1

Multivariate Logistic Regression

As in univariate logistic regression, let π(x) represent the probability of an event that
depends on p covariates or independent variables. Then, using an inv.logit formulation
for modeling the probability, we have:

eβ0 +β1 X1 +β2 X2 +...+βp Xp

π(x) =
1 + eβ0 +β1 X1 +β2 X2 +...+βp Xp

So, the form is identical to univariate logistic regression, but now with more than one
covariate. [Note: by “univariate” logistic regression, I mean logistic regression with
one independent variable; really there are two variables involved, the independent
variable and the dichotomous outcome, so it could also be termed bivariate.]

To obtain the corresponding logit function from this, we calculate (letting X represent
the whole set of covariates X1 , X2 , . . . , Xp ):

" #
π(X)
logit[π(X)] = ln
1 − π(X)
eβ0 +β1 X1 +β2 X2 +...+βp Xp
 
1+eβ0 +β1 X1 +β2 X2 +...+βp Xp
= ln  eβ0 +β1 X1 +β2 X2 +...+βp Xp

1 − 1+e β0 +β1 X1 +β2 X2 +...+βp Xp

eβ0 +β1 X1 +β2 X2 +...+βp Xp

 
β0 +β1 X1 +β2 X2 +...+βp Xp
1+e
= ln  1

1+eβ0 +β1 X1 +β2 X2 +...+βp Xp
h i
= ln eβ0 +β1 X1 +β2 X2 +...+βp Xp

= β0 + β1 X1 + β2 X2 + . . . + βp Xp

So, again, we see that the logit of the probability of an event given X is a simple
linear function.
 2

To summarize, the two basic equations of multivariate logistic regression are:

eβ0 +β1 X1 +β2 X2 +...+βp Xp

π(X) =
1 + eβ0 +β1 X1 +β2 X2 +...+βp Xp

which gives the probabilities of outcome events given the covariate values X1 , X2 , . . . , Xp ,
and

logit[π(X)] = β0 + β1 X1 + β2 X2 + . . . + βp Xp

which shows that logistic regression is really just a standard linear regression model,
once we transform the dichotomous outcome by the logit transform. This transform
changes the range of π(X) from 0 to 1 to −∞ to +∞, as usual for linear regression.

Again analogously to univariate logistic regression, the above equations are for mean
probabilities, and each data point will have an error term. Once again, we assume
that this error has mean zero, and that it follows a binomial distribution with mean
π(X), and variance π(X)(1 − π(X)). Of course, now X is a vector, whereas before it
was a scalar value.

Interpretation of the β coefficients in multiple logis-

tic regression
Interpretation of the intercept, β0 : Notice that regardless of the number of co-
variate values, if they are all set to zero, then we have

eβ0
π(x) =
1 + eβ0

exactly the same as in the univariate case. So, the interpretation of β0 remains
the same as in the simpler case: β0 sets the “baseline” event rate, through the
above function, when all covariate values are set equal to zero.
For example, if β0 = 0 then

eβ0 e0 1
π(x) = β
= 0
= = 0.5
1+e 0 1+e 1+1

and if β0 = 1 then

eβ0 e1
π(x) = = = 0.73
1 + eβ0 1 + e1
 3

and if β0 = −1 then

eβ0 e−1
π(x) = = = 0.27
1 + eβ0 1 + e−1

and so on.
As before, positive values of β0 give values greater than 0.5, while negative
values of β0 give probabilities less than 0.5, when all covariates are set to zero.

Interpretation of the slopes, β1 , β2 , . . . , βp : Recall the effect on the proba-

bility of an event as X changes by one unit in the univariate case. There, we
saw that the coefficient β1 is such that eβ1 is the odds ratio for a unit change in
z
zβ1 β1
X, and in general, for a change of z units, the OR = e = e .
Nothing much changes for the multivariate case, except:

• When there is more than one independent variable, if all variables are
completely uncorrelated with each other, then the interpretations of all
coefficients are simple, and follow the above pattern:
We have OR = ezβi for any variable Xi , i = 1, 2, . . . , p, where the OR
represents the odds ratio for a change of size z for that variable.
• When the variables are not uncorrelated, the interpretation is more diffi-
cult. It is common to say that OR = ezβi represents the odds ratio for
a change of size z for that variable adjusted for the effects of the other
variables. While this is essentially correct, we must keep in mind that
confounding and collinearity can change and obscure these estimated rela-
tionships. The way confounding operates is identical to what we saw for
linear regression.

Estimating the β coefficients given a data set

As in the univariate case, the distribution associated with logistic regression is the
binomial. For a single subject with covariate values xi = {x1i , x2i , . . . , xpi }, the like-
lihood function is:

i i
π(xi )y (1 − π(xi ))1−y

For n subjects, the likelihood function is:

n
i i
π(xi )y (1 − π(xi ))1−y
Y

i=1
 4

To derive estimates of the unknown β parameters, as in the univariate case, we need

to maximize this likelihood function. We follow the usual steps, including taking the
logarithm of the likelihood function, taking (p + 1) partial derivatives with respect
to each β parameter and setting these (p + 1) equations equal to zero, to form a set
of (p + 1) equations in (p + 1) unknowns. Solving this system of equations gives the
maximum likelihood equations.

We again omit the details here (as in the univariate case, no easy closed form formulae
exists), and will rely on statistical software to find the maximum likelihood estimates
for us.

Inferences typically rely on SE formulae for confidence intervals, and likelihood ratio
testing for hypothesis tests. Again, we will omit the details, and rely on statistical
software.

We next look at several examples.

Multiple Logistic Regression Examples

We will look at three examples:

• Logistic regression with dummy or indicator variables

• Logistic regression with many variables

• Logistic regression with interaction terms

In all cases, we will follow a similar procedure to that followed for multiple linear
regression:

1. Look at various descriptive statistics to get a feel for the data. For logistic
regression, this usually includes looking at descriptive statistics, for example
within “outcome = yes = 1” versus “outcome = no = 0” subgroups.

2. The above “by outcome group” descriptive statistics are often sufficient for
discrete covariates, but you may want to prepare some graphics for continuous
variables. Recall that we did this for the age variable when looking at the CHD
example.

3. For all continuous variables being considered, calculate a correlation matrix of

each variable against each other variable. This allows one to begin to investigate
possible confounding and collinearity.
 5

4. Similarly, for each categorical/continous independent variable pair, look at the

values for the continuous variable in each category of the other variable.

5. Finally, create tables for all categorical/categorical independent variable pairs.

6. Perform a separate univariate logistic regression for each independent variable.

This begins to investigate confounding (we will see in more detail next class),
as well as providing an initial “unadjusted” view of the importance of each
variable, by itself.

7. Think about any “interaction terms” that you may want to try in the model.

8. Perform some sort of model selection technique, or, often much better, think
about avoiding any strict model selection by finding a set of models that seem
to have something to contribute to overall conclusions.

9. Based on all work done, draw some inferences and conclusions. Carefully inter-
pret each estimated parameter, perform “model criticism”, possibly repeating
some of the above steps (for example, run further models), as needed.

10. Other inferences, such as predictions for future observations, and so on.

As with linear regression, the above should not be considered as “rules”, but rather
as a rough guide as to how to proceed through a logistic regression analysis.

Logistic regression with dummy or indicator variables

Chapter 1 (section 1.6.1) of the Hosmer and Lemeshow book described a data set
called ICU. Deleting the ID variable, there are 20 variables in this data set, which we
describe in the table below:
 6

Description Coding variable name

Vital Status 0 = Lived STA
(Main outcome) 1 = Died
Age Years AGE
Sex 0 = Male SEX
1 = Female
Race 1 = White RACE
2 = Black
3 = Other
Service at ICU Admission 0 = Medical SER
1 = Surgical
Cancer Part of Present 0 = No CAN
Problem 1 = Yes
History of Chronic Renal O = No CRN
Failure 1 = Yes
Infection Probable at ICU 0 = No INF
Admission 1 = Yes
CPR Prior to ICU Admission 0 = No CPR
1 = Yes
Systolic Blood Pressure at mm Hg SYS
ICU Admission
Heart Rate at ICU Admission Beats/min HRA
Previous Admission to an ICU 0 = No PRE
within 6 Months 1 = Yes
Type of Admission 0 = Elective TYP
1 = Emergency
Long Bone, Multiple, Neck, 0 = No FRA
Single Area, or Hip Fracture 1 = Yes
PO2 from Initial Blood Gases 0 > 60 PO2
1 ≤ 60
PH from Initial Blood Gases 0 ≥ 7.25 PH
1 < 7.25
PCO2 from initial Blood 0 ≤ 45 PCO
Gases 1 > 45
Bicarbonate from Initial 0 ≥ 18 BIC
Blood Gases 1 < 18
Creatinine from Initial Blood 0 ≤ 2.0 CRE
Gases 1 > 2.0
Level of Consciousness at ICU O = No Coma LOC
Admission or Stupor
1 = Deep
stupor
2 = Coma
 7

The main outcome is vital status, alive or dead, coded as 0/1 respectively, under the
variable name sta. For this illustrative example, we will investigate the effect of the
dichotomous variables sex, ser, and loc. Later, we will look at more of the variables.

# read the data into R

> icu.dat <- read.table(file="g:\\icudat.txt", header = T)

# Take a quick look at the data

> summary(icu.dat)

sta age sex race ser

Min. :0.0 Min. :16.00 Min. :0.00 Min. :1.000 Min. :0.000
1st Qu.:0.0 1st Qu.:46.75 1st Qu.:0.00 1st Qu.:1.000 1st Qu.:0.000
Median :0.0 Median :63.00 Median :0.00 Median :1.000 Median :1.000
Mean :0.2 Mean :57.55 Mean :0.38 Mean :1.175 Mean :0.535
3rd Qu.:0.0 3rd Qu.:72.00 3rd Qu.:1.00 3rd Qu.:1.000 3rd Qu.:1.000
Max. :1.0 Max. :92.00 Max. :1.00 Max. :3.000 Max. :1.000
can crn inf cpr sys
Min. :0.0 Min. :0.000 Min. :0.00 Min. :0.000 Min. : 36.0
1st Qu.:0.0 1st Qu.:0.000 1st Qu.:0.00 1st Qu.:0.000 1st Qu.:110.0
Median :0.0 Median :0.000 Median :0.00 Median :0.000 Median :130.0
Mean :0.1 Mean :0.095 Mean :0.42 Mean :0.065 Mean :132.3
3rd Qu.:0.0 3rd Qu.:0.000 3rd Qu.:1.00 3rd Qu.:0.000 3rd Qu.:150.0
Max. :1.0 Max. :1.000 Max. :1.00 Max. :1.000 Max. :256.0
hra pre typ fra po2
Min. : 39.00 Min. :0.00 Min. :0.000 Min. :0.000 Min. :0.00
1st Qu.: 80.00 1st Qu.:0.00 1st Qu.:0.000 1st Qu.:0.000 1st Qu.:0.00
Median : 96.00 Median :0.00 Median :1.000 Median :0.000 Median :0.00
Mean : 98.92 Mean :0.15 Mean :0.735 Mean :0.075 Mean :0.08
3rd Qu.:118.25 3rd Qu.:0.00 3rd Qu.:1.000 3rd Qu.:0.000 3rd Qu.:0.00
Max. :192.00 Max. :1.00 Max. :1.000 Max. :1.000 Max. :1.00
ph pco bic cre loc
Min. :0.000 Min. :0.0 Min. :0.000 Min. :0.00 Min. :0.000
1st Qu.:0.000 1st Qu.:0.0 1st Qu.:0.000 1st Qu.:0.00 1st Qu.:0.000
Median :0.000 Median :0.0 Median :0.000 Median :0.00 Median :0.000
Mean :0.065 Mean :0.1 Mean :0.075 Mean :0.05 Mean :0.125
3rd Qu.:0.000 3rd Qu.:0.0 3rd Qu.:0.000 3rd Qu.:0.00 3rd Qu.:0.000
Max. :1.000 Max. :1.0 Max. :1.000 Max. :1.00 Max. :2.000

# Create the subset of variables we need

> icu1.dat <- data.frame(sta=icu.dat$sta, loc=icu.dat$loc,

 8

sex=icu.dat$sex, ser=icu.dat$ser)

# Look at reduced data set

> summary(icu1.dat)
sta loc sex ser
Min. :0.0 Min. :0.000 Min. :0.00 Min. :0.000
1st Qu.:0.0 1st Qu.:0.000 1st Qu.:0.00 1st Qu.:0.000
Median :0.0 Median :0.000 Median :0.00 Median :1.000
Mean :0.2 Mean :0.125 Mean :0.38 Mean :0.535
3rd Qu.:0.0 3rd Qu.:0.000 3rd Qu.:1.00 3rd Qu.:1.000
Max. :1.0 Max. :2.000 Max. :1.00 Max. :1.000

# Notice that loc, sex, and ser need to be made into factor variables

icu1.dat <- data.frame(sta=icu.dat$sta, loc=as.factor(icu.dat$loc),

sex=as.factor(icu.dat$sex), ser=as.factor(icu.dat$ser))

# Look at reduced data set again, this time with factor variables

> summary(icu1.dat)
sta loc sex ser
Min. :0.0 0:185 0:124 0: 93
1st Qu.:0.0 1: 5 1: 76 1:107
Median :0.0 2: 10
Mean :0.2
3rd Qu.:0.0
Max. :1.0

# Preliminary comments:
- Not too many events, only 20% rate
- loc may not be too useful, poor variability
- sex and ser reasonably well balanced

# Create two by two tables of all variables, V1 = side, V2 = top

> table(icu1.dat$sta, icu1.dat$sex)

0 1
0 100 60
1 24 16

# Not much difference observed

# (death rates: M 24/124= 0.19 ~ F 16/76 = 0.21)
 9

> table(icu1.dat$sta, icu1.dat$ser)

0 1
0 67 93
1 26 14

# Fewer deaths (sta=1) at the surgical unit (ser=1),

# OR = 67*14/(26*93) = 0.39

> table(icu1.dat$sta, icu1.dat$loc)

0 1 2
0 158 0 2
1 27 5 8

# Probably very low accuracy here,

# but many more deaths in cats 1 and 2.

> table(icu1.dat$sex, icu1.dat$ser)

0 1
0 54 70
1 39 37

# A bit, but not too much potential for confounding here,

# especially since effect of sex is not strong to begin with

> table(icu1.dat$sex, icu1.dat$loc)

0 1 2
0 116 3 5
1 69 2 5

# Too few data points to say much, maybe females

# have higher values of loc

> table(icu1.dat$ser, icu1.dat$loc)

0 1 2
0 84 2 7
1 101 3 3

# Again hard to say much

# Overall, loc does not look too useful, not much
# effect from any of these variables except maybe for ser.
 10

# Simple logistic regression for each variable:

> output <- glm(sta ~ sex, data=icu1.dat, family=binomial)

> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ sex, family = binomial, data = icu1.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-0.6876 -0.6876 -0.6559 -0.6559 1.8123

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.4271 0.2273 -6.278 3.42e-10 ***
sex1 0.1054 0.3617 0.291 0.771
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 200.08 on 198 degrees of freedom
AIC: 204.08

Number of Fisher Scoring iterations: 4

$intercept.ci
[1] -1.8726220 -0.9816107

$slopes.ci
[1] -0.6035757 0.8142967

$OR
sex1
1.111111

$OR.ci
[1] 0.5468528 2.2575874

# Very wide CI, result is inconclusive for sex covariate,

# as both endpoints of potential clinical interest.
 11

> output <- glm(sta ~ ser, data=icu1.dat, family=binomial)

> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ ser, family = binomial, data = icu1.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-0.8098 -0.8098 -0.5296 -0.5296 2.0168

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.9466 0.2311 -4.097 4.19e-05 ***
ser1 -0.9469 0.3682 -2.572 0.0101 *
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 193.24 on 198 degrees of freedom
AIC: 197.24

Number of Fisher Scoring iterations: 4

$intercept.ci
[1] -1.3994574 -0.4937348

$slopes.ci
[1] -1.6685958 -0.2252964

$OR
ser1
0.3879239

$OR.ci
[1] 0.1885116 0.7982796

# OR < 1, being in surgical unit may save lives,

# CI shows at least a 20% effect (approximately).
 12

> output <- glm(sta ~ loc, data=icu1.dat, family=binomial)

> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ loc, family = binomial, data = icu1.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.7941 -0.5617 -0.5617 -0.5617 1.9619

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.7668 0.2082 -8.484 < 2e-16 ***
loc1 18.3328 1073.1090 0.017 0.986370
loc2 3.1531 0.8175 3.857 0.000115 ***
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 163.78 on 197 degrees of freedom
AIC: 169.78

Number of Fisher Scoring iterations: 15

$intercept.ci
[1] -2.174912 -1.358605

$slopes.ci
[,1] [,2]
[1,] -2084.922247 2121.587900
[2,] 1.550710 4.755395

$OR
loc1 loc2
9.158944e+07 2.340741e+01

$OR.ci
[,1] [,2]
[1,] 0.000000 Inf
[2,] 4.714817 116.2095
 13

# As we suspected, too few data points to analyze this variable

# results are completely non-informative

# Drop loc from further analyses

# Multivariate logistic model for sex and ser

> output <- glm(sta ~ sex + ser, data=icu1.dat, family=binomial)

> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ sex + ser, family = binomial, data = icu1.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-0.8168 -0.8048 -0.5266 -0.5266 2.0221

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.96129 0.27885 -3.447 0.000566 ***
sex1 0.03488 0.36896 0.095 0.924688
ser1 -0.94442 0.36915 -2.558 0.010516 *
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 193.23 on 197 degrees of freedom
AIC: 199.23

Number of Fisher Scoring iterations: 4

$intercept.ci
[1] -1.5078281 -0.4147469

$slopes.ci
[,1] [,2]
[1,] -0.6882692 0.758025
[2,] -1.6679299 -0.220904

$OR
sex1 ser1
 14

1.0354933 0.3889063

$OR.ci
[,1] [,2]
[1,] 0.5024449 2.1340574
[2,] 0.1886372 0.8017936

# Almost identical results as for univariate model

# As we suspected, no confounding.

# Overall conclusion: Service at ICU Admission (ser)

# seems to be associated with a strong effect, the
# surgical unit being associated with fewer deaths
# (OR = 0.39, 95% CI = (.19, 0.80). Nothing much
# can be said about sex or loc, results are inconclusive.

# Can predict some death rates depending on sex/ser categories:

# Create a data set for all possible category combinations:

> newdata <- data.frame(sex=as.factor(c(0,0,1,1)), ser=as.factor(c(0,1,0,1)))

> newdata
sex ser
1 0 0
2 0 1
3 1 0
4 1 1

# Predict outcome probabilities for these four categories:

> predict(output, newdata=newdata)

1 2 3 4
-0.9612875 -1.9057045 -0.9264096 -1.8708266

# Above are on the logit scale, more useful to do this:

> predict(output, newdata=newdata, type="response")

1 2 3 4
0.2766205 0.1294642 0.2836537 0.1334461

# Above are predictions on probability scale.

# Can see poor results when ser = 0, two sexes are similar in rate.
 15

Logistic regression example with many variables

We will continue with the same example, but now consider many more variables,
including both categorical and continuous variables.

Very shortly, we will see an excellent way to simultaneously select a model and to
investigate confounding in data sets with a large number of variables. For now, we
will take a quick look at logistic regression using four variables from the ICU data
set: Age, sex, ser, and typ.

# Read in full data set

icu.dat <- read.table(file="g:\\icudat.txt", header = T)

# Create the reduced data set we will use:

> icu2.dat <- data.frame(sta=icu.dat$sta, sex=as.factor(icu.dat$sex),

ser=as.factor(icu.dat$ser), age = icu.dat$age, typ=as.factor(icu.dat$typ))

# Check descriptive statistics

> summary(icu2.dat)
sta sex ser age typ
Min. :0.0 0:124 0: 93 Min. :16.00 0: 53
1st Qu.:0.0 1: 76 1:107 1st Qu.:46.75 1:147
Median :0.0 Median :63.00
Mean :0.2 Mean :57.55
3rd Qu.:0.0 3rd Qu.:72.00
Max. :1.0 Max. :92.00

# Quick look at correlations

# Check effect of age and typ, two variables we did

# not look at before

> table(icu2.dat$sta, icu2.dat$typ)

0 1
0 51 109
1 2 38

# Looks like a large effect: those presenting at

# emergency (typ=1) have a much higher death rate.
 16

# Let’s also look at a table between ser and typ:

> table(icu2.dat$ser, icu2.dat$typ)

0 1
0 1 92
1 52 55

# Looks like there could be some confounding here,

# these are strongly related.

# Check the association between age and the outcome, sta

> boxplot(list(icu2.dat$age[icu2.dat$sta==0], icu2.dat$age[icu2.dat$sta==1]))

It looks like those with higher ages also have higher death rates.

Let’s look at a regression with all five variables included:

 17

> output <- glm(sta ~ sex + ser + age + typ, data=icu2.dat, family=binomial)
> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ sex + ser + age + typ, family = binomial,
data = icu2.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.2753 -0.7844 -0.3920 -0.2281 2.5072

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -5.26359 1.11678 -4.713 2.44e-06 ***
sex1 -0.20092 0.39228 -0.512 0.60851
ser1 -0.23891 0.41697 -0.573 0.56667
age 0.03473 0.01098 3.162 0.00156 **
typ1 2.33065 0.80238 2.905 0.00368 **
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 172.42 on 195 degrees of freedom
AIC: 182.42

Number of Fisher Scoring iterations: 6

$intercept.ci
[1] -7.452432 -3.074752

$slopes.ci
[,1] [,2]
[1,] -0.96976797 0.56792495
[2,] -1.05615818 0.57834395
[3,] 0.01320442 0.05624833
[4,] 0.75801170 3.90328595

$OR
sex1 ser1 age typ1
0.8179766 0.7874880 1.0353364 10.2846123
 18

$OR.ci
[,1] [,2]
[1,] 0.3791710 1.764602
[2,] 0.3477894 1.783083
[3,] 1.0132920 1.057860
[4,] 2.1340289 49.565050

As expected, age has a strong effect, with an odds ratio of 1.035 per year, or 1.03510 =
1.41 per decade (95% CI per year of (1.013, 1.058), so (1.138, 1.757) per decade). Typ
also has a very strong effect, with a CI of at least 2.

There does indeed seem to be some confounding between ser and typ, as the coefficient
estimate for ser has changed drastically from when typ was not in the model. In fact,
ser no longer looks “important”, it has been “replaced” by typ. Because of the high
correlation between ser and typ, it is difficult to separate out the effects of these two
variables.

We will return to this issue when we discuss model selection for logistic regression.

Logistic regression with interaction terms

Going back to the example where we had just sex and ser in the model, what if we
wanted to investigate an interaction term between these two variables?

So far, we have seen that ser is associated with a strong effect, but the effect of sex
was inconclusive. But, what if the effect of ser is different among males and females,
i.e., what if we have an interaction (or sometimes called effect modification) between
sex and ser?

Here is how to investigate this for logistic regression in R:

# Create the variable that will be used in the interaction:

# Create a blank vector to store new variable

> ser.sex <- rep(0, length(icu1.dat$ser))

# Change value to 1 when both ser and sex are one

> for (i in 1:length(icu1.dat$ser)) {if (icu1.dat$ser[i] == 1

& icu1.dat$sex[i] == 1) ser.sex[i] <- 1 }

# Check new variable

 19

> ser.sex
[1] 0 0 0 0 1 0 0 0 0 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 1 1 1 1
[36] 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 1 0 0 1 0 1 0
[71] 0 0 0 1 0 0 0 1 0 1 0 0 0 1 0 0 0 0 0 0 1 0 0 1 1 0 0 0 1 0 0 0 0 0 0
[106] 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1
[141] 0 0 0 1 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 0
[176] 0 0 0 1 0 0 0 1 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0

# Add new variable to data.frame,

# since all 0/1, do not bother with factors here

> icu3.dat <- data.frame(sta = icu1.dat$sta,

ser = icu1.dat$ser, sex = icu1.dat$sex, ser.sex = ser.sex)

> summary(icu3.dat)
sta ser sex ser.sex
Min. :0.0 Min. :0.000 Min. :0.00 Min. :0.000
1st Qu.:0.0 1st Qu.:0.000 1st Qu.:0.00 1st Qu.:0.000
Median :0.0 Median :1.000 Median :0.00 Median :0.000
Mean :0.2 Mean :0.535 Mean :0.38 Mean :0.185
3rd Qu.:0.0 3rd Qu.:1.000 3rd Qu.:1.00 3rd Qu.:0.000
Max. :1.0 Max. :1.000 Max. :1.00 Max. :1.000

> Estimate the logistic regression model with interaction:

> output <- glm(sta ~ sex + ser + ser.sex, data=icu3.dat, family=binomial)

> logistic.regression.or.ci(output)
$regression.table

Call:
glm(formula = sta ~ sex + ser + ser.sex, family = binomial, data = icu3.dat)

Deviance Residuals:
Min 1Q Median 3Q Max
-0.8696 -0.7244 -0.4590 -0.4590 2.1460

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.7777 0.2930 -2.654 0.00795 **
sex -0.4263 0.4799 -0.888 0.37440
ser -1.4195 0.4945 -2.870 0.00410 **
ser.sex 1.1682 0.7518 1.554 0.12021
---
 20

Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 190.81 on 196 degrees of freedom
AIC: 198.81

Number of Fisher Scoring iterations: 4

$intercept.ci
[1] -1.3519790 -0.2034301

$slopes.ci
[,1] [,2]
[1,] -1.366836 0.5142995
[2,] -2.388770 -0.4502697
[3,] -0.305277 2.6416881

$OR
sex ser ser.sex
0.6529412 0.2418301 3.2162162

$OR.ci
[,1] [,2]
[1,] 0.25491223 1.6724666
[2,] 0.09174244 0.6374562
[3,] 0.73691921 14.0368799

# Looking at the confidence interval, the result is inconclusive,

# but we can certainly not rule out an interaction term, OR = 3.22.

# Possibly some confounding as well, so cannot conclude any strong result.

 21

Note on ORs in presence of interactions

Note that one needs to be careful in interpreting the Odds Ratios from the above
outputs, because of the interaction term.

The OR given above for the sex variable, 0.653, applies only within the medical service
(coded as 0 in the ser variable), and the OR given above for ser, 0.242, applies only
within males (coded 0 in the sex variable).

In order to obtain the OR for sex within the ser = 1 (surgical category), or to obtain
the OR for ser within females (sex = 1), one needs to multiply by the OR from the
interaction term. Hence, we have:

OR for ser within females = 0.242 * 3.22 = 0.779.

OR for sex within surgical unit = 0.653 * 3.22 = 2.10.