CHAPTER

Exercise Solutions

200

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

EXERCISE 9.1
From the equation for the AR(1) error model et = et 1 + vt , we have

var ( et ) = 2 var ( et 1 ) + var ( vt ) + 2 cov ( et 1 , vt )

from which we get
e2 = 2 e2 + v2 + 0

e2 (1 2 ) = v2
and hence

e2 =

v2
1 2

To find E ( et et 1 ) we note that

et et 1 = et21 + et 1vt
Taking expectations,

E ( et et 1 ) = E ( et21 ) + 0 = e2
Similarly,

et et 2 = et 1et 2 + et 2 vt
and

E ( et et 2 ) = E ( et 1et 2 ) + 0 = E ( et et 1 ) = 2 e2

201

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

202

EXERCISE 9.2
(a)

Using hand calculations

T

r1 =

et et 1
t =2
T

et21
t =2

(b)

(i)

0.0979
= 0.1946 ,
0.5032

r2 =

et et 2
t =3
T

et2 2

0.1008
= 0.2007
0.5023

t =3

The test statistic for testing H 0 : 1 = 0 against the alternative H1 : 1 0 is

Z = T r1 = 10 0.1946 = 0.6154 . Comparing this value to the critical Z values for a

two tail test with a 5% level of significance, Z (0.025) = 1.96 and Z (0.975) = 1.96 , we do
not reject the null hypothesis and conclude that r1 is not significantly different from
zero.
(ii) The test statistic for testing H 0 : 2 = 0 against the alternative H1 : 2 0 is

Z = T r2 = 10 0.2007 = 0.6347 . Comparing this value to the critical Z values for a

two tail test with a 5% level of significance, Z (0.025) = 1.96 and Z (0.975) = 1.96 , we do
not reject the null hypothesis and conclude that r2 is not significantly different from
zero.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

203

EXERCISE 9.3
(a)

Equation (9.49) can be used to conduct two Lagrange multiplier tests for AR(1) errors.
The first test is to test whether the coefficient for et 1 is significantly different from zero.
The null hypothesis is H 0 : = 0. The value of the test statistic is
t=

0.428
= 2.219
0.201

The critical t value for a 5% level of significance and 23 degrees of freedom is 2.069.
Since 2.219 > 2.069, we reject the null hypothesis and conclude that first order
autocorrelation is present.
The second LM test examines whether or not the R 2 from (9.49) is significant. The null
hypothesis is again H 0 : = 0 and the test statistic value is

LM = T R 2 = 26 0.165 = 4.29
2
-distribution with a 5% critical value of
When the null hypothesis is true, LM has a (1)

3.84. Since 4.29 > 3.84, we reject the null hypothesis and conclude that first order
autoregressive errors exist.
(b)

Ignoring autocorrelation means the estimates will be unbiased but no longer the best
estimates. Furthermore, the standard errors will be incorrect resulting in misleading
confidence intervals and hypothesis tests. The standard errors from the model with AR(1)
errors are larger than the standard errors from the least-squares estimated model. Thus, it
is likely that the least squares estimates and standard errors have overstated the precision
of the estimates in the relationship between disposer shipments and durable goods
expenditure. If autocorrelation is ignored, the confidence intervals will be narrower than
the correct confidence intervals, and hypothesis tests will have a probability of a type 1
error that is greater than the specified significance level.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

204

EXERCISE 9.4
(a)

(b)

Under the assumptions of the AR(1) model, corr(et , et k ) = k . Thus,

(i)

corr(et , et 1 ) = = 0.9

(ii)

corr(et , et 4 ) = 4 = 0.94 = 0.6561

(iii)

e2 =

(i)

corr(et , et 1 ) = = 0.4

(ii)

corr(et , et 4 ) = 4 = 0.44 = 0.0256

(iii)

e2 =

v2
1
=
= 5.263
1 2 1 0.92

v2
1
=
= 1.190
2
1 1 0.42

When the correlation between the current and previous period error is weaker, the
correlations between the current error and the errors at more distant lags die out relatively
quickly, as is illustrated by a comparison of 4 = 0.6561 in part (a)(ii) with 4 = 0.0256 in
part (b)(ii). Also, the larger the correlation , the greater the variance e2 , as is illustrated
by a comparison of e2 = 5.263 in part (a)(iii) with e2 = 1.190 in part (b)(iii).

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

205

EXERCISE 9.5
(a)

(b)

(i)

eT +1 = eT

(ii)

eT + 2 = eT +1 = 2 eT

Equation (9.25) gives us the nonlinear least squares estimates of the coefficients
1 = 3.89877 and 2 = 0.88837 . The final observation in bangla.dat is A34 = 53.86,
P34 = 0.89 . Therefore, the nonlinear least squares residual for the last observation is

e34 = ln ( 53.86 ) 3.89877 0.88837 ln ( 0.89 ) = 0.19114

(c)

Forecasts for eT +1 and eT + 2 are given by

eT +1 = eT = 0.42214 0.19114 = 0.08069

eT + 2 = eT +1 = 0.42214 0.08069 = 0.03406
(d)

Noting that eT +1 = eT , the forecast value for ln( AT +1 ) can be calculated as

ln( AT +1 ) = 1 + 2 ln( PT +1 ) + eT ,
= 3.89877 + 0.88837 ln(1) + 0.08069 = 3.97946
Similarly, the forecast value for ln( AT + 2 ) is

ln( AT + 2 ) = 1 + 2 ln( PT + 2 ) + eT + 2
= 3.89877 + 0.88837 ln(1.2) + 0.03406 = 4.09480

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

206

Exercise 9.5 (continued)

(e)

In Chapter 4 we are told that there are two ways to forecast a dependent variable when the
left-hand side of the equation is in the form of the logarithm of that variable. The first
method is to calculate the natural predictor y n , which is the better predictor to use when
the sample size is small

y n = exp(b1 + b2 x)
The second method is to calculate a corrected predictor y c , which is the better predictor
to use when the sample size is large

y c = exp(b1 + b2 x + 2 / 2) = y n e

/2

Applied to our nonlinear least squares estimation, the natural predictors are

AT +1 = exp 1 + 2 ln( PT +1 ) + eT +1

= exp ( 3.89877 + 0.88837 ln(1) + 0.08069 )

= 53.488

AT + 2 = exp 1 + 2 ln( PT + 2 ) + eT + 2

= exp ( 3.89877 + 0.88837 ln(1.2) + 0.03406 )

= 60.027
The corrected predictors are

AT +1 = exp 1 + 2 ln( PT +1 ) + eT +1 + v2 2

= exp ( 3.89877 + 0.88837 ln(1) + 0.08069 + 0.28542 2 )

= 55.711

AT + 2 = exp 1 + 2 ln( PT + 2 ) + eT + 2 + v2 2

= exp ( 3.89877 + 0.88837 ln(1.2) + 0.03406 + 0.28542 2 )

= 62.523

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

207

EXERCISE 9.6
We consider two ways to derive the lag weights, by recursive substitution and by equating
coefficients of the lag operator. Recursive substitution is tedious but does not require new
machinery. Using the lag operator requires new machinery, but is less tedious.
Recursive substitution
One way to find the required expressions for the lag weights is to use recursive
substitution on the ARDL model. Once we have substituted in enough lagged equations,
we can determine the lag weights by observation. Recursive substitution begins with the
current ARDL model

yt = + 3 xt 3 + 1 yt 1 + 2 yt 2 + 3 yt 3 + vt

(1)

We lag the equation by 1 period

yt 1 = + 3 xt 4 + 1 yt 2 + 2 yt 3 + 3 yt 4 + vt 1

(2)

and then substitute it back into (1)

yt = + 3 xt 3 + 1 ( + 3 xt 4 + 1 yt 2 + 2 yt 3 + 3 yt 4 + vt 1 )
+2 yt 2 + 3 yt 3 + vt
= + 3 xt 3 + 1 + 1 3 xt 4 + 12 yt 2 + 1 2 yt 3 + 1 3 yt 4

(3)

+1 vt 1 + 2 yt 2 + 3 yt 3 + vt
This process is repeated with the larger lags until the required xt s is reached. In this
model, we stop the process of recursive substitution after substituting in yt 3 . At this stage
it should be clear that further substitution would not involve any additional lags of the
independent variables xt s for s = 1, 2,3, 4,5 or 6 . This ensures that the expressions for the
lag weights that we determine will not change with further substation. Rearranging the
final equation should give an expression similar to

yt = + 1 + 12 + 13 + 12 + 2 + 12 + 3
+3 xt 3 + 13 xt 4 + 12 3 xt 5 + 2 3 xt 5 + 133 xt 6
+12 3 xt 6 + 12 3 xt 6 + 3 3 xt 6 + 14 yt 4 + 12 2 yt 4
+12 2 yt 4 + 12 2 yt 4 + 13 yt 4 + 22 yt 4 + 13 yt 4

(4)

+132 yt 5 + 133 yt 6 + 13vt 3 + 12 3 yt 5 + 122 yt 5

+122 yt 5 + 2 3 yt 5 + 2 3 yt 5 + 2 3 yt 5 + 12 3 yt 6
+32 yt 6 + vt + 1vt 1 + 12 vt 2 + 2 vt 2 + 12 vt 3 + 12 vt 3 + 3vt 3
By observation of (4) we can group the coefficients for each xt s remembering that s is
defined as the coefficient of xt s . Therefore,

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

208

Exercise 9.6 (continued)

0 = 1 = 2 = 0
3 = 3
4 = 13 = 13
5 = 12 3 + 2 3 = 14 + 23
6 = 133 + 12 3 + 12 3 = 15 + 24 + 33
= 1s 1 + 2 s 2 + 3s 3
Also the last result shows that

s = 1s 1 + 2s 2 + 3 s 3 , for s 6
Using the lag operator
The lag operator L operates on a variable, say yt , such that Lyt = yt 1 . Although it will
seem like magic to you at first, it is possible to do algebra with the lag operator. In
particular, raising L to the power s, written Ls has the effect of lagging yt s times. That is,
Ls yt = yt s . With this little bit of knowledge, the model

yt = + 3 xt 3 + 1 yt 1 + 2 yt 2 + 3 yt 3 + vt
can be written as

(1 1 L 2 L2 3 L3 ) yt = + 3 L3 xt + vt
A bit more faith is required for the next step where we invert the left-hand side function of
the lag operator to obtain
yt = (1 1 L 2 L2 3 L3 ) 1 ( + 3 L3 xt + vt )
Now consider the infinite lag representation

yt = + s xt s + et = + 0 xt + 1 xt 1 + 2 xt 2 + 3 xt 3 + 4 xt 4 +
s =0

+ et

It can be written in terms of the lag operator as

yt = + (0 + 1 L + 2 L2 + 3 L3 + 4 L4 +

) xt + et

We now have two different equations for the same model, where yt is the left-hand side
variable for both of them. It follows that the right-hand sides must be equal
(1 1 L 2 L2 3 L3 ) 1 ( + 3 L3 xt + vt ) = + (0 + 1 L + 2 L2 + 3 L3 + 4 L4 +

) xt + et

and that

( + 3 L3 xt + vt ) = (1 1 L 2 L2 3 L3 ) + (0 + 1 L + 2 L2 + 3 L3 + 4 L4 +
The left hand side can be written as

) xt + et

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

209

Exercise 9.6 (continued)

+ (0 L0 + 0 L1 + 0 L2 + 3 L3 + 0 L4 + 0 L5 + 0 L6 ) xt + vt
Although they have zero coefficients, we have included all powers of the lag operator up
to lag 6. As we will see, it is useful to do so. Now consider the right hand side. After
multiplying out and collecting terms, we have

(1 1 2 3 ) + (0 L0 + (1 10 ) L + (2 20 11 ) L2
+ (3 30 21 12 ) L3 + (4 31 22 13 ) L4
+(5 32 23 14 ) L5 + (6 33 24 15 ) L6 +

xt

+ (1 1 L 2 L2 3 L3 )et
Equating coefficients of like terms establishes the relationship between the ARDL model
and its infinite lag representation. For the constant and error term, we have

= (1 1 2 3 )
vt = (1 1 L 2 L2 3 L3 )et = et 1et 1 2 et 2 3et 3
For the lag weights we equate coefficients of equal powers of the lag operator

0 = 0
0 = 1 10
0 = 2 20 11
3 = 3 30 21 12
0 = 4 31 22 13
0 = 5 32 23 14
0 = 6 33 24 15
From these expressions we obtain

0 = 1 = 2 = 0
3 = 3
4 = 13
5 = 23 + 14
6 = 33 + 24 + 15
s = 3s 3 + 2s 2 + 1 s 1

for s 6

For this process to work, and for the lag weights to be valid, coefficients for long lags
must converge to zero.
The lag operator method will seem daunting at first, but it is worth the investment. You
can go crazy doing recursive substitution.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

210

EXERCISE 9.7
(a)

The forecast DIP for January 2006 is

DIPT +1 = 0.109 + 0.033DIPT + 0.236 DIPT 1 + 0.200 DIPT 2

= 0.109 + 0.033 1.041 + 0.236 1.006 + 0.200 1.221
= 0.6250
The forecast DIP for February 2006 is

DIPT + 2 = 0.109 + 0.033DIPT +1 + 0.236 DIPT + 0.200 DIPT 1

= 0.109 + 0.033 0.625 + 0.236 1.041 + 0.200 1.006
= 0.5765
The forecast DIP for March 2006 is

DIPT + 3 = 0.109 + 0.033DIPT + 2 + 0.236 DIPT +1 + 0.200 DIPT

= 0.109 + 0.033 0.5765 + 0.236 0.625 + 0.200 1.041
= 0.4837
(b)

To find the 95% confidence intervals, we first find the forecast error standard errors using
the expressions derived in Section 9.5

1 = v = 0.4293

2 = v 1 + 12 = 0.4293 1 + 0.0332 = 0.42953

3 = v

2
1

+ 2

+ 12 + 1 = 0.4293

( 0.033

+ 0.236 ) + 0.0332 + 1 = 0.44143

2

The model was estimated using monthly observations from January, 1985 to December,
2005, a total of 252 observations and 248 degrees of freedom. Confidence intervals were
constructed using the t-value t(0.975,248) = 1.9696 .

The confidence intervals, given by the expression DIPT + j j 1.9696 , are:

Month

DIPT + j

Lower bound

Upper bound

January
February

0.6250
0.5765

0.42930
0.42953

0.221
0.269

1.471
1.422

March

0.4837

0.44143

0.386

1.353

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

211

EXERCISE 9.8
Equation (9.28) is the estimated version of the model

yt = + 0 xt + 1 xt 1 + 1 yt 1 + vt
In equation (9C.6) on page 266, expressions for the lag weights of the more general model

yt = + 0 xt + 1 xt 1 + 2 xt 2 + 3 xt 3 + 1 yt 1 + 2 yt 2 + vt
are given. Expressions for the lag weights for (9.28) can be obtained from the more
general ones by setting 2 = 3 = 2 = 0 . Doing so yields
Weight

Estimate

0 = 0
1 = 1 + 10
2 = 11
3 = 12
4 = 13
5 = 14
6 = 15

0.7766
0.2969
0.1200
0.0485
0.0196
0.0079
0.0032

From the lag weight distribution, we can see that the immediate effect of a temporary 1%
increase in the price of sugar cane is an increase in the area planted of 0.77%. In
subsequent periods there are negative effects on the amount of area planted. One period
after the temporary price increase the area planted decreases by 0.30%, the second period
lagged effect is a decreases of 0.12%, the third period lagged effect is a decrease of 0.05%,
the fourth period lagged effect is a decrease of 0.02%, and the fifth period lagged effect is
a negligible decrease of less than 0.01%. This estimated lag weight distribution suggests
that producers initially overreact to the price change. If the price increase was a sustained
one, the final equilibrium change in area would be less than that which occurred in the
current period.
Notice that we predict that producers will initially overreact to a price change because
< . If > , their initial response is one of under reaction. If = , we
1

1 0

1 0

1 0

have the AR(1) error model. There is no lagged response to price by itself, but there is a
lagged response to the error in the previous period.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

212

EXERCISE 9.9
(a)

The least-squares estimated equation is

ln( JVt ) = 3.503 1.612ln(U t )

(se) (0.283) (0.156)
A 95% confidence interval for 2 is

b2 t(0.975, 22) se(b2 ) = 1.6116 2.074 0.1555 = (1.934, 1.289)

The LM test for H 0 : = 0 can be conducted as a test for the significance of from the
equation

(b)

et = 1 + 2 ln(U t ) + et 1 + vt
or by using the statistic LM = T R 2 from this equation. The value of the F-statistic for
testing the significance of is F = 5.047 with a p-value 0.036. Also, LM = T R 2 =
24 0.19376 = 4.650 with a p-value 0.031. Since both p-values are less than 0.05, we
reject H 0 : = 0 at a 5% significance level and conclude that autocorrelation exists. The
existence of autocorrelation means the assumption that the et are independent is not
correct. This problem causes the confidence interval for 2 in part (a) to be incorrect; it
could convey a false sense of the reliability of b2
.
(c)

The re-estimated model under the assumption of AR(1) errors is

ln( JVt ) = 3.503 1.600ln(U t )

(se) (0.249) (0.132)

= 0.4486
(0.2029)

The 95% confidence interval for 2 is

b2 t(0.975, 22) se(b2 ) = 1.6001 2.074 0.1315 = ( 1.873, 1.327)

This confidence interval is slightly narrower than that given in part (a). A direct
comparison with the interval in part (a) is difficult because the least squares standard
errors are incorrect in the presence of AR(1) errors. However, one could conjecture that
the least squares confidence interval is larger than it should be implying unjustified
imprecision.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

213

Exercise 9.9(c) (continued)

The correlogram of residuals from the AR(1) error model is given below. To decide
whether these correlations are significant, they are compared against the 5% significance
bounds 1.96 24 = 0.40 . None of the correlations exceed these bounds, suggesting
that the AR(1) error model has successfully modeled the autocorrelation in the errors.
.6

Correlations

.4
.2
.0
-.2
-.4
-.6
2

10

12

Lag

Figure xr9.9 Correlogram of residuals for Exercise 9.9 part (c)

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

214

EXERCISE 9.10
(a)

The estimated equation with standard errors in parentheses is

ln ( JVt ) = 1.9438 1.6028ln (U t ) + 0.7136ln (U t 1 ) + 0.4490ln ( JVt 1 )

(se)

( 0.8150 ) ( 0.1563)

( 0.3622 )

( 0.2085)

Testing the hypothesis H 0 : 1 = 10 , we find the F-value to be 0.00115 with a

corresponding p value of 0.9733. Since the p-value is greater than the level of significance
0.05, we do not reject the null hypothesis. Concluding that 1 = 10 implies the AR(1)
error model is a reasonable one. The restricted model that incorporates the restriction
1 = 10 is
ln ( JVt ) = + 0 ln (U t ) 10 ln (U t 1 ) + 1 ln ( JVt 1 ) + vt
This model is the transformed nonlinear least squares version of the AR(1) error model in
Exercise 9.9.
(b)

Forecasts for the next two periods are given by

ln ( JVT +1 ) = 1.9438 1.6028ln (U T +1 ) + 0.7136ln (U T ) + 0.4490ln ( JVT )

= 1.9438 1.6028ln ( 5 ) + 0.7136ln ( 3.72 ) + 0.4490ln ( 5.12 )
= 1.035
ln ( JVT + 2 ) = 1.9438 1.6028ln (U T + 2 ) + 0.7136ln (U T +1 ) + 0.4490ln ( JVT +1 )
= 1.9438 1.6028ln ( 5 ) + 0.7136ln ( 5 ) + 0.4490 1.035
= 0.977

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

215

EXERCISE 9.11
(a)

The marginal cost is the first derivative of the total cost with respect to quantity

dTC
= 2 + 23Q = MC
dQ
The marginal revenue is the first derivative of the total revenue with respect to quantity

dTR
= 1 + 22Q = MR
dQ
(b)

To find the profit maximising quantity Q , equate the marginal revenue and the marginal
cost
MR = MC

1 + 22Q = 2 + 2 3Q

Rearranging so that Q is the subject gives

2 1
Q =
2 ( 2 3 )
(c)

The estimated least squares regression for the total revenue function with standard errors
in parentheses is

TR = 174.2803Q 0.5024Q 2

( 4.5399 ) ( 0.0235)
The estimated least squares regression for the total cost function with standard errors in
parentheses is

TC = 2066.083 1.5784Q + 0.2277Q 2

( 727.2180 ) ( 9.4524 ) ( 0.0289 )

These estimates are appropriate under the multiple regression model assumptions. In
particular, the error term must have an expected value of zero, E (et ) = 0, must be
homoskedastic, var(et ) = 2 ,
uncorrealted, cov(et , es ) = 0.

and

the

errors

in

different

Using these estimates, the profit maximising level of output is

Q =

a2 b1
1.5784 174.2803
=
= 120
2 ( b2 a3 ) 2 ( 0.5024 0.2277 )

months

must

be

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

216

Exercise 9.11 (continued)

(d)

Assuming that the level of production for the next three months is based on the profit
maximising level of output
TR = 174.2803 120 0.5024 120 2 = 13679
TC = 2066.083 1.5784 120 + 0.2277 1202 = 5155

PROFIT = TR TC = 13679 5155 = 8524

Therefore, for the next three months we forecast total revenue to be 13,679, the total cost
to be 5,155, and the profit per month as 8,524.
Performing LM tests for AR(1) errors in the total cost function yields F = 8.553 (p-value
= 0.0054) on the significance of and LM = T R 2 = 7.812 (p-value = 0.0052) using the

(e)

2 statistic. The same tests for AR(1) errors in the total revenue function yield F = 113
(p-value = 0.0000) on the significance of and LM = T R 2 = 34.4 (p-value = 0.0000)

using the 2 statistic. We conclude that both functions have correlated errors.
Examination of the correlograms of the residuals confirms this conclusion. The
significance bounds in the figures below are at 1.96 48 = 0.283 . We find that there
are several statistically significant correlations that exceed these bounds. In particular, r1 of
the total cost model and r1 , r2 , r3 and r4 of the total revenue model are statistically
significant. They also lead us to conclude that the errors are correlated.
Total revenue

.4

1.0

.3

0.8

.2

0.6
Correlations

Correlations

Total cost

.1
.0

0.4
0.2

-.1

0.0

-.2

-0.2

-.3

-0.4
2

6
Lag

10

12

Lag

Figure xr9.11 Residual correlograms for total cost and total revenue functions

10

12

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

217

Exercise 9.11 (continued)

(f)

The re-estimated equations assuming AR(1) errors are

TR = 171.58Q 0.5085Q 2
(se)

( 8.01) ( 0.0248)

TC = 2354.71 5.7521Q + 0.2415Q 2

(se)
(g)

( 0.0634 )

et = 0.4595et 1 + vt

( 0.1521)

The profit maximising level of output suggested by the results in part (f) is

Q =

(h)

( 616.01) (8.0203) ( 0.0247 )

et = 0.9495et 1 + vt

a2 b1
5.7521 171.5814
=
= 118
2 ( b2 a3 ) 2 ( 0.5085 0.2415 )

In this case, because the errors are assumed autocorrelated, the total revenue and total cost
errors for month 48 have a bearing on the predictions, and the predictions will be different
in each of the future three months.
For the total revenue function, the estimated error for month 48 is

eTR ,48 = 8435 (171.5814 83 0.5085 (83) 2 ) = 2303.29

Therefore, given Q = 118 for the next three months, the total revenue predictions for the
next three months are given by

TRT +1 = TR 49 = 171.5814 118 0.5085 1182 + 0.9495 (2303.29) = 10979

TRT + 2 = TR 50 = 171.5814 118 0.5085 1182 + 0.94952 (2303.29) = 11090
TRT +3 = TR 51 = 171.5814 118 0.5085 1182 + 0.94953 (2303.29) = 11195
For the total cost function, the error for the last sample month is
eTC ,48 = 4829 (2354.71 5.7521(83) + 0.2415(832 )) = 1287.71

Therefore, given Q = 118 for the next three months, the total cost predictions for the next
three months are given by

TC T +1 = TC 49 = 2354.71 5.7521 118 + 0.2415 1182 + 0.4595 1287.71 = 5631

TC T + 2 = TC 50 = 2354.71 5.7521 118 + 0.2415 1182 + 0.45952 1287.71 = 5311
TC T +3 = TC 51 = 2354.71 5.7521 118 + 0.2415 1182 + 0.45953 1287.71 = 5164

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

218

Exercise 9.11(h) (continued)

Profits for the months of 49, 50 and 51 (obtained by subtracting total cost from total
revenue) are

PROFITS 49 = 5349

PROFITS 50 = 5779

PROFITS 51 = 6031

Because eTR ,48 is negative, and its impact declines as we predict further into the future, the
total revenue predictions become larger the further into the future we predict. The opposite
happens with total cost; it declines because eTC ,48 is positive. Combining these two
influences means that the predictions for profit increase over time. These predictions are,
however, much lower than 8524, the prediction for profit that was obtained when
autocorrelation was ignored. Thus, even although autocorrelation has little impact on the
optimal setting for Q , it has considerable impact on the predictions of profit. This impact
is caused by a change in the coefficient estimates, a relatively large negative residual for
revenue in month 48, and a relatively large positive residual for cost in month 48.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

219

EXERCISE 9.12
(a)

The estimated model with standard errors in parentheses is

ln( At ) = 3.8241 + 0.7746ln( Pt ) 0.2175ln( Pt 1 ) 0.0026ln( Pt 2 )

(se)

( 0.1006 )( 0.3129 )

( 0.3185 )

( 0.3221)

+ 0.5868ln( Pt 3 ) 0.0143ln( Pt 4 )

( 0.3153)

( 0.2985)

Multipliers
Lag

Delay

Interim

0
1
2
3
4

0.7746
0.2175
0.0026
0.5868
0.0143

0.7746
0.5572
0.5546
1.1414
1.1271

Only b0 , the coefficient of ln( Pt ) , is significantly different from zero at a 5% level of

significance. All coefficients for lagged values of ln( Pt ) , namely, b1 , b2 , b3 , b4 , are not
significant at a 5% level. This result is symptomatic of collinearity in the data. When
collinearity exists, least squares cannot distinguish between the individual effects of each
independent variable, resulting in large standard errors and coefficients which are not
significantly different from zero.
Interpreting the delay multipliers, if the price is increased and then decreased by 1% in
period t, there is an immediate increase of 0.77% in area planted. In period t + 1 , that is
one period after the price shock, there is a decrease in area planted of 0.22%. In period
t + 2 there is practically no change in the area planted. In period t + 3 there is an increase
in area planted by 0.59% and in period t + 4 there is a decrease of 0.01%.
The interim multipliers represent the full effect in period t + s of a sustained 1% increase
in price in period t. Thus, if the price increases by 1% in period t, there is an immediate
increase in the area planted of 0.77%. The total increase when period t + 1 is reached is
0.56%, at period t + 2 it is 0.55%, at period t + 3 it is 1.14% and, after t + 4 periods there
is a 1.13% increase.
The different signs attached to the delay multipliers, the relatively large weight at t 3 ,
and the interim multipliers that decrease and then increase are not realistic for this
example. The pattern is likely attributable to imprecise estimation.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

220

Exercise 9.12 (continued)

(b)

Using the straight line formula the lag weights are

0 = 0

i=0

1 = 0 + 1

i =1

2 = 0 + 21

i=2

3 = 0 + 31

i=3

4 = 0 + 41

i=4

Substituting these weights into the original model gives

ln( At ) = + 0 ln( Pt ) + ( 0 + 1 ) ln( Pt 1 ) + ( 0 + 21 ) ln( Pt 2 )

+ ( 0 + 31 ) ln( Pt 3 ) + ( 0 + 41 ) ln( Pt 4 ) + et
= + 0 ( ln( Pt ) + ln( Pt 1 ) + ln( Pt 2 ) + ln( Pt 3 ) + ln( Pt 4 ) )
+ 1 ( ln( Pt 1 ) + 2ln( Pt 2 ) + 3ln( Pt 3 ) + 4ln( Pt 4 ) ) + et
= + 0 zt 0 + 1 zt1 + et
(c)

The estimated equation with standard errors in parentheses is

ln( At ) = 3.8266 + 0.4247 zt 0 0.0996 zt1

(se)

(0.1056) (0.2594)

(0.1088)

The least squares estimates of 0 and 1 are 0.4247 and -0.0996 respectively.
(d)

The estimated weights are

0 = a0 = 0.42467
1 = a0 + a1 = 0.42467 0.09963 = 0.3250
2 = a0 + 2a1 = 0.42467 2 0.09963 = 0.2254
3 = a0 + 3a1 = 0.42467 3 0.09963 = 0.1258
4 = a0 + 4a1 = 0.42467 4 0.09963 = 0.0261
These lag weights satisfy expectations as they are positive and diminish in magnitude as
the lag length increases. They imply that the adjustment to a sustained price change takes
place gradually, with the biggest impact being felt immediately and with a declining
impact being felt in subsequent periods. The linear constraint has fixed the original
problem where the signs and magnitudes of the lag weights varied unexpectedly.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

221

Exercise 9.12 (continued)

(e)

The new multipliers are

Multipliers
Lag

Delay

Interim

0
1
2
3
4

0.4247
0.3250
0.2254
0.1258
0.0261

0.4247
0.7497
0.9751
1.1009
1.1270

These delay multipliers are all positive and steadily decrease as the lag becomes more
distant. This result, compared to the positive and negative multipliers obtained earlier, is a
more reasonable one. It is interesting that the total effect, given by the 4-year interim
multiplier, is almost identical in both cases, and the 3-year interim multipliers are very
similar. The earlier interim multipliers are quite different however, with the restricted
weights leading to a smaller initial impact.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

222

EXERCISE 9.13
The estimated equation with standard errors in parentheses is

y t = 1.0942 12.2998 xt 27.0635 xt 1 0.4001 yt 1

( 5.7824 ) ( 26.6145) ( 26.4483)

( 0.0601)

(i) yt is the predicted change in housing starts (in thousands) from month t 1 to month
t and xt is the change in interest rates from month t 1 to month t. The estimated
coefficients of xt and xt 1 are negative and suggest that the immediate effect of a
temporary one unit increase in xt is a decrease in yt of 12.30 and the one month
lagged effect is a decrease in yt of 27.06. The coefficient of yt 1 is negative,
suggesting that a positive change in housing starts lead to a negative change in
housing starts in the following period. These signs are generally in line with our
expectations. However, more implications of the signs and magnitudes of the
coefficients can be obtained by examining the lag weights in the infinite lag
representations as is done later in the question. The only coefficient estimate that is
significantly different from zero is 1 . Thus, with two important coefficients not
significantly different from zero (those for xt and xt 1 ), the model is not a reasonable
one.
(ii) Testing the hypothesis H 0 : 1 = 10 against the alternative H1 : 1 10 delivers
an F test value of 0.483861 with a p-value of 0.4873. Since the p value is greater that
the 0.05 level of significance, we do not reject the null hypothesis. On the basis of this
test, a restricted model of the form yt = 1 + 2 xt + et with AR(1) errors is reasonable.
(iii) A correlogram of the residuals is presented below. Significance bounds are drawn at
1.96 250 = 0.124 . Although not large, the correlations r2 , r12 and r24 are
statistically significant. Since the data are monthly, there could be annual effect that is
not being picked up. Overall, given these correlations and the insignificant coefficients
mentioned in part (a), further modeling is in order.
.15
.10
.05
Correlations

(a)

.00
-.05
-.10
-.15
-.20
-.25
5

10

15

20

Lag

Figure xr9.13a Residual correlogram for Exercise 9.13 part (a)

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

223

Exercise 9.13 (continued)

The estimated model with standard errors in parentheses is

y t = 0.4856 58.4292 xt 3 0.5227 yt 1 0.2903 yt 2 0.1641 yt 3

( 5.5671) ( 23.8935 )

( 0.0631)

( 0.0684 )

( 0.0634 )

(i) In contrast to the model estimated in part (a) all the estimated coefficients are
significant (with the exception of the intercept). There is a 3-month lagged effect on
housing starts from an initial interest rate change, and the effect is negative as one
would expect. The implications of the signs and magnitudes of the lagged y variables
are better assessed from the consequent estimates of the lagged weights.
(ii) Using the results from Exercise 9.6 we obtain the following lag weights
Lag weight estimate

0
1
2
3
4
5
6
7
8
9
10
11
12

2
=

58.4292

4 = 1 3
5 = 1 4 + 2 3
6 = 1 5 + 2 4 + 3 3
= + +
7

1 6

2 5

30.5396
1.0023
0.1963
5.4046

3 4

8 = 1 7 + 2 6 + 3 5
9 = 1 8 + 2 7 + 3 6
= + +
10

1 9

2 8

2.6034
0.1762
0.0388

3 7

0.4986

11 = 1 10 + 2 9 + 3 8
= + +
12

1 11

2 10

0.2204

3 9

40

20
Lag Weight

(b)

-20

-40

-60
0

10 11 12

Lag

Figure xr9.13b Lag weights for Exercise 9.13 part (b)

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

224

Exercise 9.13(b)(ii) (continued)

(ii) Figure 9.5 gives us a clear depiction of the distributed lag weights. The first three lag
weights are zero indicating that interest rates have no effect on changes in housing
starts over the two months following a rate change. The major impact of interest rates
appears to follow at lags 3 and 4. A change in interest rates has the largest negative
effect on housing starts at lag 3, and a positive readjustment occurs at lag 4. From lag
5 onwards, the effect on housings starts is relatively insignificant with the exception of
the small spikes at lags 7 and 8.
(iii) A correlogram of the residuals is presented in Figure xr9.13c. The new model has
eliminated the significant correlation at lag 2 that we found in the earlier model, but
those at lags 12 and 24 remain, although they are small in magnitude.
.15
.10
.05
Correlations

(b)

.00
-.05
-.10
-.15
-.20
5

10

15

20

Lag

Figure xr9.13a Residual correlogram for Exercise 9.13 part (a)

(iv) The forecast of yt for January is

yT +1 = + 3 xT 2 + 1 yT + 2 yT 1 + 3 yT 2
= 0.4856 58.4292 0.3 0.5227 (129) 0.2903 85 0.1641 (112)
= 44.08
Similarly the forecast of yt for February is

yT + 2 = + 3 xT 1 + 1 yT +1 + 2 yT + 3 yT 1
= 0.4856 58.4292 0.26 0.5227 44.08 0.2903 (129) 0.1641 85
= 14.24
The forecast of yt for March is

yT +3 = + 3 xT + 1 yT + 2 + 2 yT +1 + 3 yT
= 0.4856 58.4292 (0.06) 0.5227 ( 14.24) 0.2903 44.08
0.1641 ( 129)
= 19.80

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

Exercise 9.13(b) (continued)

(b)

(v) The forecast of housing starts, in thousands of houses, for January 2006 is

HOUSE T +1 = yT +1 + HOUSET = 44.08 + 2002 = 2046

Similarly, for February 2006,

HOUSE T + 2 = yT + 2 + HOUSE T +1 = 14.24 + 2046.08 = 2032

For March 2006 it is

HOUSE T +3 = yT + 3 + HOUSE T + 2 = 19.80 + 2031.84 = 2052

225

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

226

EXERCISE 9.14
(a)

Testing the null hypothesis H 0 : = 0 against the alternative H1 : 0 we obtain the test
statistic value LM = 4.383 with a corresponding p value of 0.0363. Since the p value is
less than a significance level of 0.05, we reject the null hypothesis and conclude that the
errors in this model are correlated.

(b)

There are a number of possible ARDL models that could be chosen here. Given the
relatively small number of observations, we have opted for the simplest one that
eliminates first-order autocorrelation, namely an ARDL(1,0) model. Also, the coefficients
of the extra terms included in ARDL(2,0) and ARDL(1,1) models were found to be
insignificant. If you experiment with more lags you will find that an ARDL(4,1) model
has a large number of significant coefficients. However, estimating such a large model
with such a small sample is too ambitious. The estimated ARDL(1,0) model is

ln(UNITCOSTt ) = 2.5088 0.1862ln(CUMPRODt ) + 0.6417 ln(UNITCOSTt 1 )

(1.2062 ) ( 0.0787 )

(se)

( 0.2005)

The LM test for AR(1) errors yields a test value of LM = 0.756 with corresponding pvalue of 0.3845, indicating that the correlation found in part (a) has been eliminated by the
inclusion of ln(UNITCOSTt 1 ) .
(c)

Using the notation

ln(UNITCOSTt ) = + 0 ln(CUMPRODt ) + 1 ln(UNITCOSTt 1 ) + vt

the lag weights are given by s = 0 1s , from which we get the following estimates.
Lag weight

estimate

0.1862

0.1195

0.0767

0.0492

0.0316

0.0203

These estimates suggest that, as cumulative production increases, most of the learning
effect that reduces unit cost occurs immediately, but there is also a gradually declining
learning effect that continues to reduce costs beyond the immediate period.
(d)

The prediction of the natural logarithm of UNITCOST for 1971 is given by

ln(UNITCOSTT +1 ) = 2.5088 0.1862ln(3800) + 0.6417 ln(16.41631)

= 2.7697
In 1970, ln(UNITCOSTT ) = 2.7983 . Thus, in line with the learning model, we expect unit
cost to be less in 1971 than in it was 1970.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

227

EXERCISE 9.15
(a)

With the trend specified as t = 1, 2, ,111 , The least squares estimated equation is

ln( POWt ) = 0.1708 + 0.0082 t 0.000037 t 2 + 0.9365ln( PROt )

(se)

(0.4147) (0.0005) (0.000004) (0.0899)

The positive sign for b2 and the negative sign for b3 , and their relative magnitudes,
suggest that the trend for ln(POW) is increasing at a decreasing rate. A positive b4 implies
the elasticity of power use with respect to productivity is positive. The residual
correlogram is depicted in the figure below. There is strong evidence of autocorrelation
with significant positive correlations exceeding the significance bound 1.96 111 = 0.186
up to lag 7, and some significant negative correlations less than the negative bound
1.96 111 = 0.186 beyond lag 19.
1.0
0.8

Correlation

0.6
0.4
0.2
0.0
-0.2
-0.4
2

10

12

14

16

18

20

22

24

Lag

Figure xr9.15a Residual correlogram for Exercise 9.15, part(a)

(b)

The estimated ARDL(1,1) model is

ln( POWt ) = 0.3662 + 0.0016 t 0.000008 t 2 + 0.9828ln( PROt )

(se)

(0.2555) (0.0006) (0.000003)

(0.1052)

0.8768ln( PROt 1 ) + 0.7966ln( POWt 1 )

(0.1110)

(0.0584)

One lag of POW and one lag of PRO were used because the coefficients of the longer lags
were not significantly different from zero. In addition, the residual correlogram in Figure
xr9.15b suggests that autocorrelation has been largely eliminated with only r5 , r6 and r21
statistically significant and even these values are relatively small.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

228

Exercise 9.15(b) (continued)

.3
.2

Correlation

.1
.0
-.1
-.2
-.3
5

10

15

20

Lag

Figure xr9.15b Residual correlogram for Exercise 9.15, part(b)

(c)

The p-values for testing the hypothesis H 0 : 4 = 1 are 0.4817 and 0.8704 for parts (a) and
(b), respectively. We do not reject H 0 in both cases. Including more lags to correct for
autocorrelation has led to a large change in the p-value, but the test decision is still the
same.

(d)

Denoting the coefficients of ln( PROt ) and ln( PROt 1 ) as 0 and 1 , respectively, and
that of ln( POWt 1 ) as 1 , the lag weights can be shown to be
0 , 10 + 1 , 1 (10 + 1 ), 12 (10 + 1 ), 13 (10 + 1 ),

The total multiplier is given by their sum

0 + 10 + 1 + 1 (10 + 1 ) + 12 (10 + 1 ) + 13 (10 + 1 ) +

= 0 (1 + 1 + 12 + 13 +
=

) + 1 (1 + 1 + 12 + 13 +

0 + 1
1 1

Using the estimates corresponding to these parameters the total multiplier is

0 + 1 0.98278 0.87683
=
= 0.5210
1 0.79663
1
1

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

229

EXERCISE 9.16
(a)

The least-squares estimated model without lags is

ln( POWt ) = 0.4306 0.0399 Dt + 0.0092 t 0.000041t 2 + 0.8035ln( PROt )

(se)

(0.4728) (0.0162)

(0.0006) (0.000005)

(0.1030)

The estimated model including lags is

ln( POWt ) = 0.5620 0.0130 Dt + 0.0021t 0.0000099 t 2 + 0.9543ln( PROt )

(se)

(0.2968) (0.0101) (0.0007) (0.0000038) (0.1072)

0.8773ln( PROt 1 ) + 0.7806ln( POWt 1 )
(0.1107)

(b)

(0.0595)

The p-value for the hypothesis test H 0 : 1 = 0 for the model is 0.0152. Since it is less than
the level of significance 0.05, we reject the null hypothesis and conclude that the dummy
variable is statistically significant.
The p-value for the hypothesis test H 0 : 1 = 0 for the model with lags is 0.2019. Since it
is greater than the level of significance 0.05, we do reject the null and therefore cannot
conclude that the dummy variable is significant.
Thus, when we do not account for autocorrelation the structural change is statistically
significant, but when we do correct for autocorrelation, the structural change is no longer
significant. In general, these results suggest that, if we do not specify the correct lag
structure, we can make misleading conclusions about the existence of structural change.

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

230

EXERCISE 9.17
(a)

The estimated model is:

ROBt = 5.6204 + 1.5614t + 0.5639 ROBt 1

( 0.2873) ( 0.0769 )

(se)

The correlogram of the residuals is shown below. The significance bounds are drawn at
1.96 117 = 0.181 . There are a few significant correlations at long lags (specifically at
lag orders 7, 12 and 22), but they are relatively small. The spike at lag 12 could indicate a
monthly seasonal effect.
.3

Correlations

.2
.1
.0
-.1
-.2
-.3
5

10

15

20

Lag

(b)

The predicted values in November and December are respectively:

ROBT +1 = 5.6204 + 1.5614 119 + 0.5639 431 = 423.2

ROBT + 2 = 5.6204 + 1.5614 120 + 0.5639 423.2 = 420.4
The standard errors of prediction are:
1 = v = 36.713
2 = v2 (1 + 12 ) = 36.7132 (1 + 0.56392 ) = 42.148
The confidence intervals are:

ROBT +1 t(0.975,114) 1 = 423.2 1.981 36.713 = (350, 496)

ROBT + 2 t(0.975,114) 2 = 420.4 1.981 42.148 = (337, 504)

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

231

EXERCISE 9.18
(i) The estimated ARDL model is

yt = 0.2552 + 0.3685 xt + 0.2299 xt 1

(se)

( 0.0543) ( 0.0455) ( 0.0456 )

Only one lag of xt was included because neither additional lags of xt nor lags of yt
were statistically significant, with the exception of yt 4 . Including yt 4 would have
unnecessarily complicated the model, particularly in light of the residual correlogram
given below, which shows no significant autocorrelations. The significance bounds in
this correlogram are at 1.96 215 = 0.134 .
.15

Correlations

.10
.05
.00
-.05
-.10
-.15
5

10

15

20

Lag

(ii) The estimated ARDL model is

zt = 0.0493 + 1.2379 xt + 1.0709 xt 1 0.1975 zt 1

(se) ( 0.2908 ) ( 0.2488 )

( 0.2549 )

( 0.0672 )

An ARDL(1,1) model was specified because additional lags of both variables were not
statistically significant, and they were also unnecessary to eliminate autocorrelation.
All correlations in the below correlogram are small, with only that at lag 16
marginally significant.
.15
.10

Correlations

(b)

.05
.00
-.05
-.10
-.15
5

10

15
Lag

20

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

232

Exercise 9.18 (continued)

(c)

(i) For (b) (i) the lag weights for 8 quarters are:

0 = 0.3685, 1 = 0.2299, 2 = 3 = 4 = 5 = 6 = 7 = 8 = 0
The lag weights beyond lag 1 are zero because there is no lagged dependent variable
in the model. The estimated lag weights suggest that a temporary 1 unit increase in xt
will cause yt to increase by 0.3685 at time t, and yt +1 to increase by 0.2299 at time
t + 1 , with no changes in y from time t + 2 onwards. In terms of the original
variables, a 1% temporary increase in the growth of disposal personal income, will
lead to an immediate increase in the growth of personal consumption of 0.3685%, and
an increase of in the following quarter of 0.2299%, with no further changes in
subsequent periods.
Using the notation

zt = + 0 xt + 1 xt 1 + 1 zt 1 + vt
the lag weights for (b) (ii) for the first 8 quarters are given by the expressions

0 = 0
= +
1

1 0

s = 1 s 1 ,

s >1

Making these calculations yields

0 = 1.2379,
4 = 0.0064,

1 = 0.8264,
= 0.0013,
5

2 = 0.1632,
6 = 0.0002,

3 = 0.0322
7 = 0.0000, 8 = 0.0000

These lag weights show the changes in growth of consumption of durable goods due
to a temporary increase in income growth. They suggest that the current and onequarter lagged effects are relatively large and positive, but after that the effects are
relatively small and oscillate in sign, the oscillation being a consequence of the
negative estimate for 1 .

Chapter 9, Exercise Solutions, Principles of Econometrics, 3e

233

Exercise 9.18(c) (continued)

(ii) The total multiplier for part (b) (i) is
1

s = 0 + 1 = 0.3685 + 0.2299 = 0.5984

s= 0

This result shows that the long run effect on y of a sustained 1-unit increase in x is
0.5984.
The total multiplier for part (b) (ii) is

s =
s =0

0 + 1 + 2 + 3 +

= 1.2379 + 0.8264 0.1632 + 0.0322 0.0064 + 0.0013 0.0002 + 0.0000 +

= 1.9280
This result can also be obtained from

s = 0 + 1 + 2 + 3 +
s =0

0 + 1
1
1

1.2379 + 1.0709
1 (0.1975)

= 1.9280
It shows that the long run effect on z of a sustained 1-unit increase in x is 1.928.