7

Binary Logistic Regression to Predict

Customer Satisfaction at a Restaurant

The owners of a busy restaurant in downtown Boston wish to improve

customer satisfaction. To this end, the restaurant manager is asked to col-
lect customer satisfaction data and suggest feasible recommendations to
the owners.
Sections 7.1 and 7.2 give brief descriptions of the dene phase and the
measure phase, respectively. Section 7.3 illustrates the analyze phase with
detailed instructions for using Minitab®. The improve phase and the control
phase are briey explained in Section 7.4.

7.1 Define Phase

The objective is to improve the proportion of satised customers at the res-
taurant. Inasmuch as customer satisfaction is a binary (“yes” or “no”) event,
the restaurant manager uses binary logistic regression that might help to
predict the probability of satisfying a given customer.
If P is the probability of satisfying a customer, then the binary regression
model is

 P 
log(odds of satisfying a customer) = log  = β 0 + β1 X 1 + β 2 X 2 + 
 1 − P 

where β0 is a constant and β1, β2, . . . are coefcients of independent factors

X1, X2, . . . In other words,

e(β0 +β1X1 +β2 X2 +)

P=
1 + e(β0 +β1X1 +β2 X2 +)

The independent factors that the restaurant manager wishes to consider

are the variety of drinks on the menu, the number of waiters, and whether a
free rell is offered.

105
106 Six Sigma Case Studies with Minitab®

FIGURE 7.1
Data collected to measure customer satisfaction.

7.2 Measure Phase

Random samples of customers are taken on different days with different
levels of these factors. These customers are asked whether they are satised
with the service.
Open the CHAPTER_7.MTW worksheet for the collected data (the work-
sheet is available at the publisher’s website; the data from the worksheet are
also provided in the Appendix). Figure 7.1 shows a screenshot of the work-
sheet. Notice that the rst row in the worksheet is on a day when 6 different
drinks are on the menu, 20 waiters are working, a free rell is offered, and 4
out of the 10 sampled customers are satised with the service.

7.3 Analyze Phase

Figure 7.2 shows how to select “Binary Logistic Regression”. Doing so opens
the dialog box shown in Figure 7.3. Select “Response in event/trial format”,
select the “Satised” column for “Number of events”, and “Sample Size”
column for “Number of trials”. For the “Model”, select “Variety of Drinks”,
“Number of Waiters”, and “Free Rell”. Because “Free Rell” is a categorical
variable (as opposed to a numerical variable), select “Free Rell” for “Factors
(optional)”. Click on “Options” and the dialog box shown in Figure 7.4 opens.
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 107

FIGURE 7.2
Selection of “Binary Logistic Regression”.

FIGURE 7.3
Selection of variables for binary logistic regression model.
108 Six Sigma Case Studies with Minitab®

FIGURE 7.4
Selection of “Logit”.

Ensure that “Logit” is selected under “Link Functions” and click on “OK”.
This takes you back to the dialog box shown in Figure 7.3. Click on “OK” and
the output shown in Figure 7.5 is the result. Inasmuch as the P-value (0.673)
for “Free Rell” is not less than 0.05, it is clear that changing “Free Rell”
from “No” to “Yes” has no signicant impact on “Satised”. Hence, remove
“Free Rell” from the model and repeat the analysis as shown in Figures 7.6
and 7.7. The revised output is shown in Figure  7.8. The P-value (0.043) for
“Variety of Drinks” and the P-value (0.023) for “Number of Waiters” are
less than 0.05, therefore both of these factors have a signicant impact on
“Satised”. Because the coefcient (0.409216) for “Variety of Drinks” is posi-
tive, an increase in “Variety of Drinks” increases the probability of customer
satisfaction. If “Number of Waiters” is constant, an increase of “Variety of
Drinks” by 1 will increase the odds of a customer being satised by 1.51.
(See the “Odds Ratio” column in Figure 7.8.) Because the coefcient (0.140952)
for “Number of Waiters” is positive, an increase in “Number of Waiters”
increases the probability of customer satisfaction. If “Variety of Drinks” is
constant, an increase of “Number of Waiters” by 1 will increase the odds of a
customer being satised by 1.15. (See the “Odds Ratio” column in Figure 7.8.)
The log-likelihood P-value (0.007) is less than 0.05, therefore not all coef-
cients in the binary logistics regression model are zero. Also, because the
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 109

Link Function: Logit

Response Information

Variable Value Count

Satisfied Event 22
Non-event 76
Sample Size Total 98

Logistic Regression Table

Odds 95% CI
Predictor Coef SE Coef Z P Ratio Lower Upper
Constant -5.33438 1.52679 -3.49 0.000
Variety of Drinks 0.408963 0.202161 2.02 0.043 1.51 1.01 2.24
Number of Waiters 0.141972 0.0620037 2.29 0.022 1.15 1.02 1.30
Free Refill
Yes -0.216707 0.513409 -0.42 0.673 0.81 0.29 2.20

Log-Likelihood = -47.080
Test that all slopes are zero: G = 10.216, DF = 3, P-Value = 0.017

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 2.09787 6 0.910
Deviance 2.59536 6 0.858
Hosmer-Lemeshow 1.87598 7 0.966

FIGURE 7.5
Output of binary logistic regression.

FIGURE 7.6
Repeat of analysis.
110 Six Sigma Case Studies with Minitab®

FIGURE 7.7
Removal of “Free Rell” from the model.

Link Function: Logit

Response Information

Variable Value Count

Satisfied Event 22
Non-event 76
Sample Size Total 98

Logistic Regression Table

Odds 95% CI
Predictor Coef SE Coef Z P Ratio Lower Upper
Constant -5.42252 1.51334 -3.58 0.000
Variety of Drinks 0.409216 0.201962 2.03 0.043 1.51 1.01 2.24
Number of Waiters 0.140952 0.0618736 2.28 0.023 1.15 1.02 1.30

Log-Likelihood = -47.170
Test that all slopes are zero: G = 10.037, DF = 2, P-Value = 0.007

Goodness-of-Fit Tests

Method Chi-Square DF P
Pearson 0.145305 2 0.930
Deviance 0.143553 2 0.931
Hosmer-Lemeshow 0.145305 3 0.986

FIGURE 7.8
Revised output of binary logistic regression.
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 111

FIGURE 7.9
Creation of new columns, “Variety 2” and “Waiters 2”.

P-value of each of the three goodness-of-t tests is not less than 0.05, the
binary logistics regression model is a good t to the data collected.
The probability of customer satisfaction can be predicted as follows, with
the equation:

e( −5.42252 + 0.409216*Variety of Drinks + 0.140952*Number of Waiters)

P=
1 + e( −5.42252 + 0.409216*Variety of Drinks + 0.140952*Number of Waiters)

For prediction of customer satisfaction for different levels of “Variety of

Drinks” and “Number of Waiters”, label two empty columns as “Variety 2”
and “Waiters 2”, as shown in Figure 7.9. Then, as shown in Figure 7.10, select
“Make Mesh Data”. Doing so opens the dialog box shown in Figure  7.11.
For “X”, select “Variety 2” for “Store in”, enter “3” for “From”, enter “6” for
“To”, and enter “4” for “Number of positions”. (This is done in order to
consider the following 4 values for “Variety of Drinks”: 3, 4, 5, 6.) For “Y”,
select “Waiters 2” for “Store in”, enter “10” for “From”, enter “20” for “To”,
and enter “3” for “Number of positions”. (This is done in order to consider
the following three values for “Number of Waiters”: 10, 15, 20.) Click on
“OK”, and all possible combinations of “Variety of Drinks” and “Number
of Waiters” appear in “Variety 2” and “Waiters 2” columns, as shown in
Figure  7.12. In order to calculate the probability of customer satisfaction
for each of these combinations, label an empty column as “Prob Satistifed”
as shown in Figure 7.12. Then, select “Calculator” as shown in Figure 7.13.
112 Six Sigma Case Studies with Minitab®

FIGURE 7.10
Selection of “Make Mesh Data”.

FIGURE 7.11
Entry of positions for mesh data.
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 113

FIGURE 7.12
Meshed data.

FIGURE 7.13
Selection of “Calculator”.
114 Six Sigma Case Studies with Minitab®

FIGURE 7.14
Selection of “Exponential”.

Doing so opens the dialog box shown in Figure 7.14. Select “Prob Satised”
for “Store result in variable”, select “Logarithm” from the drop-down menu
for “Functions”, and then double-click on the “Exponential” option. This
creates the “EXP” function in the “Expression” box. Type the formula as
shown in Figure 7.15. Click on “OK” and the probabilities are calculated as
shown in Figure 7.16.
In order to graph a scatterplot of the probabilities, select “Scatterplot” as
shown in Figure 7.17. Doing so opens the dialog box shown in Figure 7.18.
Select “With Connect and Groups” and click on “OK”. This opens the dia-
log box shown in Figure 7.19. Select “Prob Satised” for “Y variables” and
“Variety 2” for “X variables”. Then, select “Waiters 2” for “Categorical
variables for grouping (0-3)” and click on “OK”. The scatterplot shown in
Figure 7.20 is the result. As is evident, the probability of customer satisfaction
increases with an increase in the number of waiters and with an increase in
the number of different drinks on the menu.
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 115

FIGURE 7.15
Entry of formula for probability of customer satisfaction.

FIGURE 7.16
Probabilities of customer satisfaction.
116 Six Sigma Case Studies with Minitab®

FIGURE 7.17
Selection of “Scatterplot”.

FIGURE 7.18
Selection of “With Connect and Groups”.
Binary Logistic Regression to Predict Customer Satisfaction at a Restaurant 117

FIGURE 7.19
Selection of variables for scatterplot.

Scatterplot of Prob Satised vs Variety 2

0.5 Waiters 2
10
15
0.4 20
Prob Satised

0.3

0.2

0.1

0.0
3.0 3.5 4.0 4.5 5.0 5.5 6.0
Variety 2

FIGURE 7.20
Scatterplot.
118 Six Sigma Case Studies with Minitab®

7.4 Improve and Control Phases

The manager recommends to the owners to stop offering rells on drinks,
increase the variety of drinks on the menu, and increase the number of wait-
ers working in the restaurant. The owners consider implementing these rec-
ommendations and encourage the manager to continue collecting feedback
from the customers in the future so that quality is continuously improved
and controlled.