Applied Statistics and Econometrics, Specimen

Exam Questions with Answers.

Ron Smith: r.smith@bbk.ac.uk
2016-17

Economics: Graduate Diplomas and BScs

1 Question 1
1.1 A
For a sample of countries, i = 1; 2; :::; 189 the logarithm of life expectancy, LLE,
measured in years, was regressed on the logarithm of per capita income, LPCI,
measured in purchasing power parity ination adjusted dollars. The results
(with standard errors in parentheses) are:

LLEi = 3:58 +0:07 LP CIi +b

ui
(0:038) (0:004)
With coe cient of determination R2 = 0:623 and standard error of regression
SER=0.073.
(a) Explain how R2 , and are ubt calculated and what they measure.
(b) Which coe cients are signicantly dierent from zero? Explain how you
would interpret the coe cient of LPCI.
(c) What assumptions are required for least squares to give good estimates?
(d) A test for normality of the residuals gave a p-value of 0.000 and a test
for heteroskedasticity gave a p-value of 0.264. Do these p-values indicate mis-
specication? Explain what heteroskedasticity is and how you test for it.
(e) When the level of life expectancy, rather than the logarithm, was re-
gressed on LPCI, R2 = 0:646. Does this mean that the level specication is
better? Explain your answer.
Answer
a.The coe cient of determination tells you the proportion of the variation
in the dependent variable, yi = LLEi that the equation explains
P 2
2 bi
u
R =1 P
(yi y)2

1
bi . is the dierence between the actual and the predicted value (3:58 + 0:07LP CI)
u
and measures the part not explained by the model.
(b) Both coe cients are signicant, 3.58/0.038=94.2 and 0.07/0.004=17.5.
The coe cient on LPCI is an elasticity. A one percent increase in income
causes a 0.07 percent increase in life expectancy.
(c) The errors have mean zero, E(ui ) = 0; constant variance E(u2i ) = 2
and are not correlated with each other E(ui uj ) = 0:
The regressors are either exogenous (independent of the errors) or predeter-
mined (uncorrelated with the errors) and there are no exact linear dependences
between the regressors (which in this case means that the variance of LPCI is
non-zero.
(d) The null hypothesis of normally distributed error is rejected, the null
hypothesis of homoscedasticity is not rejected. Heteroskedasticity means that
the variance of the errors is not constant. It is tested by running a regression
of the squared residuals on a constant and source of potential variation.
(e) No. The R2 are not comparable because they relate to dierent depen-
dent variables.

1.2 B
Quarterly Colombian data for 1994Q2 to 2005Q3 on the logarithm of the number
of guerrilla attacks, at , the amount of illegal drug production, dt and a dummy
variable for quarter 1, Q1t were used to estimate the following equation by least
squares (standard errors in parentheses)
at = 5:2 +0:33 at 1 +0:66 dt 1 +0:27 Q1t +^
ut
(2:30) (0:14) (0:20) (0:10)
R2 = 0:396, SER=0.268
(a) Explain how the coe cient of determination, R2 , and standard error of
regression, SER, are calculated and what they measure.
(b) Which coe cients are signicantly dierent from zero? Explain how you
would interpret this equation.
(c) What assumptions are required for least squares to give good estimates?
(d) A test for fourth order serial correlation gave a p value of 0.58 and a
test for heteroskedasticity gave a p value of 0.85. Do these p values indicate
misspecication?
(e) Explain what fourth order serial correlation and heteroskedasticity are.
Answer
1(a). Coe cient of determination tells you the proportion of the variation
in the dependent variable that the equation explains
P 2
2 bt
u
R =1 P
(at a)2
SER tells you the average size of the error (square root of the average squared
residual); qX
s= b2t =(T k)
u

2
where k is the number of regression coe cients estimated, here 4, and T the
number of observations.
(b). All coe cients are signicant.
The interpretation is that attacks are 27% higher in Q1, and higher drug
production increases attacks with a lag of a quarter and there is partial adjust-
ment.
The long run eect of drug production on attacks is 0:66=(1 0:33) = 0:985,
so a one percent increase in drug production causes a 0.985 percent increase in
attacks. Attacks are almost proportional to drug production in the long run.
(c). The errors have mean zero, E(ut ) = 0; constant variance E(u2t ) = 2
and are not correlated with each other E(ut ut i ) = 0:
The regressors are either exogenous (independent of the errors) or predeter-
mined (uncorrelated with the errors) and there are no exact linear dependences
between the regressors.
(d). No evidence of misspecication at the 5% or 10% level since the p values
are much bigger than 0.1, so we do not reject the null of no misspecication.
(e). Fourth order serial correlation means the errors in this quarter are
related to the errors 4 quarters ago (or up to 4 quarters ago). Heteroskedasticity
means the variance of the errors is not constant.

1.3 C
US data 1971-1999 were used to estimate a relationship between the rate of
interest Rt and the rate of ination t ;

Rt = + t + ut :

The least squares estimates (with standard errors in parenthesese) were

Rt = 6:37 +0:33 t +b
ut
:
(0:66) (0:10)

The standard error of regression, s = 2:75, the coe cient of determination

R2 = 0:29; and the Durbin Watson Statistic DW = 0:63:
(a) Explain what u bt measures.
(b) Explain what R2 ; Standard Error of the Regression and Durbin Watson
statistic measure and what they tell you about this regression.
(c) Interpret the estimates of and : What do they tell you about the
relationship between ination and interest rates?
(d) What assumptions are required for least squares to give good estimates.
Answer
(a) The estimated residual u bt measures the dierence between the actual
and predicted value of
P 2 PR t :
(b) R2 = 1 bt = (Rt R)2 gives the proportion of the variation in the
u
dependentpP variable explained by the regression, 29% in this case so quite low.
s= b2t =T 2 is a measure of the average error, 2.75 percentage points
u
in this case, quite a large error in predicting the interest rate;

3
 P P 2
DW = b2t = u
u bt is a test for serial correlation, it should be close to
two, so at 0.63 this regression suers from severe positive serial correlation.
(c) is the value the interest rate would take if ination were zero, interest
rates would be 6.37%; is the eect of ination on interest rates: a 1 percentage
point increase in ination raises interest rates by 0.33 percentage points.
(d) The errors have mean zero, E(ut ) = 0; constant variance E(u2t ) = 2
and are not correlated with each other E(ut ut i ) = 0: The regressors are either
exogenous (independent of the errors) or predetermined (uncorrelated with the
errors), E( t ut ) = 0, and there are no exact linear dependences between the
regressors: the variance of t not equal zero, in the case of a single regressor.

1.4 D
Using the results in the previous question and assuming that the 95% critical
value for a t test is 2 :
(a) Test the hypothesis = 1 at the 5% level and explain why this might be
an interesting hypothesis.
(b) Explain what Type I and Type II errors are. What is the probability of
Type I error in your test in part (a).
(c) Give a 95% condence interval for b . Explain what a condence interval
is.
(d) What indicates that the estimated equation is clearly misspecied? How
would you remedy this problem?
(e) An alternative model is suggested:

Rt = 0 + 1 Rt 1 + 0 t + 1 t 1 + ut :

How would you test whether the lagged values of Rt 1 and t 1 are signicant
(i) individually and (ii) jointly?
(f) Using the model in section (e) how would you estimate the long-run eect
of ination on interest rates and test that this long-run eect is equal to one?
Answer
(a) t( = 1) = (0:33 1)=0:10 = 6:7 reject the hypothesis that equals
one. If the real interest rate (It = Rt t ) is constant plus a random error
(Fisher Hypothesis) It = I + ut then Rt = I + t + ut then in the regression
= I and = 1:
(b) Type I error is rejecting the null when it is true, Type II error is accepting
the null when it is false. The probability of type I error in (a) is 5%.
(c) The 95% condence interval is 0:33 2 0:10 i.e. the range 0:13 to 0:53:
We are 95% condent that this range covers the true value of :
(d) The very low DW indicates massive serial correlation, respecify by in-
cluding lags as in the model in (e).
(e) Individually: use the t statistics on the estimated coe cients 1 and
1 : Jointly: rst take the restricted residual sum of squares from the equation
in question 1, RRSS and the unrestricted residual sum of squares from the

4
equation in question 4(e), U RSS and calculate the F statistic
(RRSS U RSS)=2
F (2; T 4) = :
U RSS=(T 4)
If this was greater than the relevant critical value reject the joint hypothesis
that the coe cients of both the lagged values are zero.
(f) The long run eect is measured as = ( 0 + 1 )=(1 1 ): If = 1; then
(1 1 ) = ( 0 + 1 ) or 0 + 1 + 1 1 = 0: We can reparameterise the equation
as
Rt = 0 +( 1 1)Rt 1 + 0 t +( 0 + 1) t 1 + ut
Rt = 0+( 1 1)(Rt 1 t 1) + 0 t+( 0+ 1 + 1 1) t 1 + ut :
So if we estimate this and test that the coe cient on t 1 equals zero, this is a
test of a long-run unit eect. You could also do an F test using the restricted
model where t 1 is excluded.

2 Question 2
2.1 A
Suppose the average real return on equities over 100 years was 5% and its
standard deviation 10% and it appeared independently normally distributed.
For a Random Variable Z following a standard normal distribution

z 0:5 1 1:5 2 2:5 3

P (Z < z) 0:6915 0:8413 0:9332 0:9772 0:9938 0:9987
(a) Explain what P (Z < z) means. What is P (Z < 0)?
(b) Is the mean return over the 100 years signicantly dierent from zero?
(c) What is the probability of a negative return?
(d) What is the probability of a negative return, three years running?
(e) Explain how you would measure skewness and kurtosis, what values
should they take for a normal distribution.
Answer
(a) P (Z < z) is the probability that a random variable Z takes a value less
than a specied value z. P (Z < 0) = 0:5:
(b) se (mean) =10/10= 1. z= 5/1=5. Reject Ho, mean return is signicantly
dierent from zero.
(c) The probability of a negative return is z=(0-5)/10=-0.5. P(Z<-0.5)=P(Z>0.5)=1-
0.6915=0.3085.
(d) The probability of a negative return, three years running is (0.3085)^3=0.0293
(e) To measure skewness and kurtosis standardise the data to say zt , by
subtracting
P 3 mean and dividing by standard deviation. Coe cient of skewness
is
P 4 t zt =T; should be zero for a normal distribution, coe cient of kurtosis,
t zt =T should be 3.

5
2.2 B
Suppose that you have a machine for lling 1 kilogram bags of sugar. The
machine is set so that the weight of a bag should be a normally distributed
random variable with an expected value of 1005g and a standard deviation of 2g.
For a random variable Z with a standard normal, the cumulative probabilities
are:
z 0 0:33 0:5 1 1:5 2 2:5
P (Z < z) 0:5 0:6293 0:6915 0:8413 0:9332 0:9772 0:9938
(a) Expain what P (Z < z) tells you.
(b) What is the probability that a bag of sugar will weight (i) less than 1000g
(ii) between 1004g and 1006g?
(c) You take a sample of 16 bags of sugar, which have a mean of 1004g:
What is (i) the standard error of the mean and (ii) the probability of getting a
sample mean of 1004g or less?
(d) On the basis of this estimate of the mean would you conclude that the
machine was still working correctly?
(e) How would you investigate whether the variance of the machine might
have increased?
Answer
(a) The probability that a random variable Z takes a value less than a par-
ticular value z.
(b)
(i) z=(1000-1005)/2=-2.5. P(z<-2.5)=1-P(z<2.5)=1-0.9938=0.0062=, 0.6%,
roughly one chance in 200
(ii) z=(1004-1005)/2=0.5. P(-0.5<Z<0.5)=2x(0.6915-0.5)=0.383=38%.
p
(c) (i) standard error of the mean is 2= 16 = 2=4 = 0:5;
(ii) P (Z < 1004) : z=(1004-1005)/0.5=-2. P(z<-2)=1-P(z<2)=1-0.9772=0.0228,
2.28%.
(d) From c(ii) the probability of getting this value or less is 2.28%, which
is a small number, so it is probably not working correctly. This is not needed
for the answer but if you wanted to test at the 5% level the null hypothesis
that it was working properly ( = 1005), you would need to be careful whether
the alternative was 6= 1005; in which case there would be 2.5% in each tail;
or < 1005; in which case there would be 5% in the lower tail. Since the
probability is less than 2.5%, you would reject either on a two tailed or one tail
test. P
(e) Estimate the sample variance from your sample of 16 s2 = (16 1) 1 (xi
x)2 ; to check whether the variance seemed to have increased from 4. This is not
needed for the answer but you would use a variance ratio F test.

6
3 Question 3
3.1 A
Consider the table below giving data on the annual rates of ination of the
Consumer Price Index, CPI and Retail Price Index RPI and the sources of the
dierences between them.
Annual Rates of ination, year to August, and sources of dier-
ences CPI and RPI, percentages.
dierence due to
CPI RPI CPI-RPI Housing formula other
2007 1.8 4.09 -2.29 -2.20 -0.57 0.48
2008 4.7 4.77 -0.07 -0.06 -0.49 0.49
2009 1.6 -1.29 2.89 3.06 -0.55 0.38
2010 3.1 4.68 -1.58 -0.72 -0.86 0.00
2011 4.5 5.71 -0.67 0.05 -0.99 0.27
2012 2.5 2.95 -0.45 -0.15 -0.88 0.58
2013 2.7 3.29 -0.59 -0.24 -0.94 0.60

(a) Explain the main dierences between the GDP deator, CPI and RPI
as measures of ination.
(b) The Bank of England has a target range for annual CPI ination. What
is the range and in which years were they outside this range?
(c) Explain the main purposes for which the RPI is used.
(d) Explain the source of the formula eect on the dierence.
(e) Comment on the average dierence and the movement of the various
components of the dierence over this period.
Answer
(a) The GDP deator is the ratio of current price GDP to constant price
GDP and covers all goods in the economy, the CPI and RPI only cover con-
sumer goods. The CPI excludes housing and uses a geometric formula, the RPI
includes housing and uses an arithmetic formula. The RPI excludes expenditure
by the very rich and very poor.
(b) The Bank of England target is 2% CPI ination and range is 1% to 3%,
outside which the Governor has to write a letter to the Chancellor. On this
data it was outside the range in 2008, 2010 and 2011.
(c) The RPI is used to determine the amounts payable on index linked gilts,
and the basis of setting some wages.
(d) In order to aggregate prices at the most basic level to form the elementary
aggregates the RPI uses the (Carli) arithmetic means of the price relatives (price
in year t to price in year zero) whereas the CPI uses the (Jevons) geometric
mean. The arithmetic mean over-estimates relative to the geometric mean and
the upward bias is greater when the variance of ination is greater.
(e) Over this period the average dierence was -0.39, the RPI was roughly
0.4 percentage points higher, but the range was very large -2.29 to +2.89. The
bulk of the volatility came from housing, particularly the house price boom

7
in 2007 and bust in 2009. There is a trend increase in the size of the formula
eect which was assumed to be 0.5 percentage points at the time that the target
changed from RPIX to CPI.

3.2 B
The following data were taken from Economic Trends, February 2004.

N DY RDY RP I CP I HP T BY
2000 654; 649 654; 649 170:3 105:6 87:7 5:69
2001 700; 538 685; 263 173:3 106:9 95:1 3:87
2002 721; 044 696; 224 176:2 108:3 111:2 3:92

N DY is nominal household disposable income in current prices; RDY is real

household disposable income in constant 2000 prices; RP I is the retail price
index, 1987=100; CP I is the consumer price index, 1996=100; HP is a house-
price index; T BY is the Treasury Bill Yield a short term interest rate expressed
as percent per annum.
(a) Calculate the price index (implicit deator) for disposable income.
(b) Calculate the rate of ination for 2001-2 for disposable income, RP I;
CP I; and HP: Comment on the relation between them.
(c) Suppose that you owned a house, whose price increased at the average
rate, and had a 100% mortgage paying the Treasury Bill Rate on the mortgage,
what would be the real return on owning the house over 2001-2.
(d) The RPI includes housing prices in its measure of ination, the CPI does
not. Should house prices be included in ination measures?
(e) What other dierences are there between the RPI and CPI?
Answer
(a) The disposable income deator is P DY = 100 N DY =RDY
N DY RDY P DY
2000 654; 649 654; 649 100
2001 700; 538 685; 263 102:2
2002 721; 044 696; 224 103:6
(b) Ination, 2001-2:
PDY=1.37% , RPI=1.67%; CPI=1.31; HP=16.92.
Massive house price boom, RPI slightly higher than CPI or PDY.
(c). Real return is capital gain on house prices, less interest cost, less rate
of ination. Using CPI (others are acceptable)= 16.92-3.92-1.31=11.69%.
(d) Yes even if you own your own home and have paid o your mortgage
there is an implicit rental cost of home ownership and this will increase when
house prices increase.
(e) CPI uses a geometric mean, RPI an arithmetic mean which means that
CPI tended to be about 0.5% lower, though the dierence has increased.

8
3.3 C
The table below provides data on UK prices, output and oil prices in Quarter
1 of each year and their percentage changes.

RPI CPI ABMI YBHA Brent

2010 219.3 112.9 400001 383623 76.17
2011 230.9 117.6 408601 403445 96.52
2012 239.6 121.7 414835 412819 110.69
2013 247.4 125.1 420779 425516 112.96
2014 253.9 127.2 432660 446691 108.12
2015 256.4 127.4 443590 463411 33.14
Percentage Changes from previous year
2011 5.29 4.16 2.15 5.17 26.72
2012 3.77 3.49 1.53 2.32 14.68
2013 3.26 2.79 1.43 3.08 2.05
2014 2.63 1.68 2.82 4.98 -4.28
2015 0.98 0.16 2.53 3.74 -69.35

RPI: retail price index all items index: Jan 1987=100

CPI: consumer price index all items Index: 2005=100
ABMI: GDP chained volume measure 2012 constant price base, m
YBHA: GDP at current market prices, m
Brent: crude oil price in $ per barrel.
(a) Explain how current price GDP, constant price GDP and the GDP
deator are measured.
(b) Calculate the GDP deator for 2010-15 and its rate of ination for
2011-2015.
(c) Explain the dierence between the three measures of ination based
on RPI, CPI and the GDP deator.
(d) Give examples of what the three dierent measures of ination are
used for.
(e) Comment on the dierences between the three measures over this
period and discuss the extent to which ination can be explained by growth
and oil prices over the period.

3.4 Answer
PN
a. Denote current price GDP, total expenditure, in year t , i=1 pit qit ;
PN
constant price GDP, valued in year zero prices, is i=1 pi0 qit ; the GDP deator

9
is the ratio of current to constant
PN
i=1 pit qit
Pt = PN
i=1 pi0 qit
b.

GDP
RPI CPI ABMI YBHA Brent deflator
2010 219.3 112.9 400001 383623 76.17 0.959055
2011 230.9 117.6 408601 403445 96.52 0.987381
2012 239.6 121.7 414835 412819 110.69 0.99514
2013 247.4 125.1 420779 425516 112.96 1.011258
2014 253.9 127.2 432660 446691 108.12 1.03243
2015 256.4 127.4 443590 463411 33.14 1.044683
Percentage Changes from previous year
2011 5.29 4.16 2.15 5.17 26.72 2.95
2012 3.77 3.49 1.53 2.32 14.68 0.79
2013 3.26 2.79 1.43 3.08 2.05 1.62
2014 2.63 1.68 2.82 4.98 -4.28 2.09
2015 0.98 0.16 2.53 3.74 -69.35 1.19

c. GDP covers all output in the economy including government and in-
vestment, the other two only cover consumers expenditure. RPI is based on
arithmetic means and includes housing CPI is based on geometric means and
excludes housing.
d. Give some examples such as the GDP deator is used for general
ination, e.g. deating government expenditure, the CPI is the basis for the
Bank of England target, the RPI is used for index linked gilts.
e. Give some general comments such as ination is steadily coming down
on CPI and RPI but GDP shows more variation. RPI is the highest of the three
in every year except 2015. The GDP deator is the lowest for the rst three
years. There does not seem much association with growth. Oil price increses at
the beginning may have increased ination and reduced it at the end.

4 Question 4
4.1 A
Suppose the long run relationship between the desired consumption, Ct , and
income, Yt ; is Ct = KY ; and actual consumption is determined by Ct =Ct 1 =
(Ct =Ct 1 ) eut , where ut is a white noise error independent of income.

10
 (a) Explain how you would convert the equation into a form that could be
estimated by least squares.
(b) Explain how you would calculate the long-run coe cient, , the speed
of adjustment, ; and the constant of proportionality K.
(c) Interpret and explain why one might expect it to take the value one.
(d) Explain how you would test whether the estimate of , b, is signicantly
dierent from one.
(e) Suppose people adjusted their hours of work to generate the income
required to nance their desired consumption. What problems would this cause
for least squares estimation of this consumption equation?
Answer
(a) Using lower case for logs, ct = ln Ct ; take logs to get

ct ct 1 = (k + yt ct 1) + ut ;
ct = 1 + 2 yt + 3 ct 1 + ut ;

which is a linear relationship that can be estimated by least squares.

(b) Use = 3; = 2= 3; k = 1 = 3 ; K = exp(k):
(c) is the long-run elasticity of consumption to income, if the elasticity is
one, the ratio of consumption to income is constant at K in the long run. Since
the consumption income ratio is quite constant one might expect = 1:
(d).To test whether is signicantly dierent from one, either impose the
restriction to give
ct = k + (yt ct 1 ) + ut
and use an F statistic to test this against the unrestricted model or reparame-
terise equation as

ct = k + (yt ct 1) + ( 1)ct 1 + ut

and use a t test on the coe cient of lagged consumption.

(e) If people adjusted their hours of work to generate the income required
to nance their desired consumption, this would make income is endogenous
correlated with the error causing the least squares estimator to be inconsistent.

4.2 B
The output of rm i = 1; 2; :::; N , Yi is produced by a Cobb-Douglas production
function of capital, Ki ; and labour, Li , and a random error term, ui , where exp
indicates exponential.
Yi = AKi Li exp(ui )
(a) How would you estimate the parameters by least squares?
(b) Explain why the estimate of ui can be interpreted as a measure of e -
ciency. If a rm was more e cient than average would the estimate be positive
or negative?
(c) Explain what and measure.

11
 (d) How would you test the hypothesis of constant returns to scale?
(e) Suppose more e cient rms were also more capital intensive. What eect
would this have on the least squares estimates of the production function?
Answer
(a). Take logarithms to give the linear relation

ln Yi = ln A + ln Ki + ln Li + ui

which can be estimated by least squares.

(b) The error ui measures how much more or less output the rm produces
than you would expect from its inputs. More e cient rms produce more than
you would expect so have a positive error.
(c) Elasticities of output with respect to labour and capital.
(d).CRS implies + = 1 so reparameterise as

ln Yi ln Li = ln A + (ln Ki ln Li ) + ( + 1) ln Li + ui

and do a t test that the coe cient of ln Li equals zero. Alternatively use
and F test and write out the restricted version, without ln Li and give F(1,N-3)
formula.
(e).The error is now correlated with ln Ki so estimates are biased and incon-
sistent.

4.3 C
Quarterly UK data for the change in the percentage unemployment rate, Ut ,
and the growth rate, at annual percent rates, Gt , were used to estimate the fol-
lowing regressions. Standard errors are given in parentheses, R2 is the coe cient
of determination and SER the standard error of regression.
For the period 1971Q4-1993Q1, 86 observations
Ut = 0:102 0:044 Gt +0:683 Ut 1 +0:010 Gt 1 +b et
(0:030) (0:009) (0:076) (0:010)
R2 = 0:83; SER=0.128.
For the period 1993Q2-2013Q1, 80 observations
Ut = 0:080 0:086 Gt +0:186 Ut 1 +0:039 Gt 1 +b et
(0:029) (0:020) (0:112) (0:020)
R2 = 0:49; SER=0.163.

(a) Explain why you would expect these variables to be related and comment
on the signicance of the coe cients.
(b) For each period calculate (i) the long-run eect of growth on the change
in unemployment and (ii) the growth rate at which unemployment is constant.
(c) Comment on the main dierences between the two periods, including the
speed of adjustment and the error variances.
(d) A Chow breakpoint test for structural stability gave a test statistic of
F(4,158)=7.22, p value 0.0000. Explain the hypothesis tested and whether the
relationship is stable.

12
 (e) Explain what the degrees of freedom of the F distribution, (4,158), mea-
sure and why they take those values.
Answer
(a) The variables are related because if output grows the demand for labour
increases and unemployment falls, perhaps with a lag. In the rst period every-
thing is signicant except lagged growth. In the second period the intercept
and growth are signicant the lagged change in unemployment is not and lagged
growth is just on the edge of signicance.
(b) (i) Call the long run eect in rst period 1 and second period 2 : 1 =(-
0.044+0.01)/(1-0.683) =-0.107; second period 2 =(-0.086+0.039)/(1-0.186) =-
0.0577 (ii) Similarly for g1 = 0.102/0.0343=3.4; second period g2 =0.080/0.047
=1.7.
(c) The speed of adjustment 0.317= 1-0.683 was much slower in the rst
period than the second 0.814=1-0.186. Although the impact eect of growth
was larger in the second period, the long run eect of growth was half as large
in the second period. The unexplained variance of unemployment was greater
in the second period 0.163 as compared to 0.128 and this is reected in a lower
R squared.
(d) The null hypothesis is that the coe cients in the two periods are the
same, (conditional on the variances being the same, which may not be the case
here). The null is rejected p value <0.05, so the relationship is not stable
between the two periods.
The rst degree of freedom, 4 is the number of restrictions tested, that
the four coe cients are the same in each period, the second is the number
of observations minus the number of parameters estimated in the unrestricted
model, 2 4 coe cients: 158=166-8.

5 Question 5
5.1 A
Suppose that E(Yi ) = ; V ar(Yi ) = E(Yi E(Yi ))2 = 2 ; with an independent
sample i = 1; 2; :::; N:
(a) Show this can be written Yi = + ui with E(ui ) = 0; E(u2i ) = P2
:
P (b) Derive the least squares
P estimator of say b that minimises b2i =
u
2
(Yi b ) : Show that (Yi b ) = 0:
(c) Show that b is unbiased: E(b ) = :
(d) Continuing to assume E(ui ) = 0; E(u2i ) = 2 ; E(uj uk ) = 0; j 6= k;
derive the variance of b : E(b E(b ))2 :
(e) How you would obtain an unbiased estimator of 2 ? You do not need to
prove this.
Answer
(a) Write Yi = E(Yi ) + (Yi E(Yi ) = + ui : E(ui ) = E [(Yi E(Yi )] =
E(Yi ) E(Yi ) = 0: E(u2i ) = E(Yi E(Yi ))2 = 2

13
 (b)
X X X X
S = b2i =
u (Yi b )2 = Yi2 + N b 2 2b Yi
@S X
= 2N b 2 Yi = 0
@b X
b = Yi =N

The second line can be written

X
2 (Yi b) = 0
P
implying (Yi b ) = 0:
(c)
X X X X
b = Yi =N = b = ( + ui )=N = N =N + ui =N = + ui =N
X
E(b ) = + E( ui =N ) =

Since E(ui ) = 0:
(d) X
E(b E(b ))2 = E(b )2 = E( ui =N )2
from the rst line of (c)
P 2
[ ui ] [u1 + u2 + :: + uN ] [u1 + u2 + :: + uN ]
E = E
N2 N2
2 2 2
u + u2 + :: + uN + u1 u2 ::: N 2 2
= E 1 2
= 2
=
N N N
P
(e) The unbiased estimator of 2 is given by s2 = ub2i =(N 1):

5.2 B
Consider the model for t = 1; 2; :::; T

yt = xt + ut

where the observed dependent variable, yt ; and the observed independent vari-
able xt ; are measured as deviations from their means, is an unknown parameter
and ut an unobserved disturbance, with V ar(xt ) > 0.
(a) Derive the least squares estimator b both by method of moments from
the moment condition E(xt ut ) = 0 and from minimising the sum of squared
errors. What is the relationship between the two procedures?
(b) Suppose that in addition to E(xt ut ) = 0; E(ut ) = 0; E(u2t ) = 2 ;
E(ut ut i ) = 0; Show that b is unbiased and derive its variance.
(c) Under the assumptions of (b) how would you estimate the standard error
b
of ?

14
 (d) Suppose the other assumptions held but variance was not constant, but
E(u2t ) = 2 x2t ; what is the variance of b ?
(e) Under the assumptions of (d) how would you repond if you knew that
E(u2t ) = 2 x2t :
Answer
(a) Method of moments matches the sample
X X
bt =
xt u xt (yt b xt ) = 0
X X
xt yt b x2t = 0 (1)
X X
b = xt yt = x2t

Least Squares chooses b to minimise

X X X 2X 2 X
S = b2t =
u (yt b xt )2 = yt2 + b xt 2b xt yt
@S X X
= 2b x2t 2 xt yt = 0 (2)
@b
X X
b = xt yt = x2 t

@2S X
2 = 2 x2t > 0
@b
The First order condition for least squares, (2) ; is just the moment condition,
(1) multiplied by 2: So they give the same result.
(b) Since
P P P
b = Pxt yt = xt ( xt + ut )
P = + P 2
xt ut
x2t x2t xt
then E( b ) = ; and it is unbiased; since because of independence
P P
xt ut xt
E P 2 = E P 2 E(ut )
xt xt
and E(ut ) = 0: To derive the variance of b ; note since b is unbiased and treating
xt as xed
P
b b xt ut
V ( ) = E( 2
) = E( P 2 )2 (3)
xt
1 X 2
= P 2 2E xt ut
( xt )
P 2
We can write E ( xt ut ) as
E(x1 u1 + x2 u2 ::: + xT uT )(x1 u1 + x2 u2 ::: + xT uT )
E(x21 u21 + x22 u22 + ::: + x2T u2T + 2x1 u1 x2 u2 + ::::)
x21 2
+ x2 2
+ ::: + x2T 2
+ 0 + ::::
X 2
2
x2t

15
since for t = 1; 2; :::; T; E(u2t ) = 2
and E(ut ut i ) = 0:So
1 X
V (b) = P 2 2
2
x2t
( xt )
2
= P
x2t

(c) Estimate the variance by

X
s2 = b2t =(T
u 2)
P
then the estimated standard error of b is se( b ) = s= x2t :
(d) From above

E(x21 u21 + x22 u22 + ::: + x2T u2T + 2x1 u1 x2 u2 + ::::)

= x21 2 2
x + x22 2 2
x1 + ::: + x2T 2 2
x1 +0+
X1
2
x4t

1 X
V (b) = P 2
2
x4t
( x2t )
Doesnt cancel.
(e) Divide equation by xt
yt xt ut
= +
xt xt xt
with
2 2
ut 2 x
E( ) = 2t = 2
xt xt
constant variance.

6 Question 6
6.1 A
Data for 189 countries were used to explain life expectancy in each country, LE,
which ranged from 48.3 to 83.4 years by the logarithm of per capita income,
LPCI. The data were ordered by LPCI starting with the lowest and going to
the highest. The EViews output is given below.

16
 Dependent Variable: LE
Method: Least Squares
Date: 02/06/15 Time: 10:43
Sample: 1 189
Included observations: 189

Variable Coefficient Std. Error t-Statistic Prob.

C 24.27829 2.568390 9.452729 0.0000

LPCI 5.172650 0.279985 18.47472 0.0000

R-squared 0.646045 Mean dependent var 71.28771

Adjusted R-squared 0.644152 S.D. dependent var 8.050099
S.E. of regression 4.802125 Akaike info criterion 5.986520
Sum squared resid 4312.296 Schwarz criterion 6.020824
Log likelihood -563.7261 Hannan-Quinn criter. 6.000417
F-statistic 341.3155 Durbin-Watson stat 1.908097
Prob(F-statistic) 0.000000

Analysis of the residuals showed a skewness coe cient of -1.32 and a kurtosis
coe cient of 5.76 and some large outliers. Dummy variables were added for
the observations corresponding to the four largest residuals: 39 (Lesotho), 71
(Swaziland), 129 (Gabon) and 162 (Equatorial Guinea) . The results are shown
below. This regression had skewness of -0.64 and a kurtosis of 3.3

17
Dependent Variable: LE
Method: Least Squares
Date: 02/09/16 Time: 09:59
Sample: 1 189
Included observations: 189

Variable Coefficient Std. Error t-Statistic Prob.

C 24.26248 2.187408 11.09188 0.0000

LPCI 5.216575 0.238511 21.87140 0.0000
D39 -16.63978 4.082461 -4.075918 0.0001
D71 -19.06107 4.071816 -4.681222 0.0000
D129 -16.32016 4.074528 -4.005411 0.0001
D162 -20.43737 4.085622 -5.002266 0.0000

R-squared 0.752412 Mean dependent var 71.28771

Adjusted R-squared 0.745648 S.D. dependent var 8.050099
S.E. of regression 4.059935 Akaike info criterion 5.671442
Sum squared resid 3016.402 Schwarz criterion 5.774355
Log likelihood -529.9513 Hannan-Quinn criter. 5.713135
F-statistic 111.2264 Durbin-Watson stat 1.858698
Prob(F-statistic) 0.000000

(a) What values would you expect for the coe cients of skewness and
kurtosis if the distribution was normal? Comment on the skewness and kurtosis
in the two regressions. Did including the dummy variables make the distribution
of the residuals more normal?
(b) What eect does non-normality of the errors have on the properties
of the least squares estimate of the coe cient of LPCI? Are the results consistent
with this theoretical result?
(c) Explain how the dummy variables for the outlier observations are
constructed.
(d) Explain how you would interpret the coe cients of the dummy vari-
ables.
(e) Explain what the Durbin-Watson statistic tells you in this case where
the data are not time series but ordered by the independent variable.
Answer
(a) You would expect skewness equal to zero and kurtosis equal to 3.
There is negative skewness and excess kurtosis in both cases, but it is less after
the inclusion of the dummies, so closer to normal.
(b) Normality is not an assumption required for the estimates to be
BLUE so least squares retains their properties. The results are consistent with
this since the coe cient of LPCI hardly changes when the dummies are included.

18
 (c) The dummy variable for observation 39 takes the value zero for each
observation except 39 where it takes the value one.
(d) It says that life expectancy in Lesotho is 16.6 years less than would
be expected in a country with a similar income.
(e) Since the data are ordered by income the DW could pick up non-
linearity, correlation between errors in closely related income levels.

6.2 B
You have data on average earnings by age and education in the US in 1991 for
30 groups of men, i = 1; 2; :::; N: Dene wi the logarithm of earnings, Ai age in
years (divided by 100), Ei years (divided by 100) of education. It is suggested
that an appropriate model is:
2
wi = + 1 Ai + 2 Ei + 1 Ai + 2 Ai Ei + ui :

The estimated coe cients, with their standard errors in parentheses are:
bi = 7:46 +6:6Ai +9:0Ei
w 6:7A2i +4:1Ai Ei
(0:19) (0:62) (1:01) (0:56) (1:95)

R2 = 0:99; SER = 0:05:

(a) What is the stochastic component in this model? How would you inter-
pret it?
(b) Explain the role of the A2i term?
(c) Comment on these results. Are the signs what you would expect?
(d) Given these estimates of the parameters at what age do men with zero
and twenty years of education earn the maximum amount? Are these sensible
numbers?
Answer
(a) The stochastic component is ui it is the part of log earnings not explained
by the regressors, will reect unmeasured ability and other factors inuencing
eartnings.
(b) We would expect earnings to rise and then fall with age, the quadratic
terms captures this feature.
(c) Yes. For earnings to rise and fall with age, we need 1 > 0 and 1 < 0:
You earn more with better education so 2 > 0: The positive coe cient on the
interaction term 2 makes peak earnings later for more highly educated men,
which is likely.
(d) To get the maximum

@w
= 1 +2 1A + 2E =0
@A
1
A = (2 1) ( 1 + 2 E)
= 0:0746(6:6 + 4:1E)
= 0:49 + 0:31E

19
note Ai and Ei are divided by 100. For Ei = 0 max w =100x0.49=49 years. For
Ei = 0:2 w=100x(0.49+0.31x0.2)=55 years. Although the idea of somebody
with zero years of education is unlikely, these are plausible values for peak
earning age.

6.3 Question 6C.

Below are given the least squares estimates of a Phillips Curvefor two periods.
The dependent variable, INF is ination measured as 100 times the change in
the logarithm of the GDP deator over four quarters; RU is the reciprocal of
the percentage unemployment rate, U; INF(-1) means INF lagged one quarter.

(a) Why do you think the reciprocal of the unemployment rate was used
rather than the unemployment rate? How do you interpret the negative coef-
cient on INF(-4)?
(b) In the rst period what is (i) the long-run eect of RU on INF (ii) the
long run eect of U on INF when U=1 and when U=9?
(c) A Chow breakpoint test with the null hypothesis of no break at 1992Q4
gave an F(5,162) statistic with a p value of 0.0008. What is the null hypothesis?
Is it rejected? Explain the degrees of freedom of the test.
(d) What were the main dierences between the two periods? Comment on
the mean, the inuence of unemployment, the persistence of ination and the
volatility of ination.

20
Sample: 1971Q2 1992Q4 included observations 87 Dep Var INF

Variable Coefficient Std. Error t-Statistic Prob.

C -0.435689 0.456684 -0.954027 0.3429

RU 8.558871 2.813995 3.041537 0.0032
INF(-1) 1.141808 0.052029 21.94573 0.0000
INF(-4) -0.654936 0.114733 -5.708359 0.0000
INF(-5) 0.413743 0.095748 4.321145 0.0000

R-squared 0.932439 Mean dependent var 9.059201

Adjusted R-squared 0.929143 S.D. dependent var 5.373721
S.E. of regression 1.430429 Akaike info criterion 3.609579
Sum squared resid 167.7824 Schwarz criterion 3.751298
Log likelihood -152.0167 Hannan-Quinn criter. 3.666645
F-statistic 282.9284 Durbin-Watson stat 2.105356
Prob(F-statistic) 0.000000

Sample: 1993Q1 2014Q1 Included observations: 85 Dep Var INF

Variable Coefficient Std. Error t-Statistic Prob.

C 0.509042 0.398926 1.276033 0.2056

RU 2.226641 1.805614 1.233177 0.2211
INF(-1) 0.647122 0.085640 7.556296 0.0000
INF(-4) -0.461409 0.095570 -4.827967 0.0000
INF(-5) 0.402966 0.089375 4.508708 0.0000

R-squared 0.471667 Mean dependent var 2.110508

Adjusted R-squared 0.445251 S.D. dependent var 0.772657
S.E. of regression 0.575486 Akaike info criterion 1.789820
Sum squared resid 26.49477 Schwarz criterion 1.933505
Log likelihood -71.06735 Hannan-Quinn criter. 1.847614
F-statistic 17.85493 Durbin-Watson stat 2.014359
Prob(F-statistic) 0.000000

21
 Answer
(a) (i) to allow unemployment to have a large eect on ination when un-
employment was low and a small eect when unemployment was high. (ii) it
may be induced by the fact that ination is measured as lnP(t)-lnP(t-4).
(b) (i) The sum of the coe cients on lagged ination is 0.90, so the long run
eect of RU = U 1 is = 8:56=(1 0:9) = 85:6. (ii) The unemployment term in
the long-run is U 1 so the derivative of ination with respect to unemployment
in the long-run is U 2 so at U=1% it is 85.6, while at U=9% it is 85:6=92 =
85:6=81 = 1:06. The eect of a change in unemployment on ination is very
large when unemployment is very low and quite small when unemployment is
high.
(c ) The null hypothesis is that the ve coe cients in the rst period equal
the ve coe cients in the second period. It is rejected. The 5 refers to the
5 restrictions that the coe cients are equal, 162=(85+87-10) the degrees of
freedom of the unrestricted model.
(d) Both the unconditional mean and variance of ination were much higher
in the rst period, as was the conditional variance, the standard error falling
from 1.4 to 0.5. Persistence (sum of coe cients of lagged ination) dropped
from 0.9 in the rst period to 0.59 in the second period. The short-run (and
long run) eect of unemployment reduced, becoming insignicant in the second
period.

22