Chapter 24

Logistic Regression

Content list

Purpose of logistic regression 568

Assumptions of logistic regression 569
The logistic regression equation 570
Interpreting log odds and odds ratio 573
Model fit and likelihood function 574
SPSS activity – a logistic regression analysis 575

By the end of this chapter you will understand:

1 The purposes of logistic regression.
2 How to use SPSS to perform logistic regression.
3 How to interpret the SPSS printout of logistic regression.

Introduction
This chapter extends our ability to conduct regression, in this case where the DV is a
nominal variable. Our previous studies on regression have been limited to scale data DVs.

The purpose of logistic regression

The crucial limitation of linear regression is that it cannot deal with DV’s that are dichoto-
mous and categorical. Many interesting variables in the business world are dichotomous: for
example, consumers make a decision to buy or not buy, a product may pass or fail quality
control, there are good or poor credit risks, an employee may be promoted or not. A range of
regression techniques have been developed for analysing data with categorical dependent
variables, including logistic regression and discriminant analysis (DA) (Chapter 25).
Logistical regression is regularly used rather than DA when there are only two categories
of the dependent variable. Logistic regression is also easier to use with SPSS than DA when
there is a mixture of numerical and categorical IV’s, because it includes procedures for
generating the necessary dummy variables automatically, requires fewer assumptions, and
is more statistically robust. DA strictly requires the continuous independent variables
 LOGISTIC REGRESSION 569

(though dummy variables can be used as in multiple regression). Thus, in instances where
the independent variables are categorical, or a mix of continuous and categorical, and the
DV is categorical, logistic regression is necessary.

Logistic regression. Determines the impact of multiple independent variables

presented simultaneously to predict membership of one or other of the two
dependent variable categories.

Since the dependent variable is dichotomous we cannot predict a numerical value for it
using logistic regression, so the usual regression least squares deviations criteria for best fit
approach of minimizing error around the line of best fit is inappropriate. Instead, logistic
regression employs binomial probability theory in which there are only two values to
predict: that probability (p) is 1 rather than 0, i.e. the event/person belongs to one group
rather than the other. Logistic regression forms a best fitting equation or function using the
maximum likelihood method, which maximizes the probability of classifying the observed
data into the appropriate category given the regression coefficients. We will avoid the more
complicated mathematics of this.
Like ordinary regression, logistic regression provides a coefficient ‘b’, which measures
each IV’s partial contribution to variations in the DV. The goal is to correctly predict
the category of outcome for individual cases using the most parsimonious model. To
accomplish this goal, a model (i.e. an equation) is created that includes all predictor
variables that are useful in predicting the response variable. Variables can, if necessary,
be entered into the model in the order specified by the researcher in a stepwise fashion
like regression.
There are two main uses of logistic regression:

• The first is the prediction of group membership. Since logistic regression calculates the
probability of success over the probability of failure, the results of the analysis are in
the form of an odds ratio.
• Logistic regression also provides knowledge of the relationships and strengths among
the variables (e.g. marrying the boss’s daughter puts you at a higher probability for job
promotion than undertaking five hours unpaid overtime each week).

Assumptions of logistic regression

• Logistic regression does not assume a linear relationship between the dependent and
independent variables.
• The dependent variable must be a dichotomy (2 categories).
• The independent variables need not be interval, nor normally distributed, nor linearly
related, nor of equal variance within each group.
• The categories (groups) must be mutually exclusive and exhaustive; a case can only be
in one group and every case must be a member of one of the groups.
570 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

• Larger samples are needed than for linear regression because maximum likelihood
coefficients are large sample estimates. A minimum of 50 cases per predictor is
recommended.

A non-mathematical illustration of logistic regression

The dependent variable we are attempting to predict is whether an unsolicited email

providing an emailed offer is opened or not by the recipient. There are two independent
variables:

• Presence or absence of recipient’s first name in the subject line.

• Presence or absence of the offer in the subject line.

Thus there are four possible combinations of interest:

• Presence of first name, presence of offer in subject line.

• Absence of first name, presence of offer in subject line.
• Presence of first name, absence of offer in subject line.
• Absence of first name, absence of offer in subject line

For each of the options, an email is sent to a random sample of 5,000 recipients to ensure a
low margin of error and high level of confidence. After a minimum of eight hours, results
of whether the email was opened are analyzed and a logistical regression equation is gener-
ated for each of the options. Each equation predicts the impact of the independent variables
on encouraging the consumer to take the desired action: opening and reading the email.
You are able to determine which combination of the independent variables leads consumers
to open the email most frequently. This analysis could also be run against individual customer
segments. Some customer segments may respond to a combination of independent variables
to which other customer segments are not responsive. For example, you might find that the
independent variable of first name present only had a positive impact on recipients who
were 39 and younger; the independent variable of first name present with no offer had
no impact on recipients 40 and older. Then you know which variables to use or not when
sending emails to recipients of different age groups.

The logistic regression equation

While logistic regression gives each predictor (IV) a coefficient ‘b’ which measures its
independent contribution to variations in the dependent variable, the dependent variable
can only take on one of the two values: 0 or 1. What we want to predict from a knowledge
of relevant independent variables and coefficients is therefore not a numerical value of a
dependent variable as in linear regression, but rather the probability (p) that it is 1 rather
than 0 (belonging to one group rather than the other).
But even to use probability as the dependent variable is unsound, mainly because numerical
predictors may be unlimited in range. If we expressed p as a linear function of investment,
 LOGISTIC REGRESSION 571

we might then find ourselves predicting that p is greater than 1 (which cannot be true, as
probabilities can only take values between 0 and 1). Additionally, because logistic regression
has only two y values – in the category or not in the category – a straight line best fit (as in
linear regression) is not possible to draw. Consider the following hypothetical example:
200 accountancy first year students are graded on a pass-fail dichotomy on the end of the
semester accountancy exam. At the start of the course, they all took a maths pre-test with
results reported in interval data ranging from 0–50 – the higher the pretest score the more
competency in maths. Logistic regression is applied to determine the relationship between
maths pretest score (IV or predictor) and whether a student passed the course (DV). Students
who passed the accountancy course are coded 1 while those who failed are coded 0.
We can see from Figure 24.1 of the plotted ‘x’s’ that there is somewhat greater likelihood
that those who obtained above average to high score on the maths test passed the accoun-
tancy course, while below average to low scorers tended to fail. There is also an overlap in
the middle area. But if we tried to draw a straight (best fitting) line, as with linear regression,
it just would not work, as intersections of the maths results and pass/fail accountancy results
form two lines of x’s, as in Figure 24.1.
The solution is to convert or transform these results into probabilities. We might compute
the average of the Y values at each point on the X axis. We could then plot the probabilities
of Y at each value of X and it would look something like the wavy graph line superimposed
on the original data in Figure 24.2. This is a smoother curve, and it is easy to see that the
probability of passing the accountancy course (Y axis) increases as values of X increase.
What we have just done is transform the scores so that the curve now fits a cumulative
probability curve, i.e. adding each new probability to the existing total. As you can see, this
curve is not a straight line; it is more of an s-shaped curve. Predicted values are interpreted as
probabilities and are now not just two conditions with a value of either 0 or 1 but continuous
data that can take any value from 0 to 1.
The slope of the curve in Figure 24.2 is low at the lower and upper extremes of the
independent variable and greatest in the middle where it is most sensitive. In the middle, of
course, are a number of cases that are out of order, in the sense that there is an overlap with
average maths scores in both accountancy pass and fail categories, while at the extremes are

Pass 1 xx x x xx xxxxxxxxxxxxxx

Fail 0 xxxxxxxxxxxx x xx x x xx x x

Independent Variable Maths Score (X)

Figure 24.1 Maths results of 200 accountancy students plotted against the
pass-fail categories.
572 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

1 xx x x xx xxxxxxxxxxxxxx

0 xxxxxxxxxxxx x xx x x xx x x
Independent Variable Maths Score

Figure 24.2 Maths results plotted against probability of allocation to pass/fail

categories.

cases which are almost universally allocated to their correct group. The outcome is not a predic-
tion of a Y value, as in linear regression, but a probability of belonging to one of two conditions
of Y, which can take on any value between 0 and 1 rather than just 0 and 1 in Figure 24.1.
Unfortunately a further mathematical transformation – a log transformation – is needed
to normalize the distribution. You met this in Chapter 8, where log transformations and
sq. root transformations moved skewed distributions closer to normality. So what we are
about to do is not uncommon. This log transformation of the p values to a log distribution
enables us to create a link with the normal regression equation. The log distribution
(or logistic transformation of p) is also called the logit of p or logit(p).
Logit(p) is the log (to base e) of the odds ratio or likelihood ratio that the dependent
variable is 1. In symbols it is defined as:

logit (p) = log[p / (1 − p)] = ln [p / (1 − p)]

Whereas p can only range from 0 to 1, logit(p) scale ranges from negative infinity to positive
infinity and is symmetrical around the logit of .5 (which is zero). Formula 24.1 below
shows the relationship between the usual regression equation (a + bx … etc.), which is a
straight line formula, and the logistic regression equation.
The form of the logistic regression equation is:

⎡ p(x) ⎤
logit [p(x)] = log ⎢ ⎥ = a + b1x1 + b2 x 2 + b3 x3 …
⎢ 1 − p(x) ⎥

This looks just like a linear regression and although logistic regression finds a ‘best
fitting’ equation, just as linear regression does, the principles on which it does so are
rather different. Instead of using a least-squared deviations criterion for the best fit, it
uses a maximum likelihood method, which maximizes the probability of getting the
observed results given the fitted regression coefficients. A consequence of this is that the
goodness of fit and overall significance statistics used in logistic regression are different
 LOGISTIC REGRESSION 573

from those used in linear regression. p can be calculated with the following formula
(formula 24.2) which is simply another rearrangement of formula 24.1:

(a + b1 x1 + b2 x 2 + b3 x3 …)
exp
p= (a + b1 x1 + b2 x 2 + b3 x3 …)
1 + exp formula 24.2

Where:

p = the probability that a case is in a particular category,

exp = the base of natural logarithms (approx 2.72),
a = the constant of the equation and,
b = the coefficient of the predictor variables.

Formula 24.2 involves another mathematic function, exp, the exponential function. ln, the
natural logarithm, and exp are opposites. The exponential function is a constant with the value
of 2.71828182845904 (roughly 2.72). When we take the exponential function of a number, we
take 2.72 raised to the power of the number. So, exp(3) equals 2.72 cubed or (2.72)3 = 20.09.
The natural logarithm is the opposite of the exp function. If we take ln(20.09), we get the
number 3. These are common mathematical functions on many calculators.
Don’t worry, you will not have to calculate any of this mathematical material by hand.
I have simply been trying to show you how we get from a regression formula for a line to
the logistic analysis.

Logistic regression – involves fitting an equation of the form to the data:

logit(p) = a + b1x1 + b2x2 + b3x3 + …

Interpreting log odds and the odds ratio

What has been presented above as mathematical background will be quite difficult for
some of you to understand, so we will now revert to a far simpler approach to consider how
SPSS deals with these mathematical complications.
The Logits (log odds) are the b coefficients (the slope values) of the regression equation.
The slope can be interpreted as the change in the average value of Y, from one unit of
change in X. In SPSS the b coefficients are located in column ‘B’ in the ‘Variables in the
Equation’ table. Logistic regression calculates changes in the log odds of the dependent,
not changes in the dependent value as OLS regression does. For a dichotomous variable the
odds of membership of the target group are equal to the probability of membership in the
target group divided by the probability of membership in the other group. Odds value can
range from 0 to infinity and tell you how much more likely it is that an observation is a
member of the target group rather than a member of the other group. If the probability is
0.80, the odds are 4 to 1 or .80/.20; if the probability is 0.25, the odds are .33 (.25/.75).
If the probability of membership in the target group is .50, the odds are 1 to 1 (.50/.50), as
in coin tossing when both outcomes are equally likely.
574 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

Another important concept is the odds ratio (OR), which estimates the change in the
odds of membership in the target group for a one unit increase in the predictor. It is calcu-
lated by using the regression coefficient of the predictor as the exponent or exp. Assume in
the example earlier where we were predicting accountancy success by a maths competency
predictor that b = 2.69. Thus the odds ratio is exp2.69 or 14.73. Therefore the odds of passing
are 14.73 times greater for a student, for example, who had a pre-test score of 5, than for a
student whose pre-test score was 4.
SPSS actually calculates this value of the ln(odds ratio) for us and presents it as EXP(B)
in the results printout in the ‘Variables in the Equation’ table. This eases our calculations
of changes in the DV due to changes in the IV. So the standard way of interpreting a ‘b’ in
logistic regression is using the conversion of it to an odds ratio using the corresponding
exp(b) value. As an example, if the logit b = 1.5 in the B column of the ‘Variables in the
Equation’ table, then the corresponding odds ratio (column exp(B)) quoted in the SPSS
table will be 4.48. We can then say that when the independent variable increases one unit,
the odds that the case can be predicted increase by a factor of around 4.5 times, when other
variables are controlled. As another example, if income is a continuous explanatory variable
measured in ten thousands of dollars, with a ‘b’ value of 1.5 in a model predicting home
ownership = 1, no home ownership = 0. Then since exp(1.5) = 4.48, a 1 unit increase in income
(one $10,000 unit) increases the odds of home ownership about 4.5 times. This process will
become clearer through following through the SPSS logistic regression activity below.

Model fit and the likelihood function

Just as in linear regression, we are trying to find a best fitting line of sorts but, because the
values of Y can only range between 0 and 1, we cannot use the least squares approach. The
Maximum Likelihood (or ML) is used instead to find the function that will maximize our
ability to predict the probability of Y based on what we know about X. In other words, ML
finds the best values for formula 24.1 above.
Likelihood just means probability. It always means probability under a specified hypothesis.
In logistic regression, two hypotheses are of interest:

• the null hypothesis, which is when all the coefficients in the regression equation take
the value zero, and
• the alternate hypothesis that the model with predictors currently under consideration is
accurate and differs significantly from the null of zero, i.e. gives significantly better
than the chance or random prediction level of the null hypothesis.

We then work out the likelihood of observing the data we actually did observe under each
of these hypotheses. The result is usually a very small number, and to make it easier to
handle, the natural logarithm is used, producing a log likelihood (LL). Probabilities are
always less than one, so LL’s are always negative. Log likelihood is the basis for tests of a
logistic model.
The likelihood ratio test is based on –2LL ratio. It is a test of the significance of the
difference between the likelihood ratio (–2LL) for the researcher’s model with
predictors (called model chi square) minus the likelihood ratio for baseline model with
 LOGISTIC REGRESSION 575

only a constant in it. Significance at the .05 level or lower means the researcher’s model
with the predictors is significantly different from the one with the constant only (all ‘b’
coefficients being zero). It measures the improvement in fit that the explanatory variables
make compared to the null model. Chi square is used to assess significance of this ratio.

The likelihood ratio test. This tests the difference between –2LL for the full
model with predictors and –2LL for initial chi-square in the null model.

When probability fails to reach the 5% significance level, we retain the null hypothesis
that knowing the independent variables (predictors) has no increased effects (i.e. make no
difference) in predicting the dependent.
We will now have a look at the concepts and indices introduced above by running an
SPSS logistic regression and see how it all works in practice.

SPSS activity – a logistic regression analysis

Access SPSS Ch 24 Data File A. This file contains data from a survey of home owners
conducted by an electricity company about an offer of roof solar panels with a 50% subsidy
from the state government as part of the state’s environmental policy. The variables involve
household income measured in units of a thousand dollars, age, monthly mortgage, size of
family household, and whether the householder would take or decline the offer. Please
follow the instructions below and conduct a logistic regression to determine whether family
size and monthly mortgage will predict taking or declining the offer.

1 Click Analyze >> Regression >> Binary Logistic.

2 Select the grouping variable (the variable to be predicted) which must be a dichoto-
mous measure and place it into the Dependent box (Fig. 24.3). For this example it is
‘take solar panel offer’. The convention for binomial logistic regression is to code the
dependent class of greatest interest as 1 and the other class as 0, because the coding will
affect the odds ratios and slope estimates.
3 Enter your predictors (IV’s) into the Covariates box. These are ‘family size’ and
‘mortgage’. Do not alter the default method of Enter unless you wish to run a Stepwise
logistic regression. Forward selection is the usual option for a stepwise regression,
starting with the constant-only model and adding variables one at a time. The forward
stepwise logistic regression method utilizes the likelihood ratio test (chi square differ-
ence) which tests the change in –2LL between steps to determine automatically which
variables to add or drop from the model. This is considered useful only for exploratory
purposes. Selecting model variables on a theoretic basis and using the Enter method is
preferred.
4 Should you have any categorical predictor variables, click on ‘Categorical’ button and
enter it (there is none in this example). The define categorical variables box is
displayed as in Figure 24.4. SPSS asks what coding methods you would like for the
576 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

Figure 24.3 Logistic regression dialogue box.

Figure 24.4 Categorical variables box.

 LOGISTIC REGRESSION 577

predictor variable (under the Categorical button), and there are several options to
choose from. For most situations, choose the ‘indicator’ coding scheme (it is the
default). You can choose to have the first or last category of the variable as your
baseline reference category. Usually, the absence of the factor is coded as 0, and the
presence of the factor is coded 1. If so, you want to make sure that the first category
(the one of the lowest value) is designated as the reference category in the categorical
dialogue box. SPSS will convert categorical variables to dummy variables automati-
cally. The class of greatest interest should be the last class (1 in a dichotomous variable
for example).
5 Click on Options button and select Classification Plots, Hosmer-Lemeshow Goodness
Of Fit, Casewise Listing Of Residuals and select Outliers Outside 2sd. Retain default
entries for probability of stepwise, classification cutoff and maximum iterations
(Fig. 24.5).
6 Continue then OK.

Figure 24.5 Options dialogue box.

Interpretation of printout Table 24.1

The first few tables in your printout are not of major importance. The first one to take note
of is the Classification table (Table 24.1) in Block 0 Beginning Block.

1 Block 0: Beginning Block. Block 0 presents the results with only the constant included
before any coefficients (i.e. those relating to family size and mortgage) are entered into
the equation. Logistic regression compares this model with a model including all the
predictors (family size and mortgage) to determine whether the latter model is more
appropriate. The table suggests that if we knew nothing about our variables and guessed
578 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

Table 24.1 The classification table

Classification Tablea,b
Predicted
Take solar panel offer
Percentage
Observed Decline offer Take offer correct

Step 0 take solar panel decline offer 0 14 .0

offer take offer 0 16 100.0
Overall Percentage 53.3
a
Constant is included in the model.
b
The cut value is .500.

Table 24.2 Variables in the equation table

Variables in the Equation

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

Step 0 Constant .134 .366 .133 1 .715 1.143

Table 24.3 Variables not in the equation table

Variables not in the Equation

Score df Sig.

Step 0 Variables Famsize 14.632 1 .000

Mortgage 6.520 1 .011
Overall Statistics 15.085 2 .001

that a person would not take the offer we would be correct 53.3% of the time. The
variables not in the equation table tells us whether each IV improves the model
(Table 24.3). The answer is yes for both variables, with family size slightly better than
mortgage size, as both are significant and if included would add to the predictive power
of the model. If they had not been significant and able to contribute to the prediction,
then termination of the analysis would obviously occur at this point.
2 Block 1 Method = Enter. This presents the results when the predictors ‘family size’ and
‘mortgage’ are included. Later SPSS prints a classification table which shows how the
classification error rate has changed from the original 53.3%. By adding the variables
we can now predict with 90% accuracy (see Classification Table 24.8 below). The
model appears good, but we need to evaluate model fit and significance as well. SPSS will
offer you a variety of statistical tests for model fit and whether each of the independent
variables included make a significant contribution to the model.
 LOGISTIC REGRESSION 579

Table 24.4 Omnibus tests of model coefficients

Omnibus Tests of Model Coefficients

Chi-square df Sig.

Step 1 Step 24.096 2 .000

Block 24.096 2 .000
Model 24.096 2 .000

3 Model chi-square. The overall significance is tested using what SPSS calls the Model
Chi square, which is derived from the likelihood of observing the actual data under the
assumption that the model that has been fitted is accurate. There are two hypotheses to
test in relation to the overall fit of the model:
H0 The model is a good fitting model.
H1 The model is not a good fitting model (i.e. the predictors have a significant effect).
The difference between –2LL for the best-fitting model and –2LL for the null hypoth-
esis model (in which all the b values are set to zero in block 0) is distributed like
chi squared, with degrees of freedom equal to the number of predictors; this difference
is the Model chi square that SPSS refers to. Very conveniently, the difference between –
2LL values for models with successive terms added also has a chi squared distribution,
so when we use a stepwise procedure, we can use chi-squared tests to find out if adding
one or more extra predictors significantly improves the fit of our model. The –2LL
value from the Model Summary table below is 17.359.
In our case model chi square has 2 degrees of freedom, a value of 24.096 and a prob-
ability of p < 0.000 (Table 24.4). Thus, the indication is that the model has a poor fit,
with the model containing only the constant indicating that the predictors do have a
significant effect and create essentially a different model. So we need to look closely at
the predictors and from later tables determine if one or both are significant predictors.
This table has 1 step. This is because we are entering both variables and at the same
time providing only one model to compare with the constant model. In stepwise logis-
tic regression there are a number of steps listed in the table as each variable is added or
removed, creating different models. The step is a measure of the improvement in the
predictive power of the model since the previous step.
4 Model Summary. Although there is no close analogous statistic in logistic regression to
the coefficient of determination R2 the Model Summary Table 24.5 provides some
approximations. Cox and Snell’s R-Square attempts to imitate multiple R-Square based

Table 24.5 Model Summary

Model Summary
Step −2 Log likelihood Cox & Snell R square Nagelkerke R square
a
1 17.359 .552 .737
a
Estimation terminated at iteration number 8 because parameter estimates changed by less than .001.
580 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

on ‘likelihood’, but its maximum can be (and usually is) less than 1.0, making it difficult
to interpret. Here it is indicating that 55.2% of the variation in the DV is explained by the
logistic model. The Nagelkerke modification that does range from 0 to 1 is a more reliable
measure of the relationship. Nagelkerke’s R2 will normally be higher than the Cox and
Snell measure. Nagelkerke’s R2 is part of SPSS output in the ‘Model Summary’ table and
is the most-reported of the R-squared estimates. In our case it is 0.737, indicating a
moderately strong relationship of 73.7% between the predictors and the prediction.
5 H-L Statistic. An alternative to model chi square is the Hosmer and Lemeshow test
which divides subjects into 10 ordered groups of subjects and then compares the number
actually in the each group (observed) to the number predicted by the logistic regression
model (predicted) (Table 24.7). The 10 ordered groups are created based on their esti-
mated probability; those with estimated probability below .1 form one group, and so on,
up to those with probability .9 to 1.0. Each of these categories is further divided into two
groups based on the actual observed outcome variable (success, failure). The expected
frequencies for each of the cells are obtained from the model. A probability (p) value is
computed from the chi-square distribution with 8 degrees of freedom to test the fit of the
logistic model. If the H-L goodness-of-fit test statistic is greater than .05, as we want for
well-fitting models, we fail to reject the null hypothesis that there is no difference
between observed and model-predicted values, implying that the model’s estimates fit
the data at an acceptable level. That is, well-fitting models show non-significance on the
H-L goodness-of-fit test. This desirable outcome of non-significance indicates that the
model prediction does not significantly differ from the observed.
The H-L statistic assumes sampling adequacy, with a rule of thumb being enough
cases so that 95% of cells (typically, 10 decile groups times 2 outcome categories = 20
cells) have an expected frequency > 5. Our H-L statistic has a significance of .605
which means that it is not statistically significant and therefore our model is quite a
good fit (Table 24.6)
6 Classification Table. Rather than using a goodness-of-fit statistic, we often want to look
at the proportion of cases we have managed to classify correctly. For this we need to
look at the classification table printed out by SPSS, which tells us how many of the
cases where the observed values of the dependent variable were 1 or 0 respectively have
been correctly predicted. In the Classification table (Table 24.8), the columns are the
two predicted values of the dependent, while the rows are the two observed (actual)
values of the dependent. In a perfect model, all cases will be on the diagonal and the
overall percent correct will be 100%. In this study, 87.5% were correctly classified for
the take offer group and 92.9% for the decline offer group. Overall 90% were correctly
classified. This is a considerable improvement on the 53.3% correct classification with the
constant model so we know that the model with predictors is a significantly better mode.

Table 24.6 Hosmer and Lemeshow test

Hosmer and Lemeshow Test

Step Chi-square df Sig.
1 6.378 8 .605
 LOGISTIC REGRESSION 581

Table 24.7 Contingency table for Hosmer and Lemeshow test

Contingency Table for Hosmer and Lemeshow Test

Take solar panel offer = Take solar panel offer =

decline offer take offer
Observed Expected Observed Expected Total

Step 1 1 3 2.996 0 .004 3

2 3 2.957 0 .043 3
3 3 2.629 0 .371 3
4 2 2.136 1 .864 3
5 2 1.881 1 1.119 3
6 0 .833 3 2.167 3
7 0 .376 3 2.624 3
8 1 .171 2 2.829 3
9 0 .019 3 2.981 3
10 0 .000 3 3.000 3

But are both predictor variables responsible or just one of them? This is answered by
the Variables in the Equation table.

Table 24.8 Classification table

Classification Tablea

Predicted

Take solar panel offer

Observed Decline offer Take offer Percentage correct

Step 1 take solar panel decline offer 13 1 92.9

offer take offer 2 14 87.5
Overall Percentage 90.0
a
The cut value is .500.

7 Variables in the Equation. The Variables in the Equation table (Table 24.9) has several
important elements. The Wald statistic and associated probabilities provide an index of
the significance of each predictor in the equation. The Wald statistic has a chi-square
distribution.
The simplest way to assess Wald is to take the significance values and if less
than .05 reject the null hypothesis as the variable does make a significant contribution.
In this case, we note that family size contributed significantly to the prediction
(p = .013) but mortgage did not (p = .075). The researcher may well want to drop
independents from the model when their effect is not significant by the Wald statistic
(in this case mortgage).
582 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

Table 24.9 Variables in the equation

Variables in the Equation

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

Step 1a Famsize 2.399 .962 6.215 1 .013 11.007

Mortgage .005 .003 3.176 1 .075 1.005
Constant −18.627 8.654 4.633 1 .031 .000
a
Variable(s) entered on step 1: Famsize, Mortgage.

The Exp(B) column in Table 24.9 presents the extent to which raising the corre-
sponding measure by one unit influences the odds ratio. We can interpret EXP(B) in
terms of the change in odds. If the value exceeds 1 then the odds of an outcome occur-
ring increase; if the figure is less than 1, any increase in the predictor leads to a drop in
the odds of the outcome occurring. For example, the EXP(B) value associated with
family size is 11.007. Hence when family size is raised by one unit (one person) the
odds ratio is 11 times as large and therefore householders are 11 more times likely to
belong to the take offer group.
The ‘B’ values are the logistic coefficients that can be used to create a predictive
equation (similar to the b values in linear regression) formula 24.2, page 573. In this
example:

e{( 2.399 × family size ) + (.0005 × mortgage ) − 18.627}

Probability of a case =
1 + e{( 2.399 × family size ) + (.005 × mortgage ) − 18.627}

Here is an example of the use of the predictive equation for a new case. Imagine a
householder whose household size including themselves was seven and paying
a monthly mortgage of $2,500. Would they take up the offer, i.e. belong to category 1?
Substituting in we get:

e{(2.399 × 7) + (.005 × 2500) – 18.627}

Probability of a case taking offer =
1 + e{(2.399 × 7) + (.005 × 2500) – 18.627}
e10.66
=
1 + e10.66
= 0.99

Therefore, the probability that a householder with seven in the household and a mort-
gage of $2,500 p.m. will take up the offer is 99%, or 99% of such individuals will be
expected to take up the offer.
Note that, given the non-significance of the mortgage variable, you could be justified
in leaving it out of the equation. As you can imagine, multiplying a mortgage value by
B adds a negligible amount to the prediction as its B value is so small (.005).
 LOGISTIC REGRESSION 583

Effect size. The odds ratio is a measure of effect size. The ratio of odds ratios of the
independents is the ratio of relative importance of the independent variables in terms of
effect on the dependent variable’s odds. In our example family size is 11 times as
important as monthly mortgage in determining the decision.

8 Classification Plot. The classification plot or histogram of predicted probabilities

provides a visual demonstration of the correct and incorrect predictions (Table 24.10).
Also called the ‘classplot’ or the ‘plot of observed groups and predicted probabilities’,
it is another very useful piece of information from the SPSS output when one chooses
‘Classification plots’ under the Options button in the Logistic Regression dialogue
box. The X axis is the predicted probability from .0 to 1.0 of the dependent being
classified ‘1’. The Y axis is frequency: the number of cases classified. Inside the plot
are columns of observed 1’s and 0’s (or equivalent symbols). The resulting plot is very
useful for spotting possible outliers. It will also tell you whether it might be better to
separate the two predicted categories by some rule other than the simple one SPSS
uses, which is to predict value 1 if logit(p) is greater than 0 (i.e. if p is greater than .5).
A better separation of categories might result from using a different criterion. We
might also want to use a different criterion if the a priori probabilities of the two
categories were very different (one might be winning the national lottery, for example),
or if the costs of mistakenly predicting someone into the two categories differ
(suppose the categories were ‘found guilty of fraudulent share dealing’ and ‘not
guilty’, for example).

Look for two things:

(1) A U-shaped rather than normal distribution is desirable. A U-shaped distribution
indicates the predictions are well-differentiated with cases clustered at each end
showing correct classification. A normal distribution indicates too many predictions
close to the cut point, with a consequence of increased misclassification around the
cut point which is not a good model fit. For these around .50 you could just as well
toss a coin.
(2) There should be few errors. The ‘t’s’ to the left are false positives. The ‘d’s’ to the right
are false negatives. Examining this plot will also tell such things as how well the model
classifies difficult cases (ones near p = .5).

9 Casewise List. Finally, the casewise list produces a list of cases that didn’t fit the model
well. These are outliers. If there are a number of cases this may reveal the need for fur-
ther explanatory variables to be added to the model. Only one case (No. 21) falls into this
category in our example (Table 24.11) and therefore the model is reasonably sound. This
is the only person who did not fit the general pattern. We do not expect to obtain a perfect
match between observation and prediction across a large number of cases.
No excessive outliers should be retained as they can affect results significantly. The
researcher should inspect standardized residuals for outliers (ZResid in Table 24.11) and
consider removing them if they exceed > 2.58 (outliers at the .01 level). Standardized
residuals are requested under the ‘Save’ button in the binomial logistic regression dialog
box in SPSS. For multinomial logistic regression, checking ‘Cell Probabilities’ under the
‘Statistics’ button will generate actual, observed, and residual values.
584 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

Table 24.10 Classification plot

Step number: 1

Observed Groups and Predicted Probabilities

F
R 6 t
E t
Q d t
U d t
E 4 d t
N d t
C d t t
Y d t t
2 dd dt t t t
dd dt t t t
dd d d dd d tt t ttd t t
dd d d dd d tt t ttd t t
Predicte
Prob : 0 .25 .5 .75 1
Group: ddddddddddddddddddddddddddddddtttttttttttttttttttttttttttttt
Predicted Probability is of Membership for take offer
The Cut Value is .50
Symbols: d - decline offer
t - take offer
Each Symbol Represents .5 Cases

Table 24.11 Casewise list

Casewise Listb

Observed Temporary variable

Take solar
Case Selected statusa panel offer Predicted Predicted group Resid ZResid
21 S d** .924 t −.924 −3.483
a
S = Selected, U = Unselected cases, and = Misclassified cases.
**

b
Cases with studentized residuals greater than 2.000 are listed.
 LOGISTIC REGRESSION 585

How to report your results

You could model a write up like this:
‘A logistic regression analysis was conducted to predict take-up of a solar panel subsidy
offer for 30 householders using family size and monthly mortgage payment as predictors.
A test of the full model against a constant only model was statistically significant, indicat-
ing that the predictors as a set reliably distinguished between acceptors and decliners of
the offer (chi square = 24.096, p < .000 with df = 2).
Nagelkerke’s R2 of .737 indicated a moderately strong relationship between prediction
and grouping. Prediction success overall was 90% (92.9% for decline and 87.5% for
accept. The Wald criterion demonstrated that only family size made a significant contribu-
tion to prediction (p = .013). Monthly mortgage was not a significant predictor. EXP(B)
value indicates that when family size is raised by one unit (one person) the odds ratio is 11
times as large and therefore householders are 11 more times likely to take the offer’.

To relieve your stress

Some parts of this chapter may have seemed a bit daunting. But remember, SPSS does all
the calculations. Just try and grasp the main principles of what logistic regression is all
about. Essentially, it enables you to:

• see how well you can classify people/events into groups from a knowledge of inde-
pendent variables; this is addressed by the classification table and the goodness-of-fit
statistics discussed above;
• see whether the independent variables as a whole significantly affect the dependent
variable; this is addressed by the Model Chi-square statistic.
• determine which particular independent variables have significant effects on the
dependent variable; this can be done using the significance levels of the Wald statistics,
or by comparing the –2LL values for models with and without the variables concerned
in a stepwise format.

SPSS Activity. Now access SPSS Chapter 24 Data File A on the website, and
conduct your own logistic regression analysis using age and family size as
predictors for taking or declining the offer. Write out an explanation of the
results and discuss in class.

What you have learned from this chapter

Binomial (or binary) logistic regression is a form of regression which is used when the dependent
is a dichotomy and the independents are of any type. The goal is to find the best set of
coefficients so that cases that belong to a particular category will, when using the equation,

(Continued)
586 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

have a very high calculated probability that they will be allocated to that category. This enables
new cases to be classified with a reasonably high degree of accuracy as well.
Logistic regression uses binomial probability theory, does not assume linearity of relationship
between the independent variables and the dependent, does not require normally distributed
variables, and in general has no stringent requirements.
Logistic regression has many analogies to linear regression: logit coefficients correspond to
b coefficients in the logistic regression equation, the standardized logit coefficients correspond
to beta weights, and the Wald statistic, a pseudo R2 statistic, is available to summarize the
strength of the relationship. The success of the logistic regression can be assessed by looking
at the classification table, showing correct and incorrect classifications of the dependent. Also,
goodness-of-fit tests such as model chi- square are available as indicators of model appropri-
ateness, as is the Wald statistic to test the significance of individual independent variables. The
EXP(B) value indicates the increase in odds from a one unit increase in the selected variable.

Review questions

Qu. 24.1
Why are p values transformed to a log value in logistic regression?

(a) because p values are extremely small

(b) because p values cannot be analyzed
(c) because p values only range between 0 and 1
(d) because p values are not normally distributed
(e) none of the above

Qu. 24.2
The Wald statistic is:

(a) a beta weight

(b) equivalent to R2
(c) a measure of goodness of fit
(d) a measure of the increase in odds

Qu. 24.3
Logistic regression is based on:

(a) normal distribution

(b) Poisson distribution
(c) the sine curve
(d) binomial distribution
 LOGISTIC REGRESSION 587

Qu. 24.4
Logistic regression is essential where,

(a) both the dependent variable and independent variable(s) are interval
(b) the independent variable is interval and both the dependent variables are categorical
(c) the sole dependent variable is categorical and the independent variable is not interval
(d) there is only one dependent variable irrespective of the number or type of the
independent variable(s)

Qu. 24.5
exp(B) in the SPSS printout tells us

(a) the probability of obtaining the dependent variable value

(b) the significance of beta
(c) for each unit change in X what the change factor is for Y
(d) the significance of the Wald statistic

Qu. 24.6
Explain briefly why a line of best fit approach cannot be applied in logistic regression.
Check your response in the material above.

Qu. 24.7
Under what circumstances would you choose to use logistic regression?
Check your response with the material above.

Qu. 24.8
What is the probability that a householder with only two in the household and a monthly
mortgage of $1,700 will take up the offer?
Check your answer on the web site for this chapter.

Now access the Web page for Chapter 24 and check your answers to the above
questions. You should also attempt the SPSS activity.

Additional reading
Agresti, A. 1996. An Introduction to Categorical Data Analysis. John Wiley and Sons.
Fox, J. 2000. Multiple and Generalized Nonparametric Regression. Thousand
Oaks, CA: Sage Publications. Quantitative applications in the social sciences Series
588 EXTENSION CHAPTERS ON ADVANCED TECHNIQUES

No. 131. Covers non-parametric regression models for GLM techniques like logistic
regression.
Menard, S. 2002. Applied Logistic Regression Analysis (2nd edn). Thousand Oaks, CA:
Sage Publications. Series: Quantitative Applications in the Social Sciences,
No. 106 (1st edn), 1995.
O’Connell, A. A. 2005. Logistic Regression Models for Ordinal Response Variables.
Thousand Oaks, CA: Sage Publications. Quantitative applications in the social sciences,
Volume 146.
Pampel, F. C. 2000. Logistic Regression: A Primer. Sage quantitative applications in the
Social Sciences Series #132. Thousand Oaks, CA: Sage Publications.

Useful website
Alan Agresti’s website, with all the data from the worked examples in his book: http://lib.
stat.cmu.edu/datasets/agrest.