UVA-M-0859

Rev. Feb. 3, 2014

LOGISTIC REGRESSION

Almost all of us are familiar with odds. What are the chances one thing will happen
versus another? What are the chances you will succeed at work today? What are the chances
your favorite game-show contestant will win today versus the chances he or she will lose?

What we might not be familiar with is how odds can be applied to marketing analytics.
What are the chances a customer will buy your product versus the chances he or she wont?
What are the chances you will retain a customer versus the chances you will lose him or her?

When you are using odds, you are examining two opposing outcomes. Any such
unknown (i.e., one that can only be one thing or another) is known as a dummy variable. But if
you know how to examine dummy variables properly, the results are anything but dumb.

When Logistic Regression Trumps Linear Regression

A logistic regression is similar to any linear regression but with one important variation
that has critical consequences.

Think about an important metric in marketing: customer retention. If Keepmoney Bank

wants to use a regression analysis to examine whether it will retain a customer, it will set
retention as its dependent variable. Rather than being normally distributed in a bell curve in the
manner of continuous variables (Figure 1), however, a 1 will be assigned to represent customer
retention and a 0 will represent customer loss. Only those two outcomes are possible. Again, this
is a dummy variable, wherein what you are trying to predict is one of two options.

This technical note was prepared by Shea Gibbs, Research Assistant, and Rajkumar Venkatesan, Bank of America
Research Professor of Business Administration. Copyright 2013 by the University of Virginia Darden School
Foundation, Charlottesville, VA. All rights reserved. To order copies, send an e-mail to
[email protected]. No part of this publication may be reproduced, stored in a retrieval system,
used in a spreadsheet, or transmitted in any form or by any meanselectronic, mechanical, photocopying,
recording, or otherwisewithout the permission of the Darden School Foundation.

T h i s d o c u m e n t i s a u t h o r i z e d f o r u s e
A n y u n a u t h o r i z e d u s e o r r e p r
 -2- UVA-M-0859

Figure 1. A normal distribution.

Profits

0 +
Source: All figures created by case writer unless otherwise specified.

Studies have shown that logistic regression is the best model for examining dummy
variables such as customer retention.1 But why cant Keepmoney use its trusty linear regression
to determine the likelihood of customer retention given a set of independent variables? Again,
linear regressions assume a bell-curve distribution of outcomes (what is known as a normal
distribution) from negative infinity to infinity. Most things in life follow this sort of distribution.
Think of human height or school gradesa few people typically earn Cs, a few more earn a B,
the majority will earn Bs, and a very few will earn an A+. But when examining a dummy
variable such as customer retention, there is no curve across a range of outcomes. The outcome
can only be 1 or 0.

If Keepmoney attempts to use a linear regression to examine customer retention,

nonsensical predictions may result. The bank may find its chances of customer retention are
greater than 1, meaning it has even better than a 100% chance of retaining a customer. Or the
bank may find its chances are less than 0. One can round up for those predictions that are less
than 0 or round down for those greater than 1, but the results of the regression will not be
precise.

Choice Behavior

The objective of logistic regression in this example is to represent consumers choice

behavior as accurately as possible. When individual consumers choose products, the value they
place on the product does not typically increase linearly with increases in a preferred feature of
the product. Instead, research indicates consumer valuation of a product typically follows an S-
shaped curve with increases in the levels of a preferred attribute.

1
Scott A. Neslin et al., Defection Detection: Measuring and Understanding the Predictive Accuracy of
Customer Churn Models. Journal of Marketing Research 43, no. 2 (2006): 204211.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 11/8/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -3- UVA-M-0859

We can test whether the S-shaped curve represents consumers choice behavior with a
simple exercise. Imagine that on the x-axis we have the level of discount on a $300 plane ticket
from Charlottesville, Virginia, to New York. Ask a group of your friends how many of them
would purchase the flight. Then offer a discount of $20. How many additional people said they
would buy the ticket? Probably not many. Increase the discount to $40. Maybe one person half-
heartedly jumps in. At $60, you are likely to see a spike in purchasers. And from $60 to $100,
the number of purchasers should increase at every level; however, at about $100, the number of
additional purchasers will taper off, as you have reached the upper threshold.

In most real-life situations, this S-shaped curve represents how people make decisions.
As a discount (i.e., promotion) increases, the odds that people will make the choice to buy will
increase. In this example, at a $60 discount, 2 in 10 people are likely to purchase the flight to
New York; 8 in 10 are unlikely to purchase the flight.

The Logistic Transformation

We now see that a linear regression would be insufficient to accurately represent

individual consumers choices. In Figure 2, we show a distribution of probabilities from 0 to 1
representing the logistic function
1

where up = utility consumer obtains from product p = a + b1X.

Figure 2. Distribution of probabilities for a logistic distribution.

The utility function (up), otherwise known as a value function, is used to describe the
value a person places on a certain good or service. Take coffee, for example. To find the utility,
or value, you might derive from a cup of coffee, you must consider all of the variables that might
go into the decision to buy that particular cup: the taste, the price, the logo, the location of the
store from which you buy it, your personal habits, and the jolt it gives you in the morning. For
convenience purposesand based on behavioral studies indicating how people process variables
in an additive waythe value function is assumed to be linear.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 8/11/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -4- UVA-M-0859

The logistic function used to describe the ways in which consumers make choices takes
the form of the exponent of the value function over 1 plus the exponent of the value function.
The resulting distribution looks like an S-shaped curve, as shown in Figure 2. The predictions
from this function are bound between 0 and 1 (meaning if one outcome is 0.1, the opposite
outcome is 0.9).

Furthermore, the probability of success (retention) versus failure (churn) is P (1 P),

where P is the probability of retention. For example, if there are 10 outcomes with 1 success and
9 failures, the odds are 1/9. This (P (1 P)) is what is known as the odds function.
Substituting for P using the logistic function above, the odds function is equal to e(a+b1X). If we
are to make a transformation of this exponential function to a linear function via the natural log,2
we will find the log odds function, which is ln[P (1 P)] = a + b1X (Figure 3). This is
equivalent to the value function.

Figure 3. Log odds function.

Essentially, we have assumed a person has a linear value function or utility underlying
his or her decision, then we have transformed that value into something useful about the chances
he or she will make a decision. Therefore, the critical output of a logistic regression is the
probability, or percent chance, a customer will stay with a company or leave the company, and
that probability is defined in terms of the value the customer places on the companys product.

Assessing Video Game Purchasers

How can a marketing manager use logistic regression techniques to find useful
information about the ways people behave? Consider the data in Figure 4, which tally the
number of sales of Xbox games through Best Buys mobile app, as reported by Kaggle.3

2
See Appendix 1 for more information on transforming an exponential function to a linear function via the
natural log.
3
Kaggle is a user-generated business analytics community. For more information, visit http://www.kaggle.com.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 11/8/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -5- UVA-M-0859

Figure 4. Sales of Xbox games through Best Buys mobile app.

customer customer
regular review review
sku game numsales abmedian browsetime new price count average
1004622 Sniper: Ghost WarriorXbox 360 53 1 (0.00017) 0 19.99 7 3.4
1010544 Monopoly StreetsXbox 360 12 1 (0.00285) 0 29.99 3 4
1011067 MySims SkyHeroesXbox 360 3 1 (0.00157) 0 19.99 1 2
1011491 FIFA Soccer 11Xbox 360 85 1 (479.80822) 0 12.99 18 4.6
1011831 Hasbro Family Game Night 3Xbox 360 6 1 0.00094 0 9.99 2 3.5
1012721 The Sims 3Xbox 360 140 1 (0.00031) 0 19.99 13 3.8
1012876 Two Worlds IIXbox 360 5 1 0.00047 0 39.99 8 3.4
1013666 Call of Duty: The War CollectionXbox 360 41 1 0.00115 0 68.18 2 4.5
1014064 Castlevania: Lords of ShadowXbox 360 15 1 (0.00235) 0 7.99 4 4.8
1032361 Need for Speed: Hot PursuitXbox 360 168 1 (0.00039) 0 19.99 45 4.2
1052221 Marvel vs. Capcom 3: Fate of Two WorldsXbox 360 28 1 (0.00092) 0 19.99 11 4

Data source: Kaggle, Data Mining Hackathon on BIG DATA (7GB) Best Buy mobile web site,
http://www.kaggle.com/c/acm-sf-chapter-hackathon-big (accessed November 5, 2013).

Each of the games shown in this data set boasts above-median sales compared with the
other games available. In other words, a dummy variable has been set where above-median
sales is represented by a 1, and below-median sales is represented by a 0. Now, which
independent variables shown in the chart (time browsed, whether the game is new, price, number
of reviews, and review average) are good predictors of being a 1that is, above-median sales?

The output of a logistic regression of this data (Figures 5 and 6) looks similar to the
output of a linear regression, and the most important data points, in addition to the coefficients,
are r squared and p-value; other predictors of accuracy and significance go by a variety of names.

Figure 5. Output of logistic regression.

Summarystatistics:

Variable Categories Frequencies %

nrx_ind 0 1128 44.183
1 1425 55.817

Variable Observations Obs.withmissingdata Obs.withoutmissingdata

salescalls 2553 0 2553
Minimum Maximum Mean Std.deviation
0.000 12.000 2.396 2.128

Goodnessoffitstatistics(Variablenrx_ind):

Statistic Independent Full

Observations 2553 2553
Sumofweights 2553.000 2553.000
DF 2552 2551
2Log(Likelihood) 3504.580 3216.666
R(McFadden) 0.000 0.082
R(CoxandSnell) 0.000 0.107
R(Nagelkerke) 0.000 0.000
AIC 3508.580 3220.666
SBC 3520.270 3232.356
Iterations 0 6

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 8/11/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -6- UVA-M-0859

Figure 6. Model estimates.

Modelparameters(Variableabmedian):

Source Value SE WaldChiSquare Pr>Chi

Intercept (1.097) 0.502 4.769 0.029
New (1.595) 1.467 1.182 0.277
Regularprice 0.006 0.011 0.279 0.597
Customerreviewcount 0.066 0.030 4.943 0.026
Customerreviewaverage 0.399 0.116 11.878 0.001

The key difference in the logistic regression output is that the coefficients are not
interpreted as such. In order for the coefficients to add value to your analysis, you must calculate
the odds ratio. For example, if a logistic regression yields a coefficient b of 2.303, the odds ratio
says that for every one unit increase in the independent variable (e.g., number of promotions),
the odds that the dependent variable will be equal to 1 (e.g., the product is purchased) will
increase by a factor determined by taking the exponent of the coefficient: eb = e2.303 = 10. This is
not the same as a direct linear transformation.

So, examining the p-values shown in the far-right column of Figure 6, which variables
can we say are predictive of whether a game will be a top seller? Customer review average,
followed by the number of customer reviews, is the most significant variable. Price is relatively
insignificant, in this case most likely due to the fact that the price range of the games is small.

Using the coefficients determined in the regression analysis, the marketing manager can
then determine how much the odds of a game being a top seller increase if review average
increases by one point (Figure 7). In other words, if a customer review average of 3 yields a
certain probability of success, what happens if the average increases to 4? On average, the
coefficient of customer review (coefficient b, the slope of the line) is 0.399, and the exponent of
b is 1.49, which means that a single-point increase in reviews increases the odds by a factor of
about 1.5.4

4
For more information on how the odds ratio can be calculated, please see Appendix 2.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 11/8/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -7- UVA-M-0859

Figure 7. Equivalence of log odds ratio and logistic probabilities.

Coefficient of Customer
Review Average (b) 0.399
For a unit increase in customer review
exp(b) 1.490
score, the odds of selling a game
increases by 49% (holding everything
else constant).
Customer Customer
Review Review
Average = 3 Average = 4
exp(bx) 3.310 4.933
Probability of Choice 0.768 0.831
Odds 3.310 4.933
Odds Ratio 1.490
Difference in Probability 0.063

Prob (sale) when customer review average is

0.768 (1 0.768) 3 = exp(0.399 3) [1 + exp(0.399 3)]

Prob (sale) when customer review average is

4 = exp(0.399 4) [1 + exp(0.399 4)]

Conclusion

Marketing managers often want to predict customer behaviors that are not distributed
across a range of outcomes. These are cases where only one of two behaviors is possible: buy or
dont buy, customer retention versus customer loss, and so on. Here, if the manager attempts to
use a traditional linear regression to examine the behaviors, nonsensical predictions can result.

But a logistic regression can be used to represent consumers choice behavior. By

transforming the value function into a logistic function, we can model how the value a consumer
places on a product increases with a preferred feature of the product. The critical output of the
logistic regression is therefore the increase (or decrease) in the percent chance a customer will
perform a behavior based on a unit increase in a variable correlated with that behavior.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 8/11/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -8- UVA-M-0859

Appendix 1
LOGISTIC REGRESSION
Understanding Exponential Functions

In order to understand logistic regressions, it is helpful to first examine exponential

functions. Figure 1 shows the classic example of an exponential distribution. When considering
the cumulative sales of a product that has gained market acceptance over time (such as
ultrasound machines), we see that sales are slow at first but begin to increase at a greater and
greater rate once they have reached critical mass. In the graph, the red line is the actual data, or
number of sales per year since introduction. What stands out is that the curve is not a straight
line, whereas the ones used in linear regressions are. This is an exponential distribution.

Figure 1. An example of an exponential distribution.

Source: Created by case writer.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 11/8/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -9- UVA-M-0859

Appendix 1 (continued)

The black line represents a function, created using a computer program,1 which best
accounts for the data shown in the graph. The regression analysis of the available data has
produced a line defined by the form y = 4.0858e0.3225x, where 4.0858 is the intercept of the line
and the slope (0.3225) changes exponentially. (The constant e is an irrational number
approximately equal to 2.71828, which is related to the rate of change in an exponential function
and is the base of the natural logarithm. This function is found in a similar way as a straight-line
function when performing a linear regression analysis.

One thing to note about this analysis is that the regression line fits almost perfectly.
Because of the volume of data used, r squareds of up to 99% are possible, as compared with the r
squareds of 20% to 30% one finds when running linear analyses. This is because the data are
aggregate and viewed retrospectively, whereas linear regressions attempt to describe the
behavior of individuals. If the same analysis of cumulative ultrasound sales was conducted in
year two, however, it would be difficult to predict what would happen in years three, four, or
five, because r squared breaks down at that point.

What does this have to do with logistic regressions? Consider the green line in Figure 1,
which represents the natural log of cumulative sales at each time period x. The line is nearly
straight, meaning a linear regression analysis could produce an accurate function describing the
data. In other words, a logistic transformation of exponentially distributed data allows you to
view the outputs of the regression in the same way you would a linear regression.2

1
For more information on how to perform a logistic regression using computer software, please visit
http://dmanalytics.org/.
2
In algebraic terms, if y = 4.0858e03225x, the natural log of y will equal 4.0858 + 3.225x, a linear function where
the intercept is 4.0858 and the slope is 3.225.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 8/11/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.
 -10- UVA-M-0859

Appendix 2
LOGISTIC REGRESSION
Calculating Odds Ratio

Let us consider the log odds ratios presented in Figure 7 and the logistic regression
output in Figure 6. The log odds ratio is defined as the probability of observing an event (p)
versus the probability of not observing and event (1 p). In the context of the choice of games
on the mobile app, we are considering the factor by which the log odds of purchasing a game
increases when the review for the product increases from 3 to 4. A simple way to calculate this
would be to take the exponent of the coefficient of reviews from the logistic regression output. In
our case, the coefficient of reviews equals 0.399. So the log odds will increase by a factor of 1.49
or 149% (exp(0.399)) when the reviews for a product increases by one unit.

In Figure 7 we show that formula for calculating the log odds factor is equivalent to (a)
computing the predicted probability of product choice when the reviews for the products are 3
and 4, and (b) then taking the ratio of these respective probabilities. The probability of product
choice when average product review equal 3 is 0.768 and the corresponding log odds is 3.3.
Similarly, the probability of choice when average product review equals 4 is 0.831 and the log
odds is 4.933. The ratio of log odds (4.933 3.3) equals 1.4. Hence the log odds increases by a
factor of 1.4 or 140% when the average reviews for the product increases by one unit.

This document is authorized for use by 5960604 IMD enrolled in the program IMD MBA 2017 (cc - 700) from 11/8/2016 to 31/12/2017
Any unauthorized use or reproduction of this document is strictly prohibited.