10.

11 Chapter Exercises 379

• Check the sample size. A small sample might produce multicollinearity even if
the variables are not multicollinear throughout the population. If you suspect this
could be the case, include additional observations in the sample.
• Reconsider the theoretical assumptions of the model. In particular, ask whether
your regression model is overparameterized.
• Not infrequently, correlating variables can be combined into a single variable
with the aid of factor analysis (see Sect. 13).

10.11 Chapter Exercises

Exercise 1
You’re an employee in the market research department of a coffee roasting company
who is given the job of identifying the euro price of the company’s coffee in various
markets and the associated market share. You discover that market share ranges
between 0.20 and 0.55. Based on these findings, you try to estimate the influence of
price on market share using the regression indicated below.
Regression function: Market share ^y ¼ 1.26–0.298  price

(a) What average market share can be expected when the coffee price is 3 euros?
(b) You want to increase the market share to 40%. At what average price do you
need to set your coffee to achieve this aim?
(c) The regression yields an R2 of 0.42. What does this parameter tell us?
(d) How large is the total sum of squares when the error sum of squares of the
regression is 0.08?

Exercise 2
You have a hunch that the product sales mentioned in Exercise 5 (Chap. 3) are not
determined by price alone. So you perform a multivariate regression using Excel
(or statistics software like SPSS). The results of the regression are listed in
Fig. 10.18.

(a) Derive the regression function in algebraic form from the data in the table.
(b) Does the model serve to explain sales? Which metric plays a role in the
explanation and what is its value?
(c) Assume you lower the price in every country by 1000 monetary units. How
many more products would you sell?
(d) What is the effect of increasing advertising costs by 100,000 monetary units?
Explain the result and propose measures for improving the estimating equation.

Exercise 3
You’re given the job of identifying the market share of a product in various markets.
You determine the market share ranges between 51.28% and 61.08%. You try to
estimate the factors influencing market share using the regression in Fig. 10.19:
380 10 Regression Analysis

Model Summary
Model R R Square Adjusted R Square Std. Error of the Estimate
1 .975a .951 .927 .510
a. Predictors: (Constant), Advertising budget [in 100,000s MUs], Number of dealerships, Unit
price [in 1,000s of MUs]

ANOVAa
Model Sum of Squares df Mean Square F Sig.
Regression 30.439 3 10.146 39.008 .000b
1 Residual 1.561 6 0.260
Total 32.000 9
a. Dependent Variable: Sales [in 1,000s of units]
b. Predictors: (Constant), Advertising budget [in 100,000s MUs], Number of dealerships, Unit
price [in 1,000s of MUs]

Model Unstandardized
Coefficients t Sig.
B Std. Error
(Constant) 24.346 3.107 7.84 .000
Number of dealerships .253 .101 2.50 .047
1
Unit price [in 1,000s of MUs] -.647 .080 -8.05 .000
Advertising budget [in 100,000s MUs] -.005 .023 -0.24 .817

Fig. 10.18 Regression results (1)

Model Summary
Model R R Square Adjusted R Square Std. Error of the Estimate
1 ??? ??? ??? .652

ANOVAa
Model Sum of Squares df Mean Square F Sig.
Regression 124.265 2 ??? 145.971 .000
1 Residual ??? 24 ???
Total 134.481 26

Model Unstandardized
Coefficients t Sig.
B Std. Error
(Constant) 38.172 1.222 31.24 .000
1 price .-7.171 .571 -12.56 .000
ln(price) .141 .670 -0.21 .835

Fig. 10.19 Regression results (2)

10.11 Chapter Exercises 381

Model Summary
Model R R Square Adjusted R Square Std. Error of the Estimate
1 .883 .780 .771 187.632

ANOVAa
Model Sum of Squares df Mean Square F Sig.
Regression 18627504.189 6 ??? 84.000 .000
1 Residual 5245649.061 149 ???
Total 13423873153.250 155

Model Unstandardized Coefficients Standardized

Coefficients Sig.
B Std. Error
BETA
(Constant) 9897.875 146.52 .000
Price of Senso White -949.518 59.094 -.64 .000
Senso White advertised with
338.607 188.776 .19 .075
leaflets
Other toothpaste brands
-501.432 74.345 -.27 .000
advertised with leaflets
1
Other toothpaste brands
-404.053 87.042 -.18 .000
advertised in daily newspapers
Senso White advertised in daily
245.758 73.186 .13 .001
newspapers
Senso White advertised with
286.195 202.491 .15 .160
leaflets that contain images

Fig. 10.20 Regression toothpaste

(a) Derive the regression function in algebraic form from the above table.
(b) Determine R2 and the adjusted R2.
(c) How large is the residual sum of squares?
(d) Does the model have an explanatory value for determining market share?
(e) What’s a reasonable way to improve the model?
(f) What happens when the product price is raised by one monetary unit?

Exercise 4
You’re an employee in the market research department of a company that
manufactures oral hygiene products. You’re given the task of determining weekly
sales of the toothpaste Senso White at a specific drugstore chain over the past 3 years.
You attempt to estimate the factors influencing weekly market share using the
regression in Fig. 10.20. The potential factors include:

• The price of Senso White (in €)

• Senso White advertised with leaflets by the drugstore chain (0 ¼ no; 1 ¼ yes)
• Other toothpaste brands advertised with leaflets by the drugstore chain (0 ¼ no;
1 ¼ yes)
382 10 Regression Analysis

• Other toothpaste brands advertised in daily newspapers by the drugstore chain

(0 ¼ no; 1 ¼ yes)
• Senso White advertised in daily newspapers by the drugstore chain (0 ¼ no;
1 ¼ yes)
• Senso White advertised with leaflets that contain images by the drugstore chain
(0 ¼ no; 1 ¼ yes)

(a) Derive the regression equation in algebraic form from Fig. 10.20.
(b) What sales can be expected with a toothpaste price of €2.50 when the drugstore
chain uses no advertising for Senso White and uses leaflets for a competing
toothpaste?
(c) Interpret R, R2, and adjusted R2. Explain the purpose of the adjusted R2.
(d) What is the beta needed for?
(e) Assume you want to improve the model by introducing a price threshold effect
to account for sales starting with €2.50. What is the scale of the price threshold
effect? Which values should be used to code this variable in the regression?

Exercise 5
The fast-food chain Burger Slim wants to introduce a new children’s meal. The
company decides to test different meals at its 2261 franchises for their effect on total
revenue. Each meal variation contains a slim burger and, depending on the franchise,
some combination of soft drink (between 0.1 and 1.0 litre), salad, ice cream, and a
toy. These are the variables:

• Revenue: Revenue through meal sales in the franchise [in MUs]

• Salad: Salad, 1 (salad); salad, 0 (no salad)
• Ice cream: Ice cream, 1 (ice cream); ice cream, 0 (no ice cream)
• Toy: Toy, 1 (toy); toy, 0 (no toy)
• Sz_Drink: Size of soft drink
• Price: Price of meal

You perform two regressions with the results in Fig. 10.21:

(a) Calculate R2 from regression one.

(b) What is the adjusted R2 needed for?
(c) Using regression one determine the average revenue generated by a meal that
costs five euros and contains a slim burger, a 0.5 l soft drink, a salad, and a toy.
(d) Using regression two determine which variable has the largest influence.
Explain your answer.
(e) Compare the results of regressions one and two. Which of the solutions would
you consider in a presentation for the client?
10.11 Chapter Exercises 383

Regression 1:
Model Summary
Model R R Square Adjusted R Square Std. Error of the Estimate
1 ??? ??? .747 3911.430

ANOVAa
Model Sum of Squares df Mean Square F Sig.
Regression ??? 4 ??? 1668.726 .000
1 Residual 34515190843.303 2256 ???
Total 136636463021.389 2260

Model Unstandardized Standardized

Coefficients Coefficients t Sig.
B Std. Error BETA
(Constant) 25949.520 265.745 97.648 .000
Price 4032.796 73.255 .58 55.051 .000
1 Salad -7611.182 164.631 -.49 -46.232 .000
Ice Cream 3708.259 214.788 .18 17.265 .000
Toy 6079.439 168.553 .38 36.068 .000

Regression 2:
Model Summary
Model R R Square Adjusted R Square Std. Error of the Estimate
1 .866 .750 .750 3891.403

ANOVAa
Model Sum of Squares df Mean Square F Sig.
Regression 102488948863.420 5 ??? 1353.613 .000
1 Residual 34147514157.969 2255 ???
Total 136636463021.389 2260

Model Unstandardized Standardized

Coefficients Coefficients Sig. Tolerance VIF
B Std. Error BETA
(Constant) 25850.762 265.143 .000
Price -30.079 827.745 -.004 .971 .008 129.174
Sz_Drink 24583.927 4989.129 .590 .000 .008 129.174
1
Salad 7619.569 163.797 -.490 .000 .999 1.001
Ice Cream 3679.932 213.765 .182 .000 .997 1.003
Toy 6073.666 167.694 .382 .000 .999 1.001

Fig. 10.21 Regression results Burger Slim

384 10 Regression Analysis

Revenue through meal sales [in MUs]

Price

Fig. 10.22 Scatterplot

(f) Consider the scatterplot in Fig. 10.22. What’s the problem? Describe the effects
of the results from regressions one and two on interpretability. How can the
problem be eliminated?

10.12 Exercise Solutions

Solution 1

(a) Market share ¼ 1.26–0.298  Price ¼ 1.26–0.298  3 ¼ 36.6%.

0:298 ¼ €2:89.
(b) 0.40 ¼ 1.26–0.298  price , price ¼ 0:40
1:26
(c) 42% of the variance in market share is explained by variance in the independent
variable price.
(d) R2 ¼ 1
ESS
TSS , TSS ¼ 1
R2 ¼ 0:58 ¼ 0:14.
ESS 0:08

Solution 2

(a) ^y ¼ 24:346 þ 0:253  x1
0:647  x2
0:005  x3 , where:

x1: number of locations
x2: item price [in 1000s of MUs]
x3: advertising budget [in 100,000 s of MUs]
10.12 Exercise Solutions 385

The p-value of the F-Test is p ¼ 0.000. There is at least one variable with a
significant influence. These are the variables number of dealerships
( p ¼ 0.047 < 0.05) and unit price ( p ¼ 0.000 < 0.05). The insignificant
influence of advertising budget ( p ¼ 0.817 > 0.05) would, in practice, eliminate
the variable x3 from the regression (see part d) of the exercise, yielding the
following result: ^y ¼ 24:346 þ 0:253  x1
0:647  x2 .
(b) The p-value of the F-Test is p ¼ 0.000. There is at least one variable with a
significant influence. We already know the coefficient of determination:
R2 ¼ 0.951.
(c) The regression coefficient for the item price is α2 ¼
0:647. Since the item price
is measured by 1000s of units, a price decrease of 1000 MUs affects sales as
follows: Δsales ¼ (
1)  (
0.647) ¼ 0.647. Sales is also measured by 1000s of
units, which means that total sales increase by 1000  0.647 ¼ 647 units.
(d) The regression coefficient for advertising expenses is α3 ¼
0:005. Since the
advertising expenses are measured by 100,000 s of MUs, an increase of adver-
tising expenses by 100,000 MUs affects sales as follows: Δsales ¼ (+1) 
(
0.005) ¼ (
0.005). Sales are measured in 1000s of units, which means
they will sink by 1000  (
0.005) ¼ (
5). The result arises because the variable
advertising budget is an insignificant influence (close to zero); advertising
appears to play no role in determining sales.

Solution 3

(a) ^y ¼ 38:172
7:171  x1 þ 0:141  x2 , where:

x1: price of the company’s product.
x2: price of the competition’s product put through the logarithmic function.
The p-value of the F-Test is p ¼ 0.000. There is at least one variable with a
significant influence. This is the variable price ( p ¼ 0.000 < 0.05). The
insignificant influence of the competition’s price put through the logarithmic
function ( p ¼ 0.835 > 0.05) would, in practice, eliminate the variable x2 from
the regression [see part (e) of the exercise], yielding the following result:
^y ¼ 38:172
7:171  x1 .
Sum of Squares ðRSSÞ
(b) R2 ¼ Explained
Total Sum of Squares ðTSSÞ ¼ 134:481 ¼ 0:924;
124:265


R2adj ¼ 1
1
R2 n
1 n
k ¼ 1
ð1
0:924Þ 27
3 ¼ 0:918.
27
1

(c) RSS + ESS ¼ TSS , ESS ¼ TSS—RSS ¼ 10.216.

(d) Yes, because R2 has a very high value.
(e) By eliminating the price subjected to the logarithmic function [see exercise
section (a)].
(f) The regression coefficient for the price is α1 ¼
7:171. This means sales would
sink by (+1)(
7.171) ¼
7.171 percentage points.
386 10 Regression Analysis

Solution 4

(a) ^y¼ 9898
949.5price + 338.6LLsw
501.4LLOT
404.1NPOT + 245.8NPsw +

286.2LLSW_image.
(b) ^y ¼ 9898
949.52.5 + 338.60
501.41
404.10 + 245.80 +
286.20  7023.
(c) R equals the correlation coefficient; R2 is the model’s coefficient of determina-
tion and expresses the percentage of variance in sales explained by variance in
the independent variables (right side of the regression function). When creating
the model, a high variance explanation needs to be secured with as few variables
as possible. The value for R2 will stay the same even if more independent
variables are added. The adjusted R2 is used to prevent an excessive number
of independent variables. It is a coefficient of determination corrected by the
number of regressors.
(d) Beta indicates the influence of standardized variables. Standardization is used to
make the independent variables independent from the applied unit of measure
and thus commensurable. The standardized beta coefficients that arise in the
regression thus have commensurable sizes. Accordingly, the variable with the
largest coefficient has the largest influence.
(e) Create a new metric variable with the name Price_low. The following
conditions apply: Price_low ¼ 0 (when the price is smaller than €2.50); other-
wise Price_low ¼ Price. Another possibility: create a new variable with the
name Price_low. The following conditions apply here: Price_low ¼ 0 (when the
price is less than €2.50); otherwise Price_low ¼ 1.

Solution 5
RSS ESS 34,515,190,843:303
R2 ¼ ¼1
¼1
¼ 0:7474
TSS TSS 136,636,463,021:389

(a) In order to compare regressions with varying numbers of independent variables.

(b) Average proceeds ¼ 25,949.5 + 54032.79—
7611.182 + 6079.44 ¼ 44,581.752 MU.
(c) Lettuce, because the standardized beta value has the second largest value.
(d) The price and size of the beverage in regression two show a high VIF value,
i.e. a low tolerance. In addition, the R2 of regression one to regression two has
barely increased. The independent variables in regression two are multicollinear,
distorting significances and coefficients. The decision impinges on
regression one.
(e) Nonlinear association exists. As a result, systematic errors occur in certain areas
of the x-axis in the linear regression. The residuals are autocorrelated. The
systematic distortion can be eliminated by using a logarithmic function or by
inserting a quadratic term.