Applied Logistic Regression

Overview

Logistic regression is similar to ordinary multiple regression – except that logistic regression is
used when the dependent variable is binary and assumes only two discrete values. Examples
include ‘yes-no’ dependent variables such as whether a customer responded to a marketing
campaign or not, whether a person is a homeowner or not, whether a business goes bankrupt
or not, or whether a person votes guilty or not guilty. Like ordinary multiple regression, the
predictor variables can be metric variables (e.g., age, income, or sales units) or categorical
(e.g., gender, religion, or region). Indicator or ‘dummy’ variables are used to include categorical
variables as predictors.

While the basic concepts are similar for multiple linear regression and logistic regression, the
interpretation of the regression equation and the coefficients are somewhat different. In multiple
regression, the dependent variable is a continuous or metric variable – sales or profits, for
example – and can assume many values. The multiple regression coefficients are multiplied by
the values of the predictor variables to yield the predicted value for the dependent variable.

In logistic regression, the observed values for the dependent variable take on only two values
and are usually represented using a 0-1 dummy variable. The mean of a 0-1 dummy variable is
equal to the proportion of observations with a value of 1 – and can be interpreted as a
probability. The predicted values in a logistic regression will always range between 0 and 1
and are also interpreted as probabilities.

Suppose we are modeling home ownership (where 1 indicates a homeowner and 0 a non-
owner) as a function of income. Each individual in the dataset is either a homeowner or not so
the observed values for the dependent variable will be 0 or 1. The predicted value based on
the model is interpreted as the probability that the individual is a homeowner. For example, for

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 2

a person with an income of $35,000, the predicted probability may be .22 compared with a
predicted probability of .95 for a person with a $250,000 income.

Like linear regression, once a logistic regression model has been estimated it can be used to
make predictions for new observations. Assume a bank has data from a past marketing
campaign promoting a ‘gold’ credit card – including whether the customer signed up for the offer
or not (the dependent variable) as well as information on other bank services the customer used
plus financial and demographic customer information (the predictor variables). These data can
be used to estimate a logistic regression model. Then the bank could use this model to identify
which additional customers to target with this or a similar offer. By inputting values for the
predictor variables for each new customer – the logistic model will yield a predicted probability.
Customers with high predicted probabilities may be chosen to receive the offer since they seem
more likely to respond positively.

A difference between linear and logistic regression is the shape of the model as shown in
Exhibit 1. The simple linear regression is represented by a straight line. For the linear
regression, an increase of one unit in the predictor variable has a constant effect – equal to the
slope of the line. In logistic regression, the relationship between the dependent variable and the
predictor variables is assumed to be nonlinear. A logistic regression model with a single
predictor is represented by an s-shaped curve. Moreover, the curve never falls below 0 or
exceeds 1 – regardless of the values of the predictor variables. Thus, the predicted values can
always be interpreted as probabilities.

Exhibit 1 Simple Linear Regression versus Logistic Regression

Linear Regression Logistic Regression

In logistic regression, the effect on the predicted probability of a one-unit increase in the
predictor variable varies. At the extremes, a one-unit change has very little effect, but has a
larger effect in the middle. In many situations, this is intuitive. For example, consider the effect
that a $20,000 increase in income might have on the probability of home ownership. The
difference in the likelihood that an individual owns a home may not change much as their
income increases from $10,000 to $30,000 or from $1,000,000 to $1,020,000 – but may
increase quite a bit if income increases from $50,000 to $70,000. Unfortunately, this non-
linearity complicates the interpretation of the regression coefficients. In a linear regression, the
interpretation of the coefficient for X is straightforward: an increase of 1 unit in X results in a

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 3

change in the expected value of Y equal to B (the coefficient for X). However, in a logistic
regression, we cannot say that a 1 unit increase in X will result in an expected change in Y
equal to B. Rather, it depends on where on the curve the value of X is located.

The Simple Logistic Regression Curve

Consider the simple case with a single predictor variable. For now, we assume that the
predictor variable is a continuous variable. In a simple linear regression, the model would be:

Y = B0 + B1X

where B0 is the intercept or constant term (equal to the predicted value of Y when X=0), and
B1 is the slope of the regression line.

In a simple logistic regression, the model is:

e B0 + B1X
Prob(Y = 1) =
1 + e B0 + B1X

which can be also be written as:

1
Prob(Y = 1) = − ( B0 + B1X )
.
1+ e

Thus, as in linear regression, there are two coefficients, B0 and B1, in a simple logistic
regression. These coefficients determine the specific shape of the curve. The B0 coefficient
(also referred to as the constant) determines the location of the logistic curve along the X axis.
As the constant increases, the logistic curve shifts left on the X axis. The B1 coefficient
determines the steepness and direction of the curve. A positive B1 means the curve will
increase as X increases. If B1 is negative, the curve decreases as X increases. Larger values
for B1 indicate a steeper curve.

Exhibit 2 Logistic Regression Coefficients Control the Shape of the Curve

B0 coefficient ‘shifts’ the curve B1 coefficient controls the steepness

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 4

Interpreting Logistic Regression Coefficients

In simple linear regression, interpretation of the coefficients is straightforward. The constant

term estimates the value of Y when X=0. The B1 coefficient estimates the change in Y for a one
unit increase in X. Because of the nonlinear nature of the logistic regression model,
interpretation of the coefficients is more complex. To interpret logistic regression, we start with
a discussion of probabilities, odds and odds ratios.

A probability is the likelihood of an event and is bounded between 0 and 1. If the weather
forecast says the probability of rain is 0.25, then there is a 25% chance of rain. Odds are the
ratios of two probabilities: the probability that the event will occur divided by the probability that
the event will not occur. If the probability of rain is 0.25, then the odds are:

Prob (event) 0.25 1

Odds = = = = .333
Prob (no event) 0.75 3

Since odds are the ratio of two probabilities, odds are always positive, but may be greater than
one. In fact, odds can range from 0 to infinity. When the odds are less than 1, the probability of
the event (say, rain) is lower than the probability of no event (no rain). Conversely, odds greater
than 1 indicate the probability of the event is greater than the probability of no event. Odds of 1
indicate equal (that is, .50) probabilities of event and no event – meaning that both outcomes
are equally likely.

Finally, an odds ratio is the ratio of two odds. In logistic regression, the odds ratio for the
predictor variable X indicates the expected change in the odds that Prob(Y=1) for a one unit
increase in X. The odds ratio is particularly important in logistic regression because, unlike
linear regression, the ‘slope’ of the curve is not constant. However, the odds ratio for a predictor
variable is constant. The odds ratio for the predictor variable is computed by raising e to the
B
power B1, or e 1 . For example, consider a logistic regression to predict the probability of
purchase of a newly released movie DVD (the dependent variable) using the value (in dollars) of
an ‘instant coupon’. If the B1 coefficient is 0.7, we know – since the coefficient is positive - that
increasing the value of the instant coupon increases the probability of purchase. However,
because of the s-shaped curve, the magnitude of the increase will depend on whether the
increase is, say, from $1 to $2 or from $4 to $5. Since e 0.7 = 2.01 , we can say that for every
dollar increase in the instant coupon value, the odds of purchase increase by a factor of 2.01.
That is, the odds of purchase are twice a large for a $5 coupon compared with a $4 coupon.

Simple Logistic Regression Example Using SPSS

In December 1998 the U.S. Senate voted on two articles of impeachment against President
Clinton. The vote of each senator is a binary variable – taking on only two values, guilty or not
guilty. Exhibit 3 shows the number of guilty and not guilty votes on the first article of
impeachment.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 5

Exhibit 3 Number of Guilty and Not Guilty Votes on Impeachment Article 1

Vote on Article I

Cumulative
Frequency Percent Valid Percent Percent
Valid not guilty 55 55.0 55.0 55.0
guilty 45 45.0 45.0 100.0
Total 100 100.0 100.0

In addition to the each senator’s vote (guilty or not guilty) on each of the two articles of
impeachment, we have data on variables that might be predictive of how a senator voted.
These include:

• Political Party (Republican or Democrat).

• Percent of the vote Clinton received in the 1996 presidential election in the senator’s
state.
• Degree of ideological conservatism (0 – 100 where 100 is most conservative). These
ratings are issued by the American Conservative Union and are based on a senator’s
voting records.

The simple cross-tabs in Exhibit 4 shows a strong association between political party and the
first vote. All of the democratic senators voted not guilty whereas the great majority of
republican senators voted guilty.

Exhibit 4 Cross-tabs of Political Party and Vote on Article 1

Count
Vote on Article I
not guilty guilty Total
political party democrat 45 45
republican
10 45 55
Total 55 45 100

Exhibit 5 reports the correlations between the vote on the first article, degree of ideological
conservatism and percent of the vote Clinton received in the senator’s state. Degree of
conservatism has a high positive correlation (.866) with vote – guilty votes are associated with
more conservative senators. In contrast, percent of state vote shows a moderate negative
correlation (-.429) with the vote – as a larger percent of the 1996 state vote was for Clinton, the
more not guilty votes.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 6

Exhibit 5 Correlations between vote on article 1, conservatism and % of state vote for Clinton

% of vote for
Vote on degree of Clinton in
Article I conservativism state
Vote on Article I Pearson Correlation 1.000 .866** -.429**
Sig. (2-tailed) . .000 .000
N 100 100 100
degree of conservativism Pearson Correlation .866** 1.000 -.447**
Sig. (2-tailed) .000 . .000
N 100 100 100
% of vote for Clinton in Pearson Correlation -.429** -.447** 1.000
state Sig. (2-tailed) .000 .000 .
N
100 100 100

**. Correlation is significant at the 0.01 level (2-tailed).

Because the variable we want to predict is binary (vote of guilty or not guilty), logistic regression
is appropriate. While the other three variables (political party, degree of conservatism and % of
state vote for Clinton in 1996 election) are potential predictors of the vote on the first
impeachment article, high correlations or collinearity between these predictors is likely –
particularly between political party and degree of conservatism. The correlation between these
two variables is .906, suggesting that it would be unwise to include both as predictors. Since
the ideological conservatism variable captures a broader range than simple party affiliation, let
us start by considering that as a single predictor.

A simple scatter plot of vote (0=not guilty, 1=guilty) versus degree of conservatism is shown in
Exhibit 6. Because of the 0-1 nature of the dependent variable, this plot is not as insightful as
one might hope – although it does show that senators with lower ratings on conservatism
tended to vote not guilty, whereas nearly all senators with high conservatism ratings voted
guilty.

Exhibit 6 Scatterplot of senate vote (guilty =1) by degree of conservatism

1.0

.8

.6

.4
Vote on first article

.2

0.0
0 20 40 60 80 100

degree of conservativism

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 7

A logistic regression using the vote on Article 1 as the dependent variable and degree of
ideological conservatism as the predictor will allow us to quantify the relationship between these
variables and to assign a specific probability of a guilty vote to each senator.

Use the following steps to specify a logistic regression in SPSS:

• Select Analyze/Regression/Binary logistic

• Select your dependent variable
• Select your independent variable(s) (called covariates by SPSS)
• (optional)To save the predicted values (this will create a new column of data with the
predicted probabilities):
o Click on Save…
o Click on Probabilities under Predicted Values
o Click on Continue
• Click on OK.

Interpreting the SPSS Output

Exhibit 7 contains the full SPSS output from the logistic regression. Key parts of the output
have been labeled and are described below.

A: This indicates how many cases are included in the analysis and how many, if any, have
missing variables. Here, we see that all 100 cases have been included.

B: This indicates how the dependent variable is coded:

1 indicates guilty and 0 not guilty.

C: Block 0: Beginning Block:

This reports results from an ‘intercept-only’ model and we will skip directly to the Block 1
results. In essence, this model is a basis of comparison for the model specified.

D: Omnibus Tests of Model Coefficients:

These values test whether or not all of the predictor variables entered in this step, in this
block, or in the model have a significant effect. Here these values are all the same – they
will differ if you request a stepwise regression or enter the variables in ‘blocks’ (i.e., groups
of variables). Higher chi-square values indicate more significance which is reported in the
column labeled ‘Sig.’ An 0.05 cutoff for significance level is commonly used – here the
0.000 is less than 0.05 so we conclude that the model is statistically significant.

E: Model Summary:
-2 Log likelihood is used to measure how well the model fits the data. Smaller values are
better – although we will not focus on this statistic. The two R-square measures are
attempts to come up a measure comparable to the familiar R2 from ordinary multiple
regression. However because the observed values for the dependent variable in a logistic
regression can take on only two values, 0 or 1, there isn’t a measure that is totally

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 8

analogous. However, as in linear regression, higher values are ‘better’ and these measures
can be used to compare models with different sets of predictor variables.

F: Classification Table:
The classification table compares the predicted values for the dependent variable, based on
the model, with the actual observed values in the data. SPSS uses the predicted
probabilities for each case (or senator) to compute the predicted value. By default, the
SPSS cutoff is 0.50 – although a different value can be chosen.

If the predicted probability is greater than the cutoff of 0.50, SPSS predicts a value of 1
(which would be a guilty vote in this case). If the predicted probability is less than 0.50,
SPSS predicts a value of 0 (or a not guilty vote for this data). The actual and predicted
values are summarized in a 2x2 table along with the percentages of correct classifications.
Here we see that, overall, the model correctly predicts or classifies 94% (=94/100) of the
votes. The percent of not guilty votes correctly classified is 90.9% (=50/55) and the percent
of guilty votes correctly classified is 97.8% (=44/45).

G: Variables in the Equation:

The Variables in the Equation table summarizes the coefficients, standard errors,
significance tests and odds ratios for each variable in the model as well as the constant
term.

B: These are the coefficients in the logistic regression equation. In this case the final model
equation is: P(Guilty) = 1/[1+exp(-Σ BX)]
= 1/[1+exp(6.207 - .108 conservative)]

S.E.: The standard error of B.

Wald: A measure of the significance of the predictor variables. Higher values in conjunction
with the degrees of freedom (df) indicate significance.

Sig.: The significance of the Wald test. If the common 0.05 cutoff for significance level is
used, values less than 0.05 indicate statistical significance.

Exp(B): This is the odds ratio and is useful for interpreting the effects of the predictor
variables.

In summary, a senator’s degree of ideological conservatism is a significant predictor of his or

her vote on the first article of impeachment. The coefficient for the conservatism variable is
positive – so more conservative senators were more likely to cast a guilty vote. Using this
single predictor, the model correctly classified 94% of the senators, with a slightly higher
prediction accuracy for guilty votes than for not guilty. Exhibit 8 lists the predicted probabilities
for different ratings of ideological conservatism. Exhibit 9 shows a plot (based on the numbers
in Exhibit 8) of the logistic regression model.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 9

Exhibit 7 Logistic Regression SPSS Output

Logistic Regression
Case Processing Summary

Unweighted Cases a N Percent

Selected Cases Included in Analysis 100 100.0
Missing Cases 0 .0
Total 100 100.0
A
Unselected Cases 0 .0
Total 100 100.0
a. If weight is in effect, see classification table for the total
number of cases.

Dependent Variable Encoding

Original Value Internal Value

not guilty 0 B
guilty 1

Block 0: Beginning Block C

Classification Table a,b

Predicted

VOTE1 Percentage
Observed not guilty guilty Correct
Step 0 VOTE1 not guilty 55 0 100.0
guilty 45 0 .0
Overall Percentage 55.0
a. Constant is included in the model.
b. The cut value is .500

Variables in the Equation

B S.E. Wald df Sig. Exp(B)

Step 0 Constant -.201 .201 .997 1 .318 .818

Variables not in the Equation

Score df Sig.
Step 0 Variables CONSERV 75.041 1 .000
Overall Statistics 75.041 1 .000

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 10

Exhibit 6 (continued)

Block 1: Method = Enter

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 100.520 1 .000
Block 100.520 1 .000 D
Model 100.520 1 .000

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square E
1 37.107 .634 .848

Classification Table a

Predicted

VOTE1 Percentage
Observed not guilty guilty Correct F
Step 1 VOTE1 not guilty 50 5 90.9
guilty 1 44 97.8
Overall Percentage 94.0
a. The cut value is .500

Variables in the Equation

B S.E. Wald df Sig. Exp(B)

Step
a
CONSERV .108 .024 20.642 1 .000 1.114 G
1 Constant -6.207 1.566 15.698 1 .000 .002
a. Variable(s) entered on step 1: CONSERV.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 11

We see that a senator with a conservatism rating of 0 (e.g. Boxer of California) has a predicted
probability of voting guilty of 0.002. At the other extreme, a senator with a conservatism rating
of 100 (e.g. Helms of North Carolina) has a 0.99 predicted probability of voting guilty. The
‘dividing line’ appears to be between 57 and 58. A conservatism rating of 57 leads to a
predicted probability of 0.4873 – just under the 0.50 cutoff. An increase of one to a 58 rating
increases the predicted probability to 0.5142. For a simple logistic regression (i.e., one with
only one predictor variable), it is easy to compute the smallest value of the predictor needed for
a predicted probability of 0.50 or greater. This 50% cutoff value is equal to − B 0 B1 . In this
example − B 0 B1 = 6.207/.108 = 57.47. Thus, any senator with a conservatism rating of 58 or
higher would be expected to vote guilty, those with a rating of 57 or lower are expected to vote
not guilty.

In Exhibit 8 note that although conservatism increases in equal 5-unit increments, the increases
in predicted probabilities are not equal. A 5-unit increase from 0 to 5 raises the probability of a
guilty vote by only 0.0014 (=0.0034 - 0.0020), whereas an increase from 50 to 55 raises the
probability by 0.1251 (=0.4336 - 0.3085).

Exhibit 8 Table of Conservatism Ratings and Predicted Probabilities

Degree of Predicted
conservatism Probabilitya
0 0.0020
5 0.0034
10 0.0059
15 0.0101
20 0.0172
25 0.0291
30 0.0489
35 0.0811
40 0.1316
45 0.2064
50 0.3085
55 0.4336
60 0.5678
65 0.6927
70 0.7946
75 0.8691
80 0.9193
85 0.9513
90 0.9711
95 0.9829
100 0.9900
a
Predicted Probability = 1/[1+exp(-ΣBX)] = 1/[1 + exp(6.207 - .108X)]

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 12

Exhibit 9 Plot of Logistic Regression Model

Predicted Probability of Guilty Vote

1.00
0.90
Predicted Probability

0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
0 20 40 60 80 100
Degree of Ideological Conservatism

Interpreting Logistic Regression Coefficients

In a linear regression model, we use the coefficients and standardized coefficients to describe
the magnitude of the effect of that predictor. For example, if the coefficient for a predictor X is
4.5, we would say that for every unit increase in X, the predicted value of Y will increase by 4.5.
Unfortunately, the interpretation is more complicated in a logistic regression.

To interpret logistic regression results, you must understand probabilities, odds and odds ratios:

• A Probability is the likelihood of an event and is bounded between 0 and 1.

• An Odds is the ratio of two probabilities – for example, the probability of voting guilty
divided by the probability of voting not guilty. Odds cannot be negative (because
probabilities cannot be negative), but odds have no upper bound. In this example, the
odds of a senator with a conservatism rating of 100 voting guilty is .99/.01 = 99.

• An Odds Ratio is the ratio of two odds. Like odds, the odds ratio cannot be negative,
and has no upper bound.

Exhibit 10 shows the probabilities, odds and odds ratios for different values of ideological
conservatism. As we noted before, an increase of 1 in ideological conservatism does not have a
constant effect on the dependent variable (predicted probability). When conservatism increases
from 40 to 41, the probability of a guilty vote increases by 0.1444 - 0.1316 = 0.0128. When
conservatism increases from 80 to 81, the probability of a guilty vote increases by 0.9270 -
0.9193 = 0.0077. The effect of an increase in conservatism depends on where you are on the
curve in Exhibit 8.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 13

Exhibit 10 Probabilities, Odds and Odds Ratios

Conservatism Prob(Guilty) Prob(Not Guilty) Odds Odds Ratio

Rating =1-Pr(Guilty) =Pr(Guilty)/Pr(Got Guilty)
40 .1316 .8684 .1515
41 .1444 .8556 .1668 .1668/.1515=1.114
: : : :
80 .9193 .0807 11.3916
81 .9270 .0730 12.6986 12.6986/11.3916=1.114

However the odds ratio is constant. In this data, every increase of one in conservatism rating
increases the odds of a guilty vote by a factor of 1.114, or 11.4%. So while a one-unit change in
a predictor variable does not have a constant effect on the predicted probability, it does have a
constant effect on the odds.

The odds ratio for a specific predictor is found using the following formula:

Odds Ratio =exp(B) = eB

For this example, B=0.108 and exp(B) = 1.114. SPSS reports the odds ratio for each predictor
in the last column of the ‘Variables in the Equation’ output. Now consider what the odds ratio will
be for positive coefficients, negative coefficients and coefficients of zero:

• Positive coefficients: For any value of X > 0, exp(X ) >1. So a positive coefficient will
have an odds ratio greater than one, indicating that an increase in that variable will
multiply the odds by a factor greater than one – increasing the odds that Y=1.

• Negative coefficients: For any value of X < 0, 0< exp(X) <1. So a negative coefficient
will have an odds ratio less than one, indicating that an increase in that variable will
multiple the odds by a factor less than one – decreasing the odds that Y=1.

• Coefficients equal to zero: If the coefficient for a predictor variable is 0, then exp(0) = 1,
indicating that a one unit change will the odds that Y=1 by a factor of 1. Since
multiplying something by one leaves it unchanged – an odds ratio of 1 indicates no
effect (corresponding to the zero coefficient).

Note that the magnitudes of positive and negative coefficients can be compared using the
absolute values of the coefficients themselves or by taking the inverse of the odds ratio of the
negative (or positive) coefficient. For example, an odds ratio of 2.0 has the same ‘magnitude’
as an odds ratio of 0.5 (1/2 = .5, or 1/.5 = 2). One is a doubling of the odds ratio, the other is a
halving.

Adding Another Predictor to the Model

Exhibit 11 reports the key SPSS output from a second logistic regression model that includes
the percent of state vote for Clinton in the1996 election in addition to degree of ideological
conservatism. The correlations reported earlier showed that senators from states showing
strong support for Clinton in the 1996 election were less likely to vote guilty. Will including it as
a predictor improve the model?

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 14

Exhibit 11 SPSS Output from Second Logistic Regression Model

Omnibus Tests of Model Coefficients

Chi-square df Sig.
Step 1 Step 100.856 2 .000
Block 100.856 2 .000
Model 100.856 2 .000

Model Summary

-2 Log Cox & Snell Nagelkerke

Step likelihood R Square R Square
1 36.771 .635 .850

Classification Table a

Predicted

VOTE1 Percentage
Observed not guilty guilty Correct
Step 1 VOTE1 not guilty 50 5 90.9
guilty 1 44 97.8
Overall Percentage 94.0
a. The cut value is .500

Variables in the Equation

B S.E. Wald df Sig. Exp(B)

Step
a
CONSERV .105 .024 18.892 1 .000 1.111
1 STATEVOT -.040 .070 .324 1 .569 .961
Constant -4.216 3.774 1.248 1 .264 .015
a. Variable(s) entered on step 1: CONSERV, STATEVOT.

Conclusions drawn from the results in Exhibit 11 are summarized below:

• There is virtually no difference in the chi-square values or overall significance of the

model.

• The –2 Log likelihood is nearly the same, as are the R square measures.

• The classification results are identical to the model using only degree of ideological
conservatism.

• The ‘variables in the equation’ table shows that the state vote variable is not statistically
significant. The coefficient for conservatism is essentially unchanged.

In summary, this additional predictor does not improve the model and need not be included.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.
Page 15

Summary

Logistic regression is used when the dependent variable is binary. Like linear regression, the
predictor variable can be metric or categorical. In a logistic regression, the predicted values are
bounded between 0 and 1 and are interpreted as the probability that the dependent variable
equals one. Like linear regression, the coefficients and statistical tests will indicate whether the
predictor variables are statistically significant – and whether they have a positive or negative
effect on the probability that the dependent variable is one. However, unlike linear regression,
the effect of a one-unit change in X on Y is not linear – rather it depends on the value of X. The
odds ratios, in combination with the coefficients, are used to interpret the effects of individual
predictor variables.

This document is authorized for use by Julia DeNardis, from 2/8/2021 to 2/26/2021, in the course:
AEM 4015 / NBA 6340: Customer Analytics & Strategy - Park (Spring 2021), Cornell University.
Any unauthorized use or reproduction of this document is strictly prohibited*.