Solutions For Chapter 7
Solutions For Chapter 7
Exercise Solutions
141
142
EXERCISE 7.1
(a)
When a GPA is increased by one unit, and other variables are held constant, average
starting salary will increase by the amount $1643 ( t = 4.66 , and the coefficient is
significant at = 0.001). Students who take econometrics will have a starting salary
which is $5033 higher, on average, than the starting salary of those who did not take
econometrics ( t = 11.03 , and the coefficient is significant at = 0.001). The intercept
suggests the starting salary for someone with a zero GPA and who did not take
econometrics is $24,200. However, this figure is likely to be unreliable since there would
be no one with a zero GPA. The R2 = 0.74 implies 74% of the variation of starting salary is
explained by GPA and METRICS
(b)
if FEMALE = 0
1 + 2GPA + 3 METRICS
E ( SAL ) =
( 1 + 4 ) + 2GPA + 3 METRICS if FEMALE = 1
(c)
To see if the value of econometrics is the same for men and women, we change the model
to
1 + 2 GPA + 3 METRICS
E ( SAL ) =
( 1 + 4 ) + 2 GPA + ( 3 + 5 ) METRICS
if FEMALE = 0
if FEMALE = 1
143
EXERCISE 7.2
(a)
(b)
There are very little changes in the remaining coefficients, or their standard errors, when
FULLMOON and NEWMOON are omitted. The equation goodness-of-fit statistic
decreases slightly, as expected when variables are omitted. Based on these casual
observations the consequences of omitting FULLMOON and NEWMOON are negligible.
(c)
H 0 : 6 = 7 = 0
H1 : 6 or 7 is nonzero.
F=
( SSER SSEU ) 2
SSEU (229 7)
where SSER = 27424.19 is the sum of squared errors from the estimated equation with
FULLMOON and NEWMOON omitted and SSEU = 27108.82 is the sum of squared errors
from the estimated equation with these variables included. The calculated value of the F
statistic is 1.29. The .05 critical value is F(0.95, 2, 222) = 3.307 , and corresponding p-value is
0.277. Thus, we do not reject the null hypothesis that new and full moons have no impact
on the number of emergency room cases.
144
EXERCISE 7.3
(a)
The estimated coefficient of the price of alcohol suggests that, if the price of pure alcohol
goes up by $1 per liter, the average number of days (out of 31) that alcohol is consumed
will fall by 0.045.
(b)
24.78
q p
= 0.045
= 0.320
3.49
p q
(c)
To compute this elasticity, we need q for married black males in the 21-30 age range. It is
given by
q = 4.099 0.045 24.78 + 0.000057 12425 + 1.637 0.807 + 0.035 0.580
= 3.97713
Thus, the price elasticity is
q p
24.78
= 0.045
= 0.280
p q
3.97713
(d)
The coefficient of income suggests that a $1 increase in income will increase the average
number of days on which alcohol is consumed by 0.000057. If income was measured in
terms of thousand-dollar units, which would be a sensible thing to do, the estimated
coefficient would change to 0.057.
(e)
The effect of GENDER suggests that, on average, males consume alcohol on 1.637 more
days than women. On average, married people consume alcohol on 0.807 less days than
single people. Those in the 12-20 age range consume alcohol on 1.531 less days than those
who are over 30. Those in the 21-30 age range consume alcohol on 0.035 more days than
those who are over 30. This last estimate is not significantly different from zero, however.
Thus, two age ranges instead of three (12-20 and an omitted category of more than 20), are
likely to be adequate. Black and Hispanic individuals consume alcohol on 0.580 and 0.564
less days, respectively, than individuals from other races. Keeping in mind that the critical
t-value is 1.960, all coefficients are significantly different from zero, except that for the
dummy variable for the 21-30 age range.
145
EXERCISE 7.4
(a)
The estimated coefficient for SQFT suggests that an additional square foot of floor space
will increase the price of the house by $72.79. The positive sign is as expected, and the
estimated coefficient is significantly different from zero. The estimated coefficient for
AGE implies the house price is $179 less for each year the house is older. The negative
sign implies older houses cost less, other things being equal. The coefficient is
significantly different from zero.
(b)
The estimated coefficients for the dummy variables are all negative and they become
increasingly negative as we move from D92 to D96. Thus, house prices have been
steadily declining in Stockton over the period 1991-96, holding constant both the size and
age of the house.
(c)
Including a dummy variable for 1991 would have introduced exact collinearity unless the
intercept was omitted. Exact collinearity would cause least squares estimation to fail. The
collinearity arises between the dummy variables and the constant term because the sum of
the dummy variables equals 1; the value of the constant term.
146
EXERCISE 7.5
(a)
E (YIELD )
( NITRO )
when PHOS = 1
when PHOS = 2
when PHOS = 3
The effect of additional nitrogen on yield depends on both the level of nitrogen and the
level of phosphorus. For a given level of phosphorus, marginal yield is positive for small
values of NITRO but becomes negative if too much nitrogen is applied. The level of
NITRO that achieves maximum yield for a given level of PHOS is obtained by setting the
first derivative equal to zero. For example, when PHOS = 1 the maximum yield occurs
when NITRO = 7.444/3.888 = 1.915. The larger the amount of phosphorus used, the
smaller the amount of nitrogen required to attain the maximum yield.
(b)
E (YIELD )
( PHOS )
when NITRO = 1
when NITRO = 2
when NITRO = 3
Comments similar to those made for part (a) are also relevant here.
(c)
t=
b2 + 2b4 + b6
8.011 2 1.944 0.567
=
= 7.367
se ( b2 + 2b4 + b6 )
0.233
t=
b2 + 4b4 + b6
8.011 4 1.944 0.567
=
= 1.660
se ( b2 + 4b4 + b6 )
0.040
Since |t| < 2.080 = t(0.975, 21) , we do not reject the null hypothesis. A zero marginal yield
with respect to nitrogen is compatible with the data when NITRO = 1 and PHOS = 2.
147
t=
b2 + 6b4 + b6
8.011 6 1.944 0.567
=
= 8.742
se ( b2 + 6b4 + b6 )
0.233
Since |t| > 2.080 = t(0.975, 21) , we reject the null hypothesis and conclude that the
marginal product of yield to nitrogen is not zero when NITRO = 3 and PHOS = 1.
(d)
The maximizing levels NITRO and PHOS are those values for NITRO and PHOS such
that the first-order partial derivatives are equal to zero.
E (YIELD )
( PHOS )
E (YIELD )
( NITRO )
= 3 + 25 PHOS + 6 NITRO = 0
= 2 + 24 NITRO + 6 PHOS = 0
NITRO =
PHOS =
The yield maximizing levels of fertilizer are not necessarily the optimal levels. The
optimal levels are those where the marginal cost of the inputs is equal to the marginal
value product of those inputs. Thus, the optimal levels are those for which
E (YIELD )
( PHOS )
PRICEPHOS
PRICEPEANUTS
and
E (YIELD )
( NITRO )
PRICENITRO
PRICEPEANUTS
148
EXERCISE 7.6
(a)
( 0.0264 ) ( 0.0359 )
( 0.00104 )
( 0.001414 )
( 0.000218 )
(b)
( 0.00510 )
R 2 = 0.8619
( 0.004140 )
dy 1
= 2
dx y
or
dy
= 2 dx
y
PRICE
1
= 2 + UTOWN
SQFT PRICE
PRICE
1
= 3
AGE PRICE
Using this result for the coefficients of SQFT and AGE, we find that an additional 100
square feet of floor space increases price by 3.6% for a house not in University town; a
house which is a year older leads to a reduction in price of 0.0904%. Both estimated
coefficients are significantly different from zero.
(c)
( ln( PRICE
pool
an approximation of the percentage change in price due to the presence of a pool is 1.90%.
Using the results in Section7.5.1b,
PRICEnopool
100 = ( e 2 1) 100
the exact percentage change in price due to the presence of a pool is 1.92%.
149
( ln( PRICE
fireplace
PRICEnofireplace
100 = e 3 1 100
the exact percentage change in price due to the presence of a fireplace is also 0.66%.
(e)
n
ln
( PRICEutown )
SQFT = 25
n
ln
( PRICEnoutown )
SQFT = 25
(e
0.2477
1) 100 = 28.11%
150
EXERCISE 7.7
(a)
( 0.6464 ) ( 0.5765)
( 0.4486 )
+ 1.4313DISPAD
( 0.6053)
( 0.1052 )
R 2 = 0.8428
( 0.1562 )
(b)
The estimates of 2 , 3 and 4 are all significant and have the expected signs. The sign
of 2 is negative, while the signs of the other two coefficients are positive. These signs
imply that Brands 2 and 3 are substitutes for Brand 1. If the price of Brand 1 rises, then
sales of Brand 1 will fall, but a price rise for Brand 2 or 3 will increase sales of Brand 1.
Furthermore, with the log-linear function, the coefficients are interpreted as proportional
changes in quantity from a 1-unit change in price. For example, a one-unit increase in the
price of Brand 1 will lead to a 375% decline in sales; a one-unit increase in the price of
Brand 2 will lead to a 115% increase in sales.
These percentages are large because prices are measured in dollar units. If we wish to
consider a 1 cent change in price a change more realistic than a 1-dollar change then
the percentages 375 and 115 become 3.75% and 1.15%, respectively.
(c)
n 2 SAL1
n1
Q exp{b5 } Q
SAL1
100 =
100 = (e0.4237 1) 100 = 52.8%
n
Q
SAL11
151
n 3 SAL1
n1
Q exp{b6 } Q
SAL1
100 =
100 = (e1.4313 1) 100 = 318%
n
Q
SAL11
The signs and relative magnitudes of b5 and b6 lead to results consistent with economic
logic. A display increases sales; a display and an advertisement increase sales by an even
larger amount.
(d)
(e)
H0
(i)
(ii)
(iii)
(iv)
5 = 0
6 = 0
5 = 6 = 0
6 5
Test
Value
Degrees of
Freedom
5% Critical
Value
Decision
t = 4.03
t = 9.17
F = 42.0
t = 6.86
46
46
(2,46)
46
2.01
2.01
3.20
1.68
Reject H0
Reject H0
Reject H0
Reject H0
The test results suggest that both a store display and a newspaper advertisement will
increase sales, and that both forms of advertising will increase sales by more than a store
display by itself.
152
EXERCISE 7.8
(a)
(0.2537) (0.0868)
R 2 = 0.4391
(0.2629)
All estimated coefficients are significantly different from zero. The intercept suggests that
the average price of CDs that have a 1999 copyright and are not sold on the internet is
$15.46. For every year the copyright date is earlier than 1999, the price increases by 27
cents. For CDs sold through the internet, the price is $2.36 cheaper. The positive
coefficient of AGE supports Mixon and Resslers hypothesis.
(b)
(0.2424) (0.2567)
R 2 = 0.4380
(0.2632)
Again, all estimated coefficients are significantly different from zero. They suggest that
the average price of new releases, not sold on the internet, is $15.53. If the CD is not a
new release, the price is 79 cents higher. If it is purchased over the internet, the price is
$2.36 less. The positive coefficient of OLD supports Mixon and Resslers hypothesis.
153
EXERCISE 7.9
The estimated coefficients and their standard errors (in parenthesis) for the various parts of
this question are given in the following table.
Variable
Constant (1)
(a)
(b)
(c)
(f)
(g)
128.98*
(34.59)
342.88*
(72.34)
161.47
(120.7)
109.72
(135.6)
98.48
(179.1)
7.5756*
(2.317)
2.9774
(3.352)
2.0383
(3.542)
1.7200
(4.842)
2.3822*
(0.6036)
9.0739*
(3.670)
18.325
(11.49)
22.104
(40.26)
0.1602
(0.0867)
0.6115
(0.5381)
0.9087
(3.079)
0.0055
(0.0064)
0.0131
(0.0784)
568869
35
0.000065
(0.000663)
568708
34
AGE (2 )
INC (3 )
1.4577*
(0.5974)
AGE INC (4 )
AGE 2 INC (5 )
AGE 3 INC (6 )
819286
38
SSE
N K
635637
37
580609
36
See table.
(b)
The signs of the estimated coefficients suggest that pizza consumption responds positively
to income and negatively to age, as we would expect. All estimated coefficients are greater
than twice their standard errors, indicating they are significantly different from zero using
one or two-tailed tests. We note that scaling the income variable (dividing by 1000) has
increased the coefficient 1000 times.
(c)
E ( PIZZA )
( AGE )
= 2 + 4 INC
E ( PIZZA )
INC
= 3 + 4 AGE
We expect 3 > 0 and 4 < 0 implying that the response of pizza consumption to income
will be positive, but that it will decline with age. The estimates agree with these
expectations. Negative signs for b2 and b4 imply that, as someone ages, his or her pizza
consumption will decline, and the decline will be greater the higher the level of income.
154
(d)
and
H1: 2 0 and/or 4 0
= 7.40
The 5% critical value for (2, 36) degrees of freedom is Fc = 3.26 and the p-value of the test
is 0.002. Thus, we reject H0 and conclude that age does affect pizza expenditure.
(e)
E ( PIZZA )
INC
= 3 + 4 AGE
Point estimates, standard errors and 95% interval estimates for this quantity, for different
ages, are given in the following table.
Age
Point
Estimate
Standard
Error
20
30
40
50
5.870
4.268
2.665
1.063
1.977
1.176
0.605
0.923
Confidence Interval
Lower
Upper
1.861
1.882
1.439
0.809
9.878
6.653
3.892
2.935
The interval estimates were calculated using tc = t(0.975,36) = 2.0281 . As an example of how
the standard errors were calculated, consider age 30. We have
n
n
n
n
var
( b3 + 30b4 ) = var
( b3 ) + 302 var
( b4 ) + 2 30 cov
( b3 , b4 )
= 13.466 + 900 0.0075228 60 0.31421 = 1.38392
n
se ( b3 + 30b4 ) = var
( b3 + 30b4 ) = 1.176
The corresponding interval estimate is
4.268 2.028 1.176 = (1.882, 6.653)
155
The point estimates for the marginal propensity to spend on pizza decline as age increases,
as we would expect. However, the confidence intervals are relatively wide indicating that
our information on the marginal propensities is not very reliable. Indeed, all the
confidence intervals do overlap.
(f)
E ( PIZZA )
INC
= 2 + 4 AGE + 5 AGE 2
If this marginal effect is to increase with age, up to a point, and then decline, then 5 < 0.
The sign of the estimated coefficient b5 = 0.0055 did not agree with this anticipation.
However, with a t value of t = 0.0055/0.0064 = 0.86, it is not significantly different from
zero.
(g)
Two ways to check for collinearity are (i) to examine the simple correlations between each
pair of variables in the regression, and (ii) to examine the R2 values from auxiliary
regressions where each explanatory variable is regressed on all other explanatory variables
in the equation. In the tables below there are 3 simple correlations greater than 0.94 in part
(f) and 5 in part (g). The number of auxiliary regressions with R2s greater than 0.99 is 3 for
part (f) and 4 for part (g). Thus, collinearity is potentially a problem. Examining the
estimates and their standard errors confirms this fact. In both cases there are no t-values
which are greater than 2 and hence no coefficients are significantly different from zero.
None of the coefficients are reliably estimated. In general, including squared and cubed
variables can lead to collinearity if there is inadequate variation in a variable.
Simple Correlations
INC
AGE
AGE INC
AGE 2 INC
AGE
AGE INC
AGE 2 INC
AGE 3 INC
0.4685
0.9812
0.5862
0.9436
0.6504
0.9893
0.8975
0.6887
0.9636
0.9921
R2 in part (f)
R2 in part (g)
INC
AGE
AGE INC
AGE 2 INC
AGE 3 INC
0.99796
0.68400
0.99956
0.99859
0.99983
0.82598
0.99999
0.99999
0.99994
156
EXERCISE 7.10
(a)
The estimated equation with gender (FEMALE) included, and with standard errors written
in parentheses, is
( 51.3441) (1.5501)
( 0.0004 )
( 27.7681)
The t-value for gender is t = 190.2581 27.7681 = 6.8517 indicating that it is a relevant
explanatory variable. Including it in the model has led to substantial changes in the
coefficients of the remaining variables.
(b)
When level of educational attainment is included the estimated model, with the standard
errors in parentheses, becomes
( 83.3909 ) ( 2.3263)
( 0.0007 )
( 57.8402 )
1.6802COLLEGE 73.2047GRAD
( 62.6621)
( 92.0859 )
None of the dummy variable coefficients are significant, casting doubt on the relevance of
education as an explanatory variable. Also, including the education dummies has had little
impact on the remaining coefficient estimates. To confirm the lack of evidence supporting
the inclusion of education, we need to use an F test to jointly test whether the coefficients
of HS, COLLEGE and GRAD are all zero. The value of this statistic is
F=
= 2.0209
The 5% critical value for (3, 34) degrees of freedom is Fc = 3.05; the p-value is 0.13. We
cannot conclude that level of educational attainment influences pizza consumption.
(c)
To test this hypothesis we estimate a model where the dummy variable gender (FEMALE)
interacts with every other variable in the equation. The estimated equation, with standard
errors in parentheses, is
( 63.9450 ) (1.9155)
( 0.0007 )
( 96.5078)
( 3.0871)
( 0.0008)
157
To test the hypothesis that the regression equations for males and females are identical, we
test jointly whether the coefficients of FEMALE, AGEFEMALE, INCOMEFEMALE are
all zero. Note that individual t tests on each of these coefficients do not suggest gender is
relevant. However, when we take all variables together, the F value for jointly testing their
coefficients is
F=
= 18.1344
This value is greater than Fc = 2.866 which is the 5% critical value for (3, 36) degrees of
freedom. The p-value is 0.0000. Thus, we reject the null hypothesis that males and females
have identical pizza expenditure equations. This result implies different equations should
be used to model pizza expenditure for males, and that for females. It does not say how the
equations differ. For example, all their coefficients could be different, or simply modelling
different intercepts might be adequate.
158
EXERCISE 7.11
(a)
(120.6634 ) ( 3.3521)
( 0.00367 )
( 0.0000867 )
From the sample we obtain average age = 33.475 and average income = 42,925. Thus, the
required marginal effect of income is
n
E
( PIZZA)
= 0.00907 0.00016 33.475 = 0.00371
INCOME
Using computer software, we find the standard error of this estimate to be 0.000927, and
the t value for testing whether the marginal effect is significantly different from zero is
t = 0.00371 0.000927 = 4.00 . The corresponding p-value is 0.0003 leading us to conclude
that the marginal income effect is statistically significant at a 1% level of significance.
(c)
(d)
n
E
( PIZZA)
= 2.9774 0.00016INCOME
( AGE )
= 2.9774 0.00016 42,925 = 9.854
Using computer software, we find the standard error of this estimate to be 2.5616, and the
t value for testing whether the marginal effect is significantly different from zero is
159
Important pieces of information for Gutbusters are the responses of pizza consumption to
age and income. It is helpful to know the demand for pizzas in young and old communities
and in high and low income areas. A good starting point in an investigation of this kind is
to evaluate the responses at average age and average income. Such an evaluation will
indicate whether there are noticeable responses, and, if so, give some idea of their
magnitudes. The two responses are estimated as
n
E
( PIZZA)
= 0.0037
INCOME
n
E
( PIZZA)
= 9.85
( AGE )
Both these estimates are significantly different from zero at a 1% level of significance.
They suggest that increasing income will increase pizza consumption, but, as a community
ages, its demand for pizza declines. Interval estimates give an indication of the reliability
of the estimated responses. In this context, we estimate that the income response lies
between 0.0018 and 0.0056, while the age response lies between 15.05 and 4.66.
160
EXERCISE 7.12
The estimated model is
(se)
(28.153) (27.810)
(8.901)
(0.925)
The within sample predictions, with age expressed in terms of years (not units of 10 years)
are graphed in the following figure. They are also given in a table on the next page.
15
10
5
0
SCORE
SCOREHAT
-5
-10
-15
20
24
28
32
36
40
44
AGE_UNITS
(a)
To test the hypothesis that a quadratic function is adequate we test H 0 : 4 = 0. The tstatistic is 2.972 and is greater than 2. We therefore reject H 0 and conclude that the
quadratic function is not adequate. For suitable values of 2 , 3 and 4 , the cubic function
can decrease at an increasing rate, then go past a point of inflection after which it
decreases at a decreasing rate, and then it can reach a minimum and increase. These are
characteristics worth considering for a golfer. That is, the golfer improves at an increasing
rate, then at a decreasing rate, and then declines in ability. These characteristics are
displayed in Figure xr7.12.
(b)
(i)
(ii) Between the ages of 20 and 25, where the differences between the predictions are
increasing.
(iii) Between the ages of 25 and 30, where the differences between the predictions are
declining.
(iv) At the age of 36.
(v) At the age of 40.
(c)
At the age of 70, the predicted score (relative to par) for Lion Forrest is 241.71. To break
100 it would need to be less than 28. Thus, he will not be able to break 100 when he is 70.
predicted scores
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
4.4403
4.5621
4.7420
4.9633
5.2097
5.4646
5.7116
5.9341
6.1157
6.2398
6.2900
6.2497
6.1025
5.8319
5.4213
4.8544
4.1145
3.1852
2.0500
0.6923
0.9042
2.7561
4.8799
7.2921
10.0092
161
162
EXERCISE 7.13
(a)
(3.15)
(18.04)
(6.62)
(7.52)
(2.24)
(3.36)
The 5% critical t-value for testing the significance of the coefficients and for other
hypothesis tests is tc = t(0.975,990) = 1.962 . Considering the variables individually:
The intercept estimate tells us that the average earnings per hour for someone with no
education, no experience and who is male, white, not married, not in a union, not from the
south, not working fulltime and not living in the city is 1.3568. Using the corrected
predictor from Chapter 4.4 in POE this estimate is 1.4916. This is the base from which all
our dummy variables are measured. This estimate is significantly different from zero at a
5% level of significance because t > 1.962. This estimate is taken from a range with no
data points so we expect this estimate to be unreliable.
EDUC An increase in education by one year is associated with an approximate 10.42%
increase in hourly wages. This estimate is significantly different from zero at a 5% level
of significance. It is expected that as the level of education increases, ones hourly wage
will increase also.
EXPER One year of extra experience is associated with an approximate 0.99% increase
in hourly wages. This is an expected result, although it does seem small in magnitude.
We do expect that hourly wages are positively related to years of experience, and intuition
tells us that our estimate should be less than 2 . This estimate is significantly different
from zero at a 5% level of significance because t > tc .
FEMALE Using the exact calculation, the hourly wage for females is 17.22% less than
that for males. This is an unsurprising result since we expect that there would be more
stay at home mothers than fathers at the time the data were collected. This estimate is
significantly different from zero at a 5% level of significance because t > tc .
163
BLACK Wages for blacks are, on average, 10.77% less than they are for whites, using an
exact calculation. Since we have controlled for education and experience and other worker
characteristics, it could be indicative of discrimination. The estimate is significantly
different from zero at a 5% level of significance because t > tc ; however, it is not
significantly different from zero at a 1% level of significance.
MARRIED Wages for married workers are, on average, 12.84% higher than they are for
those who are not married. This is a surprising result because one would not expect that
marital status would be associated with hourly wages. In fact, one might expect married
workers would not be able to work as many hours as those not married, and this would
exclude them from the higher wage brackets. Perhaps married workers have other
characteristics that lead them to have higher wages. The estimate is significant at a 5%
level of significance since t > tc .
UNION Wages for workers who are part of a union are 12.18% higher than they are for
workers who are not. This estimate has the expected sign because we suspect that workers
in a union have more bargaining power with their employers and can therefore demand a
higher wage. The estimate is significant at a 5% level of significance since t > tc .
SOUTH Wages for southerners are 6.204% less than they are non-southerners. This
estimate is significantly different from zero at a 5% level of significance because t > tc ;
however, it is not significantly different from zero at a 1% level of significance.
FULLTIME The hourly wage for full time workers is 30.12% higher than it is for those
who do not work full time. Normally we would expect that casual workers have a higher
hourly rate than full time workers when comparing hourly rates within businesses.
However, most workers with degrees or specialized training have full time positions, and
these workers often earn much more than other workers. In this sample, the latter reason
outweighs the former, and the estimate therefore has a positive sign. This estimate is
significant at a 5% level of significance since t > tc .
METRO The hourly wage for someone who lives in a metropolitan area is 12.59%
higher than it is for those who do not. This estimate has the expected sign because we
expect that those in metropolitan areas would have higher costs and therefore would need
to be earning a higher wage. Metropolitan areas would also have a greater range of higherpaying jobs. The estimate is significant at a 5% level of significance since t > tc .
164
(6.46) (38.74)
(14.90)
(15.21)
(4.92)
(9.13)
( 3.30)
+0.2183FULLTIME + 0.1306METRO
(10.69)
(8.25)
There are only slight differences in the estimated coefficient values, and the signs of the
coefficients are the same.
What is evident is that the t-statistic values are all much larger in magnitude for estimation
from the cps.dat data. This reflects the use of a larger sample size of 4733 observations in
cps.dat relative to cps2.dat which has only 1000 observations. Using a larger sample size
improves the reliability of our estimated coefficients because we have more information
about our regression function. The larger t-values also mean that the estimates have
smaller p-values and will therefore be significantly different from zero at a smaller level of
significance.
The decline in the standard errors that leads to the increased precision is given in the
following table.
Variable
cps2.dat
standard
error
cps.dat
standard
error
Decrease
se(cps2.dat)
se(cps.dat)
C
EDUC
EXPER
FEMALE
BLACK
MARRIED
UNION
SOUTH
FULLTIME
METRO
0.096812
0.005777
0.001321
0.028558
0.050843
0.030156
0.038633
0.031215
0.042851
0.03528
0.044882
0.002764
0.0006
0.013155
0.022098
0.013772
0.01762
0.014216
0.020421
0.015817
0.05193
0.003013
0.000721
0.015403
0.028745
0.016384
0.021013
0.016999
0.02243
0.019463
165
EXERCISE 7.14
(a)
( 0.4960 ) ( 0.0338 )
( 0.0236 )
( 0.000552 )
R 2 = 0.2591
We notice that these estimated coefficients are similar to those reported in Table 7.1. As
shown in the following table, the major difference between the two regression outputs is in
the standard errors. The larger sample size has led to a decrease in the standard errors of
around 50%, highlighting the impact of extra information on the reliability of our
estimates.
Standard error
from Table 7.1
Variable
1.0550
0.0702
0.0514
0.00120
C
EDUC
EXPER
EXPER2
Standard error
from cps.dat
Difference in
standard errors
0.4960
0.0338
0.0236
0.00055
0.5590
0.0364
0.0278
0.00065
n
E (WAGE )
EXPER
Evaluating this derivative at the sample median experience of 19 years, yields 0.139.
Thus, we estimate that for a worker with 19 years of experience, an additional year of
experience increases hourly wage by 13.9 cents. The marginal effect of experience
calculated using the data used or Table 7.1 was 15.76 cents.
(b)
( 0.4594 ) ( 0.0336 )
( 0.4038)
+1.2685BLACK FEMALE
( 0.1666 )
R 2 = 0.2365
( 0.5355)
There are 2 major differences between these estimates and those in Table 7.4. Firstly, the
coefficient of BLACK FEMALE is different. The larger sample size could have changed
this result since it has more observations on black females and hence more information on
the relevant parameter. Secondly, the standard errors are again much smaller in this
estimation compared to those in Table 7.4, reflecting the increased reliability from a larger
sample size.
166
To test whether there is interaction between BLACK and FEMALE, we test the hypothesis
H 0 : 5 = 0 against the alternative H1 : 5 0. The quickest way to do this is to consider
the p-value of BLACK FEMALE given by the computer output. The p-value is 0.018.
Since this value is less than 0.05, we reject the null at a 5% level of significance and, in
contrast to the result in Table 7.4, we conclude that there is a significant interaction
between BLACK and FEMALE.
(c)
( 0.4888) ( 0.0335)
( 0.4070 )
( 0.1663)
( 0.5343)
( 0.2209 )
0.7343MIDWEST 0.7775WEST
( 0.2310 )
R 2 = 0.2408
( 0.2368)
Comparing the results with Table 7.5, we find the coefficient estimates are quite similar
and the signs of all the coefficient estimates are exactly the same. The major difference
lies in the standard errors and the p-values. All of the standard errors are significantly
smaller in this estimation compared to those in Table 7.5. Again, this is a reflection of the
use of more information from the larger sample size which increases the reliability of our
estimates. The ability to better estimate the regression with a larger sample size is also
highlighted by the p-values. The p-values allow us to conclude that all the estimated
coefficients are significant at a 5% level of significance. This is not the case in Table 7.5
where the coefficients for BLACK, BLACK FEMALE, MIDWEST and WEST are not
significantly different from zero at a 5% level of significance.
To test the hypothesis that there is no regional effect, we test H 0 : 6 = 7 = 8 = 0 against
the alternative that at least one i 0 for i = 6 to 8. To do this we use a joint test. The
unrestricted model is the model estimated above. The restricted model is the model
estimated in part (b).
F=
= 8.7913
The critical F-value for a 5% level of significance, and degrees of freedom of 3 and 4725,
is 2.6068. Since F > Fc , we reject the null and conclude that there is a regional effect.
167
( 0.5618) ( 0.0408)
(se)
( 0.6293)
( 0.1959 )
( 0.8193)
( 0.9738)
( 0.0717 )
( 0.8303)
( 0.3712 )
R 2 = 0.2397
(1.0989 )
To test whether there is a difference between the wage equation for southern and nonsouthern workers we use the Chow test. We therefore test the hypothesis H 0 : i = 0 for
i = 1 to 5, against the alternative that at least one i 0. Note that the i coefficients are
those specific to the variable SOUTH and all the interaction variables related to SOUTH.
The value of the F-statistic is
F=
The critical F-value for a 5% level of significance, and degrees of freedom of 5 and 4723,
is 2.2150. Since F > Fc , we reject the null and conclude that there is a difference between
the wage equations for southern and non-southern workers.
(e)
(se)
( 0.0400 ) ( 0.0029 )
R 2 = 0.2530
( 0.0138)
The estimated percentage difference in wages between males and females is given by
168
EXERCISE 7.15
(a)
( 3.3803) ( 0.1444 )
( 0.4108)
( 0.3568)
(1.7821)
(1.4983)
( 0.8435)
( 3.7963)
We expect the parameter estimate for the dummy variable PERSON to be positive because
of reputation and knowledge of incumbent. However, it could be negative if incumbents
were, on average, unpopular. We expect the parameter estimate for WAR to be positive
reflecting national feeling during and immediately after first and second world wars.
(b)
(se)
( 2.7444 ) ( 0.1028)
( 0.2866 )
( 0.2684 )
(1.3009 )
(1.1914 )
( 0.5837 )
( 2.6335)
The signs are as expected. We expect that the parameter for GROWTH to be positive
because society rewards good economic growth. For the same reason we expect the
parameter for GOODNEWS to be positive. We expect a negative sign for the parameter of
INFLATION because increased prices impact negatively on society. We expect the
parameter for PERSON to be positive because a party is usually in power for more than
one term, therefore we expect the incumbent party to get the majority vote for most of the
elections. We expect that for each subsequent term it is more likely that the presidency
will change hands; therefore we expect the parameter for DURATION to be negative. The
sign for PARTY is as expected if one knows that the Democratic party was in power for
most of the period 1916-2000. We expect the parameter for WAR to be positive because
voters are more likely to stay with an incumbent party in a time of war.
169
All the estimates are statistically significant at a 1% level of significance except for
INFLATION, PERSON and WAR. The coefficients of INFLATION and PERSON are
statistically significant at a 5% level of significance, however. The coefficient of WAR is
statistically insignificant even at a level of 10%. Lastly, an R2 of 0.9231 suggests that the
model fits the data very well.
(d)
Using the estimated equation from part (c), a forecast for the 2004 vote is
n
VOTE
2004 = 49.6074 + 0.6909 2.0 0.7751 1.7 + 0.8374 1 + 3.2510 1
n
VOTE
2004 t(0.975,14) se( f ) = 56.473 2.145 3.0707 = (49.89, 63.06)
(f)
For the 2008 election the Republican party has been in power for one more term and so we
set DURATION = 1. Also, the incumbent, George Bush, is not running for election and so
we set PERSON = 0. In the absence of precise information we leave GROWTH and
INFLATION at their earlier values of 2.0 and 1.7, respectively. Also, at the time of writing
these solutions a recession was forecast for the U.S. economy, and so we set GOODNEWS
= 0. These values lead to the following prediction
n
VOTE
2008 = 49.6074 + 0.6909 2.0 0.7751 1.7 + 0.8374 0 + 3.2510 0
n
VOTE
2008 t(0.975,14) se( f ) = 48.757 2.145 3.0343 = (42.25, 55.27)
This interval suggests there is considerable uncertainty about who will win the 2008
election.
170
EXERCISE 7.16
(a)
Mean
Median
Std. Dev.
Skewness
AGE
Kurtosis
19.57407
18
17.19425
0.93851
3.561539
BATHS
1.973148
0.612067
0.912199
6.55344
BEDROOMS
3.17963
0.709496
0.537512
5.751031
FIREPLACE
0.562963
0.49625
-0.25387
1.064451
OWNER
0.488889
0.500108
0.044455
1.001976
POOL
0.07963
0.270844
3.105585
10.64466
PRICE
154863.2
130000
122912.8
6.291909
60.94976
SQFT
2325.938
2186.5
1008.098
1.599577
7.542671
TRADITIONAL
0.538889
0.498716
-0.15603
1.024345
500
400
300
200
100
0
0
500000
1000000
1500000
171
R 2 = 0.7373
3.980833
0.006215
0.190119
0.031506
0.084275
0.067465
0.004275
0.029901
0.056093
0.109970
Std. Error
t-value
0.045895
0.000518
0.020558
0.016611
0.019015
0.017746
0.031581
0.001406
0.017027
0.033355
86.738
11.999
9.248
1.897
4.432
3.802
0.135
21.269
3.294
3.297
p-value
0.0000
0.0000
0.0000
0.0581
0.0000
0.0002
0.8924
0.0000
0.0010
0.0010
SSE = 77.9809
Our initial expectation for the coefficients of the variables BATH, BEDROOMS, and SQFT
is that they should be positive because increasing their values should increase price. We
expect the parameter for AGE to be negative because older houses sell for less. The
parameters for the dummy variables FIREPLACE, POOL and WATERFRONT should be
positive because these features are expected to increase the selling price of a house. With
the exception of BEDROOMS and POOL, all these coefficients have their expected signs
and are significant at a 5% level of significance. Although the coefficients of BEDROOMS
and POOL have the wrong signs, they are not significant at a 5% level. Also, further
thought suggests a negative sign for BEDROOMS may be reasonable. If SQFT is kept
constant and the number of bedrooms is increased, they will be smaller bedrooms that may
reduce the price. With respect to the other variables, traditional houses sell for less and
owner-occupied houses sell for more.
The coefficient of WATERFRONT can be used to tell us the percentage increase or
decrease associated with a waterfront house. On average, a waterfront house sells for
100 ( exp ( 0.10997 ) 1) = 11.62% higher than a house that is not waterfront.
The model fits the data well. We have 73.73% of the variation in ln( PRICE ) explained by
the model and the generalized R2 value, calculated as the squared correlation between
n , PRICE )]2 = 0.8092 .
price and its predictor, is [corr( PRICE
172
After including the variable TRADITIONAL WATERFRONT , the results from estimating
the regression model are:
Dependent Variable: ln(PRICE/1000)
Coefficient
C
AGE
BATHS
BEDROOMS
FIREPLACE
OWNER
POOL
SQFT/100
TRADITIONAL
WATERFRONT
TRAD WFRONT
3.971113
0.006147
0.188258
0.031333
0.087314
0.068370
0.002394
0.030031
0.044913
0.165374
0.172175
R 2 = 0.7373
Std. Error
t-value
0.045946
0.000517
0.020521
0.016570
0.019007
0.017706
0.031513
0.001403
0.017561
0.039951
0.068716
86.430
11.881
9.174
1.891
4.594
3.861
0.076
21.399
2.557
4.139
2.506
p-value
0.0000
0.0000
0.0000
0.0589
0.0000
0.0001
0.9395
0.0000
0.0107
0.0000
0.0124
SSE = 77.9809
Let ln( P0 ) be the mean log-price for a non-traditional house that is not on the waterfront,
and let 9 , 10 and 11 be the coefficients of TRADITIONAL, WATERFRONT and
TRADITIONAL WATERFRONT , respectively. Then the mean log-price for a traditional
house not on the waterfront is
ln( PT ) = ln( P0 ) + 9
The mean log-price for a non-traditional house on the waterfront is
ln( PW ) = ln( P0 ) + 10
The mean log-price for a traditional house on the waterfront is
173
Thus, traditional houses on the waterfront sell for less than traditional houses elsewhere.
The price advantage from being on the waterfront is lost if the house is a traditional style.
The approximate proportional difference in price for houses which are both traditional and
on the waterfront cannot be obtained by simply summing the traditional and waterfront
effects 9 and 10 . The extra effect from both characteristics, 11 , must also be added. Its
estimate is significant at a 5% level of significance.
The corresponding exact percentage price differences are as follows.
For traditional houses not on the waterfront:
The Chow test requires the original model plus an interaction variable of TRADITIONAL
with every other variable. We want to test the hypothesis H 0 : i = 0 for i = 1 to 9, against
the alternative that at least one i 0. Rejecting the null indicates that the equations for
traditional and non-traditional home prices are not the same. The unrestricted model is:
F=
( SSER SSEU )
SSEU ( N K )
( 78.7719-75.7995 ) 9 = 4.6272
75.7995 (1080 18 )
Since 4.627 > F(0.95,9,1062) = 1.889 , the null hypothesis is rejected at a 5% level of
significance. We conclude that there are different regression functions for traditional and
non-traditional styles. Note that SSEU = 75.7995 is equal to the sum of the SSE from
traditional houses (31.058) and the SSE from non-traditional houses (44.741).
174
Given the outcome of the hypothesis test in part (d), for this prediction exercise we use an
equation estimated from observations on traditional houses only. Also, for the number of
bedrooms and the number of bathrooms we use the median values BEDROOMS = 3 and
BATHS = 2.
A prediction for ln ( PRICE 1000 ) is
n
ln
( PRICE 1000 ) = 3.732185 0.00675898 20 + 0.214191 2 + 0.027533 3
+ 0.122849 + 0.097535 + 0.0271036 25
= 5.005958
The natural predictor is
n
n n = exp ln
PRICE
( PRICE 1000 ) 1000 = exp(5.005958) 1000 = 149300
The corrected predictor is
n = PRICE
n n exp ( 2 / 2 ) = 149300 ( 0.2328152 2 ) = 153402
PRICE
c
175
EXERCISE 7.17
(a)
The histogram for PRICE is positively skewed. On the other hand, the logarithm of PRICE
is much less skewed and is more symmetrical. Thus, the histogram of the logarithm of
PRICE is closer in shape to a normal distribution than the histogram of PRICE.
300
250
200
150
100
50
0
100000
200000
300000
400000
500000
120
100
80
60
40
20
0
11.0
11.5
12.0
12.5
13.0
(b)
( 0.0020 )
( 0.0219 )
( 0.0132 )
All coefficients are significant with the exception of those for BATHS and STORIES. All
signs are reasonable, although those for BEDS and STORIES deserve closer scrutiny. They
suggest that, for a given floor area, houses with more bedrooms and/or more stories are
cheaper.
176
Vacancy at the time of house sale reduces the price by approximately 8%. Using the exact
formula this price reduction is
( 0.0554 ) ( 0.0005)
(se)
( 0.0252 )
( 0.0200 )
( 0.0030 )
( 0.0337 )
R 2 = 0.7407
SSEO = 17.5203
( 0.0472 ) ( 0.0005)
( 0.0265)
( 0.0028)
( 0.0285 )
( 0.0178)
R 2 = 0.7257
SSEV = 14.08308
There are some large differences in the coefficient estimates. In particular, the coefficients
of BATHS and STORIES have changed sign. However, in both of these cases the
coefficients are not significant. The remaining variables, whose coefficient are all
significant, have similar effects.
(e)
To carry out a Chow test, we use the sum of squared errors from the restricted model that
does not distinguish between vacant and occupied houses, SSER = 33.38128 and the sum
of squared errors from the unrestricted model that is given by
F=
( SSER SSEU )
SSEU ( N K ) /
( 33.3813-31.6034 ) 6 = 8.14
31.6034 ( 880 12 )
The 5% critical F value is F(0.95,6,868) = 2.109 . Thus, we reject H 0 . The equations for
houses vacant at the time of sale and occupied at the time of sale are not the same.