS residuals ut , consistent estimators are
T
X
= T 1
A x0t xt
t=1
and
T
X
= T 1
B u2t x0t xt
t=1
such that
=A
V 1 B
A 1 .
20
Gordins Central Limit Theorem: If {zt } is a stationary and ergodic stochastic process
of dimension K 1 that satisfies the following conditions:
(1) Asymptotic uncorrelatedness: E[zt |ztk , ztk1 , . . .] converges in mean square to zero as
k .
(2) Summability of autocovariances: the asymptotic variance is finite, where

X

X X
lim Var( T zT ) = Cov(zt , zs ) = Cov(zt , ztk ) = .
T
t=0 s=0 k=
(3) Asymptotic negligibility of innovations: information eventually becomes negligible as it

fades far back in time from the current observation.
Then
T
X d
T zT = T 1/2 zt Normal(0, ).
t=1
Asymptotic distribution of the OLS estimator when ut is temporally dependent:

If x0t ut satisfies the conditions of Gordins CLT,
T
d
X
1/2
T x0t ut Normal(0, B),
t=1
P
where B k= Cov(x0t ut , x0tk utk ). From this follows
d
T ( ) = Normal(0, V),
where V = A1 BA1 and A E(x0t xt ).
Estimation of the variance of the OLS estimator when ut is temporally dependent:

A consistent estimate is
=A
V 1 B
A 1 ,
where
T
p
X
= T 1
A x0t xt A.
t=1
and
q T
0 ),
X X
=
B 0 + k +
( with k = T 1
ut utk x0t xtk .
k
k=1 t=k+1
21
Newey-West estimator: A positive semidefinite estimator of B is

q
=
0 +
X k 0 ).
k +
B 1 ( k
k=1
q+1
7 Autoregressive Models
7.1 Properties of the Autoregressive Model of Order 1
Autoregressive model of order 1 without constant:
iid
ut = ut1 + et , et (0, 2 )
Moving average representation:

X
ut = et + et1 + 2 et2 + = i eti
i=0
Conditional moments:
E(ut |ut1 ) = ut1
Var(ut |ut1 ) = 2
Unconditional moments: The mean is

E(ut ) = 0
and, if || < 1, the second moments are
1
Var(ut ) = 2
1 2
k
Cov(ut , utk ) = 2
1 2
Corr(ut , utk ) = k .
7.2 Estimation of the Autoregressive Model of Order 1

Consistency of OLS when the disturbance is white noise: The OLS estimator of the
AR(1) model without constant,
T
!1 T
X X
= u2t1 ut1 ut ,
t=2 t=2
is consistent if et is white noise and thus E(ut1 et ) = 0.
22
Asymptotic normality of OLS when the disturbance is white noise: By the martin-
gale difference CLT,
d
) Normal(0, V).
T (
If the disturbance et is homoscedastic, the asymptotic covariance can be estimated as
T
!1 T
!1
X X
V = e2 T 1 2
yt1 Avar(
[ ) = e2 2
yt1 .
t=2 t=2
Inconsistency of OLS when the disturbance is autocorrelated: If the disturbance et

is autocorrelated, the OLS estimator of the AR(1) model is inconsistent.
Asymptotic bias of OLS when the disturbance follows an AR(1) process: If et

follows itself an AR(1) process,
et = et1 + t ,
where t is iid white noise, the asymptotic bias of the OLS estimator of is
+
plim = .
1 +
7.3 The Autoregressive Model of Order p

Lag operator: The lag operator is defined as Lyt = yt1 and thus
Lp yt = ytp , p {. . . , 2, 1, 0, 1, 2, . . .}.
Lag polynomial: A polynomial in L is called a lag polynomial. The lag polynomial

a(L) a0 + a1 L + . . . + ap Lp
can be applied to a time series variable yt to yield
a(L)yt = a0 yt + a1 yt1 + . . . + ap ytp .
Rules for lag polynomials: Lag polynomials can be multiplied. For example, define a(L) =
a0 + a1 L and b(L) = b0 + b1 L + b2 L2 , then
a(L)b(L) = a0 b0 + (a0 b1 + a1 b0 )L + (a0 b2 + a1 b1 )L2 + a1 b2 L3 .
Lag polynomials can be evaluated at 1,
p
X
a(1) = a0 + a1 + . . . + ap = ai .
i=0
A lag polynomial applied to a time invariant quantity yields

a(L) = a0 + a1 L + . . . + ap Lp = a0 + a1 + . . . + ap = a(1).
23
The AR(p) model: The autoregressive model of order p is
yt = + a1 yt1 + . . . + ap ytp + et ,
or
a(L)yt = + et ,
where the disturbance process et is iid white noise with mean zero.
OLS Estimation of the AR(p) model: The OLS estimator

a
1

.. = (Z0 Z)1 Z0 Y.
.
a
p
is based on the observation matrices

yp+1 1 yp y1
.. .. .. .. .
Y = . , Z = . . .
yT 1 yT 1 yT p
If the disturbance is white noise, the OLS estimator is consistent and asymptotically normal.
8 Dynamic Regression
8.1 Autoregressive Distributed Lag Model
Autoregressive distributed lag (ADL) model:
yt = + a1 yt1 + a2 yt2 + . . . + ap ytp + b0 xt + b1 xt1 + . . . + bq xtq + t
or, using lag polynomials,
a(L)yt = + b(L)xt + t .
OLS estimation of the ADL model: If E[t |yt1 , yt2 , . . . , xt , xt1 , xt2 , . . .] = 0, the joint
process {(yt , xt )} is stationary and ergodic, and there is no perfect multicollinearity among the
regressors 1, yt1 , . . . , ytp , xt , . . . , xtq , then the OLS estimator of the ADL model is consistent
and asymptotically normally distributed.
24
Interpretation of the ADL coefficients as partial effects:

E(yt |1, yt1 , . . . , ytp , xt , . . . , xtq )
= ai i [1, . . . , p]
yti
E(yt |1, yt1 , . . . , ytp , xt , . . . , xtq )
= bi i [0, . . . , q]
xti
Long-run effect of a permanent shift in x on y:

E(y|x) b(1) b0 + b1 + . . . + bq
= =
x a(1) 1 a1 . . . ap
8.2 Error-Correction Model

Error-correction model (ECM): The ECM is a reparameterization of the ADL model:
yt = (yt1 xt1 ) + a p1 ytp+1 + b0 xt + . . . + bq1 xtq+1 + t

1 yt1 + . . . + a
where = 1 L and
= a(1)
= b(1)/a(1)
p
X
i =
a ai , for i = 1, . . . , p 1
k=i+1
b0 = b0
q
X
bi = bi , for i = 1, . . . , q 1
k=i+1
Estimation of the ECM by OLS: Writing the ECM as
yt = + () yt1 + xt1 + lagged differences + t ,

| {z } |{z}
1 2
it can be estimated by OLS. The OLS estimator is, under the conditions stated for the ADL
model, consistent and asymptotically normal.
Estimation of long-run parameters of the ECM: Calculate

2
= .
1
The asymptotic standard errors of can be obtained using the delta method.
25
9 Tests of Autocorrelation and Lag Order Selection

Durbin-Watson test: Test of the null hypothesis that the regression disturbance t is not
autocorrelated against the alternative that it is AR(1). Test statistic based on an estimation
sample t = 1, . . . , T :
PT
t=2 (t t1 )2
DW = PT 2 .
t=1
t
LM test for autocorrelation (Breusch-Godfrey test): In the model
y t = xt + u t , ut = 1 ut1 + . . . + r utr + t , t iid
it tests H0 : 1 = . . . = r = 0 against H1 : H0 . The test statistic is
0 PZ

LM = T 0 ,

where = Y X are the residuals obtained under H0 , Z = [X,
1 , . . . ,
r ] and PZ =
d
Z(Z0 Z)1 Z0 . Under H0 , LM 2r .
Information criteria: Let k be the number of parameters in the ADL model. The following
information criteria can be used:
2k
2) +
AIC(k) = log( ,
T
2 2k 2k 2 + 2k
AICc(k) = log(
)+ + ,
T T k1
2k log log T
HQ(k) = log(2) + ,
T
k log T
2) +
BIC(k) = log( .
T
26
Part C. Econometrics for Integrated I(1)

Time Series Processes
10 Stochastic Trends
10.1 Integrated Processes
A process {yt } is called integrated of order d (I(d)) if it is nonstationary and differencing it at
least d times is necessary to obtain a stationary process. In particular, a nonstationary process
is I(1) if its first difference, yt = (1 L)yt = yt yt1 is I(0).
10.2 Stochastic Trends with iid Increments (Random Walks)

Stochastic trend without drift: Let {t } be a iid white noise process with mean zero and
finite variance 2 . A driftless stochastic trend with iid increments is defined as
t
X
s t = s 0 + 1 + + t = s 0 + t .
t=1
Assuming s0 = 0, it has mean E(st ) = 0 and variance Var(st ) = t 2 .
Stochastic trend with drift: Let {t } be a iid white noise process with mean zero and
finite variance 2 . A stochastic trend with drift and iid increments is defined as
t
X
st = s0 + t + 1 + + t = s0 + ( + t ).
t=1
Assuming s0 = 0, it has mean E(st ) = t and variance Var(st ) = t 2 .
10.3 Stochastic Trends with Autocorrelated Increments

Stochastic trend without drift: Let {ut } be a stationary linear process with mean zero
and long-run variance 2 . A driftless stochastic trend with autocorrelated increments is defined
as
t
X
st = s0 + u1 + + ut = s0 + ut .
t=1

Assuming s0 =0 and normalizing by t, it has mean E(
st / t) = 0 and asymptotic variance
st / t) = 2 .
limt Var(
27
Stochastic trend with drift: Let {ut } be a stationary linear process with mean zero and
long-run variance 2 . A stochastic trend with drift and autocorrelated increments is defined as
t
X
st = s0 + t + u1 + + ut = s0 + ( + ut ).
t=1

Assuming s0 = 0 and normalizing by t, it has mean E(
st / t) = t and asymptotic variance
limt Var(st / t) = 2 .
10.4 Beveridge-Nelson Decomposition

st is a linear I(0) process with long-run variance 2 ,
Let st be I(1) with drift so that
t
X
st = s0 + t + ui
st = + ut = + (L)t ,
i=1
where Var(t ) = 2 . Then st can be linearly decomposed into (a) a linear deterministic trend,
(b) a random walk, (c) a stationary process , and (d) an initial condition z0 = s0 0 :
st = t + (/) s + t + z0 .
|{z} | {z }t
|{z} |{z}
divergence at rate t divergence at rate t bounded in probability bounded in probability
11 Unit Root Asymptotics

Standard Brownian motion (Wiener process): The standard Brownian motion W (t),
t [0, 1] is defined as a stochastic process with the following properties.
Initialization: W (0) = 0.
Normality: W (t) is distributed as N (0, t).
Independent increments: For given points in time 0 t1 < t2 < . . . tk 1, the increments
(W (t2 ) W (t1 )), (W (t3 ) W (t2 )), . . . , (W (tk ) W (tk1 )) are stochastically independent.
Any increment W (s) W (t), s > t, is normally distributed with mean zero and variance
s t.
The process is continuous in t.
28
Functional central limit theorem: Let 1 , . . . , T be iid white noise random variables with
mean zero and and variance 2 . Furthermore define [T r], 0 r 1, as the largest integer
smaller or equal to T r. Then the step-function
[T r]
1 X
XT (r) = t , 0 r 1
T t=1
with XT (r) = 0 for [T r] < 1 converges weakly towards a standard Brownian motion, XT (r)
W (r), as T .
Continuous mapping theorem: By the continuous mapping theorem, if XT (r) W (r)

then f (XT (r)) f (W (r)) for continuous functions f .
Convergence of sample moments of random walks without drift: Let 1 , . . . , T be

iid white noise random variables with mean zero and and variance 2 and define the random
walk without drift yt = yt1 + t , y0 = 0. Then the following convergence results hold:
R1
(1) T 3/2 Tt=1 yt1
P
0 W (r)dr
PT R1
(2) T 2 t=1
2
yt1 2 0
(W (r))2 dr
PT
(3) T 1 t=1 yt yt1 0.5 2 (W (1)2 1)
Convergence of sample moments of a stochastic trend with correlated increments:

Let {ut } be a linear I(0) process with mean zero, finite variance 0 and finite long-run variance
2 > 0 and define the driftless stochastic trend yt = yt1 + ut , ut = (L)t , y0 = 0. Then the
following convergence results hold:
R1
(1) T 3/2 Tt=1 yt1
P
0 W (r)dr
PT R1
(2) T 2 2t1
t=1 y 2 0
(W (r))2 dr
PT
(3) T 1 t=1 y t yt1 0.52 W (1)2 0.50

12 Unit Root Tests

12.1 Dickey-Fuller (DF) Tests
Dickey-Fuller test without intercept and trend: Estimate by OLS the population model
yt = yt1 + t ,
29
where t is iid white noise with variance 2 and compute the test statistics
PT !
y y
t t1
DF- = T ( 1) = T Pt=1
T 2
1
t=1 yt1
1 1
DF-t = = PT 2 .
SE( ) ( t=1 yt1 )1/2
Under the null hypothesis = 1, the test statistics have limiting distribution
0.5(W (1)2 1)
DF- DF R 1
0
(W (r))2 dr
0.5(W (1)2 1)
DF-t DFt qR .
1 2 dr
0
(W (r))
Dickey-Fuller test with intercept: Estimate by OLS the population model
yt = + yt1 + t ,
DF- = T ( 1)
1
DF-t = .
SE( )
0.5([W (1)]2 W (0)2 1)

DF- DF R1
0
(W (r))2 dr
0.5([W (1)]2 W (0)2 1)
DF-t DFt qR .
1 (r))2 dr
0
(W
Dickey-Fuller test with intercept and trend: Estimate by OLS the population model
yt = 0 + 1 t + yt1 + t ,
DF- = T ( 1)
1
DF-t = .
SE( )
30
0.5([W (1)]2 W (0)2 1)

DF- DF R1
0
(W (r))2 dr
0.5([W (1)]2 W (0)2 1)
DF-t DFt qR .
1 2
0
(W (r)) dr
12.2 Phillips-Perron (PP) Test

Phillips-Perron test without intercept and trend: Estimate the DF-regression
yt = yt1 + ut ,
where ut is potentially autocorrelated. In addition, estimate the autocovariances of ut , k , up

to order q, from which an estimate of the long-run variance, 2 , is computed. Then, under the
null of = 1,
1) 0.5 T 2 (SE(
))2 /
u2 (
2
PP- T ( u2 )
converges to DF and
s
u2 1
2
u2
PP-t 0.5 (T SE(
)/
u )
2 SE(
)

converges to DFt .
12.3 Augmented Dickey-Fuller (ADF) Tests

Augmented Dickey-Fuller test without intercept and trend: Estimate by OLS the
augmented model
y t1 + + p
yt = yt1 + 1 y tp + vt
and compute the t-statistic

1
ADF-t = .
SE()
Under the null hypothesis = 1, the ADF-t statistic has limiting distribution
ADF-t DFt .
31
Augmented Dickey-Fuller test with intercept: Estimate by OLS the augmented model
y t1 + + p
yt = + yt1 + 1 y tp + vt

1
ADF-t = .
SE( )
ADF-t DFt .
Augmented Dickey-Fuller test with intercept and trend: Estimate by OLS the aug-
mented model
y t1 + + p
yt = 0 + 1 t + yt1 + 1 y tp + vt

1
ADF-t = .
SE( )
ADF-t DFt .
Choice of the lag order: Information criteria such as the AIC or BIC can be used to
determine the lag order p. The maximum lag order pmax in this search can be chosen as the
integer part of 12 (T /100)1/4 .
13 Cointegration
13.1 Multivariate Beveridge-Nelson decomposition
Linear vector I(0) process: A linear n dimensional zero-mean vector I(0) process is
ut = t + 1 t1 + 2 t2 + . . . = (L)t , 0 = I,
where t is iid with mean zero and positive definite variance matrix , if
(C1) the sequence I, 1 , 2 , . . . is one-summable so that (1) = I + 1 + 2 + is finite,

(C2) at least one element of (1) is nonzero.
32
Long-run variance matrix: The long-run variance matrix of a linear vector I(0) process is

T ) = (1)(1)0 .
lim Var( T u
T
Vector I(1) process: A vector I(1) process of dimension n is
yt = + ut = + (L)t ,
where ut = (L)t is a linear zero-mean vector I(0) process.
Multivariate Beveridge-Nelson decomposition: A vector I(1) process of dimension n

can be decomposed into
t
X
yt = t + (1) i + t + z 0 ,
i=1
where t is a deterministic trend, (1)[1 + + t ] is a stochastic trend, t is a linear vector

I(0) process, and z0 is an initial condition.
13.2 Cointegration and common stochastic trends

Cointegration: Let yt be an n-dimensional vector I(1) process with stochastic increment ut
that is a linear vector I(0) process. Then yt is cointegrated with cointegration vector 1 if
1 6= 0 and 01 yt is stationary. (In a strict sense, stationarity of 01 yt requires a suitable choice
of initial conditions.)
Cointegration rank: The cointegration rank is the number, r, of linearly independent cointe-
gration vectors 1 , . . . , r , and the cointegration space is the space spanned by the cointegration
vectors.
Cointegration matrix: The (n r) cointegration matrix = [ 1 , . . . , r ] is the matrix of

all cointegration vectors. The cointegration matrix has rank r.
Rank of (1): The matrix (1) of the multivariate Beveridge-Nelson decomposition has
rank k n r.
Common stochastic trends: A cointegrated n-dimensional vector I(1) process with coin-
tegration rank r is driven by k = n r common stochastic trends.
33
13.3 Engle-Granger test of the null of no cointegration

Spurious regression: Suppose yt is an n-dimensional vector I(1) process and not cointe-
grated. Then the regression
y1,t = + 2 y2,t + + 2 y2,t + zt
is called spurious. The disturbance zt is I(1) and the OLS estimator has a non-standard
asymptotic distribution.
Cointegrating regression: Suppose yt is an n-dimensional vector I(1) process and cointe-

grated with cointegration rank r = 1. Then the regression
y1,t = + 2 y2,t + + 2 y2,t + zt ,
is called cointegrating regression. The disturbance zt is I(0). The OLS estimator is supercon-
sistent but has a non-standard asymptotic distribution.
Engle-Granger test: Suppose yt is an n-dimensional vector I(1) and either cointegrated

with cointegration rank r = 1 or not cointegrated. The hypotheses H0 : r = 0 versus H1 : r = 1
can be tested as follows.
Stage 1. Estimate by OLS the regression
+ 2 y2,t + + n yn,t + zt .
y1,t =
Stage 2. Estimate the ADF regression without intercept and trend on the first-stage
residuals
zt =
zt1 + 1
zt1 + + p
ztp + vt .
Stage 3. Construct the ADF statistic

1
ADF-t =
SE()
and compare it with the appropriate critical values.
13.4 The Vector Error Correction Model

Vector autoregressive (VAR) model: The n-dimensional VAR(p) model is
yt = + t + A1 yt1 + A2 yt2 + + Ap ytp + t ,
where the disturbances are iid with mean zero and variance matrix .
34
Vector error correction model (VECM): An n-dimensional VAR(p) model can be repa-
rameterized as the VECM
yt = + t + yt1 + 1 yt1 + + p1 ytp+1 + t ,
where = In + pj=1 Aj and i = pj=i+1 Aj .

P P
Reduced rank of : Suppose yt is an n-dimensional vector I(1) process and cointegrated

with cointegration rank r. Then its VECM matrix has reduced rank r and can be decomposed
into two n r matrices and of full column rank r such that = 0 .
Identification of the cointegration vectors: The decomposition = 0 is not unique.

To achieve identification, 1 normalization and at least r 1 restrictions per cointegration vector
have to be imposed.
Decomposition of the deterministic components: The trend parameters and the inter-
cept of the VECM can be decomposed into
= 1 + 2
= 1 + 2
where is the n k orthogonal complement to
Five models of the deterministic part: Based on how the parameters of the deterministic
part are restricted, five models emerge:
(1) 2 , 1 , 1 , 2 unrestricted: quadratic trend in the levels.
(2) 2 = 0; 1 , 1 , 2 unrestricted: linear trend in the cointegration relationships and in the
levels.
(3) 2 = 0, 1 = 0; 1 , 2 unrestricted: intercept in first differences, linear trend in the levels.
(4) 2 = 0, 1 = 0, 2 = 0; 1 unrestricted: intercept in the cointegration relationships and
in the levels.
(5) 2 = 0, 1 = 0, 2 = 0, 1 = 0: zero mean and no trend in the levels.
13.5 Maximum likelihood estimation of the VECM

The likelihood function in summation form: Assume the VECM disturbances t are
iid with mean zero and variance matrix , and follow a normal distribution. Then the log
likelihood function is, up to a constant,
T
T 1X
L = log | | (yt 0 yt1 zt1 )0 1 0
(yt yt1 zt1 ),
2 2 t=1
0 0
where zt1 = [yt1 , . . . , ytp+1 ]0 .
35
Observation matrices: The observations for all variables can be stacked into the matrices
Y = [y1 , . . . , yT ], Y1 = [y0 , . . . , yT 1 ], Z1 = [z0 , . . . , zT 1 ], E = [1 , . . . , T ].
The likelihood function in observation matrix form: In terms of observation matrices,

the log likelihood function can be rewritten as
T 1 0 0
log | | tr 1 0

L= (Y Y1 Z1 )(Y Y1 Z1 ) ,
2 2
where tr denotes the trace of a matrix.
Concentrating the likelihood function with respect to : Defining the auxiliary re-
gressions Y = 0 Z1 + R0 and Y1 = 1 Z1 + R1 , the concentrated likelihood function
is
T 1 0 0
log | | tr 1 0

L= (R0 R1 )(R0 R1 ) .
2 2
Concentrating the likelihood function with respect to and : Defining the moment
matrices S00 = R0 R00 /T , S01 = R0 R01 /T , S10 = R1 R00 /T , and S11 = R1 R01 /T , the concentrated
likelihood function is
T ()| nT = T log |S00 S01 ( 0 S11 )1 0 S10 | nT .
L= log |
2 2 2 2
Maximizing the concentrated likelihood function: The concentrated likelihood function

is maximized by solving the generalized eigenvalue problem
|S11 S10 S1
00 S01 | = 0
i and eigenvectors v
for eigenvalues 1 > >
i , i = 1, . . . , n. Sorting the eigenvalues as 1 >
n > 0, the ML estimator for is, given cointegration rank r, constructed from the eigenvectors
that are associated with the r largest eigenvalues,
= ( r ).
v1 , . . . , v
Maximized log likelihood function: The maximized log likelihood function, given cointe-
gration rank r, is
" r
#
T X
i ) nT .
L(r) = log |S00 | + log(1
2 i=1
2
36
Estimators of , , and : Given cointegration rank r, the remaining estimators are

0 S11 )
= S01 (
1
= S00 S01 (
1 0 S10
0 S11 )
1 1 0 Y1 Z01 (Z1 Z01 )1 .

= YZ0 (Z1 Z0 )1
13.6 Determining the Cointegration Rank

Trace test: To test the hypotheses H0 : r = r0 against H1 : r = n, use the trace statistic
n
X
LR = T i ).
log(1
i=r0 +1
The asymptotic distribution of the trace statistic is a functional of Wiener processes and tab-
ulated below.
Lambda-max test: To test the hypotheses H0 : r = r0 against H1 : r = r0 + 1, use the

lambda-max statistic
r +1 ).
LR = T log(1 0
The asymptotic distribution of the lambda-max statistic is a functional of Wiener processes

and tabulated below.
13.7 Specification of the VECM

Information criteria: To choose the lag order p of the VAR, one of the following information
criteria can be used
e e (p)| + 2 pn2 ,
AIC(m) = ln |
T
2 4 2
e e (p)| + 2 pn2 + 2p n + 2pn ,
AICc(m) = ln |
T T pn2 1
HQ(m) = ln |e e (p)| + 2 ln ln T pn2 ,

T
e e (p)| + ln T pn2 .
BIC(m) = ln |
T
37
Multivariate Breusch-Godfrey LM test: In the model of the VAR disturbances
t = Rs ts + t
it tests H0 : Rs = 0 against H1 : H0 . Based on the auxiliary regression
p ytp + Cs
t = B1 yt1 + + B ts + t
the LM statistic is
LM = (T np n 0.5)(log | | log | |),
= P 0 = P 0t /T . The LM statistic follows asymptotically a 2

where t t t /T and t t
distribution with n2 degrees of freedom.
38
Part D. Tables
Percentiles of the 2distribution
F 2 0.0100 0.0250 0.0500 0.1000 0.9000 0.9500 0.9750 0.9900
r=1 0.0002 0.0010 0.0039 0.0158 2.7055 3.8415 5.0239 6.6349
2 0.0201 0.0506 0.1026 0.2107 4.6052 5.9915 7.3778 9.2103
3 0.1148 0.2158 0.3518 0.5844 6.2514 7.8147 9.3484 11.3449
4 0.2971 0.4844 0.7107 1.0636 7.7794 9.4877 11.1433 13.2767
5 0.5543 0.8312 1.1455 1.6103 9.2364 11.0705 12.8325 15.0863
6 0.8721 1.2373 1.6354 2.2041 10.6446 12.5916 14.4494 16.8119
7 1.2390 1.6899 2.1673 2.8331 12.0170 14.0671 16.0128 18.4753
8 1.6465 2.1797 2.7326 3.4895 13.3616 15.5073 17.5345 20.0902
9 2.0879 2.7004 3.3251 4.1682 14.6837 16.9190 19.0228 21.6660
10 2.5582 3.2470 3.9403 4.8652 15.9872 18.3070 20.4832 23.2093
11 3.0535 3.8157 4.5748 5.5778 17.2750 19.6751 21.9200 24.7250
12 3.5706 4.4038 5.2260 6.3038 18.5493 21.0261 23.3367 26.2170
13 4.1069 5.0088 5.8919 7.0415 19.8119 22.3620 24.7356 27.6882
14 4.6604 5.6287 6.5706 7.7895 21.0641 23.6848 26.1189 29.1412
15 5.2293 6.2621 7.2609 8.5468 22.3071 24.9958 27.4884 30.5779
16 5.8122 6.9077 7.9616 9.3122 23.5418 26.2962 28.8454 31.9999
17 6.4078 7.5642 8.6718 10.0852 24.7690 27.5871 30.1910 33.4087
18 7.0149 8.2307 9.3905 10.8649 25.9894 28.8693 31.5264 34.8053
19 7.6327 8.9065 10.1170 11.6509 27.2036 30.1435 32.8523 36.1909
20 8.2604 9.5908 10.8508 12.4426 28.4120 31.4104 34.1696 37.5662
21 8.8972 10.2829 11.5913 13.2396 29.6151 32.6706 35.4789 38.9322
22 9.5425 10.9823 12.3380 14.0415 30.8133 33.9244 36.7807 40.2894
23 10.1957 11.6886 13.0905 14.8480 32.0069 35.1725 38.0756 41.6384
24 10.8564 12.4012 13.8484 15.6587 33.1962 36.4150 39.3641 42.9798
25 11.5240 13.1197 14.6114 16.4734 34.3816 37.6525 40.6465 44.3141
26 12.1981 13.8439 15.3792 17.2919 35.5632 38.8851 41.9232 45.6417
27 12.8785 14.5734 16.1514 18.1139 36.7412 40.1133 43.1945 46.9629
28 13.5647 15.3079 16.9279 18.9392 37.9159 41.3371 44.4608 48.2782
29 14.2565 16.0471 17.7084 19.7677 39.0875 42.5570 45.7223 49.5879
30 14.9535 16.7908 18.4927 20.5992 40.2560 43.7730 46.9792 50.8922
40 22.1643 24.4330 26.5093 29.0505 51.8051 55.7585 59.3417 63.6907
50 29.7067 32.3574 34.7643 37.6886 63.1671 67.5048 71.4202 76.1539
60 37.4849 40.4817 43.1880 46.4589 74.3970 79.0819 83.2977 88.3794
70 45.4417 48.7576 51.7393 55.3289 85.5270 90.5312 95.0232 100.4252
80 53.5401 57.1532 60.3915 64.2778 96.5782 101.8795 106.6286 112.3288
90 61.7541 65.6466 69.1260 73.2911 107.5650 113.1453 118.1359 124.1163
100 70.0649 74.2219 77.9295 82.3581 118.4980 124.3421 129.5612 135.8067
39
Percentiles of the tdistribution

Ft 0.9000 0.9500 0.9750 0.9900 0.9950
r=1 3.0777 6.3138 12.7062 31.8205 63.6567
2 1.8856 2.9200 4.3027 6.9646 9.9248
3 1.6377 2.3534 3.1824 4.5407 5.8409
4 1.5332 2.1318 2.7764 3.7469 4.6041
5 1.4759 2.0150 2.5706 3.3649 4.0321
6 1.4398 1.9432 2.4469 3.1427 3.7074
7 1.4149 1.8946 2.3646 2.9980 3.4995
8 1.3968 1.8595 2.3060 2.8965 3.3554
9 1.3830 1.8331 2.2622 2.8214 3.2498
10 1.3722 1.8125 2.2281 2.7638 3.1693
11 1.3634 1.7959 2.2010 2.7181 3.1058
12 1.3562 1.7823 2.1788 2.6810 3.0545
13 1.3502 1.7709 2.1604 2.6503 3.0123
14 1.3450 1.7613 2.1448 2.6245 2.9768
15 1.3406 1.7531 2.1314 2.6025 2.9467
16 1.3368 1.7459 2.1199 2.5835 2.9208
17 1.3334 1.7396 2.1098 2.5669 2.8982
18 1.3304 1.7341 2.1009 2.5524 2.8784
19 1.3277 1.7291 2.0930 2.5395 2.8609
20 1.3253 1.7247 2.0860 2.5280 2.8453
21 1.3232 1.7207 2.0796 2.5176 2.8314
22 1.3212 1.7171 2.0739 2.5083 2.8188
23 1.3195 1.7139 2.0687 2.4999 2.8073
24 1.3178 1.7109 2.0639 2.4922 2.7969
25 1.3163 1.7081 2.0595 2.4851 2.7874
26 1.3150 1.7056 2.0555 2.4786 2.7787
27 1.3137 1.7033 2.0518 2.4727 2.7707
28 1.3125 1.7011 2.0484 2.4671 2.7633
29 1.3114 1.6991 2.0452 2.4620 2.7564
30 1.3104 1.6973 2.0423 2.4573 2.7500
1.2816 1.6449 1.9600 2.3263 2.5758
40
Percentiles of the F distribution
The Table shows the values k, for which P (v k) = F (k) = 0.95 holds. (r1 = degrees of freedom of the nominator,r2 = degrees of
freedom of the denominator)
F (k) = 0.95 r1 = 1 2 3 4 5 6 7 8 9 10 20 120
r2 = 1 161.4476 199.5000 215.7073 224.5832 230.1619 233.9860 236.7684 238.8827 240.5433 241.8817 248.0131 253.2529
2 18.5128 19.0000 19.1643 19.2468 19.2964 19.3295 19.3532 19.3710 19.3848 19.3959 19.4458 19.4874
3 10.1280 9.5521 9.2766 9.1172 9.0135 8.9406 8.8867 8.8452 8.8123 8.7855 8.6602 8.5494
4 7.7086 6.9443 6.5914 6.3882 6.2561 6.1631 6.0942 6.0410 5.9988 5.9644 5.8025 5.6581
5 6.6079 5.7861 5.4095 5.1922 5.0503 4.9503 4.8759 4.8183 4.7725 4.7351 4.5581 4.3985
6 5.9874 5.1433 4.7571 4.5337 4.3874 4.2839 4.2067 4.1468 4.0990 4.0600 3.8742 3.7047
7 5.5914 4.7374 4.3468 4.1203 3.9715 3.8660 3.7870 3.7257 3.6767 3.6365 3.4445 3.2674

8 5.3177 4.4590 4.0662 3.8379 3.6875 3.5806 3.5005 3.4381 3.3881 3.3472 3.1503 2.9669
9 5.1174 4.2565 3.8625 3.6331 3.4817 3.3738 3.2927 3.2296 3.1789 3.1373 2.9365 2.7475
10 4.9646 4.1028 3.7083 3.4780 3.3258 3.2172 3.1355 3.0717 3.0204 2.9782 2.7740 2.5801
11 4.8443 3.9823 3.5874 3.3567 3.2039 3.0946 3.0123 2.9480 2.8962 2.8536 2.6464 2.4480
12 4.7472 3.8853 3.4903 3.2592 3.1059 2.9961 2.9134 2.8486 2.7964 2.7534 2.5436 2.3410
13 4.6672 3.8056 3.4105 3.1791 3.0254 2.9153 2.8321 2.7669 2.7144 2.6710 2.4589 2.2524
14 4.6001 3.7389 3.3439 3.1122 2.9582 2.8477 2.7642 2.6987 2.6458 2.6022 2.3879 2.1778
15 4.5431 3.6823 3.2874 3.0556 2.9013 2.7905 2.7066 2.6408 2.5876 2.5437 2.3275 2.1141
16 4.4940 3.6337 3.2389 3.0069 2.8524 2.7413 2.6572 2.5911 2.5377 2.4935 2.2756 2.0589
17 4.4513 3.5915 3.1968 2.9647 2.8100 2.6987 2.6143 2.5480 2.4943 2.4499 2.2304 2.0107
18 4.4139 3.5546 3.1599 2.9277 2.7729 2.6613 2.5767 2.5102 2.4563 2.4117 2.1906 1.9681
19 4.3807 3.5219 3.1274 2.8951 2.7401 2.6283 2.5435 2.4768 2.4227 2.3779 2.1555 1.9302
20 4.3512 3.4928 3.0984 2.8661 2.7109 2.5990 2.5140 2.4471 2.3928 2.3479 2.1242 1.8963
21 4.3248 3.4668 3.0725 2.8401 2.6848 2.5727 2.4876 2.4205 2.3660 2.3210 2.0960 1.8657
22 4.3009 3.4434 3.0491 2.8167 2.6613 2.5491 2.4638 2.3965 2.3419 2.2967 2.0707 1.8380
23 4.2793 3.4221 3.0280 2.7955 2.6400 2.5277 2.4422 2.3748 2.3201 2.2747 2.0476 1.8128
24 4.2597 3.4028 3.0088 2.7763 2.6207 2.5082 2.4226 2.3551 2.3002 2.2547 2.0267 1.7896
25 4.2417 3.3852 2.9912 2.7587 2.6030 2.4904 2.4047 2.3371 2.2821 2.2365 2.0075 1.7684
26 4.2252 3.3690 2.9752 2.7426 2.5868 2.4741 2.3883 2.3205 2.2655 2.2197 1.9898 1.7488
27 4.2100 3.3541 2.9604 2.7278 2.5719 2.4591 2.3732 2.3053 2.2501 2.2043 1.9736 1.7306
28 4.1960 3.3404 2.9467 2.7141 2.5581 2.4453 2.3593 2.2913 2.2360 2.1900 1.9586 1.7138
29 4.1830 3.3277 2.9340 2.7014 2.5454 2.4324 2.3463 2.2783 2.2229 2.1768 1.9446 1.6981
30 4.1709 3.3158 2.9223 2.6896 2.5336 2.4205 2.3343 2.2662 2.2107 2.1646 1.9317 1.6835
40 4.0847 3.2317 2.8387 2.6060 2.4495 2.3359 2.2490 2.1802 2.1240 2.0772 1.8389 1.5766
60 4.0012 3.1504 2.7581 2.5252 2.3683 2.2541 2.1665 2.0970 2.0401 1.9926 1.7480 1.4673
120 3.9201 3.0718 2.6802 2.4472 2.2899 2.1750 2.0868 2.0164 1.9588 1.9105 1.6587 1.3519
3.8416 2.9958 2.6050 2.3720 2.2142 2.0987 2.0097 1.9385 1.8800 1.8308 1.5706 1.2216
41
Critical values for the DF tests
Table 1: Empirical cumulative distribution of the DF- statistic under H0 : = 1

Probability that the statistic is less than entry
Sample size (T)
0.01 0.025 0.05 0.10 0.90 0.95 0.975 0.99
(a) No intercept, no trend
25 -11.8 -9.3 -7.3 -5.3 1.01 1.41 1.78 2.28
50 -12.8 -9.9 -7.7 -5.5 0.97 1.34 1.69 2.16
100 -13.3 -10.2 -7.9 -5.6 0.95 1.31 1.65 2.09
250 -13.6 -10.4 -8.0 -5.7 0.94 1.29 1.62 2.05
500 -13.7 -10.4 -8.0 -5.7 0.93 1.28 1.61 2.04
-13.8 -10.5 -8.1 -5.7 0.93 1.28 1.60 2.03
(b) Intercept, no trend

25 -17.2 -14.6 -12.5 -10.2 -0.76 0.00 0.65 1.39
50 -18.9 -15.7 -13.3 -10.7 -0.81 -0.07 0.53 1.22
100 -19.8 -16.3 -13.7 -11.0 -0.83 -0.11 0.47 1.14
250 -20.3 -16.7 -13.9 -11.1 -0.84 -0.13 0.44 1.08
500 -20.5 -16.8 -14.0 -11.2 -0.85 -0.14 0.42 1.07
-20.7 -16.9 -14.1 -11.3 -0.85 -0.14 0.41 1.05
(c) Intercept and trend

25 -22.5 -20.0 -17.9 -15.6 -3.65 -2.51 -1.53 -0.46
50 -25.8 -22.4 -19.7 -16.8 -3.71 -2.60 -1.67 -0.67
100 -27.4 -23.7 -20.6 -17.5 -3.74 -2.63 -1.74 -0.76
250 -28.5 -24.4 -21.3 -17.9 -3.76 -2.65 -1.79 -0.83
500 -28.9 -24.7 -21.5 -18.1 -3.76 -2.66 -1.80 -0.86
-29.4 -25.0 -21.7 -18.3 -3.77 -2.67 -1.81 -0.88
Source: Hayashi (2000), p. 576.
42
Table 2: Empirical cumulative distribution of the DF-t statistic under H0 : = 1

Probability that the statistic is less than entry
Sample size (T)
0.01 0.025 0.05 0.10 0.90 0.95 0.975 0.99
(a) No intercept, no trend
25 -2.65 -2.26 -1.95 -1.60 0.92 1.33 1.70 2.15
50 -2.62 -2.25 -1.95 -1.61 0.91 1.31 1.66 2.08
100 -2.60 -2.24 -1.95 -1.61 0.90 1.29 1.64 2.04
250 -2.58 -2.24 -1.95 -1.62 0.89 1.28 1.63 2.02
500 -2.58 -2.23 -1.95 -1.62 0.89 1.28 1.62 2.01
-2.58 -2.23 -1.95 -1.62 0.89 1.28 1.62 2.01
(b) Intercept, no trend

25 -3.75 -3.33 -2.99 -2.64 -0.37 0.00 0.34 0.72
50 -3.59 -3.23 -2.93 -2.60 -0.41 -0.04 0.28 0.66
100 -3.50 -3.17 -2.90 -2.59 -0.42 -0.05 0.26 0.63
250 -3.45 -3.14 -2.88 -2.58 -0.42 -0.06 0.24 0.62
500 -3.44 -3.13 -2.87 -2.57 -0.44 -0.07 0.24 0.61
-3.42 -3.12 -2.86 -2.57 -0.44 -0.08 0.23 0.60
(c) Intercept and trend

25 -4.38 -3.95 -3.60 -3.24 -1.14 -0.81 -0.50 -0.15
50 -4.15 -3.80 -3.50 -3.18 -1.19 -0.87 -0.58 -0.24
100 -4.05 -3.73 -3.45 -3.15 -1.22 -0.90 -0.62 -0.28
250 -3.98 -3.69 -3.42 -3.13 -1.23 -0.92 -0.64 -0.31
500 -3.97 -3.67 -3.42 -3.13 -1.24 -0.93 -0.65 -0.32
-3.96 -3.66 -3.41 -3.12 -1.25 -0.94 -0.66 -0.33
Source: Hayashi (2000), p. 578.
Critical values for the ADF test for cointegration

(Engle-Granger test)
Number of regressors,
1% 2.5% 5% 10%
excluding constant
A. Regressors have no drift
1 -3.96 -3.64 -3.37 -3.07
2 -4.31 -4.02 -3.77 -3.45
3 -4.73 -4.37 -4.11 -3.83
4 -5.07 -4.71 -4.45 -4.16
5 -5.28 -4.98 -4.71 -4.43
B. Some regressors have drift
1 -3.96 -3.67 -3.41 -3.13
2 -4.36 -4.07 -3.80 -3.52
3 -4.65 -4.39 -4.16 -3.84
4 -5.04 -4.77 -4.49 -4.20
5 -5.36 -5.02 -4.74 -4.46
Source: Hayashi (2000), Table 10.1, p. 646.
Critical values for the trace and lambda-max tests

Source: Osterwald-Lenum (1992)
Case 1: no deterministics
Note: the number of variables is indicated by p in the Table.

Case 2: restricted intercept Case 3: unrestricted intercept

Case 4: intercept and restricted trend Case 5: intercept and unrestricted trend


Formulary Eco II

Uploaded by

Copyright:

Available Formats

Formulary Eco II

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Formulary Eco II

Uploaded by

Copyright:

Available Formats

Formulas and Tables

April 27, 2016

Part A. Cross Section Econometrics

Uniform convergence: Let w be a random vector taking values in W RM , let be a

f (b) f (a) = f 0 (c) (b a).

f (b) f (a) = x f (c) (b a).

1.2 Vector differentiation

and the Hessian is

Example 1: Consider the K 1 vector function g() = f (), where f () is a L 1 vector

1.3 Delta Method

Now let c : RQ , Q P , be a continuously differentiable function and assume that is in

It is estimated by the M-estimator that minimizes the sample function

First order condition:

2.2 Asymptotic Properties

Bo = E[s(w, o )s(w, o )0 ] = Var[s(w, o )].

2.3 Estimators of the Asymptotic Variance

A(xi , o ) = E[H(wi , o )|x].

Take the square roots of the elements on the main

2.4 Generalized information matrix equality

E[s(w, o )s(w, o )0 ] = o2 E[H(w, o )], o2 > 0.

2.5 Nonlinear Least Squares

In terms of the M-estimator, the objective function specializes to

q(w, ) = [y m(x, )]2 /2.

NLS variance matrix: The asymptotic variance matrix is

where the matrix Ao becomes

Homoscedasticity: Under homoscedasticity, i.e., under Assumption NLS.3, Var(y|x) =

Wald test of nonlinear hypotheses: To test the Q nonlinear hypotheses H0 : c() = 0

2.7 Optimization Methods

the Newton-Raphson method uses the iteration

Generalized Gauss-Newton method: The generalized Gauss-Newton method uses the

3 Conditional Maximum Likelihood Estimation

`i () `(yi , xi , ) = log f (yi |xi ; ),

Sample log likelihood function:

First order condition:

Expectation of the Hessian:

Fisher information matrix:

Bo = Var[si ( o )] = E[si ( o )si ( o )0 ].

Conditional information matrix equality (CIME): Under fairly general conditions, in

E[Hi ( o )|xi ] = E[si ( o )si ( o )0 |xi ]

Unconditional information matrix equality (UIME): By the law of iterated expecta-

Ao = E[Hi ( o )] = E[si ( o )si ( o )0 ] = Bo .

3.2 Asymptotic Properties

3.3 Estimators of the Asymptotic Variance

(b) Using the conditional expectation:

where A(xi , o ) E[H(yi , xi , o )|xi )].

(c) Outer product of the score:

Take the square roots of the elements on the main

Wald test of nonlinear hypotheses: To test the Q nonlinear hypotheses H0 : c() = 0

is the P 1 score evaluated at the restricted estimate and A

4 Generalized Method of Moments Estimation

Generalized method of moments (GMM) estimator: The GMM estimator minimizes

is an L L symmetric, positive semidefinite weighting matrix.

First order condition:

Expected gradient of the moment condition:

Variance of the moment condition:

o = E [g(wi , o )g(wi , o )0 ] = Var [g(wi , o )] .

4.2 Asymptotic Properties

4.3 Estimation of the Variance

Take the square roots of the elements on the main