Solutions To Exercises in Chapter 8: 1.szm, xr8-1.wf1 and Xr8-1.sas, Respectively
Solutions To Exercises in Chapter 8: 1.szm, xr8-1.wf1 and Xr8-1.sas, Respectively
(b) The estimated version of equation 8.4.2 and corresponding sum of squared errors are
∧
trt = 111.71 − 5.057 pt SSER = 20907
(8.85) (4.012)
(c) The SHAZAM, EViews and SAS instructions can be found in the files xr8-1.xls, xr8-
1.szm, xr8-1.wf1 and xr8-1.sas, respectively.
(d) In Section 8.4.3 the model and restrictions are
trt = β1 + β 2 pt + β 3a t + β 4 a t2 + et
β 3 + 80β 4 = 1 β1 + 2β 2 + 40β 3 + 1600β 4 = 175
Substituting these restrictions into the model yields
trt = [175 − 2β 2 − 40(1 − 80β 4 ) − 1600β 4 ] + β 2 pt + (1 − 80β 4 ) a t + β 4 a t2 + et
F=
(SSER − SSEU ) J = (27151. − 2592.3) 2
= 1.753
SSEU (T − K ) 2592.3 74
(e) Using computer software, we find that, for p = 2 and a = 20, the predicted total revenue is
∧
tr = 146.59. The standard error of the forecast error is se( f ) = 6.064, and a 95%
prediction interval is (134.5, 158.7).
(f) The following table gives the R2s from auxiliary regressions between the explanatory
variables as well as the sample correlations between the explanatory variables. The
results suggest that there are no substantial correlations between pt and the advertising
variables, but the correlation between at and at2 is relatively high. Whether or not it is a
problem depends on whether we have been able to estimate the effect of advertising on
total revenue sufficiently precisely. In this regard the standard error of the coefficient of
2
at is relatively small, but that for the coefficient of at2 is relatively large, suggesting
collinearity could be preventing us from getting a precise estimate of this coefficient.
(g) The response of total revenue to price and advertising expenditure is an important
relationship to quantify because it has implications for price setting and for the
appropriate level of advertising for the hamburger chain. The estimated equation in part
(a) suggests the responses of total revenue to price and advertising are, respectively,
∂(tr ) ∂(tr )
= −10.198 = 3.361 − 0.0536a
∂p ∂a
Thus, demand is price elastic, and there are diminishing returns to advertising
expenditure. On the basis of one-tailed tests, all estimated coefficients are significantly
less than (or greater than) zero at a 5% level of significance.
To see whether advertising contributes significantly to the equation, we need to test
whether the coefficients of at and at2 are both zero. The p-value for this F-test is
0.00000, indicating that the hypothesis is rejected at any conventional significance level.
For testing whether 40 could be the optimal level of advertising expenditure, we obtain a
p-value of 0.801, and thus this hypothesis is not rejected. An F test of the joint
hypothesis that the optimal values for p and a, respectively, are 2 and 40, yields a p-value
of 0.180. Hence, at a 5% significance level, we cannot reject these values as optimal. A
95% prediction interval for total revenue when p = 2 and a = 20 is (134.6, 158.6). Given
our earlier test conclusion, that for p = 2 and a = 40, E(tr) = 175 is a reasonable value, it
would make sense not to set a as low as 20.
Whether or not collinearity between a and a2 is a problem depends on whether
∂(tr ) ∂ a is estimated accurately. For a = 20 and a = 40, this response and the
corresponding standard errors are:
∂(tr )
When a = 20, = 2.291 se(b3 + 40b4) = 0.259
∂a
∂(tr )
When a = 40, = 1.221 se(b3 + 80b4) = 0.874
∂a
For small values of a, the response is estimated reasonably accurately, but, for larger a, it
becomes quite unreliable.
3
8.2 The results of the tests in parts (a) to (e) appear in the following table. Note that, in all cases,
there is insufficient evidence to reject the null hypothesis at the 5% level of significance.
(f) The auxiliary R2s and the explanatory-variable correlations that are exhibited in the
following table suggest a high degree of collinearity in the model.
To see whether the collinearity has an impact we need to examine its effect on the
reliability of estimation. The estimated equation, with t values in parentheses, is
!
ln(Yt ) = 0.035 + 0.056 ln(Kt) + 0.226 ln(Lt) + 0.044 ln(Et) + 0.670 ln(Mt)
(0.800) (0.216) (0.511) (0.112) (1.855)
R2 = 0.952
The very small t-values for all variables except ln(Mt), our inability to reject any of the
null hypotheses in parts (a) through (e), and the high R2, are indicative of high
collinearity. Collectively, all the variables produce a model with high level of
explanation and a good predictive ability. Furthermore, our economic theory tells us that
all the variables are important ones in a production function. However, we have not been
able to estimate the effects of the individual explanatory variables with any reasonable
degree of precision.
SSE 979.8306
And R2 = 1 − = 1− = 0.8612
SST 7057.5267
4
At α = 0.05 , the critical F0.05, 2, 37 ≈ 3.23 . Since the calculated F is greater than the
critical F we reject H 0 . There is no evidence from the data to suggest that β 2 = β3 = 0 .
8.4 The model from Exercise 8.3 is yt = β1 + β 2 xt + β 3 zt + et . The SSE from estimating this model
is 979.8306.
The model after augmenting with the squares and the cubes of predictions yˆ t2 and yˆ t3 is
At α = 0.05, F0.05,2,35 ≈ 3.275 . Since the calculated F is greater than the critical F we reject H 0
and conclude that the model is misspecified.
8.5 (a) Let the total variation, unexplained variation and explained variation be denoted by SST,
SSE and SSR, respectively. Then, we have
SSE = ∑ e"t2 = (T − K ). σ" 2 = (20 − 3) × 2.5193 = 42.8281
SSE
Also, R2 = 1 − and hence
SST
SSE 42.8281
SST = = = 802.0243
1− R 2
1 − 0.9466
and
SSR = SST − SSE = 802.0243 − 42.8281 = 759.1962
(b) A 95% confidence interval for β 2 is
b2 ± tcse(b2) = 0.69914 ± (2.11)( 0.048526 ) = (0.2343, 1.1639)
A 95% confidence interval for β 3 is
b3 ± tcse(b3) = 1.7769 ± (2.11)( 0.03712 ) = (1.3704, 2.1834)
(c) To test H0: β 2 ≥ 1 against the alternative H1: β 2 < 1, we calculate
b2 − β 2 0.69914 − 1
t= = = −1.3658
se(b2 ) 0.048526
5
At a 5% significance level, we reject H0 if t < −tc = −1.74. Since t > −tc we fail to reject
H0 and conclude that β 2 ≥ 1.
(d) To test H 0 : β 2 = β 3 = 0 against the alternative H1: β 2 or β 3 or both are nonzero, we
calculate
explained variation (K − 1) 759.1962 2
F= = = 151
unexplained variation (T − K ) 42.8281 17
At a 5% level of significance Fc = 3.59. Since F > Fc we reject H0 and conclude that the
evidence from the data does not support β 2 = β 3 = 0.
Estimating the cubic model and augmenting it with the squares of the predictions, and
squares and cube of the predictions, yields the RESET test results in the top half of the
table. The F-values are 0.9874 and 0.5456 with the p-values of 0.3307 and 0.5871. Both
of the p-values are above the significance level of 0.05. Thus, the data suggest that the
cubic model is adequate. Turning to the log-log cost model, the RESET test results are in
the bottom half of the table. Both the F-statistics are very big giving the p-values of
0.000027 and 0.0000 which are below 0.05. This suggest the log-log model is inadequate.
Based on the RESET tests we prefer the cubic model.
The marginal cost function from the log-log total cost function will be ever-decreasing or
ever-increasing depending on whether α2 < 1 or α2 > 1; it cannot have decreasing and
increasing segments as it does when the total cost function is cubic. It is thus less
flexible, and, in this sense, less reasonable than the cubic. Furthermore, for this particular
data set, α2 < 1 implying the marginal cost function is ever-decreasing. Whether this is
reasonable for a given industry will depend on the marginal revenue function. It is not
reasonable to have functions that imply output can be profitably expanded to infinity.
8.7 We use t tests for testing all the hypotheses. In each case there are 60 degrees of freedom and
the 5% critical value is tc = 2.00.
(a) The value of the t statistic for testing the null hypothesis that β 2 = 0 is 1.5. This value is
less than tc = 2; therefore, we fail to reject H0 and conclude that there is no sample
evidence to suggest that β 2 ≠ 0.
(b) For testing H0: β 1 + 2β 2 = 5 against the alternative H1: β 1 + 2β 2 ≠ 5, we use the statistic
t=
(b1 + 2b2 ) − (β1 + 2β 2 )
se(b1 + 2b2 )
= 3 + 16 + (−8) = 3.3166
7
8−5
Therefore, t = = 0.9045. Since t is less than tc = 2.0, we fail to reject H0. It can be
3.3166
concluded that there is no sample evidence to suggest that β 1 + 2β 2 ≠ 5.
(c) For testing H 0 : β1 − β 2 + β 3 = 4 against the alternative H1: β1 − β 2 + β 3 ≠ 4 , we use the
statistic
(b1 − b2 + b3 ) − (β1 − β 2 + β 3 )
t=
se(b1 − b2 + b3 )
= 3 + 4 + 3 + 4 + 2 − 0 = 16
−2 − 4
Therefore, se(b1 − b2 + b3) = 4, and t = = −1.5. This t value is less than two in
4
absolute value. Therefore, we fail to reject H0 and conclude that there is no sample
evidence to suggest that β 1 − β 2 + β 3 = 4 is incorrect.
8.8 (a) The restricted and unrestricted least squares estimates and their standard errors appear in
the following table. The most noticeable change in the results is the change in the
estimate for β 3 , from −0.583 to 0.187. The positive sign of the restricted estimate is more
in line with our a priori views about the cross-price elasticity with respect to liquor.
However, its magnitude may still be smaller than we would suspect; it is only slightly
greater than 0.167, the cross-price elasticity with respect to the remaining goods and
services. There is not a large difference between the two alternative estimates of the other
elasticities. With one exception, the standard errors for the restricted estimates are less
than their counterparts for the unrestricted estimates. Thus, apart from the case of β 5 , the
standard errors support the theoretical result that restricted least squares estimates have
lower variances than unrestricted least squares estimates.
β1 β2 β3 β4 β5
(b) To examine the presence of collinearity we look at auxiliary R2s and the sample
correlations between the explanatory variables that appear in the following table. There
are relatively high correlations between the price of beer, the price of other liquor, and
income. The relatively large standard error and the wrong sign for ln(pL) are a likely
consequence of this correlation. Although introducing the nonsample information has not
made the estimated coefficient for ln(pL) significantly different from zero, it has reduced
the standard error considerably and corrected the wrong sign.
8
(c) and (d) The results for these two parts appear in the following table. The two 95%
prediction intervals are (70.6, 127.9) and (59.6, 116.7). The effect of the nonsample
restriction has been to increase both endpoints of the interval by approximately 10 liters.
ln(q) q
ln( q") se( f ) tc lower upper lower upper
(c) Restricted 4.5541 0.14446 2.056 4.257 4.851 70.6 127.9
(d) Unrestricted 4.4239 0.16285 2.060 4.088 4.759 59.6 116.7
8.9 (a) The estimated Cobb-Douglas production function with standard errors in parentheses is
ln( Q" ) = - 0.129 + 0.559 ln(L) + 0.488 ln(K) R2 = 0.688
(0.546) (0.816) (0.704)
The magnitudes of the elasticities of production (coefficients of ln(L) and ln(K)) seem
reasonable, but their standard errors are very large, implying the estimates are unreliable.
The sample correlation between ln(L) and ln(K) is 0.986. It seems that labor and capital
are used in a relatively fixed proportion, leading to a collinearity problem which has
produced the unreliable estimates.
(b) After imposing constant returns to scale the estimated function is
ln( Q" ) = 0.020 + 0.398 ln(L) + 0.602 ln(K)
(0.053) (0.559) (0.559)
We note that the relative magnitude of the elasticities of production with respect to
capital and labor has changed, and the standard errors have declined. However, the
standard errors are still relatively large, implying that estimation is still imprecise.
yt = β1 + β 2 xt + β 3 zt + γ yˆ t + et
8.11 (a) The least squares result for the estimation of yt = β1 + β 2 xt + β 3 wt + et gives b3 = 0.4979
with se(b3 ) = 0.1174 . This suggests that b3 is significantly different from zero and
therefore wt should be included in the model. Additionally, the RESET test based on the
equation yt = β1 + β 2 xt + et gives F-statistics of 17.982 and 8.7234 which are much
higher than the critical values of 4.17 and 3.32 at a 0.05 level of significance. Thus, the
model omitting wt is inadequate.
(b) When wt is omitted from the model, β 2 is overestimated and se(b2 ) is much smaller.The
omitted variable bias has led to a change in sign for b2 , and the apparent change in the
significance of b2 from being insignificant to being significant is misleading.
(c) The correlation between xt and wt is very high and likely to be positive. To see why it is
likely to be positive, note that, from Section 8.6.1a
ˆ xt , wt )
cov(
E (b2* ) = β 2 + β3
ˆ xt )
var(
where b2* is the estimator of β 2 in the model with wt omitted. Because b3 > 0 , it is likely
that β3 > 0 . Furthermore, since b2 < b2* , we expect that β3 cov( ˆ xt ) > 0 .
ˆ xt , wt ) var(
ˆ xt , wt ) > 0 . That is, xt and wt are positively
Bringing all these facts together implies cov(
correlated. The correlation is likely to be high because the difference between b2 and b2*
is dramatic and because the RESET test successfully detected the omitted variable. The
high standard error for b2 is likely to be attributable to the high correlation.
( SSE R − SSEU ) J
F = = 2.7985
SSEU (T − K )
At α = 0.05, F0.05,2,43 ≈ 3.23 . Since the F-statistic is less than the critical value we do not
reject H 0 ; the response of yield is the same irrespective of whether the rain falls during
germination, development or flowering.
10
The test outcomes are not consistent with our intuition; the number of cylinders and the
engine size should definitely have an impact on miles per gallon.
(c) The high correlations and auxiliary R 2 ' s in the table below indicate that the estimation
in part (a) suffers from the problem of collinearity. Its consequence is that some of the
standard errors of the estimators are large. This imprecise estimation explains why the t-
statistics in parts (b)(i) and (ii), and the F-statistic in part (b)(iii), are small leading to H 0
not being rejected.
Correlations R2 s
cyl 1.0000 0.9507
eng 0.9508 1.0000 0.9482
hp 0.8429 0.8973 1.0000 0.8125
wgt 0.8975 0.9330 0.8645 0.8758
cyl eng hp
11
8.14 (a) The coefficients of LY, LK and LPF are 0.6792, 0.3503 and 0.3219, respectively. Since
the model is in log-log form the coefficient are elasticities. The interpretations of these
coefficients are that the percentage changes in VC when Y, K, and PF change by 1% are
0.6792, 0.3503 and 0.3219, respectively.
(b) We expect all the coefficients to have positive signs. An increase in any of the
explanatory variables should lead to an increase in variable cost. All except the
coefficients for LPM and LSTAGE have the expected sign.
(c) The coefficient for LPM has a p-value of 0.4966 which is higher than 0.05. This indicates
that this coefficient is not significantly different from zero.
(d) Augmenting the equation with the squares of the predictions, and squares and cubes of
the predictions, yields the RESET test F-statistics of 3.3803 and 1.8601 with the
corresponding p-values of 0.0671 and 0.1577, respectively. These two p-values are
higher than the conventional 0.05 level of significance indicating that the model is
adequate.
(e) From the middle part of Table 8.10 the F-statistic for testing H 0 : β 2 + β 3 = 1 is 6.1048
with a p-value of 0.014. This p-value is less than the significance level of 0.05. We reject
H 0 and conclude that constant returns to scale do not exist.
(f) The F-statistic and the p-value for testing H 0 : β 4 + β5 + β 6 = 1 can be read from the
bottom part of Table 8.10. The F value is very large and the corresponding p-value of
0.00000 is below the significance level of 0.05. We reject H 0 and conclude that there is
no evidence to suggest that if all input prices increase by the same proportion, variable
cost will increase by the same proportion.
(g) To test H 0 : β 2 + β 3 = 1 the t statistic is defined as
b2 + b3 − 1
t=
se(b2 + b3 )
b4 + b5 + b6 − 1
t=
se(b4 + b5 + b6 )
se(b4 + b5 + b6 ) = var(
ˆ b4 ) + var(
ˆ b5 ) + var(
ˆ b6 ) + 2cov(
ˆ b4 , b5 ) + 2cov(
ˆ b4 , b6 ) + 2cov(
ˆ b5 , b6 )
= 0.0019 + 0.0013 + 0.01 + 2( −0.000) + 2( −0.0022) + 2( −0.0029)
= 0.0548
And
0.2754 + 0.3219 − 0.0683 − 1
t = = − 8.6
0.0548
At α = 0.05 , t0.025 = 1.96 . We reject H 0 . In this case t 2 = (8.6) 2 = 74 which again is
approximately equal to the F-value of 75. The difference can be attributed to rounding
error.