CHAPTER

12
SELECTED TOPICS IN SINGLE EQUATION
REGRESSION MODELS

QUESTIONS

12.1. (a), (b), and (c) A dynamic model is a regression model that explicitly
allows for intertemporal (i.e., over time) analysis of economic phenomena
such as the consumption function, the demand for money function, etc. A
dynamic model containing the current values of the explanatory variables as
well as their past values is called a distributed lag model, whereas a model
that includes among the explanatory variables the lagged value(s) of the
dependent variable is called an autoregressive model.
12.2. Psychological reasons, technological reasons, and institutional reasons can
lead to a lagged response of the dependent variable to explanatory variables.
12.3. Such a strategy may be ad hoc in that it has no theoretical underpinning.
Since the lagged terms are likely to be correlated, every time you add a new
term, the model has to be re-estimated. Not only that, but the values of the
coefficients already estimated change every time you reestimate the model.
12.4. True. Because of collinearity, individual coefficients cannot be estimated
precisely, but their sums or differences can be estimated correctly.
12.5. Simplicity at any cost should not be the goal of the model builder. Since the
LPM has serious problems, the logit and probit models are the preferred
alternatives in regression models that involve dichotomous dependent
variables. These models guarantee that the estimated probabilities will be
non-negative.
12.6. True.
12.7 A spurious regression may result if we regress a nonstationary time series
on one or more nonstationary time series. However, such a regression may be
meaningful if there is a stable, long-term relationship between these time series, even

1
though individually they are non-stationary. In that case we say that the time series
are cointegrated.

PROBLEMS

12.8. (1) PĈE t = -117.36 + 0.9614 PDIt

t = (-7.03) (194.74) r 2 = 0.999

(2) PĈE t = 10.11 + 0.2041 PDIt + 0.8256 PCE t −1

t = (0.45) (2.83) (10.52) R 2 = 1.000

(a) In regression (1), the marginal propensity to consume out of PDI is
about 1 (0.9614).
(b) From regression (2), the short-run MPC is only 0.2041 and the long-run
MPC is (0.204 / 0.1744) = 1.170. Since the coefficient of the lagged PCE in
regression (2) is statistically significant, there appears to be a distributed lag
effect.
12.9. (1) lnP̂CE t = -0.3834 + 1.0346 ln PDIt

t = (-11.73) (257.97) r 2 = 0.999

(2) lnP̂CE t = 0.0516 + 0.3155 ln PDIt + 0.6803 lnPCE t − 1

t = (0.92) (3.78) (8.64) R 2 = 1.000

(a) Since we are using a double-log model, all the slope coefficients are
elasticities.
(b) From regression (2), the short-run elasticity is 0.3155 and the long-run
elasticity is (0.3155 / 0.3197) = 0.9869.
12.10. (a) Current and lagged capacity utilization rate each has a positive effect on
the inflation rate as measured by the GNP deflator. An increasing capacity
utilization, ceteris paribus, signifies demand pressure, which puts upward
pressure on prices. Hence the positive relationship is expected.
(b) The short-run impact is 0.1408 and the long-run impact is found by
summing: (0.1408 + 0.2360) = 0.3768.
(c) Yes. For 15 d.f., the 5% (two-tailed) critical t value is 2.131.

2
 (d) Use the R 2 variant of the F test. In the present example, the F value is
19.9725, which is significant beyond the 1% level; the 1% critical F for 2
and 15 d.f. is 6.36 ( F2,15 = 6.36).

(e) Recent data can be found in the latest Economic Report of the President.
12.11. We can use the logit model. The results are as follows:
 Pˆ 
ln  i  = -4.8341 + 3.0609 ln X i
 1 − Pi 

t = (-10.8876) (11.9046) r 2 = 0.9793

These results show that if the dosage increases by 1%, on the average, the
odds in favor of death increase by about 3.06%.
12.12. Here is an illustrative calculation. For X = 50, Equation (12.36) gives:
 P 
ln  i  = -3.2438 + 0.0792 (50) = 0.7162
 1 − Pi 

Pi
Therefore, = antilog(0.7162) = 2.0466.
1 − Pi

Solving for Pi , we obtain Pi = 0.6718.

12.13. (a) Holding other things constant, if income goes up by a unit, say, a
thousand dollars, the log of the odds in favor of restaurant usage goes up by
about 0.37 units. Likewise, holding other things constant, the log of the
odds in favor of restaurant usage goes down by about -1.1 units if a couple
needs a baby sitter. Both these coefficients have the correct signs.
(b) Logit = [-9.456 + 0.3638 (44) – 1.107 (1) ] = 5.4442.
Pi
(c) Since = antilog(5.4442) = 231.4121, we can solve for Pi in order
1 − Pi

to get Pi = 0.9957.

Therefore, Pi = 0.9957 (almost 100%) that a couple with an income of

$44,000 and who needs a baby sitter will eat out.
12.14. (a) The plots will show that the dividend series is generally upward trending
in a rather smooth manner but the profits series, although generally upward

3
trending, shows much more volatility. In both cases, however, the
impression one gets is that the two time series are non-stationary.
(b) The unit root test was applied to the two series, including the constant
and trend terms. The results were as follows:
Dividends
∆Dividends t = 0.0411 + 0.0004t − 0.0180 Dividends t-1
τ = (2.09) (1.61) (−1.42)*
R 2 = 0.0217
Profits
∆ log Pr ofitst = 0.1412 + 0.0010t − 0.0524 Pr ofits t-1
τ = (2.69) (2.81) (−2.60)*
R 2 = 0.0354
Note: ∆ is the first-difference operator and τ is the Dickey-Fuller tau
statistic.
The 1%, 5%, and 10% Augmented Dickey-Fuller critical tau statistic values
are -4.0673, -3.4620, and -3.1570, respectively. If the tau values of the
lagged terms, Dividend t − 1 and Profits t − 1 , are smaller in absolute value than
the critical tau values in absolute terms, we conclude that the series under
consideration has a unit root, i.e., it is not stationary. Therefore, both our
time series are nonstationary.
Note: There is a more advanced version of the unit root Dickey-Fuller (DF)
test called the Augmented Dickey-Fuller test (ADF). This is available as an
option in EViews (it is in the “View” drop-down menu after a variable is
opened). The critical values shown above are from there. The ADF critical
values can be used because, asymptotically, the DF and ADF tests are the
same.
(c) If one nonstationary time series (dividends) is regressed on another
nonstationary time series (profits), the results are likely to be spurious.
However, if the two series are cointegrated, the regression results would not
be spurious. Following the discussion in the text, dividends were regressed
on profits and the residuals from this regression were subjected to the unit

4
 root test, as described in the text, namely, e t was regressed on e t − 1 (without

intercept and trend), where e t − 1 is the lagged residual term, and the

coefficient of the lagged e t − 1 had a tau ( τ ) value which was below the
critical tau values given in (b) . This suggests that the two time series are
not cointegrated. Therefore, the regression of dividends on profits is
spurious.
(d) Using EViews, and applying the Augmented Dickey-Fuller unit root test
(with intercept and trend) to the first differences of the two series, it can be
observed that the first-differenced time series are stationary. The tau
statistic for the appropriate coefficient (lagged value of the first difference of
the two time series) is -3.14 (profits) and -1.42 (dividends). In absolute
terms, neither of these tau values exceed the critical tau values (in absolute
terms) given in (b) previously.
12.15. Assign this as a classroom exercise.
12.16. Assign this as a classroom exercise.
12.17. Since the data given in Table 6.10 were artificial, you may have difficulty in
estimating the (ungrouped) logit model for these data using the method of
maximum-likelihood.
12.18. (a) Using EViews , we obtained the following LPM regression:

Dependent Variable: Y
Sample: 1 32

Variable Coefficient Std. Error t-Statistic Prob.

C -1.498017 0.523889 -2.859419 0.0079
GPA 0.463852 0.161956 2.864054 0.0078
PSI 0.378555 0.139173 2.720035 0.0111
TUCE 0.010495 0.019483 0.538685 0.5944
R-squared 0.415900
Adjusted R-squared 0.353318
S.E. of regression 0.388057

Note: The adjusted R 2 and the standard error of the regression are also
shown here.

5
(b) EViews calculates a logit model in a few short steps. In the Table 16-7
workfile, you can proceed as follows:
• Select “Objects” and then “New Object”;
• Select “Equation” and click “OK”;
• In the Equation Specification Menu , under “Estimation Settings”, select
“BINARY” from the Method drop-down menu;
• Check the “Logit” box under “Binary estimation method”;
• Just type the names of variables in the large Equation Specification box
(do not forget the constant C).
The output of the logit function is as follows:

Dependent Variable: Y
Method: ML – Binary Logit
Sample: 1 32

Variable Coefficient Std. Error z-Statistic Prob.

C -13.02135 4.931317 -2.640541 0.0083
GPA 2.826113 1.262940 2.237726 0.0252
PSI 2.378688 1.064563 2.234426 0.0255
TUCE 0.095158 0.141554 0.672235 0.5014
S.E. of regression 0.384716 McFadden R-squared 0.374038
Obs with Dep=0 21 Total obs 32
Obs with Dep=1 11

Note: The conventional R 2 is not very meaningful in models that have a

qualitative dependent variable. There are similar measures, collectively
known as pseudo R 2 . The McFadden R 2 shown above is one such measure.
Alternatively, one can use the Count R 2 defined in Problem 10.24. When
using the Count R 2 , if the predicted probability is greater than 0.5 we
classify it as 1 and if it is less than 0.5 we classify it as 0. Then the number
of correct predictions is counted and the Count R 2 is computed accordingly.
The McFadden R 2 and the Count R 2 may sometimes differ substantially.
Do keep in mind, however, that goodness-of-fit measures for qualitative
dependent variable models are of secondary importance.
(c) The outputs of LPM and Logit models are not directly comparable for
reasons discussed in the text, although qualitatively the coefficients have the

6
 same signs. In view of the deficiencies of the LPM model, it is better to
use the logit or probit model. For instance, in the LPM 5 out of 32
estimated Y values were negative.
12.19 This is a class exercise.
12.20 (a) Scatterplot is as follows:

1.7

1.6

1.5

1.4

1.3

1.2

1.1

1
1-Sep-02 14-Jan-04 28-May-05 10-Oct-06 22-Feb-08 6-Jul-09
Date

There is definitely some cyclicality in this graph. There also appears to be a

general upward trend over time.

7
 (b) The scatterplot of difference is as follows:

0.06

0.04

0.02

0
1-Sep-02 14-Jan-04 28-May-05 10-Oct-06 22-Feb-08 6-Jul-09

-0.02

-0.04

-0.06
Date

This transformed dataset seems much more stationary.

(c) Regression results from Minitab are:

Regression Analysis: Rate diff versus Rate lag

The regression equation is

Rate diff = 0.000060 Rate lag

Predictor Coef SE Coef T P

Noconstant
Rate lag 0.0000605 0.0001753 0.35 0.730

S = 0.00822769

Analysis of Variance

Source DF SS MS F P
Regression 1 0.00000806 0.00000806 0.12 0.730
Residual Error 1259 0.08522795 0.00006769
Total 1260 0.08523601

Since the coefficient of the lag term is not statistically significant, we cannot
reject the idea that it might be 0. Therefore, this data may represent a random
walk.

(d) Including a drift term, the Minitab results become:

8
Regression Analysis: Rate diff versus Rate lag

The regression equation is

Rate diff = 0.00371 - 0.00273 Rate lag

Predictor Coef SE Coef T P

Constant 0.003705 0.002788 1.33 0.184
Rate lag -0.002732 0.002108 -1.30 0.195

S = 0.00822519 R-Sq = 0.1% R-Sq(adj) = 0.1%

Analysis of Variance

Source DF SS MS F P
Regression 1 0.00011361 0.00011361 1.68 0.195
Residual Error 1258 0.08510842 0.00006765
Total 1259 0.08522203

Since the constant term here is not statistically significant, it is not likely that this
data follows a random walk with drift.

(e) Including a trend variable, the results are:

Regression Analysis: Rate diff versus time, Rate lag

The regression equation is

Rate diff = 0.00738 + 0.000001 time - 0.00621 Rate lag

Predictor Coef SE Coef T P

Constant 0.007378 0.003677 2.01 0.045
time 0.00000143 0.00000094 1.53 0.126
Rate lag -0.006205 0.003096 -2.00 0.045

S = 0.00822080 R-Sq = 0.3% R-Sq(adj) = 0.2%

Analysis of Variance

Source DF SS MS F P
Regression 2 0.00027197 0.00013598 2.01 0.134
Residual Error 1257 0.08495006 0.00006758
Total 1259 0.08522203

(f) The last model indicates there may be a general overall trend over time
since the constant and lagged rate variable now appear to be statistically
significant.

12.21 (a) The logistic regression model to predict whether a woman works or not is:

9
Binary Logistic Regression: work versus age, married, ...

Response Information

Variable Value Count

work 1 1343 (Event)
0 657
Total 2000

Logistic Regression Table

Odds 95% CI
Predictor Coef SE Coef Z P Ratio Lower Upper
Constant -4.15925 0.332040 -12.53 0.000
age 0.0579303 0.0072210 8.02 0.000 1.06 1.04 1.07
married 0.741777 0.126470 5.87 0.000 2.10 1.64 2.69
children 0.764488 0.0515287 14.84 0.000 2.15 1.94 2.38
education 0.0982513 0.0186522 5.27 0.000 1.10 1.06 1.14

Log-Likelihood = -1027.914
Test that all slopes are zero: G = 476.616, DF = 4, P-Value = 0.000

(b) All the independent variables appear to be statistically significant because the
p-values are essentially 0. For a one year increase in age, a woman is 6% more
likely to be in the work force, having an extra child increases the likelihood of a
woman being in the workforce by 115%, and one extra year of education
increases that likelihood by 10% (all other variables held constant). Since married
is a dummy variable, if a woman is married, she is 210% more likely to be in the
workforce than if she were not married. These values can be found in the odds
ratio column of the output.

12.22 (a) The logistic results are:

Variable Value Count

Cancer 1 40 (Event)
0 138
Total 178

Logistic Regression Table

Odds 95% CI
Predictor Coef SE Coef Z P Ratio Lower Upper
Constant -5.56459 3.98530 -1.40 0.163
Age -0.0518865 0.0394548 -1.32 0.188 0.95 0.88 1.03
HIGD -0.201064 0.232776 -0.86 0.388 0.82 0.52 1.29
Degree 0.436700 0.526090 0.83 0.406 1.55 0.55 4.34
CHK -1.49170 0.560094 -2.66 0.008 0.22 0.08 0.67
AGPI 0.0982714 0.0596606 1.65 0.100 1.10 0.98 1.24
MenAge 0.394304 0.135180 2.92 0.004 1.48 1.14 1.93
Miscarriages 0.420903 0.272058 1.55 0.122 1.52 0.89 2.60
Births 0.204228 0.141441 1.44 0.149 1.23 0.93 1.62
Weight -0.0262376 0.0107538 -2.44 0.015 0.97 0.95 0.99

10
AgePeriod 0.103813 0.0542296 1.91 0.056 1.11 1.00 1.23
Marital Status -0.107249 0.226161 -0.47 0.635 0.90 0.58 1.40

Log-Likelihood = -67.814
Test that all slopes are zero: G = 54.056, DF = 11, P-Value = 0.000

It is apparent that not all the variables are statistically significant. Especially with
logistic regression models, removal of a single variable can potentially drastically
change the significance of coefficients of the other independent variables. It is left
to the reader to follow through and decide on the best model.

11