Lecture Notes for Econometrics 2002 (first year

PhD course in Stockholm)

Paul Söderlind1

June 2002 (some typos corrected and some material added later)

1 Universityof St. Gallen. Address: s/bf-HSG, Rosenbergstrasse 52, CH-9000 St. Gallen,
Switzerland. E-mail: [email protected]. Document name: EcmAll.TeX.
Contents

1 Introduction 5
1.1 Means and Standard Deviation . . . . . . . . . . . . . . . . . . . . . 5
1.2 Testing Sample Means . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Covariance and Correlation . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 The Distribution of ˇO . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.8 Testing Hypotheses about ˇO . . . . . . . . . . . . . . . . . . . . . . 14

A Practical Matters 16

B A CLT in Action 17

2 Univariate Time Series Analysis 21

2.1 Theoretical Background to Time Series Processes . . . . . . . . . . . 21
2.2 Estimation of Autocovariances . . . . . . . . . . . . . . . . . . . . . 22
2.3 White Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Moving Average . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Autoregression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 ARMA Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.7 Non-stationary Processes . . . . . . . . . . . . . . . . . . . . . . . . 36

3 The Distribution of a Sample Average 44

3.1 Variance of a Sample Average . . . . . . . . . . . . . . . . . . . . . 44
3.2 The Newey-West Estimator . . . . . . . . . . . . . . . . . . . . . . . 48

1
 3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4 Least Squares 53
4.1 Definition of the LS Estimator . . . . . . . . . . . . . . . . . . . . . 53
4.2 LS and R2  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3 Finite Sample Properties of LS . . . . . . . . . . . . . . . . . . . . . 57
4.4 Consistency of LS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5 Asymptotic Normality of LS . . . . . . . . . . . . . . . . . . . . . . 60
4.6 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.7 Diagnostic Tests of Autocorrelation, Heteroskedasticity, and Normality 66

5 Instrumental Variable Method 74

5.1 Consistency of Least Squares or Not? . . . . . . . . . . . . . . . . . 74
5.2 Reason 1 for IV: Measurement Errors . . . . . . . . . . . . . . . . . 74
5.3 Reason 2 for IV: Simultaneous Equations Bias (and Inconsistency) . . 76
5.4 Definition of the IV Estimator—Consistency of IV . . . . . . . . . . 80
5.5 Hausman’s Specification Test . . . . . . . . . . . . . . . . . . . . . 86
5.6 Tests of Overidentifying Restrictions in 2SLS . . . . . . . . . . . . 87

6 Simulating the Finite Sample Properties 89

6.1 Monte Carlo Simulations . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2 Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

7 GMM 100
7.1 Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2 Generalized Method of Moments . . . . . . . . . . . . . . . . . . . . 101
7.3 Moment Conditions in GMM . . . . . . . . . . . . . . . . . . . . . . 101
7.4 The Optimization Problem in GMM . . . . . . . . . . . . . . . . . . 104
7.5 Asymptotic Properties of GMM . . . . . . . . . . . . . . . . . . . . 108
7.6 Summary of GMM . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.7 Efficient GMM and Its Feasible Implementation . . . . . . . . . . . . 114
7.8 Testing in GMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.9 GMM with Sub-Optimal Weighting Matrix . . . . . . . . . . . . . . 117
7.10 GMM without a Loss Function . . . . . . . . . . . . . . . . . . . . 118

2
 7.11 Simulated Moments Estimator . . . . . . . . . . . . . . . . . . . . . 119

8 Examples and Applications of GMM 122

8.1 GMM and Classical Econometrics: Examples . . . . . . . . . . . . . 122
8.2 Identification of Systems of Simultaneous Equations . . . . . . . . . 132
8.3 Testing for Autocorrelation . . . . . . . . . . . . . . . . . . . . . . . 134
8.4 Estimating and Testing a Normal Distribution . . . . . . . . . . . . . 138
8.5 Testing the Implications of an RBC Model . . . . . . . . . . . . . . . 142
8.6 IV on a System of Equations . . . . . . . . . . . . . . . . . . . . . 143

12 Vector Autoregression (VAR) 145

12.1 Canonical Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
12.2 Moving Average Form and Stability . . . . . . . . . . . . . . . . . . 146
12.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
12.4 Granger Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
12.5 Forecasts Forecast Error Variance . . . . . . . . . . . . . . . . . . . 150
12.6 Forecast Error Variance Decompositions . . . . . . . . . . . . . . . 151
12.7 Structural VARs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
12.8 Cointegration and Identification via Long-Run Restrictions . . . . . 165

12 Kalman filter 172

12.1 Conditional Expectations in a Multivariate Normal Distribution . . . . 172
12.2 Kalman Recursions . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

13 Outliers and Robust Estimators 179

13.1 Influential Observations and Standardized Residuals . . . . . . . . . . 179
13.2 Recursive Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . 180
13.3 Robust Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
13.4 Multicollinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

14 Generalized Least Squares 187

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
14.2 GLS as Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . 188
14.3 GLS as a Transformed LS . . . . . . . . . . . . . . . . . . . . . . . 191
14.4 Feasible GLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

3
15 Nonparametric Regressions and Tests 193
15.1 Nonparametric Regressions . . . . . . . . . . . . . . . . . . . . . . . 193
15.2 Estimating and Testing Distributions . . . . . . . . . . . . . . . . . . 201

16 Alphas /Betas and Investor Characteristics 208

16.1 Basic Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
16.2 Calendar Time and Cross Sectional Regression . . . . . . . . . . . . 208
16.3 Panel Regressions, Driscoll-Kraay and Cluster Methods . . . . . . . . 209
16.4 From CalTime To a Panel Regression . . . . . . . . . . . . . . . . . 216
16.5 The Results in Hoechle, Schmid and Zimmermann . . . . . . . . . . 217
16.6 Monte Carlo Experiment . . . . . . . . . . . . . . . . . . . . . . . . 219
16.7 An Empirical Illustration . . . . . . . . . . . . . . . . . . . . . . . . 222

21 Some Statistics 226

21.1 Distributions and Moment Generating Functions . . . . . . . . . . . . 226
21.2 Joint and Conditional Distributions and Moments . . . . . . . . . . . 228
21.3 Convergence in Probability, Mean Square, and Distribution . . . . . . 231
21.4 Laws of Large Numbers and Central Limit Theorems . . . . . . . . . 233
21.5 Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
21.6 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
21.7 Special Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 235
21.8 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

22 Some Facts about Matrices 248

22.1 Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
22.2 Vector Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
22.3 Systems of Linear Equations and Matrix Inverses . . . . . . . . . . . 248
22.4 Complex matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
22.5 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . 252
22.6 Special Forms of Matrices . . . . . . . . . . . . . . . . . . . . . . . 252
22.7 Matrix Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . 254
22.8 Matrix Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
22.9 Miscellaneous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

4
1 Introduction

1.1 Means and Standard Deviation

The mean and variance of a series are estimated as

PT PT
xN D t D1 x t =T and O 2 D t D1 .x t N 2 =T:
x/ (1.1)

The standard deviation (here denoted Std.x t /), the square root of the variance, is the most
common measure of volatility.
The mean and standard deviation are often estimated on rolling data windows (for in-
stance, a “Bollinger band” is ˙2 standard deviations from a moving data window around
a moving average—sometimes used in analysis of financial prices.)
If x t is iid (independently and identically distributed), then it is straightforward to find
the variance of the sample average. Then, note that
P  P
T
Var x
t D1 t =T D TtD1 Var .x t =T /

D T Var .x t / =T 2
D Var .x t / =T: (1.2)

The first equality follows from the assumption that x t and xs are independently distributed
(so the covariance is zero). The second equality follows from the assumption that x t and
xs are identically distributed (so their variances are the same). The third equality is a
trivial simplification.
A sample average is (typically) unbiased, that is, the expected value of the sample
average equals the population mean. To illustrate that, consider the expected value of the
sample average of the iid x t
PT PT
E t D1 x t =T D tD1 E x t =T
D E xt : (1.3)

The first equality is always true (the expectation of a sum is the sum of expectations), and

5
 a. Distribution of sample average b. Distribution of √T times sample average
3 0.4
T=5 T=5
2 T=25 T=25
T=50 T=50
T=100 0.2 T=100
1

0 0
−2 0 2 −5 0 5
Sample average √T times sample average
Figure p
1.1: Sampling distributions. This figure shows the distribution of the sample mean
and of T times the sample mean of the random variable z t 1 where z t  2 .1/.

the second equality follows from the assumption of identical distributions which implies
identical expectations.

1.2 Testing Sample Means

The law of large numbers (LLN) says that the sample mean converges to the true popula-
tion mean as the sample size goes to infinity. This holds for a very large class of random
variables, but there are exceptions. A sufficient (but not necessary) condition for this con-
vergence is that the sample average is unbiased (as in (1.3)) and that the variance goes to
zero as the sample size goes to infinity (as in (1.2)). (This is also called convergence in
mean square.) To see the LLN in action, see Figure 1.1.
p
The central limit theorem (CLT) says that T xN converges in distribution to a normal
distribution as the sample size increases. See Figure 1.1 for an illustration. This also
holds for a large class of random variables—and it is a very useful result since it allows
us to test hypothesis. Most estimators (including LS and other methods) are effectively
some kind of sample average, so the CLT can be applied.
The basic approach in testing a hypothesis (the “null hypothesis”), is to compare the
test statistics (the sample average, say) with how the distribution of that statistics (which
is a random number since the sample is finite) would look like if the null hypothesis is
true. For instance, suppose the null hypothesis is that the population mean is  Suppose
also that we know that distribution of the sample mean is normal with a known variance
h2 (which will typically be estimated and then treated as if it was known). Under the null
hypothesis, the sample average should then be N.; h2 /. We would then reject the null

6
hypothesis if the sample average is far out in one the tails of the distribution. A traditional
two-tailed test amounts to rejecting the null hypothesis at the 10% significance level if
the test statistics is so far out that there is only 5% probability mass further out in that
tail (and another 5% in the other tail). The interpretation is that if the null hypothesis is
actually true, then there would only be a 10% chance of getting such an extreme (positive
or negative) sample average—and these 10% are considered so low that we say that the
null is probably wrong.

Density function of N(0.5,2) Density function of N(0,2)

0.4 0.4
Pr(x ≤ −1.83) = 0.05 Pr(y ≤ −2.33) = 0.05
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4
x y = x−0.5

Density function of N(0,1)

0.4
Pr(z ≤ −1.65) = 0.05
0.3

0.2

0.1

0
−4 −3 −2 −1 0 1 2 3 4
z = (x−0.5)/√2

Figure 1.2: Density function of normal distribution with shaded 5% tails.

See Figure 1.2 for some examples or normal distributions. recall that in a normal
distribution, the interval ˙1 standard deviation around the mean contains 68% of the
probability mass; ˙1:65 standard deviations contains 90%; and ˙2 standard deviations
contains 95%.
In practice, the test of a sample mean is done by “standardizing” the sampe mean so

7
that it can be compared with a standard N.0; 1/ distribution. The logic of this is as follows

Pr.xN  2:7/ D Pr.xN   2:7 / (1.4)

 
xN  2:7 
D Pr  : (1.5)
h h

If xN  N.; h2 /, then .xN /= h  N.0; 1/, so the probability of xN  2:7 can be calcu-
lated by calculating how much probability mass of the standard normal density function
there is above .2:7 /= h.
To construct a two-tailed test, we also need.the probability that xN is above some num-
ber. This number is chosen to make the two-tailed tst symmetric, that is, so that there
is as much probability mass below lower number (lower tail) as above the upper number
(upper tail). With a normal distribution (or, for that matter, any symmetric distribution)
this is done as follows. Note that .xN /= h  N.0; 1/ is symmetric around 0. This means
that the probability of being above some number, .C /= h, must equal the probability
of being below 1 times the same number, or
   
xN  C  C  xN 
Pr  D Pr  : (1.6)
h h h h

A 10% critical value is the value of .C /= h that makes both these probabilities
equal to 5%—which happens to be 1.645. The easiest way to look up such critical values
is by looking at the normal cumulative distribution function—see Figure 1.2.

1.3 Covariance and Correlation

The covariance of two variables (here x and y) is typically estimated as

b
Cov .x t ; z t / D
PT
t D1 .x t N .z t
x/ zN / =T: (1.7)

Note that this is a kind of sample average, so a CLT can be used.

The correlation of two variables is then estimated as

b
Corr .x t ; z t / D
b
Cov .x t ; z t /
; (1.8)
c .x t / Std
Std c .z t /

c t / is an estimated standard deviation. A correlation must be between 1 and 1

where Std.x

8
 Pdf of t when true β=−0.1 Pdf of t when true β=0.51
0.4 0.4

0.3 Probs: 0.15 0.01 0.3 Probs: 0.05 0.05

0.2 0.2

0.1 0.1

0 0
−4 −2 0 2 4 −4 −2 0 2 4
t t

H0: β=0.5
Pdf of t when true β=2
0.4 t = (b−0.5)/σ
If b ~ N(β,σ2), then t ~ N((β−0.5)/σ,1)
0.3 Probs: 0.00 0.44
Probabilities (power) are shown for
0.2
t ≤ −1.65 and t > 1.65 (10% critical values)
0.1

0
−4 −2 0 2 4
t

Figure 1.3: Power of two-sided test

(try to show it). Note that covariance and correlation measure the degree of linear relation
only. This is illustrated in Figure 1.4.
The pth autocovariance of x is estimated by

b
Cov x t ; x t p


D
PT
t D1 .x t N xt
x/ p


xN =T; (1.9)

where we use the same estimated (using all data) mean in both places. Similarly, the pth
autocorrelation is estimated as

b
Cov x t ; x t p
Corr x t ; x t p D
b


: (1.10)
c .x t /2
Std
Compared with a traditional estimate of a correlation (1.8) we here impose that the stan-
dard deviation of x t and x t p are the same (which typically does not make much of a
difference).

9
 Correlation of 0.9 Correlation of 0
2 2
1 1
0 0
y

y
−1 −1
−2 −2
−5 0 5 −5 0 5
x x

Correlation of −0.9 Correlation of 0

2 2
1 1
0 0
y

−1 −1
−2 −2
−5 0 5 −5 0 5
x x

Figure 1.4: Example of correlations on an artificial sample. Both subfigures use the same
sample of y.

1.4 Least Squares

Consider the simplest linear model

y t D x t ˇ0 C u t ; (1.11)

where all variables are zero mean scalars and where ˇ0 is the true value of the parameter
we want to estimate. The task is to use a sample fy t ; x t gTtD1 to estimate ˇ and to test
hypotheses about its value, for instance that ˇ D 0.
If there were no movements in the unobserved errors, u t , in (1.11), then any sample
would provide us with a perfect estimate of ˇ. With errors, any estimate of ˇ will still
leave us with some uncertainty about what the true value is. The two perhaps most impor-
tant issues in econometrics are how to construct a good estimator of ˇ and how to assess

10
the uncertainty about the true value.
O we get a fitted residual
For any possible estimate, ˇ,

uO t D y t O
x t ˇ: (1.12)

One appealing method of choosing ˇO is to minimize the part of the movements in y t that
we cannot explain by x t ˇ, O that is, to minimize the movements in uO t . There are several
candidates for how to measure the “movements,” but the most common is by the mean of
squared errors, that is, ˙ tTD1 uO 2t =T . We will later look at estimators where we instead use
˙ tTD1 juO t j =T .
With the sum or mean of squared errors as the loss function, the optimization problem
T
1X
min .y t x t ˇ/2 (1.13)
ˇ T t D1

has the first order condition that the derivative should be zero as the optimal estimate ˇO
T
1X  
xt yt x t ˇO D 0; (1.14)
T tD1

which we can solve for ˇO as

1
T T
!
1X 2 1X
ˇO D x x t y t , or (1.15)
T t D1 t T t D1
D Var
c .x t / 1 b
Cov .x t ; y t / ; (1.16)

where a hat indicates a sample estimate. This is the Least Squares (LS) estimator.

1.5 Maximum Likelihood

A different route to arrive at an estimator is to maximize the likelihood function. If u t in

(1.11) is iid N 0;  2 , then the probability density function of u t is

1
u2t = 2 2


pdf .u t / D p exp : (1.17)
2 2

11
Since the errors are independent, we get the joint pdf of the u1 ; u2 ; : : : ; uT by multiplying
the marginal pdfs of each of the errors. Then substitute y t x t ˇ for u t (the derivative of
the transformation is unity) and take logs to get the log likelihood function of the sample
T
T T 1X
ln  2 x t ˇ/2 = 2 :


ln L D ln .2/ .y t (1.18)
2 2 2 t D1

This likelihood function is maximized by minimizing the last term, which is propor-
tional to the sum of squared errors - just like in (1.13): LS is ML when the errors are iid
normally distributed.
Maximum likelihood estimators have very nice properties, provided the basic dis-
tributional assumptions are correct. If they are, then MLE are typically the most effi-
cient/precise estimators, at least asymptotically. ML also provides a coherent framework
for testing hypotheses (including the Wald, LM, and LR tests).

1.6 The Distribution of ˇO

Equation (1.15) will give different values of ˇO when we use different samples, that is
different draws of the random variables u t , x t , and y t . Since the true value, ˇ0 , is a fixed
constant, this distribution describes the uncertainty we should have about the true value
after having obtained a specific estimated value.
O use (1.11) in (1.15) to substitute for y t
To understand the distribution of ˇ,
1
T T
!
1X 2 1X
ˇO D x x t .x t ˇ0 C u t /
T t D1 t T t D1
T
! 1 T
1X 2 1X
D ˇ0 C xt xt ut ; (1.19)
T tD1 T t D1

where ˇ0 is the true value.

The first conclusion from (1.19) is that, with u t D 0 the estimate would always be
perfect — and with large movements in u t we will see large movements in ˇ. O The second
conclusion is that not even a strong opinion about the distribution of u t , for instance that
u t is iid N 0;  2 , is enough to tell us the whole story about the distribution of ˇ. O The

reason is that deviations of ˇO from ˇ0 are a function of x t u t , not just of u t . Of course,

12
when x t are a set of deterministic variables which will always be the same irrespective
of which sample we use, then ˇO ˇ0 is a time invariant linear function of u t , so the
distribution of u t carries over to the distribution of ˇ. O This is probably an unrealistic
case, which forces us to look elsewhere to understand the properties of ˇ. O
There are two main routes to learn more about the distribution of ˇ: O (i) set up a small
“experiment” in the computer and simulate the distribution or (ii) use the asymptotic
distribution as an approximation. The asymptotic distribution can often be derived, in
contrast to the exact distribution in a sample of a given size. If the actual sample is large,
then the asymptotic distribution may be a good approximation.
A law of large numbers would (in most cases) say that both TtD1 x t2 =T and TtD1 x t u t =T
P P

in (1.19) converge to their expected values as T ! 1. The reason is that both are sample
averages of random variables (clearly, both x t2 and x t u t are random variables). These ex-
pected values are Var.x t / and Cov.x t ; u t /, respectively (recall both x t and u t have zero
means). The key to show that ˇO is consistent, that is, has a probability limit equal to ˇ0 , is
that Cov.x t ; u t / D 0. This highlights the importance of using good theory to derive not
only the systematic part of (1.11), but also in understanding the properties of the errors.
For instance, when theory tells us that y t and x t affect each other (as prices and quanti-
ties typically do), then the errors are likely to be correlated with the regressors - and LS
is inconsistent. One common way to get around that is to use an instrumental variables
technique. More about that later. Consistency is a feature we want from most estimators,
since it says that we would at least get it right if we had enough data.
Suppose that ˇO is consistent. Can we say anything more about the asymptotic distri-
bution? Well, the distributionof ˇO converges to a spike with all the mass at ˇ0 , but the
p p 
O O
distribution of T ˇ, or T ˇ ˇ0 , will typically converge to a non-trivial normal
distribution. To see why, note from (1.19) that we can write

T
! 1 p T
p   1X 2 T X
T ˇO ˇ0 D x xt ut : (1.20)
T t D1 t T tD1

The first term on the right hand side will typically converge to the inverse of Var.x t /, as
p
discussed earlier. The second term is T times a sample average (of the random variable
x t u t ) with a zero expected value, since we assumed that ˇO is consistent. Under weak
p
conditions, a central limit theorem applies so T times a sample average converges to
p
a normal distribution. This shows that T ˇO has an asymptotic normal distribution. It

13
turns out that this is a property of many estimators, basically because most estimators are
some kind of sample average. For an example of a central limit theorem in action, see
Appendix B

1.7 Diagnostic Tests

p
Exactly what the variance of T .ˇO ˇ0 / is, and how it should be estimated, depends
mostly on the properties of the errors. This is one of the main reasons for diagnostic tests.
The most common tests are for homoskedastic errors (equal variances of u t and u t s ) and
no autocorrelation (no correlation of u t and u t s ).
When ML is used, it is common to investigate if the fitted errors satisfy the basic
assumptions, for instance, of normality.

1.8 Testing Hypotheses about ˇO

Suppose we now assume that the asymptotic distribution of ˇO is such that

p   d
T ˇO ˇ0 ! N 0; v 2 or


(1.21)

We could then test hypotheses about ˇO as for any other random variable. For instance,
consider the hypothesis that ˇ0 D 0. If this is true, then
p  p 
Pr O < 2 D Pr
T ˇ=v O > 2  0:025;
T ˇ=v (1.22)

which says that there is only a 2.5% chance that a random sample will deliver a value of
p
O less than -2 and also a 2.5% chance that a sample delivers a value larger than 2,
T ˇ=v
assuming the true value is zero.
We then say that we reject the hypothesis that ˇ0 D 0 at the 5% significance level
p
O
(95% confidence level) if the test statistics j T ˇ=vj is larger than 2. The idea is that,
if the hypothesis is true (ˇ0 D 0), then this decision rule gives the wrong decision in
5% of the cases. That is, 5% of all possible random samples will make us reject a true
hypothesis. Note, however, that this test can only be taken to be an approximation since it
relies on the asymptotic distribution, which is an approximation of the true (and typically
unknown) distribution.

14
 a. Pdf of N(0,1) b. Pdf of Chi−square(n)
0.4 1
n=1
n=2
n=5
0.2 0.5

0 0
−2 0 2 0 5 10
x x
Figure 1.5: Probability density functions

p
O
The natural interpretation of a really large test statistics, j T ˇ=vj D 3 say, is that
it is very unlikely that this sample could have been drawn from a distribution where the
hypothesis ˇ0 D 0 is true. We therefore choose to reject the hypothesis. We also hope that
the decision rule we use will indeed make us reject false hypothesis more often than we
reject true hypothesis. For instance, we want the decision rule discussed above to reject
ˇ0 D 0 more often when ˇ0 D 1 than when ˇ0 D 0.
There is clearly nothing sacred about the 5% significance level. It is just a matter of
convention that the 5% and 10% are the most widely used. However, it is not uncommon
to use the 1% or the 20%. Clearly, the lower the significance level, the harder it is to reject
a null hypothesis. At the 1% level it often turns out that almost no reasonable hypothesis
can be rejected.
The t-test described above works only if the null hypothesis contains a single restric-
tion. We have to use another approach whenever we want to test several restrictions
jointly. The perhaps most common approach is a Wald test. To illustrate the idea, suppose
p d
ˇ is an m  1 vector and that T ˇO ! N .0; V / under the null hypothesis , where V is a
covariance matrix. We then know that
p p d
T ˇO 0 V 1
T ˇO ! 2 .m/ : (1.23)

The decision rule is then that if the left hand side of (1.23) is larger that the 5%, say,
critical value of the 2 .m/ distribution, then we reject the hypothesis that all elements in
ˇ are zero.

15
A Practical Matters
A.0.1 Software

 Gauss, MatLab, RATS, Eviews, Stata, PC-Give, Micro-Fit, TSP, SAS

 Software reviews in The Economic Journal and Journal of Applied Econometrics

A.0.2 Useful Econometrics Literature

1. Greene (2000), Econometric Analysis (general)

2. Hayashi (2000), Econometrics (general)

3. Johnston and DiNardo (1997), Econometric Methods (general, fairly easy)

4. Pindyck and Rubinfeld (1998), Econometric Models and Economic Forecasts (gen-
eral, easy)

5. Verbeek (2004), A Guide to Modern Econometrics (general, easy, good applica-

tions)

6. Davidson and MacKinnon (1993), Estimation and Inference in Econometrics (gen-

eral, a bit advanced)

7. Ruud (2000), Introduction to Classical Econometric Theory (general, consistent

projection approach, careful)

8. Davidson (2000), Econometric Theory (econometrics/time series, LSE approach)

9. Mittelhammer, Judge, and Miller (2000), Econometric Foundations (general, ad-

vanced)

10. Patterson (2000), An Introduction to Applied Econometrics (econometrics/time se-

ries, LSE approach with applications)

11. Judge et al (1985), Theory and Practice of Econometrics (general, a bit old)

12. Hamilton (1994), Time Series Analysis

16
 13. Spanos (1986), Statistical Foundations of Econometric Modelling, Cambridge Uni-
versity Press (general econometrics, LSE approach)

14. Harvey (1981), Time Series Models, Philip Allan

15. Harvey (1989), Forecasting, Structural Time Series... (structural time series, Kalman
filter).

16. Lütkepohl (1993), Introduction to Multiple Time Series Analysis (time series, VAR
models)

17. Priestley (1981), Spectral Analysis and Time Series (advanced time series)

18. Amemiya (1985), Advanced Econometrics, (asymptotic theory, non-linear econo-

metrics)

19. Silverman (1986), Density Estimation for Statistics and Data Analysis (density es-
timation).

20. Härdle (1990), Applied Nonparametric Regression

B A CLT in Action
This is an example of how we can calculate the limiting distribution of a sample average.
p
Remark B.1 If T .xN /=  N.0; 1/ then xN  N.;  2 =T /.
p
Example B.2 (Distribution of ˙ tTD1 .z t 1/ =T and T ˙ tTD1 .z t 1/ =T when z t 
2 .1/:) When z t is iid 2 .1/, then ˙ tTD1 z t is distributed as a 2 .T / variable with pdf
T
fT ./. We now construct a new variable by transforming ˙ tD1 z t as to a sample mean
around one (the mean of z t )

zN 1 D ˙ tTD1 z t =T 1 D ˙ tTD1 .z t 1/ =T:

Clearly, the inverse function is ˙ tTD1 z t D T zN 1 C T , so by the “change of variable” rule

we get the pdf of zN 1 as
g.zN1 / D fT .T zN 1 C T / T:

17
Example B.3 Continuing the previous example, we now consider the random variable
p
zN2 D T zN1 ;
p
with inverse function zN 1 D zN2 = T . By applying the “change of variable” rule again, we
get the pdf of zN 2 as
p p p p
h .zN 2 / D g.Nz2 = T /= T D fT T zN 2 C T T:

Example B.4 When z t is iid 2 .1/, then ˙ tTD1 z t is 2 .T /, which we denote f .˙ tTD1 z t /.
T
We now construct two new variables by transforming ˙ tD1 zt

zN1 D ˙ tTD1 z t =T T
1 D ˙ tD1 .z t 1/ =T , and
p
zN2 D T zN1 :

Example B.5 We transform this distribution by first subtracting one from z t (to remove
p
the mean) and then by dividing by T or T . This gives the distributions of the sample
p
mean and scaled sample mean, zN2 D T zN1 as
1
f .zN 1 / D y T =2 1
exp . y=2/ with y D T zN 1 C T , and
2T =2 .T =2/
1 p
f .zN 2 / D y T =2 1
exp . y=2/ with y D T zN 1 C T .
2T =2 .T =2/

These distributions are shown in Figure 1.1. It is clear that f .zN 1 / converges to a spike
at zero as the sample size increases, while f .zN 2 / converges to a (non-trivial) normal
distribution.
p
Example B.6 (Distribution of ˙ tTD1 .z t 1/ =T and T ˙ tTD1 .z t 1/ =T when z t 
T
2 .1/:) When z t is iid 2 .1/, then ˙ tD1 z t is 2 .T /, that is, has the probability density
function
1
T =2 1
f ˙ tTD1 z t D ˙ tTD1 z t ˙ tTD1 z t =2 :



exp
2T =2 .T =2/
We transform this distribution by first subtracting one from z t (to remove the mean) and
p
then by dividing by T or T . This gives the distributions of the sample mean, zN 1 D

18
 p
˙ tTD1 .z t 1/ =T , and scaled sample mean, zN 2 D T zN 1 as

1
f .zN 1 / D y T =2 1
exp . y=2/ with y D T zN 1 C T , and
2T =2 .T =2/
1 p
f .zN 2 / D y T =2 1
exp . y=2/ with y D T zN 1 C T .
2T =2 .T =2/

These distributions are shown in Figure 1.1. It is clear that f .zN 1 / converges to a spike
at zero as the sample size increases, while f .zN 2 / converges to a (non-trivial) normal
distribution.

Bibliography
Amemiya, T., 1985, Advanced econometrics, Harvard University Press, Cambridge, Mas-
sachusetts.

Davidson, J., 2000, Econometric theory, Blackwell Publishers, Oxford.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Härdle, W., 1990, Applied nonparametric regression, Cambridge University Press, Cam-
bridge.

Harvey, A. C., 1989, Forecasting, structural time series models and the Kalman filter,
Cambridge University Press.

Hayashi, F., 2000, Econometrics, Princeton University Press.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

Lütkepohl, H., 1993, Introduction to multiple time series, Springer-Verlag, 2nd edn.

19
Mittelhammer, R. C., G. J. Judge, and D. J. Miller, 2000, Econometric foundations, Cam-
bridge University Press, Cambridge.

Patterson, K., 2000, An introduction to applied econometrics: a time series approach,

MacMillan Press, London.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Priestley, M. B., 1981, Spectral analysis and time series, Academic Press.

Ruud, P. A., 2000, An introduction to classical econometric theory, Oxford University

Press.

Silverman, B. W., 1986, Density estimation for statistics and data analysis, Chapman and
Hall, London.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

20
2 Univariate Time Series Analysis
Reference: Greene (2000) 13.1-3 and 18.1-3
Additional references: Hayashi (2000) 6.2-4; Verbeek (2004) 8-9; Hamilton (1994); John-
ston and DiNardo (1997) 7; and Pindyck and Rubinfeld (1998) 16-18

2.1 Theoretical Background to Time Series Processes

Suppose we have a sample of T observations of a random variable

˚ i T
y t t D1 D y1i ; y2i ; :::; yTi ;
˚

where subscripts indicate time periods. The superscripts indicate that this sample is from
planet (realization) i . We could imagine a continuum of parallel planets where the same
time series process has generated different samples with T different numbers (different
realizations).
Consider period t . The distribution of y t across the (infinite number of) planets has
some density function, f t .y t /. The mean of this distribution
Z 1
Ey t D y t f t .y t / dy t (2.1)
1

is the expected value of the value in period t, also called the unconditional mean of y t .
Note that Ey t could be different from Ey tCs . The unconditional variance is defined simi-
larly.
Now consider periods t and t s jointly. On planet i we have the pair y ti s ; y ti .
˚

The bivariate distribution of these pairs, across the planets, has some density function
g t s;t .y t s ; y t /.1 Calculate the covariance between y t s and y t as usual
Z 1Z 1
Cov .y t s ; y t / D .y t s Ey t s / .y t Ey t / g t s;t .y t s ; y t / dy t dy t s (2.2)
1 1

D E .y t s Ey t s / .y t Ey t / : (2.3)
1
R1
The relation between f t .y t / and g t s;t .y t s ; yt / is, as usual, f t .y t / = 1 gt s;t .y t s ; y t / dy t s .

21
This is the s t h autocovariance of y t . (Of course, s D 0 or s < 0 are allowed.)
A stochastic process is covariance stationary if

Ey t D  is independent of t; (2.4)
Cov .y t s ; y t / D s depends only on s, and (2.5)
both  and s are finite. (2.6)

Most of these notes are about covariance stationary processes, but Section 2.7 is about
non-stationary processes.
Humanity has so far only discovered one planet with coin flipping; any attempt to
estimate the moments of a time series process must therefore be based on the realization
of the stochastic process from planet earth only. This is meaningful only if the process is
ergodic for the moment you want to estimate. A covariance stationary process is said to
be ergodic for the mean if
T
1X
plim y t D Ey t ; (2.7)
T tD1
so the sample mean converges in probability to the unconditional mean. A sufficient
condition for ergodicity for the mean is
1
X
jCov .y t s ; y t /j < 1: (2.8)
sD0

This means that the link between the values in t and t s goes to zero sufficiently fast
as s increases (you may think of this as getting independent observations before we reach
the limit). If y t is normally distributed, then (2.8) is also sufficient for the process to be
ergodic for all moments, not just the mean. Figure 2.1 illustrates how a longer and longer
sample (of one realization of the same time series process) gets closer and closer to the
unconditional distribution as the sample gets longer.

2.2 Estimation of Autocovariances

Let y t be a vector of a covariance stationary and ergodic. The sth covariance matrix is

R .s/ D E .y t Ey t / .y t s Ey t s /0 : (2.9)

22
 One sample from an AR(1) with corr=0.85 Histogram, obs 1−20
5

0.2
0
0.1

−5
0
0 500 1000 −5 0 5
period

Histogram, obs 1−1000 Mean and Std over longer and longer samples

4 Mean
0.2 Std
2

0.1 0

−2
0
−5 0 5 0 500 1000
sample length

Figure 2.1: Sample of one realization of y t D 0:85y t 1 C" t with y0 D 4 and Std." t / D 1.

Note that R .s/ does not have to be symmetric unless s D 0. However, note that R .s/ D
R . s/0 . This follows from noting that

R . s/ D E .y t Ey t / .y t Cs Ey t Cs /0
D E .y t s Ey t s / .y t Ey t /0 ; (2.10a)

where we have simply changed time subscripts and exploited the fact that y t is covariance
stationary. Transpose to get

R . s/0 D E .y t Ey t / .y t s Ey t s /0 ; (2.11)

which is the same as in (2.9). If y t is a scalar, then R .s/ D R . s/, which shows that
autocovariances are symmetric around s D 0.

23
Example 2.1 (Bivariate case.) Let y t D Œx t ; z t 0 with Ex t DEz t D 0. Then
" #
x t
h i
RO .s/ D E xt s zt s
zt
" #
Cov .x t ; x t s / Cov .x t ; z t s /
D :
Cov .z t ; x t s / Cov .z t ; x t s /

Note that R . s/ is
" #
Cov .x t ; x t Cs / Cov .x t ; z t Cs /
R . s/ D
Cov .z t ; x t Cs / Cov .z t ; x t Cs /
" #
Cov .x t s ; x t / Cov .x t s ; z t /
D ;
Cov .z t s ; x t / Cov .z t s ; x t /

which is indeed the transpose of R .s/.

The autocovariances of the (vector) y t process can be estimated as

T
1 X
O
R .s/ D .y t N .y t
y/ s N 0;
y/ (2.12)
T t D1Cs
T
1X
with yN D yt : (2.13)
T t D1

(We typically divide by T in even if we have only T s full observations to estimate

R .s/ from.)
Autocorrelations are then estimated by dividing the diagonal elements in RO .s/ by the
diagonal elements in RO .0/

O .s/ D diagRO .s/ =diagRO .0/ (element by element). (2.14)

24
2.3 White Noise

A white noise time process has

E" t D 0
Var ." t / D  2 , and
Cov ." t s ; " t / D 0 if s ¤ 0. (2.15)

If, in addition, " t is normally distributed, then it is said to be Gaussian white noise. The
conditions in (2.4)-(2.6) are satisfied so this process is covariance stationary. Moreover,
(2.8) is also satisfied, so the process is ergodic for the mean (and all moments if " t is
normally distributed).

2.4 Moving Average

A qt h -order moving average process is

y t D " t C 1 " t 1 C ::: C q " t q; (2.16)

where the innovation " t is white noise (usually Gaussian). We could also allow both y t
and " t to be vectors; such a process it called a vector MA (VMA).
We have Ey t D 0 and



Var .y t / D E " t C 1 " t C ::: C q " t
1 q " t C 1 " t 1 C ::: C q " t q

D 2 1 C 12 C ::: C q2 :

(2.17)

Autocovariances are calculated similarly, and it should be noted that autocovariances of

order q C 1 and higher are always zero for an MA(q) process.

Example 2.2 The mean of an MA(1), y t D " t C 1 " t 1, is zero since the mean of " t (and
" t 1 ) is zero. The first three autocovariance are

D  2 1 C 12


Var .y t / D E ." t C 1 " t 1 / ." t C 1 " t 1/

Cov .y t 1; yt / D E ." t 1 C 1 " t 2 / ." t C 1 " t 1/ D  2 1

Cov .y t 2; yt / D E ." t 2 C 1 " t 3 / ." t C 1 " t 1/ D 0; (2.18)

25
and Cov.y t s ; y t / D 0 for jsj  2. Since both the mean and the covariances are finite
and constant across t , the MA(1) is covariance stationary. Since the absolute value of
the covariances sum to a finite number, the MA(1) is also ergodic for the mean. The first
autocorrelation of an MA(1) is

1
Corr .y t 1; yt / D :
1 C 12

Since the white noise process is covariance stationary, and since an MA.q/ with m <
1 is a finite order linear function of " t , it must be the case that the MA.q/ is covariance
stationary. It is ergodic for the mean since Cov.y t s ; y t / D 0 for s > q, so (2.8) is
satisfied. As usual, Gaussian innovations are then sufficient for the MA(q) to be ergodic
for all moments.
The effect of " t on y t , y t C1 ; :::, that is, the impulse response function, is the same as
the MA coefficients
@y t @y t C1 @y tCq @y t CqCk
D 1, D 1 ; :::; D q ; and D 0 for k > 0. (2.19)
@" t @" t @" t @" t
This is easily seen from applying (2.16)

y t D " t C 1 " t 1 C ::: C q " t q

y t C1 D " t C1 C 1 " t C ::: C q " t qC1

::
:
y t Cq D " t Cq C 1 " t 1Cq C ::: C q " t
y t CqC1 D " t CqC1 C 1 " t Cq C ::: C q " t C1 :

The expected value of y t , conditional on f"w gtwDs 1 is

Et s yt D Et s " t C 1 " t 1 C ::: C s " t s C ::: C q " t q

D s " t s C ::: C q " t q; (2.20)

since E t s " t .s 1/ D : : : D E t s " t D 0.

Example 2.3 (Forecasting an MA(1).) Suppose the process is

y t D " t C 1 " t 1, with Var ." t / D  2 :

26
The forecasts made in t D 2 then have the follow expressions—with an example using
1 D 2; "1 D 3=4 and "2 D 1=2 in the second column

General Example
y2 D 1=2 C 2  3=4 D 2
E2 y3 D E2 ."3 C 1 "2 / D 1 "2 D 2  1=2 D 1
E2 y4 D E2 ."4 C 1 "3 / D 0 D0

Example 2.4 (MA(1) and conditional variances.) From Example 2.3, the forecasting
variances are—with the numerical example continued assuming that  2 D 1

General Example
Var.y2 E2 y2 / D 0 D0
2
Var.y3 E2 y3 / D Var."3 C 1 "2 1 "2 / D  D1
Var.y4 E2 y4 / D Var ."4 C 1 "3 / D  2 C 12  2 D5

If the innovations are iid Gaussian, then the distribution of the s period forecast error

yt E t s y t D " t C 1 " t 1 C ::: C s 1 " t .s 1/

is
E t s y t /  N 0;  2 1 C 12 C ::: C s2 1 ;


.y t (2.21)

since " t ; " t 1 ; :::; " t .s 1/ are independent Gaussian random variables. This implies that
the conditional distribution of y t , conditional on f"w gswD 1 , is

y t j f" t s ; " t s 1 ; : : :g  N ŒE t s y t ; Var.y t E t s y t / (2.22)

 N s " t s C ::: C q " t q ;  2 1 C 12 C ::: C s2 1 : (2.23)



The conditional mean is the point forecast and the variance is the variance of the forecast
error. Note that if s > q, then the conditional distribution coincides with the unconditional
distribution since " t s for s > q is of no help in forecasting y t .

Example 2.5 (MA(1) and convergence from conditional to unconditional distribution.)

From examples 2.3 and 2.4 we see that the conditional distributions change according to

27
(where ˝2 indicates the information set in t D 2)

General Example
y2 j ˝2  N .y2 ; 0/ D N .2; 0/
y3 j ˝2  N .E2 y3 ; Var.y3 E2 y3 // D N .1; 1/
y4 j ˝2  N .E2 y4 ; Var.y4 E2 y4 // D N .0; 5/

Note that the distribution of y4 j ˝2 coincides with the asymptotic distribution.

Estimation of MA processes is typically done by setting up the likelihood function

and then using some numerical method to maximize it.

2.5 Autoregression

A p t h -order autoregressive process is

y t D a1 y t 1 C a2 y t 2 C ::: C ap y t p C "t : (2.24)

A VAR.p/ is just like the AR.p/ in (2.24), but where y t is interpreted as a vector and ai
as a matrix.

Example 2.6 (VAR(1) model.) A VAR(1) model is of the following form

" # " #" # " #
y1t a11 a12 y1t 1 "1t
D C :
y2t a21 a22 y2t 1 "2t

All stationary AR(p) processes can be written on MA(1) form by repeated substitu-
tion. To do so we rewrite the AR(p) as a first order vector autoregression, VAR(1). For
instance, an AR(2) x t D a1 x t 1 C a2 x t 2 C " t can be written as
" # " #" # " #
xt a1 a2 xt 1 "t
D C , or (2.25)
xt 1 1 0 xt 2 0
y t D Ay t 1 C "t ; (2.26)

where y t is an 2  1 vector and A a 4  4 matrix. This works also if x t and " t are vectors
and. In this case, we interpret ai as matrices and 1 as an identity matrix.

28
 Iterate backwards on (2.26)

y t D A .Ay t 2 C "t 1/ C "t

D A2 y t 2 C A" t 1 C "t
::
:
K
X
KC1
DA yt K 1 C As " t s : (2.27)
sD0

Remark 2.7 (Spectral decomposition.) The n eigenvalues (i ) and associated eigenvec-
tors (zi ) of the n  n matrix A satisfy

.A i In / zi D 0n1 :

If the eigenvectors are linearly independent, then

2 3
1 0


0
6 0 2


0
6 7 h i
A D ZZ 1 , where  D 6
7
7 and Z D z1 z2


zn :
6 :: :: :
4 : :


::
7
5
0 0


n

Note that we therefore get

A2 D AA D ZZ 1
ZZ 1
D ZZ 1
D Z2 Z 1
) Aq D Zq Z 1
:
p
Remark 2.8 (Modulus of complex number.) If  D a C bi, where i D 1, then
p
jj D ja C bij D a2 C b 2 .

Take the limit of (2.27) as K ! 1. If limK!1 AKC1 y t K 1 D 0, then we have a

moving average representation of y t where the influence of the starting values vanishes
asymptotically
X1
yt D As " t s : (2.28)
sD0

We note from the spectral decompositions that AKC1 D ZKC1 Z 1 , where Z is the ma-
trix of eigenvectors and  a diagonal matrix with eigenvalues. Clearly, limK!1 AKC1 y t K 1 D
0 is satisfied if the eigenvalues of A are all less than one in modulus and y t K 1 does not
grow without a bound.

29
 Conditional moments of AR(1), y0=4 Conditional distributions of AR(1), y0=4
4 0.4
s=1
s=3
Mean s=5
2 0.2 s=7
Variance
s=7

0 0
0 10 20 −5 0 5
Forecasting horizon x

Figure 2.2: Conditional moments and distributions for different forecast horizons for the
AR(1) process y t D 0:85y t 1 C " t with y0 D 4 and Std." t / D 1.

Example 2.9 (AR(1).) For the univariate AR(1) y t D ay t 1 C " t , the characteristic
equation is .a / z D 0, which is only satisfied if the eigenvalue is  D a. The AR(1) is
therefore stable (and stationarity) if 1 < a < 1. This can also be seen directly by noting
that aKC1 y t K 1 declines to zero if 0 < a < 1 as K increases.

Similarly, most finite order MA processes can be written (“inverted”) as AR.1/. It is

therefore common to approximate MA processes with AR processes, especially since the
latter are much easier to estimate.

Example 2.10 (Variance of AR(1).) From the MA-representation y t D 1 s

P
sD0 a " t s and
the fact that " t is white noise we get Var.y t / D  2 1 2s
D  2 = 1 a2 . Note
P

sD0 a
that this is minimized at a D 0. The autocorrelations are obviously ajsj . The covariance
matrix of fy t gTtD1 is therefore (standard deviationstandard deviationautocorrelation)
2 3
1 a a2


aT 1
T 2 7
6 7
6 a 1 a


a
2 6 6 a2 T 3 7
7
a 1


a 7:
1 a2 6 :
6
6 :: : ::
7
7
4 5
aT 1 aT 2 aT 3


1

Example 2.11 (Covariance stationarity of an AR(1) with jaj < 1.) From the MA-representation
P1 s
yt D sD0 a " t s , the expected value of y t is zero, since E" t s D 0. We know that
Cov(y t ; y t s )D ajsj  2 = 1 a2 which is constant and finite.

30
Example 2.12 (Ergodicity of a stationary AR(1).) We know that Cov(y t ; y t s )D ajsj  2 = 1 a2 ,

so the absolute value is

jCov.y t ; y t s /j D jajjsj  2 = 1 a2

Using this in (2.8) gives

1 1
X 2 X s
jCov .y t s ; y t /j D jaj
sD0
1 a2 sD0
2 1
D 2
(since jaj < 1)
1 a 1 jaj

which is finite. The AR(1) is ergodic if jaj < 1.

Example 2.13 (Conditional distribution of AR(1).) For the AR(1) y t D ay t 1 C " t with
" t  N 0;  2 , we get

E t y tCs D as y t ,
 
Var .y t Cs E t y t Cs / D 1 C a2 C a4 C ::: C a2.s 1/
2
a2s 1 2
D  :
a2 1
The distribution of y t Cs conditional on y t is normal with these parameters. See Figure
2.2 for an example.

2.5.1 Estimation of an AR(1) Process

Suppose we have sample fy t gTtD0 of a process which we know is an AR.p/, y t D ay t 1 C

" t , with normally distributed innovations with unknown variance  2 .
The pdf of y1 conditional on y0 is
!
1 .y1 ay0 /2
pdf .y1 jy0 / D p exp ; (2.29)
2 2 2 2

and the pdf of y2 conditional on y1 and y0 is

!
1 .y2 ay1 /2
pdf .y2 j fy1 ; y0 g/ D p exp : (2.30)
2 2 2 2

31
 Recall that the joint and conditional pdfs of some variables z and x are related as

pdf .x; z/ D pdf .xjz/  pdf .z/ . (2.31)

Applying this principle on (2.29) and (2.31) gives

pdf .y2 ; y1 jy0 / D pdf .y2 j fy1 ; y0 g/ pdf .y1 jy0 /

2 !
.y2 ay1 /2 C .y1 ay0 /2

1
D p exp : (2.32)
2 2 2 2

Repeating this for the entire sample gives the likelihood function for the sample
T
!
 ˇ 
T =2 1 X
pdf fy t gTtD0 ˇ y0 D 2 2 .y t a1 y t 1 /2 :


exp (2.33)
ˇ
2
2 t D1

Taking logs, and evaluating the first order conditions for  2 and a gives the usual OLS
estimator. Note that this is MLE conditional on y0 . There is a corresponding exact MLE,
but the difference is usually small (the asymptotic distributions of the two estimators are
the same under stationarity; under non-stationarity OLS still gives consistent estimates).
The MLE of Var(" t ) is given by TtD1 vO t2 =T , where vO t is the OLS residual.
P

These results carry over to any finite-order VAR. The MLE, conditional on the initial
observations, of the VAR is the same as OLS estimates of each equation. The MLE of
the ij t h element in Cov(" t ) is given by TtD1 vO i t vOjt =T , where vO i t and vOjt are the OLS
P

residuals.
To get the exact MLE, we need to multiply (2.33) with the unconditional pdf of y0
(since we have no information to condition on)

y02
 
1
pdf .y0 / D p exp ; (2.34)
2 2 =.1 a2 / 2 2 =.1 a2 /

since y0  N.0;  2 =.1 a2 //. The optimization problem is then non-linear and must be
solved by a numerical optimization routine.

32
2.5.2 Lag Operators

A common and convenient way of dealing with leads and lags is the lag operator, L. It is
such that
Ls y t D y t s for all (integer) s.

For instance, the ARMA(2,1) model

yt a1 y t 1 a2 y t 2 D " t C 1 " t 1 (2.35)

can be written as
a2 L2 y t D .1 C 1 L/ " t ;


1 a1 L (2.36)

which is usually denoted

a .L/ y t D  .L/ " t : (2.37)

2.5.3 Properties of LS Estimates of an AR(p) Process

Reference: Hamilton (1994) 8.2

The LS estimates are typically biased, but consistent and asymptotically normally
distributed, provided the AR is stationary.
As usual the LS estimate is
" T
# 1 T
O 1 X
0 1X
ˇLS ˇ D xt xt x t " t , where (2.38)
T tD1 T t D1
h i
xt D yt 1 yt 2


yt p :

The first term in (2.38) is the inverse of the sample estimate of covariance matrix of
x t (since Ey t D 0), which converges in probability to ˙xx1 (y t is stationary and ergodic
for all moments if " t is Gaussian). The last term, T1 TtD1 x t " t , is serially uncorrelated,
P

so we can apply a CLT. Note that Ex t " t "0t x t0 DE" t "0t Ex t x t0 D  2 ˙xx since u t and x t are
independent. We therefore have
T
1 X
x t " t !d N 0;  2 ˙xx :


p (2.39)
T t D1

33
Combining these facts, we get the asymptotic distribution
p  
O
T ˇLS ˇ !d N 0; ˙xx1  2 :


(2.40)

Consistency follows from taking plim of (2.38)

T
  1X
plim ˇOLS ˇ D ˙xx1 plim xt "t
T tD1
D 0;

since x t and " t are uncorrelated.

2.5.4 Autoregressions versus Autocorrelations

It is straightforward to see the relation between autocorrelations and the AR model when
the AR model is the true process. This relation is given by the Yule-Walker equations.
For an AR(1), the autoregression coefficient is simply the first autocorrelation coeffi-
cient. For an AR(2), y t D a1 y t 1 C a2 y t 2 C " t , we have
2 3 2 3
Cov.y t ; y t / Cov.y t ; a1 y t 1 C a2 y t 2 C " t /
4 Cov.y t 1 ; y t / 5 D 4 Cov.y t 1 ; a1 y t 1 C a2 y t 2 C " t / 5
6 7 6 7

Cov.y t 2 ; y t / Cov.y t 2 ; a1 y t 1 C a2 y t 2 C " t /

2 3
a1 Cov.y t ; y t 1 / C a2 Cov.y t ; y t 2 / C Cov.y t ; " t /
D 4 a1 Cov.y t 1 ; y t 1 / C a2 Cov.y t 1 ; y t 2 / 5 , or
6 7

a1 Cov.y t C a2 Cov.y t
2; yt 1/ 2; yt 2/
2 3 2 3
0 a1 1 C a2 2 C Var." t /
4 1 5 D 4 a1 0 C a2 1 5: (2.41)
6 7 6 7

2 a1 1 C a2 0

To transform to autocorrelation, divide through by 0 . The last two equations are then
" # " # " # " #
1 a1 C a2 1 1 a1 = .1 a2 /
D or D : (2.42)
2 a1  1 C a2 2 a12 = .1 a2 / C a2

If we know the parameters of the AR(2) model (a1 , a2 , and Var." t /), then we can
solve for the autocorrelations. Alternatively, if we know the autocorrelations, then we

34
can solve for the autoregression coefficients. This demonstrates that testing that all the
autocorrelations are zero is essentially the same as testing if all the autoregressive coeffi-
cients are zero. Note, however, that the transformation is non-linear, which may make a
difference in small samples.

2.6 ARMA Models

An ARMA model has both AR and MA components. For instance, an ARMA(p,q) is

y t D a1 y t 1 C a2 y t 2 C ::: C ap y t p C " t C 1 " t 1 C ::: C q " t q: (2.43)

Estimation of ARMA processes is typically done by setting up the likelihood function and
then using some numerical method to maximize it.
Even low-order ARMA models can be fairly flexible. For instance, the ARMA(1,1)
model is
y t D ay t 1 C " t C " t 1 , where " t is white noise. (2.44)

The model can be written on MA(1) form as

1
X
yt D "t C as 1 .a C /" t s : (2.45)
sD1

The autocorrelations can be shown to be

.1 C a/.a C /
1 D , and s D as 1 for s D 2; 3; : : : (2.46)
1 C  2 C 2a
and the conditional expectations are

E t y t Cs D as 1 .ay t C " t / s D 1; 2; : : : (2.47)

See Figure 2.3 for an example.

35
 a. Impulse response of a=0.9 a. Impulse response of a=0
2 2

0 0
θ=−0.8
θ=0
θ=0.8
−2 −2
0 5 10 0 5 10
period period

a. Impulse response of a=−0.9

2 ARMA(1,1): yt = ayt−1 + εt + θεt−1

−2
0 5 10
period

Figure 2.3: Impulse response function of ARMA(1,1)

2.7 Non-stationary Processes

2.7.1 Introduction

A trend-stationary process can be made stationary by subtracting a linear trend. The

simplest example is
y t D  C ˇt C " t (2.48)

where " t is white noise.

A unit root process can be made stationary only by taking a difference. The simplest
example is the random walk with drift

yt D  C yt 1 C "t ; (2.49)

where " t is white noise. The name “unit root process” comes from the fact that the largest

36
eigenvalues of the canonical form (the VAR(1) form of the AR(p)) is one. Such a process
is said to be integrated of order one (often denoted I(1)) and can be made stationary by
taking first differences.

Example 2.14 (Non-stationary AR(2).) The process y t D 1:5y t 1 0:5y t 2 C " t can be
written " # " #" # " #
yt 1:5 0:5 yt 1 "t
D C ;
yt 1 1 0 yt 2 0
where the matrix has the eigenvalues 1 and 0.5 and is therefore non-stationary. Note that
subtracting y t 1 from both sides gives y t y t 1 D 0:5 .y t 1 y t 2 /C" t , so the variable
x t D y t y t 1 is stationary.

The distinguishing feature of unit root processes is that the effect of a shock never
vanishes. This is most easily seen for the random walk. Substitute repeatedly in (2.49) to
get

y t D  C . C y t 2 C "t 1/ C "t
::
:
t
X
D t C y0 C "s : (2.50)
sD1

The effect of " t never dies out: a non-zero value of " t gives a permanent shift of the level
of y t . This process is clearly non-stationary. A consequence of the permanent effect of
a shock is that the variance of the conditional distribution grows without bound as the
forecasting horizon is extended. For instance, for the random walk with drift, (2.50), the
distribution conditional on the information in t D 0 is N y0 C t; s 2 if the innova-

tions are Gaussian. This means that the expected change is t and that the conditional
variance grows linearly with the forecasting horizon. The unconditional variance is there-
fore infinite and the standard results on inference are not applicable.
In contrast, the conditional distributions from the trend stationary model, (2.48), is
N st;  2 .

A process could have two unit roots (integrated of order 2: I(2)). In this case, we need
to difference twice to make it stationary. Alternatively, a process can also be explosive,
that is, have eigenvalues outside the unit circle. In this case, the impulse response function
diverges.

37
Example 2.15 (Two unit roots.) Suppose y t in Example (2.14) is actually the first differ-
ence of some other series, y t D z t z t 1 . We then have

zt zt 1 D 1:5 .z t 1 zt 2/ 0:5 .z t 2 zt 3/ C "t

z t D 2:5z t 1 2z t 2 C 0:5z t 3 C "t ;

which is an AR(3) with the following canonical form

2 3 2 32 3 2
3
zt 2:5 2 0:5 zt 1 "t
4 zt 1 5 D 4 1 0 0 5 4 zt 5 C 4 0 5:
6 7 6 76 7 6 7
2

zt 2 0 1 0 zt 3 0

The eigenvalues are 1, 1, and 0.5, so z t has two unit roots (integrated of order 2: I(2) and
needs to be differenced twice to become stationary).

Example 2.16 (Explosive AR(1).) Consider the process y t D 1:5y t 1 C " t . The eigen-
value is then outside the unit circle, so the process is explosive. This means that the
impulse response to a shock to " t diverges (it is 1:5s for s periods ahead).

2.7.2 Spurious Regressions

Strong trends often causes problems in econometric models where y t is regressed on x t .

In essence, if no trend is included in the regression, then x t will appear to be significant,
just because it is a proxy for a trend. The same holds for unit root processes, even if
they have no deterministic trends. However, the innovations accumulate and the series
therefore tend to be trending in small samples. A warning sign of a spurious regression is
when R2 > DW statistics.
For trend-stationary data, this problem is easily solved by detrending with a linear
trend (before estimating or just adding a trend to the regression).
However, this is usually a poor method for a unit root processes. What is needed is a
first difference. For instance, a first difference of the random walk is

y t D y t yt 1

D "t ; (2.51)

which is white noise (any finite difference, like y t y t s , will give a stationary series),

38
so we could proceed by applying standard econometric tools to
y t .
One may then be tempted to try first-differencing all non-stationary series, since it
may be hard to tell if they are unit root process or just trend-stationary. For instance, a
first difference of the trend stationary process, (2.48), gives

yt yt 1 D ˇ C "t "t 1: (2.52)

Its unclear if this is an improvement: the trend is gone, but the errors are now of MA(1)
type (in fact, non-invertible, and therefore tricky, in particular for estimation).

2.7.3 Testing for a Unit Root I

Suppose we run an OLS regression of

y t D ay t 1 C "t ; (2.53)

where the true value of jaj < 1. The asymptotic distribution is of the LS estimator is
p
a2 :


T .aO a/  N 0; 1 (2.54)

(The variance follows from the standard OLS formula where the variance of the estimator
is  2 .X 0 X=T / 1 . Here plim X 0 X=T DVar.y t / which we know is  2 = 1 a2 ).

It is well known (but not easy to show) that when a D 1, then aO is biased towards
zero in small samples. In addition, the asymptotic distribution is no longer (2.54). In
fact, there is a discontinuity in the limiting distribution as we move from a stationary/to
a non-stationary variable. This, together with the small sample bias means that we have
to use simulated critical values for testing the null hypothesis of a D 1 based on the OLS
estimate from (2.53).
The approach is to calculate the test statistic

aO 1
tD ;
O
Std.a/
and reject the null of non-stationarity if t is less than the critical values published by
Dickey and Fuller (typically more negative than the standard values to compensate for the
small sample bias) or from your own simulations.
In principle, distinguishing between a stationary and a non-stationary series is very

39
difficult (and impossible unless we restrict the class of processes, for instance, to an
AR(2)), since any sample of a non-stationary process can be arbitrary well approximated
by some stationary process et vice versa. The lesson to be learned, from a practical point
of view, is that strong persistence in the data generating process (stationary or not) invali-
dates the usual results on inference. We are usually on safer ground to apply the unit root
results in this case, even if the process is actually stationary.

2.7.4 Testing for a Unit Root II

Reference: Fuller (1976), Introduction to Statistical Time Series; Dickey and Fuller (1979),
“Distribution of the Estimators for Autoregressive Time Series with a Unit Root,” Journal
of the American Statistical Association, 74, 427-431.
Consider the AR(1) with intercept

y t D C ˛y t 1 C u t ; or
y t D C ˇy t 1 C u t ; where ˇ D .˛ 1/ : (2.55)

The DF test is to test the null hypothesis that ˇ D 0, against ˇ < 0 using the usual
t statistic. However, under the null hypothesis, the distribution of the t statistics is far
from a student-t or normal distribution. Critical values, found in Fuller and Dickey and
Fuller, are lower than the usual ones. Remember to add any nonstochastic regressors
that in required, for instance, seasonal dummies, trends, etc. If you forget a trend, then
the power of the test goes to zero as T ! 1. The critical values are lower the more
deterministic components that are added.
The asymptotic critical values are valid even under heteroskedasticity, and non-normal
distributions of u t . However, no autocorrelation in u t is allowed for. In contrast, the
simulated small sample critical values are usually only valid for iid normally distributed
disturbances.
The ADF test is a way to account for serial correlation in u t . The same critical values
apply. Consider an AR(1) u t D u t 1 C e t . A Cochrane-Orcutt transformation of (2.55)
gives

y t D .1 Q t
/ C ˇy 1 C  .ˇ C 1/
y t 1 C e t ; where ˇQ D ˇ .1 / : (2.56)

Q The fact that ˇQ D ˇ .1 / is of no importance, since ˇQ is

The test is here the t test for ˇ.
zero only if ˇ is (as long as  < 1, as it must be). (2.56) generalizes so one should include

40
p lags of
y t if u t is an AR(p). The test remains valid even under an MA structure if
the number of lags included increases at the rate T 1=3 as the sample lenngth increases.
In practice: add lags until the remaining residual is white noise. The size of the test
(probability of rejecting H0 when it is actually correct) can be awful in small samples for
a series that is a I(1) process that initially “overshoots” over time, as
y t D e t 0:8e t 1 ,
since this makes the series look mean reverting (stationary). Similarly, the power (prob of
rejecting H0 when it is false) can be awful when there is a lot of persistence, for instance,
if ˛ D 0:95.
The power of the test depends on the span of the data, rather than the number of
observations. Seasonally adjusted data tend to look more integrated than they are. Should
apply different critical values, see Ghysel and Perron (1993), Journal of Econometrics,
55, 57-98. A break in mean or trend also makes the data look non-stationary. Should
perhaps apply tests that account for this, see Banerjee, Lumsdaine, Stock (1992), Journal
of Business and Economics Statistics, 10, 271-287.
Park (1990, “Testing for Unit Roots and Cointegration by Variable Addition,” Ad-
vances in Econometrics, 8, 107-133) sets up a framework where we can use both non-
stationarity as the null hypothesis and where we can have stationarity as the null. Consider
the regression
p q
X X
yt D ˇs t s C ˇs t s C u t ; (2.57)
sD0 sDpC1

where the we want to test if H0 : ˇs D 0, s D p C1; :::; q. If F .p; q/ is the Wald-statistics

for this, then J .p; q/ D F .p; q/ =T has some (complicated) asymptotic distribution
under the null. You reject non-stationarity if J .p; q/ < critical value, since J .p; q/ !p
0 under (trend) stationarity.
Now, define
Var .u t /
G .p; q/ D F .p; q/ p   p2 q under H0 of stationarity, (2.58)
Var T uN t

and G .p; q/ !p 1 under non-stationarity, so we reject stationarity

p ifG .p; q/ > critical
value. Note that Var.u t / is a traditional variance, while Var T uN t can be estimated
with a Newey-West estimator.

41
2.7.5 Cointegration

Suppose y1t and y2t are both (scalar) unit root processes, but that

z t D y1t ˇy2t (2.59)

" #
h i y
1t
D 1 ˇ
y2t

is stationary. The processes y t and x t must then share the samehcommon stochastic
i trend,
and are therefore cointegrated with the cointegrating vector 1 ˇ . Running the
regression (2.59) gives an estimator ˇOLS which converges much faster than usual (it is
“superconsistent”) and is not affected by any simultaneous equations bias. The intuition
for the second result is that the simultaneous equations bias depends on the simultaneous
reactions to the shocks, which are stationary and therefore without any long-run impor-
tance.
This can be generalized by letting y t be a vector of n unit root processes which follows
a VAR. For simplicity assume it is a VAR(2)

y t D A1 y t 1 C A2 y t 2 C "t : (2.60)

Subtract y t from both sides, add and subtract A2 y t 1 from the right hand side

yt yt 1 D A1 y t 1 C A2 y t 2 C "t yt 1 C A2 y t 1 A2 y t 1

D .A1 C A2 I / yt 1 A2 .y t 1 yt 2/ C "t (2.61)

The left hand side is now stationary, and so is y t 1 y t 2 and " t on the right hand side. It
must therefore be the case that .A1 C A2 I / y t 1 is also stationary; it must be n linear
combinations of the cointegrating vectors. Since the number of cointegrating vectors must
be less than n, the rank of A1 CA2 I must be less than n. To impose this calls for special
estimation methods.
The simplest of these is Engle and Granger’s two-step procedure. In the first step, we
estimate the cointegrating vectors (as in 2.59) and calculate the different z t series (fewer
than n). In the second step, these are used in the error correction form of the VAR

yt yt 1 D z t 1 A2 .y t 1 yt 2/ C "t (2.62)

42
to estimate and A2 . The relation to (2.61) is most easily seen in the bivariate case. Then,
by using (2.59) in (2.62) we get
h i
yt yt 1 D ˇ y t 1 A2 .y t 1 y t 2 / C " t ; (2.63)

so knowledge (estimates) of ˇ (scalar), (2  1), A2 (2  2) allows us to “back out” A1 .

Bibliography
Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Hayashi, F., 2000, Econometrics, Princeton University Press.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

43
3 The Distribution of a Sample Average
Reference: Hayashi (2000) 6.5
Additional references: Hamilton (1994) 14; Verbeek (2004) 4.10; Harris and Matyas
(1999); and Pindyck and Rubinfeld (1998) Appendix 10.1; Cochrane (2001) 11.7

3.1 Variance of a Sample Average

In order to understand the distribution of many estimators we need to get an important

building block: the variance of a sample average.
Consider a covariance stationary vector process m t with zero mean and Cov.m t ; m t s / D
R .s/ (which only depends on s). That is, we allow for serial correlation in m t , but no
heteroskedasticity. This is more restrictive than we want, but we will return to that further
on.
N D TtD1 m t =T . The sampling variance of a mean estimator of the zero mean
P
Let m
random variable m t is defined as
2
T
! T
!0 3
1 X 1 X
Cov .m/N D E4 mt m 5 : (3.1)
T t D1 T  D1

Let the covariance (matrix) at lag s be

R .s/ D Cov .m t ; m t s /
D E m t m0t s ; (3.2)

since E m t D 0 for all t .

44
Example 3.1 (m t is a scalar iid process.) When m t is a scalar iid process, then
T T
!
1X 1 X
Var mt D 2 Var .m t / /*independently distributed*/
T t D1 T t D1
1
D T Var .m t / /*identically distributed*/
T2
1
D Var .m t / :
T
N D 0. By multiplying both sides by
pcase. Clearly, limT )1 Var.m/
This is the classical iid
T we instead get Var T m N D Var.m t /, which is often more convenient for asymptotics.

Example i3.2 Let x t and z t be two scalars, with samples averages xN and z.
N Let m t D
h 0
x t z t . Then Cov.m/ N is
" #!
xN
N D Cov
Cov .m/
zN
" #
N
Var .x/ N z/
Cov .x; N
D :
N x/
Cov .z; N Var .zN /

N with T D 3.) With T D 3, we have

Example 3.3 (Cov.m/

N D
Cov .T m/
E .m1 C m2 C m3 / m01 C m02 C m03 D

E m1 m01 C m2 m02 C m3 m03 C E m2 m01 C m3 m02 C E m1 m02 C m2 m03 C Em3 m01 C Em1 m03 :




„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
3R.0/ 2R.1/ 2R. 1/ R.2/ R. 2/

The general pattern in the previous example is

T
X1
N D
Cov .T m/ .T jsj/ R.s/: (3.3)
sD .T 1/

Divide both sides by T

p T
X1  
 jsj
Cov N D
Tm 1 R.s/. (3.4)
T
sD .T 1/

This is the exact expression for a given sample size.

45
 In many cases, we use the asymptotic expression (limiting value as T ! 1) instead.
If R .s/ D 0 for s > q so m t is an MA(q), then the limit as the sample size goes to infinity
is
p  p  q
X
ACov N D lim Cov
Tm N D
Tm R.s/; (3.5)
T !1
sD q

where ACov stands for the asymptotic variance-covariance matrix. This continues to hold
even if q D 1, provided R .s/ goes to zero sufficiently quickly, as it does in stationary
VAR systems. In this case we have
p  1
X
ACov N D
Tm R.s/: (3.6)
sD 1

Estimation in finite samples will of course require some cut-off point, which is discussed
below. p 
The traditional estimator of ACov T m N is just R.0/, which is correct when m t has
no autocorrelation, that is
p 
ACov Tm N D R.0/ D Cov .m t ; m t / if Cov .m t ; m t s / D 0 for s ¤ 0: (3.7)

By comparing with (3.5) we see that this underestimates the true variance if autocovari-
ances are mostly positive, and overestimates if they are mostly negative. The errors can
be substantial.

Example 3.4 (Variance of sample mean of AR(1).) Let m t D m t 1 Cu t , where Var.u t / D

 2 . Note that R .s/ D jsj  2 = 1 2 , so

p  1
X
AVar N D
Tm R.s/
sD 1
1 1
!
 2 X jsj 2 X
D  D 1C2 s
1 2 sD 1 1 2 sD1
2
 1C
D ;
1 2 1 

which is increasing in  (provided jj < 1, as required for stationarity). The variance
N is much larger for  close to one than for  close to zero: the high autocorrelation
of m
create long swings, so the mean cannot be estimated with any good precision in a small

46
 Variance of sample mean, AR(1) Var(sample mean)/Var(series), AR(1)
100 100

50 50

0 0
−1 0 1 −1 0 1
AR(1) coefficient AR(1) coefficient

p
Figure 3.1: Variance of T times sample mean of AR(1) process m t D m t 1 C ut .

sample. If we disregard all autocovariances, then we would conclude that the variance of
p
N is  2 = 1 2 , which is smaller (larger) than the true value when  > 0 ( < 0).


Tm
For instance, with  D 0:85, it is approximately 12 times too small. See Figure 3.1 for an
illustration.

Example 3.5 (Variance of sample mean of AR(1), continued.) Part of the reason why
N increased with  in the previous examples is that Var.m t / increases with . We
Var.m/
p
can eliminate this effect by considering how much larger AVar. T m/ N is than in the iid
p
N
case, that is, AVar. T m/=Var.m t / D .1 C / = .1 /. This ratio is one for  D 0 (iid
data), less than one for  < 0, and greater than one for  > 0. This says that if relatively
more of the variance in m t comes from long swings (high ), then the sample mean is
more uncertain. See Figure 3.1 for an illustration.

Example 3.6 (Variance of sample mean of AR(1), illustration of why limT !1 of (3.4).)
For an AR(1) (3.4) is

p T
X1
2
 
 jsj jsj
Var N D
Tm 1 
1 2 T
sD .T 1/
T
X1 
" #
2 s s
D 1C2 1 
1 2 sD1
T
2
T C1 
 
 
D 1C2 C2 :
1 2 1  T .1 /2

47
The last term in brackets goes to zero as T goes to infinity. We then get the result in
Example 3.4.

3.2 The Newey-West Estimator

3.2.1 Definition of the Estimator

Newey and West (1987) suggested the following estimator of the covariance matrix in
(3.5) as (for some n < T )

1 p n  
 X jsj O
ACov N D
Tm 1 R.s/
sD n
n C 1
n  
O
X s O

O s/ ; or since R.
O s/ D RO 0 .s/;
D R.0/ C 1 R.s/ C R.
sD1
n C 1
n  
O
X s O

D R.0/ C 1 R.s/ C RO 0 .s/ , where (3.8)
sD1
n C 1

T
O 1 X 0
R.s/ D mt mt s .if E m t D 0/: (3.9)
T t DsC1
The tent shaped (Bartlett) weights in (3.8) guarantee a positive definite covariance
estimate. In contrast, equal weights (as in (3.5)), may give an estimated covariance matrix
which is not positive definite, which is fairly awkward. Newey and West (1987) showed
that this estimator is consistent if we let n go to infinity as T does, but in such a way that
n=T 1=4 goes to zero.
There are several other possible estimators of the covariance matrix in (3.5), but sim-
ulation evidence suggest that they typically do not improve a lot on the Newey-West
estimator.

Example 3.7 (m t is MA(1).) Suppose we know that m t D " t C" t 1 .Then R.s/
 D 0 for
1
s  2, so it might be tempting to use n D 1 in (3.8). This gives ACov
p
Tm O
N D R.0/ C
1 O
2
ŒR.1/ C RO 0 .1/, while the theoretical expression (3.5) is ACovD R.0/ C R.1/ C R0 .1/.
The Newey-West estimator puts too low weights on the first lead and lag, which suggests
that we should use n > 1 (or more generally, n > q for an MA(q) process).

48
 Scatter plot, iid residuals Scatter plot, Var(residual) depends on x2

20 20
10 10
0 0
y

y
−10 −10
−20 −20
−10 −5 0 5 10 −10 −5 0 5 10
x x

y = 0.03 + 1.3x + u
Solid regression lines are based on all data,
dashed lines exclude the crossed out data point

Figure 3.2: Effect of heteroskedasticity on uncertainty about regression line

O
punder quite general circumstances, S in (3.8)-(3.9) is a
It can also be shown that,
consistent estimator of ACov T m N , even if m t is heteroskedastic (on top of being au-
tocorrelated). (See Hamilton (1994) 10.5 for a discussion.)

3.2.2 How to Implement the Newey-West Estimator

Economic theory and/or stylized facts can sometimes help us choose the lag length n.
For instance, we may have a model of stock returns which typically show little autocor-
relation, so it may make sense to set n D 0 or n D 1 in that case. A popular choice of
n is to round .T =100/1=4 down to the closest integer, although this does not satisfy the
consistency requirement.
It is important to note that definition of the covariance matrices in (3.2) and (3.9)
assume that m t has zero mean. If that is not the case, then the mean should be removed
in the calculation of the covariance matrix. In practice, you remove the same number,
estimated on the whole sample, from both m t and m t s . It is often recommended to
remove the sample means even if theory tells you that the true mean is zero.

49
 Std of LS slope coefficient under heteroskedasticity
0.1
OLS formula
White’s
Simulated
0.08

0.06

0.04 Model: yt = 0.9xt + ǫt ,

where ǫt ∼ N (0, ht ), with ht = 0.5exp(αx2t )

bLS is the LS estimate of b in

0.02 yt = a + bxt + ut

Number of simulations: 25000

0
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
α (effect of regressor on variance)

Figure 3.3: Variance of OLS estimator, heteroskedastic errors

3.3 Summary

T
1X
N D
Let m m t and R .s/ D Cov .m t ; m t s / . Then
T tD1
p  1
X
ACov N D
Tm R.s/
p sD 1
ACov N D R.0/ D Cov .m t ; m t / if R.s/ D 0 for s ¤ 0
Tm

1
p n  

O
X s O O 0

Newey-West W ACov N D R.0/ C
Tm 1 R.s/ C R .s/ :
sD1
nC1

Bibliography
Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

50
 Scatter plot, iid residuals Scatter plot, autocorrelated residuals

20 20

0 0
y

y
−20 −20

−10 −5 0 5 10 −10 −5 0 5 10
x x

y = 0.03 + 1.3x + u
Solid regression lines are based on all data,
dashed lines are based on late sample (high x values).
The regressor is (strongly) autocorrelated, since
it is an increasing series (-10,-9.9,...,10).

Figure 3.4: Effect of autocorrelation on uncertainty about regression line

Harris, D., and L. Matyas, 1999, “Introduction to the generalized method of moments esti-
mation,” in Laszlo Matyas (ed.), Generalized Method of Moments Estimation . chap. 1,
Cambridge University Press.

Hayashi, F., 2000, Econometrics, Princeton University Press.

Newey, W. K., and K. D. West, 1987, “A simple positive semi-definite, heteroskedasticity

and autocorrelation consistent covariance matrix,” Econometrica, 55, 703–708.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

51
 Std of LS slope under autocorrelation Std of LS slope under autocorrelation
0.1 κ = −0.9 0.1 κ = 0
OLS formula
Newey-West
Simulated
0.05 0.05

0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS slope under autocorrelation
where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9
xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 3.5: Variance of OLS estimator, autocorrelated errors

52
4 Least Squares
Reference: Greene (2000) 6
Additional references: Hayashi (2000) 1-2; Verbeek (2004) 1-4; Hamilton (1994) 8

4.1 Definition of the LS Estimator

4.1.1 LS with Summation Operators

Consider the linear model

y t D x t0 ˇ0 C u t ; (4.1)

where y t and u t are scalars, x t a k  1 vector, and ˇ0 is a k  1 vector of the true

coefficients. Least squares minimizes the sum of the squared fitted residuals
T T
X X
2
e t2 D yt x t0 ˇ ; (4.2)
t D1 t D1

by choosing the vector ˇ. The first order conditions are

T
X  
0kx1 D xt yt x t0 ˇOLS or (4.3)
t D1
T
X T
X
xt yt D x t x t0 ˇOLS ; (4.4)
tD1 t D1

which are the so called normal equations. These can be solved as

1
T T
!
X X
ˇOLS D x t x t0 xt yt (4.5)
t D1 t D1
1
T T
!
1 X 1X
D x t x t0 xt yt (4.6)
T t D1
T t D1

53
Remark 4.1 (Summation and vectors) Let z t and x t be the vectors
2 3
" # x1t
z1t
zt D and x t D 4 x2t 5 ;
6 7
z2t
x3t

then
2 3 2 3 2 PT PT 3
T T x 1t T x 1t z1t x 1t z 2t t D1 x1t z1t x1t z2t
X X h i X
7 6 PT PTt D1
x t z t0 D 4 x2t 5 z1t z2t D D
4 x2t z1t x2t z2t 5 4 t D1 x2t z1t tD1 x2t z2t 5 :
6 7 6 7
t D1 tD1 t D1
PT PT
x3t x3t z1t x3t z2t t D1 x3t z1t t D1 x3t z2t

4.1.2 LS in Matrix Form

Define the matrices

2 3 2 3 2 3 2 3
y1 u1 x10 e1
6 y2 7 u2 x20 e2
6 7 6 7 6 7 6 7
6 7 6 7 6 7
Y D6 6 :: 7
7 ; u D 6 :: 7 ; X D6 :: 7 ; and e D 6 :: 7 :
4 : 5 : : :
6 7 6 7 6 7
4 5 4 5 4 5
yT T 1
uT T 1
xT0 T k
eT T 1
(4.7)
Write the model (4.1) as
2 3 2 3 2 3
y1 x10 u1
y2 x20 u2
6 7 6 7 6 7
6 7 6 7 6 7
6 :: 7D6 :: 7 ˇ0 C 6 :: 7 or (4.8)
: : :
6 7 6 7 6 7
4 5 4 5 4 5
yT xT0 uT
Y D Xˇ0 C u: (4.9)

Remark 4.2 Let x t be a k  1 and z t an m  1 vector. Define the matrices

2 3 2 3
x10 z10
6 0 7 6 0 7
6 x2 7 6 z2 7
X D6 6 :: 7
7 and Z D 6 6 :: 7
7 :
4 : 5 4 : 5
xT0 T k
zT0 T m

54
We then have
T
X
x t z t0 D X 0 Z:
t D1

We can then rewrite the loss function (4.2) as e 0 e, the first order conditions (4.3) and
(4.4) as (recall that y t D y t0 since it is a scalar)
 
0kx1 D X 0 Y X ˇOLS (4.10)

X 0 Y D X 0 X ˇOLS ; (4.11)

and the solution (4.5) as

1
ˇOLS D X 0 X X 0 Y:


(4.12)

4.2 LS and R2 

The first order conditions in LS are

T
X
x t uO t D 0, where uO t D y t O
yO t , with yO t D x t0 ˇ: (4.13)
t D1

T
This implies that the fitted residuals and fitted values are orthogonal, ˙ tD1 T
yO t uO t D ˙ tD1 ˇO 0 x t uO t D
0. If we let x t include a constant, then (4.13) also implies that the fitted residuals have a
zero mean, ˙ tTD1 uO t =T D 0. We can then decompose the sample variance (denoted Var) c
of y t D yO t C uO t as
c .y t / D Var
Var c .yO t / C Var
c .uO t / ; (4.14)

since yO t and uO t are uncorrelated in this case. (Note that Cov.yO t ; uO t / D EyO t uO t EyO t EuO t so
the orthogonality is not enough to allow the decomposition; we also need EyO t EuO t D 0—
this holds for sample moments as well.)
We define R2 as the fraction of Var c .y t / that is explained by the model

c .yO t /
Var
R2 D (4.15)
c .y t /
Var
c .uO t /
Var
D1 : (4.16)
c .y t /
Var

55
 c .uO t /, so it
LS minimizes the sum of squared fitted errors, which is proportional to Var
maximizes R2 .
We can rewrite R2 by noting that

b b
Cov .y t ; yO t / D Cov .yO t C uO t ; yO t / D Var
c .yO t / : (4.17)

Use this to substitute for Var b

c .yO t / in (4.15) and then multiply both sides with Cov .y t ; yO t / =Var
c .yO t / D
1 to get

2
R D
b
Cov .y t ; yO t /2
c .yO t /
c .y t / Var
Var
b
D Corr .y t ; yO t /2 (4.18)

which shows that R2 is the square of correlation coefficient of the actual and fitted value.
b
Note that this interpretation of R2 relies on the fact that Cov .yO t ; uO t / D 0. From (4.14)
this implies that the sample variance of the fitted variables is smaller than the sample
variance of y t . From (4.15) we see that this implies that 0  R2  1.
To get a bit more intuition for what R2 represents, suppose the estimated coefficients

2
equal the true coefficients, so yO t D x t0 ˇ0 . In this case, R2 D Corr x t0 ˇ0 C u t ; x t0 ˇ0 ,
that is, the squared correlation of y t with the systematic part of y t . Clearly, if the model
is perfect so u t D 0, then R2 D 1. On contrast, when there is no movements in the
systematic part (ˇ0 D 0), then R2 D 0.

Remark 4.3 In a simple regression where y t D a C bx t C u t , where x t is a scalar,

b
R2 D Corr .y t ; x t /2 . To see this, note that, in this case (4.18) can be written

2
b 
O
Cov y ; bx
t
 2
b
bO 2 Cov .y t ; x t /2
t
R D  D ;
c .y t / Var
Var c bx O t bO 2 Var
c .y t / Var
c .x t /

so the bO 2 terms cancel.

Remark 4.4 Now, consider the reverse regression x t D c C dy t C v t . The LS estimator

b
of the slope is dOLS D Cov .y t ; x t / =Var b
c .y t /. Recall that bOLS D Cov .y t ; x t / =Var
c .x t /.

b
We therefore have
Cov .y t ; x t /2
bOLS dOLS D D R2 :
Var .y t / Var .x t /
c c

56
This shows that dOLS D 1=bOLS if (and only if) R2 D 1.

4.3 Finite Sample Properties of LS

Use the true model (4.1) to substitute for y t in the definition of the LS estimator (4.6)
1
T T
!
1X 1X
ˇOLS D x t x t0 x t x t0 ˇ0 C u t


T t D1 T t D1
T
! 1 T
1X 0 1X
D ˇ0 C xt xt xt ut : (4.19)
T t D1 T tD1

It is possible to show unbiasedness of the LS estimator, even if x t stochastic  and u t is 

autocorrelated and heteroskedastic—provided E.u t jx t s / D 0 for all s. Let E u t j fx t gTtD1
denote the expectation of u t conditional on all values of x t s . Using iterated expectations
on (4.19) then gives
2
T
! 1 T
3
1 X 1 X  
EˇOLS D ˇ0 C Ex 4 x t x t0 x t E u t j fx t gTtD1 5 (4.20)
T t D1 T t D1

D ˇ0 ; (4.21)

since E.u t jx t s / D 0 for all s. This is, for instance, the case when the regressors are
deterministic. Notice that E. u t j x t / D 0 is not enough for unbiasedness since (4.19)
contains terms involving x t s x t u t from the product of . T1 TtD1 x t x t0 / 1 and x t u t .
P

Example 4.5 (AR(1).) Consider estimating ˛ in y t D ˛y t 1 C u t . The LS estimator is

1
T T
!
1X 2 1X
˛O LS D y yt 1yt
T t D1 t 1
T tD1
T
! 1 T
1X 2 1X
D˛C y yt 1ut :
T t D1 t 1 T t D1

In this case, the assumption E.u t jx t s / D 0 for all s (that is, s D :::; 1; 0; 1; :::) is false,
since x t C1 D y t and u t and y t are correlated. We can therefore not use this way of
proving that ˛O LS is unbiased. In fact, it is not, and it can be shown that ˛O LS is downward-
biased if ˛ > 0, and that this bias gets quite severe as ˛ gets close to unity.

57
 The finite sample distribution of the LS estimator is typically unknown.
Even in the most restrictive case where u t is iid N 0;  2 and E.u t jx t s / D 0 for all

s, we can only get that

2
T
! 13
1 X
ˇOLS j fx t gTtD1  N 4ˇ0 ;  2 x t x t0 5: (4.22)
T t D1

This says that the estimator, conditional on the sample of regressors, is normally dis-
tributed. With deterministic x t , this clearly means that ˇOLS is normally distributed in a
small sample. The intuition is that the LS estimator with deterministic regressors is just
a linear combination of the normally distributed y t , so it must be normally distributed.
However, if x t is stochastic, then we have to take into account the distribution of fx t gTtD1
to find the unconditional distribution of ˇOLS . The principle is that
  Z 1   Z 1  
pdf ˇ DO O
pdf ˇ; x dx D O
pdf ˇ jx pdf .x/ dx;
1 1

so the distribution in (4.22) must be multiplied with the probability density function of
fx t gTtD1 and then integrated over fx t gTtD1 to give the unconditional distribution (marginal)
of ˇOLS . This is typically not a normal distribution.
Another way to see the same problem is to note that ˇOLS in (4.19) is a product of
T
two random variables, .˙ tD1 x t x t0 =T / 1 and ˙ tTD1 x t u t =T . Even if u t happened to be
normally distributed, there is no particular reason why x t u t should be, and certainly no
strong reason for why .˙ tTD1 x t x t0 =T / 1 ˙ tTD1 x t u t =T should be.

4.4 Consistency of LS

Reference: Greene (2000) 9.3-5 and 11.2; Hamilton (1994) 8.2; Davidson (2000) 3
We now study if the LS estimator is consistent.

Remark 4.6 Suppose the true parameter value is ˇ0 . The estimator ˇOT (which, of course,
depends on the sample size T ) is said to be consistent if for every " > 0 and ı > 0 there
exists N such that for T  N
 
O
Pr ˇT ˇ0 > ı < ":

58
 p
(kxk D x 0 x, the Euclidean distance of x from zero.) We write this plim ˇOT D ˇ0 or
just plim ˇO D ˇ0 , or perhaps ˇO !p ˇ0 . (For an estimator of a covariance matrix, the
most convenient is to stack the unique elements in a vector and then apply the definition
above.)

Remark 4.7 (Slutsky’s theorem.) If g .:/ is a continuous function, then plim g .zT / D
g .plim zT /. In contrast, note that Eg .zT / is generally not equal to g .EzT /, unless g .:/
is a linear function.

Remark 4.8 (Probability limit of product.) Let xT and yT be two functions of a sample
of length T . If plim xT D a and plim yT D b, then plim xT yT D ab.

Assume
T
1X
plim x t x t0 D ˙xx < 1; and ˙xx invertible. (4.23)
T tD1

The plim carries over to the inverse by Slutsky’s theorem.1 Use the facts above to write
the probability limit of (4.19) as
T
O 1 1X
plim ˇLS D ˇ0 C ˙xx plim xt ut : (4.24)
T t D1

To prove consistency of ˇOLS we therefore have to show that

T
1X
plim x t u t D Ex t u t D Cov.x t ; u t / D 0: (4.25)
T t D1

This is fairly easy to establish in special cases, for instance, when w t D x t u t is iid or
when there is either heteroskedasticity or serial correlation. The case with both serial
correlation and heteroskedasticity is just a bit more complicated. In other cases, it is clear
that the covariance the residuals and the regressors are not all zero—for instance when
some of the regressors are measured with error or when some of them are endogenous
variables.
An example of a case where LS is not consistent is when the errors are autocorrelated
and the regressors include lags of the dependent variable. For instance, suppose the error
1
This puts non-trivial restrictions on the data generating processes. For instance, if x t include lagged
values of y t , then we typically require y t to be stationary and ergodic, and that u t is independent of x t s
for s  0.

59
is a MA(1) process
u t D " t C 1 " t 1; (4.26)

where " t is white noise and that the regression equation is an AR(1)

y t D y t 1 C ut : (4.27)

This is an ARMA(1,1) model and it is clear that the regressor and error in (4.27) are
correlated, so LS is not a consistent estimator of an ARMA(1,1) model.

4.5 Asymptotic Normality of LS

Reference: Greene (2000) 9.3-5 and 11.2; Hamilton (1994) 8.2; Davidson (2000) 3
Remark 4.9 (Continuous mapping theorem.) Let the sequences of random matrices fxT g
d p
and fyT g, and the non-random matrix faT g be such that xT ! x, yT ! y, and aT ! a
d
(a traditional limit). Let g.xT ; yT ; aT / be a continuous function. Then g.xT ; yT ; aT / !
g.x; y; a/. Either of yT and aT could be irrelevant in g.
d
Remark 4.10 From the previous remark: if xT ! x (a random variable) and plim QT D
d
Q (a constant matrix), then QT xT ! Qx.
p
Premultiply (4.19) by T and rearrange as

T
! 1 p T
p   1X T X
T ˇOLS ˇ0 D x t x t0 xt ut : (4.28)
T t D1 T t D1

If the first term on the right hand side converges in probability to a finite matrix (as as-
sumed in (4.23)), and the vector of random variables x t u t satisfies a central limit theorem,
then
p d
T .ˇOLSˇ0 / ! N 0; ˙xx1 S0 ˙xx1 , where


(4.29)
T p T !
1X T X
˙xx D plim x t x t0 and S0 D Cov xt ut :
T t D1 T t D1

The last matrix in the covariance matrix does not need to be transposed since it is sym-
metric (since ˙xx is). This general expression is valid for both autocorrelated and het-
eroskedastic residuals—all such features are loaded into the S0 matrix. Note that S0 is

60
 p
the variance-covariance matrix of T times a sample average (of the vector of random
variables x t u t ), which can be complicated to specify and to estimate. In simple cases,
we can derive what it is. To do so, we typically need to understand the properties of the
residuals. Are they autocorrelated and/or heteroskedastic? In other cases we will have to
use some kind of “non-parametric” approach to estimate it.
A common approach is to estimate ˙xx by ˙ tTD1 x t x t0 =T and use the Newey-West
estimator of S0 .

4.5.1 Special Case: Classical LS assumptions

Reference: Greene (2000) 9.4 or Hamilton (1994) 8.2.

We can recover the classical expression for the covariance,  2 ˙xx1 , if we assume that
the regressors are stochastic, but require that x t is independent of all u t Cs and that u t is
iid. It rules out, for instance, that u t and x t 2 are correlated and also that the variance of
u t depends on x t . Expand the expression for S0 as Expand the expression for S0 as
p T ! p T !
T X T X
S0 D E xt ut u t x t0 (4.30)
T t D1 T t D1
1
D E .::: C xs 1 us 1 C xs us C :::/ ::: C us 1 xs0 1 C us xs0 C ::: :


T
Note that

Ex t s u t s u t x t0 D Ex t s x t0 Eu t s u t (since u t and x t s independent)

(
0 if s ¤ 0 (since Eu t s u t D 0 by iid u t )
D (4.31)
Ex t x t0 Eu t u t else.

This means that all cross terms (involving different observations) drop out and that we
can write
T
1X
S0 D Ex t x t0 Eu2t (4.32)
T tD1
T
1 X2
D E x t x t0 (since u t is iid and  2 D Eu2t ) (4.33)
T t D1
D  2 ˙xx : (4.34)

61
Using this in (4.29) gives
p
Asymptotic CovŒ T .ˇOLS ˇ0 / D ˙xx1 S0 ˙xx1 D ˙xx1  2 ˙xx ˙xx1 D  2 ˙xx1 :

4.5.2 Special Case: White’s Heteroskedasticity

Reference: Greene (2000) 12.2 and Davidson and MacKinnon (1993) 16.2.
This section shows that the classical LS formula for the covariance matrix is valid
even if the errors are heteroskedastic—provided the heteroskedasticity is independent of
the regressors.
The only difference compared with the classical LS assumptions is that u t is now
allowed to be heteroskedastic, but this heteroskedasticity is not allowed to depend on the
moments of x t . This means that (4.32) holds, but (4.33) does not since Eu2t is not the
same for all t.
However, we can still simplify (4.32) a bit more. We assumed that Ex t x t0 and Eu2t
(which can both be time varying) are not related to each other, so we could perhaps multi-
ply Ex t x t0 by ˙ tTD1 Eu2t =T instead of by Eu2t . This is indeed true asymptotically—where
any possible “small sample” relation between Ex t x t0 and Eu2t must wash out due to the
assumptions of independence (which are about population moments).
In large samples we therefore have
T T
! !
1X 2 1X 0
S0 D Eu Ex t x t
T t D1 t T t D1
T T
! !
1X 2 1X
D Eu E x t x t0
T t D1 t T tD1
D ! 2 ˙xx ; (4.35)

where ! 2 is a scalar. This is very similar to the classical LS case, except that ! 2 is
the average variance of the residual rather than the constant variance. In practice, the
estimator of ! 2 is the same as the estimator of  2 , so we can actually apply the standard
LS formulas in this case.
This is the motivation for why White’s test for heteroskedasticity makes sense: if the
heteroskedasticity is not correlated with the regressors, then the standard LS formula is
correct (provided there is no autocorrelation).

62
4.6 Inference

Consider some estimator, ˇOk1 , with an asymptotic normal distribution

p d
T .ˇO ˇ0 / ! N .0; V / : (4.36)

Suppose we want to test the null hypothesis that the s linear restrictions Rˇ0 D r hold,
where R is an s  k matrix and r is an s  1 vector. If the null hypothesis is true, then
p d
T .RˇO r/ ! N.0; RVR0 /; (4.37)

since the s linear combinations are linear combinations of random variables with an
asymptotic normal distribution as in (4.37).

Remark 4.11 If the n  1 vector x  N.0; ˙/, then x 0 ˙ 1

x  2n .

Remark 4.12 From the previous remark and Remark (4.9), it follows that if the n  1
d d
vector x ! N.0; ˙/, then x 0 ˙ 1
x ! 2n .

From this remark, it follows that if the null hypothesis, Rˇ0 D r, is true, then Wald
test statistics converges in distribution to a 2s variable

1 d
T .RˇO r/0 RVR0 .RˇO r/ ! 2s :


(4.38)

Values of the test statistics above the x% critical value of the 2s distribution mean that
we reject the null hypothesis at the x% significance level.
p
When there is only one restriction (s D 1), then T .RˇO r/ is a scalar, so the test
can equally well be based on the fact that
p
T .RˇO r/ d
p ! N.0; 1/:
RVR0
In this case, we should reject the null hypothesis if the test statistics is either very low
(negative) or very high (positive). In particular, let ˚./ be the standard normal cumulative
distribution function. We then reject the null hypothesis at the x% significance level if the
test statistics is below xL such that ˚.xL / D .x=2/% or above xH such that ˚.xH / D
1 .x=2/% (that is with .x=2/% of the probability mass in each tail).

63
Example 4.13 (TR2 =.1 R2 / as a test of the regression.) Recall from (4.15)-(4.16) that
R2 D Var
c .yO t / =Var
c .y t / D 1 Varc .uO t / =Var
c .y t /, where yO t and uO t are the fitted value
and residual respectively. We therefore get

TR2 =.1 R2 / D T Var

c .yO t / =Var
c .uO t / :

To simplify the algebra, assume that both y t and x t are demeaned and that no intercept is
used. (We get the same results, but after more work, if we relax this assumption.) In this
O so we can rewrite the previous equation as
case, yO t D x t0 ˇ,

TR2 =.1 R2 / D T ˇO 0 ˙xx ˇ=

O Var
c .uO t / :

This is identical to (4.38) when R D Ik and r D 0k1 and the classical LS assumptions
are fulfilled (so V D Var .uO t / ˙xx1 /. The TR2 =.1 R2 / is therefore a 2k distributed
statistics for testing if all the slope coefficients are zero. More generally, the test statistic
is distributed as 2k 1 when there are k 1 slopes and one intercept.

Example 4.14 (F version of the test.) There is also an Fk;T k version of the test in the
previous example: .R2 =k/=Œ.1 R2 /=.T k/. Note that k times an Fk;T k variable
converges to a 2k variable as T k ! 1. This means that the 2k form in the previous
example can be seen as an asymptotic version of the (more common) F form. More
generally, the test statistic ŒR2 =.k 1/=Œ.1 R2 /=.T k/ is distributed as Fk 1;T k
when there are k 1 slopes and one intercept.

4.6.1 Tests of Non-Linear Restrictions

To test non-linear restrictions, we can use the delta method which gives the asymptotic
distribution of a function of a random variable.

Fact 4.15 Remark 4.16 (Delta method) Consider an estimator ˇOk1 which satisfies
p   d
T ˇO ˇ0 ! N .0; ˝/ ;

and suppose we want the asymptotic distribution of a transformation of ˇ

q1 D g .ˇ/ ;

64
where g .:/ is has continuous first derivatives. The result is
p h   i d
T g ˇO


g .ˇ0 / ! N 0; qq ; where
@g .ˇ0 / @g .ˇ0 /0 @g .ˇ0 /
D 0 ˝ , where is q  k:
@ˇ @ˇ @ˇ 0

Proof. By the mean value theorem we have

  @g .ˇ  /  O 
g ˇO D g .ˇ0 / C ˇ ˇ0 ;
@ˇ 0
where 2 3
@g1 .ˇ / @g1 .ˇ /
@ˇ1



@ˇk
@g .ˇ/ 6 :: :: :: 7
D6 : : : 7 ;
@ˇ 0 4
@gq .ˇ / @gq .ˇ /
5
@ˇ1



@ˇk qk
p
and we evaluate it at ˇ  which is (weakly) between ˇO and ˇ0 . Premultiply by T and
rearrange as
p h   i @g .ˇ  / p  
T g ˇO g .ˇ0 / D T ˇO ˇ0 .
@ˇ 0
If ˇO is consistent (plim ˇO D ˇ0 ) and @g .ˇ  / =@ˇ 0 is continuous, then by Slutsky’s theorem
plim @g .ˇ  / =@ˇ 0 D @g .ˇ0 / =@ˇ 0 , which is a constant. The result then follows from the
continuous mapping theorem.

4.6.2 On F Tests

F tests are sometimes used instead of chi–square tests. However, F tests rely on very spe-
cial assumptions and typically converge to chi–square tests as the sample size increases.
There are therefore few compelling theoretical reasons for why we should use F tests.2
This section demonstrates that point.

Remark 4.17 If Y1  2n1 , Y2  2n2 ; and if Y1 and Y2 are independent, then Z D

d
.Y1 =n1 /=.Y1 =n1 /  Fn1 ;n2 . As n2 ! 1, n1 Z ! 2n1 (essentially because the denomi-
nator in Z is then equal to its expected value).
2
However, some simulation evidence suggests that F tests may have better small sample properties than
chi-square test.

65
 To use the F test to test s linear restrictions Rˇ0 D r, we need to assume that the
p
small sample distribution of the estimator is normal, T .ˇO ˇ0 /  N.0;  2 W /, where
 2 is a scalar and W a known matrix. This would follow from an assumption that the
residuals are normally distributed and that we either consider the distribution conditional
on the regressors or that the regressors are deterministic. In this case W D ˙xx1 .
Consider the test statistics
1
F D T .RˇO r/0 RO 2 WR0 .RˇO


r/=s:

This is similar to (4.38), expect that we use the estimated covariance matrix O 2 W instead
of the true  2 W (recall, W is assumed to be known) and that we have divided by the
number of restrictions, s. Multiply and divide this expressions by  2
1
T .RˇO r/0 R 2 WR0 .RˇO


r/=s
F D :
O 2 = 2

The numerator is an 2s variable divided by its degrees of freedom, s. The denominator
can be written O 2 = 2 D ˙.uO t =/2 =T , where uO t are the fitted residuals. Since we just
assumed that u t are iid N.0;  2 /, the denominator is an 2T variable divided by its degrees
of freedom, T . It can also be shown that the numerator and denominator are independent
(essentially because the fitted residuals are orthogonal to the regressors), so F is an Fs;T
variable.
We need indeed very strong assumptions to justify the F distributions. Moreover, as
d
T ! 1, sF ! 2n which is the Wald test—which do not need all these assumptions.

4.7 Diagnostic Tests of Autocorrelation, Heteroskedasticity, and Normality

Reference: Greene (2000) 12.3, 13.5 and 9.7; Johnston and DiNardo (1997) 6; and Pindyck
and Rubinfeld (1998) 6, Patterson (2000) 5
LS and IV are still consistent even if the residuals are autocorrelated, heteroskedastic,
and/or non-normal, but the traditional expression for the variance of the parameter esti-
mators is invalid. It is therefore important to investigate the properties of the residuals.
We would like to test the properties of the true residuals, u t , but these are unobserv-
able. We can instead use residuals from a consistent estimator as approximations, since
the approximation error then goes to zero as the sample size increases. The residuals from

66
an estimator are

uO t D y t x t0 ˇO
 
D x t0 ˇ0 ˇO C u t : (4.39)

If plim ˇO D ˇ0 , then uO t converges in probability to the true residual (“pointwise consis-

tency”). It therefore makes sense to use uO t to study the (approximate) properties of u t .
We want to understand if u t are autocorrelated and/or heteroskedastic, since this affects
the covariance matrix of the least squares estimator and also to what extent least squares is
efficient. We might also be interested in studying if the residuals are normally distributed,
since this also affects the efficiency of least squares (remember that LS is MLE is the
residuals are normally distributed).
It is important that the fitted residuals used in the diagnostic tests are consistent. With
poorly estimated residuals, we can easily find autocorrelation, heteroskedasticity, or non-
normality even if the true residuals have none of these features.

4.7.1 Autocorrelation

Let Os be the estimate of the sth autocorrelation coefficient of some variable, for instance,
the fitted residuals. The sampling properties of Os are complicated, but there are several
useful large sample results for Gaussian processes (these results typically carry over to
processes which are similar to the Gaussian—a homoskedastic process with finite 6th
moment is typically enough). When the true autocorrelations are all zero (not 0 , of
course), then for any i and j different from zero
" # " # " #!
p Oi 0 1 0
T !d N ; : (4.40)
Oj 0 0 1

This result can be used to construct tests for both single autocorrelations (t-test or 2 test)
and several autocorrelations at once (2 test).

Example 4.18 (t-test) We want to test the hypothesis that 1 D 0. Since the N.0; 1/
distribution has 5% of the probability mass below -1.65 and another 5% above 1.65, we
p
can reject the null hypothesis at the 10% level if T jO1 j > 1:65. With T D 100, we
p
therefore need jO1 j > 1:65= 100 D 0:165 for rejection, and with T D 1000 we need

67
 p
jO1 j > 1:65= 1000  0:0:53.
p
The Box-Pierce test follows directly from the result in (4.40), since it shows that T Oi
p
and T Oj are iid N(0,1) variables. Therefore, the sum of the square of them is distributed
as an 2 variable. The test statistics typically used is
L
X
QL D T Os2 !d L
2
: (4.41)
sD1

Example 4.19 (Box-Pierce) Let O1 D 0:165, and T D 100, so Q1 D 100  0:1652 D
2:72. The 10% critical value of the 21 distribution is 2.71, so the null hypothesis of no
autocorrelation is rejected.

The choice of lag order in (4.41), L, should be guided by theoretical considerations,

but it may also be wise to try different values. There is clearly a trade off: too few lags may
miss a significant high-order autocorrelation, but too many lags can destroy the power of
the test (as the test statistics is not affected much by increasing L, but the critical values
increase).

Example 4.20 (Residuals follow an AR(1)process) If u t D 0:9u t 1 C " t , then the true
autocorrelation coefficients are j D 0:9j .

A common test of the serial correlation of residuals from a regression is the Durbin-
Watson test
d D 2 .1 O1 / ; (4.42)

where the null hypothesis of no autocorrelation is


not rejected if d > dupper

rejected if d < dlower (in favor of positive autocorrelation)
else inconclusive

where the upper and lower critical values can be found in tables. (Use 4 d to let
negative autocorrelation be the alternative hypothesis.) This test is typically not useful
when lagged dependent variables enter the right hand side (d is biased towards showing
no autocorrelation). Note that DW tests only for first-order autocorrelation.

68
Example 4.21 (Durbin-Watson.) With O1 D 0:2 we get d D 1:6. For large samples,

the 5% critical value is dlower  1:6, so O1 > 0:2 is typically considered as evidence of
positive autocorrelation.

The fitted residuals used in the autocorrelation tests must be consistent in order to in-
terpret the result in terms of the properties of the true residuals. For instance, an excluded
autocorrelated variable will probably give autocorrelated fitted residuals—and also make
the coefficient estimator inconsistent (unless the excluded variable is uncorrelated with
the regressors). Only when we know that the model is correctly specified can we interpret
a finding of autocorrelated residuals as an indication of the properties of the true residuals.

4.7.2 Heteroskedasticity

Remark 4.22 (Kronecker product.) If A and B are matrices, then

2 3
a11 B


a1n B
6 : :: 7
A˝B D6 ::
4 : 75:
am1 B


amn B

Example 4.23 Let x1 and x2 be scalars. Then

2 " # 3 2 3
x1 x1 x1
" # " # 6 x1 7 6 7
x1 x1 6
" x2 #
7 6 x1 x2 7
˝ D6 7D6 7:
x2 x2 6 x1 7 6 x2 x1 7
x2
4 5 4 5
x2 x2 x2

White’s test for heteroskedasticity tests the null hypothesis of homoskedasticity against
the kind of heteroskedasticity which can be explained by the levels, squares, and cross
products of the regressors. Let w t be the unique elements in x t ˝ x t , where we have
added a constant to x t if there was not one from the start. Run a regression of the squared
fitted LS residuals on w t
uO 2t D w t0 C " t (4.43)

and test if all elements (except the constant) in are zero (with a 2 or F test). The
reason for this specification is that if u2t is uncorrelated with x t ˝ x t , then the usual LS
covariance matrix applies.

69
 Breusch-Pagan’s test is very similar, except that the vector w t in (4.43) can be any
vector which is thought of as useful for explaining the heteroskedasticity. The null hy-
pothesis is that the variance is constant, which is tested against the alternative that the
variance is some function of w t .
The fitted residuals used in the heteroskedasticity tests must be consistent in order to
interpret the result in terms of the properties of the true residuals. For instance, if some
of the of elements in w t belong to the regression equation, but are excluded, then fitted
residuals will probably fail these tests.

4.7.3 Normality

We often make the assumption of normally distributed errors, for instance, in maximum
likelihood estimation. This assumption can be tested by using the fitted errors. This works
since moments estimated from the fitted errors are consistent estimators of the moments
of the true errors. Define the degree of skewness and excess kurtosis for a variable z t
(could be the fitted residuals) as
T
O 1X
3 D .z t zN /3 =O 3 ; (4.44)
T t D1
T
1X
O4 D .z t zN /4 =O 4 3; (4.45)
T t D1

where zN is the sample mean and O 2 is the estimated variance.

Remark 4.24 (2 .n/ distribution.) If xi are independent N.0; i2 / variables, then ˙inD1 xi2 =i2 
2 .n/.

In a normal distribution, the true values are zero and the test statistics O3 and O4 are
themselves normally distributed with zero covariance and variances 6=T and 24=T , re-
spectively (straightforward, but tedious, to show). Therefore, under the null hypothesis
of a normal distribution, T O32 =6 and T O42 =24 are independent and both asymptotically
distributed as 2 .1/, so the sum is asymptotically a 2 .2/ variable
 
O 2 O
W D T 3 =6 C 4 =24 !d 2 .2/:
2
(4.46)

This is the Jarque and Bera test of normality.

70
 Histogram of 100 draws from a U(0,1) distribution
0.2
θ3 = −0.14, θ4 = −1.4, W = 8

0.15

0.1

0.05

0
0 0.2 0.4 0.6 0.8 1
Figure 4.1: This figure shows a histogram from 100 draws of iid uniformly [0,1] dis-
tributed variables.

Bibliography
Davidson, J., 2000, Econometric theory, Blackwell Publishers, Oxford.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Hayashi, F., 2000, Econometrics, Princeton University Press.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

Patterson, K., 2000, An introduction to applied econometrics: a time series approach,

MacMillan Press, London.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

71
 Average LS estimate of ρ Std of LS estimate of ρ

simulation
0.9 asymptotic
0.1
0.85
0.8
0.05
0.75
0.7 0
0 100 200 300 400 500 0 100 200 300 400 500
Sample size, T Sample size, T

√
T × Std of LS estimate of ρ True model: yt = 0.9yt−1 + ǫt ,
0.8 where ǫt is iid N(0,2)

0.7 Estimated model: yt = a + ρyt−1 + ut

0.6 Number of simulations: 25000

0.5

0.4
0 100 200 300 400 500
Sample size, T

Figure 4.2: Distribution of LS estimator of autoregressive parameter.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

72
 Distribution of LS t-stat, T = 5 Distribution of LS t-stat, T = 100
t = (b̂ − 0.9)/Std(b̂)
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−4 −2 0 2 4 −4 −2 0 2 4

Model: Rt = 0.9ft + ǫt , ǫt = vt − 2,
where vt has a χ22 distribution Probability density functions

Estimated model: yt = a + bft + ut N(0,1)

0.4
Number of simulations: 25000 χ22 − 2
0.3
T = 5 T = 100 0.2
Kurtosis of t-stat: 46.753 3.049
Frequency of |t-stat| > 1.65 0.294 0.105 0.1
Frequency of |t-stat| > 1.96 0.227 0.054 0
−4 −2 0 2 4

Figure 4.3: Distribution of LS estimator when residuals have a t3 distribution.

73
5 Instrumental Variable Method
Reference: Greene (2000) 9.5 and 16.1-2
Additional references: Hayashi (2000) 3.1-4; Verbeek (2004) 5.1-4; Hamilton (1994) 8.2;
and Pindyck and Rubinfeld (1998) 7

5.1 Consistency of Least Squares or Not?

Consider the linear model

y t D x t0 ˇ0 C u t ; (5.1)

where y t and u t are scalars, x t a k  1 vector, and ˇ0 is a k  1 vector of the true

coefficients. The least squares estimator is
1
T T
!
1X 1X
ˇOLS D x t x t0 xt yt (5.2)
T t D1 T t D1
T
! 1 T
1X 1X
D ˇ0 C x t x t0 xt ut ; (5.3)
T t D1 T t D1

where we have used (5.1) to substitute for y t . The probability limit is

1
T T
!
1X 1X
plim ˇOLS ˇ0 D plim x t x t0 plim xt ut : (5.4)
T t D1 T t D1

In many cases the law of large numbers applies to both terms on the right hand side. The
first term is typically a matrix with finite elements and the second term is the covariance of
the regressors and the true residuals. This covariance must be zero for LS to be consistent.

5.2 Reason 1 for IV: Measurement Errors

Reference: Greene (2000) 9.5.

74
 Suppose the true model is
y t D x t0 ˇ0 C ut : (5.5)

Data on y t and x t is not directly observable, so we instead run the regression

y t D x t0 ˇ C u t ; (5.6)

where y t and x t are proxies for the correct variables (the ones that the model is true for).
We can think of the difference as measurement errors

y t D y t C v ty and (5.7)
x t D x t C v tx ; (5.8)

where the errors are uncorrelated with the true values and the “true” residual ut .
Use (5.7) and (5.8) in (5.5)

0
yt v ty D x t v tx ˇ0 C ut or
y t D x t0 ˇ0 C " t where " t D v tx0 ˇ0 C v ty C ut : (5.9)

Suppose that x t is a measured with error. From (5.8) we see that v tx and x t are corre-
lated, so LS on (5.9) is inconsistent in this case. To make things even worse, measurement
errors in only one of the variables typically affect all the coefficient estimates.
To illustrate the effect of the error, consider the case when x t is a scalar. Then, the
probability limit of the LS estimator of ˇ in (5.9) is

plim ˇOLS D Cov .y t ; x t / =Var .x t /

D Cov x t ˇ0 C ut ; x t =Var .x t /

D Cov x t ˇ0 v tx ˇ0 C ut ; x t =Var .x t /

Cov .x t ˇ0 ; x t / C Cov v tx ˇ0 ; x t C Cov ut ; x t

D
Var .x t /
Cov v tx ˇ0 ; x t v tx


Var .x t /
D ˇ0 C
Var .x t / Var .x t /
x


D ˇ0 ˇ0 Var v t =Var .x t /
" #
Var v tx


D ˇ0 1 : (5.10)
Var .x t / C Var .v tx /

75
since x t and v tx are uncorrelated with ut and with each other. This shows that ˇOLS goes
to zero as the measurement error becomes relatively more volatile compared with the true
value. This makes a lot of sense, since when the measurement error is very large then the
regressor x t is dominated by noise that has nothing to do with the dependent variable.
Suppose instead that only y t is measured with error. This not a big problem since this
measurement error is uncorrelated with the regressor, so the consistency of least squares
is not affected. In fact, a measurement error in the dependent variable is like increasing
the variance in the residual.

5.3 Reason 2 for IV: Simultaneous Equations Bias (and Inconsis-

tency)

Suppose economic theory tells you that the structural form of the m endogenous variables,
y t , and the k predetermined (exogenous) variables, z t , is

F y t C Gz t D u t , where u t is iid with Eu t D 0 and Cov .u t / D ˙; (5.11)

where F is m  m, and G is m  k. The disturbances are assumed to be uncorrelated with

the predetermined variables, E.z t u0t / D 0.
Suppose F is invertible. Solve for y t to get the reduced form

1 1
yt D F Gz t C F ut (5.12)
D ˘ z t C " t , with Cov ." t / D ˝: (5.13)

The reduced form coefficients, ˘ , can be consistently estimated by LS on each equation

since the exogenous variables z t are uncorrelated with the reduced form residuals (which
are linear combinations of the structural residuals). The fitted residuals can then be used
to get an estimate of the reduced form covariance matrix.
The j th line of the structural form (5.11) can be written

Fj y t C Gj z t D ujt ; (5.14)

where Fj and Gj are the j th rows of F and G, respectively. Suppose the model is nor-
malized so that the coefficient on yjt is one (otherwise, divide (5.14) with this coefficient).

76
Then, rewrite (5.14) as

yjt D Gj1 zQ t Fj1 yQ t C ujt

D x t0 ˇ C ujt , where x t0 D zQ t0 ; yQ t0 ;
 
(5.15)

where zQ t and yQ t are the exogenous and endogenous variables that enter the j th equation,
which we collect in the x t vector to highlight that (5.15) looks like any other linear re-
gression equation. The problem with (5.15), however, is that the residual is likely to be
correlated with the regressors, so the LS estimator is inconsistent. The reason is that a
shock to ujt influences yjt , which in turn will affect some other endogenous variables in
the system (5.11). If any of these endogenous variable are in x t in (5.15), then there is a
correlation between the residual and (some of) the regressors.
Note that the concept of endogeneity discussed here only refers to contemporaneous
endogeneity as captured by off-diagonal elements in F in (5.11). The vector of predeter-
mined variables, z t , could very well include lags of y t without affecting the econometric
endogeneity problem.

Example 5.1 (Supply and Demand. Reference: GR 16, Hamilton 9.1.) Consider the
simplest simultaneous equations model for supply and demand on a market. Supply is

q t D p t C ust ; > 0;

and demand is
q t D ˇp t C ˛A t C udt ; ˇ < 0;

where A t is an observable demand shock (perhaps income). The structural form is there-
fore " #" # " # " #
1 qt 0 ust
C At D :
1 ˇ pt ˛ udt
The reduced form is " # " # " #
qt 11 "1t
D At C :
pt 21 "2t
If we knew the structural form, then we can solve for q t and p t to get the reduced form in

77
terms of the structural parameters
" # " # " #" #
ˇ
qt ˇ
˛ ˇ ˇ ust
D 1
At C 1 1
:
pt ˇ
˛ ˇ ˇ
udt

Example 5.2 (Supply equation with LS.) Suppose we try to estimate the supply equation
in Example 5.1 by LS, that is, we run the regression

q t D p t C " t :

If data is generated by the model in Example 5.1, then the reduced form shows that p t is
correlated with ust , so we cannot hope that LS will be consistent. In fact, when both q t
and p t have zero means, then the probability limit of the LS estimator is

Cov .q t ; p t /
plim O D
Var .p t /
 
Cov ˛ˇ A t C ˇ udt ˇ
us ; ˛ A C 1 ˇ udt
ˇ t ˇ t
1
ud
ˇ t
D  
Var ˛ ˇ A t C 1 ˇ udt 1
u
ˇ t
s
;

where the second line follows from the reduced form. Suppose the supply and demand
shocks are uncorrelated. In that case we get
˛2
udt C ˇ 2 Var ust



Var .A t / C Var
. ˇ /2 . ˇ /2 . ˇ /
plim O D ˛2 1
Var u t C 1 2 Var .ust /
d


Var .A t / C
. ˇ /2 . ˇ /2 . ˇ /
2 d
˛ Var .A t / C Var u t C ˇVar ust



D :
˛ 2 Var .A t / C Var udt C Var .ust /

First, suppose the supply shocks are zero, Var ust D 0, then plim O D , so we indeed

estimate the supply elasticity, as we wanted. Think of a fixed supply curve, and a demand
curve which moves around. These point of p t and q t should trace out the supply curve. It
is clearly ust that causes a simultaneous equations problem in estimating the supply curve:
ust affects both q t and p t and the latter is the regressor in the supply equation. With no
movements in ust there is no correlation between the shock and the regressor. Second, now
suppose instead that the both demand shocks are zero (both A t D 0 and Var udt D 0).

Then plim O D ˇ, so the estimated value is not the supply, but the demand elasticity. Not
good. This time, think of a fixed demand curve, and a supply curve which moves around.

78
 Distribution of LS estimator, T=200 Distribution of LS estimator, T=900

0.88 0.89
0.4 Mean and std:
0.03
0.4 0.01

0.2 0.2

0 0
0.6 0.8 1 0.6 0.8 1

True model: yt = ρyt−1 + εt, where εt = νt + θνt−1,

where ρ=0.8, θ=0.5 and νt is iid N(0,2)

Estimated model: yt = ρyt−1 + ut

Figure 5.1: Distribution of LS estimator of autoregressive parameter

Example 5.3 (A flat demand curve.) Suppose we change the demand curve in Example
5.1 to be infinitely elastic, but to still have demand shocks. For instance, the inverse
demand curve could be p t D A t C uD t . In this case, the supply and demand is no longer
a simultaneous system of equations and both equations could be estimated consistently
with LS. In fact, the system is recursive, which is easily seen by writing the system on
vector form " #" # " # " #
1 0 pt uDt
C At D :
1 qt 0 ust
A supply shock, ust , affects the quantity, but this has no affect on the price (the regressor
in the supply equation), so there is no correlation between the residual and regressor in
the supply equation. A demand shock, uD t , affects the price and the quantity, but since
quantity is not a regressor in the inverse demand function (only the exogenous A t is) there
is no correlation between the residual and the regressor in the inverse demand equation
either.

79
 LS, T=200 IV, T=200 ML, T=200

0.88 0.78 0.78

0.4 0.03
0.4 0.05
0.4 0.05

0.2 0.2 0.2

0 0 0
0.6 0.8 1 0.6 0.8 1 0.6 0.8 1

LS, T=900 IV, T=900 ML, T=900

0.89 0.80 0.80

0.4 0.01
0.4 0.02
0.4 0.02

0.2 0.2 0.2

0 0 0
0.6 0.8 1 0.6 0.8 1 0.6 0.8 1

Figure 5.2: Distribution of LS, IV and ML estimators of autoregressive parameter. See

Figure 5.1 for details.

5.4 Definition of the IV Estimator—Consistency of IV

Reference: Greene (2000) 9.5; Hamilton (1994) 8.2; and Pindyck and Rubinfeld (1998)
7.
Consider the linear model
y t D x t0 ˇ0 C u t ; (5.16)

where y t is a scalar, x t a k  1 vector, and ˇ0 is a vector of the true coefficients. If

we suspect that x t and u t in (5.16) are correlated, then we may use the instrumental
variables (IV) method. To do that, let z t be a k  1 vector of instruments (as many
instruments as regressors; we will later deal with the case when we have more instruments
than regressors.) If x t and u t are not correlated, then setting x t D z t gives the least
squares (LS) method.

80
 Recall that LS minimizes the variance of the fitted residuals, uO t D y t x t0 ˇOLS . The
first order conditions for that optimization problem are
T
1X  
0kx1 D xt yt x t0 ˇOLS : (5.17)
T t D1

If x t and u t are correlated, then plim ˇOLS ¤ ˇ0 . The reason is that the probability limit of
the right hand side of (5.17) is Cov.x t ; y t x t0 ˇOLS /, which at ˇOLS D ˇ0 is non-zero, so the
first order conditions (in the limit) cannot be satisfied at the true parameter values. Note
that since the LS estimator by construction forces the fitted residuals to be uncorrelated
with the regressors, the properties of the LS residuals are of little help in deciding if to
use LS or IV.
The idea of the IV method is to replace the first x t in (5.17) with a vector (of similar
size) of some instruments, z t . The identifying assumption of the IV method is that the
instruments are uncorrelated with the residuals (and, as we will see, correlated with the
regressors)

0kx1 D Ez t u t (5.18)
x t0 ˇ0 :


D Ez t y t (5.19)

The intuition is that the linear model (5.16) is assumed to be correctly specified: the
residuals, u t , represent factors which we cannot explain, so z t should not contain any
information about u t .
The sample analogue to (5.19) defines the IV estimator of ˇ as1
T
1X  
0kx1 D zt yt x t0 ˇOI V ; or (5.20)
T t D1
1
T T
!
1X 1X
ˇOI V D z t x t0 zt yt : (5.21)
T t D1 T t D1

It is clearly necessay for ˙z t x t0 =T to have full rank to calculate the IV estimator.

Remark 5.4 (Probability limit of product) For any random variables yT and xT where
plim yT D a and plim xT D b (a and b are constants), we have plim yT xT D ab.
1
1
In matrix notation where z t0 is the t t h row of Z we have ˇOI V D .Z 0 X=T / .Z 0 Y =T /.

81
 To see if the IV estimator is consistent, use (5.16) to substitute for y t in (5.20) and
take the probability limit
T T T
1X 1X 1X
plim 0
z t x t ˇ0 C plim z t u t D plim z t x t0 ˇOI V : (5.22)
T t D1 T t D1 T t D1

Two things are required for consistency of the IV estimator, plim ˇOI V D ˇ0 . First, that
plim ˙ z t u t =T D 0. Provided a law of large numbers apply, this is condition (5.18).
Second, that plim ˙z t x t0 =T has full rank. To see this, suppose plim ˙z t u t =T D 0 is
satisfied. Then, (5.22) can be written
T
!
1X  
plim z t x t0 ˇ0 plim ˇOI V D 0: (5.23)
T t D1

If plim ˙z t x t0 =T has reduced rank, then plim ˇOI V does not need to equal ˇ0 for (5.23) to
be satisfied. In practical terms, the first order conditions (5.20) do then not define a unique
value of the vector of estimates. If a law of large numbers applies, then plim ˙z t x t0 =T D
Ez t x t0 . If both z t and x t contain constants (or at least one of them has zero means), then
a reduced rank of Ez t x t0 would be a consequence of a reduced rank of the covariance
matrix of the stochastic elements in z t and x t , for instance, that some of the instruments
are uncorrelated with all the regressors. This shows that the instruments must indeed be
correlated with the regressors for IV to be consistent (and to make sense).

Remark 5.5 (Second moment matrix) Note that Ezx 0 D EzEx 0 C Cov.z; x/. If Ez D 0
and/or Ex D 0, then the second moment matrix is a covariance matrix. Alternatively,
suppose both z and x contain constants normalized to unity: z D Œ1; zQ 0 0 and x D Œ1; xQ 0 0
where zQ and xQ are random vectors. We can then write
" # " #
1 h i 0 0
Ezx 0 D 1 ExQ 0 C
EQz Q x/
0 Cov.z; Q
" #
1 ExQ 0
D :
EQz EQz ExQ 0 C Cov.z;
Q x/
Q

For simplicity, suppose zQ and xQ are scalars. Then Ezx 0 has reduced rank if Cov.z;
Q x/
Q D 0,
0
Q x/
since Cov.z; Q is then the determinant of Ezx . This is true also when zQ and xQ are vectors.

82
Example 5.6 (Supply equation with IV.) Suppose we try to estimate the supply equation
in Example 5.1 by IV. The only available instrument is A t , so (5.21) becomes
1
T T
!
1X 1X
OI V D At pt At qt ;
T t D1 T t D1

so the probability limit is

1
plim OI V D Cov .A t ; p t / Cov .A t ; q t / ;

since all variables have zero means. From the reduced form in Example 5.1 we see that
1
Cov .A t ; p t / D ˛Var .A t / and Cov .A t ; q t / D ˛Var .A t / ;
ˇ ˇ
so
  1  
1
plim OI V D ˛Var .A t / ˛Var .A t /
ˇ ˇ
D :

This shows that OI V is consistent.

5.4.1 Asymptotic Normality of IV

Little is known about the finite sample distribution of the IV estimator, so we focus on the
asymptotic distribution—assuming the IV estimator is consistent.
d
Remark 5.7 If xT ! x (a random variable) and plim QT D Q (a constant matrix),
d
then QT xT ! Qx.

Use (5.16) to substitute for y t in (5.20)

1
T T
!
1X 1X
ˇOI V D ˇ0 C z t x t0 zt ut : (5.24)
T t D1 T t D1
p
Premultiply by T and rearrange as

T
! 1 p T
p 1X T X
T .ˇOI V ˇ0 / D z t x t0 zt ut : (5.25)
T tD1 T tD1

83
If the first term on the right hand side converges in probability to a finite matrix (as as-
sumed in in proving consistency), and the vector of random variables z t u t satisfies a
central limit theorem, then
p d
T .ˇOI Vˇ0 / ! N 0; ˙zx1 S0 ˙xz1 , where


(5.26)
T p T !
1X T X
˙zx D z t x t0 and S0 D Cov zt ut :
T t D1 T t D1
0
The last matrix in the covariance matrix follows from .˙zx1 /0 D .˙zx / 1 D ˙xz1 . This
general expression is valid for both autocorrelated and heteroskedastic residuals—all such
features are loaded into the S0 matrix. Note that S0 is the variance-covariance matrix of
p
T times a sample average (of the vector of random variables x t u t ).

Example 5.8 (Choice of instrument in IV, simplest case) Consider the simple regression

y t D ˇ1 x t C u t :

The asymptotic variance of the IV estimator is

p T
!
p T X
AVar. T .ˇOI V ˇ0 // D Var z t u t = Cov .z t ; x t /2
T t D1
p
If z t and u t is serially uncorrelated and independent of each other, then Var.˙ tTD1 z t u t = T / D
Var.z t / Var.u t /. We can then write
p Var.z t / Var.u t /
AVar. T .ˇOI V ˇ0 // D Var.u t / 2
D :
Cov .z t ; x t / Var.x t /Corr .z t ; x t /2
An instrument with a weak correlation with the regressor gives an imprecise estimator.
With a perfect correlation, then we get the precision of the LS estimator (which is precise,
but perhaps not consistent).

5.4.2 2SLS

Suppose now that we have more instruments, z t , than regressors, x t . The IV method does
not work since, there are then more equations than unknowns in (5.20). Instead, we can
use the 2SLS estimator. It has two steps. First, regress all elements in x t on all elements

84
in z t with LS. Second, use the fitted values of x t , denoted xO t , as instruments in the IV
method (use xO t in place of z t in the equations above). In can be shown that this is the
most efficient use of the information in z t . The IV is clearly a special case of 2SLS (when
z t has the same number of elements as x t ).
It is immediate from (5.22) that 2SLS is consistent under the same condiditons as
IV since xO t is a linear function of the instruments, so plim TtD1 xO t u t =T D 0, if all the
P

instruments are uncorrelated with u t .

The name, 2SLS, comes from the fact that we get exactly the same result if we replace
the second step with the following: regress y t on xO t with LS.
Example 5.9 (Supply equation with 2SLS.). With only one instrument, A t , this is the
same as Example 5.6, but presented in another way. First, regress p t on A t

Cov .p t ; A t / 1
p t D ıA t C u t ) plim ıOLS D D ˛:
Var .A t / ˇ
Construct the predicted values as

pO t D ıOLS A t :

Second, regress q t on pO t

q t D pO t C e t , with plim O2SLS D plim

b
Cov .q t ; pO t /
:
c .pO t /
Var

Use pO t D ıOLS A t and Slutsky’s theorem

b
plim Cov q t ; ıOLS A t


plim O2SLS D  
c ıOLS A t
plim Var
Cov .q t ; A t / plim ıOLS
D
Var .A t / plim ıOLS
2
h ih i
1
ˇ
˛Var .A t / ˇ
˛
D h i2
Var .A t / ˇ 1 ˛

D :

Note that the trick here is to suppress some the movements in p t . Only those movements

85
that depend on A t (the observable shifts of the demand curve) are used. Movements in p t
which are due to the unobservable demand and supply shocks are disregarded in pO t . We
know from Example 5.2 that it is the supply shocks that make the LS estimate of the supply
curve inconsistent. The IV method suppresses both them and the unobservable demand
shock.

5.5 Hausman’s Specification Test

Reference: Greene (2000) 9.5

This test is constructed to test if an efficient estimator (like LS) gives (approximately)
the same estimate as a consistent estimator (like IV). If not, the efficient estimator is most
likely inconsistent. It is therefore a way to test for the presence of endogeneity and/or
measurement errors.
Let ˇOe be an estimator that is consistent and asymptotically efficient when the null
hypothesis, H0 , is true, but inconsistent when H0 is false. Let ˇOc be an estimator that is
consistent under both H0 and the alternative hypothesis. When H0 is true, the asymptotic
distribution is such that    
Cov ˇOe ; ˇOc D Var ˇOe : (5.27)

Proof. Consider the estimator ˇOc C .1 / ˇOe , which is clearly consistent under H0
since both ˇOc and ˇOe are. The asymptotic variance of this estimator is
     
2 Var ˇOc C .1 /2 Var ˇOe C 2 .1 / Cov ˇOc ; ˇOe ;

which is minimized at  D 0 (since ˇOe is asymptotically efficient). The first order condi-
tion with respect to 
     
2Var ˇOc 2 .1 / Var ˇOe C 2 .1 2/ Cov ˇOc ; ˇOe D 0

should therefore be zero at  D 0 so

   
Var ˇOe D Cov ˇOc ; ˇOe :

(See Davidson (2000) 8.1)

86
 This means that we can write
       
O O O
Var ˇe ˇc D Var ˇe C Var ˇc O O O
2Cov ˇe ; ˇc
   
D Var ˇOc Var ˇOe : (5.28)

We can use this to test, for instance, if the estimates from least squares (ˇOe , since LS
is efficient if errors are iid normally distributed) and instrumental variable method (ˇOc ,
since consistent even if the true residuals are correlated with the regressors) are the same.
In this case, H0 is that the true residuals are uncorrelated with the regressors.
All we need for this test are the point estimates and consistent estimates of the vari-
ance matrices. Testing one of the coefficient can be done by a t test, and testing all the
parameters by a 2 test
 0   1  
ˇOe ˇOc Var ˇOe ˇOc ˇOe ˇOc  2 .j / ; (5.29)

where j equals the number of regressors that are potentially endogenous or measured
with error. Note that the covariance matrix in (5.28) and (5.29) is likely to have a reduced
rank, so the inverse needs to be calculated as a generalized inverse.

5.6 Tests of Overidentifying Restrictions in 2SLS

When we use 2SLS, then we can test if instruments affect the dependent variable only
via their correlation with the regressors. If not, something is wrong with the model since
some relevant variables are excluded from the regression.

Bibliography
Davidson, J., 2000, Econometric theory, Blackwell Publishers, Oxford.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Hayashi, F., 2000, Econometrics, Princeton University Press.

87
Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

88
6 Simulating the Finite Sample Properties
Reference: Greene (2000) 5.3 and Horowitz (2001)
Additional references: Cochrane (2001) 15.2; Davidson and MacKinnon (1993) 21; Davi-
son and Hinkley (1997); Efron and Tibshirani (1993) (bootstrapping, chap 9 in particular);
and Berkowitz and Kilian (2000) (bootstrapping in time series models)
We know the small sample properties of regression coefficients in linear models with
fixed regressors and iid normal error terms. Monte Carlo simulations and bootstrapping
are two common techniques used to understand the small sample properties when these
conditions are not satisfied.
How they should be implemented depends crucially on the properties of the model
and data: if the residuals are autocorrelated, heteroskedastic, or perhaps correlated across
regressions equations. These notes summarize a few typical cases.
The need for using Monte Carlos or bootstraps varies across applications and data
sets. For a case where it is not needed, see Figure 6.1.

6.1 Monte Carlo Simulations

6.1.1 Monte Carlo Simulations in the Simplest Case

Monte Carlo simulations is essentially a way to generate many artificial (small) samples
from a parameterized model and then estimating the statistic on each of those samples.
The distribution of the statistic is then used as the small sample distribution of the estima-
tor.
The following is an example of how Monte Carlo simulations could be done in the
special case of a linear model with a scalar dependent variable

y t D x t0 ˇ C u t ; (6.1)

where u t is iid N.0;  2 / and x t is stochastic but independent of u t ˙s for all s. This means
that x t cannot include lags of y t .
Suppose we want to find the small sample distribution of a function of the estimate,

89
 alpha t LS t NW t boot
US industry portfolios, 1970:1-2012:12 all NaN NaN NaN NaN
15 A (NoDur) 3.62 2.72 2.71 2.70
Mean excess return

B (Durbl) -1.21 -0.58 -0.59 -0.59

C (Manuf ) 0.70 0.72 0.71 0.69
10 D (Enrgy) 4.06 1.80 1.81 1.81
D
A E (HiTec) -1.82 -1.00 -1.00 -0.98
GC
I FH JB E F (Telcm) 1.82 1.07 1.06 1.05
5 G (Shops) 1.37 0.94 0.94 0.94
H (Hlth ) 2.13 1.22 1.24 1.24
0 I (Utils) 2.87 1.61 1.58 1.56
0 0.5 1 1.5 J (Other) -0.65 -0.61 -0.60 -0.61
β (against the market)
NW uses 1 lag
The bootstrap samples pairs of (yt , xt )
3000 simulations

Figure 6.1: CAPM, US industry portfolios, different t-stats

O To do a Monte Carlo experiment, we need information on (i) the coefficients ˇ; (ii)

g.ˇ/.
the variance of u t ;  2 ; (iii) and a process for x t .
The process for x t is typically estimated from the data on x t (for instance, a VAR
system x t D A1 x t 1 C A2 x t 2 C e t ). Alternatively, we could simply use the actual
sample of x t and repeat it.
The values of ˇ and  2 are often a mix of estimation results and theory. In some
case, we simply take the point estimates. In other cases, we adjust the point estimates
so that g.ˇ/ D 0 holds, that is, so you simulate the model under the null hypothesis
in order to study the size of asymptotic tests and to find valid critical values for small
samples. Alternatively, you may simulate the model under an alternative hypothesis in
order to study the power of the test using either critical values from either the asymptotic
distribution or from a (perhaps simulated) small sample distribution.
To make it a bit concrete, suppose you want to use these simulations to get a 5%
critical value for testing the null hypothesis g.ˇ/ D 0. The Monte Carlo experiment
follows these steps.

1. Construct an artificial sample of the regressors (see above), xQ t for t D 1; : : : ; T .

Draw random numbers uQ t for t D 1; : : : ; T and use those together with the artificial

90
 sample of xQ t to calculate an artificial sample yQ t for t D 1; : : : ; T from

yQ t D xQ t0 ˇ C uQ t ; (6.2)

by using the prespecified values of the coefficients ˇ.

2. Calculate an estimate ˇO and record it along with the value of g.ˇ/

O and perhaps also
the test statistic of the hypothesis that g.ˇ/ D 0.

3. Repeat the previous steps N (3000, say) times. The more times you repeat, the
better is the approximation of the small sample distribution.

4. Sort your simulated ˇ, O g.ˇ/,

O and the test statistic in ascending order. For a one-
sided test (for instance, a chi-square test), take the (0:95N )th observations in these
sorted vector as your 5% critical values. For a two-sided test (for instance, a t-
test), take the (0:025N )th and (0:975N )th observations as the 5% critical values.
You may also record how many times the 5% critical values from the asymptotic
distribution would reject a true null hypothesis.
O g.ˇ/,
5. You may also want to plot a histogram of ˇ, O and the test statistic to see if there
is a small sample bias, and how the distribution looks like. Is it close to normal?
How wide is it?

See Figures 6.2–6.3 for an example.

We have the same basic procedure when y t is a vector, except that we might have
to consider correlations across the elements of the vector of residuals u t . For instance,
we might want to generate the vector uQ t from a N.0; ˙/ distribution—where ˙ is the
variance-covariance matrix of u t .

Remark 6.1 (Generating N.; ˙/ random numbers) Suppose you want to draw an n  1
vector " t of N.; ˙/ variables. Use the Cholesky decomposition to calculate the lower
triangular P such that ˙ D PP 0 . Draw u t from an N.0; I / distribution, and define
" t D  C P u t . Note that Cov." t / D E P u t u0t P 0 D PIP 0 D ˙ .

6.1.2 Monte Carlo Simulations when x t Includes Lags of y t

If x t contains lags of y t , then we must set up the simulations so that feature is preserved in
every artificial sample that we create. For instance, suppose x t includes y t 1 and another

91
 Average LS estimate of ρ Std of LS estimate of ρ

simulation
0.9 asymptotic
0.1
0.85
0.8
0.05
0.75
0.7 0
0 100 200 300 400 500 0 100 200 300 400 500
Sample size, T Sample size, T

√
T × Std of LS estimate of ρ True model: yt = 0.9yt−1 + ǫt ,
0.8 where ǫt is iid N(0,2)

0.7 Estimated model: yt = a + ρyt−1 + ut

0.6 Number of simulations: 25000

0.5

0.4
0 100 200 300 400 500
Sample size, T

Figure 6.2: Results from a Monte Carlo experiment of LS estimation of the AR coeffi-
cient.

vector z t of variables which are independent of u t ˙s for all s. We can then generate an
artificial sample as follows. First, create a sample zQ t for t D 1; : : : ; T by some time series
model (for instance, a VAR) or by taking the observed sample itself. Second, observation
t of .xQ t ; yQ t / is generated as
" #
yQ t 1
xQ t D and yQ t D xQ t0 ˇ C uQ t for t D 1; : : : ; T (6.3)
zQ t

We clearly need the initial value yQ0 to start up the artificial sample—and then the rest of
the sample (t D 1; 2; :::) is calculated recursively.
For instance, for a VAR(2) model (where there is no z t )

y t D A1 y t 1 C A2 y t 2 C ut ; (6.4)

92
 Distribution of LS estimator, T = 25 Distribution of LS estimator, T = 100
Mean and std: Mean and std:
0.1 0.74 0.16 0.1 0.86 0.06

0.05 0.05

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

True model: yt = 0.9yt−1 + ǫt , ǫt is iid N(0,2)

Estimated model: yt = a + ρyt−1 + ut
Number of simulations: 25000

Figure 6.3: Results from a Monte Carlo experiment of LS estimation of the AR coeffi-
cient.

the procedure is straightforward. First, estimate the model on data and record the esti-
mates (A1 ; A2 ; Var.u t /). Second, draw a new time series of residuals, uQ t for t D 1; : : : ; T
and construct an artificial sample recursively (first t D 1, then t D 2 and so forth) as

yQ t D A1 yQ t 1 C A2 yQ t 2 C uQ t : (6.5)

(This requires some starting values for y 1 and y0 .) Third, re-estimate the model on the
the artificial sample, yQ t for t D 1; : : : ; T .

6.1.3 Monte Carlo Simulations with more Complicated Errors

It is straightforward to sample the errors from other distributions than the normal, for in-
stance, a student-t distribution. Equipped with uniformly distributed random numbers,
you can always (numerically) invert the cumulative distribution function (cdf) of any
distribution to generate random variables from any distribution by using the probability
transformation method. See Figure 6.4 for an example.

Remark 6.2 Let X  U.0; 1/ and consider the transformation Y D F 1 .X/, where
F 1 ./ is the inverse of a strictly increasing cumulative distribution function F , then Y
has the cdf F .

93
 Distribution of LS t-stat, T = 5 Distribution of LS t-stat, T = 100
t = (b̂ − 0.9)/Std(b̂)
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−4 −2 0 2 4 −4 −2 0 2 4

Model: Rt = 0.9ft + ǫt , ǫt = vt − 2,
where vt has a χ22 distribution Probability density functions

Estimated model: yt = a + bft + ut N(0,1)

0.4
Number of simulations: 25000 χ22 − 2
0.3
T = 5 T = 100 0.2
Kurtosis of t-stat: 46.753 3.049
Frequency of |t-stat| > 1.65 0.294 0.105 0.1
Frequency of |t-stat| > 1.96 0.227 0.054 0
−4 −2 0 2 4

Figure 6.4: Results from a Monte Carlo experiment with thick-tailed errors.

Example 6.3 The exponential cdf is x D 1 exp. y/ with inverse y D ln .1 x/ =.
Draw x from U.0:1/ and transform to y to get an exponentially distributed variable.

It is more difficult to handle non-iid errors, like those with autocorrelation and het-
eroskedasticity. We then need to model the error process and generate the errors from that
model.
If the errors are autocorrelated, then we could estimate that process from the fitted
errors and then generate artificial samples of errors (here by an AR(2))

uQ t D a1 uQ t 1 C a2 uQ t 2 C "Q t : (6.6)

Alternatively, heteroskedastic errors can be generated by, for instance, a GARCH(1,1)

model
u t  N.0;  t2 /, where  t2 D ! C ˛u2t 1 C ˇ t2 1 : (6.7)

94
However, this specification does not account for any link between the volatility and the
regressors (squared)—as tested for by White’s test. This would invalidate the usual OLS
standard errors and therefore deserves to be taken seriously. A simple, but crude, approach
is to generate residuals from a N.0;  t2 ) process, but where  t2 is approximated by the
fitted values from
"2t D c 0 w t C  t ; (6.8)

where w t include the squares and cross product of all the regressors.

6.2 Bootstrapping

6.2.1 Bootstrapping in the Simplest Case

Bootstrapping is another way to do simulations, where we construct artificial samples by

sampling from the actual data. The advantage of the bootstrap is then that we do not
have to try to estimate the process of the errors and regressors (as we do in a Monte Carlo
experiment). The real benefit of this is that we do not have to make any strong assumption
about the distribution of the errors.
The bootstrap approach works particularly well when the errors are iid and indepen-
dent of x t s for all s. This means that x t cannot include lags of y t . We here consider
bootstrapping the linear model (6.1), for which we have point estimates (perhaps from
LS) and fitted residuals. The procedure is similar to the Monte Carlo approach, except
that the artificial sample is generated differently. In particular, Step 1 in the Monte Carlo
simulation is replaced by the following:

1. Construct an artificial sample yQ t for t D 1; : : : ; T by

yQ t D x t0 ˇ C uQ t ; (6.9)

where uQ t is drawn (with replacement) from the fitted residual and where ˇ is the
point estimate.

Example 6.4 With T D 3, the artificial sample could be

2 3 2 0 3
.yQ1 ; xQ 1 / .x1 ˇ0 C u2 ; x1 /
4 .yQ2 ; xQ 2 / 5 D 4 .x20 ˇ0 C u1 ; x2 / 5 :
6 7 6 7

.yQ3 ; xQ 3 / .x30 ˇ0 C u2 ; x3 /

95
 The approach in (6.9) works also when y t is a vector of dependent variables—and
will then help retain the cross-sectional correlation of the residuals.

6.2.2 Bootstrapping when x t Includes Lags of y t

When x t contains lagged values of y t , then we have to modify the approach in (6.9) since
uQ t can become correlated with x t . For instance, if x t includes y t 1 and we happen to
sample uQ t D u t 1 , then we get a non-zero correlation. The easiest way to handle this
is as in the Monte Carlo simulations in (6.3), but where uQ t are drawn (with replacement)
from the sample of fitted residuals. The same carries over to the VAR model in (6.4)–(6.5).

6.2.3 Bootstrapping when Errors Are Heteroskedastic

Suppose now that the errors are heteroskedastic, but serially uncorrelated. If the het-
eroskedasticity is unrelated to the regressors, then we can still use (6.9).
On contrast, if the heteroskedasticity is related to the regressors, then the traditional LS
covariance matrix is not correct (this is the case that White’s test for heteroskedasticity
tries to identify). It would then be wrong to pair x t with just any uQ t D us since that
destroys the relation between x t and the variance of the residual.
An alternative way of bootstrapping can then be used: generate the artificial sample
by drawing (with replacement) pairs .ys ; xs /, that is, we let the artificial pair in t be
.yQ t ; xQ t / D .xs0 ˇ0 C us ; xs / for some random draw of s so we are always pairing the
residual, us , with the contemporaneous regressors, xs . Note that we are always sampling
with replacement—otherwise the approach of drawing pairs would be to just re-create the
original data set.
This approach works also when y t is a vector of dependent variables.

Example 6.5 With T D 3, the artificial sample could be

2 3 2 0 3
.yQ1 ; xQ 1 / .x2 ˇ0 C u2 ; x2 /
4 .yQ2 ; xQ 2 / 5 D 4 .x30 ˇ0 C u3 ; x3 / 5
6 7 6 7

.yQ3 ; xQ 3 / .x30 ˇ0 C u3 ; x3 /

It could be argued (see, for instance, Davidson and MacKinnon (1993)) that bootstrap-
ping the pairs .ys ; xs / makes little sense when xs contains lags of ys , since the random
sampling of the pair .ys ; xs / destroys the autocorrelation pattern on the regressors.

96
6.2.4 Autocorrelated Errors

It is quite hard to handle the case when the errors are serially dependent, since we must
the sample in such a way that we do not destroy the autocorrelation structure of the data.
A common approach is to fit a model for the residuals, for instance, an AR(1), and then
bootstrap the (hopefully iid) innovations to that process.
Another approach amounts to resampling blocks of data. For instance, suppose the
sample has 10 observations, and we decide to create blocks of 3 observations. The first
block is .uO 1 ; uO 2 ; uO 3 /, the second block is .uO 2 ; uO 3 ; uO 4 /, and so forth until the last block,
.uO 8 ; uO 9 ; uO 10 /. If we need a sample of length 3 , say, then we simply draw  of those
block randomly (with replacement) and stack them to form a longer series. To handle
end point effects (so that all data points have the same probability to be drawn), we also
create blocks by “wrapping” the data around a circle. In practice, this means that we add
a the following blocks: .uO 10 ; uO 1 ; uO 2 / and .uO 9 ; uO 10 ; uO 1 /. The length of the blocks should
clearly depend on the degree of autocorrelation, but T 1=3 is sometimes recommended as
a rough guide. An alternative approach is to have non-overlapping blocks. See Berkowitz
and Kilian (2000) for some other approaches.
See Figures 6.5–6.6 for an illustration.

6.2.5 Other Approaches

There are many other ways to do bootstrapping. For instance, we could sample the re-
gressors and residuals independently of each other and construct an artificial sample of
the dependent variable yQ t D xQ t0 ˇO C uQ t . This clearly makes sense if the residuals and
regressors are independent of each other and errors are iid. In that case, the advantage of
this approach is that we do not keep the regressors fixed.

Bibliography
Berkowitz, J., and L. Kilian, 2000, “Recent developments in bootstrapping time series,”
Econometric-Reviews, 19, 1–48.

Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

97
 Std of LS slope under autocorrelation Std of LS slope under autocorrelation
0.1 κ = −0.9 0.1 κ = 0
OLS formula
Newey-West
Simulated
Bootstrapped
0.05 0.05

0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS slope under autocorrelation where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9 xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
The block bootstrap uses blocks of size 20
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 6.5: Standard error of OLS estimator, autocorrelated errors

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Davison, A. C., and D. V. Hinkley, 1997, Bootstrap methods and their applications, Cam-
bridge University Press.

Efron, B., and R. J. Tibshirani, 1993, An introduction to the bootstrap, Chapman and Hall,
New York.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Horowitz, J. L., 2001, “The Bootstrap,” in J.J. Heckman, and E. Leamer (ed.), Handbook
of Econometrics . , vol. 5, Elsevier.

98
 Std of LS intercept under autocorrelation Std of LS intercept under autocorrelation
0.1 κ = −0.9 0.1 κ = 0

0.05 0.05
OLS formula
Newey-West
Simulated
Bootstrapped
0 0
−0.5 0 0.5 −0.5 0 0.5
ρ (autocorrelation of residual) ρ (autocorrelation of residual)

Model: yt = 0.9xt + ǫt ,
Std of LS intercept under autocorrelation where ǫt = ρǫt−1 + ξt , ξt is iid N
0.1 κ = 0.9 xt = κxt−1 + ηt , ηt is iid N

ut is the residual from LS estimate of

yt = a + bxt + ut
0.05
NW uses 15 lags
The block bootstrap uses blocks of size 20
Number of simulations: 25000
0
−0.5 0 0.5
ρ (autocorrelation of residual)

Figure 6.6: Standard error of OLS estimator, autocorrelated errors

99
7 GMM
References: Greene (2000) 4.7 and 11.5-6
Additional references: Hayashi (2000) 3-4; Verbeek (2004) 5; Hamilton (1994) 14; Ogaki
(1993), Johnston and DiNardo (1997) 10; Harris and Matyas (1999); Pindyck and Rubin-
feld (1998) Appendix 10.1; Cochrane (2001) 10-11

7.1 Method of Moments

Let m .x t / be a k 1 vector valued continuous function of a stationary process, and let the
probability limit of the mean of m .:/ be a function .:/ of a k  1 vector ˇ of parameters.
We want to estimate ˇ. The method of moments (MM, not yet generalized to GMM)
estimator is obtained by replacing the probability limit with the sample mean and solving
the system of k equations
T
1X
m .x t / .ˇ/ D 0k1 (7.1)
T t D1

for the parameters ˇ.

It is clear that this is a consistent estimator of ˇ if is continuous. (Proof: the sample
mean is a consistent estimator of .:/, and by Slutsky’s theorem plim .ˇ/ O D .plim ˇ/O
if is a continuous function.)

Example 7.1 (Moment conditions for variances and covariance) Suppose the series x t
and y t have zero means. The following moment conditions define the traditional variance
and covariance estimators
1 XT
x2 xx D 0
T t D1 t
1 XT
y2 yy D 0
T tD1 t
1 XT
xt yt xy D 0:
T t D1

100
It does not matter if the parameterers are estimated separately or jointly. In contrast, if
we want the correlation, xy , instead of the covariance, then we change the last moment
condition to
1 XT p p
x t y t xy xx yy D 0;
T t D1

which must be estimated jointly with the first two conditions.

Example 7.2 (MM for an MA(1).) For an MA(1), y t D  t C  t 1, we have

2
Ey t2 D 2 1 C  2


D E . t C  t 1/

E .y t y t 1 / D E Œ. t C  t 1 / . t 1 C  t 2 / D 2 :

The moment conditions could therefore be

" P
# " #
1 T 2 2 2
T t D1 y t   1 C  0
1
PT 2
D ;
T t D1 y t y t 1   0

which allows us to estimate  and  2 .

7.2 Generalized Method of Moments

GMM extends MM by allowing for more orthogonality conditions than parameters. This
could, for instance, increase efficiency and/or provide new aspects which can be tested.
Many (most) traditional estimation methods, like LS, IV, and MLE are special cases
of GMM. This means that the properties of GMM are very general, and therefore fairly
difficult to prove.

7.3 Moment Conditions in GMM

Suppose we have q (unconditional) moment conditions,

2 3
Em1 .w t ; ˇ0 /
6 :: 7
Em.w t ; ˇ0 / D 6
4 : 7
5
Emq .w t ; ˇ0 /
D 0q1 ; (7.2)

101
from which we want to estimate the k  1 (k  q) vector of parameters, ˇ. The true
values are ˇ0 . We assume that w t is a stationary and ergodic (vector) process (otherwise
the sample means does not converge to anything meaningful as the sample size increases).
The sample averages, or “sample moment conditions,” evaluated at some value of ˇ, are
T
1X
N
m.ˇ/ D m.w t ; ˇ/: (7.3)
T t D1

The sample average m N .ˇ/ is a vector of functions of random variables, so they are ran-
dom variables themselves and depend on the sample used. It will later be interesting to
N .ˇ/. Note that m.ˇ
calculate the variance of m N 1 / and m.ˇ
N 2 / are sample means obtained
by using two different parameter vectors, but on the same sample of data.
Example 7.3 (Moments conditions for IV/2SLS.) Consider the linear model y t D x t0 ˇ0 C
u t , where x t and ˇ are k  1 vectors. Let z t be a q  1 vector, with q  k. The moment
conditions and their sample analogues are
T
1X
0q1 D Ez t u t D EŒz t .y t x t0 ˇ0 /; N .ˇ/ D
and m z t .y t x t0 ˇ/;
T t D1

(or Z 0 .Y Xˇ/=T in matrix form). Let q D k to get IV; let z t D x t to get LS.
Example 7.4 (Moments conditions for MLE.) The maximum likelihood estimator maxi-
mizes the log likelihood function, T1 ˙ tTD1 ln L .w t I ˇ/, which requires T1 ˙ tTD1 @ ln L .w t I ˇ/ =@ˇ D
0. A key regularity condition for the MLE is that E@ ln L .w t I ˇ0 / =@ˇ D 0, which is just
like a GMM moment condition.

7.3.1 Digression: From Conditional to Unconditional Moment Conditions

Suppose we are instead given conditional moment restrictions

E Œu.x t ; ˇ0 /jz t  D 0m1 ; (7.4)

where z t is a vector of conditioning (predetermined) variables. We want to transform this

to unconditional moment conditions.
Remark 7.5 (E.ujz/ D 0 versus Euz D 0:) For any random variables u and z,

Cov .z; u/ D Cov Œz; E .ujz/ :

102
The condition E.ujz/ D 0 then implies Cov.z; u/ D 0. Recall that Cov.z; u/ D Ezu EzEu,
and that E.ujz/ D 0 implies that Eu D 0 (by iterated expectations). We therefore get that
" #
Cov .z; u/ D 0
E .ujz/ D 0 ) ) Euz D 0:
Eu D 0

Example 7.6 (Euler equation for optimal consumption.) The standard Euler equation
for optimal consumption choice which with isoelastic utility U .C t / D C t1 = .1 / is

C t C1
   ˇ 
ˇ
E R t C1 ˇ 1ˇˇ ˝ t D 0;
Ct
where R t C1 is a gross return on an investment and ˝ t is the information set in t . Let
z t 2 ˝ t , for instance asset returns or consumption t or earlier. The Euler equation then
implies
C t C1
   
E R t C1 ˇ z t z t D 0:
Ct

Let z t D .z1t ; :::; znt /0 , and define the new (unconditional) moment conditions as
2 3
u1 .x t ; ˇ/z1t
6 7
6 u1 .x t ; ˇ/z2t 7
::
6 7
:
6 7
6 7
6 7
m.w t ; ˇ/ D u.x t ; ˇ/ ˝ z t D 6 6 u 1 .x t ; ˇ/znt
7
7 ; (7.5)
6 u2 .x t ; ˇ/z1t 7
6 7
6 :: 7
:
6 7
4 5
um .x t ; ˇ/znt q1

which by (7.4) must have an expected value of zero, that is

Em.w t ; ˇ0 / D 0q1 : (7.6)

This a set of unconditional moment conditions—just as in (7.2). The sample moment con-
ditions (7.3) are therefore valid also in the conditional case, although we have to specify
m.w t ; ˇ/ as in (7.5).
Note that the choice of instruments is often arbitrary: it often amounts to using only
a subset of the information variables. GMM is often said to be close to economic theory,

103
but it should be admitted that economic theory sometimes tells us fairly little about which
instruments, z t , to use.

Example 7.7 (Euler equation for optimal consumption, continued) The orthogonality
conditions from the consumption Euler equations in Example 7.6 are highly non-linear,
and theory tells us very little about how the prediction errors are distributed. GMM has
the advantage of using the theoretical predictions (moment conditions) with a minimum
of distributional assumptions. The drawback is that it is sometimes hard to tell exactly
which features of the (underlying) distribution that are tested.

7.4 The Optimization Problem in GMM

7.4.1 The Loss Function

The GMM estimator ˇO minimizes the weighted quadratic form

2 30 2 32 3
mN 1 .ˇ/ W11





W1q m N 1 .ˇ/
6 : 7 6 : :: :: 7 6 :: 7
6 :: 7 6 :: : : 7 6 : 7
J D6 : (7.7)
6 7 6 76 7
7 6 : :: :: 7 ::
6 :: 7 6 ::
6 7
4 5 4 : : 76
54 : 7
5
mN q .ˇ/ W1q





Wqq m N q .ˇ/
0
D m.ˇ/
N N
W m.ˇ/; (7.8)

N
where m.ˇ/ is the sample average of m.w t ; ˇ/ given by (7.3), and where W is some
q  q symmetric positive definite weighting matrix. (We will soon discuss a good choice
of weighting matrix.) There are k parameters in ˇ to estimate, and we have q moment
N
conditions in m.ˇ/. We therefore have q k overidentifying moment restrictions.
With q D k the model is exactly identified (as many equations as unknowns), and it
should be possible to set all q sample moment conditions to zero by a choosing the k D q
parameters. It is clear that the choice of the weighting matrix has no effect in this case
N ˇ/
since m. O D 0 at the point estimates ˇ.
O

Example 7.8 (Simple linear regression.) Consider the model

y t D x t ˇ0 C u t ; (7.9)

104
where y t and x t are zero mean scalars. The moment condition and loss function are
T
1X
N .ˇ/ D
m x t .y t x t ˇ/ and
T t D1
" T
#2
1X
J DW x t .y t x t ˇ/ ;
T t D1

so the scalar W is clearly irrelevant in this case.

Example 7.9 (IV/2SLS method continued.) From Example 7.3, we note that the loss func-
tion for the IV/2SLS method is
T
" #0 " T
#
0 1X 1X
N
m.ˇ/ N
W m.ˇ/ D z t .y t x t0 ˇ/ W z t .y t x t0 ˇ/ :
T t D1 T t D1

When q D k, then the model is exactly identified, so the estimator could actually be found
by setting all moment conditions to zero. We then get the IV estimator
T
1X
0D z t .y t x t0 ˇOI V / or
T tD1
1
T T
!
1X 1X
ˇOI V D z t x t0 zt yt
T t D1 T t D1
D ˙O zx1 ˙O zy ;

where ˙O zx D ˙ tD1 T
z t x t0 =T and similarly for the other second moment matrices. Let
z t D x t to get LS
ˇOLS D ˙O xx1 ˙O xy :

7.4.2 First Order Conditions

Remark 7.10 (Matrix differentiation of non-linear functions.) Let the vector yn1 be a
function of the vector xm1
2 3 2 3
y1 f1 .x/
6 : 7
6 :: 7 D f .x/ D 6 :::
6 7
7:
4 5 4 5
yn fn .x/

105
Then, @y=@x 0 is an n  m matrix
2 3 2 3
@f1 .x/ @f1 .x/ @f1 .x/
@x 0 @x1



@xm
@y 6 :
6 ::
7 6 :
7 D 6 :: :: 7
D : 7:
@x 0 4
@f
5 4 5
1 .x/ @fn .x/ @fn .x/
@x 0 @x1



@xm

(Note that the notation implies that the derivatives of the first element in y, denoted y1 ,
with respect to each of the elements in x 0 are found in the first row of @y=@x 0 . A rule to
help memorizing the format of @y=@x 0 : y is a column vector and x 0 is a row vector.)

Remark 7.11 When y D Ax where A is an n  m matrix, then fi .x/ in Remark 7.10 is

a linear function. We then get @y=@x 0 D @ .Ax/ =@x 0 D A.

Remark 7.12 As a special case of the previous remark y D z 0 x where both z and x are
vectors. Then @ .z 0 x/ =@x 0 D z 0 (since z 0 plays the role of A).

Remark 7.13 (Matrix differentiation of quadratic forms.) Let xn1 , f .x/m1 , and Amm
symmetric. Then
@f .x/0 Af .x/ @f .x/ 0
 
D2 Af .x/ :
@x @x 0
Remark 7.14 If f .x/ D x, then @f .x/ =@x 0 D I , so @ .x 0 Ax/ =@x D 2Ax.

The k first order conditions for minimizing the GMM loss function in (7.8) with re-
spect to the k parameters are that the partial derivatives with respect to ˇ equal zero at the
O
estimate, ˇ,

N ˇ/
@m. O 0 W m. O
N ˇ/
0k1 D
@ˇ
30 2
N 1 .ˇO / N 1 .ˇO /
2 32 3
@m @m O



W11





W1q m N 1 .ˇ/
6 : @ˇ1 ::
@ˇk
7 6 :: :: :: 76 :: 7
6 :: : 7 6 : : : 76 : 7
D6 : 7 (with ˇOk1 );
6 7 6 76 7
6 :: :: 7 6 :: ::: :: 76 ::
4 : 7
5
6
4 : : 76
54 : 7
5
@mN q .ˇO / N q .ˇO /
@m O
@ˇ1



@ˇk
W1q





Wqq m N q .ˇ/
(7.10)
!0
N ˇ/
@m. O
D W m. O :
N ˇ/ (7.11)
@ˇ 0 „ƒ‚…„ƒ‚…
„ ƒ‚ … qq q1
kq

106
We can solve for the GMM estimator, ˇ,O from (7.11). This set of equations must often be
solved by numerical methods, except in linear models (the moment conditions are linear
functions of the parameters) where we can find analytical solutions by matrix inversion.

Example 7.15 (First order conditions of simple linear regression.) The first order condi-
tions of the loss function in Example 7.8 is
" T
#2
d 1X O
0D W x t .y t x t ˇ/
dˇ T t D1
T T
" # " #
1X 2 1X O ; or
D x W x t .y t x t ˇ/
T t D1 t T t D1
T
! 1 T
O 1 X
2 1X
ˇD x xt yt :
T t D1 t T t D1

Example 7.16 (First order conditions of IV/2SLS.) The first order conditions correspond-
ing to (7.11) of the loss function in Example 7.9 (when q  k) are
" #0
N ˇ/
@m. O
0k1 D W m. N ˇ/ O
@ˇ 0
" T
#0 T
@ 1X 0 O 1X O
D 0
z t .y t x t ˇ/ W z t .y t x t0 ˇ/
@ˇ T tD1 T tD1
" T
#0 T
1X 0 1X O
D zt xt W z t .y t x t0 ˇ/
T t D1 T tD1
D ˙O xz W .˙O zy O
˙O zx ˇ/:

We can solve for ˇO from the first order conditions as

  1
O O O
ˇ2SLS D ˙xz W ˙zx ˙O xz W ˙O zy :

When q D k, then the first order conditions can be premultiplied with .˙O xz W / 1 , since
˙O xz W is an invertible k  k matrix in this case, to give

0k1 D ˙O zy O so ˇOI V D ˙O 1 ˙O zy :
˙O zx ˇ, zx

This shows that the first order conditions are just the same as the sample moment condi-

107
tions, which can be made to hold exactly since there are as many parameters as there are
equations.

7.5 Asymptotic Properties of GMM

We know very little about the general small sample properties, including bias, of GMM.
We therefore have to rely either on simulations (Monte Carlo or bootstrap) or on the
asymptotic results. This section is about the latter.
GMM estimates are typically consistent and normally distributed, even if the series
m.w t ; ˇ/ in the moment conditions (7.3) are serially correlated and heteroskedastic—
provided w t is a stationary and ergodic process. The reason is essentially that the esti-
mators are (at least as a first order approximation) linear combinations of sample means
which typically are consistent (LLN) and normally distributed (CLT). More about that
later. The proofs are hard, since the GMM is such a broad class of estimators. This
section discusses, in an informal way, how we can arrive at those results.

7.5.1 Consistency

Sample moments are typically consistent, so plim m .ˇ/ D E m.w t ; ˇ/. This must hold
at any parameter vector in the relevant space (for instance, those inducing stationarity and
variances which are strictly positive). Then, if the moment conditions (7.2) are true only at
the true parameter vector, ˇ0 , (otherwise the parameters are “unidentified”) and that they
are continuous in ˇ, then GMM is consistent. The idea is thus that GMM asymptotically
solves

0q1 D plim m.
N ˇ/ O
O
D E m.w t ; ˇ/;

which only holds at ˇO D ˇ0 . Note that this is an application of Slutsky’s theorem.

Remark 7.17 (Slutsky’s theorem.) If fxT g is a sequence of random matrices such that
plim xT D x and g.xT / a continuous function, then plim g.xT / D g.x/.

108
Example 7.18 (Consistency of 2SLS.) By using y t D x t0 ˇ0 Cu t , the first order conditions
in Example 7.16 can be rewritten
T
O 1X O
0k1 D ˙xz W z t .y t x t0 ˇ/
T t D1
T
1X h  i
D ˙O xz W z t u t C x t0 ˇ0 ˇO
T t D1
 
O O O O
D ˙xz W ˙zu C ˙xz W ˙zx ˇ0 O
ˇ :

Take the probability limit

 
0k1 O O O O
D plim ˙xz W plim ˙zu C plim ˙xz W plim ˙zx ˇ0 O
plim ˇ :

In most cases, plim ˙O xz is some matrix of constants, and plim ˙O zu D E z t u t D 0q1 . It

then follows that plim ˇO D ˇ0 . Note that the whole argument relies on that the moment
condition, E z t u t D 0q1 , is true. If it is not, then the estimator is inconsistent. For
instance, when the instruments are invalid (correlated with the residuals) or when we
use LS (z t D x t ) when there are measurement errors or in a system of simultaneous
equations.

7.5.2 Asymptotic Normality

p
To give the asymptotic distribution of T .ˇO ˇ0 /, we need to define three things. (As
p
usual, we also need to scale with T to get a non-trivial asymptotic distribution; the
asymptotic distribution of ˇO ˇ0 is a spike at zero.) First, let S0 (a q  q matrix) denote
p
the asymptotic covariance matrix (as sample size goes to infinity) of T times the sample
moment conditions evaluated at the true parameters
hp i
S0 D ACov Tm N .ˇ0 / (7.12)
T
" #
1 X
D ACov p m.w t ; ˇ0 / ; (7.13)
T t D1

where we use the definition of m N .ˇ0 / in (7.3). (To estimate S0 it is important to recognize
that it is a scaled sample average.) Let R .s/ be the q  q covariance (matrix) of the vector

109
m.w t ; ˇ0 / with the vector m.w t 2 ; ˇ0 /

R .s/ D Cov Œm.w t ; ˇ0 /; m.w t s ; ˇ0 /

D E m.w t ; ˇ0 /m.w t s ; ˇ0 /0 : (7.14)

Then, it is well known that

hp i 1
X
ACov N .ˇ0 / D
Tm R.s/: (7.15)
sD 1

In practice, we often estimate this by using the Newey-West estimator (or something
similar).
Second, let D0 (a q  k matrix) denote the probability limit of the gradient of the
sample moment conditions with respect to the parameters, evaluated at the true parameters

N 0/
@m.ˇ
D0 D plim , where (7.16)
@ˇ 0
2 3
@m
N 1 .ˇ / @m
N 1 .ˇ /
@ˇ1



@ˇk
6 : :: 7
N 0/ 6
@m.ˇ 6 :: : 7
D 7 at the true ˇ vector. (7.17)
7
6 : ::
@ˇ 0 6 :: : 7
4 5
@m
N q .ˇ / @m
N q .ˇ /
@ˇ1



@ˇk

O also shows up in the first order conditions

Note that a similar gradient, but evaluated at ˇ,
(7.11). Third, let the weighting matrix be the inverse of the covariance matrix of the
moment conditions (once again evaluated at the true parameters)

W D S0 1 : (7.18)

It can be shown that this choice of weighting matrix gives the asymptotically most ef-
ficient estimator for a given set of orthogonality conditions. For instance, in 2SLS, this
means a given set of instruments and (7.18) then shows only how to use these instruments
in the most efficient way. Of course, another set of instruments might be better (in the
O
sense of giving a smaller Cov(ˇ)).
With the definitions in (7.12) and (7.16) and the choice of weighting matrix in (7.18)
and the added assumption that the rank of D0 equals k (number of parameters) then we

110
can show (under fairly general conditions) that
p d 1
T .ˇO ˇ0 / ! N.0k1 ; V /, where V D D00 S0 1 D0


: (7.19)

This holds also when the model is exactly identified, so we really do not use any weighting
matrix.
To prove this note the following.

Remark 7.19 (Continuous mapping theorem.) Let the sequences of random matrices
d p
fxT g and fyT g, and the non-random matrix faT g be such that xT ! x, yT ! y,
and aT ! a (a traditional limit). Let g.xT ; yT ; aT / be a continuous function. Then
d
g.xT ; yT ; aT / ! g.x; y; a/. Either of yT and aT could be irrelevant in g. (See Mittel-
hammer (1996) 5.3.)
p
Example 7.20 For instance, the sequences in Remark 7.19 could be xT D T ˙ tTD w t =T ,
the scaled sample average of a random variable w t ; yT D ˙ tTD w t2 =T , the sample second
moment; and aT D ˙ tTD1 0:7t .

d
Remark 7.21 From the previous remark: if xT ! x (a random variable) and plim QT D
d
Q (a constant matrix), then QT xT ! Qx.

Proof. (The asymptotic distribution (7.19). Sketch of proof.) This proof is essentially
an application of the delta rule. By the mean-value theorem the sample moment condition
evaluated at the GMM estimate, ˇ, O is

O D m.ˇ N 1/ O
@m.ˇ
N ˇ/
m. N 0/ C .ˇ ˇ0 / (7.20)
@ˇ 0

for some values ˇ1 between ˇO and ˇ0 . (This point is different for different elements in
N Premultiply with Œ@m.
m.) O
N ˇ/=@ˇ 0 0
 W . By the first order condition (7.11), the left hand
side is then zero, so we have
!0 !0
N ˇ/
@m. O N ˇ/
@m. O N 1/ O
@m.ˇ
0k1 D 0
N 0/ C
W m.ˇ 0
W .ˇ ˇ0 /: (7.21)
@ˇ @ˇ @ˇ 0

111
 p
Multiply with T and solve as
" !0 # 1 !0
p   N ˇ/
@m. O N 1/
@m.ˇ N ˇ/
@m. O p
T ˇO ˇ0 D W W N 0 /:
T m.ˇ (7.22)
@ˇ 0 @ˇ 0 @ˇ 0
„ ƒ‚ …

If
N ˇ/
@m. O N 0/
@m.ˇ N 1/
@m.ˇ
plim 0
D 0
D D0 ; then plim D D0 ;
@ˇ @ˇ @ˇ 0
O Then
since ˇ1 is between ˇ0 and ˇ.
1
D00 WD0 D00 W:


plim D (7.23)
p p
The last term in (7.22), T m.ˇN 0 /, is T times a vector of sample averages, so by a CLT
it converges in distribution to N.0; S0 /, where S0 is defined as in (7.12). By the rules of
limiting distributions (see Remark 7.19) we then have that
p   d
T ˇO ˇ0 ! plim  something that is N .0; S0 / ; that is,
p   d
O
T ˇ ˇ0 ! N 0k1 ; .plim /S0 .plim 0 / :
 

The covariance matrix is then

p
ACovŒ T .ˇO ˇ0 / D .plim/S0 .plim 0 /

1 0
1 0
D D00 WD0 D0 W S0 Œ D00 WD0 D0 W 0 (7.24)

1 0
1
D D00 WD0 D0 W S0 W 0 D0 D00 WD0 : (7.25)

If W D W 0 D S0 1 , then this expression simplifies to (7.19). (See, for instance, Hamilton

(1994) 14 (appendix) for more details.)
It is straightforward to show that the difference between the covariance matrix in

1
(7.25) and D00 S0 1 D0 (as in (7.19)) is a positive semi-definite matrix: any linear com-
bination of the parameters has a smaller variance if W D S0 1 is used as the weighting
matrix.
All the expressions for the asymptotic distribution are supposed to be evaluated at the
true parameter vector ˇ0 , which is unknown. However, D0 in (7.16) can be estimated by
O
N ˇ/=@ˇ
@m. 0
, where we use the point estimate instead of the true value of the parameter

112
vector. In practice, this means plugging in the point estimates into the sample moment
conditions and calculate the derivatives with respect to parameters (for instance, by a
numerical method).
Similarly, S0 in (7.13) can be estimated by, for instance, Newey-West’s estimator of
p
CovΠT m. O once again using the point estimates in the moment conditions.
N ˇ/,

Example 7.22 (Covariance matrix of 2SLS.) Define

hp p T !
i T X
S0 D ACov Tm N .ˇ0 / D ACov zt ut
T t D1
T
!
N 0/
@m.ˇ 1X
D0 D plim 0
D plim z t x t0 D ˙zx :
@ˇ T t D1
p
This gives the asymptotic covariance matrix of T .ˇO ˇ0 /
1 1
V D D00 S0 1 D0 0
S0 1 ˙zx



D ˙zx :

7.6 Summary of GMM

Economic model W Em.w t ; ˇ0 / D 0q1 , ˇ is k  1

T
1X
Sample moment conditions W m.ˇ/
N D m.w t ; ˇ/
T t D1
0
Loss function W J D m.ˇ/
N N
W m.ˇ/
!0
@m. O 0 W m.
N ˇ/ O
N ˇ/ N ˇ/
@m. O
First order conditions W 0k1 D D O
N ˇ/
W m.
@ˇ @ˇ 0
Consistency W ˇO is typically consistent if Em.w t ; ˇ0 / D 0
hp i N 0/
@m.ˇ
Define W S0 D Cov TmN .ˇ0 / and D0 D plim
@ˇ 0
Choose: W D S0 1
p d 1
Asymptotic distribution W T .ˇO ˇ0 / ! N.0k1 ; V /, where V D D00 S0 1 D0

113
7.7 Efficient GMM and Its Feasible Implementation

The efficient GMM (remember: for a given set of moment conditions) requires that we
use W D S0 1 , which is tricky since S0 should be calculated by using the true (unknown)
parameter vector. However, the following two-stage procedure usually works fine:

 First, estimate model with some (symmetric and positive definite) weighting matrix.
The identity matrix is typically a good choice for models where the moment con-
ditions are of the same order of magnitude (if not, consider changing the moment
conditions). This gives consistent estimates of the parameters ˇ. Then a consistent
estimate SO can be calculated (for instance, with Newey-West).

 Use the consistent SO from the first step to define a new weighting matrix as W D
SO 1 . The algorithm is run again to give asymptotically efficient estimates of ˇ.

 Iterate at least once more. (You may want to consider iterating until the point esti-
mates converge.)

Example 7.23 (Implementation of 2SLS.) Under the classical 2SLS assumptions, there is
no need for iterating since the efficient weighting matrix is ˙zz1 = 2 . Only  2 depends
on the estimated parameters, but this scaling factor of the loss function does not affect
ˇO2SLS .

One word of warning: if the number of parameters in the covariance matrix SO is

large compared to the number of data points, then SO tends to be unstable (fluctuates a lot
between the steps in the iterations described above) and sometimes also close to singular.
The saturation ratio is sometimes used as an indicator of this problem. It is defined as the
number of data points of the moment conditions (qT ) divided by the number of estimated
parameters (the k parameters in ˇO and the unique q.q C 1/=2 parameters in SO if it is
estimated with Newey-West). A value less than 10 is often taken to be an indicator of
problems. A possible solution is then to impose restrictions on S, for instance, that the
autocorrelation is a simple AR(1) and then estimate S using these restrictions (in which
case you cannot use Newey-West, or course).

114
7.8 Testing in GMM

The result in (7.19) can be used to do Wald tests of the parameter vector. For instance,
suppose we want to test the s linear restrictions that Rˇ0 D r (R is s  k and r is s  1)
then it must be the case that under null hypothesis
p d
T .RˇO r/ ! N.0s1 ; RVR0 /: (7.26)

Remark 7.24 (Distribution of quadratic forms.) If the n  1 vector x  N.0; ˙/, then
x 0 ˙ 1 x  2n .

From this remark and the continuous mapping theorem in Remark (7.19) it follows
that, under the null hypothesis that Rˇ0 D r, the Wald test statistics is distributed as a 2s
variable

1 d
T .RˇO r/0 RVR0 .RˇO r/ ! 2s : (7.27)

We might also want to test the overidentifying restrictions. The first order conditions
(7.11) imply that k linear combinations of the q moment conditions are set to zero by
solving for ˇ:O Therefore, we have q k remaining overidentifying restrictions which
should also be close to zero if the model is correct (fits data). Under the null hypothe-
sis that the moment conditions hold (so the overidentifying restrictions hold), we know
p
that T m N .ˇ0 / is a (scaled) sample average and therefore has (by a CLT) an asymptotic
normal distribution. It has a zero mean (the null hypothesis) and the covariance matrix in
(7.12). In short,
p d

TmN .ˇ0 / ! N 0q1 ; S0 : (7.28)
O 0 S 1 m.
N ˇ/ O
If would then perhaps be natural to expect that the quadratic form T m. 0 N ˇ/
should be converge in distribution to a 2q variable. That is not correct, however, since ˇO
chosen is such a way that k linear combinations of the first order conditions always (in
every sample) are zero. There are, in effect, only q k nondegenerate random variables in
the quadratic form (see Davidson and MacKinnon (1993) 17.6 for a detailed discussion).
The correct result is therefore that if we have used optimal weight matrix is used, W D
S0 1 , then
d
O 0 S 1 m.
N ˇ/
T m. O !
N ˇ/ 2 ; if W D S 1 : (7.29)
0 q k 0

115
The left hand side equals T times of value of the loss function (7.8) evaluated at the point
estimates, so we could equivalently write what is often called the J test

O  2
TJ.ˇ/ if W D S0 1 :
q k; (7.30)

This also illustrates that with no overidentifying restrictions (as many moment conditions
as parameters) there are, of course, no restrictions to test. Indeed, the loss function value
is then always zero at the point estimates.
Example 7.25 (Test of overidentifying assumptions in 2SLS.) In contrast to the IV method,
2SLS allows us to test overidentifying restrictions (we have more moment conditions than
parameters, that is, more instruments than regressors). This is a test of whether the residu-
als are indeed uncorrelated with all the instruments. If not, the model should be rejected.
It can be shown that test (7.30) is (asymptotically, at least) the same as the traditional
(Sargan (1964), see Davidson (2000) 8.4) test of the overidentifying restrictions in 2SLS.
In the latter, the fitted residuals are regressed on the instruments; TR2 from that regres-
sion is 2 distributed with as many degrees of freedom as the number of overidentifying
restrictions.
Example 7.26 (Results from GMM on CCAPM; continuing Example 7.6.) The instru-
ments could be anything known at t or earlier could be used as instruments. Actually,
Hansen and Singleton (1982) and Hansen and Singleton (1983) use lagged Ri;t C1 c t C1 =c t
as instruments, and estimate to be 0:68 to 0:95, using monthly data. However, TJT .ˇ/O
is large and the model can usually be rejected at the 5% significance level. The rejection
is most clear when multiple asset returns are used. If T-bills and stocks are tested at the
same time, then the rejection would probably be overwhelming.
Another test is to compare a restricted and a less restricted model, where we have
used the optimal weighting matrix for the less restricted model in estimating both the less
restricted and more restricted model (the weighting matrix is treated as a fixed matrix in
the latter case). It can be shown that the test of the s restrictions (the “D test”, similar in
flavour to an LR test), is

T ŒJ.ˇO rest ri ct ed / J.ˇO less rest ri ct ed /  2s ; if W D S0 1 : (7.31)

The weighting matrix is typically based on the unrestricted model. Note that (7.30) is a
special case, since the model with allows q non-zero parameters (as many as the moment

116
conditions) always attains J D 0, and that by imposing s D q k restrictions we get a
restricted model.

7.9 GMM with Sub-Optimal Weighting Matrix

When the optimal weighting matrix is not used, that is, when (7.18) does not hold, then
the asymptotic covariance matrix of the parameters is given by (7.25) instead of the result
in (7.19). That is,
p d 1
1
T .ˇO ˇ0 / ! N.0k1 ; V2 /, where V2 D D00 WD0 D00 W S0 W 0 D0 D00 WD0


:
(7.32)
The consistency property is not affected.
The test of the overidentifying restrictions (7.29) and (7.30) are not longer valid. In-
stead, the result is that
p
O !d N 0q1 ; 2 , with


N ˇ/
T m. (7.33)

1 0 1
2 D ŒI D0 D00 WD0 D0 D00 WD0 D00 W 0 :


D0 W S0 ŒI (7.34)

This covariance matrix has rank q k (the number of overidentifying restriction). This
distribution can be used to test hypotheses about the moments, for instance, that a partic-
ular moment condition is zero.
Proof. (Sketch of proof of (7.33)-(7.34)) Use (7.22) in (7.20) to get
p p p @m.ˇ
N 1/
T m. O D
N ˇ/ N 0/ C
T m.ˇ
T N 0/
m.ˇ
@ˇ 0
p
 
N 1/
@m.ˇ
D IC N 0 /:
T m.ˇ
@ˇ 0
1
The term in brackets has a probability limit, which by (7.23) equals I D0 D00 WD0 D00 W .


p
N 0 / !d N 0q1 ; S0 we get (7.33).


Since T m.ˇ

Remark 7.27 If the n  1 vector X  N.0; ˙/, where ˙ has rank r  n then Y D
X 0 ˙ C X  2r where ˙ C is the pseudo inverse of ˙.

Remark 7.28 The symmetric ˙ can be decomposed as ˙ D ZZ 0 where Z are the
0
orthogonal eigenvectors (Z Z D I ) and  is a diagonal matrix with the eigenvalues

117
along the main diagonal. The pseudo inverse can then be calculated as ˙ C D ZC Z 0 ,
where " #
1
 11 0
C D ;
0 0
with the reciprocals of the non-zero eigen values along the principal diagonal of 111 .

This remark and (7.34) implies that the test of overidentifying restrictions (Hansen’s
J statistics) analogous to (7.29) is

O 0 C m.
N ˇ/ O d 2
T m. 2 N ˇ/ ! q k: (7.35)

It requires calculation of a generalized inverse (denoted by superscript C ), but this is fairly

straightforward since 2 is a symmetric matrix. It can be shown (a bit tricky) that this
simplifies to (7.29) when the optimal weighting matrix is used.

7.10 GMM without a Loss Function

Suppose we sidestep the whole optimization issue and instead specify k linear combi-
nations (as many as there are parameters) of the q moment conditions directly. That is,
instead of the first order conditions (7.11) we postulate that the estimator should solve

0k1 D „ƒ‚…
A m. O (ˇO is k  1).
N ˇ/ (7.36)
„ƒ‚…
kq q1

The matrix A is chosen by the researcher and it must have rank k (lower rank means that
we effectively have too few moment conditions to estimate the k parameters in ˇ). If A
is random, then it should have a finite probability limit A0 (also with rank k). One simple
case when this approach makes sense is when we want to use a subset of the moment
conditions to estimate the parameters (some columns in A are then filled with zeros), but
we want to study the distribution of all the moment conditions.
By comparing (7.11) and (7.36) we see that A plays the same role as Œ@m. O
N ˇ/=@ˇ 0 0
W,
but with the difference that A is chosen and not allowed to depend on the parameters.
In the asymptotic distribution, it is the probability limit of these matrices that matter, so
we can actually substitute A0 for D00 W in the proof of the asymptotic distribution. The

118
covariance matrix in (7.32) then becomes

1 1
V3 D .A0 D0 / A0 S0 Œ.A0 D0 / A0 0
1
D .A0 D0 / A0 S0 A00 Œ.A0 D0 / 1 0 ; (7.37)

which can be used to test hypotheses about the parameters.

Similarly, the asymptotic distribution of the moment conditions is
p
O !d N 0q1 ; 3 , with


N ˇ/
T m. (7.38)
1 1
3 D ŒI D0 .A0 D0 / A0 S0 ŒI D0 .A0 D0 / A0  0 ; (7.39)

where 3 has reduced rank. As before, this covariance matrix can be used to construct
both t type and 2 tests of the moment conditions. For instance, the test of overidentifying
restrictions (Hansen’s J statistics)
d
O 0 C m.
N ˇ/ O 2
T m. 3 N ˇ/ ! q k; (7.40)

where 3C is a generalized inverse of 3 .

7.11 Simulated Moments Estimator

Reference: Ingram and Lee (1991)

It sometimes happens that it is not possible to calculate the theoretical moments in
GMM explicitly. For instance, suppose we want to match the variance of the model with
the variance of data

E m.w t ; ˇ0 / D 0, where (7.41)

2
m.w t ; ˇ/ D .w t / Var_in_model .ˇ/ ; (7.42)

but the model is so non-linear that we cannot find a closed form expression for Var_of_model.ˇ0 /.
Similary, we could match a covariance of
The SME involves (i) drawing a set of random numbers for the stochastic shocks in
the model; (ii) for a given set of parameter values generate a model simulation with Tsi m
observations, calculating the moments and using those instead of Var_of_model.ˇ0 / (or
similarly for other moments), which is then used to evaluate the loss function JT . This is

119
repeated for various sets of parameter values until we find the one which minimizes JT .
Basically all GMM results go through, but the covariance matrix should be scaled up
with 1 C T =Tsi m , where T is the sample length. Note that the same sequence of random
numbers should be reused over and over again (as the parameter values are changed).

Example 7.29 Suppose w t has two elements, x t and y t , and that we want to match both
variances and also the covariance. For simplicity, suppose both series have zero means.
Then we can formulate the moment conditions
2 2 3
x t Var(x)_in_model.ˇ/
m.x t ; y t ; ˇ/ D 4 y t2 Var(y)_in_model.ˇ/ 5 : (7.43)
6 7

xt yt Cov(x,y)_in_model.ˇ/

Bibliography
Cochrane, J. H., 2001, Asset pricing, Princeton University Press, Princeton, New Jersey.

Davidson, J., 2000, Econometric theory, Blackwell Publishers, Oxford.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Hansen, L., and K. Singleton, 1982, “Generalized instrumental variables estimation of

nonlinear rational expectations models,” Econometrica, 50, 1269–1288.

Hansen, L., and K. Singleton, 1983, “Stochastic consumption, risk aversion and the tem-
poral behavior of asset returns,” Journal of Political Economy, 91, 249–268.

Harris, D., and L. Matyas, 1999, “Introduction to the generalized method of moments esti-
mation,” in Laszlo Matyas (ed.), Generalized Method of Moments Estimation . chap. 1,
Cambridge University Press.

Hayashi, F., 2000, Econometrics, Princeton University Press.

120
Ingram, B.-F., and B.-S. Lee, 1991, “Simulation estimation of time-series models,” Jour-
nal of Econometrics, 47, 197–205.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

Mittelhammer, R. C., 1996, Mathematical statistics for economics and business, Springer-
Verlag, New York.

Ogaki, M., 1993, “Generalized method of moments: econometric applications,” in G. S.

Maddala, C. R. Rao, and H. D. Vinod (ed.), Handbook of Statistics, vol. 11, . chap. 17,
pp. 455–487, Elsevier.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

121
8 Examples and Applications of GMM

8.1 GMM and Classical Econometrics: Examples

8.1.1 The LS Estimator (General)

The model is
y t D x t0 ˇ0 C u t ; (8.1)

where ˇ is a k  1 vector.
The k moment conditions are
T T T
1X 1X 1X
N .ˇ/ D
m x t .y t x t0 ˇ/ D xt yt x t x t0 ˇ: (8.2)
T t D1 T t D1 T t D1

The point estimates are found by setting all moment conditions to zero (the model is
N .ˇ/ D 0k1 , which gives
exactly identified), m
1
T T
!
1X 1X
ˇO D x t x t0 x t y t ˇ: (8.3)
T t D1 T t D1

If we define
hp p T !
i T X
S0 D ACov Tm N .ˇ0 / D ACov xt ut (8.4)
T t D1
T
!
N 0/
@m.ˇ 1X
D0 D plim 0
D plim x t x t0 D ˙xx : (8.5)
@ˇ T t D1
p
then the asymptotic covariance matrix of T .ˇO ˇ0 /
1 1
VLS D D00 S0 1 D0 0
S0 1 ˙xx D ˙xx1 S0 ˙xx1 :



D ˙xx (8.6)

We can then either try to estimate S0 by Newey-West, or make further assumptions to

simplify S0 (see below).

122
8.1.2 The IV/2SLS Estimator (General)

The model is (8.1), but we use an IV/2SLS method. The q moment conditions (with
q  k) are
T T T
1X 1X 1X
N .ˇ/ D
m z t .y t x t0 ˇ/ D zt yt z t x t0 ˇ: (8.7)
T t D1 T t D1 T t D1

The loss function is (for some positive definite weighting matrix W , not necessarily
the optimal)

T
" #0 " T
#
0 1X 1X
N
m.ˇ/ N
W m.ˇ/ D z t .y t x t0 ˇ/ W z t .y t x t0 ˇ/ ; (8.8)
T t D1 T tD1

O
N ˇ/=@ˇ
and the k first order conditions, .@m. 0 0 O D 0, are
N ˇ/
/ W m.
" T
#0 T
@ 1X 0 O 1X O
0k1 D z t .y t x t ˇ/ W z t .y t x t0 ˇ/
@ˇ 0 T tD1 T tD1
" T
#0 T
1X 1X O
D z t x t0 W z t .y t x t0 ˇ/
T t D1 T tD1
D ˙O xz W .˙O zy O
˙O zx ˇ/: (8.9)

We solve for ˇO as
  1
O O O
ˇ D ˙xz W ˙zx ˙O xz W ˙O zy : (8.10)

Define
hp p T !
i T X
S0 D ACov Tm N .ˇ0 / D ACov zt ut (8.11)
T t D1
T
!
N 0/
@m.ˇ 1X
D0 D plim D plim z t x t0 D ˙zx : (8.12)
@ˇ 0 T t D1
p
This gives the asymptotic covariance matrix of T .ˇO ˇ0 /
1 1
V D D00 S0 1 D0 0
S0 1 ˙zx



D ˙zx : (8.13)

When the model is exactly identified (q D k/, then we can make some simplifications

123
since ˙O xz is then invertible. This is the case of the classical IV estimator. We get
1
ˇO D ˙O zx1 ˙O zy and V D ˙zx1 S0 ˙zx
0


if q D k. (8.14)

1 1 1 1
(Use the rule .ABC / DC B A to show this.)

8.1.3 Classical LS Assumptions

Reference: Greene (2000) 9.4 and Hamilton (1994) 8.2.

This section returns to the LS estimator in Section (8.1.1) in order to highlight the
classical LS assumptions that give the variance matrix  2 ˙xx1 .
We allow the regressors to be stochastic, but require that x t is independent of all u t Cs
and that u t is iid. It rules out, for instance, that u t and x t 2 are correlated and also that
the variance of u t depends on x t . Expand the expression for S0 as
p T ! p T !
T X T X
S0 D E xt ut u t x t0 (8.15)
T tD1 T t D1
1
D E .::: C xs 1 us 1 C xs us C :::/ ::: C us 1 xs0 1 C us xs0 C ::: :


T
Note that

E x t s u t s u t x t0 D E x t s x t0 E u t s u t (since u t and x t s independent)

(
0 if s ¤ 0 (since Eus 1 us D 0 by iid u t )
D (8.16)
E x t x t0 E u t u t else.

This means that all cross terms (involving different observations) drop out and that we
can write
T
1X
S0 D E x t x t0 E u2t (8.17)
T t D1
T
1 X
D 2 E x t x t0 (since u t is iid and  2 D E u2t ) (8.18)
T t D1
D  2 ˙xx : (8.19)

124
 Using this in (8.6) gives
V D  2 ˙xx1 : (8.20)

8.1.4 Almost Classical LS Assumptions: White’s Heteroskedasticity.

Reference: Greene (2000) 12.2 and Davidson and MacKinnon (1993) 16.2.
The only difference compared with the classical LS assumptions is that u t is now
allowed to be heteroskedastic, but this heteroskedasticity is not allowed to depend on the
moments of x t . This means that (8.17) holds, but (8.18) does not since Eu2t is not the
same for all t.
However, we can still simplify (8.17) a bit more. We assumed that Ex t x t0 and Eu2t
(which can both be time varying) are not related to each other, so we could perhaps multi-
ply E x t x t0 by ˙ tTD1 E u2t =T instead of by E u2t . This is indeed true asymptotically—where
any possible “small sample” relation between E x t x t0 and E u2t must wash out due to the
assumptions of independence (which are about population moments).
In large samples we therefore have
T T
! !
1X 1 X
S0 D E u2t E x t x t0
T tD1 T t D1
T T
! !
1X 2 1X 0
D E ut E xt xt
T tD1 T t D1
D ! 2 ˙xx ; (8.21)

where ! 2 is a scalar. This is very similar to the classical LS case, except that ! 2 is
the average variance of the residual rather than the constant variance. In practice, the
estimator of ! 2 is the same as the estimator of  2 , so we can actually apply the standard
LS formulas in this case.
This is the motivation for why White’s test for heteroskedasticity makes sense: if the
heteroskedasticity is not correlated with the regressors, then the standard LS formula is
correct (provided there is no autocorrelation).

125
8.1.5 Estimating the Mean of a Process

Suppose u t is heteroskedastic, but not autocorrelated. In the regression y t D ˛ C u t ,

x t D z t D 1. This is a special case of the previous example, since E u2t is certainly
unrelated to E x t x t0 D 1 (since it is a constant). Therefore, the LS covariance matrix
is the correct variance of the sample mean as an estimator of the mean, even if u t are
heteroskedastic (provided there is no autocorrelation).

8.1.6 The Classical 2SLS Assumptions

Reference: Hamilton (1994) 9.2.

The classical 2SLS case assumes that z t is independent of all u tCs and that u t is iid.
The covariance matrix of the moment conditions are
T T
! !
1 X 1 X 0
S0 D E p zt ut p ut zt ; (8.22)
T t D1 T tD1

so by following the same steps in (8.16)-(8.19) we get S0 D  2 ˙zz :The optimal weight-
ing matrix is therefore W D ˙zz1 = 2 (or .Z 0 Z=T / 1 = 2 in matrix form). We use this
result in (8.10) to get
  1
ˇO2SLS D ˙O xz ˙O zz1 ˙O zx ˙O xz ˙O zz1 ˙O zy ; (8.23)

which is the classical 2SLS estimator.

Since this GMM is efficient (for a given set of moment conditions), we have estab-
lished that 2SLS uses its given set of instruments in the efficient way—provided the clas-
sical 2SLS assumptions are correct. Also, using the weighting matrix in (8.13) gives
  1
1
V D ˙xz 2 ˙zz1 ˙zx : (8.24)


8.1.7 Non-Linear Least Squares

Consider the non-linear regression

y t D F .x t I ˇ0 / C " t ; (8.25)

126
where F .x t I ˇ0 / is a potentially non-linear equation of the regressors x t , with a k  1
vector of parameters ˇ0 . The non-linear least squares (NLS) approach is minimize the
sum of squared residuals, that is, to solve

ˇO D arg min TtD1 Œy t F .x t I ˇ/2 :

P
(8.26)

To express this as a GMM problem, use the first order conditions for (8.26) as moment
conditions
1 PT @F .x t I ˇ/
mN .ˇ/ D Œy t F .x t I ˇ/ : (8.27)
T tD1 @ˇ
The model is then exactly identified so the point estimates are found by setting all moment
N .ˇ/ D 0k1 .
conditions to zero , m
p
As usual, S0 D CovŒ T m N .ˇ0 /, while the Jacobian is

N 0/
@m.ˇ
D0 D plim
@ˇ 0
1 PT @F .x t I ˇ/ @F .x t I ˇ/ 1 PT @2 F .x t I ˇ/
D plim plim Œy t F .x t I ˇ/ :
T t D1 @ˇ @ˇ 0 T t D1 @ˇ@ˇ 0
(8.28)

Example 8.1 (With two parameters) With ˇ D Œˇ1 ; ˇ2 0 we have

" #
@F .x t I ˇ/ @F .x t I ˇ/=@ˇ1 @F .x t I ˇ/ h i
D ; D @F .x t I ˇ/=@ˇ 1 @F .x t I ˇ/=@ˇ2 :
@ˇ @F .x t I ˇ/=@ˇ2 @ˇ 0

The moment conditions are

" #
1 PT @F .x t I ˇ/=@ˇ1
N .ˇ/ D
m Œy t F .x t I ˇ/ ;
T t D1 @F .x t I ˇ/=@ˇ2

which is a 2  1 vector. Notice that the outer product of the gradient (first term) in (8.28)
is a 2  2 matrix. Similarly, the matrix with the second derivatives (the Hessian) is also a
2  2 matrix " 2
@2 F .x t Iˇ /
#
@ F .x t Iˇ /
@2 F .x t I ˇ/
D @2@ˇ 1 @ˇ1
F .x t Iˇ /
@ˇ1 @ˇ2
@2 F .x t Iˇ /
:
@ˇ@ˇ 0 @ˇ @ˇ @ˇ @ˇ
2 1 2 2

Example 8.2 (Linear regression function as a special case) When F .x t I ˇ/ D x t0 ˇ, then

PT
x t0 ˇ =T .


@F .x t I ˇ/=@ˇ D x t , so the moment conditions are m
N .ˇ/ D tD1 x t y t
Since the second derivatives are zero, (8.28) becomes D0 D plim TtD1 x t x t0 =T , which is
P

127
the same in the LS case (except possibly for the sign of D0 , but that is of no consequence
since it is only the square of D0 that matters.)

Example 8.3 (Logistic smooth transition regression) Let G.z/ be a logistic (increasing
but “S-shaped”) function
1
G.z/ D ;
1 C expŒ .z c/

where the parameter c is the central location (where G.z/ D 1=2) and > 0 determines
the steepness of the function (a high implies that the function goes quickly from 0 to 1
around z D c.) See Figure 8.1 for an illustration. A logistic smooth transition regression
is

G.z t / ˇ10 C G.z t /ˇ20 x t C " t

˚
y t D Œ1
„ ƒ‚ …
F .x t Iˇ / in (8.25)

D Œ1 G.z t / ˇ10 x t C G.z t /ˇ20 x t C " t :

The regression coefficients vary smoothly with z t : from ˇ1 at low values of z t to ˇ2 at high
values of z t . See Figure 8.1 for an illustration. The parameter vector ( ; c; ˇ1 ; ˇ2 —called
just ˇ in (8.25)) is easily estimated by NLS by concentrating the loss function: optimize
(numerically) over . ; c/ and let (for each value of . ; c/) the parameters (ˇ1 ; ˇ2 ) be the
OLS coefficients on the vector of “regressors” .Œ1 G.z t / x t ; G.z t /x t /. The most com-
mon application of this model is obtained by letting x t D y t s (this is the LSTAR model—
logistic smooth transition auto regression model), see Franses and van Dijk (2000).

8.1.8 Moment Conditions with Spuriously Extended Sample 1

One way to handle unbalanced panels (when there is more data on sone variables than
on others), is to artificially expand the sample and then interact the moment conditions
with a dummy variable to pick out the correct subsample. This example illustrates how
and why that works. To keep it simple, the example discusses the case of estimating a
sample mean of x t —for which we have data over the sample t D 1 to  and the sample
is artificially extended to include T  data points.

128
 Logistic function, different γ Coefficient on xt, different β2
1
0.4 β2=0.5

0.3 β2=0.25
0.5
0.2 β2=0
γ=1
γ=5 0.1
0 0
−2 0 2 −2 0 2
z z
G(z) = 1/[1+exp(−γ(z−c))], c=0 yt = [1−G(zt)]β1xt + G(zt)β2xt + εt, β1=0.25
G(zt) is a logistic function with γ=2 and c=0

Figure 8.1: Logistic function and the effective slope coefficient in a Logistic smooth
transition regression

To estimate the mean we specify the moment condition

(
1 t D 1; ::; 
m t D d t .x t / , with d t D (8.29)
0 t D  C 1; :::; T

so the moment conditions look like

2 3
x1 
6 :: 7
6
6 : 7
7
6 7
6 x  7
6 7 (8.30)
6
6 0 7
7
6 :: 7
6
4 : 7
5
0
p
If Var.x t / D  2 , then the variance CovΠT m
N .ˇ0 / is

 2 C .T /0  2
S0 D D  (8.31)
T T
N 0 /=@ˇ 0 ) is
and the Jacobian (plim @m.ˇ

D0 D : (8.32)
T

129
(This is also what a numerical derivative based on a function for the moment conditions
would give.)
Combining gives
p d
T .O / ! N.0; V /; where
 1
 
   2 1
V D   
T T T
T   2  T
D   
 T 
T
D 2  : (8.33)

Therefore,
2
O D V =T D
Var./ ; (8.34)

which is the correct result—the artificial extension of the sample does not lead to a spu-
riously low uncertainty. This demonstrates that the aritificial and spurious extension of
the sample actually does no harm: the inference based on standard GMM formulas is still
correct.

8.1.9 Moment Conditions with Spuriously Extended Sample 2 (Dummies for Miss-
ing Values)

Consider the simple regression equation

y t D bx t C " t (8.35)

and suppose the sample length is T , but only the first  observations have full data, while
the last T  observations include some missing values. (Putting these observations last
is just a matter of convenience.)
Suppose we prune (“excise”) the sample by simply skipping the observations with
missing values. Under the standard iid assumptions, we then have that the LS estimate
O is distributed as
(b)
p d
T .bO b0 / ! N .0k1 ; V / ; (8.36)

130
where the covariance matrix is

! 1 
2 1X 2 1X 2
V D plim xt xt and  D plim " : (8.37)
 t D1  tD1 t

Instead, suppose we use all T observations, but let d t D 1 if there is data for period t
and zero otherwise. This gives the sample moment condition
T
1X
N D
m d t x t .y t bx t / (8.38)
T t D1

The Jacobian is
T
1X
D0 D plim dt xt xt (8.39)
T t D1
and the covariance of the moment conditions (under the standard iid assumptions)
T T T
1X 1X 1X
S0 D plim d t x t x t d t "2t D s 2 plim d t x t x t ; where s 2 D plim d t "2t :
T t D1 T t D1 T t D1
(8.40)
Combining as in (8.6) gives the covariance matrix
1
T
!
1X
V b D s2 dt xt xt : (8.41)
T tD1

To see that this is the same as in (8.37), notice that

T
X 
X
dt xt xt D x t x t , and (8.42)
t D1 t D1
T 
2 1X 2 1X 2 
s D dt "t D "t D  2:
T t D1 T t D1 T

Using this in (8.40)–(8.41) gives


! 1 
! 1
 1X 1X
V D 2
b
xt xt D 2
xt xt :
T T t D1  t D1

which is the same as in (8.37). This makes a lot of sense since the dummy approach is just

131
about nullifying the effect of the periods with missing values. In a sense this makes the
Jacobian too small, but that is compensated for by making S0 too large. This demonstrates
that the estimation could be done in either way.

8.2 Identification of Systems of Simultaneous Equations

Reference: Greene (2000) 16.1–3

This section shows how the GMM moment conditions can be used to understand if
the parameters in a system of simultaneous equations are identified or not.
The structural model (form) is

F y t C Gz t D u t ; (8.43)

where y t is a vector of endogenous variables, z t a vector of predetermined (exogenous)

variables, F is a square matrix, and G is another matrix.1 We can write the j th equation
of the structural form (8.43) as
yjt D x t0 ˇ C ujt ; (8.44)

where x t contains the endogenous and exogenous variables that enter the j th equation
with non-zero coefficients, that is, subsets of y t and z t .
We want to estimate ˇ in (8.44). Least squares is inconsistent if some of the regressors
are endogenous variables (in terms of (8.43), this means that the j th row in F contains
at least one additional non-zero element apart from coefficient on yjt ). Instead, we use
IV/2SLS. By assumption, the structural model summarizes all relevant information for
the endogenous variables y t . This implies that the only useful instruments are the vari-
ables in z t . (A valid instrument is uncorrelated with the residuals, but correlated with the
regressors.) The moment conditions for the j th equation are then
T
1X
x t0 ˇ x t0 ˇ D 0:



E z t yjt D 0 with sample moment conditions z t yjt
T t D1
(8.45)
If there are as many moment conditions as there are elements in ˇ, then this equation
is exactly identified, so the sample moment conditions can be inverted to give the Instru-
1
By premultiplying with F 1 and rearranging we get the reduced form y t D ˘ z t C" t , with ˘ D F 1

and Cov." t / D F 1 Cov.u t /.F 1 /0 .

132
mental variables (IV) estimator of ˇ. If there are more moment conditions than elements
in ˇ, then this equation is overidentified and we must devise some method for weighting
the different moment conditions. This is the 2SLS method. Finally, when there are fewer
moment conditions than elements in ˇ, then this equation is unidentified, and we cannot
hope to estimate the structural parameters of it.
We can partition the vector of regressors in (8.44) as x t0 D ŒzQ t0 ; yQ t0 , where y1t and z1t
are the subsets of z t and y t respectively, that enter the right hand side of (8.44). Partition
z t conformably z t0 D ŒzQ t0 ; z t0 , where z t are the exogenous variables that do not enter
(8.44). We can then rewrite the moment conditions in (8.45) as
" # " #0 !
zQ t zQ t
E yjt ˇ D 0: (8.46)
z t yQ t

yjt D Gj zQ t Fj yQ t C ujt
D x t0 ˇ C ujt , where x t0 D zQ t0 ; yQ t0 ;
 
(8.47)

This shows that we need at least as many elements in z t as in yQ t to have this equations
identified, which confirms the old-fashioned rule of thumb: there must be at least as
many excluded exogenous variables (z t ) as included endogenous variables (yQ t ) to have
the equation identified.
This section has discussed identification of structural parameters when 2SLS/IV, one
equation at a time, is used. There are other ways to obtain identification, for instance, by
imposing restrictions on the covariance matrix. See, for instance, Greene (2000) 16.1-3
for details.

Example 8.4 (Supply and Demand. Reference: GR 16, Hamilton 9.1.) Consider the
simplest simultaneous equations model for supply and demand on a market. Supply is

q t D p t C ust ; > 0;

and demand is
q t D ˇp t C ˛A t C udt ; ˇ < 0;

where A t is an observable exogenous demand shock (perhaps income). The only mean-

133
ingful instrument is A t . From the supply equation we then get the moment condition

E A t .q t p t / D 0;

which gives one equation in one unknown, . The supply equation is therefore exactly
identified. In contrast, the demand equation is unidentified, since there is only one (mean-
ingful) moment condition

E A t .q t ˇp t ˛A t / D 0;

but two unknowns (ˇ and ˛).

Example 8.5 (Supply and Demand: overidentification.) If we change the demand equa-
tion in Example 8.4 to

q t D ˇp t C ˛A t C bB t C udt ; ˇ < 0:

There are now two moment conditions for the supply curve (since there are two useful
instruments) " # " #
A t .q t p t / 0
E D ;
B t .q t p t / 0
but still only one parameter: the supply curve is now overidentified. The demand curve is
still underidentified (two instruments and three parameters).

8.3 Testing for Autocorrelation

This section discusses how GMM can be used to test if a series is autocorrelated. The
analysis focuses on first-order autocorrelation, but it is straightforward to extend it to
higher-order autocorrelation.
Consider a scalar random variable x t with a zero mean (it is easy to extend the analysis
to allow for a non-zero mean). Consider the moment conditions
T
" # " # " #
x t2  2 1 X x t2  2 2
m t .ˇ/ D N
; so m.ˇ/ D , with ˇ D :
x t x t 1  2 T t D1 x t x t 1  2 
(8.48)
2 2
 is the variance and  the first-order autocorrelation so  is the first-order autocovari-

134
ance. We want to test if  D 0. We could proceed along two different routes: estimate
 and test if it is different from zero or set  to zero and then test overidentifying restric-
tions. We analyze how these two approaches work when the null hypothesis of  D 0 is
true.

8.3.1 Estimating the Autocorrelation Coefficient

We estimate both  2 and  by using the moment conditions (8.48) and then test if  D
0. To do that we need to calculate the asymptotic variance of O (there is little hope of
being able to calculate the small sample variance, so we have to settle for the asymptotic
variance as an approximation).
We have an exactly identified system so the weight matrix does not matter—we can
then proceed as if we had used the optimal weighting matrix (all those results apply).
To find the asymptotic covariance matrix of the parameters estimators, we need the
probability limit of the Jacobian of the moments and the covariance matrix of the moments—
evaluated at the true parameter values. Let m N i .ˇ0 / denote the i th element of the m.ˇ/
N
vector—evaluated at the true parameter values. The probability of the Jacobian is
" # " # " #
@mN 1 .ˇ0 /=@ 2 @mN 1 .ˇ0 /=@ 1 0 1 0
D0 D plim D D ;
@mN 2 .ˇ0 /=@ 2 @mN 2 .ˇ0 /=@  2 0 2
(8.49)
since  D 0 (the true value). Note that we differentiate with respect to  2 , not  , since
we treat  2 as a parameter.
The covariance matrix is more complicated. The definition is
"p T
# "p T #0
T X T X
S0 D E m t .ˇ0 / m t .ˇ0 / :
T t D1 T t D1

Assume that there is no autocorrelation in m t .ˇ0 /. We can then simplify as

S0 D E m t .ˇ0 /m t .ˇ0 /0 .

This assumption is stronger than assuming that  D 0, but we make it here in order to
illustrate the asymptotic distribution. To get anywhere, we assume that x t is iid N.0;  2 /.

135
In this case (and with  D 0 imposed) we get
" #" #0 " #
x t2  2 x t2  2 .x t2  2 /2 .x t2  2 /x t x t 1
S0 D E DE
xt xt 1 xt xt 1 .x t2  2 /x t x t 1 .x t x t 1 /2
" # " #
E x t4 2 2 E x t2 C  4 0 2 4 0
D D : (8.50)
0 E x t2 x t2 1 0 4

To make the simplification in the second line we use the facts that E x t4 D 3 4 if x t 
N.0;  2 /, and that the normality and the iid properties of x t together imply E x t2 x t2 1 D
E x t2 E x t2 1 and E x t3 x t 1 D E  2 x t x t 1 D 0.
By combining (8.49) and (8.50) we get that
" #!
p O 2  0
1
 1
ACov T D D0 S0 D0
O
0" #0 " # 1" #1 1
4
1 0 2 0 1 0
D@ 2 4
A
0  0  0 2
" #
2 4 0
D : (8.51)
0 1
p
O
This shows the standard expression for the uncertainty of the variance and that the T .
p
Since GMM estimators typically have an asymptotic distribution we have T O !d
N.0; 1/, so we can test the null hypothesis of no first-order autocorrelation by the test
statistics
T O2  21 . (8.52)

This is the same as the Box-Ljung test for first-order autocorrelation.

This analysis shows that we are able to arrive at simple expressions for the sampling
uncertainty of the variance and the autocorrelation—provided we are willing to make
strong assumptions about the data generating process. In particular, ewe assumed that
data was iid N.0;  2 /. One of the strong points of GMM is that we could perform similar
tests without making strong assumptions—provided we use a correct estimator of the
asymptotic covariance matrix S0 (for instance, Newey-West).

136
8.3.2 Testing the Overidentifying Restriction of No Autocorrelation

We can estimate  2 alone and then test if both moment condition are satisfied at  D 0.
There are several ways of doing that, but the perhaps most straightforward is skip the loss
function approach to GMM and instead specify the “first order conditions” directly as

0 D Am
N
T
" #
h i1X x t2  2
D 1 0 ; (8.53)
T t D1 xt xt 1

which sets O 2 equal to the sample variance.

The only parameter in this estimation problem is  2 , so the matrix of derivatives
becomes " # " #
@m N 1 .ˇ0 /=@ 2 1
D0 D plim D : (8.54)
@m N 2 .ˇ0 /=@ 2 0
By using this result, the A matrix in (8.54) and the S0 matrix in (8.50,) it is straighforward
to calculate the asymptotic covariance matrix the moment conditions. In general, we have
p
ACovŒ T m. O D ŒI
N ˇ/ D0 .A0 D0 / 1
A0 S0 ŒI D0 .A0 D0 / 1
A0  0 : (8.55)

The term in brackets is here (since A0 D A since it is a matrix with constants)

0 1 1
" # " #B " #C " #
1 0 1 B h i 1 C h i 0 0
B 1 0 C 1 0 D : (8.56)
0 1 0 B
@„ ƒ‚ … 0 C
A „ ƒ‚ … 0 1
„ ƒ‚ … „ ƒ‚ … A0 „ ƒ‚ … A0
I2 D0 D0

We therefore get
" #" #" #0 " #
p 0 0 2 4 0 0 0 0 0
ACovΠT m. O D
N ˇ/ D : (8.57)
0 1 0 4 0 1 0 4

Note that the first moment condition has no sampling variance at the estimated parameters,
since the choice of O 2 always sets the first moment condition equal to zero.
The test of the overidentifying restriction that the second moment restriction is also

137
zero is  p C
0
N ACovΠT m.
Tm O
N ˇ/ m N  21 ; (8.58)

where we have to use a generalized inverse if the covariance matrix is singular (which it
is in (8.57)).
In this case, we get the test statistics (note the generalized inverse)
" #0 " #" #  T 2
0 0 0 0 ˙ t D1 x t x t 1 =T
T DT ;
˙ tTD1 x t x t 1 =T 0 1= 4 T
˙ tD1 x t x t 1 =T 4
(8.59)
which is the T times the square of the sample covariance divided by  4 . A sample cor-
relation, , O would satisfy ˙ tTD1 x t x t 1 =T D OO 2 , which we can use to rewrite (8.59) as
T O2 O 4 = 4 . By approximating  4 by O 4 we get the same test statistics as in (8.52).

8.4 Estimating and Testing a Normal Distribution

8.4.1 Estimating the Mean and Variance

This section discusses how the GMM framework can be used to test if a variable is nor-
mally distributed. The analysis cold easily be changed in order to test other distributions
as well.
Suppose we have a sample of the scalar random variable x t and that we want to test if
the series is normally distributed. We analyze the asymptotic distribution under the null
hypothesis that x t is N.;  2 /.
We specify four moment conditions
2 3 2 3
xt  xt 
T 6
6 .x t /2  2 7 6 .x t /2  2 7
6 7 7
1 X
mt D 6 3
7 so mN D 6
3
7 (8.60)
6 .x
4 t /
7
5 T t D1 6
4 .x t /
7
5
.x t /4 3 4 .x t /4 3 4

Note that E m t D 041 if x t is normally distributed.

Let mN i .ˇ0 / denote the i th element of the m.ˇ/
N vector—evaluated at the true parameter

138
values. The probability of the Jacobian is
2 3
@mN 1 .ˇ0 /=@ @m N 1 .ˇ0 /=@ 2
2 7
6 7
6 @m N 2 .ˇ0 /=@ @ N
m 2 .ˇ0 /=@
D0 D plim 6 7
2 7
6 @m
4 N 3 .ˇ0 /=@ @ N
m 3 .ˇ0 /=@ 5
2
@mN 4 .ˇ0 /=@ @m N 4 .ˇ0 /=@
2 3 2 3
1 0 1 0
T 6 7 6 7
1 X6 2.x t / 1 7 6 0 1 7
D plim 6
2
7D6 7: (8.61)
T tD1 6 4 3.x t / 0 7 6
5 4 3 2 0 7
5
3 2
4.x t / 6 0 6 2

(Recall that we treat  2 , not , as a parameter.)

The covariance matrix of the scaled moment conditions (at the true parameter values)
is "p T # "p T #0
T X T X
S0 D E m t .ˇ0 / m t .ˇ0 / ; (8.62)
T t D1 T t D1
which can be a very messy expression. Assume that there is no autocorrelation in m t .ˇ0 /,
which would certainly be true if x t is iid. We can then simplify as

S0 D E m t .ˇ0 /m t .ˇ0 /0 ; (8.63)

which is the form we use here for illustration. We therefore have (provided m t .ˇ0 / is not
autocorrelated)
02 31 02 310 2 3
xt  xt  2 0 3 4 0
2 2 2 2 4
12 6
B6 7 C B 6 7 C 6 7
B6 .x t /  7 7C B6 .x t /  7 7C D 6 0 2 0
C B6 C 6 7
S0 D E B 6 7:
B6
@4 .x t /3 7C B6
5A @4 .x t /3 7C
5A
6 3 4
4 0 15 6 0 7
5
4 4 4 4 6
.x t / 3 .x t / 3 0 12 0 96 8
(8.64)
It is straightforward to derive this result once we have the information in the following
remark.

Remark 8.6 If X  N.;  2 /, then the first few moments around the mean of a are
E.X / D 0, E.X /2 D  2 , E.X /3 D 0 (all odd moments are zero), E.X /4 D
3 4 , E.X /6 D 15 6 , and E.X /8 D 105 8 .

139
 Suppose we use the efficient weighting matrix. The asymptotic covariance matrix of
the estimated mean and variance is then (.D00 S0 1 D0 / 1 )
02 30 2 3 1 2 31 1
1 0 2 0 3 4 0 1 0 " # 1
1
2 4 12 6 7
B6 7 6 7 6 7C
B6 0 1 7 6 0 0 6 0 1 7C
2
0
B6 7 6 7 6 7C D 1
B6
@4 3 2 0 7
5
6
4 3 4 0 15 6 0 7 5
6
4 3 2 0 7C
5A 0 2 4
0 6 2 0 12 6 0 96 8 0 6 2
" #
2 0
D :
0 2 4
(8.65)

This is the same as the result from maximum likelihood estimation which use the sample
mean and sample variance as the estimators. The extra moment conditions (overidenti-
fying restrictions) does not produce any more efficient estimators—for the simple reason
that the first two moments completely characterizes the normal distribution.

8.4.2 Testing Normality

The payoff from the overidentifying restrictions is that we can test if the series is actually
normally distributed. There are several ways of doing that, but the perhaps most straight-
forward is skip the loss function approach to GMM and instead specify the “first order
conditions” directly as

0 D Am
N
2 3
xt 
T 6
" #
.x t /2 2 7
7
1 0 0 0 1 X6
D 6 7: (8.66)
0 1 0 0 T t D1 6
4 .x t /3 7
5
.x t /4 3 4

The asymptotic covariance matrix the moment conditions is as in (8.55). In this case,

140
the matrix with brackets is
0 1 1
2 3 2 3B 2 3C
1 0 0 0 1 0 B 1 0 C
6 7 6 7B " #6 7C " #
6
6 0 1 0 0 7
7
6
6 0 1 7B
7B 1 0 0 0 6
6 0 1 7C
7C 1 0 0 0
2 7B 2 7C
6
4 0 0 1 0 7
5
6
4 3 0 5B 0 1 0 0 4
B 6 3 0 5C
C 0 1 0 0
6 2 B„ 6 2
ƒ‚ … „ ƒ‚ …
0 0 0 1 0 @ A0 0 C
A A0
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
I4 D0 D0
2 3
0 0 0 0
6 7
6 0 0 0 0 7
D6 7 (8.67)
6
4 3 2 0 1 0 7
5
0 6 2 0 1

We therefore get
2 32 32 30
0 0 0 0 2 0 3 4 0 0 0 0 0
p 2 4 12 6 7
6 76 76 7
O
6 0 0 0 0 7 6 0 0 6 0 0 0 0 7
N ˇ/ D 6
ACovΠT m. 76 76 7
6
4 3 2 0 1 0 7
54
6 3 4 0 15 6 0 7 54
6 3 2 0 1 0 7
5
0 6 2 0 1 0 12 6 0 96 8 0 6 2 0 1
2 3
0 0 0 0
6 7
6 0 0 0 0 7
D6
6 0 0 6 6
7 (8.68)
4 0 7
5
0 0 0 24 8

We now form the test statistics for the overidentifying restrictions as in (8.58). In this
case, it is (note the generalized inverse)
2 30 2 32 3
0 0 0 0 0 0
6 7 6 76 7
6 0 7 6 0 0 0 0 76 0 7
T66 ˙ T .x 3
7 6
7 6 0 0 1=.6 6 /
76
7 6 ˙ T .x
7
4 t D1 t / =T 5 4 0 5 4 t D1 t /3 =T 7
5
T
˙ t D1 Œ.x t / 4 4
3 =T 0 0 0 1=.24 /8
˙ tTD1 Œ.x t /4 3 4 =T
2 2
T ˙ tTD1 .x t /3 =T T ˙ tTD1 Œ.x t /4 3 4 =T
 ˚
D C : (8.69)
6 6 24 8
When we approximate  by O then this is the same as the Jarque and Bera test of nor-

141
mality.
The analysis shows (once again) that we can arrive at simple closed form results by
making strong assumptions about the data generating process. In particular, we assumed
that the moment conditions were serially uncorrelated. The GMM test, with a modified
estimator of the covariance matrix S0 , can typically be much more general.

8.5 Testing the Implications of an RBC Model

Reference: Christiano and Eichenbaum (1992)

This section shows how the GMM framework can be used to test if an RBC model fits
data.
Christiano and Eichenbaum (1992) try to test if the RBC model predictions correspond
are significantly different from correlations and variances of data. The first step is to define
a vector of parameters and some second moments
  y 
c p
D ı; :::;  ; ; :::; Corr ;n ; (8.70)
y n

and estimate it with GMM using moment conditions. One of the moment condition is
that the sample average of the labor share in value added equals the coefficient on labor
in a Cobb-Douglas production function, another is that just the definitions of a standard
deviation, and so forth.
The distribution of the estimator for is asymptotically normal. Note that the covari-
ance matrix of the moments is calculated similarly to the Newey-West estimator.
The second step is to note that the RBC model generates second moments as a function
h .:/ of the model parameters fı; :::;  g, which are in , that is, the model generated
second moments can be thought of as h . /.
The third step is to test if the non-linear restrictions of the model (the model mapping
from parameters to second moments) are satisfied. That is, the restriction that the model
second moments are as in data
  y 
c p
H . / D h . / ; :::; Corr ; n D 0; (8.71)
y n

is tested with a Wald test. (Note that this is much like the Rˇ D 0 constraints in the linear

142
case.) From the delta-method we get
p 0
 
@Hd @H
T H. O / ! N 0; Cov. O / : (8.72)
@ 0 @
Forming the quadratic form
 0
 1
@H O / @H
TH. O / 0
Cov. H. O /; (8.73)
@ 0 @

will as usual give a 2 distributed test statistic with as many degrees of freedoms as
restrictions (the number of functions in (8.71)).

8.6 IV on a System of Equations

Suppose we have two equations

0
y1t D x1t ˇ1 C u1t
0
y2t D x2t ˇ2 C u2t ;

and two sets of instruments, z1t and z2t with the same dimensions as x1t and x2t , respec-
tively. The sample moment conditions are
T
"
#
0
1 X z1t y1t x1t ˇ1
N 1 ; ˇ2 / D
m.ˇ 0

;
T t D1 z2t y2t x2t ˇ2

Let ˇ D .ˇ10 ; ˇ20 /0 . Then

" @ 1
PT 0

@ 1
PT 0

#
N 1 ; ˇ2 /
@m.ˇ @ˇ10 T t D1 z1t y1t x1t ˇ1 @ˇ20 T t D1 z1t y1t x1t ˇ1
D @ 1
PT @ 1
PT
@ˇ 0 0 0



@ˇ10 T tD1 z2t y2t x2t ˇ2 @ˇ20 T t D1 z2t y2t x2t ˇ2
" PT #
1 0
T t D1 z1t x1t 0
D 1
PT 0
:
0 T t D1 z2t x2t

143
This is invertible so we can premultiply the first order condition with the inverse of
0 0
N
Œ@m.ˇ/=@ˇ N
 A and get m.ˇ/ D 0k1 . We can solve this system for ˇ1 and ˇ2 as
" # " PT # 1" P #
1 0 1 T
ˇ1 z x
tD1 1t 1t 0 z y
1t 1t
D T
1
PT 0
T
1
PTt D1
ˇ2 0 T t D1 z2t x2t T t D1 z2t y2t
2  P  1 3"
1 T 0 PT #
t D1 z1t x1t 0 1
T z 1t y1t
D4  P  1 5 T1 PTtD1 :
0 1 T
z 2t x 0
T t D1 z2t y2t
T tD1 2t

This is IV on each equation separately, which follows from having an exactly identified
system.

Bibliography
Christiano, L. J., and M. Eichenbaum, 1992, “Current real-business-cycle theories and
aggregate labor-market fluctuations,” American Economic Review, 82, 430–450.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

Franses, P. H., and D. van Dijk, 2000, Non-linear time series models in empirical finance,
Cambridge University Press.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

144
12 Vector Autoregression (VAR)
Reference: Hamilton (1994) 10-11; Greene (2000) 17.5; Johnston and DiNardo (1997)
9.1-9.2 and Appendix 9.2; and Pindyck and Rubinfeld (1998) 9.2 and 13.5.
Let y t be an n  1 vector of variables. The VAR(p) is

y t D  C A1 y t 1 C ::: C Ap y t p C " t ; " t is white noise, Cov(" t ) D ˝: (12.1)

Example 12.1 (VAR(2) of 2  1 vector.) Let y t D Œ x t z t 0 . Then

" # " #" # " #" # " #
xt A1;11 A1;12 xt 1 A2;11 A2;12 xt 2 "1;t
D C C :
zt A1;21 A1;22 zt 1 A2;21 A2;22 zt 2 "2;t
(12.2)

Issues:

 Variable selection

 Lag length

 Estimation

 Purpose: data description (Granger-causality, impulse response, forecast error vari-

ance decomposition), forecasting, policy analysis (Lucas critique)?

12.1 Canonical Form

A VAR(p) can be rewritten as a VAR(1). For instance, a VAR(2) can be written as

" # " # " #" # " #
yt  A1 A2 yt 1 "t
D C C or (12.3)
yt 1 0 I 0 yt 2 0
y t D  C Ay t 1 C "t : (12.4)

145
Example 12.2 (Canonical form of a univariate AR(2).)
" # " # " #" # " #
yt  a1 a2 yt 1 "t
D C C .
yt 1 0 1 0 yt 2 0

Example 12.3 (Canonical for of VAR(2) of 21 vector.) Continuing on the previous
example, we get
2 3 2 32 3 2 3
xt A1;11 A1;11 A2;11 A2;12 xt 1 "1;t
6 7 6 76 7 6 7
6 zt 7 6 A1;21 A1;22 A2;21 A2;22 76 zt 1 7 6 "2;t 7
6 7D6 76 7C6 7:
6 x 7 6 1 0 0 0 76 xt 7 6 0 7
4 t 1 5 4 54 2 5 4 5
zt 1 0 1 0 0 zt 2 0

12.2 Moving Average Form and Stability

Consider a VAR(1), or a VAR(1) representation of a VAR(p) or an AR(p)

y t D Ay t 1 C "t : (12.5)

Solve recursively backwards (substitute for y t s D Ay t s 1 C "t s , s D 1; 2;...) to get

the vector moving average representation (VMA), or impulse response function

y t D A Ay t C "t C "t


2 1

D A2 y t 2 C A"t
C "t 1

D A2 Ay t 3 C "t 2 C A"t C "t

1

D A3 y t 3 C A2 "t 2 C A"t 1 C "t

::
:
K
X
D AKC1 y t K 1C As "t s : (12.6)
sD0

Remark 12.4 (Spectral decomposition.) The n eigenvalues (i ) and associated eigen-
vectors (zi ) of the n  n matrix A satisfies

.A i In / zi D 0n1 :

146
If the eigenvectors are linearly independent, then
2 3
1 0


0
6 0 2


0
6 7 h i
A D ZZ 1 , where  D 6
7
7 and Z D z1 z2


zn :
6 :: :: :
4 : :


::
7
5
0 0


n

Note that we therefore get

A2 D AA D ZZ 1
ZZ 1
D ZZ 1
D Z2 Z 1
) Aq D Zq Z 1
:
p
Remark 12.5 (Modulus of complex number.) If  D a C bi, where i D 1, then
p
jj D ja C bij D a2 C b 2 .

We want limK!1 AKC1 y t K 1 D 0 (stable VAR) to get a moving average represen-

tation of y t (where the influence of the starting values vanishes asymptotically). We note
from the spectral decompositions that AKC1 D ZKC1 Z 1 , where Z is the matrix of
eigenvectors and  a diagonal matrix with eigenvalues. Clearly, limK!1 AKC1 y t K 1 D
0 is satisfied if the eigenvalues of A are all less than one in modulus.

Example 12.6 (AR(1).) For the univariate AR(1) y t D ay t 1 C " t , the characteristic
equation is .a / z D 0, which is only satisfied if the eigenvalue is  D a. The AR(1) is
therefore stable (and stationarity) if 1 < a < 1.

If we have a stable VAR, then (12.6) can be written

1
X
y t D As "t s (12.7)
sD0

D "t C A"t 1 C A2 "t 2 C :::

We may pick out the first n equations from (12.7) (to extract the “original” variables from
the canonical form) and write them as

y t D " t C C1 " t 1 C C2 " t 2 C :::; (12.8)

which is the vector moving average, VMA, form of the VAR.

147
Example 12.7 (AR(2), Example (12.2) continued.) Let  D 0 in 12.2 and note that the
VMA of the canonical form is
" # " # " #" # " #" #
yt "t a1 a2 "t 1 a12 C a2 a1 a2 "t 2
D C C C :::
yt 1 0 1 0 0 a1 a2 0

The MA of y t is therefore

C a12 C a2 " t


y t D " t C a1 " t 1 2 C :::

Note that
@y t @E t y t Cs
0
D Cs or D Cs ; with C0 D I (12.9)
@" t s @"0t
so the impulse response function is given by fI; C1 ; C2 ; :::g. Note that it is typically only
meaningful to discuss impulse responses to uncorrelated shocks with economic interpreta-
tions. The idea behind structural VARs (discussed below) is to impose enough restrictions
to achieve this.

Example 12.8 (Impulse response function for AR(1).) Let y t D y t 1 C " t . The MA
Pt s s
representation is y t D sD0  " t s , so @y t =@" t s D @E t y t Cs =@" t D  . Stability
requires jj < 1, so the effect of the initial value eventually dies off (lims!1 @y t =@" t s D
0).

Example 12.9 (Numerical VAR(1) of 21 vector.) Consider the VAR(1)

" # " #" # " #
xt 0:5 0:2 xt 1 "1;t
D C :
zt 0:1 0:3 zt 1 "2;t

The eigenvalues are approximately 0:52 and 0:32, so this is a stable VAR. The VMA is
" # " # " #" # " #" #
xt "1;t 0:5 0:2 "1;t 1 0:27 0:04 "1;t 2
D C C C :::
zt "2;t 0:1 0:3 "2;t 1 0:02 0:11 "2;t 2

Remark 12.10 (Generalized impulse response function) The impulse response function
makes most sense when the shocks (" t ) are uncorrelated. If they are not, then a gener-
alized impulse response function (see Pesaran and Shin (1998)) might still be useful: in
(12.8), replace Cs by Cs ˙.diag.˙// 1 , where ˙ D Cov." t /. It can be interpreted as fol-
lows: what is the typical response to "i t —when we incorporate the typical comovement

148
of the other shocks with "i t ? In practice, this means that the impulse response function
with respect to "i t is not calculated by setting " t D ei where ei is a vector of zeros except
that element i is unity. Instead, we use E." t j"i t D 1/ D ˙i =i i where ˙i is column i of
˙ (that is, Cov." t ; "i t /), so the value of the entire vector " t is predicted (assuming that
" t is normally distributed) using the information "i t D 1. Clearly, when the shocks are
uncorrelated, then this approach coincides with a traditional impulse response function
(where the impulse is "i t D 1 and "jt D 0 for j ¤ i ).

12.3 Estimation

The MLE, conditional on the initial observations, of the VAR is the same as OLS esti-
mates of each equation separately. The MLE of the ij t h element in Cov(" t ) is given by
PT
O i t vOjt =T , where vO i t and vOjt are the OLS residuals.
t D1 v
Note that the VAR system is a system of “seemingly unrelated regressions,” with the
same regressors in each equation. The OLS on each equation is therefore the GLS, which
coincides with MLE if the errors are normally distributed.

12.4 Granger Causality

Main message: Granger-causality might be useful, but it is not the same as causality.
Definition: if z cannot help forecast x, then z does not Granger-cause x; the MSE of
the forecast E. x t j x t s ; z t s ; s > 0/ equals the MSE of the forecast E.x t j x t s ; s > 0/.
Test: Redefine the dimensions of x t and z t in (12.2): let x t be n1 1 and z t is n2 1. If
the n1  n2 matrices A1;12 D 0 and A2;12 D 0, then z fail to Granger-cause x. (In general,
we would require As;12 D 0 for s D 1; :::; p.) This carries over to the MA representation
in (12.8), so Cs;12 D 0.
These restrictions can be tested with an F-test. The easiest case is when x is a scalar,
since we then simply have a set of linear restrictions on a single OLS regression.

Example 12.11 (RBC and nominal neutrality.) Suppose we have an RBC model which
says that money has no effect on the real variables (for instance, output, capital stock,
and the productivity level). Money stock should not Granger-cause real variables.

Example 12.12 (Granger causality and causality.) Do Christmas cards cause Christ-
mas?

149
Example 12.13 (Granger causality and causality II, from Hamilton 11.) Consider the
price P t of an asset paying dividends D t . Suppose the expected return (E t (P tC1 C
D t C1 /=P t ) is a constant, R. The price then satisfies P t DE t 1 s
P
sD1 R D t Cs . Suppose
D t D u t C ıu t 1 C v t , so E t D t C1 D ıu t and E t D t Cs D 0 for s > 1. This gives
P t D ıu t =R, and D t D u t C v t C RP t 1 , so the VAR is
" # " #" # " #
Pt 0 0 Pt 1 ıu t =R
D C ;
Dt R 0 Dt 1 ut C vt

where P Granger-causes D. Of course, the true causality is from D to P . Problem:

forward looking behavior.

Example 12.14 (Money and output, Sims (1972).) Sims found that output, y does not
Granger-cause money, m, but that m Granger causes y. His interpretation was that
money supply is exogenous (set by the Fed) and that money has real effects. Notice how
he used a combination of two Granger causality test to make an economic interpretation.

Example 12.15 (Granger causality and omitted information. ) Consider the VAR
2 3 2 32 3 2 3
y1t a11 a12 0 y1t 1 "1t
4 y2t 5 D 4 0 a22 0 5 4 y2t 1 5 C 4 "2t 5
6 7 6 76 7 6 7

y3t 0 a32 a33 y3t 1 "3t

Notice that y2t and y3t do not depend on y1t 1 , so the latter should not be able to
Granger-cause y3t . However, suppose we forget to use y2t in the regression and then
ask if y1t Granger causes y3t . The answer might very well be yes since y1t 1 contains
information about y2t 1 which does affect y3t . (If you let y1t be money, y2t be the (auto-
correlated) Solow residual, and y3t be output, then this is a short version of the comment
in King (1986) comment on Bernanke (1986) (see below) on why money may appear to
Granger-cause output). Also note that adding a nominal interest rate to Sims (see above)
money-output VAR showed that money cannot be taken to be exogenous.

12.5 Forecasts Forecast Error Variance

The error forecast of the s period ahead forecast is

y t Cs E t y tCs D " t Cs C C1 " t Cs 1 C ::: C Cs 1 " t C1 ; (12.10)

150
so the covariance matrix of the (s periods ahead) forecasting errors is

E .y t Cs E t y t Cs / .y t Cs E t y tCs /0 D ˝ C C1 ˝ C10 C ::: C Cs 1 ˝ Cs0 1 : (12.11)

For a VAR(1), Cs D As , so we have

y t Cs E t y t Cs D " t Cs C A" t Cs 1 C ::: C As " t C1 ; and (12.12)

0
E .y t Cs E t y t Cs / .y t Cs E t y t Cs / D ˝ C A˝A0 C ::: C As 1 ˝.As 1 /0 : (12.13)

Note that lims!1 E t y t Cs D 0, that is, the forecast goes to the unconditional mean
(which is zero here, since there are no constants - you could think of y t as a deviation
from the mean). Consequently, the forecast error becomes the VMA representation (12.8).
Similarly, the forecast error variance goes to the unconditional variance.

Example 12.16 (Unconditional variance of VAR(1).) Letting s ! 1 in (12.13) gives

1
X
Ey t y t0 D As ˝ .As /0
sD0

D ˝ C ŒA˝A0 C A2 ˝.A2 /0 C :::

D ˝ C A ˝ C A˝A0 C ::: A0

D ˝ C A.Ey t y t0 /A0 ;

which suggests that we can calculate Ey t y t0 by an iteration (backwards in time) ˚ t D

˝ C A˚ t C1 A0 , starting from ˚T D I , until convergence.

12.6 Forecast Error Variance Decompositions

If the shocks are uncorrelated, then it is often useful to calculate the fraction of Var(yi;tCs E t yi;t Cs )
due to the j t h shock, the forecast error variance decomposition. Suppose the covariance
matrix of the shocks, here ˝, is a diagonal n  n matrix with the variances !i i along the
diagonal. Let cqi be the it h column of Cq . We then have
n
X
0
Cq ˝ Cq0 D !i i cqi cqi : (12.14)
i D1

151
Example 12.17 (Illustration of (12.14) with n D 2:) Suppose
" # " #
c11 c12 !11 0
Cq D and ˝ D ;
c21 c22 0 !22

then " #
2 2
!11 c11 C !22 c12 !11 c11 c21 C !22 c12 c22
Cq ˝ Cq0 D 2 2
;
!11 c11 c21 C !22 c12 c22 !11 c21 C !22 c22
which should be compared with
" #" #0 " #" #0
c11 c11 c12 c12
!11 C !22
c21 c21 c22 c22
" # " #
2 2
c11 c11 c21 c12 c12 c22
D !11 2
C !22 2
:
c11 c21 c21 c12 c22 c22

Applying this on (12.11) gives

n
X n
X n
X
0 0 0
E .y tCs E t y t Cs / .y t Cs E t y t Cs / D !i i I C !i i c1i .c1i / C ::: C ! i i cs 1i .cs 1i /
i D1 i D1 i D1
n
X
!i i I C c1i .c1i /0 C ::: C cs 0
 
D 1i .cs 1i / ;
i D1
(12.15)

which shows how the covariance matrix for the s-period forecast errors can be decom-
posed into its n components.

12.7 Structural VARs

12.7.1 Structural and Reduced Forms

We are usually not interested in the impulse response function (12.8) or the variance
decomposition (12.11) with respect to " t , but with respect to some structural shocks, u t ,
which have clearer interpretations (technology, monetary policy shock, etc.).
Suppose the structural form of the model is

F y t D ˛ C B1 y t 1 C ::: C Bp y t p C u t ; u t is white noise, Cov(u t ) D D. (12.16)

152
This could, for instance, be an economic model derived from theory.1
Provided F 1 exists, it is possible to write the time series process as a vector autore-
gression

1 1 1 1
yt D F ˛CF B1 y t 1 C ::: C F Bp y t p CF ut (12.17)
D  C A1 y t 1 C ::: C Ap y t p C " t ; Cov ." t / D ˝, (12.18)

where

1 1 1 1 1 0


DF ˛, As D F Bs , and " t D F u t so ˝ D F D F . (12.19)

Equation (12.18) is a VAR(p) model, so a VAR can be thought of as a reduced form of

the structural model (12.16).
The key to understanding the relation between the structural model and the VAR is
the F matrix, which controls how the endogenous variables, y t , are linked to each other
contemporaneously. In fact, identification of a VAR amounts to choosing an F matrix.
Once that is done, impulse responses and forecast error variance decompositions can be
made with respect to the structural shocks. For instance, the impulse response function
of the VAR, (12.8), can be rewritten in terms of the structural shocks u t D F " t (from
(12.19))

y t D " t C C1 " t 1 C C2 " t 2 C :::

1 1 1
DF F " t C C1 F F "t 1 C C2 F F "t 2 C :::
1 1 1
DF u t C C1 F ut 1 C C2 F ut 2 C ::: (12.20)

1
Remark 12.18 The easiest way to calculate this representation is by first finding F
(see below), then writing (12.18) as

1
y t D  C A1 y t 1 C ::: C Ap y t p CF ut : (12.21)

To calculate the impulse responses to the first element in u t , set y t 1 ; :::; y t p equal to
the long-run average, .I A1 ::: Ap/ 1  (or just put them all to zero), make the
first element in u t unity and all other elements zero. Calculate the response by iterating
1
This is a “structural model” in a traditional, Cowles commission, sense. This might be different from
what modern macroeconomists would call structural.

153
forward on (12.21), but putting all elements in u t C1 ; u tC2 ; ::: to zero. This procedure can
be repeated for the other elements of u t .

We would typically pick F such that the elements in u t are uncorrelated with each
other, so they have a clear interpretation.
The VAR form can be estimated directly from data. Is it then possible to recover the
structural parameters in (12.16) from the estimated VAR (12.18)? Not without restrictions
on the structural parameters in F; Bs , ˛, and D. To see why, note that in the structural
form (12.16) we have .p C 1/ n2 parameters in fF; B1 ; : : : ; Bp g, n parameters in ˛, and
n.n C 1/=2 unique parameters in D (it is symmetric). In the VAR (12.18) we have fewer
parameters: pn2 in fA1 ; : : : ; Ap g, n parameters in in , and n.n C 1/=2 unique param-
eters in ˝. This means that we have to impose at least n2 restrictions on the structural
parameters fF; B1 ; : : : ; Bp ; ˛; Dg to identify all of them. This means, of course, that
many different structural models have can have exactly the same reduced form.

Example 12.19 (Structural form of the 2  1 case.) Suppose the structural form of the
previous example is
" #" # " #" # " #" # " #
F11 F12 xt B1;11 B1;12 xt 1 B2;11 B2;12 xt 2 u1;t
D C C :
F21 F22 zt B1;21 B1;22 zt 1 B2;21 B2;22 zt 2 u2;t

This structural form has 3  4 C 3 unique parameters. The VAR in (12.2) has 2  4 C 3.
We need at least 4 restrictions on fF; B1 ; B2 ; Dg to identify them from fA1 ; A2 ; ˝g.

12.7.2 “Triangular” Identification 1: Triangular F with Fi i D 1 and Diagonal D

Reference: Sims (1980).

The perhaps most common way to achieve identification of the structural parameters
is to restrict the contemporaneous response of the different endogenous variables, y t , to
the different structural shocks, u t . Within in this class of restrictions, the triangular iden-
tification is the most popular: assume that F is lower triangular (n.n C 1/=2 restrictions)
with diagonal element equal to unity, and that D is diagonal (n.n 1/=2 restrictions),
which gives n2 restrictions (exact identification).
A lower triangular F matrix is very restrictive. It means that the first variable can
react to lags and the first shock, the second variable to lags and the first two shocks, etc.

154
This is a recursive simultaneous equations model, and we obviously need to be careful
with how we order the variables. The assumptions that Fi i D 1 is just a normalization.
A diagonal D matrix seems to be something that we would often like to have in
a structural form in order to interpret the shocks as, for instance, demand and supply
shocks. The diagonal elements of D are the variances of the structural shocks.

Example 12.20 (Lower triangular F : going from structural form to VAR.) Suppose the
structural form is
" #" # " #" # " #
1 0 xt B11 B12 xt 1 u1;t
D C :
˛ 1 zt B21 B22 zt 1 u2;t

This is a recursive system where x t does not not depend on the contemporaneous z t , and
therefore not on the contemporaneous u2t (see first equation). However, z t does depend
on x t (second equation). The VAR (reduced form) is obtained by premultiplying by F 1
" # " #" #" # " #" #
xt 1 0 B11 B12 xt 1 1 0 u1;t
D C
zt ˛ 1 B21 B22 zt 1 ˛ 1 u2;t
" #" # " #
A11 A12 xt 1 "1;t
D C :
A21 A22 zt 1 "2;t

This means that "1t D u1t , so the first VAR shock equals the first structural shock. In
contrast, "2;t D ˛u1;t C u2;t , so the second VAR shock is a linear combination of the first
two shocks. The covariance matrix of the VAR shocks is therefore
" # " #
"1;t Var .u1t / ˛Var .u1t /
Cov D :
"2;t ˛Var .u1t / ˛ 2 Var .u1t / C Var .u2t /

This set of identifying restrictions can be implemented by estimating the structural

form with LS—equation by equation. The reason is that this is just the old fashioned fully
recursive system of simultaneous equations. See, for instance, Greene (2000) 16.3.

Remark 12.21 (Importance of the ordering of the VAR) Suppose the our objective is to
analyze the effects of policy shocks on the other variables in the VAR system, for instance,
output and prices. The identification rests on the ordering of the VAR, that is, on the struc-
ture of the contemporaneous correlations as captured by F . It is therefore important to
understand how the results on the policy shock (denoted upt ) are changed if the variables

155
are reordered. We have the following result. First, the partitioning of y t into variables
which come before, x1t , and after, x2t , the policy instrument is important for upt and the
impulse response function of all variables with respect to upt . Second, the order within
x1t and x2t does not matter for upt or the impulse response function of any variable with
respect to upt . This suggests that we can settle for partial identification in the sense that
we must take a stand on which variables that come before and after the policy instru-
ment, but the ordering within those blocks are unimportant for understanding the effects
of policy shocks.

12.7.3 “Triangular” Identification 2: Triangular F and D D I

The identifying restrictions in Section 12.7.2 is actually the same as assuming that F is
triangular and that D D I . In this latter case, the restriction on the diagonal elements of F
has been moved to the diagonal elements of D. This is just a change of normalization (that
the structural shocks have unit variance). It happens that this alternative normalization is
fairly convenient when we want to estimate the VAR first and then recover the structural
parameters from the VAR estimates.

Example 12.22 (Change of normalization in Example 12.20) Suppose the structural shocks
in Example 12.20 have the covariance matrix
" # " #
u1;t 12 0
D D Cov D :
u2;t 0 22

Premultiply the structural form in Example 12.20 by

" #
1=1 0
0 1=2

to get
" #" # " #" # " #
1=1 0 xt B11 =1 B12 =1 xt 1 u1;t =1
D C :
˛=2 1=2 zt B21 =2 B22 =2 zt 1 u2;t =2

This structural form has a triangular F matrix (with diagonal elements that can be dif-
ferent from unity), and a covariance matrix equal to an identity matrix.

156
 The reason why this alternative normalization is convenient is that it allows us to use
the widely available Cholesky decomposition.

Remark 12.23 (Cholesky decomposition) Let ˝ be an n  n symmetric positive definite

matrix. The Cholesky decomposition gives the unique lower triangular P such that ˝ D
PP 0 (some software returns an upper triangular matrix, that is, Q in ˝ D Q0 Q instead).

Remark 12.24 Note the following two important features of the Cholesky decomposition.
First, each column of P is only identified up to a sign transformation; they can be reversed
at will. Second, the diagonal elements in P are typically not unity.

Remark 12.25 (Changing sign of column and inverting.) Suppose the square matrix A2
is the same as A1 except that the i t h and j t h columns have the reverse signs. Then A2 1
is the same as A1 1 except that the i t h and j t h rows have the reverse sign.

This set of identifying restrictions can be implemented by estimating the VAR with
LS and then take the following steps.

0
 Step 1. From (12.19) ˝ D F 1 I F 1 (recall D D I is assumed), so a Cholesky
decomposition recovers F 1 (lower triangular F gives a similar structure of F 1 ,
and vice versa, so this works). The signs of each column of F 1 can be chosen
freely, for instance, so that a productivity shock gets a positive, rather than negative,
effect on output. Invert F 1 to get F .

 Step 2. Invert the expressions in (12.19) to calculate the structural parameters from
the VAR parameters as ˛ D F, and Bs D FAs .

Example 12.26 (Identification of the 2  1 case.) Suppose the structural form of the
previous example is
" #" # " #" # " #" # " #
F11 0 xt B1;11 B1;12 xt 1 B2;11 B2;12 xt 2 u1;t
D C C ;
F21 F22 zt B1;21 B1;22 zt 1 B2;21 B2;22 zt 2 u2;t
" #
1 0
with D D :
0 1

157
Step 1 above solves
" # " # 1
0" # 1 10
˝11 ˝12 F11 0 F11 0
D @ A
˝12 ˝22 F21 F22 F21 F22
2 3
1 F21
2 2
F11 F11 F22
D4 F21
2
F21 CF11 2
5
2 2 2
F11 F22 F11 F22

for the three unknowns F11 , F21 , and F22 in terms of the known ˝11 , ˝12 , and ˝22 . Note
that the identifying restrictions are that D D I (three restrictions) and F12 D 0 (one
restriction). (This system is just four nonlinear equations in three unknown - one of the
equations for ˝12 is redundant. You do not need the Cholesky decomposition to solve it,
since it could be solved with any numerical solver of non-linear equations—but why make
life even more miserable?)

A practical consequence of this normalization is that the impulse response of shock i

equal to unity is exactly the same as the impulse response of shock i equal to Std(ui t ) in
the normalization in Section 12.7.2.

12.7.4 What Do Generalized Impulse Response Functions Do?

A traditional impulse response function wit respect to the i th shock is calculated as

y t D ei t C C1 ei t 1 C C 2 ei t 2 C ::: (12.22)

where ei t is a vector of zeros, except for the i th element.

Instead, a generalized impulse response function (see Pesaran and Shin (1998)) is
calculated as

y t D ˝diag.˝/ 1 ei t C C1 ˝diag.˝/ 1 ei t 1 C C2 ˝diag.˝/ 1 ei t 2 C ::: (12.23)

where ˝ is the covariance matrix of the VAR residuals, ˝ D Cov." t /. (Notice that
diag.˝/ 1 should be understood as the inverse of diag.˝/).
The generalized impulse response function can be interpreted as follows: what is the
typical response to the i th shock, when we incorporate the typical comovement of that
shock with the other shocks? To see that, notice that if the residuals are normally dis-
tributed, then the expected value of the vector of residuals, conditional on the ith element

158
being one, is
E." t j"i t D 1/ D ˝i =!i i (12.24)

where ˝i is column i of ˝ (that is, Cov." t ; "i t /). Clearly, when the shocks are uncorre-
lated, then this approach coincides with a traditional impulse response function.
Comparing with the structural form (12.20) where the impulse response function (with
respect to the i th shock) is calculated as

1 1 1
yt D F ei t C C 1 F ei t 1 C C2 F ei t 2 C :::; (12.25)

it is clear that the generalized impulse response function assumes that

1
F D ˝diag.˝/ 1 , so F D diag.˝/˝ 1
: (12.26)

Example 12.27 (Generalized impulse response function when the true structural form is
recursive) Consider Example 12.20 and suppose Var.ui t / D 1 notice that we then have
" #" #" #0 " #
1 0 1 0 1 0 1 ˛
˝D D , so
˛ 1 0 1 ˛ 1 ˛ ˛2 C 1
" #
2
1 ˛=.˛ C 1/
˝diag.˝/ 1 D :
˛ 1

This is used in the generalized impulse response function (12.23). The generalized im-
pulse response with respect to the first shock is then calculated as
" #" # " #
1 ˛=.˛ 2 C 1/ 1 1
Cs D Cs :
˛ 1 0 ˛

The implication is that when the first shock is one, then the second shock is assumed to
be ˛, which is the same as in Example 12.20. Similarly, the generalized impulse response
with respect to the second shock is then calculated as
" #" # " #
1 ˛=.˛ 2 C 1/ 0 ˛=.˛ 2 C 1/
Cs D Cs :
˛ 1 1 1

The implication is that when the second shock is one, then the first shock is assumed to be
˛=.˛ 2 C 1/, which is different from Example 12.20 (where the first shock is then assumed
to be zero).

159
Example 12.28 (Generalized impulse response function and the implied structural form)
Using the results from the previous example in (12.26) gives
" #
 1 ˛2 C 1 ˛
F D diag.˝/˝ D ;
˛.˛ 2 C 1/ ˛ 2 C 1

which can be normalized to get unit coefficients on the diagonal by multiplying by the
inverse of diag.F  /
" #
2
1 ˛=.˛ C 1/
F D F  diag.F  / 1 D :
˛ 1

This shows the same pattern as in the previous example. The first column coincides with
the (true) structural form, while the second column (wrongly) breaks the recursive struc-
ture by assuming that the first variable depends (contemporaneously) on the second vari-
able. The reason is that the generalized impulse response function is based on the cor-
relation pattern—and it is indeed the case that the first variable is correlated with the
second. However, that is not driven by shocks to the second variable (as assumed by the
generalized impulse response function), rather the opposite.

12.7.5 Other Identification Schemes

Reference: Bernanke (1986).

Not all economic models can be written in this recursive form. However, there are
often cross-restrictions between different elements in F or between elements in F and D,
or some other type of restrictions on F which may allow us to identify the system.
Suppose we have (estimated) the parameters of the VAR (12.18), and that we want to
impose D DCov(u t ) D I . From (12.19) we then have (D D I )

1 1 0


˝DF F : (12.27)

As before we need n.n 1/=2 restrictions on F , but this time we don’t want to impose
the restriction that all elements in F above the principal diagonal are zero. Given these
restrictions (whatever they are), we can solve for the remaining elements in B, typically
with a numerical method for solving systems of non-linear equations.

160
12.7.6 What if the VAR Shocks are Uncorrelated (˝ D I )?

Suppose we estimate a VAR and find that the covariance matrix of the estimated residuals
is (almost) an identity matrix (or diagonal). Does this mean that the identification is
superfluous? No, not in general. Yes, if we also want to impose the restrictions that F is
triangular.
There are many ways to reshuffle the shocks and still get orthogonal shocks. Recall
that the structural shocks are linear functions of the VAR shocks, u t D F " t , and that we
assume that Cov." t / D ˝ D I and we want Cov.u t / D I , that, is from (12.19) we then
have (D D I )
FF 0 D I: (12.28)

There are many such F matrices: the class of those matrices even have a name: orthogonal
matrices (all columns in F are orthonormal). However, there is only one lower triangular
F which satisfies (12.28) (the one returned by a Cholesky decomposition, which is I ).
Suppose you know that F is lower triangular (and you intend to use this as the identi-
fying assumption), but that your estimated ˝ is (almost, at least) diagonal. The logic then
requires that F is not only lower triangular, but also diagonal. This means that u t D " t
(up to a scaling factor). Therefore, a finding that the VAR shocks are uncorrelated com-
bined with the identifying restriction that F is triangular implies that the structural and
reduced form shocks are proportional. We can draw no such conclusion if the identifying
assumption is something else than lower triangularity.

Example 12.29 (Rotation of vectors (“Givens rotations”).) Consider the transformation

of the vector " into the vector u; u D G 0 ", where G D In except that Gi k D c, Gi k D s,
Gki D s, and Gkk D c. If we let c D cos  and s D sin  for some angle , then
G 0 G D I . To see this, consider the simple example where i D 2 and k D 3
2 30 2 3 2 3
1 0 0 1 0 0 1 0 0
4 0 c s 5 4 0 c s 5 D 4 0 c2 C s2 0 5;
6 7 6 7 6 7
2 2
0 s c 0 s c 0 0 c Cs

161
which is an identity matrix since cos2  C sin2  D 1. The transformation u D G 0 " gives

u t D " t for t ¤ i; k
ui D " i c "k s
uk D "i s C "k c:

The effect of this transformation is to rotate the i t h and k t h vectors counterclockwise

through an angle of . (Try it in two dimensions.) There is an infinite number of such
transformations (apply a sequence of such transformations with different i and k, change
, etc.).

Example 12.30 (Givens rotations and the F matrix.) We could take F in (12.28) to be
(the transpose) of any such sequence of givens rotations. For instance, if G1 and G2 are
0
givens rotations, then F D G10 or F D G2 or F D G10 G20 are all valid.

12.7.7 Identification via Long-Run Restrictions—No Cointegration

Suppose we have estimated a VAR system (12.1) for the first differences of some variables
y t D
x t , and that we have calculated the impulse response function as in (12.8), which
we rewrite as

x t D " t C C1 " t 1 C C2 " t 2 C :::

D C .L/ " t , with Cov(" t ) D ˝: (12.29)

To find the MA of the level of x t , we solve recursively

x t D C .L/ " t C x t 1

D C .L/ " t C C .L/ " t 1 C xt 2

::
:
D C .L/ ." t C " t 1 C "t 2 C :::/
D " t C .C1 C I / " t 1 C .C2 C C1 C I / " t 2 C :::
s
X
C
D C .L/ " t , where CsC D Cs with C0 D I: (12.30)
j D0

162
 As before the structural shocks, u t , are

u t D F " t with Cov.u t / D D:

The VMA in term of the structural shocks is therefore

s
X
x t D C C .L/ F 1
u t , where CsC D Cs with C0 D I: (12.31)
j D0

The C C .L/ polynomial is known from the estimation, so we need to identify F in order to
use this equation for impulse response function and variance decompositions with respect
to the structural shocks.
As before we assume that D D I , so

1 1 0


˝DF D F (12.32)

in (12.19) gives n.n C 1/=2 restrictions.

We now add restrictions on the long run impulse responses. From (12.31) we have

@x t Cs
lim D lim CsC F 1
s!1 @u0t s!1
1
D C.1/F ; (12.33)

where C.1/ D j1D0 Cs . We impose n.n 1/=2 restrictions on these long run responses.
P

Together we have n2 restrictions, which allows to identify all elements in F .

In general, (12.32) and (12.33) is a set of non-linear equations which have to solved
for the elements in F . However, it is common to assume that (12.33) is a lower triangular
matrix. We can then use the following “trick” to find F . Since " t D F 1 u t

0
EC.1/" t "0t C.1/0 D EC.1/F 1
u t u0t F 1 C.1/0

0
C.1/˝ C.1/0 D C.1/F 1
F 1 C.1/0 : (12.34)

We can therefore solve for a lower triangular matrix

1
 D C.1/F (12.35)

by calculating the Cholesky decomposition of the left hand side of (12.34) (which is

163
 1
available from the VAR estimate). Finally, we solve for F from (12.35).

Example 12.31 (The 2  1 case.) Suppose the structural form is

" #" # " #" # " #
F11 F12
x t B11 B12
x t 1 u1;t
D C :
F21 F22
z t B21 B22
z t 1 u2;t

and we have an estimate of the reduced form

" # " # " # " #!

x t
x t 1 "1;t "1;t
DA C , with Cov D ˝.

z t
z t 1 "2;t "2;t

The VMA form (as in (12.29))

" # " # " # " #

x t "1;t "1;t 1 "1;t 2
D CA C A2 C :::

z t "2;t "2;t 1 "2;t 2

and for the level (as in (12.30))

" # " # " # " #
xt "1;t "1;t 1 "1;t 2
C A2 C A C I


D C .A C I / C :::
zt "2;t "2;t 1 "2;t 2

or since " t D F 1 u t
" # " # " # " #
xt u 1;t u1;t 1 u1;t 2
DF 1 1
C A2 C A C I F 1


C.A C I / F C:::
zt u2;t u2;t 1 u2;t 2

There are 8+3 parameters in the structural form and 4+3 parameters in the VAR, so we
need four restrictions. Assume that Cov.u t / D I (three restrictions) and that the long
run response of u1;t s on x t is zero, that is,
" # " # 1
unrestricted 0 2

F11 F12
D I C A C A C :::
unrestricted unrestricted F21 F22
" # 1
1 F11 F12
D .I A/
F21 F22
" # 1 " # 1
1 A11 A12 F11 F12
D :
A21 1 A22 F21 F22

164
The upper right element of the right hand side is
F12 C F12 A22 C A12 F11
.1 A22 A11 C A11 A22 A12 A21 / .F11 F22 F12 F21 /
1 1 0


which is one restriction on the elements in F . The other three are given by F F D
˝, that is,
2 2
2 3 "
F22 CF12 F22 F21 CF12 F11
#
4 .F11 F22 F12 F21 / 2 2
.F11 F22 F12 F21 / 5 ˝ 11 ˝ 12
F22 F21 CF12 F11
2
F21 2
CF11
D :
2 2
˝ 12 ˝ 22
.F11 F22 F12 F21 / .F11 F22 F12 F21 /

12.8 Cointegration and Identification via Long-Run Restrictions

These notes are a reading guide to Mellander, Vredin, and Warne (1992), which is well be-
yond the first year course in econometrics. See also Englund, Vredin, and Warne (1994).
(I have not yet double checked this section.)

12.8.1 Common Trends Representation and Cointegration

The common trends representation of the n variables in y t is

" # " #!
't 't
y t D y0 C   t C ˚ .L/ ; with Cov D In (12.36)
t t

t D t 1 C 't , (12.37)

where ˚ .L/ is a stable matrix polynomial in the lag operator. We see that the k  1 vector
' t has permanent effects on (at least some elements in) y t , while the r  1 (r D n k)
t does not.

The last component in (12.36) is stationary, but  t is a k  1 vector of random walks,

so the nk matrix  makes y t share the non-stationary components: there are k common
trends. If k < n, then we could find (at least) r linear combinations of y t , ˛ 0 y t where
˛ 0 is an r  n matrix of cointegrating vectors, which are such that the trends cancel each
other (˛ 0  D 0).

Remark 12.32 (Lag operator.) We have the following rules: (i) Lk x t D x t k ; (ii) if
˚ .L/ D a C bL m C cLn , then ˚ .L/ .x t C y t / D a .x t C y t / C b .x t Cm C y t Cm / C
c .x t n C y t n / and ˚ .1/ D a C b C c.

165
Example 12.33 (Söderlind and Vredin (1996)). Suppose we have
2 3 2 3
ln Y t (output) 0 1
6 7 6 7
6 ln P t (price level) 7 6 1 1 7
yt D 66 ln M (money stock)
7,  D 6
7 6 1 0 7,
7
4 t 5 4 5
ln R t (gross interest rate) 0 0
" #
money supply trend
and  t D ;
productivity trend

then we see that ln R t and ln Y t C ln P t ln M t (that is, log velocity) are stationary, so
" #
0 0 0 1
˛0 D
1 1 1 0

are (or rather, span the space of) cointegrating vectors. We also see that ˛ 0  D 022 .

12.8.2 VAR Representation

The VAR representation is as in (12.1). In practice, we often estimate the parameters in

As , ˛, the n  r matrix , and ˝ DCov." t / in the vector “error correction form”

y t D A1
y t C ::: C Ap 1
y t pC1 C ˛ 0 y t 1 C " t , with Cov(" t ) D ˝. (12.38)

This can easily be rewritten on the VAR form (12.1) or on the vector MA representation
for
y t

y t D " t C C1 " t 1 C C2 " t 2 C ::: (12.39)

D C .L/ " t : (12.40)

To find the MA of the level of y t , we recurse on (12.40)

y t D C .L/ " t C y t 1

D C .L/ " t C C .L/ " t 1 C yt 2

::
:
D C .L/ ." t C " t 1 C "t 2 C ::: C "0 / C y0 : (12.41)

166
We now try to write (12.41) in a form which resembles the common trends representation
(12.36)-(12.37) as much as possible.

12.8.3 Multivariate Beveridge-Nelson decomposition

We want to split a vector of non-stationary series into some random walks and the rest
(which is stationary). Rewrite (12.41) by adding and subtracting C.1/." t C " t 1 C :::/

y t D C .1/ ." t C " t 1 C "t 2 C ::: C "0 /CŒC.L/ C .1/ ." t C " t
C ::: C "0 / : 1 C "t 2

(12.42)
Suppose "s D 0 for s < 0 and consider the second term in (12.42). It can be written

I C C1 L C C2 L2 C ::::
 
C .1/ ." t C " t 1 C "t 2 C ::: C "0 /
D /*since C .1/ D I C C1 C C2 C :::*/
Œ C1 C2 C3 ::: " t C Œ C2 C3 ::: " t 1 C Œ C3 ::: " t 2: (12.43)

Now define the random walks

t D t 1 C "t ; (12.44)
D "t C "t 1 C "t 2 C ::: C "0 :

Use (12.43) and (12.44) to rewrite (12.42) as

y t D C .1/  t C C  .L/ " t ; where (12.45)

1
X
Cs D Cj . (12.46)
j DsC1

12.8.4 Identification of the Common Trends Shocks

Rewrite (12.36)-(12.37) and (12.44)-(12.45) as

t
X
y t D C .1/ " t C C  .L/ " t , with Cov(" t / D ˝, and (12.47)
sD0
" P # " # " #!
t
sD0 ' t 't 't
h i
D  0nr C ˚ .L/ ; with Cov D In :
t t t

(12.48)

167
 h i0
0 0
Since both " t and ' t t are white noise, we notice that the response of y t Cs to
either must be the same, that is,
" #
h i  '
t
C .1/ C Cs " t D


 0nr C ˚s for all t and s  0: (12.49)
t

This means that the VAR shocks are linear combinations of the structural shocks (as
in the standard setup without cointegration)
" #
't
D F "t
t
" #
Fk
D "t : (12.50)
Fr

Combining (12.49) and (12.50) gives that

" #
Fk
C .1/ C Cs D  Fk C ˚s (12.51)
Fr

must hold for all s  0. In particular, it must hold for s ! 1 where both Cs and ˚s
vanishes
C .1/ D  Fk : (12.52)

The identification therefore amounts to finding the n2 coefficients in F , exactly as

in the usual case without cointegration. Once that is done, we can calculate the impulse
responses and
h variance i0 decompositions with respect to the structural shocks by using
1 0 0
"t D F 't t in (12.47).2 As before, assumptions about the covariance matrix
of the structural shocks are not enough to achieve identification. In this case, we typically
rely on the information about long-run behavior (as opposed to short-run correlations) to
supply the remaining restrictions.

 Step 1. From (12.36) we see that ˛ 0  D 0rk must hold for ˛ 0 y t to be stationary.
Given an (estimate of) ˛, this gives rk equations from which we can identify rk
elements in  . (It will soon be clear why it is useful to know  ).
2
Equivalently, we can use (12.52) and (12.51) to calculate  and ˚s (for all s) and then calculate the
impulse response function from (12.48).

168
 Step 2. From (12.49) we have  ' t D C .1/ " t as s ! 1. The variances of both
sides must be equal

E ' t ' t0  0 D EC .1/ " t "0t C .1/0 , or

  0 D C .1/ ˝ C .1/0 ; (12.53)

which gives k .k C 1/ =2 restrictions on  (the number of unique elements in the

symmetric   0 ). (However, each column of  is only identified up to a sign trans-
formation: neither step 1 or 2 is affected by multiplying each element in column j
of  by -1.)

 Step 3.  has nk elements, so we still need nk rk k .k C 1/ =2 D k.k 1/=2

further restrictions on  to identify all elements. They could be, for instance, that
money supply shocks have no long run effect on output (some ij D 0). We now
know  .
" #!
't
 Step 4. Combining Cov D In with (12.50) gives
t

" # " # " #0

Ik 0 Fk Fk
D ˝ ; (12.54)
0 Ir Fr Fr

which gives n .n C 1/ =2 restrictions.

– Step 4a. Premultiply (12.52) with  0 and solve for Fk

1
Fk D  0   0 C.1/:


(12.55)

(This means that E' t ' t0 D Fk ˝Fk0 D . 0  / 1  0 C.1/˝ C .1/0  . 0  / 1 .

From (12.53) we see that this indeed is Ik as required by (12.54).) We still
need to identify Fr .
– Step 4b. From (12.54), E' t t0 D 0kr , we get Fk ˝Fr0 D 0kr , which gives
kr restrictions on the rn elements in Fr . Similarly, from E t t0 D Ir , we get
Fr ˝Fr0 D Ir , which gives r .r C 1/ =2 additional restrictions on Fr . We still
need r .r 1/ =2 restrictions. Exactly how they look does not matter for the
0
impulse response function of ' t (as long as E' t t D 0). Note that restrictions

169
 on Fr are restrictions on @y t =@ t0 , that is, on the contemporaneous response.
This is exactly as in the standard case without cointegration.

A summary of identifying assumptions used by different authors is found in Englund,

Vredin, and Warne (1994).

Bibliography
Bernanke, B., 1986, “Alternative explanations of the money-income correlation,”
Carnegie-Rochester Series on Public Policy, 25, 49–100.

Englund, P., A. Vredin, and A. Warne, 1994, “Macroeconomic shocks in an open econ-
omy - a common trends representation of Swedish data 1871-1990,” in Villy Bergström,
and Anders Vredin (ed.), Measuring and interpreting business cycles . pp. 125–233,
Claredon Press.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

King, R. G., 1986, “Money and business cycles: comments on Bernanke and related
literature,” Carnegie-Rochester Series on Public Policy, 25, 101–116.

Mellander, E., A. Vredin, and A. Warne, 1992, “Stochastic trends and economic fluctua-
tions in a small open economy,” Journal of Applied Econometrics, 7, 369–394.

Pesaran, H. H., and Y. Shin, 1998, “Generalized impulse response analysis in linear mul-
tivariate models,” Economics Letters, 58, 17–29.

Pindyck, R. S., and D. L. Rubinfeld, 1998, Econometric models and economic forecasts,
Irwin McGraw-Hill, Boston, Massachusetts, 4ed edn.

Sims, C. A., 1980, “Macroeconomics and reality,” Econometrica, 48, 1–48.

170
Söderlind, P., and A. Vredin, 1996, “Applied cointegration analysis in the mirror of
macroeconomic theory,” Journal of Applied Econometrics, 11, 363–382.

171
12 Kalman filter

12.1 Conditional Expectations in a Multivariate Normal Distribution

Reference: Harvey (1989), Lütkepohl (1993), and Hamilton (1994)

Suppose Zm1 and Xn1 are jointly normally distributed
" # " # " #!
Z ZN ˙zz ˙zx
DN ; : (12.1)
X XN ˙xz ˙xx

The distribution of the random variable Z conditional on that X D x is also normal with
mean (expectation of the random variable Z conditional on that the random variable X
has the value x)
ZN C ˙zx ˙xx1 x XN ;


E .Zjx/ D „ƒ‚… (12.2)
„ ƒ‚ … „ƒ‚…„ƒ‚…„ ƒ‚ …
m1 m1 mn nn n1

and variance (variance of Z conditional on that X D x)

n ˇ o
2ˇ
Var .Zjx/ D E ŒZ E .Zjx/ ˇ x

D ˙zz ˙zx ˙xx1 ˙xz : (12.3)

The conditional variance is the variance of the prediction error Z E.Zjx/.

Both E.Zjx/ and Var.Zjx/ are in general stochastic variables, but for the multivariate
normal distribution Var.Zjx/ is constant. Note that Var.Zjx/ is less than ˙zz (in a matrix
sense) if x contains any relevant information (so ˙zx is not zero, that is, E.zjx/ is not a
constant).
It can also be useful to know that Var.Z/ DEŒVar .ZjX/ + VarŒE .ZjX / (the X is
now random), which here becomes ˙zz ˙zx ˙xx1 ˙xz + ˙zx ˙xx1 Var.X / ˙xx1 ˙xZ =
˙zz .

172
12.2 Kalman Recursions

12.2.1 State space form

The measurement equation is

y t D Z˛ t C  t ; with Var . t / D H , (12.4)

where y t and  t are n1 vectors, and Z an nm matrix. (12.4) expresses some observable
variables y t in terms of some (partly) unobservable state variables ˛ t . The transition
equation for the states is

˛t D T ˛t 1 C u t ; with Var .u t / D Q; (12.5)

where ˛ t and u t are m  1 vectors, and T an m  m matrix. This system is time invariant
since all coefficients are constant. It is assumed that all errors are normally distributed,
and that E. t u t s / D 0 for all s.

Example 12.1 (AR(2).) The process x t D 1 x t 1 C 2 x t 2 C e t can be rewritten as

" #
h i x
t
xt D 1 0 C„ƒ‚… 0 ;
„ƒ‚… xt 1 t
yt
„ ƒ‚ …
Z „ ƒ‚ …
˛t

" # " #" # " #

xt 1 2 xt 1 et
D C ;
xt 1 1 0 xt 2 0
„ ƒ‚ … „ ƒ‚ …„ ƒ‚ … „ ƒ‚ …
˛t T ˛t 1 ut
" #
Var .e t / 0
with H D 0, and Q D : In this case n D 1, m D 2.
0 0

12.2.2 Prediction equations: E.˛ t jI t 1/

Suppose we have an estimate of the state in t 1 based on the information set in t 1,

denoted by ˛O t 1 , and that this estimate has the variance
 0
Pt 1 D E .˛O t 1 ˛t Ot 1
1 / .˛ ˛t 1/ : (12.6)

173
 Now we want an estimate of ˛ t based ˛O t 1 . From (12.5) the obvious estimate, denoted
by ˛ tjt 1 , is
˛O t jt 1 D T ˛O t 1 . (12.7)

The variance of the prediction error is

h

0 i
P t jt 1 D E ˛ t ˛O tjt 1 ˛ t ˛O tjt 1

D E ŒT ˛ t 1 C u t T ˛O t 1  ŒT ˛ t 1 C u t T ˛O t 1 0
˚

D E ŒT .˛O t 1 ˛ t 1 / u t  ŒT .˛O t 1 ˛ t 1 / u t 0
˚

D T E .˛O t 1 ˛ t 1 / .˛O t 1 ˛ t 1 /0 T 0 C Eu t u0t

 

0
D TP t 1T C Q; (12.8)

where we have used (12.5), (12.6), and the fact that u t is uncorrelated with ˛O t 1 ˛t 1.

Example 12.2 (AR(2) continued.) By substitution we get

" # " #" #
xO t jt 1 1 2 xO t 1jt 1
˛O t jt 1 D D ; and
xO t 1jt 1 1 0 xO t 2jt 1

" # " # " #

1 2 1 1 Var . t / 0
P t jt 1 D Pt 1 C
1 0 2 0 0 0
" #
Var . t / 0
If we treat x 1 and x0 as given, then P0 D 022 which would give P1j0 D .
0 0

12.2.3 Updating equations: E.˛ t jI t 1/ !E.˛ t jI t /

The best estimate of y t , given aO t jt 1, follows directly from (12.4)

yO t jt 1 D Z ˛O t jt 1, (12.9)

with prediction error

vt D yt yO t jt 1 D Z ˛t ˛O t jt 1 C t : (12.10)

174
The variance of the prediction error is

F t D E v t v t0


n

0 o
D E Z ˛ t ˛O t jt 1 C  t Z ˛ t ˛O t jt 1 C  t
h
0 i 0
D ZE ˛ t ˛O t jt 1 ˛ t ˛O t jt 1 Z C E t  t0

0
D ZP t jt 1Z C H; (12.11)

where we have used the definition of P t jt 1 in (12.8), and of H in 12.4. Similarly, the
covariance of the prediction errors for y t and for ˛ t is




Cov ˛ t ˛O t jt 1; yt yO t jt 1 D E ˛ t ˛O tjt 1 y t yO t jt 1
n

0 o
D E ˛ t ˛O t jt 1 Z ˛ t ˛O tjt 1 C  t
h
0 i
D E ˛ t ˛O t jt 1 ˛ t ˛O t jt 1 Z 0

0
D P t jt 1Z : (12.12)

Suppose that y t is observed and that we want to update our estimate of ˛ t from ˛O tjt 1

to ˛O t , where we want to incorporate the new information conveyed by y t .

Example 12.3 (AR(2) continued.) We get

" #
h i xO t jt 1
yO t jt 1 D Z ˛O tjt 1 D 1 0 D xO t jt 1 D 1 xO t 1jt 1 C 2 xO t 2jt 1 ; and
xO t 1jt 1
(" # " # " #)
h i 1 2 Var . t / 0
1 1 h i0
Ft D 1 0 Pt 1 C 1 0 :
1 0 0 2 0
0
" #
Var . t / 0
If P0 D 022 as before, then F1 D P1 D :
0 0
By applying the rules (12.2) and (12.3) we note that the expectation of ˛ t (like z in
(12.2)) conditional on y t (like x in (12.2)) is (note that y t is observed so we can use it to
guess ˛ t )
0 1 1 0 1

˛O t D ˛O tjt 1 C P t jt 1 Z 0 @ZP t jt 1 Z 0 C H A @y t Z ˛O t jt 1 A (12.13)

B C
„ƒ‚… „ƒ‚… „ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
E.zjx/ Ez ˙zx ˙xx DF t Ex

175
with variance

0 0 0

1
P t D P t jt 1 P tjt 1 Z ZP t jt 1 Z C H ZP t jt 1 , (12.14)
„ƒ‚… „ƒ‚… „ ƒ‚ …„ ƒ‚ …„ ƒ‚ …
Var.zjx/ ˙zz ˙zx ˙xx1 ˙xz

where ˛O t jt 1 (“Ez”) is from (12.7), P t jt 1 Z 0 (“˙zx ”) from (12.12), ZP tjt 1 Z 0 C H

(“˙xx ”) from (12.11), and Z ˛O t jt 1 (“Ex”) from (12.9).
(12.13) uses the new information in y t , that is, the observed prediction error, in order
to update the estimate of ˛ t from ˛O t jt 1 to ˛O t .
Proof. The last term in (12.14) follows from the expected value of the square of the
last term in (12.13)

0 0

1

0 0

1
P tjt 1Z ZP t jt 1Z CH E yt Z˛ t jt 1 yt Z˛ t jt 1 ZP t jt 1Z CH ZP t jt 1;

(12.15)
where we have exploited the symmetry of covariance matrices. Note that y t Z˛ t jt 1 D
y t yO t jt 1 , so the middle term in the previous expression is


0 0
E yt Z˛ t jt 1 yt Z˛ t jt 1 D ZP t jt 1Z C H: (12.16)

Using this gives the last term in (12.14).

12.2.4 The Kalman Algorithm

The Kalman algorithm calculates optimal predictions of ˛ t in a recursive way. You can
also calculate the prediction errors v t in (12.10) as a by-prodct, which turns out to be
useful in estimation.

1. Pick starting values for P0 and ˛0 : Let t D 1.

2. Calculate (12.7), (12.8), (12.13), and (12.14) in that order. This gives values for ˛O t
and P t . If you want v t for estimation purposes, calculate also (12.10) and (12.11).
Increase t with one step.

3. Iterate on 2 until t D T .

One choice of starting values that work in stationary models is to set P0 to the uncon-
ditional covariance matrix of ˛ t , and ˛0 to the unconditional mean. This is the matrix P

176
to which (12.8) converges: P D TP T 0 C Q. (The easiest way to calculate this is simply
to start with P D I and iterate until convergence.)
In non-stationary model we could set

P0 D 1000  Im , and ˛0 D 0m1 , (12.17)

in which case the first m observations of ˛O t and v t should be disregarded.

12.2.5 MLE based on the Kalman filter

For any (conditionally) Gaussian time series model for the observable y t the log likeli-
hood for an observation is
n 1 1 0 1
ln L t D ln .2/ ln jF t j v F vt : (12.18)
2 2 2 t t
In case the starting conditions are as in (12.17), the overall log likelihood function is
( P
T
ln L t in stationary models
ln L D PTt D1 (12.19)
t DmC1 ln L t in non-stationary models.

12.2.6 Inference and Diagnostics

We can, of course, use all the asymptotic MLE theory, like likelihood ratio tests etc. For
diagnostoic tests, we will most often want to study the normalized residuals
p
vQ i t D vi t = element i i in F t , i D 1; :::; n;

since element i i in F t is the standard deviation of the scalar residual vi t . Typical tests are
CUSUMQ tests for structural breaks, various tests for serial correlation, heteroskedastic-
ity, and normality.

Bibliography
Hamilton, J. D., 1994, Time series analysis, Princeton University Press, Princeton.

Harvey, A. C., 1989, Forecasting, structural time series models and the Kalman filter,
Cambridge University Press.

177
Lütkepohl, H., 1993, Introduction to multiple time series, Springer-Verlag, 2nd edn.

178
13 Outliers and Robust Estimators

13.1 Influential Observations and Standardized Residuals

Reference: Greene (2000) 6.9; Rousseeuw and Leroy (1987)

Consider the linear model
y t D x t0 ˇ0 C u t ; (13.1)

where x t is k  1. The LS estimator

1
T T
!
X X
ˇO D x t x t0 xt yt ; (13.2)
t D1 t D1

which is the solution to

T
X
2
min yt x t0 ˇ : (13.3)
ˇ
t D1

The fitted values and residuals are

O and uO t D y t
yO t D x t0 ˇ; yO t : (13.4)

Suppose we were to reestimate ˇ on the whole sample, except observation s. This

would give us an estimate ˇO .s/ . The fitted values and residual are then

yO t.s/ D x t0 ˇO .s/ , and uO .s/

t D yt yO t.s/ : (13.5)

A common way to study the sensitivity of the results with respect to excluding observa-
tions is to plot ˇO .s/ ˇ, O and yOs.s/ yOs . Note that we here plot the fitted value of ys using
the coefficients estimated by excluding observation s from the sample. Extreme values
prompt a closer look at data (errors in data?) and perhaps also a more robust estimation
method than LS, which is very sensitive to outliers.
Another useful way to spot outliers is to study the standardized residuals, uO s =O and
.s/
uO s =O .s/ , where O and O .s/ are standard deviations estimated from the whole sample and
excluding observation s, respectively. Values below -2 or above 2 warrant attention (recall

179
that Pr.x > 1:96/  0:025 in a N.0; 1/ distribution).
Sometimes the residuals are instead standardized by taking into account the uncer-
tainty of the estimated coefficients. Note that

uO .s/
t D yt x t0 ˇO .s/
 
D u t C x t0 ˇ ˇO .s/ ; (13.6)

since y t D x t0 ˇ C u t . The variance of uO t is therefore the variance of the sum on the

right hand side of this expression. When we use the variance of u t as we did above to
standardize the residuals, then we disregard the variance of ˇO .s/ . In general, we have
    h  i
Var uO .s/
t D Var.u t / C x 0
t Var ˇ O
ˇ .s/
x t C 2Cov u t ; x 0
t ˇ O
ˇ .s/
: (13.7)

When t D s, which is the case we care about, the covariance term drops out since ˇO .s/
cannot be correlated with us since period s is not used in the estimation (this statement
assumes that shocks are not autocorrelated). The first term is then estimated as the usual
variance of the residuals (recall that period s is not used) and the second term is the
estimated covariance matrix of the parameter vector (once again excluding period s) pre-
and postmultiplied by xs .

Example 13.1 (Errors are iid independent of the regressors.) In this case the variance of
the parameter vector is estimated as O 2 .˙x t x t0 / 1 (excluding period s), so we have
 
Var uO .s/ D O 2 1 C xs0 .˙x t x t0 / 1 xs :


t

13.2 Recursive Residuals

Reference: Greene (2000) 7.8

Recursive residuals are a version of the technique discussed in Section 13.1. They
are used when data is a time series. Suppose we have a sample t D 1; :::; T ,.and that
t D 1; :::; s are used to estimate a first estimate, ˇO Œs (not to be confused with ˇO .s/ used in
Section 13.1). We then make a one-period ahead forecast and record the fitted value and
the forecast error
Œs
yOsC1 0
D xsC1 ˇO Œs , and uO Œs
sC1 D ysC1
Œs
yOsC1 : (13.8)

180
 Rescursive residuals from AR(1) with corr=0.85
CUSUM statistics and 95% confidence band
50
2

0 0

−2
−50
0 100 200 0 100 200
period period
Figure 13.1: This figure shows recursive residuals and CUSUM statistics, when data are
simulated from y t D 0:85y t 1 C u t , with Var.u t / D 1.

This is repeated for the rest of the sample by extending the sample used in the estimation
by one period, making a one-period ahead forecast, and then repeating until we reach the
end of the sample.
A first diagnosis can be made by examining the standardized residuals, uO Œs sC1 = O Œs ,
where O Œs can be estimated as in (13.7) with a zero covariance term, since usC1 is not
correlated with data for earlier periods (used in calculating ˇO Œs ), provided errors are not
autocorrelated. As before, standardized residuals outside ˙2 indicates problems: outliers
or structural breaks (if the residuals are persistently outside ˙2).
The CUSUM test uses these standardized residuals to form a sequence of test statistics.
A (persistent) jump in the statistics is a good indicator of a structural break. Suppose we
use r observations to form the first estimate of ˇ, so we calculate ˇO Œs and uO Œs
sC1 =O Œs for
s D r; :::; T . Define the cumulative sums of standardized residuals
t
uO Œs
X
Wt D O Œs , t D r; :::; T:
sC1 = (13.9)
sDr

Under the null hypothesis that no structural breaks occurs, that is, that the true ˇ is the
same for the whole sample, W t has a zero mean and a variance equal to the number of
elements in the sum, t r C 1. This follows from the fact that the standardized resid-
uals all have zero mean and unit variance and are uncorrelated with each other. Typ-
 pW t is plotted along with
ically,
p
a 95%
 confidence interval, which can be shown to be
˙ a T r C 2a .t r/ = T r with a D 0:948. The hypothesis of no structural
break is rejected if the W t is outside this band for at least one observation. (The derivation
of this confidence band is somewhat tricky, but it incorporates the fact that W t and W t C1

181
are very correlated.)

13.3 Robust Estimation

Reference: Greene (2000) 9.8.1; Rousseeuw and Leroy (1987); Donald and Maddala
(1993); and Judge, Griffiths, Lütkepohl, and Lee (1985) 20.4.

13.3.1 Robust Means, Variances and Correlations

Outliers and other extreme observations can have very decisive influence on the estimates
of the key statistics needed for financial analysis, including mean returns, variances, co-
variances and also regression coefficients.
The perhaps best way to solve these problems is to carefully analyse the data—and
then decide which data points to exclude. Alternatively, robust estimators can be applied
instead of the traditional ones.
To estimate the mean, the sample average can be replaced by the median or a trimmed
mean (where the x% lowest and highest observations are excluded).
Similarly, to estimate the variance, the sample standard deviation can be replaced by
the interquartile range (the difference between the 75th and the 25th percentiles), divided
by 1:35
StdRobust D Œquantile.0:75/ quantile.0:25/=1:35; (13.10)

or by the median absolute deviation

StdRobust D median.jx t j/=0:675: (13.11)

Both these would coincide with the standard deviation if data was indeed drawn from a
normal distribution without outliers.
A robust covariance can be calculated by using the identity

Cov.x; y/ D ŒVar.x C y/ Var.x y/=4 (13.12)

and using a robust estimator of the variances—like the square of (13.10). A robust cor-
relation is then created by dividing the robust covariance with the two robust standard
deviations.
See Figures 13.2–13.4 for empirical examples.

182
 US industry portfolios, ERe
Monthly data 1947:1−2008:6
mean
0.11 median

0.1

0.09
β

0.08

0.07

0.06
A B C D E F G H I J

Figure 13.2: Mean excess returns of US industry portfolios

The idea of robust estimation is to give less weight to extreme observations than in
Least Squares. When the errors are normally distributed, then there should be very few ex-
treme observations, so LS makes a lot of sense (and is indeed the MLE). When the errors
have distributions with fatter tails (like the Laplace or two-tailed exponential distribution,
f .u/ D exp. juj =/=2), then LS is no longer optimal and can be fairly sensitive to
outliers. The ideal way to proceed would be to apply MLE, but the true distribution is
often unknown. Instead, one of the “robust estimators” discussed below is often used.
O Then, the least absolute deviations (LAD), least median squares
Let uO t D y t x t0 ˇ.
(LMS), and least trimmed squares (LTS) estimators solve
T
X
ˇOLAD D arg min juO t j (13.13)
ˇ
t D1

ˇOLM S D arg min median uO 2t



(13.14)
ˇ
h
X
ˇOLT S D arg min uO 2i , uO 21  uO 22  ::: and h  T: (13.15)
ˇ
i D1

Note that the LTS estimator in (13.15) minimizes of the sum of the h smallest squared
residuals.

183
 US industry portfolios, Std
Monthly data 1947:1−2008:6
std
0.2 iqr/1.35

0.18
β

0.16

0.14

0.12

A B C D E F G H I J

Figure 13.3: Volatility of US industry portfolios

These estimators involve non-linearities, so they are more computationally intensive

than LS. In some cases, however, a simple iteration may work.

Example 13.2 (Algorithm for LAD.) The LAD estimator can be written
T
X
ˇOLAD D arg min w t uO 2t , w t D 1= juO t j ;
ˇ
tD1

so it is a weighted least squares where both y t and x t are multiplied by 1= juO t j. It can be
shown that iterating on LS with the weights given by 1= juO t j, where the residuals are from
the previous iteration, converges very quickly to the LAD estimator.

It can be noted that LAD is actually the MLE for the Laplace distribution discussed
above.

13.4 Multicollinearity

Reference: Greene (2000) 6.7

When the variables in the x t vector are very highly correlated (they are “multicollinear”)
then data cannot tell, with the desired precision, if the movements in y t was due to move-

184
 Correlations, traditional Difference in correlations: traditional − robust

1
0.1
0.5 0
−0.1
0
10 10
10 10
5 5 5 5
Industry Industry

10 US industry portfolios, 1947:1−2008:6

Figure 13.4: Correlation of US industry portfolios

ments in xi t or xjt . This means that the point estimates might fluctuate wildly over sub-
samples and it is often the case that individual coefficients are insignificant even though
the R2 is high and the joint significance of the coefficients is also high. The estimators
are still consistent and asymptotically normally distributed, just very imprecise.
A common indicator for multicollinearity is to standardize each element in x t by
subtracting the sample mean and then dividing by its standard deviation

xQ i t D .xi t xN i t / =std .xi t / : (13.16)

(Another common procedure is to use xQ i t D xi t =.˙ tTD1 xi2t =T /1=2 .)

Then calculate the eigenvalues, j , of the second moment matrix of xQ t
T
1X
AD xQ t xQ t0 : (13.17)
T t D1

The condition number of a matrix is the ratio of the largest (in magnitude) of the
eigenvalues to the smallest
c D jjmax = jjmin : (13.18)

(Some authors take c 1=2 to be the condition number; others still define it in terms of the
“singular values” of a matrix.) If the regressors are uncorrelated, then the condition value
of A is one. This follows from the fact that A is a (sample) covariance matrix. If it is

185
 OLS vs LAD
2
1.5
1
0.5
0
−0.5 Data
−1 0.75*x
OLS
−1.5
LAD
−2
−3 −2 −1 0 1 2 3
x
Figure 13.5: This figure shows an example of how LS and LAD can differ. In this case
y t D 0:75x t C u t , but only one of the errors has a non-zero value.

diagonal, then the eigenvalues are equal to diagonal elements, which are all unity since
the standardization in (13.16) makes all variables have unit variances. Values of c above
several hundreds typically indicate serious problems.

Bibliography
Donald, S. G., and G. S. Maddala, 1993, “Identifying outliers and influential observations
in econometric models,” in G. S. Maddala, C. R. Rao, and H. D. Vinod (ed.), Handbook
of Statistics, Vol 11 . pp. 663–701, Elsevier Science Publishers B.V.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Judge, G. G., W. E. Griffiths, H. Lütkepohl, and T.-C. Lee, 1985, The theory and practice
of econometrics, John Wiley and Sons, New York, 2nd edn.

Rousseeuw, P. J., and A. M. Leroy, 1987, Robust regression and outlier detection, John
Wiley and Sons, New York.

186
14 Generalized Least Squares
Reference: Greene (2000) 11.3-4
Additional references: Hayashi (2000) 1.6; Johnston and DiNardo (1997) 5.4; Verbeek
(2004) 4

14.1 Introduction

Instead of using LS in the presence of autocorrelation/heteroskedasticity (and, of course,

adjusting the variance-covariance matrix), we may apply the generalized least squares
method. It can often improve efficiency.
The linear model y t D x t0 ˇ0 C u t written on matrix form (GLS is one of the cases in
econometrics where matrix notation really pays off) is

y D Xˇ0 C u; where (14.1)

2 3 2 3 2 3
y1 x10 u1
6 y2 7 x20 u2
6 7 6 7 6 7
6 7 6 7
y D 6 : 7, X D 6
6 7 :: 7 , and u D 6 :: 7.
4 :: 5 : :
6 7 6 7
4 5 4 5
yT xT0 uT

Suppose that the covariance matrix of the residuals (across time) is

2 3
Eu1 u1 Eu1 u2


Eu1 uT
u u u
6 7
Eu 2 1 Eu 2 2 Eu 2 T
Euu0 D 6
6 7
:: ::: :: 7
: :
6 7
4 5
EuT u1 EuT u2 EuT uT
D ˝T T : (14.2)

This allows for both heteroskedasticity (different elements along the main diagonal) and
autocorrelation (non-zero off-diagonal elements). LS is still consistent even if ˝ is not
proportional to an identity matrix, but it is not efficient. Generalized least squares (GLS)

187
is. The trick of GLS is to transform the variables and the do LS.

14.2 GLS as Maximum Likelihood

Remark 14.1 If the n  1 vector x has a multivariate normal distribution with mean vec-
tor  and covariance matrix ˝, then the joint probability density function is .2/ n=2 j˝j 1=2 expŒ .x
/0 ˝ 1 .x /=2.

If the T 1 vector u is N .0; ˝/, then the joint pdf of u is .2/ n=2 j˝j 1=2 expŒ u0 ˝ 1 u=2.
Change variable from u to y Xˇ (the Jacobian of this transformation equals one), and
take logs to get the (scalar) log likelihood function
n 1 1
ln L D ln .2/ ln j˝j .y Xˇ/0 ˝ 1
.y Xˇ/ : (14.3)
2 2 2
To simplify things, suppose we know ˝. It is then clear that we maximize the likelihood
function by minimizing the last term, which is a weighted sum of squared errors.
In the classical LS case, ˝ D  2 I , so the last term in (14.3) is proportional to the
unweighted sum of squared errors. The LS is therefore the MLE when the errors are iid
normally distributed.
When errors are heteroskedastic, but not autocorrelated, then ˝ has the form
2 3
2
 0


0
6 1 :: 7
6 0 2 : 7
2
˝D6 : 7: (14.4)
6 7
6 :: : :: 0 7
4 5
0


0 T2

1
In this case, we can decompose ˝ as
2 3
1=1 0


0
6 :: 7
1
6 0 1=2 : 7
˝ D P 0 P , where P D 6 7: (14.5)
6 7
:: :::
6
4 : 0 7
5
0


0 1=T

188
The last term in (14.3) can then be written
1 1
.y Xˇ/0 ˝ 1
.y Xˇ/ D .y Xˇ/0 P 0 P .y Xˇ/
2 2
1
D .P y PXˇ/0 .P y PXˇ/ : (14.6)
2
This very practical result says that if we define y t D y t = t and x t D x t = t , then we
get ML estimates of ˇ running an LS regression of y t on x t . (One of the elements in x t
could be a constant—also this one should be transformed). This is the generalized least
squares (GLS).

Remark 14.2 Let A be an n  n symmetric positive definite matrix. It can be decomposed

as A D PP 0 . There are many such P matrices, but only one which is lower triangular P
(see next remark).

Remark 14.3 Let A be an n  n symmetric positive definite matrix. The Cholesky decom-
position gives the unique lower triangular P1 such that A D P1 P10 or an upper triangular
matrix P2 such that A D P20 P2 (clearly P2 D P10 ). Note that P1 and P2 must be invertible
(since A is).

When errors are autocorrelated (with or without heteroskedasticity), then it is typ-

ically harder to find a straightforward analytical decomposition of ˝ 1 . We therefore
move directly to the general case. Since the covariance matrix is symmetric and positive
definite, ˝ 1 is too. We therefore decompose it as

1
˝ D P 0 P: (14.7)

The Cholesky decomposition is often a convenient tool, but other decompositions can
also be used. We can then apply (14.6) also in this case—the only difference is that P
is typically more complicated than in the case without autocorrelation. In particular, the
transformed variables P y and PX cannot be done line by line (y t is a function of y t ,
y t 1 , and perhaps more).

Example 14.4 (AR(1) errors, see Davidson and MacKinnon (1993) 10.6.) Let u t D
au t 1 C " t where " t is iid. We have Var.u t / D  2 = 1 a2 , and Corr.u t ; u t s / D as .

189
For T D 4, the covariance matrix of the errors is
h i0 
˝ D Cov u1 u2 u3 u4
2 3
1 a a2 a3
2 6 a 1 a a2 7
6 7
D 6 7:
1 a2 6 4 a 2
a 1 a 7
5
3 2
a a a 1

The inverse is 2 3
1 a 0 0
2
6 7
1 1 6 a 1Ca a 0 7
˝ D 2 6 7;
 6
4 0 a 1 C a2 a 7
5
0 0 a 1
and note that we can decompose it as
2 p 30 2 p 3
1 a2 0 0 0 1 a2 0 0 0
6 7 6 7
16 a 1 0 0 7 16 a 1 0 0
˝ 1D 6
7
7 6 7:
46 0 a 1 0 76
5 4 0 a 1 0 7
5
0 0 a 1 0 0 a 1
„ ƒ‚ …„ ƒ‚ …
P0 P

This is not a Cholesky decomposition, but certainly a valid decomposition (in case of
doubt, do the multiplication). Premultiply the system
2 3 2 3 2 3
y1 x10 u1
6 y2 7 6 x20 7
6 7 6 7 6 7
7 ˇ0 C 6 u 2 7
6 7
6 7D6
6 y 7 6 x0 7 6 u 7
4 3 5 4 3 5 4 3 5
0
y4 x4 u4

by P to get
2 p 3 2 p 3 2 p 3
.1 a2 /y1 .1 a2 /x10 .1 a2 /u1
x20 ax10
6 7 6 7 6 7
166 y2 ay1 7D 16 7 ˇ0 C 1 6 "2
7 6 7 6 7
7:
6 4 y3 ay2 7 6
5 4 x30 ax20 7
5 46 "3 7
5
y4 ay3 x40 ax30 "4

190
Note that all the residuals are uncorrelated in this formulation. Apart from the first ob-
servation, they are also identically distributed. The importance of the first observation
becomes smaller as the sample size increases—in the limit, GLS is efficient.

14.3 GLS as a Transformed LS

When the errors are not normally distributed, then the MLE approach in the previous
section is not valid. But we can still note that GLS has the same properties as LS has with
iid non-normally distributed errors. In particular, the Gauss-Markov theorem applies,
so the GLS is most efficient within the class of linear (in y t ) and unbiased estimators
(assuming, of course, that GLS and LS really are unbiased, which typically requires that
u t is uncorrelated with x t s for all s). This follows from that the transformed system

P y D PXˇ0 C P u
y  D X  ˇ0 C u  ; (14.8)

have iid errors, u . So see this, note that

Eu u0 D EP uu0 P 0

D P Euu0 P 0 : (14.9)

Recall that Euu0 D ˝, P 0 P D ˝ 1

and that P 0 is invertible. Multiply both sides by P 0

P 0 Eu u0 D P 0 P Euu0 P 0

1
D˝ ˝P 0
D P 0 , so Eu u0 D I: (14.10)

14.4 Feasible GLS

In practice, we usually do not know ˝. Feasible GLS (FGSL) is typically implemented by

first estimating the model (14.1) with LS, then calculating a consistent estimate of ˝, and
finally using GLS as if ˝ was known with certainty. Very little is known about the finite
sample properties of FGLS, but (the large sample properties) consistency, asymptotic
normality, and asymptotic efficiency (assuming normally distributed errors) can often be

191
established. Evidence from simulations suggests that the FGLS estimator can be a lot
worse than LS if the estimate of ˝ is bad.
To use maximum likelihood when ˝ is unknown requires that we make assumptions
about the structure of ˝ (in terms of a small number of parameters), and more gener-
ally about the distribution of the residuals. We must typically use numerical methods to
maximize the likelihood function.

Example 14.5 (MLE and AR(1) errors.) If u t in Example 14.4 are normally distributed,
then we can use the ˝ 1 in (14.3) to express the likelihood function in terms of the un-
known parameters: ˇ,  , and a. Maximizing this likelihood function requires a numerical
optimization routine.

Bibliography
Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,
Oxford University Press, Oxford.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Hayashi, F., 2000, Econometrics, Princeton University Press.

Johnston, J., and J. DiNardo, 1997, Econometric methods, McGraw-Hill, New York, 4th
edn.

Verbeek, M., 2004, A guide to modern econometrics, Wiley, Chichester, 2nd edn.

192
15 Nonparametric Regressions and Tests

15.1 Nonparametric Regressions

Reference: Campbell, Lo, and MacKinlay (1997) 12.3; Härdle (1990); Pagan and Ullah
(1999); Mittelhammer, Judge, and Miller (2000) 21

15.1.1 Introduction

Nonparametric regressions are used when we are unwilling to impose a parametric form
on the regression equation—and we have a lot of data.
Let the scalars y t and x t be related as

y t D b.x t / C " t ; (15.1)

where " t is uncorrelated over time and where E " t D 0 and E." t jx t / D 0. The function
b./ is unknown and possibly non-linear.
Suppose the sample had 3 observations (say, t D 3, 27, and 99) with exactly the same
value of x t , say 1:9. A natural way of estimating b.x/ at x D 1:9 would then be to
average over these 3 observations as we can expect average of the error terms to be close
to zero (iid and zero mean).
Unfortunately, we seldom have repeated observations of this type. Instead, we may
try to approximate the value of b.x/ (x is a single value, 1.9, say) by averaging over
observations where x t is close to x. The general form of this type of estimator is
PT
O w.x t x/y t
b.x/ D Pt D1
T
; (15.2)
t D1 w.x t x/

where w.x t x/=˙ tTD1 w.x t x/ is the weight on observation t. Note that the denominator
makes the weights sum to unity. The basic assumption behind (15.2) is that the b.x/
function is smooth so local (around x) averaging makes sense.
As an example of a w.:/ function, it could give equal weight to the k values of x t
which are closest to x and zero weight to all other observations (this is the “k-nearest

193
neighbor” estimator, see Härdle (1990) 3.2). As another example, the weight function
O
could be defined so that it trades off the expected squared errors, EŒy t b.x/ 2
, and the
2O 2 2
expected squared acceleration, EŒd b.x/=dx  . This defines a cubic spline (often used
in macroeconomics when x t D t , and is then called the Hodrick-Prescott filter).

Remark 15.1 (Easy way to calculate the “nearest neighbor” estimator, univariate case)
Create a matrix Z where row t is .y t ; x t /. Sort the rows of Z according to the second
column (x). Calculate an equally weighted centered moving average of the first column
(y).

15.1.2 Kernel Regression

A Kernel regression uses a pdf as the weight function, w.x t x/ D K Œ.x t x/= h,
where the choice of h (also called bandwidth) allows us to easily vary the relative weights
of different observations. The pdf of N.0; 1/ is often used for K./. This weighting func-
tion is positive, so all observations get a positive weight, but the weights are highest for
observations close to x and then taper off in a bell-shaped way. A low value of h means
that the weights taper off fast. See Figure 15.1 for an example.
With the N.0; 1/ kernel, we get the following estimator of b.x/ at a point x
PT xt x

2


t D1 K y t exp u =2
O
b.x/ D P h

, where K.u/ D p : (15.3)
T xt x
t D1 K h
2

Remark 15.2 (Kernel as a pdf of N.x; h2 /) If K.z/ is the pdf of an N.0; 1/ variable,
then K Œ.x t x/= h = h is the same as using an N.x; h2 / pdf of x t . Clearly, the 1= h term
would cancel in (15.3).

O
In practice we have to estimate b.x/ at a finite number of points x. This could, for
instance, be 100 evenly spread points in the interval between the minimum and the max-
imum values observed in the sample. See Figure 15.2 for an illustration. Special correc-
tions might be needed if there are a lot of observations stacked close to the boundary of
the support of x (see Härdle (1990) 4.4).

Example 15.3 (Kernel regression) Suppose the sample has three data points Œx1 ; x2 ; x3  D
Œ1:5; 2; 2:5 and Œy1 ; y2 ; y3  D Œ5; 4; 3:5. Consider the estimation of b.x/ at x D 1:9.

194
 Data and weights for b(1.7) Data and weights for b(1.9)

5 5
weights (solid)
⊗ 1 1
⊗
4 4

0 0
1.4 1.6 1.8 2 2.2 2.4 1.4 1.6 1.8 2 2.2 2.4
xt xt

Data on x: 1.5 2.0 2.5

Data and weights for b(2.1) Data on y: 5.0 4.0 3.5
5 N(0,0.252) kernel
o denotes the data
1
⊗ denotes the fitted b(x)
4 ⊗ Left y−axis: data; right y−axis: weights

0
1.4 1.6 1.8 2 2.2 2.4
xt

Figure 15.1: Example of kernel regression with three data points

With h D 1, the numerator in (15.3) is

XT  2
 p
.2 1:9/2 =2 .2:5 1:9/2 =2
K.x t x/y t D e .1:5 1:9/ =2  5 C e 4Ce  3:5 = 2
t D1
p
 .0:92  5 C 1:0  4 C 0:84  3:5/ = 2
p
D 11:52= 2:

The denominator is
XT   p
.1:5 1:9/2 =2 .2 1:9/2 =2 .2:5 1:9/2 =2
K.x t x/ D e Ce Ce = 2
t D1
p
 2:75= 2:

The estimate at x D 1:9 is therefore

O
b.1:9/  11:52=2:75  4:19:

195
 Kernel regression
5.2
Data
5
h= 0.25
4.8 h= 0.2
4.6
4.4
y

4.2
4
3.8
3.6
3.4
1.4 1.6 1.8 2 2.2 2.4
x

Figure 15.2: Example of kernel regression with three data points

Kernel regressions are typically consistent, provided longer samples are accompanied
by smaller values of h, so the weighting function becomes more and more local as the
sample size increases. It can be shown (see Härdle (1990) 3.1 and Pagan and Ullah (1999)
3.3–4) that under the assumption that x t is iid, the mean squared error, variance and bias
of the estimator at the value x are approximately (for general kernel functions)
h i n o2
O
MSE.x/ D Var b.x/ O
C BiasŒb.x/ , with
h
O
i 1  2 .x/ R 1
Var b.x/ D  1 K.u/2 du
T h f .x/
 2 
O 2 1 d b.x/ df .x/ 1 db.x/ R1
BiasŒb.x/ D h  2
C  1 K.u/u2 du: (15.4)
2 dx dx f .x/ dx

In these expressions,  2 .x/ is the variance of the residuals in (15.1) and f .x/ the marginal
density of x. The remaining terms are functions of either the true regression function or
the kernel.
With a N.0; 1/ kernel these expressions can be simplified since
R1 2 1 R1
1 K.u/ du D p and 1 K.u/u2 du D 1, if N.0; 1/ kernel. (15.5)
2 

196
In this case (15.4) becomes
h
O
i 1  2 .x/ 1
Var b.x/ D  p
T h f .x/ 2 
 2 
O 2 1 d b.x/ df .x/ 1 db.x/
BiasŒb.x/ D h  C : (15.6)
2 dx 2 dx f .x/ dx

A smaller h increases the variance (we effectively use fewer data points to estimate
b.x/) but decreases the bias of the estimator (it becomes more local to x). If h decreases
less than proportionally with the sample size (so hT in the denominator of the first term
increases with T ), then the variance goes to zero and the estimator is consistent (since the
bias in the second term decreases as h does).
The variance is a function of the variance of the residuals and the “peakedness” of the
kernel, but not of the b.x/ function. The more concentrated the kernel is (s K.u/2 du
large) around x (for a given h), the less information is used in forming the average around
x, and the uncertainty is therefore larger—which is similar to using a small h. A low
density of the regressors (f .x/ low) means that we have little data at x which drives up
the uncertainty of the estimator.
The bias increases (in magnitude) with the curvature of the b.x/ function (that is,
.d 2 b.x/=dx 2 /2 ). This makes sense, since rapid changes of the slope of b.x/ make it hard
to get b.x/ right by averaging at nearby x values. It also increases with the variance of
the kernel since a large kernel variance is similar to a large h.
It is clear that the choice of h has a major importance on the estimation results. A
lower value of h means a more “local” averaging, which has the potential of picking up
sharp changes in the regression function—at the cost of being more affected by random-
ness. See Figures 15.3–15.4 for an example.
A good (but computationally intensive) approach to choose h is by the leave-one-out
cross-validation technique. This approach would, for instance, choose h to minimize the
expected (or average) prediction error
XT h i2
EPE.h/ D yt bO t .x t ; h/ =T; (15.7)
t D1

where bO t .x t ; h/ is the fitted value at x t when we use a regression function estimated on

a sample that excludes observation t , and a bandwidth h. This means that each prediction
is out-of-sample. To calculate (15.7) we clearly need to make T estimations (for each

197
 Drift vs level, in bins Volatility vs level, in bins
1.5
0.2 Volt+1 = (∆yt+1 − fitted ∆yt+1)2
1
0.1

Volt+1
∆yt+1

0
0.5
−0.1
−0.2 0
0 5 10 15 20 0 5 10 15 20
yt yt
Daily federal funds rates 1954−2006

Figure 15.3: Federal funds rate

x t )—and then repeat this for different values of h to find the minimum.

Remark 15.4 (Speed and fast Fourier transforms) The calculation of the kernel estimator
can often be speeded up by the use of a fast Fourier transform.

If the observations are independent, then it can be shown (see Härdle (1990) 4.2,
Pagan and Ullah (1999) 3.3–6, and also (15.6)) that, with a Gaussian kernel, the estimator
at point x is asymptotically normally distributed
p h i 
1  2
.x/

O
T h b.x/ O
E b.x/ d
! N 0; p ; (15.8)
2  f .x/

where  2 .x/ is the variance of the residuals in (15.1) and f .x/ the marginal density of
x. (A similar expression holds for other choices of the kernel.) This expression assumes
that the asymptotic bias is zero, which is guaranteed if h is decreased (as T increases)
slightly faster than T 1=5 . In practice, to implement this expression  2 .x/ is estimated as
a kernel regression (15.3) but with Œy t b.x O t /2 substituted for y t , f .x/ is estimated by
a kernel density estimate as in (15.24) and it is assumed that the asymptotic bias is zero
O
(E b.x/ D b.x/). Notice that the estimation of  2 .x/ is quite computationally intensive
O
since it requires estimating b.x/ at every point x D x t in the sample.
See Figure 15.5 for an example—and Figure 15.7 for the reason why the confidence
band varies across x values.

198
 Drift vs level, kernel regression Vol vs level, kernel regression
1.5
0.2 Small h Small h
Large h 1 Large h
0.1

Volt+1
∆yt+1

Volt+1 = (∆yt+1 − fitted ∆yt+1)2

0
0.5
−0.1
−0.2 0
0 5 10 15 20 0 5 10 15 20
yt yt
Daily federal funds rates 1954−2006

Drift vs level, LS and LAD Vol vs level, LS and LAD

1.5
0.2 LS LS
LAD 1 LAD
0.1
Volt+1
∆yt+1

0
0.5
−0.1
−0.2 0
0 5 10 15 20 0 5 10 15 20
yt yt

Figure 15.4: Federal funds rate

15.1.3 Multivariate Kernel Regression

Suppose that y t depends on two variables (x t and z t )

y t D b.x t ; z t / C " t ; (15.9)

where " t is uncorrelated over time and where E " t D 0 and E." t jx t ; z t / D 0. This makes
the estimation problem much harder since there are typically few observations in every
bivariate bin (rectangle) of x and z. For instance, with as little as a 20 intervals of each
of x and z, we get 400 bins, so we need a large sample to have a reasonable number of
observations in every bin.

199
 Drift vs level, kernel regression Vol vs level, kernel regression
1.5
0.2 Volt+1 = (∆yt+1 − fitted ∆yt+1)2
Point estimate and 1
0.1

Volt+1
90% confidence band
∆yt+1

0
0.5
−0.1
−0.2 0
0 5 10 15 20 0 5 10 15 20
yt yt
Daily federal funds rates 1954−2006

Figure 15.5: Federal funds rate

In any case, the most common way to implement the kernel regressor is to let
PT    
xt x zt z
K
t D1 x hx
K z hz
yt
Ob.x; z/ D     ; (15.10)
PT xt x zt z
t D1 Kx hx
K z hz

where Kx .u/ and Kz .v/ are two kernels like in (15.3) and where we may allow hx and
hy to be different (and depend on the variance  of xt andy t ). 
In this case, the weight of
xt x zt z
the observation (x t ; z t ) is proportional to Kx hx Kz hz , which is high if both x t
and y t are close to x and y respectively.

15.1.4 Semiparametric Estimation

A possible way out of the curse of dimensionality of the multivariate kernel regression is
to specify a partially linear model

y t D z t0 ˇ C b.x t / C " t ; (15.11)

where " t is uncorrelated over time and where E " t D 0 and E." t jx t ; z t / D 0. This model
is linear in z t , but possibly non-linear in x t since the function b.x t / is unknown.
To construct an estimator, start by taking expectations of (15.11) conditional on x t

E.y t jx t / D E.z t jx t /0 ˇ C b.x t /: (15.12)

200
Subtract from (15.11) to get

yt E.y t jx t / D Œz t E.z t jx t /0 ˇ C " t : (15.13)

The “double residual” method (see Pagan and Ullah (1999) 5.2) has several steps. First,
estimate E.y t jx t / by a kernel regression of y t on x t , bOy .x/, and E.z t jx t / by a similar
kernel regression of z t on x t , bOz .x/. Second, use these estimates in (15.13)

yt bOy .x t / D Œz t bOz .x t /0 ˇ C " t (15.14)

and estimate ˇ by least squares. Third, use these estimates in (15.12) to estimate b.x t / as

O t / D bOy .x t /
b.x O
bOz .x t /0 ˇ: (15.15)

It can be shown that (under the assumption that y t , z t and x t are iid)
p
T .ˇO ˇ/ !d N 0; Var." t / Cov.z t jx t / 1 :
 
(15.16)

We can consistently estimate Var." t / by the sample variance of the fitted residuals in
(15.11)—plugging in the estimated ˇ and b.x t /: and we can also consistently estimate
Cov.z t jx t / by the sample variance of z t bOz .x t /. Clearly, this result is as if we knew the
non-parametric parts of the problem (which relies on the consistency of their estimators),
so it is only an asymptotic results. By using this result, together with the known properties
of the kernel regressions (see, for instance, 15.6), it should be possible to apply the delta
method to (15.15) to construct the asymptotic variance of b.x O t / (that is, at a given point
x t ).

15.2 Estimating and Testing Distributions

Reference: Harvey (1989) 260, Davidson and MacKinnon (1993) 267, Silverman (1986);
Mittelhammer (1996), DeGroot (1986)

201
15.2.1 Parametric Tests of Normal Distribution

The skewness, kurtosis and Bera-Jarque test for normality are useful diagnostic tools. For
an iid normally distributed variable, x t iid N ;  2 , they are

Test statistic Distribution

3
D T1 TtD1 xt  N 0; T6
P

skewness

4 (15.17)
D T1 TtD1 xt  N 3; 24
P

kurtosis T
Bera-Jarque D T6 skewness2 C 24
T
.kurtosis 3/2 22 :

This is implemented by using the estimated mean and standard deviation. The distribu-
tions stated on the right hand side of (15.17) are under the null hypothesis that x t is iid
N ;  2 .

The intuition for the 22 distribution of the Bera-Jarque test is that both the skewness
and kurtosis are, if properly scaled, N.0; 1/ variables. It can also be shown that they,
under the null hypothesis, are uncorrelated. The Bera-Jarque test statistic is therefore a
sum of the square of two uncorrelated N.0; 1/ variables, which has a 22 distribution.
The Bera-Jarque test can also be implemented as a test of overidentifying restrictions
in GMM. The moment conditions
2 3
xt 
T 6 2 2 7
7
2 1 X6 .x t / 
g.;  / D 6
3
7; (15.18)
T t D1 6
4 .x t /
7
5
.x t /4 3 4

should all be zero if x t is N.;  2 /. We can estimate the two parameters,  and  2 , by
using the first two moment conditions only, and then test if all four moment conditions
are satisfied. It can be shown that this is the same as the Bera-Jarque test if x t is indeed
iid N.;  2 /.

15.2.2 Nonparametric Tests of General Distributions

The Kolmogorov-Smirnov test is designed to test if an empirical distribution function,

EDF.x/, conforms with a theoretical cdf, F .x/. The empirical distribution function is

202
 Empirical distribution function and theoretical cdf

0.8

0.6

0.4
data
0.2 cdf of N(3,1)
EDF
0

3 3.5 4 4.5 5 5.5 6

x

Figure 15.6: Example of empirical distribution function

defined as the fraction of observations which are less or equal to x, that is,
T
(
1X 1 if q is true
EDF .x/ D ı.x t  x/; where ı.q/ D (15.19)
T tD1 0 else.

The EDF.x t / and F .x t / are often plotted against the sorted (in ascending order) sample
fx t gTtD1 . See Figure 15.6.

Example 15.5 (EDF) Suppose we have a sample with three data points: Œx1 ; x2 ; x3  D
Œ5; 3:5; 4. The empirical distribution function is then as in Figure 15.6.

Define the absolute value of the maximum distance

DT D max jEDF .x t / F .x t /j : (15.20)

xt

Example 15.6 (Kolmogorov-Smirnov test statistic) Figure 15.6 also shows the cumula-
tive distribution function (cdf) of a normally distributed variable. The test statistic (15.20)
is then the largest difference (in absolute terms) of the EDF and the cdf—among the ob-
served values of x t .

203
 p
We reject the null hypothesis that EDF.x/ D F .x/ if T D t > c, where c is a critical
value which can be calculated from
p  1
2i 2 c 2
X
lim Pr T DT  c D 1 2 . 1/i 1
e : (15.21)
T !1
i D1

It can be approximated by replacing 1 with a large number (for instance, 100). For
instance, c D 1:35 provides a 5% critical value. There is a corresponding test for com-
paring two empirical cdfs.
Pearson’s 2 test does the same thing as the K-S test but for a discrete distribution.
Suppose you have K categories with Ni values in category i. The theoretical distribution
predicts that the fraction pi should be in category i , with K
P
i D1 pi D 1. Then

K
X .Ni Tpi /2 2
 K 1: (15.22)
i D1
Tpi

There is a corresponding test for comparing two empirical distributions.

15.2.3 Kernel Density Estimation

Reference: Silverman (1986)

A histogram is just a count of the relative number of observations that fall in (pre-
specified) non-overlapping intervals. If we also divide by the width of the interval, then
the area under the histogram is unity, so the scaled histogram can be interpreted as a den-
sity function. For instance, if the intervals (“bins”) are a wide, then the scaled histogram
can be defined as
T
(
1 X1 1 if q is true
g.xjx is in bini / D ı.x t is in bini /; where ı.q/ D (15.23)
T t D1 a 0 else.

Note that the area under g.x/ indeed integrates to unity.

We can gain efficiency by using a more sophisticated estimator. In particular, using
a pdf instead of the binary function is often both convenient and more efficient. The
N.0; h2 / is often used. The kernel density estimator of the pdf at some point x is then

exp u2 =2


1 XT x
t x 
fO .x/ D K , where K.u/ D p : (15.24)
Th t D1 h 2

204
 Histogram (scaled: area=1) Kernel density estimation
0.2 0.2
Small h
0.15 0.15 Large h
0.1 0.1

0.05 0.05

0 0
0 5 10 15 20 0 5 10 15 20
yt yt

Daily federal funds rates 1954−2006

K−S (against N(µ,σ2)): sqrt(T)D = 15.32

Skewness: 1.251
kurtosis: 5.235
Bera−Jarque: 8996

Figure 15.7: Federal funds rate

The value h D Std.x t /1:06T 1=5 is sometimes recommended, since it can be shown to be
the optimal choice (in MSE sense) if data is normally distributed and the N.0; 1/ kernel
is used. Clearly, using K Œ.x t x/= h = h is the same as using pdf of N.x; h2 /. The
bandwidth h could be chosen by a leave-one-out cross-validation technique.
The results on bias and variance in (15.4) are approximately true also for the kernel
density estimation if we interpret b.x/ as the pdf of x. In particular, it can be shown that
(with iid data and a Gaussian kernel) the asymptotic distribution is
p h i 
1

O O d
T h f .x/ E f .x/ ! N 0; p f .x/ ; (15.25)
2 
The easiest way to handle a bounded support of x is to transform the variable into one
with an unbounded support, estimate the pdf for this variable, and then use the “change
of variable” technique to transform to the pdf of the original variable.
We can also estimate multivariate pdfs. Let x t be a d 1 matrix and ˝O be the estimated
covariance matrix of x t . We can then estimate the pdf at a point x by using a multivariate

205
Gaussian kernel as
1 XT
fO .x/ D expŒ .x x t /0 ˝O 1
.x x t /=.2h2 /: (15.26)
O 1=2
T hd .2/d=2 j˝j t D1

1=.d C4/
The value h D 0:96T is sometimes recommended.

Kernel Density Estimation and LAD

Reference: Amemiya (1985) 4.6

Let uO t .b/ be the residuals in a regression model for choice b of the coefficients,
uO t .b/ D y t x t0 b. The least absolute deviations (LAD) estimator solve
T
X
ˇOLAD D arg min juO t .b/j (15.27)
b
t D1

This estimator involve non-linearities, but a a simple iteration works nicely. If we assume
that the median of the true residual, u t , is zero, then we (typically) have
p XT
T .ˇOLAD ˇ0 / !d N 0; f .0/ 2 ˙xx1 =4 , where ˙xx D plim x t x t0 =T;
 
t D1
(15.28)
where f .0/ is the value of the pdf of u t at zero. Unless we know this density function
(or else we would probably have used MLE instead of LAD), we need to estimate it—for
instance with a kernel density method.
p
Example 15.7 (N.0;  2 /) When u t  N.0;  2 ), then f .0/ D 1= 2 2 , so the covari-
ance matrix in (15.28) becomes  2 ˙xx1 =2. This is =2 times larger than when using
LS.

Remark 15.8 (Algorithm for LAD) The LAD estimator can be written
T
X
ˇOLAD D arg min w t uO t .b/2 , w t D 1= juO t .b/j ;
ˇ
t D1

so it is a weighted least squares where both y t and x t are multiplied by 1= juO t .b/j. It can
be shown that iterating on LS with the weights given by 1= juO t .b/j, where the residuals
are from the previous iteration, converges very quickly to the LAD estimator.

206
Bibliography
Amemiya, T., 1985, Advanced econometrics, Harvard University Press, Cambridge, Mas-
sachusetts.

Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The econometrics of financial

markets, Princeton University Press, Princeton, New Jersey.

Davidson, R., and J. G. MacKinnon, 1993, Estimation and inference in econometrics,

Oxford University Press, Oxford.

DeGroot, M. H., 1986, Probability and statistics, Addison-Wesley, Reading, Mas-

sachusetts.

Härdle, W., 1990, Applied nonparametric regression, Cambridge University Press, Cam-
bridge.

Harvey, A. C., 1989, Forecasting, structural time series models and the Kalman filter,
Cambridge University Press.

Mittelhammer, R. C., 1996, Mathematical statistics for economics and business, Springer-
Verlag, New York.

Mittelhammer, R. C., G. J. Judge, and D. J. Miller, 2000, Econometric foundations, Cam-

bridge University Press, Cambridge.

Pagan, A., and A. Ullah, 1999, Nonparametric econometrics, Cambridge University

Press.

Silverman, B. W., 1986, Density estimation for statistics and data analysis, Chapman and
Hall, London.

207
 .D..3

16 Alphas /Betas and Investor Characteristics

16.1 Basic Setup

The task is to evaluate if alphas or betas of individual investors (or funds) are related
to investor (fund) characteristics, for instance, age or trading activity. The data set is
panel with observations for T periods and N investors. (In many settings, the panel is
unbalanced, but, to keep things reasonably simple, that is disregarded in the discussion
below.)

16.2 Calendar Time and Cross Sectional Regression

The calendar time (CalTime) approach is to first define M discrete investor groups (for
instance, age 18–30, 31–40, etc) and calculate their respective average excess returns (yNjt
for group j )
1 P
yNjt D yi t ; (16.1)
Nj i 2Groupj
where Nj is the number of individuals in group j .
Then, we run a factor model

yNjt D x t0 ˇj C vjt ; for j D 1; 2; : : : ; M (16.2)

where x t typically includes a constant and various return factors (for instance, excess re-
turns on equity and bonds). By estimating these M equations as a SURE system with
White’s (or Newey-West’s) covariance estimator, it is straightforward to test various hy-
potheses, for instance, that the intercept (the “alpha”) is higher for the M th group than
for the for first group.

Example 16.1 (CalTime with two investor groups) With two investor groups, estimate the

208
following SURE system

yN1t D x t0 ˇ1 C v1t ;
yN2t D x t0 ˇ2 C v2t :

The CalTime approach is straightforward and the cross-sectional correlations are fairly
easy to handle (in the SURE approach). However, it forces us to define discrete investor
groups—which makes it hard to handle several different types of investor characteristics
(for instance, age, trading activity and income) at the same time.
The cross sectional regression (CrossReg) approach is to first estimate the factor
model for each investor

yi t D x t0 ˇi C "i t ; for i D 1; 2; : : : ; N (16.3)

and to then regress the (estimated) betas for the pth factor (for instance, the intercept) on
the investor characteristics
ˇOpi D zi0 cp C wpi : (16.4)

In this second-stage regression, the investor characteristics zi could be a dummy variable

(for age roup, say) or a continuous variable (age, say). Notice that using a continuos
investor characteristics assumes that the relation between the characteristics and the beta
is linear—something that is not assumed in the CalTime approach. (This saves degrees of
freedom, but may sometimes be a very strong assumption.) However, a potential problem
with the CrossReg approach is that it is often important to account for the cross-sectional
correlation of the residuals.

16.3 Panel Regressions, Driscoll-Kraay and Cluster Methods

References: Hoechle (2011) and Driscoll and Kraay (1998)

16.3.1 OLS

Consider the regression model

yi t D xi0 t ˇ C "i t ; (16.5)

209
where xi t is an K  1 vector. Notice that the coefficients are the same across individuals
(and time). Define the matrices
T N
1 XX
˙xx D xi t xi0 t (an K  K matrix) (16.6)
T N t D1 i D1
T N
1 XX
˙xy D xi t yi t (a K  1 vector). (16.7)
T N t D1 i D1

The LS estimator (stacking all T N observations) is then

ˇO D ˙xx1 ˙xy : (16.8)

16.3.2 GMM

The sample moment conditions for the LS estimator are

T N
1X 1 X
hi t D 0K1 , where hi t D xi t "i t D xi t .yi t xi0 t ˇ/: (16.9)
T t D1 N i D1

Remark 16.2 (Distribution of GMM estimates) Under fairly weak assumption, the ex-
p d
actly identified GMM estimator T N .ˇO ˇ0 / ! N.0; D0 1 S0 D0 1 /, where D0 is the
p
Jacobian of the average moment conditions and S0 is the covariance matrix of T N
times the average moment conditions.

Remark 16.3 (Distribution of ˇO ˇ0 ) As long as T N is finite, we can (with some abuse

p
of notation) consider the distribution of ˇO ˇ instead of T N .ˇO ˇ0 / to write

ˇO ˇ0  N.0; D0 1 SD0 1 /;

where S D S0 =.T N / which is the same as the covariance matrix of the average moment
conditions (16.9).

To apply these remarks, first notice that the Jacobian D0 corresponds to (the probabil-
ity limit of) the ˙xx matrix in (16.6). Second, notice that
T N
!
1X 1 X
Cov.average moment conditions/ D Cov hi t (16.10)
T t D1 N i D1

210
looks differently depending on the assumptions of cross correlations.
In particular, if hi t has no correlation across time (effectively, N1 N
P
i D1 hi t is not auto-
correlated), then we can simplify as
T N
!
1 X 1 X
Cov.average moment conditions/ D 2 Cov hi t : (16.11)
T t D1 N i D1

We would then design an estimator that would consistently estimate this covariance matrix
by using the time dimension.

Example 16.4 (DK on T D 2 and N D 4) As an example, suppose K D 1, T D 2 and

N D 4. Then, (16.10) can be written
 
1 1
Cov .h1t C h2t C h3t C h4t / C .h1;tC1 C h2;tC1 C h3;tC1 C h4;tC1 / :
24 24
If there is no correlation across time periods, then this becomes
   
1 1 1 1
Cov .h1t C h2t C h3t C h4t / C 2 Cov .h1;t C1 C h2;t C1 C h3;t C1 C h4;t C1 / ;
22 4 2 4
which has the same form as (16.11).

16.3.3 Driscoll-Kraay

The Driscoll and Kraay (1998) (DK) covariance matrix is

O D ˙ 1 S˙ 1 ;
Cov.ˇ/ (16.12)
xx xx

where
T N
1 X 0 1 X
SD 2 h t h t ; with h t D hi t , hi t D xi t "i t ; (16.13)
T t D1 N i D1
where hi t is the LS moment condition for individual i . Clearly, hi t and h t are K  1, so S
is KK. Since we use the covariance matrix of the moment conditions, heteroskedasticity
is accounted for.
Notice that h t is the cross-sectional average moment condition (in t ) and that S is an

211
estimator of the covariance matrix of those average moment conditions

b
 
1 PT PN
S D Cov hi t :
T N tD1 i D1
To calculate this estimator, (16.13) uses the time dimension (and hence requires a reason-
ably long time series).

Remark 16.5 (Relation to the notation in Hoechle (2011)) Hoechle writes Cov.ˇ/ O D
.X 0 X/ 1 SOT .X 0 X/ 1 , where SOT D TtD1 hO t hO 0t ; with hO t D N
P P
i D1 hi t . Clearly, my ˙xx D
X 0 X=.T N / and my S D SOT =.T 2 N 2 /. Combining gives Cov.ˇ/ O D .˙xx T N / 1 ST 2 N 2 .˙xx T N / 1 ,

which simplifies to (16.12).

Example 16.6 (DK on N D 4) As an example, suppose K D 1 and N D 4. Then,

(16.13) gives the cross-sectional average in period t
1
ht D .h1t C h2t C h3t C h4t / ;
4
and the covariance matrix
T
1 X
SD 2 h t h0t
T t D1
T  2
1 X 1
D 2 .h1t C h2t C h3t C h4t /
T t D1 4
T
1 X 1 2
D .h C h22t C h23t C h24t ;
T 2 t D1 16 1t
C 2h1t h2t C 2h1t h3t C 2h1t h4t C 2h2t h3t C 2h2t h4t C 2h3t h4t /

so we can write
" 4
1 X
SD c it /
Var.h
T  16 i D1
b b b
C 2Cov.h1t ; h2t / C 2Cov.h1t ; h3t / C 2Cov.h1t ; h4t /
bov.h ; h / C 2Cbov.h
C 2C 2t 3t 2t ; h4t /

b i
C2Cov.h ; h / :
3t 4t

212
Notice that S is the (estimate of) the variance of the cross-sectional average, Var.h t / D
VarŒ.h1t C h2t C h3t C h4t /=4.

A cluster method puts restrictions on the covariance terms (of hi t ) that are allowed
to enter the estimate S. In practice, all terms across clusters are left out. This can be
implemented by changing the S matrix. In particular, instead of interacting all i with
each other, we only allow for interaction within each of the G clusters (g D 1; :::; G/
G T
X 1 X g g
0 1
h t h t , where hgt D
X
SD 2
hi t : (16.14)
gD1
T t D1 N i 2 cluster g

(Remark: the cluster sums should be divided by N , not the number of individuals in the
cluster.)

Example 16.7 (Cluster method on N D 4, changing Example 16.6 directly) Reconsider

Example 16.6, but assume that individuals 1 and 2 form cluster 1 and that individuals 3
and 4 form cluster 2—and disregard correlations across clusters. This means setting the
covariances across clusters to zero,
T
1 X 1 2
SD 2 .h1t C h22t C h23t C h24t ;
T t D1 16
2h1t h2t C 2h1t h3t C 2h1t h4t C 2h2t h3t C 2h2t h4t C 2h3t h4t /
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
0 0 0 0

so we can write

b b
" 4 #
1 X
SD c i t / C 2Cov.h1t ; h2t / C 2Cov.h3t ; h4t / :
Var.h
T  16 i D1

Example 16.8 (Cluster method on N D 4) From (16.14) we have the cluster (group)
averages
1 1
h1t D .h1t C h2t / and h2t D .h3t C h4t / :
4 4
T
0
Assuming only one regressor (to keep it simple), the time averages, T1 hgt hgt , are
P
t D1

213
then (for cluster 1 and then 2)
T T  2 T
1 X 1 1
0 1X 1 1X 1 2
h1t C h22t C 2h1t h2t , and


ht ht D .h1t C h2t / D
T t D1 T t D1 4 T t D1 16
T T
1 X 2 2
0 1X 1 2
h3t C h24t C 2h3t h4t :


ht ht D
T t D1 T t D1 16

Finally, summing across these time averages gives the same expression as in Example
16.7. The following 4  4 matrix illustrates which cells that are included (assumption: no
dependence across time)

i 1 2 3 4
2
1 h1t h1t h2t 0 0
2 h1t h2t h22t 0 0
2
3 0 0 h3t h3t h4t
4 0 0 h3t h4t h24t

In comparison, the iid case only sums up the principal diagonal, while the DK method
fills the entire matrix.

Instead, we get White’s covariance matrix by excluding all cross terms. This can be
accomplished by defining
T N
1 X 1 X
SD 2 2
hi t h0i t : (16.15)
T t D1 N i D1

Example 16.9 (White’s method on N D 4) With only one regressor (16.15) gives
T
1 X 1 2
h1t C h22t C h23t C h24t


SD 2
T tD1 16
4
1 X
D c it /
Var.h
T  16 i D1

Finally, the traditional LS covariance matrix assumes that E hi t h0i t D ˙xx  E "2it , so
we get
T N
O 1 2 2 1 XX 2
CovLS .ˇ/ D ˙xx s =T N , where s D " : (16.16)
T N t D1 i D1 i t

214
Remark 16.10 (Why the cluster method fails when there is a missing “time fixed effect”—
and one of the regressors indicates the cluster membership) To keep this remark short,
assume yi t D 0qi t C "i t , where qi t indicates the cluster membership of individual i (con-
stant over time). In addition, assume that all individual residuals are entirely due to an
(excluded) time fixed effect, "i t D w t . Let N D 4 where i D .1; 2/ belong to the first
cluster (qi D 1) and i D .3; 4/ belong to the second cluster (qi D 1). (Using the values
qi D ˙1 gives qi a zero mean, which is convenient.) It is straightforward to demon-
strate that the estimated (OLS) coefficient in any sample must be zero: there is in fact no
uncertainty about it. The individual moments in period t are then hi t D qi t  w t
2 3 2 3
h1t wt
6 7 6 7
6 h2t 7 6 w t 7
6 h 7 D 6 w 7:
6 7 6 7
4 3t 5 4 t 5
h4t wt

The matrix in Example 16.8 is then

i 1 2 3 4
2 2
1 wt wt 0 0
2 2
2 wt wt 0 0
3 0 2
0 w t w t2
4 0 0 w t2 w t2

These elements sum up to a positive number—which is wrong since N

P
i D1 hi t D 0 by
definition, so its variance should also be zero. In contrast, the DK method adds the off-
diagonal elements which are all equal to w t2 , so summing the whole matrix indeed gives
zero. If we replace the qi t regressor with something else (eg a constant), then we do not
get this result.
To see what happens if the qi variable does not coincide with the definitions of the clus-
ters change the regressor to qi D . 1; 1; 1; 1/ for the four individuals. We then get
.h1t ; h2t ; h3t ; h4t / D . w t ; w t ; w t ; w t /. If the definition of the clusters (for the covari-

215
ance matrix) are unchanged, then the matrix in Example 16.8 becomes

i 1 2 3 4
1 w t2 w t2 0 0
2 w t2 w t2 0 0
3 0 0 w t2 w t2
4 0 0 w t2 w t2

which sum to zero: the cluster covariance estimator works fine. The DK method also
works since it adds the off-diagonal elements which are

i 1 2 3 4
1 w t2 w t2
2 w t2 w t2
3 w t2 w t2
4 w t2 w t2

which also sum to zero. This suggests that the cluster covariance matrix goes wrong
only when the cluster definition (for the covariance matrix) is strongly related to the qi
regressor.

16.4 From CalTime To a Panel Regression

The CalTime estimates can be replicated by using the individual data in the panel. For
instance, with two investor groups we could estimate the following two regressions

yi t D x t0 ˇ1 C u.1/
i t for i 2 group 1 (16.17)
yi t D x t0 ˇ2 C u.2/
i t for i 2 group 2. (16.18)

More interestingly, these regression equations can be combined into one panel regres-
sion (and still give the same estimates) by the help of dummy variables. Let zj i D 1 if
individual i is a member of group j and zero otherwise. Stacking all the data, we have

216
(still with two investor groups)

yi t D .z1i x t /0 ˇ1 C .z2i x t /0 ˇ2 C ui t
" #!0 " #
z1i x t ˇ1
D C ui t
z2i x t ˇ2
# "
z 1i
D .zi ˝ x t /0 ˇ C ui t , where zi D : (16.19)
z2i

This is estimated with LS by stacking all N T observations.

Since the CalTime approach (16.2) and the panel approach (16.19) give the same
coefficients, it is clear that the errors in the former are just group averages of the errors in
the latter
1 P .j /
vjt D i 2Group j ui t : (16.20)
Nj
We know that
1
Var.vjt / D . i i  ih / C  ih ; (16.21)
Nj
where  i i is the average Var.u.j / .j / .j /
i t / and  ih is the average Cov.ui t ; uht /. With a large
cross-section, only the covariance matters. A good covariance estimator for the panel
approach will therefore have to handle the covariance with a group—and perhaps also
the covariance across groups. This suggests that the panel regression needs to handle the
cross-correlations (for instance, by using the cluster or DK covariance estimators).

16.5 The Results in Hoechle, Schmid and Zimmermann

Hoechle, Schmid, and Zimmermann (2009) (HSZ) suggest the following regression on all
data (t D 1; : : : ; T and also i D 1; : : : ; N )

yi t D .zi t ˝ x t /0 d C vi t (16.22)
D .Œ1; z1i t ; : : : ; zmi t  ˝ Œ1; x1t ; : : : ; xk t /0 d C vi t ; (16.23)

where yi t is the return of investor i in period t , zqi t measures characteristics q of investor

i in period t and where xpt is the pth pricing factor. In many cases zj i t is time-invariant
and could even be just a dummy: zj i t D 1 if investor i belongs to investor group j
(for instance being 18–30 years old). In other cases, zj i t is still time invariant and con-

217
tains information about the number of fund switches as well as other possible drivers of
performance like gender. The x t vector contains the pricing factors. In case the charac-
teristics z1i t ; : : : ; zmi t sum to unity (for a given individual i and time t ), the constant in
Œ1; z1i t ; : : : ; zmi t  is dropped.
This model is estimated with LS (stacking all N T observations), but the standard
errors are calculated according to Driscoll and Kraay (1998) (DK)—which accounts for
cross-sectional correlations, for instance, correlations between the residuals of different
investors (say, v1t and v7t ).
HSZ prove the following two propositions.

Proposition 16.11 If the zi t vector in (16.22) consists of dummy variables indicating

exclusive and constant group membership (z1i t D 1 means that investor i belongs to
group 1, so zj i t D 0 for j D 2; :::; m), then the LS estimates and DK standard errors
of (16.22) are the same as LS estimates and Newey-West standard errors of the CalTime
approach (16.2). (See HSZ for a proof.)

Proposition 16.12 (When zi t is a measure of investor characteristics, eg number of fund

switches) The LS estimates and DK standard errors of (16.22) are the same as the LS
estimates of CrossReg approach (16.4), but where the standard errors account for the
cross-sectional correlations, while those in the CrossReg approach do not. (See HSZ for
a proof.)

Example 16.13 (One investor characteristic and one pricing factor). In this case (16.22)
is
2 30
1
6 7
6 x1t 7
yi t D 6
6 z 7 d C vi t ;
7
4 it 5
zi t x1t
D d0 C d1 x1t C d2 zi t C d3 zi t x1t C vi t :

In case we are interested in how the investor characteristics (zi t ) affect the alpha (inter-
cept), then d2 is the key coefficient.

218
16.6 Monte Carlo Experiment

16.6.1 Basic Setup

This section reports results from a simple Monte Carlo experiment. We use the model

yi t D ˛ C ˇf t C ıgi C "i t ; (16.24)

where yi t is the return of individual i in period t, f t a benchmark return and gi is the

(demeaned) number of the cluster ( 2; 1; 0; 1; 2) that the individual belongs to. This is
a simplified version of the regressions we run in the paper. In particular, ı measures how
the performance depends on the number of fund switches.
The experiment uses 3000 artificial samples with t D 1; : : : ; 2000 and i D 1; : : : ; 1665.
Each individual is a member of one of five equally sized groups (333 individuals in each
group). The benchmark return f t is iid normally distributed with a zero mean and a stan-
p
dard deviation equal to 15= 250, while "i t is a also normally distributed with a zero
mean and a standard deviation of one (different cross-sectional correlations are shown in
the table). In generating the data, the true values of ˛ and ı are zero, while ˇ is one—and
these are also the hypotheses tested below. To keep the simulations easy to interpret, there
is no autocorrelation or heteroskedasticity.
Results for three different GMM-based methods are reported: Driscoll and Kraay
(1998), a cluster method and White’s method. To keep the notation short, let the re-
gression model be yi t D xi0 t b C "i t , where xi t is a K  1 vector of regressors. The (least
squares) moment conditions are
1 PT PN
hi t D 0K1 , where hi t D xi t "i t : (16.25)
T N t D1 i D1
Standard GMM results show that the variance-covariance matrix of the coefficients is

O D ˙ 1 S˙ 1 , where ˙xx D 1 PT PN
Cov.b/ xx xx xi t xi0 t ; (16.26)
T N t D1 i D1
and S is covariance matrix of the moment conditions.

219
 The three methods differ with respect to how the S matrix is estimated
1 PT 0
PN
SDK D t D1 h t h t , where h t D i D1 hi t ;
T 2N 2
1 PT PM j j 0 j
X
SC l D t D1 j D1 h t .h t / , where h t D hi t ;
T 2N 2 i 2 cluster j
1 PT PN 0
SW h D t D1 i D1 hi t hi t : (16.27)
T 2N 2
To see the difference, consider a simple example with N D 4 and where i D .1; 2/ belong
to the first cluster and i D .3; 4/ belong to the second cluster. The following matrix shows
the outer product of the moment conditions of all individuals. White’s estimator sums up
the cells on the principal diagonal, the cluster method adds the underlined cells, and the
DK method adds also the remaining cells
2 3
i 1 2 3 4
6 1 h1t h01t h1t h02t h1t h03t h1t h04t 7
6 7
6 7
0 0 0 0 7
2 h h h h h h h h (16.28)
6
6 2t 1t 2t 2t 2t 3t 2t 4t 7
6 7
6 3 h3t h0 h3t h0 h3t h0 h3t h0 7
4 1t 2t 3t 4t 5
4 h4t h1t h4t h2t h4t h3t h4t h04t
0 0 0

16.6.2 MC Covariance Structure

To generate data with correlated (in the cross-section) residuals, let the residual of indi-
vidual i (belonging to group j ) in period t be

"i t D ui t C vjt C w t ; (16.29)

where uit  N.0; u2 ), vjt  N.0; v2 ) and w t  N.0; w2 )—and the three components
are uncorrelated. This implies that

Var."i t / D u2 C v2 C w2 ;

" #
v2 C w2 if individuals i and k belong to the same group
Cov."i t ; "k t / D
w2 otherwise.
(16.30)

220
Clearly, when w2 D 0 then the correlation across groups is zero, but there may be corre-
lation within a group. If both v2 D 0 and w2 D 0, then there is no correlation at all across
individuals. For CalTime portfolios (one per activity group), we expect the ui t to average
out, so a group portfolio has the variance v2 C w2 and the covariance of two different
group portfolios is w2 .
The Monte Carlo simulations consider different values of the variances—to illustrate
the effect of the correlation structure.

16.6.3 Results from the Monte Carlo Simulations

Table 16.1 reports the fraction of times the absolute value of a t-statistics for a true null
hypothesis is higher than 1.96. The table has three panels for different correlation patterns
the residuals ("i t ): no correlation between individuals, correlations only within the pre-
specified clusters and correlation across all individuals.
In the upper panel, where the residuals are iid, all three methods have rejection rates
around 5% (the nominal size).
In the middle panel, the residuals are correlated within each of the five clusters, but
there is no correlation between individuals that belong to the different clusters. In this
case, but the DK and the cluster method have the right rejection rates, while White’s
method gives much too high rejection rates (around 85%). The reason is that White’s
method disregards correlation between individuals—and in this way underestimates the
uncertainty about the point estimates. It is also worth noticing that the good performance
of the cluster method depends on pre-specifying the correct clustering. Further simula-
tions (not tabulated) shows that with a completely random cluster specification (unknown
to the econometrician), gives almost the same results as White’s method.
The lower panel has no cluster correlations, but all individuals are now equally cor-
related (similar to a fixed time effect). For the intercept (˛) and the slope coefficient on
the common factor (ˇ), the DK method still performs well, while the cluster and White’s
methods give too many rejects: the latter two methods underestimate the uncertainty since
some correlations across individuals are disregarded. Things are more complicated for the
slope coefficient of the cluster number (ı). Once again, DK performs well, but both the
cluster and White’s methods lead to too few rejections. The reason is the interaction of
the common component in the residual with the cross-sectional dispersion of the group
number (gi ).

221
 To understand this last result, consider a stylised case where yi t D ıgi C "i t where
ı D 0 and "i t D w t so all residuals are due to an (excluded) time fixed effect. In this
case, the matrix above becomes
2 3
i 1 2 3 4
6 1 w t2 2 2 2 7
6 7
6 w t w t w t 7
6 2 w t2 w t2 w t2 w t2 7 (16.31)
6 7
6 7
6 3
4 w t2 w t2 w t2 w t2 75
4 w t2 w t2 w t2 w t2

(This follows from gi D . 1; 1; 1; 1/ and since hi t D gi w t we get .h1t ; h2t ; h3t ; h4t / D
. w t ; w t ; w t ; w t /.) Both White’s and the cluster method sums up only positive cells,
so S is a strictly positive number. (For this the cluster method, this result relies on the as-
sumption that the clusters used in estimating S correspond to the values of the regressor,
gi .) However, that is wrong since it is straightforward to demonstrate that the estimated
coefficient in any sample must be zero. This is seen by noticing that N
P
i D1 hi t D 0 at
a zero slope coefficient holds for all t, so there is in fact no uncertainty about the slope
coefficient. In contrast, the DK method adds the off-diagonal elements which are all equal
to w t2 , giving the correct result S D 0.

16.7 An Empirical Illustration

See 16.2 for results on a ten-year panel of some 60,000 Swedish pension savers (Dahlquist,
Martinez and Söderlind, 2011).

Bibliography
Driscoll, J., and A. Kraay, 1998, “Consistent Covariance Matrix Estimation with Spatially
Dependent Panel Data,” Review of Economics and Statistics, 80, 549–560.

Hoechle, D., 2011, “Robust Standard Errors for Panel Regressions with Cross-Sectional
Dependence,” The Stata Journal forhcoming.

Hoechle, D., M. M. Schmid, and H. Zimmermann, 2009, “A Generalization of the Cal-

222
endar Time Portfolio Approach and the Performance of Private Investors,” Working
paper, University of Basel.

223
 Driscoll-
White Cluster Kraay

A. No cross-sectional correlation

˛ 0.049 0.049 0.050

ˇ 0.044 0.045 0.045
0.050 0.051 0.050

B. Within-cluster correlations

˛ 0.853 0.053 0.054

ˇ 0.850 0.047 0.048
0.859 0.049 0.050

C. Within- and between-cluster correlations

˛ 0.935 0.377 0.052

ˇ 0.934 0.364 0.046
0.015 0.000 0.050

Table 16.1: Simulated size of different covariance estimators This table presents the
fraction of rejections of true null hypotheses for three different estimators of the co-
variance matrix: White’s (1980) method, a cluster method, and Driscoll and Kraay’s
(1998) method. The model of individual i in period t and who belongs to cluster j is
ri t D ˛ C ˇf t C gi C "i t , where f t is a common regressor (iid normally distributed)
and gi is the demeaned number of the cluster that the individual belongs to. The sim-
ulations use 3000 repetitions of samples with t D 1; : : : ; 2000 and i D 1; : : : ; 1665.
Each individual belongs to one of five different clusters. The error term is constructed as
"i t D ui t C vjt C w t , where ui t is an individual (iid) shock, vjt is a shock common to
all individuals who belong to cluster j , and w t is a shock common to all individuals. All
shocks are normally distributed. In Panel A the variances of .ui t ; vjt ; w t / are (1,0,0), so
the shocks are iid; in Panel B the variances are (0.67,0.33,0), so there is a 33% correlation
within a cluster but no correlation between different clusters; in Panel C the variances are
(0.67,0,0.33), so there is no cluster-specific shock and all shocks are equally correlated,
effectively having a 33% correlation within a cluster and between clusters.

224
 Table 16.2: Investor activity, performance, and characteristics

I II III IV

Constant –0.828 –1.384 –0.651 –1.274

(2.841) (3.284) (2.819) (3.253)
Default fund 0.406 0.387 0.230 0.217
(1.347) (1.348) (1.316) (1.320)
1 change 0.117 0.125
(0.463) (0.468)
2– 5 changes 0.962 0.965
(0.934) (0.934)
6–20 changes 2.678 2.665
(1.621) (1.623)
21–50 changes 4.265 4.215
(2.074) (2.078)
51– changes 7.114 7.124
(2.529) (2.535)
Number of changes 0.113 0.112
(0.048) (0.048)
Age 0.008 0.008
(0.011) (0.011)
Gender 0.306 0.308
(0.101) (0.101)
Income –0.007 0.009
(0.033) (0.036)
R-squared (in %) 55.0 55.1 55.0 55.1

The table presents the results of pooled regressions of an individual’s daily excess return on return factors,
and measures of individuals’ fund changes and other characteristics. The return factors are the excess
returns of the Swedish stock market, the Swedish bond market, and the world stock market, and they are
allowed to across the individuals’ characteristics. For brevity, the coefficients on these return factors are
not presented in the table. The measure of fund changes is either a dummy variable for an activity category
(see Table ??) or a variable counting the number of fund changes. Other characteristics are the individuals’
age in 2000, gender, or pension rights in 2000, which is a proxy for income. The constant term and
coefficients on the dummy variables are expressed in % per year. The income variable is scaled down by
1,000. Standard errors, robust to conditional heteroscedasticity and spatial cross-sectional correlations as
in Driscoll and Kraay (1998), are reported in parentheses. The sample consists of 62,640 individuals
225
followed daily over the 2000 to 2010 period.
21 Some Statistics
This section summarizes some useful facts about statistics. Heuristic proofs are given in
a few cases.
Some references: Mittelhammer (1996), DeGroot (1986), Greene (2000), Davidson
(2000), Johnson, Kotz, and Balakrishnan (1994).

21.1 Distributions and Moment Generating Functions

Most of the stochastic variables we encounter in econometrics are continuous. For a

continuous random variable X , the range is uncountably infinite and the probability that
Rx
X  x is Pr.X  x/ D 1 f .q/dq where f .q/ is the continuous probability density
function of X . Note that X is a random variable, x is a number (1.23 or so), and q is just
a dummy argument in the integral.

Fact 21.1 (cdf and pdf) The cumulative distribution function of the random variable X is
Rx
F .x/ D Pr.X  x/ D 1 f .q/dq. Clearly, f .x/ D dF .x/=dx. Note that x is just a
number, not random variable.

Fact 21.2 (Moment generating function of X) The moment generating function of the
random variable X is mgf .t/ D E e tX . The rth moment is the rth derivative of mgf .t/
evaluated at t D 0: E X r D d mgf .0/=dt r . If a moment generating function exists (that
is, E e tX < 1 for some small interval t 2 . h; h/), then it is unique.

Fact 21.3 (Moment generating function of a function of X) If X has the moment generat-
ing function mgfX .t/ D E e tX , then g.X/ has the moment generating function E e tg.X / .
The affine function a C bX (a and b are constants) has the moment generating func-
tion mgfg.X / .t/ D E e t .aCbX / D e t a E e t bX D e t a mgfX .bt/. By setting b D 1 and
a D E X we obtain a mgf for central moments (variance, skewness, kurtosis, etc),
mgf.X E X / .t/ D e t E X mgfX .t/.

Example 21.4 When X  N.;  2 /, then mgfX .t/ D exp t C  2 t 2 =2 . Let Z D

.X /= so a D = and b D 1=. This gives mgfZ .t/ D exp. t=/mgfX .t=/ D

226
exp t 2 =2 . (Of course, this result can also be obtained by directly setting  D 0 and

 D 1 in mgfX .)

Fact 21.5 (Characteristic function and the pdf) The characteristic function of a random
variable x is

g./ D E exp.ix/
R
D x exp.ix/f .x/dx;

where f .x/ is the pdf. This is a Fourier transform of the pdf (if x is a continuous random
variable). The pdf can therefore be recovered by the inverse Fourier transform as
1 R1
f .x/ D exp. ix/g./d:
2 1
In practice, we typically use a fast (discrete) Fourier transform to perform this calcula-
tion, since there are very quick computer algorithms for doing that.

Fact 21.6 The charcteristic function of a N.;  2 / distribution is exp.i  2  2 =2/

j
and of a lognormal(;  2 ) distribuion (where ln x  N.;  2 /) j1D0 .i j/Š exp.j C
P

j 2  2 =2/.

Fact 21.7 (Change of variable, univariate case, monotonic function) Suppose X has the
probability density function fX .c/ and cumulative distribution function FX .c/. Let Y D
g.X / be a continuously differentiable function with dg=dX > 0 (so g.X/ is increasing
for all c such that fX .c/ > 0. Then the cdf of Y is

FY .c/ D PrŒY  c D PrŒg.X/  c D PrŒX  g 1 .c/ D FX Œg 1 .c/;

1
where g is the inverse function of g such that g 1 .Y / D X. We also have that the pdf
of Y is
ˇ dg 1 .c/ ˇ
ˇ ˇ
1
fY .c/ D fX Œg .c/ ˇˇ ˇ:
dc ˇ
If, instead, dg=dX < 0 (so g.X/ is decreasing), then we instead have the cdf of Y

FY .c/ D PrŒY  c D PrŒg.X/  c D PrŒX  g 1 .c/ D 1 FX Œg 1 .c/;

but the same expression for the pdf.

227
 Proof. Differentiate FY .c/, that is, FX Œg 1 .c/ with respect to c.

Example 21.8 Let X  U.0; 1/ and Y D g.X/ D F 1 .X/ where F .c/ is a strictly
increasing cdf. We then get
dF .c/
fY .c/ D :
dc
The variable Y then has the pdf dF .c/=dc and the cdf F .c/. This shows how to gen-
erate random numbers from the F ./ distribution: draw X  U.0; 1/ and calculate
Y D F 1 .X/.

Example 21.9 Let Y D exp.X/, so the inverse function is X D ln Y with derivative

1=Y . Then, fY .c/ D fX .ln c/=c. Conversely, let Y D ln X , so the inverse function is
X D exp.Y / with derivative exp.Y /. Then, fY .c/ D fX Œexp.c/ exp.c/.

Example 21.10 Let X  U.0; 2/, so the pdf and cdf of X are then 1=2 and c=2 respec-
tively. Now, let Y D g.X/ D X gives the pdf and cdf as 1=2 and 1 C y=2 respectively.
The latter is clearly the same as 1 FX Œg 1 .c/ D 1 . c=2/.

Fact 21.11 (Distribution of truncated a random variable) Let the probability distribution
and density functions of X be F .x/ and f .x/, respectively. The corresponding functions,
conditional on a < X  b are ŒF .x/ F .a/=ŒF .b/ F .a/ and f .x/=ŒF .b/ F .a/.
Clearly, outside a < X  b the pdf is zero, while the cdf is zero below a and unity above
b.

21.2 Joint and Conditional Distributions and Moments

21.2.1 Joint and Conditional Distributions

Fact 21.12 (Joint and marginal cdf) Let X and Y be (possibly vectors of) random vari-
ables and let x and y be two numbers. The joint cumulative distribution function of
Rx Ry
X and Y is H.x; y/ D Pr.X  x; Y  y/ D 1 1 h.qx ; qy /dqy dqx , where
2
h.x; y/ D @ F .x; y/=@x@y is the joint probability density function.

Fact 21.13 (Joint and marginal pdf) The marginal cdf of X is obtained by integrating out
R x R 1 
Y : F .x/ D Pr.X  x; Y anything/ D 1 1 h.qx ; qy /dqy dqx . This shows that the
R1
marginal pdf of x is f .x/ D dF .x/=dx D 1 h.qx ; qy /dqy .

228
Fact 21.14 (Conditional distribution) The pdf of Y conditional on X D x (a number) is
g.yjx/ D h.x; y/=f .x/. This is clearly proportional to the joint pdf (at the given value
x).

Fact 21.15 (Change of variable, multivariate case, monotonic function) The result in
Fact 21.7 still holds if X and Y are both n  1 vectors, but the derivative are now
@g 1 .c/=@dc 0 which is an n  n matrix. If gi 1 is the i th function in the vector g 1
then
@g1 1 .c/ @g1 1 .c/
2 3
@c1



@cn
@g 1 .c/ 6 : : 7
D 6 :
: :
: 7:
@dc 0 4 5
@gn 1 .c/ @gn 1 .c/
@c1



@cm

21.2.2 Moments of Joint Distributions

Fact 21.16 (Caucy-Schwartz) .E XY /2  E.X 2 / E.Y 2 /:

Proof. 0  EŒ.aXCY /2  D a2 E.X 2 /C2a E.XY /CE.Y 2 /. Set a D E.XY /= E.X 2 /

to get
ŒE.XY /2 2 ŒE.XY /2
0 2
C E.Y /, that is, 2
 E.Y 2 /:
E.X / E.X /

Fact 21.17 ( 1  Corr.X; y/  1). Let Y and X in Fact 21.16 be zero mean variables
(or variables minus their means). We then get ŒCov.X; Y /2  Var.X/ Var.Y /, that is,
1  Cov.X; Y /=ŒStd.X/Std.Y /  1.

21.2.3 Conditional Moments

R R
Fact 21.18 (Conditional moments) E .Y jx/ D yg.yjx/dy and Var .Y jx/ D Œy
E .Y jx/g.yjx/dy.

Fact 21.19 (Conditional moments as random variables) Before we observe X , the condi-
tional moments are random variables—since X is. We denote these random variables by
E .Y jX /, Var .Y jX/, etc.

Fact 21.20 (Law of iterated expectations) E Y D EŒE .Y jX/. Note that E .Y jX/ is a
random variable since it is a function of the random variable X . It is not a function of Y ,
however. The outer expectation is therefore an expectation with respect to X only.

229
 R R  RR RR
Proof. EŒE .Y jX / D yg.yjx/dy f .x/dx D yg.yjx/f .x/dydx D yh.y; x/dydx D
E Y:

Fact 21.21 (Conditional vs. unconditional variance) Var .Y / D Var ŒE .Y jX /CE ŒVar .Y jX /.

Fact 21.22 (Properties of Conditional Expectations) (a) Y D E .Y jX/CU where U and

E .Y jX / are uncorrelated: Cov .X; Y / D Cov ŒX; E .Y jX/ C U  D Cov ŒX; E .Y jX /.
It follows that (b) CovŒY; E .Y jX / D VarŒE .Y jX/; and (c) Var .Y / D Var ŒE .Y jX/ C
Var .U /. Property (c) is the same as Fact 21.21, where Var .U / D E ŒVar .Y jX /.
RR R R 
Proof. Cov .X; Y / D x.y E y/h.x; y/dydx D x .y E y/g.yjx/dy f .x/dx,
but the term in brackets is E .Y jX / E Y .

Fact 21.23 (Conditional expectation and unconditional orthogonality) E .Y jZ/ D 0 )

E Y Z D 0.

Proof. Note from Fact 21.22 that E.Y jX/ D 0 implies Cov .X; Y / D 0 so E XY D
E X E Y (recall that Cov .X; Y / D E XY E X E Y ). Note also that E .Y jX / D 0 implies
that E Y D 0 (by iterated expectations). We therefore get
" #
Cov .X; Y / D 0
E .Y jX / D 0 ) ) E YX D 0:
EY D 0

21.2.4 Regression Function and Linear Projection

Fact 21.24 (Regression function) Suppose we use information in some variables X to

predict Y . The choice of the forecasting function YO D k.X/ D E .Y jX/ minimizes
EŒY k.X/2 : The conditional expectation E .Y jX / is also called the regression function
of Y on X . See Facts 21.22 and 21.23 for some properties of conditional expectations.

Fact 21.25 (Linear projection) Suppose we want to forecast the scalar Y using the k  1
vector X and that we restrict the forecasting rule to be linear YO D X 0 ˇ. This rule is a
linear projection, denoted P .Y jX/, if ˇ satisfies the orthogonality conditions EŒX.Y
X 0 ˇ/ D 0k1 , that is, if ˇ D .E XX 0 / 1 E XY . A linear projection minimizes EŒY
k.X /2 within the class of linear k.X/ functions.

230
Fact 21.26 (Properties of linear projections) (a) The orthogonality conditions in Fact
21.25 mean that
Y D X 0 ˇ C ";

where E.X"/ D 0k1 . This implies that EŒP .Y jX/" D 0, so the forecast and fore-
cast error are orthogonal. (b) The orthogonality conditions also imply that EŒXY  D
EŒXP .Y jX/. (c) When X contains a constant, so E " D 0, then (a) and (b) carry over to
covariances: CovŒP .Y jX/; " D 0 and CovŒX; Y  D CovŒXP; .Y jX/.

Example 21.27 (P .1jX/) When Y t D 1, then ˇ D .E XX 0 / 1

E X . For instance, sup-
pose X D Œx1t ; x t 2 0 . Then
" # 1 " #
2
E x1t E x1t x2t E x1t
ˇD 2
:
E x2t x1t E x2t E x2t

If x1t D 1 in all periods, then this simplifies to ˇ D Œ1; 00 .

Remark 21.28 Some authors prefer to take the transpose of the forecasting rule, that is,
to use YO D ˇ 0 X . Clearly, since XX 0 is symmetric, we get ˇ 0 D E.YX 0 /.E XX 0 / 1 .

Fact 21.29 (Linear projection with a constant in X) If X contains a constant, then P .aY C
bjX / D aP .Y jX/ C b.

Fact 21.30 (Linear projection versus regression function) Both the linear regression and
the regression function (see Fact 21.24) minimize EŒY k.X/2 , but the linear projection
imposes the restriction that k.X/ is linear, whereas the regression function does not im-
pose any restrictions. In the special case when Y and X have a joint normal distribution,
then the linear projection is the regression function.

Fact 21.31 (Linear projection and OLS) The linear projection is about population mo-
ments, but OLS is its sample analogue.

21.3 Convergence in Probability, Mean Square, and Distribution

Fact 21.32 (Convergence in probability) The sequence of random variables fXT g con-
verges in probability to the random variable X if (and only if) for all " > 0

lim Pr.jXT X j < "/ D 1:

T !1

231
 p
We denote this XT ! X or plim XT D X (X is the probability limit of XT ). Note: (a)
X can be a constant instead of a random variable; (b) if XT and X are matrices, then
p
XT ! X if the previous condition holds for every element in the matrices.

Example 21.33 Suppose XT D 0 with probability .T 1/=T and XT D T with prob-

ability 1=T . Note that limT !1 Pr.jXT 0j D 0/ D limT !1 .T 1/=T D 1, so
limT !1 Pr.jXT 0j D "/ D 1 for any " > 0. Note also that E XT D 0  .T 1/=T C
T  1=T D 1, so XT is biased.

Fact 21.34 (Convergence in mean square) The sequence of random variables fXT g con-
verges in mean square to the random variable X if (and only if)

lim E.XT X/2 D 0:

T !1

m
We denote this XT ! X . Note: (a) X can be a constant instead of a random variable;
m
(b) if XT and X are matrices, then XT ! X if the previous condition holds for every
element in the matrices.

Distribution of sample avg. Distribution of √T × sample avg.

3 0.4
T=5
T=25
T=100 0.3
2
0.2
1
0.1

0 0
−2 −1 0 1 2 −5 0 5
Sample average √T × sample average

Sample average of zt−1 where zt has a χ2(1) distribution

Figure 21.1: Sampling distributions

Fact 21.35 (Convergence in mean square to a constant) If X in Fact 21.34 is a constant,

m
then then XT ! X if (and only if)

lim .E XT X/2 D 0 and lim Var.XT 2 / D 0:

T !1 T !1

232
This means that both the variance and the squared bias go to zero as T ! 1.

Proof. E.XT X/2 D E XT2 2X E XT C X 2 . Add and subtract .E XT /2 and recall

that Var.XT / D E XT2 .E XT /2 . This gives E.XT X/2 D Var.XT / 2X E XT C X 2 C
.E XT /2 D Var.XT / C .E XT X/2 .

Fact 21.36 (Convergence in distribution) Consider the sequence of random variables

fXT g with the associated sequence of cumulative distribution functions fFT g. If limT !1 FT D
F (at all points), then F is the limiting cdf of XT . If there is a random variable X with
d
cdf F , then XT converges in distribution to X : XT ! X . Instead of comparing cdfs, the
comparison can equally well be made in terms of the probability density functions or the
moment generating functions.
m
Fact 21.37 (Relation between the different types of convergence) We have XT ! X )
p d
XT ! X ) XT ! X . The reverse implications are not generally true.

Example 21.38 Consider the random variable in Example 21.33. The expected value is
E XT D 0.T 1/=T C T =T D 1. This means that the squared bias does not go to zero,
so XT does not converge in mean square to zero.

Fact 21.39 (Slutsky’s theorem) If fXT g is a sequence of random matrices such that plim XT D
X and g.XT / a continuous function, then plim g.XT / D g.X/.

Fact 21.40 (Continuous mapping theorem) Let the sequences of random matrices fXT g
d p
and fYT g, and the non-random matrix faT g be such that XT ! X , YT ! Y , and aT ! a
d
(a traditional limit). Let g.XT ; YT ; aT / be a continuous function. Then g.XT ; YT ; aT / !
g.X; Y; a/.

21.4 Laws of Large Numbers and Central Limit Theorems

Fact 21.41 (Khinchine’s theorem) Let X t be independently and identically distributed

p
(iid) with E X t D  < 1. Then ˙ tTD1 X t =T ! .
T
Fact 21.42 (Chebyshev’s theorem) If E X t D 0 and limT !1 Var.˙ tD1 X t =T / D 0, then
p
T
˙ t D1 X t =T ! 0.

Fact 21.43 (The Lindeberg-Lévy theorem) Let X t be independently and identically dis-
d
tributed (iid) with E X t D 0 and Var.X t / < 1. Then p1T ˙ tTD1 X t = ! N.0; 1/.

233
21.5 Stationarity

Fact 21.44 (Covariance stationarity) X t is covariance stationary if

E X t D  is independent of t;
Cov .X t s ; X t / D s depends only on s, and
both  and s are finite.

Fact 21.45 (Strict stationarity) X t is strictly stationary if, for all s, the joint distribution
of X t ; X t C1 ; :::; X tCs does not depend on t.

Fact 21.46 (Strict stationarity versus covariance stationarity) In general, strict station-
arity does not imply covariance stationarity or vice versa. However, strict stationary with
finite first two moments implies covariance stationarity.

21.6 Martingales

Fact 21.47 (Martingale) Let ˝ t be a set of information in t , for instance Y t ; Y t 1 ; ::: If

E jY t j < 1 and E.Y t C1 j˝ t / D Y t , then Y t is a martingale.

Fact 21.48 (Martingale difference) If Y t is a martingale, then X t D Y t Yt 1 is a mar-

tingale difference: X t has E jX t j < 1 and E.X t C1 j˝ t / D 0.

Fact 21.49 (Innovations as a martingale difference sequence) The forecast error X t C1 D

Y t C1 E.Y t C1 j˝ t / is a martingale difference.

Fact 21.50 (Properties of martingales) (a) If Y t is a martingale, then E.Y tCs j˝ t / D Y t

for s  1. (b) If X t is a martingale difference, then E.X t Cs j˝ t / D 0 for s  1.

Proof. (a) Note that E.Y t C2 j˝ t C1 / D Y t C1 and take expectations conditional on ˝ t :

EŒE.Y t C2 j˝ t C1 /j˝ t  D E.Y t C1 j˝ t / D Y t . By iterated expectations, the first term equals
E.Y t C2 j˝ t /. Repeat this for t C 3, t C 4, etc. (b) Essentially the same proof.

Fact 21.51 (Properties of martingale differences) If X t is a martingale difference and

g t 1 is a function of ˝ t 1 , then X t g t 1 is also a martingale difference.

234
 Proof. E.X tC1 g t j˝ t / D E.X t C1 j˝ t /g t since g t is a function of ˝ t .

Fact 21.52 (Martingales, serial independence, and no autocorrelation) (a) X t is serially

uncorrelated if Cov.X t ; X t Cs / D 0 for all s ¤ 0. This means that a linear projection of
X t Cs on X t ; X t 1;::: is a constant, so it cannot help predict X t Cs . (b) X t is a martingale
difference with respect to its history if E.X t Cs jX t ; X t 1 ; :::/ D 0 for all s  1. This means
that no function of X t ; X t 1 ; ::: can help predict X t Cs . (c) X t is serially independent if
pdf.X t Cs jX t ; X t 1 ; :::/ D pdf.X tCs /. This means than no function of X t ; X t 1 ; ::: can
help predict any function of X t Cs .

Fact 21.53 (WLN for martingale difference) If X t is a martingale difference, then plim ˙ tTD1 X t =T D
0 if either (a) X t is strictly stationary and E jx t j < 0 or (b) E jx t j1Cı < 1 for ı > 0 and
all t. (See Davidson (2000) 6.2)

Fact 21.54 (CLT for martingale difference) Let X t be a martingale difference. If plim ˙ tTD1 .X t2
E X t2 /=T D 0 and either

(a) X t is strictly stationary or

.E jX t j2Cı /1=.2Cı/
(b) max t2Œ1;T  T
˙ tD1 E X t2 =T
< 1 for ı > 0 and all T > 1;

p d
then .˙ tTD1 X t = T /=.˙ tTD1 E X t2 =T /1=2 ! N.0; 1/. (See Davidson (2000) 6.2)

21.7 Special Distributions

21.7.1 The Normal Distribution

Fact 21.55 (Univariate normal distribution) If X  N.;  2 /, then the probability den-
sity function of X, f .x/ is
1 1 x  2
f .x/ D p e 2.  / :
2 2

The moment generating function is mgfX .t/ D exp t C  2 t 2 =2 and the moment gen-

erating function around the mean is mgf.X / .t/ D exp  2 t 2 =2 .

Example 21.56 The first few moments around the mean are E.X / D 0, E.X /2 D
 2 , E.X /3 D 0 (all odd moments are zero), E.X /4 D 3 4 , E.X /6 D 15 6 ,
and E.X /8 D 105 8 .

235
 Pdf of N(0,1) Pdf of bivariate normal, corr=0.1

0.4 0.2

0.1
0.2
2
0
2 0
0 0
−2 −1 0 1 2 y −2 −2 x
x

Pdf of bivariate normal, corr=0.8

0.4

0.2
2
0
2 0
0
y −2 −2 x
Figure 21.2: Normal distributions

Fact 21.57 (Standard normal distribution) If X  N.0; 1/, then the moment generating
function is mgfX .t/ D exp t 2 =2 . Since the mean is zero, m.t/ gives central moments.

The first few are E X D 0, E X 2 D 1, E X 3 D 0 (all odd moments are zero), and
p
EX 4 D 3. The distribution function, Pr.X  a/ D ˚.a/ D 1=2 C 1=2 erf.a= 2/,
Rz
where erf./ is the error function, erf.z/ D p2 0 exp. t 2 /dt). The complementary error
function is erfc.z/ D 1 erf.z/. Since the distribution is symmetric around zero, we have
˚. a/ D Pr.X  a/ D Pr.X  a/ D 1 ˚.a/. Clearly, 1 ˚. a/ D ˚.a/ D
p
1=2 erfc. a= 2/.

Fact 21.58 (Multivariate normal distribution) If X is an n1 vector of random variables

with a multivariate normal distribution, with a mean vector  and variance-covariance
matrix ˙, N.; ˙/, then the density function is
 
1 1 0 1
f .x/ D exp .x / ˙ .x / :
.2/n=2 j˙j1=2 2

236
 Pdf of bivariate normal, corr=0.1 Conditional pdf of y, corr=0.1
0.8
x=−0.8
0.6 x=0
0.2
0.4
0.1
2 0.2
0
2 0
0 0
y −2 −2 x −2 −1 0 1 2
y

Pdf of bivariate normal, corr=0.8 Conditional pdf of y, corr=0.8

0.8
x=−0.8
0.6 x=0
0.4
0.4
0.2
2 0.2
0
2 0
0 0
y −2 −2 x −2 −1 0 1 2
y

Figure 21.3: Density functions of normal distributions

Fact 21.59 (Conditional normal distribution) Suppose Zm1 and Xn1 are jointly nor-
mally distributed
" # " # " #!
Z Z ˙ZZ ˙ZX
N ; :
X X ˙XZ ˙XX

The distribution of the random variable Z conditional on that X D x (a number) is also

normal with mean
E .Zjx/ D Z C ˙ZX ˙XX1 .x X / ;

and variance (variance of Z conditional on that X D x, that is, the variance of the
prediction error Z E .Zjx/)

Var .Zjx/ D ˙ZZ ˙ZX ˙XX1 ˙XZ :

237
Note that the conditional variance is constant in the multivariate normal distribution
(Var .ZjX / is not a random variable in this case). Note also that Var .Zjx/ is less than
Var.Z/ D ˙ZZ (in a matrix sense) if X contains any relevant information (so ˙ZX is
not zero, that is, E .Zjx/ is not the same for all x).

Example 21.60 (Conditional normal distribution) Suppose Z and X are scalars in Fact
21.59 and that the joint distribution is
" # " # " #!
Z 3 1 2
N ; :
X 5 2 6

The expectation of Z conditional on X D x is then

2 1
E .Zjx/ D 3 C .x 5/ D 3 C .x 5/ :
6 3
Similarly, the conditional variance is
22 1
Var .Zjx/ D 1 D :
6 3
Fact 21.61 (Stein’s lemma) If Y has normal distribution and h./ is a differentiable func-
tion such that E jh0 .Y /j < 1, then CovŒY; h.Y / D Var.Y / E h0 .Y /.
R1
Proof. EŒ.Y /h.Y / D 1 .Y /h.Y /.Y I ;  2 /d Y , where .Y I ;  2 / is the
pdf of N.;  2 /. Note that d.Y I ;  2 /=d Y D .Y I ;  2 /.Y /= 2 , so the integral
R1
can be rewritten as  2 1 h.Y /d.Y I ;  2 /. Integration by parts (“ udv D uv
R
ˇ1 R1
vdu”) gives  2 h.Y /.Y I ;  2 /ˇ 1 2 0 2 0
R  
1 .Y I ;  /h .Y /d Y D  E h .Y /.

Fact 21.62 (Stein’s lemma 2) It follows from Fact 21.61 that if X and Y have a bivariate
normal distribution and h./ is a differentiable function such that E jh0 .Y /j < 1, then
CovŒX; h.Y / D Cov.X; Y / E h0 .Y /.

Example 21.63 (a) With h.Y / D exp.Y / we get CovŒX; exp.Y / D Cov.X; Y / E exp.Y /;
(b) with h.Y / D Y 2 we get CovŒX; Y 2  D Cov.X; Y /2 E Y so with E Y D 0 we get a
zero covariance.

Fact 21.64 (Stein’s lemma 3) Fact 21.62 still holds if the joint distribution of X and Y is
a mixture of n bivariate normal distributions, provided the mean and variance of Y is the
same in each of the n components. (See Söderlind (2009) for a proof.)

238
Fact 21.65 (Truncated normal distribution) Let X  N.;  2 /, and consider truncating
the distribution so that we want moments conditional on a < X  b. Define a0 D
.a /= and b0 D .b /= . Then,

.b0 / .a0 /
E.Xja < X  b/ D   and
˚.b0 / ˚.a0 /
(  2 )
b0 .b0 / a0 .a0 / .b0 / .a0 /
Var.Xja < X  b/ D  2 1 :
˚.b0 / ˚.a0 / ˚.b0 / ˚.a0 /

Fact 21.66 (Lower truncation) In Fact 21.65, let b ! 1, so we only have the truncation
a < X. Then, we have
.a0 /
E.Xja < X/ D  C  and
1 ˚.a0 /
(  2 )
2 a0 .a0 / .a0 /
Var.Xja < X/ D  1C :
1 ˚.a0 / 1 ˚.a0 /

(The latter follows from limb!1 b0 .b0 / D 0.)

Example 21.67 Suppose X  N.0;  2 / and we want to calculate E jxj. This is the same
as E.XjX > 0/ D 2.0/.

Fact 21.68 (Upper truncation) In Fact 21.65, let a ! 1, so we only have the trunca-
tion X  b. Then, we have
.b0 /
E.XjX  b/ D   and
˚.b0 /
(  2 )
b0 .b0 / .b0 /
Var.XjX  b/ D  2 1 :
˚.b0 / ˚.b0 /

(The latter follows from lima! 1 a0 .a0 / D 0.)

Fact 21.69 (Delta method) Consider an estimator ˇOk1 which satisfies

p   d
T ˇO ˇ0 ! N .0; ˝/ ;

and suppose we want the asymptotic distribution of a transformation of ˇ

q1 D g .ˇ/ ;

239
where g .:/ is has continuous first derivatives. The result is
p h   i d
T g ˇO


g .ˇ0 / ! N 0; qq ; where
@g .ˇ0 / @g .ˇ0 /0 @g .ˇ0 /
D 0 ˝ , where is q  k:
@ˇ @ˇ @ˇ 0

Proof. By the mean value theorem we have

  @g .ˇ  /  O 
g ˇO D g .ˇ0 / C ˇ ˇ0 ;
@ˇ 0
where 2 3
@g1 .ˇ / @g1 .ˇ /
@ˇ1



@ˇk
@g .ˇ/ 6 :: :: :: 7
D6 : : : 7 ;
@ˇ 0 4
@gq .ˇ / @gq .ˇ /
5
@ˇ1



@ˇk qk
p
and we evaluate it at ˇ  which is (weakly) between ˇO and ˇ0 . Premultiply by T and
rearrange as
p h   i @g .ˇ  / p  
T g ˇO g .ˇ0 / D T ˇO ˇ0 .
@ˇ 0
If ˇO is consistent (plim ˇO D ˇ0 ) and @g .ˇ  / =@ˇ 0 is continuous, then by Slutsky’s theorem
plim @g .ˇ  / =@ˇ 0 D @g .ˇ0 / =@ˇ 0 , which is a constant. The result then follows from the
continuous mapping theorem.

21.7.2 The Lognormal Distribution

Fact 21.70 (Univariate lognormal distribution) If x  N.;  2 / and y D exp.x/ then

the probability density function of y, f .y/ is
1 1 ln y  2
f .y/ D p e 2.  /
, y > 0:
y 2 2

The rth moment of y is E y r D exp.r C r 2  2 =2/. See 21.4 for an illustration.

Example 21.71 The first two moments are E y D exp  C  2 =2 and E y 2 D exp.2 C

2 2 /. We therefore get Var.y/ D exp 2 C  2 exp  2



1 and Std .y/ = E y D
p
exp. 2 / 1.

240
 Normal distributions Lognormal distributions

N(0,1)
0.6 0.6
N(0,0.5)
0.4 0.4

0.2 0.2

0 0
−4 −2 0 2 4 −4 −2 0 2 4
lnx x

Figure 21.4: Lognormal distribution

Fact 21.72 (Moments of a truncated lognormal distribution) If x  N.;  2 / and y D

exp.x/ then E.y r jy > a/ D E.y r /˚.r a0 /=˚. a0 /, where a0 D .ln a / =.
Note that the denominator is Pr.y > a/ D ˚. a0 /. In contrast, E.y r jy  b/ D
E.y r /˚. r C b0 /=˚.b0 /, where b0 D .ln b / =. The denominator is Pr.y  b/ D
˚.b0 /. Clearly, E.y r / D exp.r C r 2  2 =2/

Fact 21.73 (Moments of a truncated lognormal distribution, two-sided truncation) If x 

N.;  2 / and y D exp.x/ then

˚.r a0 / ˚.r b0 /
E.y r ja > y < b/ D E.y r / ;
˚.b0 / ˚.a0 /

where a0 D .ln a / = and b0 D .ln b / =. Note that the denominator is Pr.a >
y < b/ D ˚.b0 / ˚.a0 /. Clearly, E.y r / D exp.r C r 2  2 =2/.

Example 21.74 The first two moments of the truncated (from below) lognormal distri-
bution are E.yjy > a/ D exp  C  2 =2 ˚. a0 /=˚. a0 / and E.y 2 jy > a/ D

exp 2 C 2 2 ˚.2 a0 /=˚. a0 /.

Example 21.75 The first two moments of the truncated (from above) lognormal distri-
bution are E.yjy  b/ D exp  C  2 =2 ˚.  C b0 /=˚.b0 / and E.y 2 jy  b/ D

exp 2 C 2 2 ˚. 2 C b0 /=˚.b0 /.

Fact 21.76 (Multivariate lognormal distribution) Let the n  1 vector x have a mulivari-

241
ate normal distribution
2
3 2 3
1 11


1n
6 : 7
:: 7 and ˙ D 6 ::: : : : :: 7
6
x  N.; ˙/, where  D 6
4 5 4 5:
: 7
n n1


nn

Then y D exp.x/ has a lognormal distribution, with the means and covariances

E yi D exp .i C i i =2/

  
Cov.yi ; yj / D exp i C j C .i i C jj /=2 exp.ij / 1
  q  
Corr.yi ; yj / D exp.ij / 1 = Œexp.i i / 1 exp.jj / 1 :

Cleary, Var.yi / D exp Œ2i C i i  Œexp.i i / 1. Cov.y1 ; y2 / and Corr.y1 ; y2 / have the
same sign as Corr.xi ; xj / and are increasing in it. However, Corr.yi ; yj / is closer to zero.

21.7.3 The Chi-Square Distribution

Fact 21.77 (The 2n distribution) If Y  2n , then the pdf of Y is f .y/ D 2n=2 1.n=2/ y n=2 1 e y=2
,
where ./ is the gamma function. The moment generating function is mgfY .t/ D .1
2t / n=2 for t < 1=2. The first moments of Y are E Y D n and Var.Y / D 2n.

Fact 21.78 (Quadratic forms of normally distribution random variables) If the n  1

vector X  N.0; ˙/, then Y D X 0 ˙ 1 X  2n . Therefore, if the n scalar random
variables Xi , i D 1; :::; n, are uncorrelated and have the distributions N.0; i2 /, i D
1; :::; n, then Y D ˙inD1 Xi2 =i2  2n .

Fact 21.79 (Distribution of X 0 AX) If the n1 vector X  N.0; I /, and A is a symmetric
idempotent matrix (A D A0 and A D AA D A0 A) of rank r, then Y D X 0 AX  2r .

Fact 21.80 (Distribution of X 0 ˙ C X) If the n  1 vector X  N.0; ˙/, where ˙ has

rank r  n then Y D X 0 ˙ C X  2r where ˙ C is the pseudo inverse of ˙ .

Proof. ˙ is symmetric, so it can be decomposed as ˙ D CC 0 where C are the

orthogonal eigenvector (C 0 C D I ) and  is a diagonal matrix with the eigenvalues along
the main diagonal. We therefore have ˙ D CC 0 D C1 11 C10 where C1 is an n  r

242
matrix associated with the r non-zero eigenvalues (found in the r  r matrix 11 ). The
generalized inverse can be shown to be
" #
h i  1 0 h i0
C 11
˙ D C1 C2 C1 C2 D C1 111 C10 ;
0 0

We can write ˙ C D C1 111=2 111=2 C10 . Consider the r  1 vector Z D 111=2 C10 X, and
note that it has the covariance matrix

E ZZ 0 D 111=2 C10 E XX 0 C1 111=2 D 111=2 C10 C1 11 C10 C1 111=2 D Ir ;

since C10 C1 D Ir . This shows that Z  N.0r1 ; Ir /, so Z 0 Z D X 0 ˙ C X  2r .

Fact 21.81 (Convergence to a normal distribution) Let Y  2n and Z D .Y n/=n1=2 .

d
Then Z ! N.0; 2/.

Example 21.82 If Y D ˙inD1 Xi2 =i2 , then this transformation means Z D .˙inD1 Xi2 =i2
1/=n1=2 .

Proof. We can directly note from the moments of a 2n variable that E Z D .E Y
n/=n1=2 D 0, and Var.Z/ D Var.Y /=n D 2. From the general properties of moment
generating functions, we note that the moment generating function of Z is

p
  n=2
t n t
mgfZ .t/ D e 1 2 with lim mgfZ .t/ D exp.t 2 /:
n1=2 n!1

d
This is the moment generating function of a N.0; 2/ distribution, which shows that Z !
N.0; 2/. This result should not come as a surprise as we can think of Y as the sum of
n variables; dividing by n1=2 is then like creating a scaled sample average for which a
central limit theorem applies.

21.7.4 The t and F Distributions

Fact 21.83 (The F .n1 ; n2 / distribution) If Y1  2n1 and Y2  2n2 and Y1 and Y2 are
independent, then Z D .Y1 =n1 /=.Y2 =n2 / has an F .n1 ; n2 / distribution. This distribution
has no moment generating function, but E Z D n2 =.n2 2/ for n > 2.

243
 a. Pdf of Chi−square(n) b. Pdf of F(n1,10)
1 1
n=1 n1=2
n=2 n1=5
n=5
0.5 0.5 n1=10
n=10

0 0
0 5 10 0 2 4 6
x x

c. Pdf of F(n1,100) d. Pdf of N(0,1) and t(n)

1 0.4
n1=2 N(0,1)
n1=5 t(10)
t(50)
0.5 n1=10 0.2

0 0
0 2 4 6 −2 0 2
x x
Figure 21.5: 2 , F, and t distributions

Fact 21.84 (Convergence of an F .n1 ; n2 / distribution) In Fact (21.83), the distribution of

n1 Z D Y1 =.Y2 =n2 / converges to a 2n1 distribution as n2 ! 1. (The idea is essentially
that n2 ! 1 the denominator converges to the mean, which is E Y2 =n2 D 1. Only the
numerator is then left, which is a 2n1 variable.)

Fact 21.85 (The tn distribution) If X  N.0; 1/ and Y  2n and X and Y are indepen-
dent, then Z D X=.Y =n/1=2 has a tn distribution. The moment generating function does
not exist, but E Z D 0 for n > 1 and Var.Z/ D n=.n 2/ for n > 2.

Fact 21.86 (Convergence of a tn distribution) The t distribution converges to a N.0; 1/

distribution as n ! 1.

Fact 21.87 (tn versus F .1; n/ distribution) If Z  tn , then Z 2  F .1; n/.

244
21.7.5 The Bernouilli and Binomial Distributions

Fact 21.88 (Bernoulli distribution) The random variable X can only take two values:
1 or 0, with probability p and 1 p respectively. The moment generating function is
mgf .t / D pe t C 1 p. This gives E.X/ D p and Var.X/ D p.1 p/.

Example 21.89 (Shifted Bernoulli distribution) Suppose the Bernoulli variable takes the
values a or b (instead of 1 and 0) with probability p and 1 p respectively. Then E.X/ D
pa C .1 p/b and Var.X/ D p.1 p/.a b/2 .

Fact 21.90 (Binomial distribution). Suppose X1 ; X2 ; :::; Xn all have Bernoulli distribu-
tions with the parameter p. Then, the sum Y D X1 C X2 C ::: C Xn has a Binomial
distribution with parameters p and n. The pdf is pdf.Y / D nŠ=ŒyŠ.n y/Šp y .1 p/n y
for y D 0; 1; :::; n. The moment generating function is mgf .t/ D Œpe t C 1 pn . This
gives E.Y / D np and Var.Y / D np.1 p/.

Example 21.91 (Shifted Binomial distribution) Suppose the Bernuolli variables X1 ; X2 ; :::; Xn
take the values a or b (instead of 1 and 0) with probability p and 1 p respectively.
Then, the sum Y D X1 C X2 C ::: C Xn has E.Y / D nŒpa C .1 p/b and Var.Y / D
nŒp.1 p/.a b/2 .

21.7.6 The Skew-Normal Distribution

Fact 21.92 (Skew-normal distribution) Let  and ˚ be the standard normal pdf and cdf
respectively. The pdf of a skew-normal distribution with shape parameter ˛ is then

f .z/ D 2.z/˚.˛z/:

If Z has the above pdf and

Y D  C !Z with ! > 0;

then Y is said to have a SN.; ! 2 ; ˛/ distribution (see Azzalini (2005)). Clearly, the pdf
of Y is
f .y/ D 2 Œ.y / =! ˚ Œ˛ .y / =! =!:

The moment generating function is mgfy .t/ D 2 exp t C ! 2 t 2 =2 ˚.ı!t/ where ı D

p
˛= 1 C ˛ 2 . When ˛ > 0 then the distribution is positively skewed (and vice versa)—and

245
when ˛ D 0 the distribution becomes a normal distribution. When ˛ ! 1, then the
density function is zero for Y  , and 2 Œ.y / =! =! otherwise—this is a half-
normal distribution.
p
Example 21.93 The first three moments are as follows. First, notice that E Z D 2=ı,
p
Var.Z/ D 1 2ı 2 = and E.Z E Z/3 D .4= 1/ 2=ı 3 . Then we have

EY D  C ! EZ
Var.Y / D ! 2 Var.Z/
E .Y E Y /3 D ! 3 E.Z E Z/3 :

Notice that with ˛ D 0 (so ı D 0), then these moments of Y become , ! 2 and 0
respecively.

21.7.7 Generalized Pareto Distribution

Fact 21.94 (Cdf and pdf of the generalized Pareto distribution) The generalized Pareto
distribution is described by a scale parameter (ˇ > 0) and a shape parameter (). The
cdf (Pr.Z  z/, where Z is the random variable and z is a value) is
(
1 .1 C z=ˇ/ 1= if  ¤ 0
G.z/ D
1 exp. z=ˇ/  D 0;

for 0  z and z  ˇ= in case  < 0. The pdf is therefore

(
1
ˇ
.1 C z=ˇ/ 1= 1 if  ¤ 0
g.z/ D 1
ˇ
exp. z=ˇ/  D 0:

The mean is defined (finite) if  < 1 and is then E.z/ D ˇ=.1 /, the median is
.2 1/ˇ= and the variance is defined if  < 1=2 and is then ˇ 2 =Œ.1 /2 .1 2/.

21.8 Inference

Fact 21.95 (Comparing variance-covariance matrices) Let Var.ˇ/ O and Var.ˇ  / be the
variance-covariance matrices of two estimators, ˇO and ˇ  , and suppose Var.ˇ/ O Var.ˇ  /
O 
is a positive semi-definite matrix. This means that for any non-zero vector R that R0 Var.ˇ/R

246
R0 Var.ˇ  /R, so every linear combination of ˇO has a variance that is as large as the vari-
ance of the same linear combination of ˇ  . In particular, this means that the variance of
every element in ˇO (the diagonal elements of Var.ˇ/)O is at least as large as variance of
the corresponding element of ˇ  .

Bibliography
Azzalini, A., 2005, “The skew-normal distribution and related Multivariate Families,”
Scandinavian Journal of Statistics, 32, 159–188.

Davidson, J., 2000, Econometric theory, Blackwell Publishers, Oxford.

DeGroot, M. H., 1986, Probability and statistics, Addison-Wesley, Reading, Mas-

sachusetts.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

Johnson, N. L., S. Kotz, and N. Balakrishnan, 1994, Continuous univariate distributions,

Wiley, New York, 2nd edn.

Mittelhammer, R. C., 1996, Mathematical statistics for economics and business, Springer-
Verlag, New York.

Söderlind, P., 2009, “An extended Stein’s lemma for asset pricing,” Applied Economics
Letters, forthcoming, 16, 1005–1008.

247
22 Some Facts about Matrices
Some references: Greene (2000), Golub and van Loan (1989), Björk (1996), Anton
(1987), Greenberg (1988).

22.1 Rank

Fact 22.1 (Submatrix) Any matrix obtained from the m  n matrix A by deleting at most
m 1 rows and at most n 1 columns is a submatrix of A.

Fact 22.2 (Rank) The rank of the m  n matrix A is  if the largest submatrix with non-
zero determinant is   . The number of linearly independent row vectors (and column
vectors) of A is then .

22.2 Vector Norms

Fact 22.3 (Vector p-norm) Let x be an n  1 matrix. The p-norm is defined as/

n
!1=p
X
kxkp D jxi jp :
i D1

The Euclidian norm corresponds to p D 2

n
!1=2
X p
kxk2 D xi2 D x 0 x:
i D1

22.3 Systems of Linear Equations and Matrix Inverses

Fact 22.4 (Linear systems of equations) Consider the linear system Ax D c where A is
m  n, x is n  1, and c is m  1. A solution is a vector x such that Ax D c. It has
a unique solution if and only if rank.A/ D rank.Œ A c / D n; an infinite number of
solutions if and only if rank.A/ D rank.Œ A c / < n; and no solution if and only if
rank.A/ ¤ rank.Œ A c /.

248
Example 22.5 (Linear systems of equations, unique solution when m D n) Let x be 21,
and consider the linear system
" # " #
1 5 3
Ax D c with A D and c D :
2 6 6

Here rank .A/ D 2 and rank.Œ A c / D 2. The unique solution is x D Œ 3 0 0 :

Example 22.6 (Linear systems of equations, no solution when m > n) Let x be a scalar,
and consider the linear system
" # " #
1 3
Ax D c with A D and c D :
2 7

Here rank .A/ D 1 and rank.Œ A c / D 2. There is then no solution.

Example 22.7 (Inverse of 2  2 matrices). For a 2  2 matrix we have
" # 1 " #
a b 1 d b
D :
c d ad bc c a

In particular, for a triangular matrix we have

" # 1 " #
a 0 1=a 0
D :
c d c=.ad / 1=d
Fact 22.8 (Least squares) Suppose that no solution exists to Ax D c. The best approxi-
mate solution, in the sense of minimizing (the square root of) the sum of squared errors,
1=2
O 0 .c Ax/ O 2 , is xO D .A0 A/ 1 A0 c, provided the inverse exist.

.c Ax/ O D kc Axk
This is obviously the least squares solution. In the example with c D Œ 3 7 0 , it is
0" #0 " #1 1 " #0 " #
1 1 1 3
xO D @ A
2 2 2 7
17
D or 3:4:
5
This is illustrated in Figure 22.1. (Translation to OLS notation: c is the vector of depen-
dent variables for m observations, A is the matrix with explanatory variables with the t t h
observation in row t , and x is the vector of parameters to estimate).

249
 Sum of squared errors of solution to A*x=c, scalar x
5
With uniqe solution
4 No solutions

3 A=[1 2]′ and c=[3,7]′ A=[1 2]′ and c=[3,6]′

0
2.5 3 3.5 4
x

Figure 22.1: Value of qudratic loss function.

Fact 22.9 (Pseudo inverse or generalized inverse) Suppose that no solution exists to
Ax D c, and that A0 A is not invertible. There are then several approximations, x,
O which
all minimize kc Axk O 2 is given by xO D A c, where AC
O 2 . The one with the smallest kxk C

is the Moore-Penrose pseudo (generalized) inverse of A. See Fact 22.56.

Example 22.10 (Linear systems of equations, unique solution when m > n) Change c
in Example 22.6 to c D Œ 3 6 0 . Then rank .A/ D 1 and rank.Œ A c / D 1, and the
unique solution is x D 3:

Example 22.11 (Linear systems of equations, infinite number of solutions, m < n) Let x
be 2  1, and consider the linear system
h i
Ax D c with A D 1 2 and c D 5:

Here rank .A/ D 1 and rank.Œ A c / D 1. Any value of x1 on the line 5 2x2 is a
solution.

Example 22.12 (Pseudo inverses again) In the previous example, there is an infinite
number of solutions along the line x1 D 5 2x2 . Which one has the smallest norm

250
 O 2 D Œ.5 2x2 /2 C x22 1=2 ? The first order condition gives x2 D 2, and therefore
kxk
x1 D 1. This is the same value as given by xO D AC c, since AC D Œ0:2; 0:4 in this case.

Fact 22.13 (Rank and computers) Numerical calculations of the determinant are poor
indicators of whether a matrix is singular or not. For instance, det.0:1  I20 / D 10 20 .
Use the condition number instead (see Fact 22.53).

1
Fact 22.14 (Some properties of inverses) If A, B, and C are invertible, then .ABC / D
1
C 1 B 1 A 1 ; .A 1 /0 D .A0 / 1 ; if A is symmetric, then A 1 is symmetric; .An / D

n
A 1 .

Fact 22.15 (Changing sign of column and inverting) Suppose the square matrix A2 is the
same as A1 except that the i t h and j t h columns have the reverse signs. Then A2 1 is the
same as A1 1 except that the i t h and j t h rows have the reverse sign.

22.4 Complex matrices

p
Fact 22.16 (Modulus of complex number) If  D a C bi, where i D 1, then jj D
p
ja C bi j D a2 C b 2 .

Fact 22.17 (Complex matrices) Let AH denote the transpose of the complex conjugate of
A, so that if " #
h i 1
A D 1 2 C 3i then AH D :
2 3i
A square matrix A is unitary (similar to orthogonal) if AH D A 1 , for instance,
" # " #
1Ci 1Ci 1 i 1Ci
2 2 H 1 2 2
AD 1 i 1Ci
gives A DA D 1 i 1 i
:
2 2 2 2

and it Hermitian (similar to symmetric) if A D AH , for instance

" #
1 1Ci
2 2
AD 1 i 1
:
2 2

A Hermitian matrix has real elements along the principal diagonal and Aj i is the complex
conjugate of Aij . Moreover, the quadratic form x H Ax is always a real number.

251
22.5 Eigenvalues and Eigenvectors

Fact 22.18 (Homogeneous linear system). Consider the linear system in Fact 22.4 with
c D 0: Amn xn1 D 0m1 . Then rank.A/ D rank.Œ A c /, so it has a unique solution
if and only if rank.A/ D n; and an infinite number of solutions if and only if rank.A/ < n.
Note that x D 0 is always a solution, and it is the unique solution if rank.A/ D n. We
can thus only get a nontrivial solution (not all elements are zero), only if rank .A/ < n.

Fact 22.19 (Eigenvalues) The n eigenvalues, i , i D 1; : : : ; n, and associated eigenvec-

tors, zi , of the n  n matrix A satisfy

.A i I / zi D 0n1 :

We require the eigenvectors to be non-trivial (not all elements are zero). From Fact 22.18,
an eigenvalue must therefore satisfy

det.A i I / D 0:

Fact 22.20 (Right and left eigenvectors) A “right eigenvector” z (the most common) sat-
isfies Az D z, and a “left eigenvector” v (seldom used) satisfies v 0 A D v 0 , that is,
A0 v D v.

Fact 22.21 (Rank and eigenvalues) For any m  n matrix A, rank .A/ D rank .A0 / D
rank .A0 A/ D rank .AA0 / and equals the number of non-zero eigenvalues of A0 A or AA0 .

Example 22.22 Let x be an n  1 vector, so rank .x/ D 1. We then have that the outer
product, xx 0 also has rank 1.

Fact 22.23 (Determinant and eigenvalues) For any n  n matrix A, det.A/ D ˘inD1 i .

22.6 Special Forms of Matrices

22.6.1 Triangular Matrices

Fact 22.24 (Triangular matrix) A lower (upper) triangular matrix has zero elements
above (below) the main diagonal.

252
Fact 22.25 (Eigenvalues of triangular matrix) For a triangular matrix A, the eigenvalues
equal the diagonal elements of A. This follows from that

det.A I / D .A11 / .A22 / : : : .Ann / :

Fact 22.26 (Squares of triangular matrices) If T is lower (upper) triangular, then T T is

as well.

22.6.2 Orthogonal Vector and Matrices

Fact 22.27 (Orthogonal vector) The n  1 vectors x and y are orthogonal if x 0 y D 0.

Fact 22.28 (Orthogonal matrix) The nn matrix A is orthogonal if A0 A D I . Properties:

If A is orthogonal, then det .A/ D ˙1; if A and B are orthogonal, then AB is orthogonal.

Example 22.29 (Rotation of vectors (“Givens rotations”).) Consider the matrix G D In

except that Gi k D c, Gi k D s, Gki D s, and Gkk D c. If we let c D cos  and
s D sin  for some angle , then G 0 G D I . To see this, consider the simple example
where i D 2 and k D 3
2 30 2 3 2 3
1 0 0 1 0 0 1 0 0
4 0 c s 5 4 0 c s 5 D 4 0 c2 C s2 0 5;
6 7 6 7 6 7

0 s c 0 s c 0 0 c2 C s2

which is an identity matrix since cos2  C sin2  D 1. G is thus an orthogonal matrix. It

is often used to “rotate” an n  1 vector " as in u D G 0 ", where we get

u t D " t for t ¤ i; k
ui D " i c "k s
uk D "i s C "k c:

The effect of this transformation is to rotate the i t h and k t h vectors counterclockwise

through an angle of .

22.6.3 Positive Definite Matrices

Fact 22.30 (Positive definite matrix) The n  n matrix A is positive definite if for any
non-zero n  1 vector x, x 0 Ax > 0. (It is positive semidefinite if x 0 Ax  0.)

253
Fact 22.31 (Some properties of positive definite matrices) If A is positive definite, then
all eigenvalues are positive and real. (To see why, note that an eigenvalue satisfies Ax D
x. Premultiply by x 0 to get x 0 Ax D x 0 x. Since both x 0 Ax and x 0 x are positive real
numbers,  must also be.)

Fact 22.32 (More properties of positive definite matrices) If B is a r  n matrix of rank r

and A is a n  n positive definite matrix, then BAB 0 is also positive definite and has rank
r. For instance, B could be an invertible n  n matrix. If A D In , then we have that BB 0
is positive definite.

Fact 22.33 (More properties of positive definite matrices) If A is positive definite, then
det .A/ > 0 and all diagional elements are positive; if A is positive definite, then A 1 is
too.

Fact 22.34 (Cholesky decomposition) See Fact 22.42.

22.6.4 Symmetric Matrices

Fact 22.35 (Symmetric matrix) A is symmetric if A D A0 .

Fact 22.36 (Properties of symmetric matrices) If A is symmetric, then all eigenvalues are
real, and eigenvectors corresponding to distinct eigenvalues are orthogonal.
1
Fact 22.37 If A is symmetric, then A is symmetric.

22.6.5 Idempotent Matrices

Fact 22.38 (Idempotent matrix) A is idempotent if A D AA. If A is also symmetric, then

A D A0 A.

22.7 Matrix Decompositions

Fact 22.39 (Diagonal decomposition) An n  n matrix A is diagonalizable if there exists

a matrix C such that C 1 AC D  is diagonal. We can thus write A D CC 1 . The
n  n matrix A is diagonalizable if and only if it has n linearly independent eigenvectors.
We can then take C to be the matrix of the eigenvectors (in columns), and  the diagonal
matrix with the corresponding eigenvalues along the diagonal.

254
Fact 22.40 (Spectral decomposition.) If the eigenvectors are linearly independent, then
we can decompose A as
h i
A D ZZ 1 , where  D diag.1 ; :::; 1 / and Z D z1 z2


zn ;

where  is a diagonal matrix with the eigenvalues along the principal diagonal, and Z is
a matrix with the corresponding eigenvalues in the columns.

Fact 22.41 (Diagonal decomposition of symmetric matrices) If A is symmetric (and pos-

sibly singular) then the eigenvectors are orthogonal, C 0 C D I , so C 1 D C 0 . In this
case, we can diagonalize A as C 0 AC D , or A D CC 0 . If A is n  n but has rank
r  n, then we can write
" #
h i  0 h i0
1
A D C1 C2 C1 C2 D C1 1 C10 ;
0 0

where the n  r matrix C1 contains the r eigenvectors associated with the r non-zero
eigenvalues in the r  r matrix 1 .

Fact 22.42 (Cholesky decomposition) Let ˝ be an n  n symmetric positive definite

matrix. The Cholesky decomposition gives the unique lower triangular P such that
˝ D PP 0 (some software returns an upper triangular matrix, that is, Q in ˝ D Q0 Q
instead). Note that each column of P is only identified up to a sign transformation; they
can be reversed at will.

Example 22.43 (2  2 matrix) For a 2  2 matrix we have the following Cholesky decom-
position " #! " p #
a b a 0
chol D p p :
b d b= a d b 2 =a

Fact 22.44 (Triangular Decomposition) Let ˝ be an n  n symmetric positive definite

matrix. There is a unique decomposition ˝ D ADA0 , where A is lower triangular with
ones along the principal diagonal, and D is diagonal with positive diagonal elements.
This decomposition is usually not included in econometric software, but it can easily be
calculated from the commonly available Cholesky decomposition since P in the Cholesky

255
decomposition is of the form
2 p 3
D11 0


0
p p
D11 A21 D22 0
6 7
6 7
P D6 :: :: :: 7:
6
: : :
7
p p p
4 5
D11 An1 D22 An2


Dnn

Fact 22.45 (Schur decomposition) The decomposition of the n  n matrix A gives the
n  n matrices T and Z such that

A D ZT Z H

where Z is a unitary n  n matrix and T is an n  n upper triangular Schur form with the
eigenvalues along the diagonal. Note that premultiplying by Z 1 D Z H and postmulti-
plying by Z gives
T D Z H AZ;

which is upper triangular. The ordering of the eigenvalues in T can be reshuffled, al-
though this requires that Z is reshuffled conformably to keep A D ZT Z H , which involves
a bit of tricky “book keeping.”

Fact 22.46 (Generalized Schur Decomposition) The decomposition of the n  n matrices

G and D gives the n  n matrices Q, S , T , and Z such that Q and Z are unitary and S
and T upper triangular. They satisfy

G D QSZ H and D D QT Z H :

The generalized Schur decomposition solves the generalized eigenvalue problem Dx D

Gx, where  are the generalized eigenvalues (which will equal the diagonal elements in
T divided by the corresponding diagonal element in S ). Note that we can write

QH GZ D S and QH DZ D T:

Example 22.47 If G D I in the generalized eigenvalue problem Dx D Gx, then we

are back to the standard eigenvalue problem. Clearly, we can pick S D I and Q D Z in
this case, so G D I and D D ZT Z H , as in the standard Schur decomposition.

256
Fact 22.48 (QR decomposition) Let A be m  n with m  n. The QR decomposition is

Amn D Qmm Rmn

" #
h i R
1
D Q1 Q2
0
D Q1 R1 :

where Q is orthogonal (Q0 Q D I ) and R upper triangular. The last line is the “thin
QR decomposition,” where Q1 is an m  n orthogonal matrix and R1 an n  n upper
triangular matrix.

Fact 22.49 (Inverting by using the QR decomposition) Solving Ax D c by inversion of

A can be very numerically inaccurate (no kidding, this is a real problem). Instead, the
problem can be solved with QR decomposition. First, calculate Q1 and R1 such that
A D Q1 R1 . Note that we can write the system of equations as

Q1 Rx D c:

Premultply by Q10 to get (since Q10 Q1 D I )

Rx D Q10 c:

This is an upper triangular system which can be solved very easily (first solve the first
equation, then use the solution is the second, and so forth.)

Fact 22.50 (Singular value decomposition) Let A be an m  n matrix of rank . The

singular value decomposition is

0
A D Umm Smn Vnn

where U and V are orthogonal and S is diagonal with the first  elements being non-zero,
that is, 2 3
" # s11


0
S1 0 6 : :
: : ::: 7 :
7
SD , where S1 D 4 ::
6
0 0 5
0


s

257
Fact 22.51 (Singular values and eigenvalues) The singular values of A are the nonnega-
tive square roots of AAH if m  n and of AH A if m  n.

Remark 22.52 If the square matrix A is symmetric and idempotent (A D A0 A), then
the singular values are the same as the eigevalues. From Fact (22.41) we know that a
symmetric A can be decomposed as A D CC 0 . It follows that this is the same as the
singular value decomposition.

Fact 22.53 (Condition number) The condition number of a matrix is the ratio of the
largest (in magnitude) of the singular values to the smallest

c D jsi i jmax = jsi i jmin :

For a square matrix, we can calculate the condition value from the eigenvalues of AAH
or AH A (see Fact 22.51). In particular, for a square matrix we have
ˇp ˇ ˇp ˇ
c D ˇ i ˇ = ˇ i ˇ ;
ˇ ˇ ˇ ˇ
max min

where i are the eigenvalues of AAH and A is square.

Fact 22.54 (Condition number and computers) The determinant is not a good indicator
of the realibility of numerical inversion algorithms. Instead, let c be the condition number
of a square matrix. If 1=c is close to the a computer’s floating-point precision (10 13 or
so), then numerical routines for a matrix inverse become unreliable. For instance, while
det.0:1I20 / D 10 20 , the condition number of 0:1I20 is unity and the matrix is indeed
easy to invert to get 10  I20 .

Fact 22.55 (Inverting by using the SVD decomposition) The inverse of the square matrix
A is found by noting that if A is square, then from Fact 22.50 we have

1
AA D I or
USV 0 A 1
D I , so
1 1
A D VS U 0;

provided S is invertible (otherwise A will not be). Since S is diagonal, S 1 is also

diagonal with the inverses of the diagonal elements in S, so it is very easy to compute.

258
Fact 22.56 (Pseudo inverse or generalized inverse) The Moore-Penrose pseudo (gener-
alized) inverse of an m  n matrix A is defined as
" #
C C 0 C S111 0
A D V S U ; where Snxm D ;
0 0

where V and U are from Fact 22.50. The submatrix S111 contains the reciprocals of the
non-zero singular values along the principal diagonal. AC satisfies the AC satisfies the
Moore-Penrose conditions

0
0
AAC A D A, AC AAC D AC , AAC D AAC , and AC A D AC A:

See Fact 22.9 for the idea behind the generalized inverse.

Fact 22.57 (Some properties of generalized inverses) If A has full rank, then AC D A 1 ;
.BC /C D C C B C ; if B, and C are invertible, then .BAC / 1 D C 1 AC B 1 ; .AC /0 D
.A0 /C ; if A is symmetric, then AC is symmetric.

Example 22.58 (Pseudo inverse of a square matrix) For the matrix

" # " #
1 2 0:02 0:06
AD , we have AC D :
3 6 0:04 0:12

Fact 22.59 (Pseudo inverse of symmetric matrix) If A is symmetric, then the SVD is
identical to the spectral decomposition A D ZZ 0 where Z is a matrix of the orthogonal
0
eigenvectors (Z Z D I ) and  is a diagonal matrix with the eigenvalues along the main
diagonal. By Fact 22.56) we then have AC D ZC Z 0 , where
" #
1
 11 0
C D ;
0 0

with the reciprocals of the non-zero eigen values along the principal diagonal of 111 .

259
22.8 Matrix Calculus

Fact 22.60 (Matrix differentiation of non-linear functions, @y=@x 0 ) Let the vector yn1
be a function of the vector xm1
2 3 2 3
y1 f1 .x/
6 : 7
6 :: 7 D f .x/ D 6 :::
6 7
7:
4 5 4 5
yn fn .x/

Then, let @y=@x 0 be the n  m matrix

2 3 2 3
@f1 .x/ @f1 .x/ @f1 .x/
@x 0 @x1



@xm
@y 6 :
6 ::
7 6 :
7 D 6 :: :: 7
D : 7:
@x 0 4
@f
5 4 5
n .x/ @fn .x/ @fn .x/
@x 0 @x1



@xm

This matrix is often called the Jacobian of the f functions. (Note that the notation implies
that the derivatives of the first element in y, denoted y1 , with respect to each of the
elements in x 0 are found in the first row of @y=@x 0 . A rule to help memorizing the format
of @y=@x 0 : y is a column vector and x 0 is a row vector.)

Fact 22.61 (@y 0 =@x instead of @y=@x 0 ) With the notation in the previous Fact, we get
2 3
@f1 .x/ @fn .x/
i 6 @x1


@x1 0
@y 0

h
@f1 .x/ @fn .x/ : : 7 @y
D


D6 : :: 7D :
@x @x @x 4 : 5 @x 0
@f1 .x/
@xm



@f@xn .x/
m

Fact 22.62 (Matrix differentiation of linear systems) When yn1 D Anm xm1 ; then
f .x/ is a linear function
2 3 2 32 3
y1 a11


a1m x1
6 : 7 6 : :: 76 : 7
6 :: 7 D 6 :: : 7 6 :: 7 :
4 5 4 54 5
yn an1


anm xm

In this case @y=@x 0 D A and @y 0 =@x D A0 .

Fact 22.63 (Matrix differentiation of inner product) The inner product of two column
vectors, y D z 0 x, is a special case of a linear system with A D z 0 . In this case we get

260
@ .z 0 x/ =@x 0 D z 0 and @ .z 0 x/ =@x D z. Clearly, the derivatives of x 0 z are the same (a
transpose of a scalar).

Example 22.64 (@ .z 0 x/ =@x D z when x and z are 2  1 vectors)

" #! " #
@ h i x z1
1
z1 z2 D :
@x x2 z2

Fact 22.65 (First order Taylor series) For each element fi .x/ in the n vector f .x/, we
can apply the mean-value theorem

@fi .bi /
fi .x/ D fi .c/ C .x c/ ;
@x 0
for some vector bi between c and x. Stacking these expressions gives
2 3 2 3 2 32 3
@f1 .b1 / @f1 .b1 /
f1 .x/ f1 .c/


x1
6 : 7 6 : 7 6 : @x1 ::
@xm
76 :: 7
6 :: 7 D 6 :: 7 C 6 :: : : 75 or
76
4 5 4 5 4 54
@fn .bn / @fn .bn /
fn .x/ fn .c/ @x1



@xm
xm
@f .b/
f .x/ D f .c/ C .x c/ ;
@x 0
where the notation f .b/ is a bit sloppy. It should be interpreted as that we have to
evaluate the derivatives at different points for the different elements in f .x/.

Fact 22.66 (Matrix differentiation of quadratic forms) Let xm1 be a vector, Amm a
matrix, and f .x/n1 a vector of functions. Then,

@f .x/0 Af .x/ @f .x/ 0

 
A C A0 f .x/


D 0
@x @x
@f .x/ 0
 
D2 Af .x/ if A is symmetric.
@x 0

If f .x/ D x, then @f .x/ =@x 0 D I , so @ .x 0 Ax/ =@x D 2Ax if A is symmetric.

261
Example 22.67 (@ .x 0 Ax/ =@x D 2Ax when x is 2  1 and A is 2  2)
" # " #! " # " #! " #
@ h i A A x A A A A x1
11 12 1 11 12 11 21
x1 x2 D C ;
@x A21 A22 x2 A21 A22 A12 A22 x2
" #" #
A11 A12 x1
D2 if A21 D A12 :
A12 A22 x2

Example 22.68 (Least squares) Consider the linear model Ym1 D Xmn ˇn1 C um1 .
We want to minimize the sum of squared fitted errors by choosing the n  1 vector ˇ. The
fitted errors depend on the chosen ˇ: u .ˇ/ D Y Xˇ, so quadratic loss function is

L D u.ˇ/0 u.ˇ/
D .Y Xˇ/0 .Y Xˇ/ :

In thus case, f .ˇ/ D u .ˇ/ D Y Xˇ, so @f .ˇ/ =@ˇ 0 D X. The first order condition
for u0 u is thus  
O O
2X Y X ˇ D 0n1 or X 0 Y D X 0 X ˇ;
0

which can be solved as

1
ˇO D X 0 X X 0 Y:

Fact 22.69 (Matrix of 2nd order derivatives of of a non-linear function, @2 y=@x@x 0 ) Let
the scalar y be a function of the vector xm1

y D f .x/ :

Then, let @2 y=@x@x 0 be the m  m matrix with @2 y=@xi @xj in cell .i; j /
2 2 3
@ f .x/ @2 f .x/
@x1 @x1



@x 1 @xm
2
@ y 6 :
6 :: :: 7
D : 7:
@x@x 0 4 5
@2 f .x/ @2 f .x/
@xm @x1



@x m @xm

This matrix is often called the Hessian of the f function. This is clearly a symmetric
matrix.

262
22.9 Miscellaneous

Fact 22.70 (Some properties of transposes) .A C B/0 D A0 C B 0 ; .ABC /0 D C 0 B 0 A0 (if

conformable).

Fact 22.71 (Kronecker product) If A and B are matrices, then

2 3
a11 B


a1n B
6 : :: 7
A˝B D6 ::
4 : 7 5:
am1 B


amn B

Some properties: .A ˝ B/ 1 D A 1 ˝ B 1 (if conformable); .A ˝ B/.C ˝ D/ D

AC ˝ BD (if conformable); .A ˝ B/0 D A0 ˝ B 0 ; if a is m  1 and b is n  1, then
a˝b D .a˝In /b; if A is symmetric and positive definite, then chol.A˝I / Dchol.A/˝I
and chol.I ˝ A/ D I ˝chol.A/.

Fact 22.72 (Cyclical permutation of trace) Trace.ABC / DTrace.BCA/ DTrace.CAB/,

if the dimensions allow the products.

Fact 22.73 (The vec operator). vec A where A is m  n gives an mn  1 vector2with the
3
" a
# 6 11 7
a11 a12 6 a21 7
columns in A stacked on top of each other. For instance, vec D6 6 a 7.
7
a21 a22 4 12 5
a22
Properties: vec .A C B/ D vec AC vec B; vec .ABC / D .C 0 ˝ A/ vec B; if a and b
are column vectors, then vec .ab 0 / D b ˝ a.

Fact 22.74 (The vech operator) vechA where A is m  m gives an m.m C 1/=2  1 vector
2 A stacked
with the elements on and below the principal diagonal 3 on top of each other
" # a11
a11 a12
(columnwise). For instance, vech D 4 a21 5, that is, like vec, but uses
6 7
a21 a22
a22
only the elements on and below the principal diagonal.

Fact 22.75 (Duplication matrix) The duplication matrix Dm is defined such that for any
symmetric m  m matrix A we have vec A D Dm vechA. The duplication matrix is

263
therefore useful for “inverting” the vech operator (the step from vec A to A is trivial).
For instance, to continue the example of the vech operator
2 3 2 3
1 0 0 2 3 a11
7 a
6 0 1 0 7 6 11 7 6 a21 7
6 6 7
7 4 a21 5 D 6
6 a 7 or D2 vechA D vec A:
6 7
6 0 1 0 7
4 21 5
a22
4 5
0 0 1 a22

Fact 22.76 (OLS notation) Let x t be k  1 and y t be m  1. Suppose we have T such

vectors. The sum of the outer product (a k  m matrix) is
T
X
SD x t y t0 :
t D1

Create matrices XT k and YT m by letting x t0 and y t0 be the t t h rows

2 3 2 3
x10 y10
6 : 7
:: 7 and YT m D 6 ::: 7 :
6 7
XT k D 64 5 4 5
0 0
xT yT

We can then calculate the same sum of outer product, S , as

S D X 0 Y:

(To see this, let X.i; W/ be the i th row of X , and similarly for Y , so
T
X
0
XY D X.t; W/0 Y.t; W/;
t D1

which is precisely ˙ tTD1 x t y t0 .) For instance, with

2 3
" # pt
at
xt D and y t D 4 q t 5 ;
6 7
bt
rt

and T D 2 we have
" #" # " #
0 a 1 a 2 p 1 q1 r 1 PT at h i
XY D D t D1 pt qt rt :
b1 b2 p2 q2 r2 bt

264
Fact 22.77 (Matrix geometric series) Suppose the eigenvalues to the square matrix A are
all less than one in modulus. Then,

1
I C A C A2 C


D .1 A/ :

To see why this makes sense, consider .1 A/ ˙ tTD1 At (with the convention that A0 D I ).
It can be written as

A/ ˙ tTD1 At D I C A C A2 C


A I C A C A2 C


D I AT C1 :



.1

If all the eigenvalues are stable, then limT !1 AT C1 D 0, so taking the limit of the
previous equation gives
.1 A/ lim ˙ tTD1 A D I:
T !1

Fact 22.78 (Matrix exponential) The matrix exponential of an n  n matrix A is defined

as
1
X .At/s
exp .At/ D :
sD0
sŠ

Bibliography
Anton, H., 1987, Elementary linear algebra, John Wiley and Sons, New York, 5th edn.

Björk, Å., 1996, Numerical methods for least squares problems, SIAM, Philadelphia.

Golub, G. H., and C. F. van Loan, 1989, Matrix computations, The John Hopkins Univer-
sity Press, Baltimore, 2nd edn.

Greenberg, M. D., 1988, Advanced engineering mathematics, Prentice Hall, Englewood

Cliffs, New Jersey.

Greene, W. H., 2000, Econometric analysis, Prentice-Hall, Upper Saddle River, New
Jersey, 4th edn.

265