Examiners’ commentaries 2009

Examiners’ commentaries 2009

20 Elements of econometrics

Specific comments on questions – Zone A

Section A

Answer all eight parts of question 1 (15 marks each).

Question 1

(a) RF corresponding to zt and yt are:

β2 u2t − u1t
zt = xt + ··· , (1.1)
α − β1 α − β1

αβ2 α(u2t − u1t )

yt = xt + + u1t
α − β1 α − β1
αβ2 αu2t − βu1t
= xt + ··· . (1.2)
α − β1 α − β1

( 1.1) and ( 1.2) will exist only if α − β1 6= 0.

In the reduced form, the explanatory variable and the disturbance term are independent, hence
reduced form parameters can be consistently estimated by OLS.
[See Dougherty, 2007, Chapter 9.1.]
(b) A time series is stationary (weakly) if the mean, variance and autocovariance of the series are
independent of time. Autocovariance may depend on length of the lag.
In this example Xt = θXt−1 + ut ; t = 1, 2, . . . , T , where E(ut ) = 0; Var (ut ) = σ 2 and
E(us ut ) = 0 for all s and t, s 6= t. Assume X0 is fixed. We can write:

t=1: X1 = θX0 + u1 ,
t=2: X2 = θX1 + u2 = θ(θX0 + u1 ) + u2 = θ2 X0 + θu1 + u2 ,
t=3: X3 = θX2 + u3 = θ3 X0 + θ2 u1 + θu2 + u3 ,
..
.

Doing these recursive substitutions, we can write:

Xt = θXt−1 + ut = θt X0 + ut + θut−1 + · · · + θt−1 u1 .

Therefore:
E(Xt ) = E(θt X0 + ut + θut−1 + · · · + θt−1 u1 ) = θt X0 ,

Var (Xt ) = Var θt X0 + ut + θut−1 + · · · + θt−1 u1

= σε2 (1 + θ2 + θ4 + · · · + θ2(t−1) )
t−1
X
= σ2 θ2s .
s=0

1
20 Elements of econometrics

If θ ≥ 1 then for large ‘t’, it is easy to see that E(Xt ) → ∞ and Var (Xt ) → ∞. So the variable
Xt is non-stationary and standard analysis test statistics are not valid as these processes assume
stationarity and finite mean and variance for the random variable. If this is the case then the
variable is growing at the exponential rate which is rare for economic variables.
If |θ| < 1, then for large ‘t’:
E(Xt ) = θt X0 = 0,
and:
t−1
X σ2
Var (Xt ) = σ 2 θ2s = σ 2 (1 + θ2 + θ4 + · · · ) = ,
s=0
1 − θ2
which is a constant.
For large t, covariance is given by:
Cov (Xt , Xt−s ) = E[Xt − E(Xt )][Xt−s − E(Xt−s )] = E[Xt Xt−s ]
= E[ut + θut−1 + θ2 ut−2 · · · + θt−1 u1 ][ut−s + θut−s−1 + · · · + θt−1 u−s ]
= θs (1 + θ2 + θ4 + · · · )σ 2
θs σ2
= ,
1 − θ2
which depends only upon the value of ‘s’. Therefore if −1 < θ < 1. Hence the variable Xt is
stationary if |θ| < 1.
[See Dougherty, 2007, Chapter 13.1.]
(c)
Var ((aX + bY )) = E[(aX + bY ) − E(aX + bY )]2
= E[(aX − E(aX)) + (bY − E(bY ))]2
= E[(aX − E(aX))]2 + E[(bY − E(bY ))]2 + 2E[(aX − E(aX))(bY − E(bY ))]
= a2 E[(X − E(X))]2 + b2 E[(Y − E(Y ))]2 + 2abE[(X − E(X))(Y − E(Y ))]
= a2 Var (X) + b2 Var (Y ) + 2abCov (X, Y ) .
If X and Y are independent then Cov (X, Y ) = 0 and:
Var (aX + bY ) = a2 Var (X) + b2 Var (Y ) .
[See Dougherty, 2007, Chapter R.4.]
(d) Given the equations:
Yt = β1 + β2 Xt + ut ,
ut = ρut−1 + vt ,
where v has zero mean, constant variance and zero autocovariance, combine the two equations to
give:
Yt = β1 (1 − ρ) + ρYt−1 + β2 Xt − β2 ρXt−1 + vt ,
which is the restricted of the general form (an (ADL(1, 1)) model):
Yt = λ1 + λ2 Yt−1 + λ3 Xt + λ4 Xt−1 + vt ,
and is subject to the restriction λ4 = −λ2 λ3 . The test of this restriction is the common factor
test.
Note that the usual F test of the restriction is not appropriate because the restriction is
non-linear so we have to use the test statistic:
 
RSSR
N log ∼ χ21 ,
RSSU
where RSSR and RSSU are residual sum of squares from the restricted and unrestricted models
respectively. N is the sample size and the test statistic is asymptotically chi-square with one
degree of freedom.
[See Dougherty, 2007, Chapter 12.6.]

2
 Examiners’ commentaries 2009

(e) Type I error is the error of rejecting the null hypothesis when it is true. Probability of Type 1
error is usually denoted by α. Probability of Type 1 error is the level of significance.
Type II error is the error of not rejecting the null hypothesis when the alternative hypothesis is
true. Probability of Type II error is usually denoted by β.
Power of a test is defined as (1 – Probability of Type II error). A good answer here will give a
diagrammatic representation of the relationship between α, β and the power of the test. A good
answer here will give a diagrammatic representation of the relationship between α, β and the
power of the test.
[See Dougherty, 2007, Chapter 2.8.]
(f) Concept of sufficient condition for consistency has to be used.
[See Dougherty, 2007, Chapter R.8.]
i.
P 
Xi Nµ µ
E = 2 = ⇒ biased,
N2 N N
N σ2 σ2
P 
Xi
Var = = ,
N2 N4 N3
P 
Xi
lim E → 0 ⇒ asymptotically not unbiased.
n→∞ N2
This estimator is biased and also asymptotically not biased, inefficient (no need to
compare variance with variances of other estimators as biased) and inconsistent.
ii.
P 
Xi Nµ µ
E = = ⇒ biased,
N −1 N −1 1 − 1/N
P 
Xi
As N → ∞, E → µ ⇒ asymptotically unbiased,
N −1
N σ2 N σ2 σ2
P 
Xi
Var = 2
= 2 = → 0,
N −1 (N − 1) N − 2N + 1 N − 2 + 1/N
as N → 0.
This estimator is biased, inefficient but consistent as it is asymptotically unbiased and
variance tends to 0 as N → ∞.
iii.
P 
Xi Nµ
E = = µ ⇒ unbiased,
N −1 N
N σ2 σ2
P 
Xi
Var = = → 0,
N N2 N
as N → ∞.
This estimator is unbiased, efficient and consistent.
(g) Drawback of LPM and brief discussion of logit or probit model is required. Some technical
analysis should be used.
[See Dougherty, 2007, Chapter 10.1.]
(h) Likelihood is function is:
P P
L = px1 (1 − p)1−x1 · · · pxn (1−xn ) = p xi (1 − p)n− xi
,
X   X 
ln L = xi ln p + n − xi ln (1 − p),
P P
∂ ln L xi n − xi
= + = 0,
∂p p̂ 1 − p̂
or: P P
(1 − p̂) xi − p̂(n − xi )
= 0.
p̂(1 − p̂)

3
20 Elements of econometrics

Solving, we get: P
xi
p̂ = .
n
This is the maximum likelihood estimate.
Corresponding maximum likelihood estimator (MLE) of p is:
P
Xi
p̂M LE = = X̄.
n
[See Dougherty, 2007, Chapter 10.6.]

Section B

Answer three questions from this section (20 marks each).

Question 2

(a) There are many reasons why economic relationships would be lagged, especially if the data is
collected quarterly or monthly. There are relationships where inertia is strong, e.g. expenditure
on housing. Another area where lagged relationships are important is that of expectations
whereby economic agents’ actions are dependent on expectations (which are modelled by
extrapolation of past behaviour). Other economic relationships mentioned by Dougherty include
Friedman’s permanent income hypothesis and Brown’s habit-persistence model of aggregate
demand. All these models use lagged values of variables to explain economic behaviour.
[See Dougherty, 2007, Chapter 11.2.]
(b)
Yt = α + βXt + βλXt−1 + βλ2 Xt−2 + βλ3 Xt−3 + . . . + εt . (2.1)
If we multiply this equation through by λ and lag we get:

λYt−1 = αλ + βλXt−1 + βλ2 Xt−2 + βλ3 Xt−3 + βλ4 Xt−4 . . . + λεt−1 . (2.2)

Now subtract ( 2.2) from ( 2.1) to give:

Yt − λYt−1 = α(1 − λ) + βXt + εt − λεt−1 ,

Yt = α(1 − λ) + βXt + λYt−1 + (εt − λt−1 ),

which we could estimate by OLS except that Yt−1 and t−1 are correlated hence a RHS variable
is correlated with the error term so OLS produces inconsistent parameter estimates.
(c) i. Yt can be expressed as:

Yt = β1 + β2 λxt + β2 (1 − λ)Xt−1 + β2 (1 − λ)2 Xt−2 + · · · + β2 (1 − λ)s Xt−s+1

e
+ ut .

As 0 ≤ λ ≤ 1, higher terms can be neglected. OLS is not appropriate as the parameters

are non-linear hence some no-linear technique or grid search has to used.
ii. Short run effect is β2 λ. Let Y and X in equilibrium be Y and X , then Yt = Y and
Xt = Xt−1 = · · · = X. Therefore, the equation in ( 2.1) can be expressed as

Y = β1 + β2 X[λ + λ(1 − λ) + λ(1 − λ)2 + · · · ]

= β1 + β2 X.

Hence long run effect of X on Y is β2 .

[See Dougherty, 2007, Chapter 11.3.]

4
 Examiners’ commentaries 2009

Question 3

(a) The effect is that OLS gives unbiased, consistent but inefficient estimates and the standard errors
(and hence t values) are incorrect ⇒ t tests (and F tests) are invalid. If explanatory variables
include a lagged dependent variable then estimates are also inconsistent.
[See Dougherty, 2007, Chapter 12.3.]
(b) Discussion of Durbin-Watson test should be given.
The assumptions required are that the equation should include a constant term, that there
should be no lagged dependent variable and that the serial correlation is of first-order.
[See Dougherty, 2007, Chapter 12.3.]
(c) Cochrane-Orcutt iterative procedure can be used. Technical details are required.
[See Dougherty, 2007, Chapter 12.4.]

Question 4

(a) If after removing the trend from a non-stationary series the resulting variables becomes
stationary, then the variable is called trend stationary. Let:

Zt = Xt − α1 t = α0 + ut ,

where E(ut ) = 0, Var (ut ) = σ 2 and E(ut ut−s ) = 0 for all s and t. Then:

E(Zt ) = E(α0 + ut ) = α0 ,
Var (Zt ) = Var (α0 + ut ) = σ 2 ,
Cov (Zt , Zt−s ) = E[Zt − E(Zt )][Zt−s − E(Zt−s )] = E(ut ut−s ) = 0.

This means that Zt has constant mean and variance for all t, and covariance is zero for all s. It
implies that the series is trend-stationary.
If a non-stationary process can be transformed into a stationary process by differencing then the
series is said to be difference stationary.
Let Xt be a random walk with a drift:

Xt = β0 + Xt−1 + εt , (4.1)

where E(εt ) = 0, Var (εt ) = σ 2 and E(εt εs ) = 0 for all s and t, s 6= t.

Subtract Xt−1 from both sides of ( 4.1) to get:

∆Xt = Xt − Xt−1 = β0 + εt .

It can be easily checked that E(∆Xt ) = β0 , Var (∆Xt ) = σε2 and Cov (∆Xt , ∆Xt−s ) = 0 for all s
and t. This means that ∆Xt is stationary. This implies that Xt is difference stationary.
It is important to know whether a variable is difference or trend stationary because for difference
stationary variables shocks have a permanent effect whereas for trend stationary variables shocks
are transitory.
[See Dougherty, 2007, Chapter 13.1.]
(b) Description of Dickey-Fuller test and ADF test is required.
[See Dougherty, 2007, Chapter 13.3.]
(c) Consider a simple ADL(1,1) [this is also known as ARDL(1, 1)] model:

Yt = α1 + α2 Yt−1 + α3 Xt + α4 Xt−1 + ut . (4.2)

5
20 Elements of econometrics

Rewrite ( 4.2) as:

∆Yt = Yt − Yt−1
= α1 + α2 Yt−1 − Yt−1 + α3 Xt − α3 Xt−1 + α3 Xt−1 + α4 Xt−1 + ut
= α1 − (1 − α2 )Yt−1 + α3 ∆Xt + (α3 + α4 )Xt−1 + ut
 
α1 (α3 + α4 )
= α3 ∆Xt − (1 − α2 ) Yt−1 − − Xt−1 + ut
(1 − α2 ) (1 − α2 )
= α3 ∆Xt − (1 − α2 ) [Yt−1 − β1 − β2 Xt−1 ] + ut ,

or:
∆Yt = α3 ∆Xt − π [Yt−1 − β1 − β2 Xt−1 ] + ut , (4.3)
where:
α1 (α3 + α4 )
π = (1 − α2 ), β1 = , β2 = .
(1 − α2 ) (1 − α2 )
Equation ( 4.3) is the ECM.
When the two variables Y and X are cointegrated the ECM incorporates not only the short-run
but also log-run effects. The long run equilibrium Yt−1 − β1 − β2 Xt−1 is included in the model
together with the short-run effect captured by the differenced term.
All the terms in the ECM, given by ( 4.3), are stationary. As Y and X are I(1), then ∆X and
∆Y are I(0). As Y and X are cointegrated their linear combination:

ut−1 = Yt−1 − β1 − β2 Xt−1 ∼ I(0).

The coefficient π provides us with the information about the speed of adjustment in cases of
disequilibrium:
i. If π = 1 then 100% of the adjustment takes place within the period. In other words
adjustment is instantaneous and full.
ii. If π = 0.5 then 50% adjustment takes place each period.
iii. If π = 0 then there is no adjustment.
[See Dougherty, 2007, Chapter 13.5.]

Question 5

(a) The quarterly dummies are attached to the level of sales implying that the sales effect on profits
varies over quarters. In this case quarter 4 is the base quarter (hence Q4 is not needed in the
equation) and the sales effect on profits is increased in quarters 1 and 3 but reduced in quarter 2.
If Q4 is included in the equation then we will face a situation of perfect multicollinearity (dummy
variable trap).
(b) With sample size 24 the critical t value is 2.093 under a 2 tail test. The coefficients are all
significantly different from 0 with the exception of quarter 2. The slope increases over the
quarter 4 figure in quarters 1 and 3.
(c) The effect of using the quarterly dummies by themselves is to allow for intercept adjustments
between quarters. This is often a sensible procedure when variables, such as profits or sales, vary
seasonally. The quarterly dummies are one way of effectively seasonally adjusting the data.
(d) With both slope and intercept variation the effect is to effectively produce a different model for
each quarter.
[See Dougherty, 2007, Chapters 5.1, 5.2 and 5.3.]

6
 Examiners’ commentaries 2009

Question 6

(a) Spurious regression was first demonstrated by Granger and Newbold who showed, using Monte
Carlo techniques, that a regression involving two non-stationary series could give rise to spurious
results in that the t-statistics over-rejected the null hypothesis of a zero coefficient for two
independent random walk series.
If Yt and Xt are non-stationary and we regress Yt and Xt , i.e.

Yt = π0 + π1 Xt + vt ,

then even if there is no relationship between Y and X, the regression will produce a t-ratio which
will reject the null hypothesis H0 : π1 = 0.
The reason for this result is that if H0 : π1 = 0 then:

Yt = π0 + vt ,

and since Yt is I(1) and π0 is constant it follows that vt must be I(1). This violates the standard
distributional theory based on the assumption that vt is stationary, i.e. vt is I(0). Hence the
misleading result.
[See Dougherty, 2007, Chapter 13.2.]
(b) In general a linear combination of two time series will be non-stationary if one or more of them is
non-stationary. The degree of integration of the combination will be equal to that of most highly
integrated individual series. For example a combination of I(1) and I(0) series will be I(1) and a
combination of I(1) and I(1) series will be I(1).
If a long run relationship exists between the time series then the result may be different. Suppose
Yt and Xt are both I(1). A linear combination Yt and Xt may be written as ut = Yt − βXt . If
the linear combination ut is I(0), then Yt and Xt are said to be cointegrated.
If Yt and Xt are cointegrated then it implies that a long run relationship exists between Yt and
Xt . This concept can be generalised. Consider a general linear model:

Yt = β1 + β2 X2t + · · · + βK XKt + ut .

Then the disturbance term ut can be thought of as measuring the deviation between components
of the model. In the short run the divergence between the components will fluctuate, but if the
model is correctly specified there will be a limit to the divergence. Hence though the Y s and Xs
are non-stationary ‘u’ will be stationary. If there are K variables in the model the maximum
number of cointegrating relationships will be K − 1.
Cointegration is an overriding requirement for any economic model using non-stationary time
series data. If the variables do not cointegrate then we have the problem of spurious regression
and econometric work becomes almost meaningless.
If a cointegrating relationship exists then OLS estimators can be shown to be superconsistent.
[See Dougherty, 2007, Chapter 13.4.]
(c) Consider the model:
Yt = βXt + ut , t = 1, 2, . . . , T.
If Xt is not independently distributed of ut then the OLS estimator of β will be inconsistent.
Consider a variable Z that is correlated with u but not correlated with X. Z can be considered
as an instrumental variable. An estimator of β based on Z is known an instrumental variable
(IV) estimator. It is defined as: P
Zt Yt
β̂IV = P .
Zt Xt
It can be shown that β̂IV is a consistent estimator of β.
P P P
Zt Yt Zt (βXt + ut ) Zt ut
β̂IV = P = P =β+P ,
Zt Xt Zt Xt Zt Xt
P
plim ( Zt ut /T )
p lim(β̂IV ) = β + P = β.
plim ( Zt Xt /T )

7
20 Elements of econometrics

Thus β̂ is a consistent estimator of β.

P P
Note: plim ( Zt ut /T ) = 0 and plim ( Zt Xt /T ) 6= 0.
[See Dougherty, 2007, Chapter 8.6.]
(d) [See Dougherty, 2007, Chapter 5.4.]
(e) To answer this question it has to be shown that measurement error in explanatory variables
induces the OLS estimator to be inconsistent. Technical details are required.
[See Dougherty, 2007, Chapter 8.4.]

Question 7

(a) t values for equation (1) are 4.20, 4.70, 5.25, 4.57, 1.10 and 0.72, with critical value 2.032; and for
equation (2), 4.24, 4.66, 5.25 and 4.73, with critical value 2.028. The coefficients are of expected
sign – since we have a log-log model the coefficients are elasticities. The second equation is better
because there are no non-significant variables but need to check for multicollinearity (and
heteroskedasticity).
[See Dougherty, 2007, Chapter 2.8.]
(b) Definition of heteroskedasticity should be given.
Might expect heteroskedasticity because the data set is cross section and cities are probably very
different in size (POP and DEN).
(c) Use Goldfeldt-Quandt test.
POP and DEN as possible variables causing heteroskedasticity.
Please note that in Goldfeldt-Quandt test it is important to specify which variable is being used
for ordering the sample.
(d) OLS estimators are unbiased, consistent but inefficient. Standard errors of OLS estimators are
biased hence t and F tests are invalid.
[See Dougherty, 2007, Chapter 7.1.]

Question 8

(a) Omitted variable bias occurs when a valid variable is omitted from the estimated model. Suppose
the ‘true’ model is yt = β1 x1t + β2 x2t + ut but the estimated model is yt = β1 x1t + vt . The OLS
estimator of β1 from the estimated model is:
P
x1t yt
β̂1 = P 2
x1t
P
x1t (β1 x1t + β2 x2t + ut )
= P 2
x1t
P P
x1t x2t x1t ut
= β1 + β2 P 2 + P 2 ,
x1t x1t
with: P
  x1t x2t
E β̂1 = β1 + β2 P 2 ,
x1t
since E(ut ) = 0.
P
x1t x2t
The bias is β2 P 2 which depends on the value of β2 and the covariance between x1t and x2t .
x1t P
If x1 and x2 are orthogonal to each other then x1t x2t = 0 and the OLS estimator will unbiased.
If a relevant variable is omitted from the equation then the standard errors of the coefficients and
the test statistics are, in general,t invalidated.
[See Dougherty, 2007, Chapter 6.2.]

8
 Examiners’ commentaries 2009

(b) i. Model SS = 7.7556.

ii. Model df = 3.
iii. R-squared = 0.741.
iv. Avetemp t value = 6.55.
v. Harvrain 95% CI = [−0.0056, −0.0016].
[See Dougherty, 2007, Chapters 1.7, 2.8 and 2.9.]
(c) Critical t value is t27−4 = 2.069 hence all coefficients are significantly different from 0. Winter
rain increases the log price by .001 per mill, avetemp increases log price by .712 per degree and
harvrain reduces log price by .0036 per mill.
[See Dougherty, 2007, Chapters 1.6 and 2.8.]
(d) i. The equation predicts that as wine ages the log price increases by 0.23 per year – the
critical t value is t27−5 = 2.074.
ii. The equation in (b) has a specification error in that age appears to be part of the true
model. The omitted variable bias will be related to the true coefficient and the correlation
between the omitted variable and the remaining variables – the true coefficient is
presumably non-zero and positive, but age is unlikely to be correlated to any of the
included variables hence one would not expect any measurable bias – and this is what you
observe. The difference between the two models is very small.
[See Dougherty, 2007, Chapters 1.6, 2.8 and 6.2.]

Changes in the format of the examination

This examination is three hours long. The paper is divided into two sections. The format will
change in 2010. In 2009, Section A, which was compulsory, contained eight questions which were
intended to examine the whole syllabus. Section B was designed to examine a selection of topics in
greater depth. Section B had seven questions, of which students were required to answer three.
From 2010 the new format will be: In Section A, Question 1 will have 5 parts (instead of 8) worth 5
marks each. Thus Question 1 will be worth 25 marks (instead of 40). Question 1 will remain
compulsory. Section B will consist of 5 questions (instead of 7), of which students have to answer
three. Each question in Section B is worth 25 marks.