A.H. Studenmund-Using Econometrics Solution PDF

Answers to Odd-Numbered Exercises

Chapter One: An Overview of Regression Analysis

1-3. (a) Positive, (b) negative, (c) positive, (d) negative, (e) ambiguous, (f) negative.
1-5. (a) The coefficients in the new equation are not the same as those estimated in the previous
equation because the sample is different. When the sample changes, so too can the estimated
coefficients. In particular, the constant term can change substantially between samples; in our
research for this exercise, for instance, we found one sample that had a negative intercept (and
a very steep slope).
(b) Equation 1.21 has the steeper slope (6.38 > 4.30) while Equation 1.24 has the greater intercept
(125.1 > 103.4). They intersect at 9.23 inches above 5 feet (162.3 pounds).
(c) Equation 1.24 misguesses by more than ten pounds on exactly the same three observations
that Equation 1.21 does, but the sum of the squared residuals is greater for Equation 1.24 than
for Equation 1.21. This is not a surprise, because the coefficients of Equation 1.21 were
calculated using these data.
(d) If it were our last day on the job, wed probably use an equation that wed calculate from both
equations by taking the mean, or by taking an average weighted by sample size, of the two.

1-7. (a) The estimated slope coefficient of 5.77 represents the change in the size of a house (in square
feet) given a one thousand dollar increase in the price of the house. The estimated intercept of
72.2 is the value of SIZE when PRICE equals zero; it includes the constant value of any
omitted variables, any measurement errors, and/or an incorrect functional form.
(b) No. All we have shown is that a statistical relationship exists between the price of a house and
its size.
(c) The new slope coefficient would be 0.00577 (5.77/1000), but nothing else would change.
1-9. (a) 101.40 is the estimated constant term, and it is an estimate of the bond price when the Federal
funds rate equals zero, but since zero is outside the normal range of observation, this
interpretation doesnt mean much (as is often the case). 4.78 is an estimate of the slope
coefficient, and it tells us by how much bond prices will change when the federal funds rate
increases by one unit (one percentage point). The sign of the estimated slope coefficient is as
expected; as interest rates rise, the price of existing bonds is expected to fall. We typically do
not develop hypotheses involving constant terms.
(b) See equation 1.17; the equation could have been equivalently stated as Yi = 101.40
4.78Xi + ei
(c) An error term is unobservable and shouldnt be included in an estimated equation from which
we actually calculate a Y. If the question is reworded to ask why a residual isnt included,
then the answer becomes the same as for part (b) above.
2 Studenmund Fifth Edition
(d) If the Federal funds rate increases by 1 percentage point, bond prices will fall by $4.78.
Possible criticisms are: (1) The appropriate interest rate to use when explaining the value of a
long-term asset is the long-term interest rate, not the short-term interest rate. (2) It seems
reasonable that there are more than just one relevant explanatory variable, and (3) Given how
responsive capital markets are to interest rate changes, it seems likely that a monthly (or even
more frequent) data set would provide more observations (and therefore a better fit) for the
same number of years.
1-11. (a) The error term is the theoretical, unobservable, difference between the true (population)
regression line and the observed point. The residual is the measured difference between the
observed point and the estimated regression line.
(b)
Yi 2 6 3 8 5 4
Xi 1 4 2 5 3 4
ei 0.20 0.24 0.12 0.92 0.56 1.76
,i 0.50 0.00 0.00 0.50 0.50 2.00
1-13. (a) Negative, because a high unemployment rate would imply that jobs are scarce; as a result,
employers would not need to advertise much, if at all, to attract applicants for a job opening.
Yes.
(b) While 97% of the variation of HWI around its mean can be explained by UR in this sample, it
seems possible that other variables might matter as well, especially over a longer time period.
When reviewing suggested additional independent variables, the key at this point is to make
sure that every variable (like the cost of help-wanted advertising or the efficiency of
networking alternatives to advertising) is time-series, not cross-sectional.
(c) Yes. The EViews estimates are = 364.0404 and = 46.39477.
0 1
Chapter Two: Ordinary Least Squares
2-3. (a) Because of Equation 2.5, 0 and 1 tend to compensate for each other in a two-variable
model. Thus if 1 is too high, then 0 is likely to be too low.
(b) There are only two possibilities:
i) 0 = 0 + 2 X2
ii) 2 = 0 (and therefore 0 = 0 )
2-5. (a) The squares are least in the sense that they are being minimized.
(b) If R = 0, then RSS = TSS, and ESS = 0. If R is calculated as ESS/TSS, then it cannot be
2 2
2
negative. If R is calculated as 1 RSS/TSS, however, then it can be negative if RSS > TSS,
which can happen if Y is a worse predictor of Y than Y (possible only with a non-OLS
estimator or if the constant term is omitted).
(c) We prefer Model T because it has estimated signs that meet expectations and also because it
includes an important variable (assuming that interest rates are nominal) that Model A omits.
2
A higher R does not automatically mean that an equation is preferred.
Answers to Odd-Numbered Exercises 3
2-7. (a) Even though the fit in Equation A is better, most researchers would prefer equation B because
the signs of the estimated coefficients are as would be expected. In addition, X4 is a
theoretically sound variable for a campus track, while X3 seems poorly specified because an
especially hot or cold day would discourage fitness runners.
(b) The coefficient of an independent variable tells us the impact of a one-unit increase in that
variable on the dependent variable holding constant the other explanatory variables in the
equation. If we change the other variables in the equation, were holding different variables
constant, and so the has a different meaning.
2-9. As the students will learn in Chapters 6 and 7, theres a lot more to specifying an equation than
maximizing R 2 .
2-11. (a) It might seem that the higher the percentage body fat, the higher the weight, holding constant
height, but muscle weighs more than fat, so its possible that a lean, highly muscled man
could weigh more than a less well-conditioned man of the same height.
(b) We prefer Equation 1.24 because we dont think F belongs in the equation on theoretical
grounds. The meaning of the coefficient of X changes in that F now is held constant.
(c) The fact that R 2 drops when the percentage body fat is introduced to the equation strengthens
our preference for 1.24.
(d) This is subtle, but since 0.28 times 12.0 equals 3.36, we have reason to believe that the impact
of bodyfat on weight (holding constant height) is very small indeed. That is, moving from
average bodyfat to no bodyfat would lower your weight by only 3.36 pounds.
The results do indeed match. R = .744, and R 2 = .733.

2
2-13.
Chapter Three: Learning to Use Regression Analysis

3-3. (a) A male professor in this sample earns $817 more than a female professor, holding constant the
other independent variables in the equation.
(b) Most students will expect a negative coefficient, so they will call this an unexpected sign.
Most professors and administrators will expect a positive sign because of the growing
competition among colleges for African-American professors, so they will call this an
expected sign. A key point here is not to change expectations based solely on this result.
(c) R is not a dummy variable because it takes on more than two values. For each additional year
in rank, the ith professors salary will rise by $406, holding constant the other independent
variables in the equation.
(d) Yes. The coefficient is large and, as well learn in Chapter 5, statistically significantly greater
than zero. (In addition, its quite robust.)
(e) S = 36721 + 817(1) + 426(0) + 406(3) + 3539(1) = S = $42,295. This is an approximation
i i
because the actual research included a number of other independent variables. Luckily,
$42,295 is fairly close to what the entire equation cranks out, because most third-year
assistant professors have zeroes for the other variables in the equation. (By the way, this
estimate is ten years old, so its a bit low.)
3-5. (a) Positive.

(b) Obviously, the best equation includes the actual traffic data (which, it turns out, are available).
Since the traffic dummy variable is correlated with the actual traffic variable and since the
new equation has expected signs and reasonable coefficients, it seems slightly better than
Equation 3.5.
(c) No! The theoretical underpinnings of the model are much more important. Of course, the
higher R 2 is certainly a plus.
3-7. Most students at this stage can exactly duplicate the Woodys results.
3-9. (a) Negative, positive, none

(b) Holding all other included explanatory variables constant, a car with an automatic
transmission gets 2.76 miles less per gallon than a model with a manual transmission, and a
car with a diesel engine gets 3.28 miles more per gallon than one without a diesel engine.
(c) Lovell added the EPA variable because he wanted to test the accuracy of EPA estimates. If
these estimates were perfectly accurate, then the EPA variable would explain all the variation
in miles per gallon.
Chapter Four: The Classical Model

4-3. (a) An additional pound of fertilizer per acre will cause corn yield (bushels/acre) to increase by
0.10 bushel/acre, holding rainfall constant. An additional inch of rain will increase corn yield
(bushels/acre) by 5.33 bushels/acre holding fertilizer/acre constant.
(b) No. (This is a typical student mistake.) First, since its hard to imagine zero inches of rain
falling in an entire year, this intercept has no real-world meaning. In addition, recall that the
OLS estimate of the intercept includes the non-zero mean of the error term in order to validate
Classical Assumption II (as explained in the text), so even if rainfall were zero, it wouldnt
make sense to attempt to draw inferences from the estimate of the 0 term unless it was
known that the mean of the (unobservable) error term was zero.
(c) No; this could be an unbiased estimate. 0.10 is the estimated coefficient for this sample, but
the mean of the coefficients obtained for the population could still equal the true F .
(d) Not necessarily; 5.33 could still be close to or even equal to the true value. An estimated
coefficient produced by an estimator that is not BLUE could still be accurate. If the estimator
is biased, its bias could be small and its variance smaller still.
4-5. (a) Classical Assumption II.

(b) Classical Assumption VI.
(c) R: A one-unit increase in yesterdays R will result in a 0.1% increase in todays Dow Jones
average, holding constant the other independent variables in the equation.
M: The Dow Jones will fall by 0.017% on Mondays, holding constant the other variables in
the equation.
(d) Technically, C is not a dummy variable because it can take on three different values. Saunders
assumed (at least implicitly) that all levels of cloud cover between 0% and 20% have the same
impact on the Dow and also that all levels of cloud cover between 21% and 99% have the
same impact on the Dow. In addition, by using the same variable to represent both sunny and
cloudy days, he constrained the coefficient of sun and cloud to be equal.
(e) In our opinion, this particular equation does little to support Saunders conclusion. The poor
fit and the constrained specification combine to outweigh the significant coefficients of Rt1
and M.
4-7. (a) The estimated coefficient of C shows that (for this sample) a one percent increase in the
nonwhite labor force in the ith city adds 0.002 percentage points to the overall labor force
participation rate in that city, holding constant all the other independent variables in the
equation. The estimated coefficient of the dummy variable, D, shows that if a city is in the
South, the labor force participation rate will be 0.80 percentage points lower than in other
cities, holding constant the other explanatory variables in the equation.
(b) Perfect collinearity is virtually impossible in a cross-section like this one because no variable
is a perfect linear function of another; some are closely related, but none is a perfect linear
function.
(c) This does not imply that one of the estimates is biased. The estimates were taken from two
different samples and are quite likely to differ. In addition, the true value may have changed
between decades.
(d) Disagree. Beginners often confuse the constant term with the mean of the dependent variable.
While the estimated constant term shows the value of the dependent variable in the unlikely
case that all of the explanatory variables equal zero, it also includes the mean of the
observations of the error term as mentioned in question 3. part (b) above.
4-9. (a) This possibly could violate Assumption III, but its likely that the firm is so small that no
simultaneity is involved. Well cover simultaneous equations in Chapter 14.
(b) Holding constant the other independent variables, the store will sell 134.4 more frozen
yogurts per fortnight if it places an ad. If we ignore long-run effects, this means that the owner
should place the ad as long as the cost of the ad is less that the increase in profits brought
about by selling 134.4 more frozen yogurts.
(c) The result doesnt disprove the owners expectation. School is not in session during the prime
yogurt-eating summer months, so the variable might be picking up the summer time increases
demand for frozen yogurt from non-students.
(d) Answers will vary wildly, so perhaps its best just to make sure that all suggested variables
are time-series for two-week periods. For students who have read Chapters 14 only, the best
answer would be any variable that measures the existence of, prices of, or advertising of local
competition. Students who have read Chapter 6 might reasonably be expected to try to find a
variable whose expected omitted-variable bias on the coefficient of C is negative. Examples
include the number of rainy days in the period or the number of college students returning
home for vacation in the period.
Chapter Five: Hypothesis Testing

5-5. For N : Reject H:0 0 if 4.42 > tC and 4.42 is negative.
For P : Reject H:0 0 if 4.88 > tC and 4.88 is positive.

For I : Reject H:0 0 if 2.37 > tC and 2.37 is positive.
(a) tC = 1.943; reject the null hypothesis for all three coefficients.
(b) tC = 1.311; reject H0 for all three coefficients.
(c) tC = 6.965; cannot reject the null hypothesis for any of the three coefficients.
5-7. This is a concern for part (a) but not for parts (b) and (c). In part (a), 160 probably is the
coefficient we expect; after all, if our expectation was something else, why did we specify 160 in
the null? In parts (b) and (c), however, it seems unlikely that wed expect zero.
5-9. (a) For all three, H0: 0, HA: > 0, and the critical 5% one-sided t-value for 24 degrees of
freedom is 1.711. For LOT, we can reject H0 because +7.0 > 1.711 and +7.0 is positive. For
BED, we cannot reject H0 because +1.0 < 1.711 even though +1.0 is positive. For BEACH,
we can reject H0 because +10.0 > 1.711 and +10.0 is positive.
(b) H0: 0, HA: < 0, and the critical 10% one-sided t-value for 24 degrees of freedom is
1.318, so we reject H0 because 2.0 > 1.318 and 2.0 is negative.
(c) H0: = 0, HA: 0, and the critical 5% two-sided t-value for 24 degrees of freedom is
2.064, so we cannot reject H0 because 1.0 < 2.064. Note that we dont check the sign
because the test is two-sided and both signs are in the alternative hypothesis.
(d) The main problems are that the coefficients of BED and FIRE are insignificantly different
from zero.
(e) Given that we werent sure what sign to expect for the coefficient of FIRE, the insignificant
coefficient for BED is the most worrisome.
(f) Unless the students have read Chapter 6, this will be a difficult question for them to answer.
Its possible that the dataset is unrepresentative, or that theres an omitted variable causing
bias in the estimated coefficient of BED. Having said that, the most likely answer is that BED
is an irrelevant variable if LOT also is in the equation. Beach houses on large lots tend to have
more bedrooms than beach houses on small lots, so BED might be irrelevant if LOT is
included.
5-11. (a) For the t-tests:
Coefficient: P M S T
Hypoth. Sign: + +
t-value: 5.8 6.3 1.0 3.3
tC = 1.671 reject reject do not reject
(5% one-sided reject
with 60 d.f.,
as close to 73 as Table B1 goes)
(b) No. We still agree with the authors original expectations despite the contrary result.
(c) Keynes point is well taken; empirical results will indeed allow an econometrician to discover
a theoretical mistake now and then. Unfortunately, far too many beginning researchers use
this loophole to change expectations to get right signs without enough thinking or analysis.
(d) Holding all other included explanatory variables constant, an increase in winning percentage
of 150 points will increase revenues by $7,965,000 ($53.1 times 150 times 1000) and thus it
would be profitable for this team to hire a $4,000,000 free agent who can raise its winning
percentage to 500 from 350.
5-13. (a) All the expected signs for the coefficients are positive, so: H0: 0, HA: > 0. (Some
students will see the Wall Street Journal as a competitor and will hypothesize a negative
coefficient for J, but the authors intended the Journals circulation as a measure of what the
Posts circulation would have been without Watergate.)
(b) Coefficient: J S W
Hypoth. Sign: + + +
t-value: 14.27 6.07 1.31
tC = 1.717 reject reject do not
with 22 d.f.)
(c) Assuming that the specification is correct, Watergate had a positive but statistically
insignificant effect on the Posts circulation.
Chapter Six: Specification: Choosing the Independent Variables

6-3. (a) Coefficient: C E M
Hypoth. Sign: + + +
t-value: 4.0 4.0 2.0
with 27 d.f.)
The problem with the coefficient of M is that it is significant in the unexpected direction, one
indicator of a possible omitted variable.
(b) The coefficient of M is unexpectedly negative, so were looking for a variable the omission of
which would cause negative bias in the estimate of M. We thus need a variable that is
negatively correlated with meat consumption with a positive expected coefficient or a variable
that is positively correlated with meat consumption with a negative expected coefficient.
So:
Possible Expected Correlation Direction
Omitted Variable Sign of with M of Bias
+ + +
*
B
F + + +
+ + +
*
W
R +
H +
O +
*
indicates a weak expected sign or correlation.
(c) The only one of the above variables that seems likely to have caused any of the observed
negative bias in M is H, and that variable is a good proxy for the obvious omission of the
general level of quality of cardiac health care, so wed choose to add it. We wouldnt overly
penalize a student who chose to add another variable as long as the alternative was an annual
aggregate variable. (Students who add disaggregate variables would probably benefit from
some extra review sessions.)
6-7. Some students will come to the conclusion that sequential specification searches are perfectly
reasonable in business applications, and they need to be reminded of the biased estimated
coefficients generated by such searches.
6-9. (a) In a supply equation, the coefficient of price will have a positive expected sign because the
higher the price, holding all else constant, the more suppliers would be willing to produce.
(b) The price of inputs (such as labor, seeds, transporation, machinery, and fertilizer), the price of
a production substitute (a crop or product that could be produced instead of the crop or
product being modeled), and exogenous factors (like local growing conditions, local strikes
that dont have an impact on the price, etc.) are just a few examples of important variables in a
supply-side equation.
(c) Lag those independent variables that influence the production decision on a delayed basis. In
particular, lag them by the length of time it takes for that particular event to have an impact on
production. For example, if growers must make production decisions a year before the crop is
harvested, then price should be lagged one year, etc. If a product can be stored at a fairly low
cost, then such a lag might not be appropriate because producers could choose to wait until
prices rose before going to market.
6-11. (a) Coefficient: W T C L

Hypoth. sign: + + + +
t-value: 4.0 3.0 2.0 1.0
tC = 1.697 reject reject reject do not
(5% one sided reject
with 30 d.f.)
(b) Almost any equation could potentially have an omitted variable, and this one is no exception.
In addition, L might be irrelevant. Finally, the coefficient of C seems far too large, suggesting
at least one omitted variable. C appears to be acting as a proxy for other luxury options or the
general quality of the car. (Some students will claim that they shouldnt be expected to know
that 5.8 is too large. Even students who dont know what cruise control is, however, should be
expected to know that $5,800 is too large a portion of a cars price to be a reasonable amount
to add to that price if the car has cruise control, whatever it is!)
(c) We prefer T. Using the four criteria:
Theory: Engine size seems a small component of car price, especially for sedans.
t-score: Ls coefficient is insignificant at the 5% level.
R2 : The overall fit adjusted for degrees of freedom does not change when L is dropped.
Bias: The other coefficients do not change significantly when L is dropped.
(d) The coefficient of L had a t-score of exactly 1.0, so dropping L will have no impact on R 2 . If
the t-score had been less than 1.0, dropping L would have increased R2 , and if the t-score had
been greater than 1.0, dropping L would have decreased R 2 .
6-13. This question is similar to Exercise 6-8 but since the answers to odd-numbered exercises are not in
the back of the text, 13 is more suitable for use on a problem set or exam.
(a) X1 = either dummy variable
X2 = either dummy variable
X3 = Parents educational background
X4 = Iowa Test score
(b) We have two variables for which we expect positive coefficients (Iowa score and Parents
education) and two positive estimated coefficients ( 3 and 4 ), so wed certainly expect X3
and X4 to be those two variables. In choosing between them, its fair to expect a more
significant coefficient for Iowa than for Parents. Next, we have two variables for which we
expect a zero coefficient (the dummies) and two estimated coefficients ( 1 and 2 ) that are
not significantly different from zero, so wed certainly expect X1 and X2 to be the dummies.
There is no evidence to allow us to distinguish which dummy is X1 and which is X2. (Students
who justify this answer by expecting negative signs for coefficients of the two dummies are
ignoring the presence of the Iowa test score variable in the equation that holds constant the
test-taking skills of the student.)
(c) Coefficient: D D PE IT
Hypoth. sign: 0 0 + +
t-value: 1.0 0.25 +2.0 +12.0
tC = 2.093 do not do not
(5% two-sided reject reject
with 19 d.f.)
tC = 1.729 reject reject
(5% one-sided
with 19 d.f.)
(d) As you can see, we used a one-sided test for those coefficients for which we had a specific
prior expectation but a two-sided test around zero for those coefficients for which we did not.
6-15. (a) (i) The coefficient of CV is 0.19 with a SE ( ) of 0.23 and a t-score of 0.86. The R 2
is .773, and the rest of the equation is extremely similar to Equation 5.14 except that the
coefficient of CVN falls to 0.48 with a t-score of 1.86.
(ii) The coefficient of N is 0.00054 with a SE ( ) of 0.063 and a t-score of 0.0086. The
R 2 is .766, and the rest of the equation is identical (for all intents and purposes) to
Equation 5.10.
(b) Theory: P is a price ratio, and while its possible that a price ratio would be a function of the
size of a market or a country, its not at all obvious that either variable would add
anything since CVN is already in the equation.
t-score: Both t-scores are insignificant.
R2 : R 2 falls when either variable is added.
bias: None of the coefficients change at all when N is added, so it clearly is irrelevant. The
omission of CV does change the coefficient of CVN somewhat, making it likely that
CV is redundant since CVN is in the equation.
(c) Since CVN = f[CV/N], it would make little theoretical sense to include all three variables in
an equation, even though technically you dont violate Classical Assumption VI by doing so.
(d) Its good econometric practice to report all estimated equations in a research report, especially
those that were undertaken for specification choice or sensitivity analysis.
Chapter Seven: Specification: Choosing a Functional Form

7-3. (a) Linear in the coefficients but not the variables
(b) Linear in the coefficients but not the variables
(c) Linear in the coefficients but not the variables
(d) Nonlinear in both
(e) Nonlinear in both
7-5. (a) For beginning econometrics students, its usually more interesting go back and read the
articles that started the controversy than to wade through the current literature. See, for
starters, A. W. Phillips, The Relation Between Unemployment and the Rate of Change of
Money Wage Rates in the United Kingdom, 18611957, Economica, Nov. 1958, pp. 283
299 and R. G. Lipsey, A Further Analysis, Economica, Feb. 1960, pp. 131. As indicated in
the text, our favorite analysis of the issue is Nancy Wulwicks Phillips Approximate
Regression, in Neil de Marchi and Christopher Gilbert (eds.) History and Methodology of
Econometrics (Oxford: Oxford University Press, 1989), pp. 170188.
(b) A linear form, a double-log form, a polynomial form (with no 1U t term), and a semi-log
form all could have been used (with varying success).
(c) There is evidence that macropolicies and external shocks were shifting the short-run Phillips
curve upward and to the right during this period, causing a plotting of the annual
inflation/unemployment rates to appear to be positively sloped. While there is strong
theoretical support for a vertical long-run Phillips curve, students who let the data convince
them that a short-run Phillips curve was positively sloped during this time period are making
the same mistake that many Keynesians did for over a decade.
7-7. (a) This question contains a hidden difficulty in that the sample size is purposely not given. D
students will give up, while C students will use an infinite sample size. B students will
state the lowest sample size at which each of the coefficients would be significantly different
from zero (listed below), and A students will look up the article in Econometrica and
discover that there were 125 cotton producers and 26 sugar producers, leading to the tCs and
hypothesis results listed below.
Coefficient: 1C 2C 1S 2S
t-value: 30.667 3.000 4.214 1.914
Lowest d.f. at
which signif. (5%) 1 2 2 7
5% tC given
actual d.f. 1.645 1.645 1.714 1.714
(So all four coefficients are significantly different from zero in the expected direction.)
(b) A double-log function seems quite appropriate.
(c) Since the equations are double-log, the elasticities are the coefficients themselves:
Industry Labor Capital

Cotton 0.92 0.12
Sugar 0.59 0.33
(d) The sum indicates whether or not returns to scale are constant, increasing or decreasing. In
this example, Cotton is experiencing increasing returns to scale while Sugar is experiencing
decreasing returns to scale.
7-9. (a) and both have positive hypothesized signs because as either L or K increases holding
the other constant, output should increase.
(b) H0: 0; HA: > 0; for both.
(c) Reject H0 if t-value > 1.708 and t-value is positive.
(d) L: t = 2.02; K: t = 5.86, so reject H0 for both.
(e) The relative prices of the two inputs need to be known.
7-11. (a) polynomial (second-degree, with a negative expected coefficient for age and a positive
expected coefficient for age squared)
(b) double-log (We would not quibble with those who chose a linear form to avoid the constant
elasticity properties of a double-log.)
(c) semilog (lnX)
(d) linear (All intercept dummies have a linear functional relationship with the dependent variable
by definition.)
(e) inverse (Most students will remember from the text that a U-shaped polynomial typically is
used to model a cost-curve and will want to apply it here. The problem is that the telephone
industry appears to be an industry in which costs continually decrease as size increases,
making an inverse our choice.)
7-13. (a) Coefficient: LQ A V

Hypoth. sign: +
t-value: 4.0 2.0 0.13
with 20 d.f.)
(b) Q constant, A and V non-constant.

(c) No. The coefficient of V is quite insignificant, and the equation (simplified from an
unpublished article) is flawed to boot. Note, however, that the violence may be causing the
absentee rate to rise, so that the significant coefficient for A does indicate some support for
the charge.
(d) In our opinion, this is a classic case of spurious correlation because actual total output
appears on both sides of the equation, causing almost all of the fit by definition. If we could
make one change, wed drop LQ from the equation, but we worry that little will be left when
we do.
7-15. (a) Coefficient: S G D SD GD

Hypoth. sign: + + + +? +
t-value: 4.5 0.4 2.9 5.0 2.3
tC = 1.645 reject do not reject do not reject
(5% one-sided reject reject
with infinite d.f.)
(b) As Primeaux puts it (on page 622 of his article), A duopoly firm of small size spends more
than a monopoly firm of the same size. However, as scale increases, eventually, the duopoly
firm spends less.
(c) Again, from page 622, There is no difference between monopoly and duopoly firms at zero
rates of growth in sales. However, as growth takes place, the duopoly firms engage in more
sales promotion activity.
Chapter Eight: Multicollinearity

8-3. Perfect multicollinearity; each can be stated as an exact function of the other two. To solve the
perfect multicollinearity problem, one of the three explanatory variables must be dropped.
8-5. (a) Using the mean method:

X3i = 16HEIGHTi + WEIGHTi
Using the regression method:
X3i = 6.4HEIGHTi + WEIGHTi
(b) Using the mean method:
X3i = 5Ii + Pi
X3i = 2.7Ii + Pi
(c) Using the mean method:

X3i = 2.01Yi + YDi
X3 = 3.63Yi + YDi
8-7. (a) No; no explanatory variable is a perfect function of another.

(b) Yes; income in any quarter will be strongly correlated with income in previous quarters.
(c) If all the variables were specified in terms of first differences, its likely that much of the
multicollinearity would be avoided.
8-9. (a) Coefficient: PC PQ Y C N

t-value: 0.801 1.199 0.514 1.491 1.937
tC = 1.725 at the 5% level, so only N is significantly different from zero in the expected
direction.
(b) The obviously low t-scores could be caused by irrelevant variables, by omitted variables
biasing the estimated coefficients toward zero, or by severe imperfect multicollinearity.
(c) The high simple correlation coefficient between Y and C indicates that the two are virtually
identical (redundant), which makes sense theoretically. The r between the two price variables
is not as high, but mild multicollinearity can still be shown to exist.
(d) Yt and Ct both serve as measures of the aggregate buying power of the economy, so they are
redundant, and one should be dropped. It doesnt matter statistically which one is dropped, but
Yt seems analytically more valid than Ct, so wed drop C. Dropping one of the price variables
would be a mistake, since they have opposite expected signs. While forming a relative price
variable is an option, the low level of multicollinearity, the reasonable coefficients, and the
possibility that Ct is also multicollinear with prices (so dropping it will improve things) all
argue for making just one change.
8-11. (a) Coefficient: M B A S

t-value: 5.0 1.0 1.0 2.5
tC = 1.645 reject do not do not reject
(5% one-sided reject reject
with infinite d.f.)
(b) The insignificant t-scores of the coefficients of A and B could have been caused by omitted
variables, irrelevance, or multicollinearity (a good choice, since thats the topic of this
chapter). In particular, since most students graduate at about the same age, the collinearity
between A and B must be fairly spectacular (Stanford gave us no clues).
(c) Its probably a good idea, since the improvement in GPA caused by extra maturity may
eventually be offset by a worsening in GPA due to separation from an academic environment.
(d) We believe in making just one change at a time to best be able to view the impact of the
change on the estimated regression. Thus, our first choice would be to drop either A or B
(wed prefer to drop A, but on theoretical grounds, not as a result of the unexpected sign).
Switching to a polynomial before dropping one of the redundant variables will only make
things worse, in our opinion.
8-13. (a) Coefficient: EMP UNITS LANG EXP

t-value: 0.98 +2.39 +2.08 4.97
tC = 1.725 do not reject reject reject
with 20 d.f.)
(b) The functional form is semilog left or semilog (ln Y). Semilog left is an appropriate functional
form for an equation with salary as the dependent variable because salaries often increase in
percentage terms when an independent variable (like experience) increases by one unit.
(c) Theres a chance that an omitted variable is pulling down the coefficient of EMP. However,
its more likely that EMP and EXP are redundant (since they in essence measure the same
thing) and are causing multicollinearity.
(d) These results confirm our opinion that EMP and EXP are redundant.
(e) If we knew that this particular school district doesnt give credit for teaching experience
elsewhere, then it would make sense to drop EXP. Without that specific knowledge, however,
wed drop EMP because EXP includes EMP.
(f) Theory: The two variables are redundant
t-test: The coefficient of EMP is indeed insignificantly different from zero.
R2 : R 2 does indeed increase when EMP is dropped from the equation.
Bias: Students will have to work backwards to calculate the SE( )s, but if they do, theyll
find that the coefficient of EXP does indeed change by a standard error. This is
exactly what wed expect to happen when we drop a redundant variable from an
equation; the coefficient of the remaining redundant variable will adjust to pick up the
effect of both variables.
8-15. The ever-innovative Rao and Miller used this example (developed by Professor Maurice G.
Kendall) to show that the inspection of simple correlation coefficients is not an adequate test for
multicollinearity.
(a) Since R and L are obviously correlated but R and (L R) are not, many beginning students
will want to drop either R or L from Model A, but this would leave out the difference between
leg lengths that is the inherent causal variable.
(b) As Rao and Miller point out (on page 48 of their text), the implicit estimates of the
coefficients are identical because the conditions imposed on the residuals for estimation in
either case are implicitly the same. To calculate the coefficients of one model from the other,
multiply out the 2 term of Model B, reconfigure to correspond to Model A, and solve for the
coefficients of Model B: 0 = 0 , 1 = (1 2 ), and 2 = 2 .
(c) Since the coefficient estimates are identical for every sample, their distributions must also be
identical, meaning that the two models are identically vulnerable to multicollearity.
(d) If you drop L from Model A, then the linkage between the Models cited in the answers above
is lost.
Chapter Nine: Serial Correlation

9-3. m t = 28187 + 16.86P . The Durbin-
The coefficient estimates for all three orders are the same: HS t
Watson d results differ, however:
(a) DW = 3.08
(b) DW = 3.08
(c) DW = 0.64
Note that any order change will be likely to change the DW except for the reverse order (for
which DW will always be exactly the same).
9-7. (a) dL = 1.49; DW = 0.81 < 1.49, so wed reject the null hypothesis of no positive serial
correlation.
(b) This is not necessarily a sign of pure serial correlation. Its reasonable to think that residuals
from the same country would have more in common than would residuals from other
countries (that is, the model could be consistently underestimating for France and
overestimating for Canada, producing six positive residuals followed by six negative
residuals). As a result, the DW for such pooled datasets will at times give indications of serial
correlation when it does not indeed exist. The appropriate measure is the Durbin-Watson d for
each country taken individually, since the order of the countries will influence the overall DW
statistic, and that order is arbitrary.
(c) If the serial correlation is impure, then a variable needs to be added to the equation to help
distinguish better between the countries. If the serial correlation is judged to be pure,
however, then generalized least squares might be applied one country at a time. It is possible
to specify different first-order serial correlation coefficients for each country and then
estimate one pooled regression equation.
9-9. (a) An outlier in the residuals can occur even if no outlier exists in the dataset if all the Xs are
very low (or very high) simultaneously, producing an unusually low or high Y. In such a
situation, Y would be dramatically lower (or higher) than Y.
(b) When an extreme outlier exists in the residuals, the Durbin-Watson test will not necessarily
produce an accurate measure of the existence of serial correlation because the outlier will give
the appearance of severe negative serial correlation. That is, there will be a large
(et et1) of one sign followed by a large (et et1) of the opposite sign, so the two large
squared terms will move the DW dramatically toward four. In such a circumstance, some
researchers will drop the outlier from the DW calculation (but not from the data set). A one-
time dummy equal to one in the observation with the outlier residual will solve this problem
by in essence setting the residual equal to zero; this is almost (but not quite) the same as
dropping the observation.
9-11 (a) As weve mentioned, we prefer a one-sided Durbin-Watson d test, so with K = 3 and N = 40,
the 5% critical values are dL = 1.34 and dU = 1.66. Since DW = 0.85 is less than DL, we can
reject the null hypothesis of no positive serial correlation.
(b) Coefficient: L P W
Hypoth. sign: + + +
t-value: 0.04 2.6 3.0
tC 2.423 do not reject reject
with 40 closest to 36 in Table B-1 d.f.)
(c) The estimated coefficient of P looks reasonable in terms of size and significance, but the one
for L looks pathetically small. We would never expect such similar variables to have such
dramatically different coefficients. Many students will want to drop L, pointing out that the
Lakers almost always play well, so fans may not pay much attention to exactly how the
Lakers are doing at any given point. Wed guess that a long losing streak would show the true
relevance of this variable, however.
(d) Pure serial correlation is certainly a possibility, but the fact that some fans are most
interested in games played late in the season implies that an omitted variable with a temporal
pattern exists. Wed want to include such a variable before concluding that pure serial
correlation exists.
(e) We prefer dropping the first observation to including zeroes for L and P, but an even better
alternative might be to use last seasons winning percentages as proxies for this seasons for
opening day (or even a few games thereafter). While opening day might have always sold out
in the past, there is no guarantee that it always will be sold out in the future.
9-13. (a) This is a cross-sectional dataset and we normally wouldnt expect autocorrelation, but well
test anyway since thats what the question calls for. DL for a 5% one-sided, K = 3, test is
approximately 1.61, substantially higher than the DW of 0.50. (Sample sizes in Table B-4
only go up to 100, but the critical values at those sample sizes turn out to be reasonable
estimates of those at 200.) As a result, we can reject the null hypothesis of no positive serial
correlation, which in this case seems like evidence of impure serial correlation caused by an
omitted variable or an incorrect functional form.
(b) Coefficient: G D F
Hypoth. sign: + + ?
t-value: 3.5 7.0 2.5
tC = 1.645 reject reject reject
(5% one-sided
with infinite d.f.)
We certainly have impure serial correlation. In addition, some students will conclude that F
has a coefficient that is significant in the unexpected direction. (As it turns out, the negative
coefficient could have been anticipated because the dependent variable is in percentage terms
but F is in aggregate terms. Wed guess that the more food a pig eats, the bigger it is, meaning
that its chances of growing at a high rate are low, thus the negative sign.)
(c) The coefficient of D is significant in the expected direction, but given the problems with this
equation, wed be hesitant to conclude much of anything just yet.
(d) In this case, the accidental ordering was a lucky stroke (not a mistake), because it allowed us
to realize that younger pigs will gain weight at a higher rate than their older counterparts. If
the data are ordered by age, positive residuals will be clustered at one end of the data set,
while negative ones will be clustered at the other end, giving the appearance of serial
correlation.
9-15. (a) For the first equation (standard errors in parentheses):

n = 3.9 + 0.17lnUSD + 1.12lnSY 0.037lnSP
1nSDL t t t t
(0.071) (0.062) (0.030)
t= 2.47 18.24 1.22
N = 25 (annual 19601984) R = .980 DW = 0.70
2
Coefficient: 1 2 3
Hypothesized sign: + + +
tC = 1.721, so: reject reject do not
reject
For the second equation:

n = 3.29 + 0.143lnUSD + 1.062lnSY
1nSDL t t t
(0.067) (0.035)
t= 2.15 30.50
N = 25 (annual 19601984) R 2 = 0.980 DW = 0.59
Coefficient: 1 2
Hypothesized sign: + +
tC = 1.717, so: reject reject
(b) The specification criteria now imply that SP is an irrelevant variable: R 2 increases slightly
when SP is added, but SPs coefficient is not significantly different from zero, and the other
coefficient estimates do not change more than one standard error when SP is added.
(c) For both equations, DW is far below the critical value for a 5% one-sided test, so we can
reject the null hypothesis of no positive serial correlation. (For the first equation, 0.70 < 1.12,
and for the second equation 0.59 < 1.21.)
(d) Once again, such a small improvement in the DW statistic is not evidence that the serial
correlation is impure.
(e) Since the SDL equations imply that SP is irrelevant, it makes sense to run GLS on the second
equation:
n = 2.12 + 0.094lnUSD + 0.914lnSY
1nSDL t t t
(0.058) (0.069)
t= 1.64 13.15
N = 24 (annual 19601984) R 2 = .994 = 0.61
Chapter Ten: Heteroskedasticity
10-5. m i = 1273.2 + 0.72I

(a) CO R 2 = .97
i
(0.044)
t = 16.21
where: CO = average consumption

I = average income.
(b)
n
i ) = 29.54 2.34(lnI i ) R 2 = .39
2
1n(e
(0.94)
t = 2.49
tC = 3.499 at the 1% level (two-tailed) so we cannot reject the null hypothesis of

homoskedasticity.
(c) Some students will feel that when there is only one independent variable, weighted least
squares (with X as the proportionality factor) will produce the same answer as OLS because
you could always multiply through the WLS equation by Z to get the OLS result. Since the
dependent variables are different, such a hypothesis can be proven to be incorrect. We can
illustrate this with the data from this example by using Ii as the proportionality factor,
obtaining:
COi /I i = 0 / I i + 1 + u i
(where ui is a homoskedastic error term)

weighted least squares produces:
n
COi /Ii = 2400.1/ Ii + 0.40 R 2 = .94
(0.14)
t = 2.77
Thus even with only one independent variable, weighted least squares does indeed make a
difference. Note that the relative magnitudes of the estimated coefficients have now
switched due to the WLS regression; 2400.1 in the WLS regression is comparable to
1273.2, not 0.72, in the OLS regression.
(d) It does not suggest anything should necessarily be done about it. Interestingly, there is another
potential cause of heteroskedasticity in this model. That is, the variables are means of ranges
of differing widths and different sample sizes. Thus it would seem reasonable to think that
each of these means might come from a distribution with a different variance, and some sort
of heteroskedasticity would be fairly likely.
R = .122, N = 33, so NR = 4.026 < 21.7 = the critical 1% Chi-square value with 9 d. f; so we
2 2
10-7.
cannot reject the null hypothesis of homoskedasticity. Thus both tests show evidence of
heteroskedasticity.
10-9. (a) Multicollinearity and heteroskedasticity (but not positive serial correlation) appear to exist.
Wed tackle the multicollinearity first. Since the heteroskedasticity could be impure, you
should get the best specification you can before worrying about correcting for
heteroskedasticity.
(b) For all intents and purposes, the two equations are identical. Given that, and given the
reasonably strong t-score of STU, wed stick with Equation 10.29. Note that the ratio of the
FAC/STU coefficients is far more than 10/1 in Equation 10.29. This means that Equation
10.29 overemphasizes the importance of faculty, compared to Equation 10.30. (On second
thought, whats wrong with overemphasizing the importance of faculty?)
(c) Both methods show evidence of heteroskedasticity. For instance, if TOT = Z, the Park test t =
4.85 > 2.67, the critical t-value for a two-sided, 1 percent test with 57 degrees of freedom
(interpolating).
(d) We dont provide a t-score for the coefficient of (1/TOT) because that coefficient is an
estimate of the intercept of the original equation. (Note, however, that we do provide a t-score
for the constant of the WLS equation.) This is a good example of coefficient switching.
The pathetic fit is mainly due to the fact that the dependent variable in the WLS equation
(VOL/VOT) is significantly more difficult to predict than is VOL, where size alone can
explain most of the variation. In addition, it turns out that not one of the WLS variables in this
equation has any theoretical validity at all, making a bad situation even worse.
(e) There are many possible answers to this question. Using EViews, the easiest might be to
reformulate the equation, using SAT and STU/FAC (the student/faculty ratio) as proxies for
quality:
n
VOL/TOTi = 0.067 + 0.00011SATi 0.0045STU i /FACi
(0.00007) (0.0013)
t = 1.59 3.44
R = .19 N = 60
2
DW = 2.11
10-11. (a) Coefficient: W U lnP

Hypoth. sign: + ?
t-value: 10.0 1.0 3.0
tC = 1.66 reject do not reject
with 90 d.f.,interpolating)
(b) We disagree, mainly because econometrics cannot prove that discrimination is the cause of
any differences in employment. Less importantly, we disagree that the nondiscriminatory
expected value of the coefficient of W is 1.00; for starters, a constant term and other variables
are in the equation.
(c) Heteroskedasticity seems reasonably unlikely, despite the cross-sectional nature of the dataset,
because the dependent variable is stated in per capita terms.
(d) The two-sided 1% critical t-value is approximately 2.64 (interpolating), so we cannot reject
the null hypothesis of homoskedasticity.
(e) The theory behind P or lnP seems quite weak (despite its significance). Our preference would
be to change P to a non-aggregate measure, for instance the percentage of the population that
is black in the ith city, or some measure of the unemployment benefits available in the ith city.
10-13. (a) Coefficient: P I Q A S T

Hypoth. sign: + + + +?
t-value: 0.97 6.43 3.62 1.93 1.6 2.85
tC = 1.684 do not reject reject reject do not do not
(5% one-sided with reject reject reject
40 d.f., closest to 43)
The last two variables cause some difficulties for most students when hypothesizing signs.
Our opinion is that having more suburban newspapers should hurt metropolitan newspaper
circulation but that the number of television stations is a measure more of the size of a city
than of the competition a newspaper faces. By the way, we see Q as a proxy for quality and A
as an endogenous variable (note that the authors did indeed estimate the equation with 2SLS,
a technique beyond the scope of this chapter).
(b) Heteroskedasticity seems extremely likely, since larger cities will have larger newspaper
circulation, leading to larger error term variances. Using a two-sided 1% test, we can reject
the null hypothesis of homoskedasticity since 3.13 > 2.66, the critical t-value with 60 degrees
of freedom (closest to 48).
(c) Heteroskedasticity, multicollinearity, and omitted variables all seem likely.
(d) While its tempting to divide the equation through by population (or reformulate the equation
by making the dependent variable per capita circulation), this would dramatically lessen the
equations usefulness. Instead, we would attempt to improve the specification. Reasonable
answers would include attempting to reduce some of the multicollinearity (redundancy)
among the independent variables, trying to find a better measure of quality than the number of
editorial personnel or substituting the number of major metropolitan newspaper competitors
for S and T.
10-15. (a) To test for serial correlation, first run:
St = 0.73 + 0.22It + 0.46ln(1 + Vt )

(0.05) (0.16)
t = 4.50 2.85
N = 58 (monthly) R 2 .556 DW = 1.54

Since DW = 1.54, the Durbin-Watson test is inconclusive at the 5% one-sided level. Lott and
Ray, the source of these data, reach the same inconclusion but with slightly different numbers.
This means that there is a chance that we transcribed the dataset incorrectly. If so,
comparability with Lott and Ray is reduced, but the quality of the exercise for students is
undiminished.
(b) As mentioned in the text, we do not recommend running GLS if the DW is in the inconclusive
range. Our major reason is that a poor estimate of rho can introduce bias into an equation
while pure serial correlation will not. This is especially true when the t-scores are not being
used to decide whether to drop a variable, as is the case in this example.
(c) A mere doubling in the size of the dependent variable should not, in and of itself, cause you to
be concerned about heteroskedasticity in a time series equation. If the dependent variable had
gone up ten times, then heteroskedasticity (or nonstationarity, depending on the situation)
would be a real concern.
(d) A Park test with T as a proportionality factor produces a t-score of 0.45, providing no
2 2
evidence whatsoever of heteroskedasticity. A White test produces an R of .20, for an NR of
11.6, which indicates heteroskedasticity at the 5% level but not at the 1% level we typically
suggest for heteroskedasticity tests.
(e) A Park test with T as a proportionality factor produces a t-score of 0.99, once again providing
no evidence of heteroskedasticity.
(f) Our preference would be to reformulate the equation (perhaps after testing for
nonstationarity). We do not favor blindly using WLS in a time-series equation.
Chapter Twelve: Time-Series Models

12-3. (a) The double-log functional form doesnt change the fact this is a dynamic model and that Y
and M almost surely are related by a distributed lag.
(b) In their relationship to M, both Y and R have the same distributed lag pattern over time, since
the lambda of 0.60 applies to both (in logs).
(c) Many students will run a Durbin-Watson d test on this equation and conclude that there is no
positive serial correlation because DW = 1.80, which is greater than 1.65, the 5% one-sided
critical dU. The best students wont be so sure, however, because the Durbin-Watson d is biased
toward 2 in the presence of a lagged dependent variable and the question doesnt provide
enough information to run an LMSC. One student in a hundred will do more research, find out
about Durbins h test (which we no longer include in the text), and use the h test to conclude
that we cannot reject the null hypothesis of no serial correlation because h =
(1 0.5 * 1.80)26/(1 26 * 0.10 ) = 0.59, which is less that 1.96, the 5% critical z-value.
2
12-5. n = 243 + 5.2AD + 1.9AD + 3.1AD + 1.0AD + 3.3AD

(a) SALES t t1 t2 t3 t4
n
(b) SALES = 38.86 + 2.98AD + 0.79SALES
t t1
12-7. (a) Second-order serial correlation means that the current observation of the error term is a
function of the previous two observations of the error term.
(b) et = a0 + a1Xt + a2Yt1 + a3et1 + a4et2 + ut
There would be 2 degrees of freedom in the test because there are two restrictions in the null
hypothesis (a3 = a4 = 0).
(c) e t = 11.8 0.22At + 0.04St1 0.06et1 0.25et2
R = .066 N = 23 (19782000)
2
LM = N * R = 23 * 0.066 = 1.52 < 5.99, the 5% critical Chi-square value with 2 d.f., so we
2
cannot reject the null hypothesis of no second-order serial correlation.

12-9. We suggest that the farmers rethink either the form of their equation or their expectations. Their
current equation posits that corn growth is a distributed lag function of rainfall, a not unreasonable
idea. Unfortunately, lambda is restricted to between zero and one, so the likelihood of observing a
negative lambda is small.
12-11. (a) 2.01 > 3.22, so we cannot reject the null hypothesis of nonstationarity for Y.
(b) 0.72 > 3.22, so we cannot reject the null hypothesis of nonstationarity for r.
(c) 4.16 > 3.22, so we cannot reject the null hypothesis of nonstationarity for CO.
(d) 0.04 > 3.22, so we cannot reject the null hypothesis of nonstationarity for I.
All four variables appear to be nonstationary, at least at the 2.5% level. This is a bit surprising,
because interest rate variables often are stationary, and its not a coincidence that the interest rate
is the only variable to have a negative t-score.
Chapter Thirteen: Dummy Dependent Variable Techniques

13-3. (a) This equation is a linear probability model, and as such it can encounter at least two problems
in addition to any normal specification problems it might have: R 2 is not an accurate
measure of the overall fit, and Yi is not bounded by 0 and 1.
(b) Some might question the importance of PV in the equation, and others might suggest adding
variables to measure recent changes in profits or book value, but our inclination would be to
switch to a logit before analyzing the specification much further.
(c) The best way to win this argument is to note that the linear probability model produces
nonsensical forecasts (greater than 1.0 and less than 0.0).
13-5. It wont take long for students to realize that with a logit, Di is never greater than 1 or less than 0.
13-7. (a) Its reasonable conclude that if Di = 2, then the logit computer program will balk at taking the
log of a negative number (2). As mentioned in the text, however, logit programs iterate using
Equation 13.7 (or a version thereof), so its possible that a software package could produce
some sort of estimates. (Weve never seen one, however.)
(b) With a linear probability model, the answer is unambiguous. The estimated coefficients of
WN and ME will double in size. If we divide each coefficient by two, the theoretical
meanings will be unchanged from the original model.
13-9. In all three models, we find evidence that A is an irrelevant variable. (Note that in the datafild
D = J)
(a) Di = 0.22 0.38Mi 0.001Ai + 0.09Si
(0.16) (0.007) (0.04)
t = 2.43 0.14 2.42
R = .29
2
N = 30 R 2 p = .81
n
(b) ln[D i /(1 Di )] = 5.27 2.61Mi 0.01Ai + 0.67Si
(1.20) (0.04) (0.32)
t= 2.17 0.25 2.10
R p = .76
2
n
(c) Z i = F (Pi ) = 3.05 1.45Mi 0.006Ai + 0.39Si
1
(0.63) (0.02) (0.18)

t = 2.31 0.26 2.20
R 2 p = .76 iterations = 5 LR = 13.85
13-11. (a) Three

(b) The three dependent variables are ln(Pu/Pc), ln(Pj/Pc), and ln(Pa/Pc), which are the log of the
ratio of the probability of attending the choice in question divided by the probability of
attending your college.
Chapter Fourteen: Simultaneous Equations

14-3. (a) If ,2 decreases, Y2 and then Y1 decrease.
(b) If ,D increases, QD increases, and then QS increases (equilibrium condition) and Pt increases.
(Remember that the variables are simultaneously determined, so it doesnt matter which one is
on the left-hand side.)
(c) If , increases, CO increases, and then Y increases and YD increases.
14-5. All these cases can be shown to involve a positive correlation between the ,s and the Ys.
14-7. (a) A: Predetermined = 2 < 3 = # of slope coefficients, so underidentified.

B: Predetermined = 2 = 2 = # of slope coefficients, so exactly identified.
(b) Note that X2 is endogenous to this system, so:
Y1: Predetermined = 3 < 4 = # of slope coefficients, so underidentified.
Y2: Predetermined = 3 > 1 = # of slope coefficients, so overidentified.
X2: Predetermined = 3 = 3 = # of slope coefficients, so exactly identified.
(c) Note that you can consider the change in Y to be endogenous to the system with a non-
stochastic equation in which it equals Yt Yt1. Given this, there are six predetermined
variables, Yt1, Et, Dt, Mt, Rt1 and Gt, so the identification properties of the four stochastic
equations can be determined by using the order condition (which is necessary but not
sufficient):
COt: Predetermined = 6 > 1 = # of slope coefficients, so overidentified.
It: Predetermined = 6 > 4 = # of slope coefficients, so overidentified.
Rt: Predetermined = 6 > 3 = # of slope coefficients, so overidentified.
Yt: Predetermined = 6 > 3 = # of slope coefficients, so overidentified.
14-9. (a) It = 270 + 0.19Yt 8.26rt1

(0.01) (11.62)
t= 13.93 0.71
R = .933
2
N = 28DW = 0.38
(b) Yt = 235.17 + 0.77Gt 0.51NXt + 1.51Tt + 0.67COt1 + 36.4rt1
(0.23) (0.17) (0.16) (0.10) (9.32)
t= 3.32 2.94 9.43 6.81 3.91
R = .999
2
N = 28
(c) It = 26.5 + 0.19 Yt 8.48rt1
(Standard errors obtained from this estimation are biased and should be disregarded.)
(d) It = 265.5 + 0.19 Yt 8.48rt1
(0.01) (11.6)
t= 13.9 0.73
R = .933
2
N = 28 DW = 0.38
14-11. Most reasonable models of the labor market are simultaneous and therefore potentially involve
simultaneity bias and should be estimated with 2SLS.
14-13. (a) OLS estimation will still encounter simultaneity bias because price and quantity are
simultaneously determined. Not all endogenous variables will appear on the left-hand side of
a structural equation.
(b) The direction of the bias depends on the correlation between the error term and the right-
hand-side endogenous variable. If the correlation between the error term and price is positive,
as it most likely is, then the simultaneity bias will also be positive.
(c) Three: stage one: P = f (YD, W)
stage two: QD = f ( P , YD) and QS = f ( P , W)
(d) OLS: QD = 57.3 0.86P + 1.03YD
Q S = 167.5 + 3.95P 1.42W
2SLS: QD = 95.1 6.11 P + 2.71YD
Q S = 480.2 + 13.5 P 5.50W
14-15. (a) All three variables are nonstationary. In Exercises 12-10 and 12-11, we showed that both COt
and Yt are nonstationary. If COt is nonstationary, then so too must be COt1. Since YDt and Yt
are highly correlated, its reasonable to think that if one is nonstationary then so too is the
other. As a test of this, a Dickey-Fuller test on YDt produces a t-score of 0.02 < 3.22 , the
critical value, so we cannot reject the null hypothesis. This is further evidence that YDt is
nonstationary.
(b) If we run a Dickey-Fuller test on the residuals, we get 2.8 > 2.048 , a 2.5% 1-sided 28 df
critical t value (which we use as an approximationsee the text). Thus we can reject the null
hypothesis of nonstationarity and conclude that the residuals are stationary. This implies that
Equation 14.31 is reasonably cointegrated.
(c) A lag model is more likely to be cointegrated dynamic model distributed because the lagged
values of the dependent variable that appear on the right hand side of the equation should
have the same degree of nonstationarity as the dependent variable.
(d) We agree with the majority of applied econometricians who think that the concept of
cointegration is unrelated to the estimation technique. As a result, we do not hesitate to
recommend the use of the Dickey-Fuller test when testing 2SLS residuals for cointegration.
There are those who disagree, however, by pointing out that nonstationarity in a truely
simultaneous system implies that a test for cointegration should go beyond testing the
residuals of just one of the equations in the system.
Chapter Fifteen: Forecasting

15-3. Answers to this question will depend on the forecasted X values chosen.
15-7. If the answers to Exercise 15-6 were calculated correctly, then calculating backwards will
indeed reproduce the original series.
15-9. (a) e99 = Y99 Y99 = 27 27.5 = 0.5

(b) Y100 = 0 + 1(27) 0.5( 0.5) = 27.25
Y101 = 0 + 1(27.25) = 27.25
Y102 = 0 + 1(27.25) = 27.25
Chapter Sixteen: Statistical Principles

16-3. The mean is 16.89 and the standard deviation is 6.43. Thus the 1999 P/E ratio was more than two
standard deviations above the mean.
16-5. Standardized scores: 1.9, 0.0, and 0.8, raw score: 90.
16-7. Z = (70 66)/2.5 = 1.60. P[Z > 1.60] = 0.0548.
16-9. People who have visited France for pleasure more than once during the past two years are more
likely to have had good experiences than are people who visited France just once and then never
returned and/or people making their first visit to France.
16-11. The standard deviation is the square root of the variance, or 0.686, and the 95% two-sided tc with
34 degrees of freedom is approximately 2.03, so a 95% confidence interval is 6.19 +
( )
2.03 0.686 / 35 or 6.19 + 0.24. There is a 95% probability that a confidence internal
constructed in this fashion will include the true value of the mean prediction of the population, so:
(a) No. This says nothing about how accurate or inaccurate the forecasters are.
(b) No. If anything, we might estimate that approximately 95% of the forecasts are in an interval
equal to our estimate of the mean plus or minus 2 standard deviations of the individual
observations:
6.19 + 1.96 (0.686) or 6.19 + 1.34.
16-13. The calculated t-value = (13.333 11.2)/[5.499/ 12 ] = 1.344. Since a two-sided 5% critical t-
value is 2.201, we cannot reject the null hypothesis.
16-15. All three statements are wrong:

(a) This statement confuses P[A if B] with P[B if A].
(b) Statistical significance is not the same as practical importance. We cannot tell the estimated
size of the effect from the P value.
(c) Not being able to reject the null hypothesis is not the same as proving the null hypothesis to
be true.
The correct interpretation is that the test shows that if the cold vaccine is ineffective, there is a
0.08 probability of observing a difference this large (or larger) between the fractions of each
group that caught colds.
Lecture Notes
Chapter One: An Overview of Regression Analysis

1-1. What is Econometrics?
Heres the perfect opportunity for an inspirational talk about how econometrics is used today. You
could mention anything from the high percentage of academic articles and government studies that
involve econometrics in one way or another to how knowledge of econometrics will help students
do well in business school and (hopefully) business.
1-2. What is Regression Analysis?

Many instructors feel that students learn more if an example is introduced early on and then the
nitty gritty details of this section are applied to the same example as the class progresses. One
possibility for such an example is a simple model of something that all the students are familiar
with, for instance a model of the number of students (N) that sign up for a class as a function of
how easy a grader the professor of the class is (measured by the professors average GPA) and the
number of majors in the department (NM):
+ +
N i = f (GPA i , NMi ) + ,i = 0 + 1GPA i + 2 NMi + ,i
or a model of the number of tickets sold for a football game (T) as a function of the rank (1 = first
place) of the home team in the league standings (P) and the number of miles away the visiting
teams campus is (M):

Ti = f (Pi , Mi ) + ,i = 0 + 1Pi + 2 Mi + ,i
By the way, youre doing well if only a couple students expect a positive sign for the coefficient
of P!
1-3. The Estimated Regression Equation

Then in this section you could continue the example by providing an estimated equation, complete
with numbers and hats. A real key here is to focus on the difference between the error term and
the residual.
1-4. A Simple Example of Regression Analysis

This weight/height example is worth taking the time to go into in detail; we refer back to it a
number of times later in the text. One way of improving the example is to ask your male students
to take a moment on the first day of class to write down their heights and weights for you. When
you cover this example on the second or third day, you could announce the student-estimated
coefficients and compare them to the textbooks, thus bringing the example to life and also
planting the seeds of the idea that estimated regression coefficients come from a sampling
distribution of beta-hats.
1-5. Using Regression to Explain Housing Prices

The estimates in this section represent the simplest possible model that can be estimated from the
data presented in Section 11.7, the Housing Price Interactive Exercise. Since that section discusses
the data and the theory behind hedonic models more than Section 1.5 does, you might consider
looking at Section 11.7 before you present Section 1.5 in class.
1-6. Summary and Exercises

For a problem set on this chapter, wed recommend Exercises 1-4, 1-11, and 1-12.
Chapter Two: Ordinary Least Squares

2-1. Estimating Single-Independent-Variable Models with OLS
Students often get bogged down in Section 2.1.2, and we urge you to require the students to work
through an estimation example (like Section 2.1.3 or Exercise 2-4) on their own. No matter how
hard we try, we still encounter students who have difficulty understanding what regression
estimation actually is until theyve done it for themselves.
If you want an easy estimation problem for which the coefficient estimates are not in the text, use
Ando and Modiglianis data in Exercise 10-5. The coefficient estimates for this dataset are:
m i = 1273.2 + 0.72I
CO R 2 = .97
i
(0.044)
t = 16.21
where: CO = average consumption

I = average income.
2-2. Estimating Multivariate Regression Models with OLS

The key here is understanding the meaning of multivariate regression coefficients. In our
experience, there are two levels of such understanding. At the first level, the student can recall that
all the other independent variables in the equation are held constant only if asked directly. At the
second level, the student understands the fact that the other regressors are held constant well
enough to use and interpret the estimated coefficients correctly in an example. We think its vital
to cover this concept in enough detail and from enough different points of view that all students
can reach the second level by the end of the chapter.
2-3. Evaluating the Quality of a Regression Equation

This short section lists a number of items that should be considered when evaluating the quality of
a regression equation. The most important concept to communicate is that fit is just one of many
attributes of an equation that should be considered.
2-4. Describing the Overall Fit of the Estimated Model

One of the problems with introducing R and then immediately replacing it with R 2 is that some
2
2
students ask why we bothered mentioning it at all. Our answer is that its important to know R in
its own right and that understanding R makes the advantages of R 2 more obvious.
2
Unfortunately, if you do a good enough job of presenting the benefits of R 2 then some students
will become curve-fitters for the rest of the semester. Its worthwhile to suggest to these true
believers that they reread (aloud with a friend, if they dare) the imaginary conversation on page
58 that points out the problems with relying on R 2 as the primary measure of the quality of a
regression.
Lecture Notes 29
2-5. An Example of the Misuse of R 2

We do not return to this water-demand example, and most students can understand the point of the
example without it being discussed, so it can be skipped in class if necessary. Note, however, that
this example and the interactive regression learning example at the back of this instructors
manual are based on the same research.

For a problem set on this chapter, wed recommend Exercises 2-4, 2-7, 2-8 and 2-11.
Chapter Three: Learning to Use Regression Analysis

Weve heard from some users of the text that they cover Chapter 3 after Chapter 7 or (less frequently) after
Chapter 10, and we have no objections to this practice. To help these instructors, weve made slight
revisions in order to allow Chapter 3 to be skipped without much loss of continuity. The two topics in
Chapter 3 that are needed in future chapters are some coverage of dummy variables (pages 6970) and
some exposure to the Woodys dataset.
3-1. Steps in Applied Regression Analysis

This section is best taught around an example. Indeed, if youre going to cover the Woodys
example in detail and youre short on time, Sections 3.1 and 3.2 can be presented simultaneously
in class. If you have the time, however, itd be better to present the six steps with a distinct
example (from your own research, for example, perhaps a bit simplified), before going on to the
Woodys example.
3-2. Using Regression Analysis to Pick Restaurant Locations

This Woodys restaurant location example is one of the two most important examples in the book.
We return to it in virtually every chapter, so the time you spend going over the example in class
will pay off in enhanced student understanding of a number of different topics. In this chapter, the
example puts the six steps in applied regression analysis to work, acts as an illustration of
estimation for those who are still stuck on Section 2.1.2, and is a first introduction to what the
computer output of a packaged regression program looks like. A good project assignment is to ask
the students to estimate this regression themselves on your computer system. If you grade this
assignment pass/fail and allow students to work together, it will help students with computer
phobias (or who still dont quite know what a regression is) to work with other students and make
some progress in an unthreatening environment.

For a problem set on this chapter, wed recommend Exercises 3-3, 3-5, 3-7, and 3-10.
Chapter Four: The Classical Model

4-1. The Classical Assumptions
The students wont want to hear you say it, but the classical assumptions simply must be known
cold by econometrics students. There is little value to a week spent studying serial correlation or
heteroskedasticity if the underlying assumption thats being violated isnt understood. As a result,
any trick or threat you can think of to get the students to learn the classical assumptions is
probably fair to the students in the long run. Its sad but true that the most effective way we know
to get even the weakest students to learn the assumptions is to tell them that the assumptions will
be on the next midterm (and then follow through).
4-2. The Sampling Distribution of

A very effective demonstration technique to bring this concept home quite dramatically is to give
each student his or her own dataset drawn from a larger population and to ask each student to
calculate a -hat from that sample. If you then collect the calculated s and graph them, it
produces a fairly normal-looking distribution of estimates. Such an in-class example requires you
to develop a different small subsample for each student and also requires each student to run a
small regression, but we find that the results are worth the effort.
One way to cut down on the instructor workload of the previous suggestion is to ask each student
to run a weight-height regression on a dataset (males only) that they collect on their own. To keep
the assignment manageable, of course, the sample size should be fairly low and some sort of
computer facilities should be available. Such an exercise can serve as a students first introduction
to the computer if the Woodys example of the previous chapter was not assigned to be estimated
individually.
4-3. The Gauss-Markov Theorem and the Properties of OLS Estimators

Needless to say, this short section is quite important. While we dont recommend that you
require it, note that a proof of the Gauss-Markov Theorem is in the answer to Exercise 4-8 in
Appendix A.
4-4. Standard Econometric Notation

Some students who are unfamiliar with the expectation function will have trouble with Section
4.4. These students often have an easier time if they think of an expected value as a mean. Some
instructors may want to save time by presenting Sections 4.1 and 4.4 at the same time.

For a problem set on this chapter, wed recommend Exercises 4-4, 4-5, and 4-9.
Chapter Five: Hypothesis Testing

For some groups of students, you may want to cover Chapter 16 before starting Chapter 5. For the
majority, however, we find that presenting basic statistical concepts as students need them is much more
effective for student learning than is a complete review of statistical inference, but we recognize that not
everyone agrees on this issue.
5-1. What is Hypothesis Testing?

To give students exposure to more than one example of hypothesis testing, its often worth the
effort to use a different example in class than we used in the text. For instance, consider a supply
equation for cotton:
+ + ?
Ct = f(AHt , PCt 1, PSt 1, Mt ) + i
Ct = 0 + 1AHt + 2 PCt 1 + 3PSt1 + 4 Mt + t
where: Ct = thousands of bales of cotton produced in year t

AHt = thousands of acres of cotton harvested in year t
PCt1 = the real price of a thousand pounds of cotton in year t 1
PSt1 = the real price of a thousand bushels of soybeans in year t 1
Mt = price index for farm machinery in year t
Lecture Notes 31
This is a typical student effort at such an equation. Acres harvested is not tautologically related to
production (as acres planted might be if technology was constant during the sample), but it is
likely to be jointly determined with production. Still, a positive sign would be expected. The
price of cotton is lagged one time period because of the length of time it takes to switch crops; a
positive sign is expected. Soybeans can be planted on the same kinds of land on which cotton is
planted, and so a negative sign is expected. Finally, the price of cotton machinery (negative
expected sign) was unavailable, so the student instead chose the average price of all farm
machinery, leading to an ambiguous expected sign (since cotton is fairly labor intensive, an
increase in price in farm machinery might shift farmers into cotton production).
Based on this theory, the appropriate null and alternative hypotheses would be:
1: H0: 1 0, HA: 1 > 0
2: H0: 2 0, HA: 2 > 0
3: H0: 3 0, HA: 3 < 0
4: H0: 4 = 0, HA: 4 0
5-2. The t-Test

Continuing on with the cotton example, the appropriate border value for all four null hypotheses
is zero, so the formula for the t-statistic to use is Equation 5.3:
t k = k / SE( k )
The equation will be estimated for annual U.S. data from 1947 through 1972, so there will be
21 degrees of freedom. Given this, and given a 5 percent level of significance, the appropriate
decision rules and critical t-values (from Statistical Table B.1) are:
1: Reject H0 if | t1 | > 1.721 and if t1 > 0
2: Reject H0 if | t 2 | > 1.721 and if t 2 > 0
3: Reject H0 if | t 3 | < 1.721 and if t 3 < 0
4: Reject H0 if | t 4 | > 2.080
5-3. Examples of t-Tests

Continuing on with the cotton example, the estimated equation is:
Ct = 624 + 0.33AH t + 1929PCt 1 154PSt1 + 12.4Mt

(0.10) (1078) (82) (7.4)
t = 3.16 1.79 1.89 1.67
(these t-scores are computer-calculated and differ from those calculated by hand because of
rounding.)
t-tests of the first three coefficients are all one-sided tests, and in all three cases we can reject the
null hypothesis. This is because the calculated t-score is greater in absolute value than the critical
t-value and has the same sign as the alternative hypothesis for all three.
A t-test on the cofficient of M is appropriately a two-sided test, however. We cannot reject the
null hypothesis of no effect for this variable because the calculated t-score, 1.67, is not greater
than the critical t-value of 2.080.
5-4. Limitations of the t-Test

This section sets the stage for our discussion in Chapter 6 of the problems associated with using
the t-test as a method for selecting independent variables. While the section is self-explanatory, it
also can be applied to the cotton production equation if desired. For example, the high t-score for
AH does not imply theoretical validity or importance. Indeed, it is one of the two worst variables
in the equation. By the way, weve not yet covered multicollinearity, but there is a fairly high
simple correlation coefficient in this dataset. The r between AHt and PCt1 is 0.792.

For a problem set on this chapter, wed recommend Exercises 5-2, 5-5, 5-6a, 5-9, and 5-13.
5-6. Appendix: The F-Test

Sections 5.6.1 and 5.6.2 are fairly simple, since they involve little more than the comparison of an
F-score (printed out by the computer) with a critical F-value in the back of the book. We find that
it enhances student understanding of statistical hypothesis testing and of the t-test to distinguish
between tests of hypotheses about one coefficient and tests of multiple hypotheses.
The textbook does not include any exercises for the appendices, so this instructors manual will
include a problem that can be used as an in-class teaching example (or as an exercise on a problem
set or as a question on an exam) for each appendix in the text.
For Appendix 5.5, a possible example/exercise/exam question would be:
Suppose you want to test the hypothesis that the first 20 restaurants in the Woodys example are
different in some way from the rest of the sample. Given the following two regression results, use
the F-text to test the null hypothesis that the slope coefficients are the same in the two samples
(also called a Chow test):
That is, given:

= 89,977 9693N + 0.431P + 1.67I
A:Yi i i i
(3132) (0.122) (0.77)

t= 3.10 3.52 2.18
N= 20(observations 1-20) R 2 = .554
RSSA = 3.890E + 09 (RSS for this sample)
= 119,761 9460N + 0.324P + 0.68I

B: Yi i i i
(2792) (0.091) (0.79)

t= 3.39 3.56 0.86
N= 13 (observations 21-33) R 2 = .596
RSSB = 1.68E + 09 (RSS for this sample)
Test the hypotheses:

H0: NA = NB; PA = PB; IA = IB
HA: H0 is not true.
at the 5 percent level of significance.

Lecture Notes 33
Answer: The appropriate F-statistic to use is:

F = [(RSST RSSA RSSB ) /(K + 1)]/
[(RSSA + RSSB ) /(N + M 2[K + 1])]
where: RSST = the residual sum of squares for the entire sample (6.13E + 09 as in Table 3.2 on
page 81 of the text)
N = observations in sample A (20)
M = observations in sample B (13)
K = number of explanatory variables (3)
and F has K + 1 numerator and N + M 2 (K + 1) denominator degrees of freedom. The decision
rule to use is to reject the null hypothesis if F exceeds the critical F-value of 2.76 found in
Statistical Table B.2 for 4 and 25 degrees of freedom:
Reject H0 if F > 2.76
As it turns out, substitution into the F formula produces:
F = [(0.56E + 09)/(4)]/[5.57E + 09)/(25)]
= (0.14E + 09)/(0.223E + 09) = 0.628
Since F is less than 2.76, we cannot reject the null hypothesis. That is, the coefficients are not
significantly different at the 5 percent level of significance.
Chapter Six: Specification: Choosing the Independent Variables

6-1. Omitted Variables
The demand for chicken is one of the two most important examples in the text and is referred to in
almost all the remaining chapters. As a result, we recommend that you take the time to go over the
example while discussing omitted and irrelevant variables in class.
Equation 6.7 is sometimes difficult for students to understand from just the examples in the book,
so another example (or exam question) might be helpful:
In the Woodys equation, what would be the expected bias on the coefficient of the competition
variable if we were to drop the population variable from the equation?
Well, Equation 6.7 becomes:
bias = p f(rN, P )
The expected sign for the population coefficient is positive, and youd expect the simple
correlation coefficient between population and competition to be positive (because the more
densely populated an area is, the more restaurants youd expect to be built . . . indeed, r turns out
to be + 0.726), so:
bias = p f(rN, P ) = + + = +
So wed expect dropping P from the equation to cause positive bias in N , and thats what
happens:
Yi = 84,439 148N i + 2.32I i
(1778) (0.66)
t = 0.84 3.50
N = 33 R = .258
2
Note that the coefficient of N has increased from 9075 (in Equation 3.5) to 1487, which is the
direction it would move if it had been positively biased by the omission of P.
6-2. Irrelevant Variables and
6-3. An Illustration of the Misuse of Specification Criteria

To continue the example above, if we had started with the Woodys equation with only N and I as
above, we might have concluded (mistakenly) that N is irrelevant, because without N, the equation
looks like:
Yi = 77,543 + 2.34I i R 2 = .265

(0.66)
t = 3.54
That is, if we apply the four criteria for an irrelevant variable:
Theory - Perhaps Woodys patrons are so dedicated that they eat there no matter what else
is available?
t-test - The t-score is 0.84, insignificantly different from zero, indicating irrelevance.
R2 - R2 rises when N is dropped, indicating that it is irrelevant.
Bias - When N is dropped, the coefficient of I doesnt change at all, an indication that N
is irrelevant (or that P and N are uncorrelated).
Thus wed mistakenly come to the conclusion that N was irrelevant when in fact the problem was
being caused by the omitted variable P!
6-4. Specification Searches

For more on these issues, see Edward E. Leamer, Lets Take the Con out of Econometrics,
American Economic Review, March 1983, pp. 3143; and Michael McAleer, et al., What Will
Take the Con Out of Econometrics? American Economic Review, June 1985, pp. 293313.
6-5. An Example of Choosing Independent Variables

This example turns out to be an excellent way for students to learn the principles outlined in the
previous sections. Most students first impulse is to include every variable and then drop the ones
that have insignificant coefficients. As mentioned previously, this approach is almost sure to cause
bias. By the way, this is a great place to ask students to collect their own data sets and compare
estimated coefficients with each other and with the book. Finally, this dataset was collected at a
college that avoids the use of multiple choice questions on tests.

For a problem set on this chapter, wed recommend Exercises 6-2, 6-3, 6-11, and 6-13.
6-7. Appendix: Additional Specification Criteria

Our choices of RESET, AIC and Schwarz for this section were not based on an opinion that they
are the best or even the most-used additional specification criteria. Instead, we looked for criteria
that are easy to use and that represent different approaches to specification choice. (For instance,
AIC and SC are criteria pure and simple, but RESET can more appropriately be called a test.)
Suggestions of criteria to add for the next edition would be appreciated.
Lecture Notes 35
At any rate, for Appendix 6.7, a possible exam Question/exercise/example would be:
For practice with our additional specification criteria, return to the Woodys dataset, drop I from
the equation, and attempt to detect this specification error using:
(a) Ramseys RESET.
(b) Akaikes Information Criterion
(c) The Schwarz Criterion
Answer:
(a) Using the Y version of RESET, we obtain: F = [(6.13328 10 4.92506 10 )/3]/[(4.92506
4 9 9
10 )/26] = 19.14 > 2.99 = FC, the closest in the table to a 5% critical value with 3 and 26
9
degrees of freedom. We thus reject the null hypothesis that the coefficients of the added terms
are jointly equal zero and conclude that there is a specification error according to RESET.
(b) For Y = f (N, P, I) + AIC = 22.12
For Y = f(N, P) + AIC = 22.24
Thus the original specification (with I) has a lower AIC and is preferable according to that
criterion.
(c) For Y = f (N, P, I) + SC = 22.30
For Y = f (N, P) + SC = 22.37
Thus the original specification (with I) has a lower SC and is preferable according to that
criterion.
Chapter Seven: Specification: Choosing a Functional Form

7-1. The Use and Interpretation of the Constant Term
Some instructors may disagree with the no exceptions point of view we take in the text, and we
think students at this level would probably benefit from hearing both sides of the debate. In
particular, counter-examples, while not convincing to us, may make for an interesting class
session.
7-2. Alternative Functional Forms

The translog function mentioned in footnote 6 combines three different functional forms, the
double-log, the polynomial quadratic, and an interaction term, to come up with an equation
uniquely suited to estimating various kinds of cost functions:
C = 0 + 1Q + 2Q2 + 3B + 4 B2 + 5QB +
where: C = the log of total cost of production

Q = the log of the level of output
B = the log of the number of offices or plants
The translog cost function allows the production of the item to exhibit increasing or decreasing
costs as production is increased, and it even lets different economies of scale exist over different
ranges of output. While the translog function in general is too complex to warrant detailed
consideration in class, it serves as an excellent example of the innovative ways in which various
functional forms can be combined to meet the requirements of complex economic theories. For
more, see Laurits R. Christensen and and William H. Greene, Economies of Scale in U.S.
Electrical Power Generation, Journal of Political Economy, August 1976, pp. 655676.
7-3. Lagged Independent Variables

A good first example of a lagged independent variable is that of a price variable in a supply
equation for an agricultural good. For example, Section 5.1 of this Instructors Manual estimated
the supply of cotton (C) as a function a number of variables, including the price of cotton lagged
one year:
Ct = 624 + 0.33AHt + 1929PCt 1 154PSt 1 + 12.4Mt

(0.10) (1078) (82) (7.4)
t = 3.16 1.79 1.89 1.67
7-4. Using Dummy Variables

The two key concepts here are that:
1. you must have one fewer dummy variable than you have conditions and
2. that if you have more than one dummy variable describing a situation, the coefficient of a
particular dummy represents the impact of that condition on the dependent variable compared
to the omitted condition (holding all non-related variables in the equation constant).
7-5. Slope Dummy Variables

One interesting extension of the use of slope dummies is piecewise regression (also called
segmented regression), where the D = 1 if X is greater than a specified value X* and D = 0 if not.
A piecewise regession allows the slope of a linear relationship to change, as in the old kinked
demand curve of oligopoly theory. If the equilibrium price is considered X*, then a kinked
demand curve for an individual oligopoly would predict a different slope at prices greater than X*
than at lower prices because competitors will match price decreases (changing industry supply)
but not price increases.

For a problem set on this chapter, wed recommend exercises 7.4, 7.7 (which purposely omits the
sample size), 7.15, and 7.16.
Chapter Eight: Multicollinearity

8-1. Perfect vs. Imperfect Multicollinaerity
Perfect multicollinearity is so unusual that it often can be covered in class in less than five
minutes. A typical example of perfect multicollinearity that results from a student mistake is a
variable that does not change at all during the sample period. In such a case, the variable is
perfectly multicollinear with the constant term. Similarly, other students include dummy variables
for all four seasons in an equation that also includes the constant term. They encounter perfect
multicollinearity because the sum of the four dummy variables will always equal one.
8-2. The Consequences of Multicollinearity

Both examples are used again later in the text. The student consumption function is used to
illustrate remedies for multicollinearity in Section 8.4, while the demand for gasoline example is
used as an example of heteroskedasticity in Section 10.5.
8-3. The Detection of Multicollinearity

A refresher in the simple correlation coefficient may be in order here; at a minimum, make sure
that the students have read Section 2.4.2.
Lecture Notes 37
8-4. Remedies for Multicollinearity

The last paragraph of Section 8.4.4 warns against combining cross-sectional and time series data
sets to avoid multicollinearity. The basic reason we do not advocate such pooled datasets has to
do with the dynamic or time-wise interpretation of the coefficients. Time series coefficients show
the impact on the dependent variable of a one-unit change in the independent variable for a
specified unit of time, such as a quarter or year (holding the other Xs constant). Cross-sectional
coefficients measure the same impact over an indefinite time period and therefore can be
interpreted as measuring long-run effects (where long-run is not precisely quantifiable in time-
series units). Thus mixing the two makes the interpretation of the coefficients quite difficult in
situations where the coefficients are likely to change depending on whether theyre long-run or
short-run. For more on this, see Edwin Kuh and John R. Meyer, How Extraneous are Extraneous
Estimates? Review of Economics and Statistics, Nov. 1957, pp. 380393.
Also note that such pooled data sets have potential inherent heteroskedasticity (if the variances of
the error terms of the different pooled samples are different, as would seem likely). Finally,
pooling makes it difficult to test or adjust for serial correlation without doing so for each time
series before pooling. This is because the first observation of one time series will follow the last
observation of the previous time series. In such a case, there usually is no meaning to a model
which posits that one observation of the error term is a function of the previous one.
8-5. Choosing the Proper Remedy

The fish/Pope study in Section 8.5.2 is a good one for class use because students seem to enjoy it
and because alternative specifications can not only be discussed but also estimated, since the data
are given.

For a problem set on this chapter, wed recommend exercises 8-6, 8-9, 8-11, and 8-12.
8-7. Appendix: The SAT Interactive Regression Learning Exercise

The fact that weve devoted almost 40 pages of this text to interactive regression learning
exercises lets you know how potentially valuable we feel they can be, and we urge you to take
advantage of them. One way to use the SAT interactive exercise of this section is to discuss it in
class (after assigning the reading up through but not including Section 8.7.2). We try to let the
class discuss the issues and then vote on which specification to start with, which regression run to
move to, etc. We find its best to let the students make their choices and then give them feedback
after their decisions, but others choose to give feedback as they go along. Such an in-class exercise
will promote active learning and also prepare the students to do the interactive exercise in Chapter
11 on their own.
Chapter Nine: Serial Correlation

9-1. Pure vs. Impure Serial Correlation
Instructors who choose to spend only a short amount of time on this section will suffer no loss of
continuity as long as students understand that serial correcation can be caused by a specification
error.
9-2. The Consequences of Serial Correlation

This topic is one of the hardest in elementary econometrics for beginning students to understand,
and instructors are urged to pay special attention to it.
9-3. The Durbin-Watson d Test

One way to liven this section up is to ask one student to randomly choose a number between zero
and four (the DW), then ask another student to pick a sample size and a number of explanatory
variables and finally ask the rest of the students to decide what these numbers imply about serial
correlation.
9-4. Remedies for Serial Correlation

Section 9.4.2 is certainly worth covering even if you encourage the use of GLS.

Chapter Ten: Heteroskedasticity

10-1. Pure vs. Impure Heteroskedasticity
10-2. The Consequences of Heteroskedasticity

See our comments on Sections 9.1 and 9.2.
10-3. Testing for Heteroskedasticity

We know that our decision to focus on the Park test will cause some difficulty for instructors who
typically use another test, and our inclusion of the White test is meant to offset the problem. In
addition, the rest of the text is worded flexibly enough to accomodate whichever test the instructor
decides to use.
While we understand the theoretical problems with the Park test, we find that its conclusions are
typically the same as more complex tests. More importantly, beginning students seem to
intuitively understand the Park test more readily and find it easier to calculate than other tests.
10-4. Remedies for Heteroskedasticity

The coefficient-switching that goes on when WLS is used (described at the top of page 365)
sometimes is difficult for students to understand right off the bat. In such a situation, you might
consider pulling the answer to Exercise 10-5 out of the first section of this instructors manual and
using the exercise (especially part c) as an in-class example. That is, students who have trouble
understanding Equation 10.16 sometimes see things more clearly when there is only one
independent variable and when coefficient estimates are included (so that the similarity of
magnitudes can be seen).
10-5. A More Complete Example

This reprise of the demand for gasoline example from Section 8.2.2 is an extremely useful in-class
exercise, but it can be left to the students own reading without any loss of continuity.

*
For a problem set on this chapter, wed recommend exercises 10-2, 10-4, 10-5, 10-13 and 10-7
*
( only if you want students to have experience with the White test as well as with the Park test).
Lecture Notes 39
Chapter Eleven: A Regression Users Handbook

11-1. A Regression Users Checklist and Guide
The checklist is mainly a reference for student use and doesnt deserve much class time. The
guide, however, is a useful summary of the previous five chapters and serves as a good review for
an examination or as a good preparation for doing the interactive exercise. We often find that a
class devoted to reviewing the items in Table 11.2 synthesizes student knowledge that otherwise
was fragmented by chapter and topic.
11-2. Running Your Own Regression Project

You may want to assign this section well before you get to Chapter 11, as it lays out the steps
necessary for a student to complete his or her own independent project.
Once the students have completed this section, they should be ready to run their own regressions,
following the six steps of applied regression first outlined in Chapter 3.
Its usually worth your time to require that students get their topics approved by you before they
begin any estimations. Topics should not be approved unless the student has found an adequate
data set (including all the independent variables in the first regression). This approval requirement
will help students avoid projects with inadequate data (like a study of the black market value of
the Vietnamese piaster during the Vietnam war) or which are fairly difficult without more
advanced techniques (like an aggregate investment function).
11-3. Economic Data

The internet is becoming the best source of data for students. A good place is start is the
textbooks website: <www.aw-bc.com/studenmund>. The website contains links to good data
sources, and well continually update those links as the sources (or their addresses) change.
11-4. The Ethical Econometrician

Before assigning a project or requiring that an interactive regression learning exercise be done
independently, its a good idea to review the problems associated with sequential specification
searches. Otherwise, the students may draw the wrong conclusion from the lists of many different
specifications back to back.
11-5. Practical Advice for Applied Econometricians

Were delighted that Peter Kennedy coauthored this section and allowed us to include in it some
material from his Guide to Econometric. It should be extremely useful for all applied
econometricians.
11-7. Appendix: The Housing Price Interactive Exercise

We urge you to require this interactive exercise as part (or all) of a problem set. Once a student
has finished this assignment, he or she should be ready to tackle an individually chosen research
project without too much help from you. As mentioned in the Hints in Appendix A, our favorite
model on theoretical grounds is:
+ + + +
P = f (S, N, A, A 2 , Y, CA) +
The results of the estimation of this particular specification are:
P = 182 + 0.09S 29.9N 1.86A + 0.0156A2 + 0.005Y + 0.89CA

(0.01) (5.0) (0.94) (0.0097) (0.001) (11.7)
t = 7.14 5.98 1.98 1.60 3.58 0.08
R = 0.90
2
N = 43 DW = 1.76
We admit that CAs coefficient is absurdly insignificant, but we still think the variable belongs in
the equation.
Chapter Twelve: Time-Series Models

12-1. Dynamic Models
While most students can identify Equation 12.3 as a dynamic mode (or distributed lag equation)
when discussing the chapter, they will completely miss the boat when analyzing a similar equation
outside the context of the chapter. We find that only repeated practice will help solve this
problem.
12-2. Serial Correlation and Dynamic Models

As we hint at in this section, some econometricians disagree with the majority view that serial
correlation is more likely in dynamic models. Their position is that omitted variables introduce
bias whenever they are correlated with included variables and that lagged errors therefore
introduce bias here but not in other cases.
12-3. Granger Causality

The purpose of this section is to present the idea, not to cover it fully. You will lose little by
skipping this section.
12-4. Spurious Correlation and Nonstationarity

This section attempts to walk the narrow line between being understandable by a college
sophomore and doing justice to the material in a fairly short space. Wed love to get your
feedback on this section, particularly on the usefulness of the boxes material on page 441.

For a problem set on this chapter, wed recommend exercises 12-2, 12-5, 12-6, 12-9,
and 1211.
Lecture Notes 41
Chapter Thirteen: Dummy Dependent Variable Techniques

13-1. The Linear Probability Model
We use the womens labor force example thoughout the chapter and even have an exercise on it at
the end of the chapter, so its probably worth mentioning in class.
13-2. The Binomial Logit Model

This section is the heart of the entire chapter, and its worth quite a bit of time. Students at first
will balk at equations like 13.7, but if you work through enough examples, theyll come to realize
that a logit isnt much harder than OLS. We think that going through the example in Section
13.2.3 in class will help bring this realization about.
Other good in-class examples are: 1. how students chose whether or not to attend your college or
university and 2. what determines whether or not a student will graduate within four (or five)
years. Finally, some students will be excited to learn that logits are used by marketing research
companies to plan advertising and by political campaign media consultants to help predict and
impact election votes.
13-3. Other Dummy Dependent Variable Techniques

Little will be lost if you skip this section, but we find that its well worth the time.

Chapter Fourteen: Simultaneous Equations

14-1. Structural and Reduced-Form Equations
Our notation does not change to the double-subscript 12 system typically used in the discussion
and analysis of simultaneous systems. We feel that switching notation for such a short chapter
would probably do more harm than good.
14-2. The Bias of Ordinary Least Squares (OLS)

If Section 12.2.1 and/or Section 14.6 have or will be assigned, this is the perfect time to make the
point that all three sections basically are talking about the same problem. That is, the violation of
Classical Assumption III means that the observations of the error term are no longer independent
of the explanatory variables, and OLS attributes variation in Y caused by to one of the Xs,
causing bias.
14-3. Two-Stage Least Squares (2 SLS)

Most students will learn 2SLS better and a lot more quickly if theyre asked to estimate an
equation with 2SLS by hand (that is, estimating the reduced-forms themselves with OLS and
then using the estimated Ys in place of the actual Ys on the right-hand side of an OLS estimate of
the second stage). After completing this project, most students will feel fairly comfortable with
2SLS. In addition, if the project also includes required OLS estimates and packaged 2SLS
estimates of the structural equations, then the students will also be aware of the differences
between 2SLS and OLS estimates and between 2SLS-packaged SE ( s) and 2SLS by hand
SE ( s) . Exercise 14-9 provides just such a project, and the data required for the estimation are in
Table 14.1.
14-4. The Identification Problem

Heres an in-class example you can use to give the students experience with using the order
condition:
Use the order condition to determine the likely identification properties of the equations in the
following system:
1. Y1 = f(Y2 , X1, X2 , 1 )
2. Y2 = f (Y3 , Y4 , X3 , 2 )
3. Y3 = f (X 4 , X5 , 3 )
4. Y4 = Y1 + Y3
where the s are classical error terms.
(answer: Equation three is recursive to the system and doesnt provide us with any information
that will help identify the other equations. In addition, equation 4 is a nonstochastic identity and
need not be identified. Thus there are two endogenous variables, Y1 and Y2 and four
predetermined variables, X1, X2, X3, and Y3 (X4 and X5 show up only in Equation 3). As a result,
the identification properties of only two equations need to be determined, and both of them have
one fewer slope coefficient (3) than there are predetermined variables in the system (4), so they
are overidentified.)

For a problem set on this chapter, wed recommend exercises 14-2, 14-7, 14-8, 14-9, and 14-13.
14-6. Appendix: Errors in the Variables

The key here is for students to understand that only errors in the measurement of independent
variables cause bias.
Chapter Fifteen: Forecasting

15-1. What is Forecasting?
Most students will know what forecasting is by the time they get to this section, but the examples
are nonetheless worth mentioning in class because we return to them later in the chapter and
because the data sets for all of them are available in the text.
15-2. More Complex Forecasting Problems

At some point the students should be told that forecasts often are made by an individual with a
superb feel for a market (or the economy as a whole) and then the econometric models are
adjusted to come out with the answers that the guru expects. Whether in simulation analysis or
forecasting, the use of such add factors and mull factors will catch most students by surprise
and will open their eyes to the fact that econometric forecasting is not accurate enough by itself to
place a lot of faith in.
15-3. ARIMA Models

Weve shortened this coverage to do little more than introduce the topic to the reader.

For a problem set on this chapter, wed recommend exercises 15-2, 15-4a, and 15-8.
Lecture Notes 43
Chapter Sixteen: Statistical Principles

This chapter was written by Gary Smith of Pomona College. Gary is the author of our favorite statistics
text, Statistical Reasoning (McGraw-Hill Publishers, 1998, ISBN 0-07-059276-4), and we owe a real debt
of gratitude to Gary for taking on this task.
Weve tried to set up the chapter so that it can be assigned in its entirety (in courses with a minimal stats
requirement) or assigned on a section-specific basise (for readers who need brushing up on particular
topics).
16-1. Probability Distributions

Placing heavy emphasis on the section on standardized variables, Section 16.1.4, will pay off later
in enhanced student understanding of the principle behind the normal distribution and the t-test.
16-2. Sampling
This short section is so well-written and so universally applicable that we anticipate assigning it to
all of our students when we cover data collection in Chapter 11.
16-3. Estimation
16-4. Hypothesis Tests

We purposely created some overlap between Chapter 5 and these two sections so that students
who are having trouble with Chapter 5 can get a bit more background and detail.

For a problem set on this chapter, wed recommend exercises 16-5, 16-7, 16-9, 16-11, 16-13,
and 16-15.
Sample Examinations
Each of the following sample exams follows the same format. Question one consists of four identifications
from the chapters and one short application of the material. Question two asks the student did you read
the book? about one of the most important topics covered. Question three asks the student to apply the
materials of the chapters in a manner that parallels work already done, and question four attempts to see if
the student can apply skills learned in the chapters to a regression result that he or she has not seen before.
Questions from different sample exams can thus be combined to form tests that cover more chapters or
that are harder or easier than these are. Answers to the questions (typically just textbook page references)
follow each sample exam.
By the way, if your students seem especially nervous before a particular exam, you might relax them a bit
with the following actual quote from a mayonnaise jar: Keep cool but dont freeze.
Sample Exam for Chapters 13

1. Briefly identify the following in words or equations as appropriate:
(a) Degrees of freedom
(b) Estimated regression equation
(c) The Six Steps in Applied Regression Analysis
(d) Ordinary Least Squares
(e) The meaning of 1 in:
Yi = 0 + 1 X1i + 2 X2i + i
2. Two of the most important econometric concepts to date have been the stochastic error term and
the residual. Carefully distinguish between these two concepts, being sure to:
(a) define both terms
(b) state how they are similar,
(c) state how they are different,
(d) give an example of an equation that contains a stochastic error term, and
(e) give an example of an equation that contains a residual.
3. A source of constant discussion among applied econometricians is the degree to which measures
of the overall fit of an estimated equation also measure the quality of that regression. To date, we
have introduced something like four different measures of overall fit, but the two most used are
R 2 and R 2 .
(a) Carefully distinguish between R 2 and R 2 .
(b) Of the two, which do you recommend typically using? Why?
(c) What drawbacks are there to the use of the measure you chose (as your answer to part (b)
above) as the primary determinant of the overall quality of a regression?
4. Consider the following estimated equation:

Ct = 18.5 0.07Pt + 0.93YDt 0.74D1t 1.3D2t 1.3D3t
where: Ct = Per-capita pounds of pork consumed in the U.S. in quarter t.
Pt = The price of a hundred pounds of pork (in dollars) in quarter t.
YDt = Per capita disposable income (in dollars) in quarter t.
D1t = Dummy equal to 1 in the first quarter (Jan.Mar.) of the year and 0 otherwise.
D2t = Dummy equal to 1 in the second quarter of the year and 0 otherwise.
D3t = Dummy equal to 1 in the third quarter of the year and 0 otherwise.
(a) What is the meaning of the estimated coefficient of YD?
(b) Specify expected signs for each of the coefficients. Explain your reasoning.
(c) Suppose we changed the definition of D3t so that it was equal to 1 in the fourth quarter and
zero otherwise and re-estimated the equation with all the other variables unchanged. Which of
the estimated coefficients would change? Would your answer to part (b) above change?
Explain your answer.
Answers:
1. (a) See page 54.
(b) See pages 1618.
(c) See page 67.
(d) See Section 2.1.
(e) See page 41.
2. See Sections 1.2.3 and 1.3.
3. See Sections 2.4.1 and 2.4.3.
4. Wed expect pork consumption to be the highest in the fourth quarter due to holidays, so the
expected signs of the dummy coefficients are negative. For part (c), the coefficients of P and YD
shouldnt change, but the others should, because the omitted condition is now the third quarter and
not the fourth quarter. Thus the expected signs of the coefficients of D1 and D2 are no longer
negative. (Some of the best students will note that the estimate of the coefficient of D2 will be
almost exactly zero.)

Note: To do this test, students will need to consult a t-table like Statistical Table B-1 in the text. Since the
table on page 611 is accompanied by a description of how to use it, we have repeated Table B-1 without
the description on the inside back cover of this text. This will allow students to use a t-table (as long as no
notes are written on it) without seeing the additional material if you so desire.

(a) P-value
(b) Sampling distribution of
(c) Type I Error
(d) Level of significance
(e) Given = 4.0 and SE ( ) = 2.0, test the null hypothesis that = 0 (versus the alternative
hypothesis that it does not equal zero) with 13 degrees of freedom at the 1 percent level of
significance.
Sample Examinations 47
2. In Chapter Four, we discussed seven explicit assumptions about the properties of a regression
equation. These assumptions (or, more accurately, the first six of them) are usually referred to as
the Classical Assumptions.
(a) Carefully and accurately state the seven assumptions.
(b) State what each assumption actually means in real-world terms (be extremely brief).
3. One of the shortest but most important sections in the book is that on the Gauss-Markov Theorem.
(a) What is the Gauss-Markov Theorem?
(b) Carefully explain what the precise properties specified by the Gauss-Markov theorem mean
and why they are desireable for an equation.
(c) What could cause the Gauss-Markov Theorem to no longer hold? What should we do in such
a situation?
4. Consider the following regression equation for the United States (standard errors in parentheses):
Pt = 4.00 0.010PRPt + 0.030PRBt + 0.20YD t
(0.005) (0.020) (0.04)
R = .98
2
n = 29
where: Pt = per capita pounds of pork consumed in time period t
PRPt = the price of pork in time period t
PRBt = the price of beef in time period t
YDt = per capita disposable income in time period t
(a) Hypothesize signs and specify the appropriate null and alternative hypotheses for the
coefficients of each of these variables.
(b) State your decision rules and then test your hypotheses on the above results using the t-test at
a 5% level of significance.
(c) If you could add one variable to the regression, what variable would you add? Why?
Answers:
1. (a) See pages 128130.
(b) See pages 9798.
(c) See Section 5.1.2.
(d) See Section 5.2.3.
(e) t = 2.0, and the critical two-sided, 1 percent, 13 degree of freedom critical t-value is 3.012, so
we can not reject the null hypothesis.
2. See Section 4.1.
3. See Section 4.3.

4. Coefficient PRP PRB YD

Hypothesized sign + +
Calculated t-score 2.0 +1.5 +5.0
the critical t-value
for a 5 percent
one-sided test with
25 d.f. is 1.708, so reject do not reject
reject
For part (c), the two most important criteria are whether or not the suggested variable is a time
series variable for the U.S. and whether or not that variable can be measured. Tastes, for
example, are important but hard to measure.
Sample Exam for Chapter 67

(a) The major consequence of including an irrelevant variable in a regression equation.
(b) Four valid criteria for determining whether a given variable belongs in an equation.
(c) The problem with sequential specification searches.
(d) The elasticity of Y with respect to X in:
ln Y = 0 + 1 ln X +
(e) The sign of the bias on the coefficient of age caused by omitting experience in an equation
explaining the salaries of various workers.
2. Most of Chapter Six is concerned with specification bias.

(a) What exactly is specification bias?
(b) What is known to cause specification bias?
(c) Are unbiased estimates always better than biased estimates? Why or why not?
(d) Whats the best way to avoid specification bias?
3. There are at least two different possible approaches to the problem of building a model of the
costs of production of electric power.
I: Model I hypothesizes that per unit costs (C) as a function of the number of kilowatt-hours
produced (Q) continually and smoothly falls as production is increased, but it falls at a
decreasing rate.
II: Model II hypothesizes that per unit costs (C) decrease fairly steadily as production (Q)
increases, but costs decrease at a much faster rate for hydroelectric plants than for other kinds
of facilities.
Given this information,

(a) What functional form would you recommend for estimating Model I? Be sure to write out a
specific equation.
(b) What functional form would you recommend for estimating Model II? Be sure to write out a
specific equation.
(c) Would R 2 be a reasonable way to compare the overall fits of the two equations? Why or why
not?
Note: In the following question, you may want to change the absolute size of the coefficients,
depending on the size of your school, but remember to change the size of the standard errors
proportionally.
4. One your way to the cashiers office (to pay yet another dorm damage bill), you overhear U. R.
Accepted (the Dean of Admissions) having a violent argument with I. M. Smart (the Director of
the Computer Center) about an equation that Smart built to understand the number of applications
that the school received from high school seniors. In need of an outside opinion, they turn to you
to help them evaluate the following regression results (standard errors in parentheses):
N t = 150 + 180A t + 1.50 ln Tt + 30.0Pt

(90) (1.50) (60.0)
R = .50 N = 22 (annual)
2
where: Nt = the number of high school seniors who apply for admission in year t.
At = the number of people on the admission staff who visit high schools full time
spreading information about the school in year t.
Tt = dollars of tuition in year t.
Pt = the percent of the faculty in year t that had PhDs in year t.
How would you respond if they asked you to:

(a) Discuss the expected signs of the coefficients.
(b) Compare these expectations with the estimated coefficients by using the appropriate tests.
(c) Evaluate the possible econometric problems that could have caused any observed differences
between the estimated coefficients and what you expected.
(d) Determine whether the semi-log function for T makes theoretical sense.
(e) Make any suggestions you feel are appropriate for another run of the equation.
Answers:
1. (a) See Section 6.2.1.
(b) See Section 6.2.3.
(d) Constant = 1.
(e) positive, since the expected sign of the coefficient of experience is positive and since the
simple correlation coefficient between age and experience is positive. See Section 6.1.3.
2. See Section 6.1. Note in particular that biased estimates can at times be closer to the true than
unbiased ones and that the solution to omitted variable bias is not to simply include every
justifiable variable.
3. (a) A number of forms are possible, but a reciprocal form would be perhaps the most appropriate:
Ct = 0 + 1/Q t + t
(b) Such an hypothesis calls for the use of a slope dummy defined (for instance) as D = 1 if the
plant is hydroelectric and 0 otherwise. The resulting equation would be:
Ct = 0 + 1Q t + 2D t + 3Dt Q t + t
(c) R 2 is perfectly appropriate for comparing the overall fits of the two equations because the
number of independent variables changes but the functional form of the dependent variable
does not.
4. (a)/(b) Coefficient 1 2 3
Expected Sign + +
t-score +2.0 +1.0 +0.5
Decision reject do not do not
reject reject
(given the 5 percent level and a resulting critical t-score of 1.734.)

(c) 2 has an insignificant unexpected sign, and 3 is not significant, so an irrelevant variable
and an omitted variable (s) are both possible.
(d) Many would expect that the impact of an extra dollar of tuition (on applications) decreases as
tuition increases. If this is so, the semi-log functional form makes sense; if not, then we
should use the linear form as a default.
(e) Percent PhDs appears to be intended as a proxy for the way in which the quality of the school
has changed over the years. If we could find a better proxy for quality, then it would make
sense to substitute it for X3, and that would be a possible answer. If the unexpected sign for
lnT is being caused by an omitted variable, that omitted variable must be causing positive
bias. Thus for a suggested omitted variable to be a good suggestion, it would have to be likely
to have caused positive bias. Therefore, it must be either positively correlated with tuition and
have a positive expected coefficient (like the number of high-school seniors in the country or
in the region) or else have a negative correlation with tuition and have a negative expected
coefficient (like the number of inadequate facilities, etc.).

Note: This exam uses heteroskedasticiy as the topic of the major question, question number two, so the
IDs in question number one avoid heteroskedasticity as a topic. Question number two can easily be
reworded to cover serial correlation or multicollinearity, but then question number one should also be
changed to cover heteroskedasticity.

(a) Impure serial correlation
(b) Dominant variable
(c) Variance Inflation Factor
(d) Generalized Least Squares
(e) Given a calculated Durbin-Watson d statistic of 2.58, a dL of 1.21 and a dU of 1.55, what
would you conclude?
2. Carefully outline (be brief!) a description of the problem typically referred to as pure
heteroskedasticity.
(a) What is it?
(b) What are its consequences?
(c) How do you diagnose it?
(d) What do you do to get rid of it?
3. A model of the number of cars sold in the United States from 1980 through 2004 produced the
following results (standard errors in parentheses):
Ct = 3738 48.0Pt + 10.0Yt + 6.0At 360.0Rt
(12.0) (2.0) (2.0) (120.0)
R = .85
2
DW = 1.86 N = 25 (annual)
where: Ct = thousands of cars sold in year t
Pt = price index for domestic cars in year t
Yt = disposable income (billions of dollars) in year t
At = billions of dollars of auto industry advertising expenditures in year t
Rt = the interest rate in year t
(a) Hypothesize the expected signs of the coefficients and test the appropriate null hypotheses at
the 1 percent level.
(b) What econometric problems appear to be present in this equation? Why?
(c) Suppose you were now told that the simple correlation coefficients between P, A, and Y were
all between 0.88 and 0.94 and that a Park test with Y as Z produced a t-score of 0.50. Would
your answer to part (b) above change? Why or why not? How would it change?
(d) What suggestions would you have for another run of this regression?
4. In a study of the long-run and short-run demand for money, Chow estimated the following
demand equation (standard errors in parentheses) for the U.S. from 1947:1 through 1965:4:
Mt = 0.14 + 1.05Y*t 0.01Yt 0.75Rt
(0.15) (0.10) (0.05)
R = .996
2
DW = 0.88 N = 76 (quarterly)
where: Mt = the log of the money stock in quarter t

Y*t = the log of permanent income (a moving average of previous quarters current
income) in quarter t
Yt = the log of current income in quarter t
rT = the log of the rate of interest in quarter t
(a) Hypothesize signs and test the appropriate null hypotheses at the 5 percent level of
significance.
(b) What econometric problems seem likely to be in this equation?
(c) In particular, are there are any problems related to the coefficient of Y? If so, are these
problems more likely to have been caused by multicollinearity, serial correlation, or
heteroskedasticity?
(d) What suggestions would you have for another estimation of this equation? Why?
Answers:
(e) The answer depends on whether you encourage your students to use one-sided or two-sided
tests. For a one-sided test, the correct answer is that we can reject the null hypothesis of
positive serial correlation. For a two-sided test, the correct answer is that the test is
inconclusive.
2. The best answer would be only slightly more detailed than the summary on pages 376377.
3. (a) Coefficient P Y A R
Expected Sign + +
Calculated t-score 4.0 5.0 3.0 3.0
Decision reject reject reject reject
(given a 1 percent critical t-score of 2.528) In addition, the DW = 1.86 is higher than the dU of
1.77, so there is no evidence of positive serial correlation.
(b) There may be theoretically sound variables that have been omitted, for instance some measure
of competition from foreign-made cars, but the results themselves give no indication of any
problems.
(c) These results are clear indications of multicollinearity but not of heteroskedasticity.
(d) Unless a theoretically sound variable can be added measuring competition, the regression
need not be changed at all.
4. (a) Coefficient Y* Y R
Expected Sign + +
Calculated t-score 7.0 0.1 15.0
Decision reject do not reject
reject
(given a 5 percent critical t-score of about 1.67) In addition, the DW = 0.88 is less than the dL
of 1.54, so we can reject the null hypothesis of no positive serial correlation.
(b) The poor significance of Y could have been caused by multicollinearity, an omitted
variable, or by an irrelevant variable (Y), and in addition, as mentioned above, we have serial
correlation.
(c) Of the three, only multicollinearity between Y* and Y could have caused the problem, since
serial correlation and heteroskedasticty would cause the t-scores to appear higher than they
actually should be.
(d) This is not an easy question, but we would not suggest dropping Y, nor would we consider
GLS right now. It turns out that making the equation a distributed lag by adding Mt1 to the
right-hand side switches the insignificance to Y* (and raises the DW, naturally) giving
another indication of multicollinearity. The best action right now would seem to be to review
the theory in search of an omitted variable and failing to find one, do nothing. This problem
was adapted from a source, Edward Tower, Econometrics Exams, Puzzles and Problems
(Durham, N.C.: Eno River Press, 1985) p. 167, that is just full of questions of a similar (or of
a more theoretical) nature.

(a) Problems with ad hoc distributed lags
(b) Unconditional forecasting
(c) Moving-average process
(d) How to test for serial correlation in a dynamic model.
(e) The identification properties (use the order condition only) of each of the equations in the
following system:
i. Y1t = f(Y2t, X1t, X2t)
ii Y2t = f(Y3t, Y1t)
iii. Y3t = f(X3t, X4t)
2. Virtually all of Chapter 14 is spent discussing the violation of the assumption that the error term is
independent of the explanatory variables.
(a) Under what circumstances might that assumption be violated?
(b) What would the violation of that assumption be likely to cause?
(c) What general technique is used to rid the equation of this problem? Specifically, how does it
work?
3. Suppose youve been hired by your schools admissions department to help them decide whether
to change admissions procedures. You are given the files of all of the students in the last
graduating class (including those students who didnt graduate) and told to build a model to
explain why some admitted students graduate and others dont.
(a) Specify the functional form you would use in building such a model and carefully explain
why that form is appropriate.
(b) Specify the independent variables you would include in the equation and briefly explain how
they apply to the dependent variable in question.
(c) Carefully explain the meaning of the coefficient of your first independent variable.
4. You have been hired to forecast GDP (Y) for the Caribbean island of Tabasco. Tabasco has
domestic food (F) and shelter (S) industries, a tourist (T) industry, and an export (X) industry. All
tourists come from the United States, while the exports are split between Mexico and the U.S.
Investment is virtually zero, and government expenditures (G) can be considered to be
exogenously determined. Imports (I) are a function of GDP. Thus the structural equations for a
simultaneous model of the Tabascan economy would look something like:
Y=F+S+T+X+GI
F = fF (Y, ?)
S = fS (Y, ?)
T = fT (USGNP, ?)
X = fX (USGNP, MEXICOGNP, ?)
I = fI (Y, ?)
G = exogenous
(a) Develop a theory for Tabascos economy. Then choose which predetermined variables you
would like to add to the simultaneous system and specify to which of the five stochastic
structural equations (see question marks) you would like to add them. Explain your reasoning.
(b) Comment on the identification properties of each of the five stochastic equations in the system
you outlined in your answer to part a above.
(c) How should the coefficients of the system be estimated?

(d) What technique would you use to forecast Tabascos GNP? Why?
Answers:
(c) See Section 15.3.
(e) There are three predetermined variables in the system (X1, X2, and Y3). This means that i is
exactly identified because there are also three slope coefficients to be estimated and that ii is
overidentified, since there are only two slope coefficients to be estimated.
2. See Sections 14.2 and 14.3. This answer will also depend on whether or not Chapter 12 and/or
Section 14.6 have been assigned.
3. The appropriate functional form is the logit because of the problems with the linear probability
model outlined in Section 13.1.2. The key to choosing independent variables is the type of
variable suggested; some students will misunderstand the disaggregate nature of the variables
required by such a study and will suggest variables that are constant for all observations in the
dataset. Each coefficient tells the impact of a one-unit change in the independent variable in
question (holding constant all the other independent variables in the equation) on the log of the
odds that the person graduated.
4. OK, OK, we know this will be hard to grade, since each answer will be different depending on the
exact variables and equations added, but this question tends to work well. The student will be
forced to apply the identification, estimation, and forecasting techniques to a system of their own
choosing in much the same way they will have to in their work later on.
The key to the questions on estimation and forecasting have to do with the size and importance of
the feedback loops (as compared to exogenous factors in determining GDP). In this case, there is a
good chance that the most accurate forecast of Tabascos GDP would be based on a simplified
reduced-form equation that included only USGNP and MEXICOGNP as explanatory variables.
An Additional Interactive Regression Learning Exercise
In this final section, we present a demand for water interactive learning regression exercise. This
interactive exercise is intended to be used in class to give students experience with specification issues like
omitted variables, irrelevant variables, multicollinearity, and, possibly, serial correlation. Thus the best
time to use the exercise is sometime between Chapters Eight and Ten. A hoped-for byproduct is to make
the students more comfortable with interactive exercises in general.
One way to use this exercise in class is to introduce the topic, define all the available variables, encourage
class discussion of expected signs and significance, and then have the class pick their guess at a best
specification. After you put the estimated result on the board (or hand out copies of it), the class should
then test hypotheses, discuss probable econometric difficulties, and choose whether to change the
specification. While only 26 regression runs are provided here, instructors who wish to have more
available are encouraged to use the dataset and estimate them. Indeed, one way of handling this (or any
other) interactive exercise is to bring a PC to class and estimate the regressions after theyve been
specified.
Obviously, its vital to give feedback along the way, particularly since otherwise some students might get
the idea that we can estimate as many runs as we want. Even given this, we try hard to make the students
choose a course of action before we comment on it.
Building A Model of the Demand for Water in Los Angeles
This model is an extension of the short example presented in Section 2.5. The goal is to build a model of
the demand for water in Los Angeles County as a function of population, rainfall and other variables. The
dependent variable is:
Wt = Consumption of water in Los Angeles county in year t (hundreds of millions of gallons).
A few comments about the data are in order. Population data for Los Angeles county were unavailable for
the entire sample when the data were collected, so population in Glendale was used as a proxy. Second,
conservation efforts took place only during the last two years in the sample, so if conservation efforts have
an effect but with a lag, wed be forced to use a one-time dummy. Thus there are reasons to be
concerned that population and conservation are not accurately measured in the sample. The means,
standard deviations, and simple correlation coefficients of the variables, plus the raw data themselves, are
contained in tables on the next two pages.
The available independent variables are:

POPt = Population in Glendale, California in year t (thousands of people).
Tt = Average annual temperature at the Los Angeles civic center in year t (degrees F).
Rt = Inches of rainfall at the Los Angeles civic center in year t.
PRt = Average price of a gallon of water in Los Angeles in year t (dollars).
COt = A dummy variable to measure the conservation efforts undertaken in the last two years in the
sample, when there was a severe drought in Northern California (the source of much of L.A.s
water) even though no such drought took place in Southern California.
Yt = Total personal income for Los Angeles county in year t (in billions of dollars).
Means and Variances for the Demand for Water Data
Variable Mean Standard Deviation

W 71.01 7.97
POP 126.83 9.78
T 65.15 1.24
R 13.75 6.20
PR 0.33 0.13
CO 0.08 0.27
Y 26.91 13.55
Simple Correlation Coefficients for the Data

Y, PR 0.954 CO, PR 0.764
Y, T 0.108 CO, T 0.096
Y, CO 0.639 CO, R 0.036
Y, POP 0.828 CO, POP 0.294
Y, R 0.009 T, POP 0.158
PR, POP 0.630 T, PR 0.079
PR, R 0.024 R, T 0.063
R, POP 0.124
An Additional Interactive Regression Learning Exercise 57
Data for the Demand for Water Interactive Exercise
YEAR W POP T R PR Y
1 53.38500 108.700 63.6 26.21 0.23 9.5345
2 60.31100 110.780 64.6 9.46 0.22 10.7830
3 60.06400 113.340 64.9 11.99 0.22 11.9910
4 58.93900 114.790 63.7 11.94 0.23 13.0190
5 62.33000 115.780 64.8 16.00 0.22 14.2880
6 62.36100 117.110 65.7 9.54 0.23 15.4280
7 65.40900 118.260 67.1 21.13 0.25 15.8760
8 71.91400 118.380 67.3 5.58 0.25 17.3270
9 70.41700 119.440 65.7 8.18 0.26 18.0020
10 73.54900 121.880 65.3 4.83 0.26 18.8880
11 69.09300 123.120 62.3 18.79 0.27 20.3420
12 67.47900 126.770 65.1 8.38 0.26 21.5810
13 73.27700 129.830 63.2 7.93 0.26 22.9790
14 68.94700 132.440 64.3 13.69 0.26 24.2280
15 73.43700 135.790 66.0 20.44 0.26 26.0990
16 72.34000 137.050 66.5 22.00 0.28 27.9240
17 76.04400 137.270 66.2 16.58 0.30 29.9080
18 75.15500 137.920 65.4 27.47 0.34 32.4070
19 81.66400 132.660 66.1 7.77 0.37 34.0980
20 80.54400 132.750 65.1 12.32 0.39 34.3140
21 81.41800 133.110 66.5 6.54 0.43 37.1160
22 76.70700 132.700 65.0 17.45 0.44 39.1910
23 80.61500 136.510 65.3 16.69 0.47 43.7840
24 80.06200 138.010 63.0 10.70 0.51 47.8220
25 81.46700 139.260 65.2 11.01 0.60 53.3340
26 69.31500 133.920 65.9 14.97 0.75 59.2670
How to Find the Desired Regression Run
Note that the possible variables include a theoretically strong variable (R), a probable irrevelevant variable
(T), a possible irrelevant variable (CO), two important but basically redundant variables (POP and Y), and
a potentially weak (at least theoretically) variable (PR) that is also multicollinear with POP and Y. One of
the nicest attributes of this exercise is that when you make a specification change, most statistics (from
SEs to R 2 s) act the way they should; believe us, finding such examples is not easy. With respect to
serial correlation, all classes so far have ended up choosing as their final specification regression runs one
of #26, #25, #24, #11, or #8, of which only #26 has a DW below the critical dL value (#11 is on the
borderline). A GLS version of equation #26 appears as equation #27.
Once the class has chosen its best specification, refer to the following key to find the number of the
regression run chosen. This interactive exercise has been used extensively in class, and the following
26 runs contain all the choices that have been chosen during that time, but there are a number of other
specifications that could also be chosen. If the class picks one of these, shift to the closest alternative.
Run # # of Xs KEY to Included Independent Variables

1 6 All
2 5 All but POP
3 5 All but T
4 5 All but R
5 5 All but PR
6 5 All but CO
7 5 All but Y
8 4 All but POP and T
9 4 All but Y and T
10 4 All but R and T
11 4 All but PR and T
12 4 All but CO and T
13 4 All but R and PR
14 4 All but R and CO
15 4 All but R and Y
16 4 All but PR and CO
17 4 All but PR and Y
18 4 All but CO and Y
19 4 All but POP and R
20 4 All but POP and PR
21 4 All but POP and CO
22 3 POP, R, and CO are included
23 3 POP, R, and PR are included
24 3 R, CO, and Y are included
25 3 R, PR, and Y are included
26 2 POP and R are included
Regression Run # 1
# INT POP T R PR CO Y DW R2
7.6 0.07 0.79 0.33 63.4 7.20 1.08 1.91 .84
SE 0.46 0.52 0.11 68.9 4.49 0.88
t 0.2 1.5 3.0 0.9 1.6 1.2
To drop POP see #2; to drop PR see #5; otherwise see key.
Regression Run #2
15.0 0.80 0.32 73.5 6.92 1.21 1.89 .85
SE 0.51 0.10 20.9 3.98 0.17
t 1.6 3.2 3.5 1.7 6.99
To add POP see #1; to drop T see #8; otherwise see key.
Regression Run #3
46.0 0.18 0.35 52.6 6.90 0.92 1.64 .83
SE 0.47 0.11 70.6 4.62 0.90
t 0.4 3.1 0.7 1.5 1.0
To add T see #1; to drop POP see #8; otherwise see key.
Regression Run #4
54.0 0.45 0.99 132.0 5.44 2.01 1.27 .78
SE 0.50 0.61 76.4 5.24 0.97
t 0.9 1.6 1.7 1.0 2.07
To add R see #1; to drop POP see #19; otherwise see key.
Regression Run #5
39.0 0.47 0.74 0.36 9.63 0.28 2.10 .85
SE 0.14 0.52 0.10 3.62 0.13
t 3.3 1.4 3.5 2.7 2.2
To add PR see #1; to drop T see #11; otherwise see key.
Regression Run #6
51.0 0.24 0.75 0.31 128.0 1.78 1.38 .83
SE 0.43 0.54 0.11 57.9 0.80
t 0.5 1.4 2.7 2.2 2.2
To add CO see #1; to drop POP see #21; otherwise see key.
Regression Run #7
55.0 0.62 0.71 0.38 20.0 9.92 2.10 .84
SE 0.09 0.53 0.10 10.0 3.95
t 6.6 1.3 3.6 2.0 2.5
To add Y see #1; to drop T see #9; otherwise see key.
Regression Run #8
68.0 0.33 78.4 6.14 1.26 1.58 .84
SE 0.10 21.5 4.09 0.18
t 3.2 3.7 1.5 7.1
To add POP see #3; to drop CO see #25; otherwise see key.
Regression Run #9
11.0 0.64 0.39 18.4 9.27 1.82 .83
SE 0.09 0.11 10.1 4.00
t 6.9 3.7 1.8 2.3
To add Y see #3; to drop PR see #22; otherwise see key.
Regression Run #10

107.0 0.36 124.0 4.90 1.88 1.19 .77
SE 0.51 79.0 5.43 1.00
t 0.7 1.6 0.9 1.9
To add R see #3; to add T see #4; otherwise see key.
Regression Run #11

5.5 0.51 0.38 8.95 0.25 1.81 .84
SE 0.14 0.11 3.67 0.13
t 3.6 3.6 2.4 2.0
To add PR see #3; to drop Y see #22; otherwise see key.
Regression Run #12

86.0 0.12 0.33 115.0 1.59 1.16 .82
SE 0.43 0.12 58.0 0.80
t 0.3 2.8 2.0 2.0
To add CO see #3; to drop POP see #25; otherwise see key.
Regression Run #13

43.0 0.36 0.93 10.7 0.35 1.62 .76
SE 0.17 0.64 4.5 0.15
t 2.1 1.4 2.4 2.3
To add R see #5; to add PR see #4; otherwise see key.
Regression Run #14

84.0 0.66 0.95 178. 2.50 1.03 .78
SE 0.46 0.61 62.4 0.85
t 1.4 1.6 2.9 2.9
To add R see #6; to add CO see #4; otherwise see key.
Regression Run #15

65.0 0.55 0.89 24.5 10.7 1.62 .75
SE 0.12 0.66 12.4 4.9
t 4.8 1.4 2.0 2.2
To add R see #7; to add Y see #4; otherwise see key.
Regression Run #16

47.0 0.68 0.56 0.39 0.03 1.19 .80
SE 0.13 0.59 0.12 0.10
t 5.1 1.0 3.3 0.3
To add PR see #6; to add CO see #5; otherwise see key.
Regression Run #17

56.0 0.75 0.58 0.40 3.73 1.37 .82
SE 0.07 0.56 0.11 2.63
t 10.0 1.0 3.6 1.4
Regression Run #18

50.0 0.71 0.55 0.39 0.43 1.18 .80
SE 0.10 0.59 0.12 6.96
t 7.4 0.9 3.3 0.1
To add CO see #7; to drop T see #23; otherwise see key.
Regression Run #19

2.1 0.93 66.3 7.35 1.15 1.38 .78
SE 0.61 25.1 4.79 0.21
t 1.5 2.6 1.5 5.5
To add R see #2; to add POP see #4; otherwise see key.
Regression Run #20

10.0 1.06 0.29 16.3 0.64 2.03 .77
SE 0.62 0.12 3.6 0.07
t 1.7 2.3 4.5 8.7
Regression Run #21

26.0 0.69 0.33 98.0 1.35 1.33 .84
SE 0.53 0.10 16.2 0.16
t 1.3 3.1 6.0 8.5
To add CO see #2; to drop T see #25; otherwise see key.
Regression Run #22

19.0 0.76 0.41 3.60 1.23 .82
SE 0.07 0.11 2.64
t 10.3 3.7 1.4
To add PR see #9; to drop CO see #26; otherwise see key.
Regression Run #23

15.0 0.73 0.40 0.18 1.05 .80
SE 0.10 0.12 6.93
t 7.6 3.4 0.02
To add CO see #9; to drop PR see #26; otherwise see key.
Regression Run #24

59.0 0.30 16.1 0.65 1.80 .75
SE 0.13 3.8 0.08
t 2.3 4.2 8.5
To add PR see #8; to add POP see #11; otherwise see key.
Regression Run #25

71.0 0.34 99.9 1.38 1.14 .83
SE 0.11 16.4 0.16
t 3.2 6.1 8.5
To add CO see #8; to add POP see #12; otherwise see key.
Regression Run #26

16.0 0.73 0.40 1.05 .81
SE 0.07 0.11
t 10.1 3.6
To add PR see #23; to add CO see #22; otherwise see key.
Regression Run #27: A GLS version [AR(1) method] of run #26.
# INT POP T R PR CO Y n
Rho R2
11.0 0.68 0.30 0.546 .81
SE 0.17 0.10
t 4.1 2.9

A.H. Studenmund-Using Econometrics Solution PDF

Uploaded by

Copyright:

Available Formats

A.H. Studenmund-Using Econometrics Solution PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Read this document in other languages

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A.H. Studenmund-Using Econometrics Solution PDF

Uploaded by

Copyright:

Available Formats

Answers to Odd-Numbered Exercises

Chapter One: An Overview of Regression Analysis

Chapter Two: Ordinary Least Squares

The results do indeed match. R = .744, and R 2 = .733.

Chapter Three: Learning to Use Regression Analysis

3-5. (a) Positive.

3-9. (a) Negative, positive, none

Chapter Four: The Classical Model

4-5. (a) Classical Assumption II.

Chapter Five: Hypothesis Testing

For P : Reject H:0 0 if 4.88 > tC and 4.88 is positive.

5-11. (a) For the t-tests:

Chapter Six: Specification: Choosing the Independent Variables

6-11. (a) Coefficient: W T C L

Chapter Seven: Specification: Choosing a Functional Form

Industry Labor Capital

7-13. (a) Coefficient: LQ A V

(b) Q constant, A and V non-constant.

7-15. (a) Coefficient: S G D SD GD

Chapter Eight: Multicollinearity

8-5. (a) Using the mean method:

(c) Using the mean method:

8-7. (a) No; no explanatory variable is a perfect function of another.

8-9. (a) Coefficient: PC PQ Y C N

8-11. (a) Coefficient: M B A S

8-13. (a) Coefficient: EMP UNITS LANG EXP

Chapter Nine: Serial Correlation

9-15. (a) For the first equation (standard errors in parentheses):

For the second equation:

Chapter Ten: Heteroskedasticity

10-5. m i = 1273.2 + 0.72I

where: CO = average consumption

tC = 3.499 at the 1% level (two-tailed) so we cannot reject the null hypothesis of

(where ui is a homoskedastic error term)

10-11. (a) Coefficient: W U lnP

10-13. (a) Coefficient: P I Q A S T

10-15. (a) To test for serial correlation, first run:

St = 0.73 + 0.22It + 0.46ln(1 + Vt )

N = 58 (monthly) R 2 .556 DW = 1.54

Chapter Twelve: Time-Series Models

12-5. n = 243 + 5.2AD + 1.9AD + 3.1AD + 1.0AD + 3.3AD

cannot reject the null hypothesis of no second-order serial correlation.

Chapter Thirteen: Dummy Dependent Variable Techniques

(0.63) (0.02) (0.18)

13-11. (a) Three

Chapter Fourteen: Simultaneous Equations

14-7. (a) A: Predetermined = 2 < 3 = # of slope coefficients, so underidentified.

14-9. (a) It = 270 + 0.19Yt 8.26rt1

Chapter Fifteen: Forecasting

15-9. (a) e99 = Y99 Y99 = 27 27.5 = 0.5

Chapter Sixteen: Statistical Principles

16-7. Z = (70 66)/2.5 = 1.60. P[Z > 1.60] = 0.0548.

16-15. All three statements are wrong:

Chapter One: An Overview of Regression Analysis

1-2. What is Regression Analysis?

1-3. The Estimated Regression Equation

1-4. A Simple Example of Regression Analysis

1-5. Using Regression to Explain Housing Prices