Logistic Regression: Interaction Terms

1
 Interactions in Logistic Regression

I For linear regression, with predictors X1 and X2 we saw

that an interaction model is a model where the
interpretation of the effect of X1 depends on the value of
X2 and vice versa.

I Exactly the same is true for logistic regression.

I The simplest interaction models includes a predictor

variable formed by multiplying two ordinary predictors:

logit(P(Y = 1)) = β0 + β1 × X1 + β2 × X2 + β3 × X1 × X2

I Interaction term

2
 Interactions in Logistic Regression

We will look at the interpretation of interactions in 3 cases:

1 Interaction between two dummy variables.
2 Interaction between a dummy and a continuous variable.
3 Interaction between two continuous variables.

3
 Interaction Between 2 Dummy Variables

I Consider a logistic model for the risk of suffering a heart

attack over a year in terms gender and smoking status:

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I sex indicates gender (male=1, female=0)

I smoke indicates smoking status (smokes=1, does not=0).

4
 Interpreting the Intercept

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I In order to interpret β0 we need to find a situation in

which the final three terms in the equation vanish.

I This happens when an observation corresponds to a

female non-smoker, for then sex=0 and smoke=0.

logit P(Y = 1) = β0 + β1 × 0 + β2 × 0 + β3 (0 × 0)
= β0

I Consequently, β0 is the log odds in favour of a female

non-smoker suffering from a heart attack.

5
 Interpretations of Other Quantities Involving β0

We can also give interpretations on the odds scale and on the

probability scale:
I exp(β0 ) is the odds in favour of a female non-smoker
suffering from a heart attack.
exp(β0 )
I
1+exp(β0 )is the probability of a female non-smoker
suffering from a heart attack.

6
 Interpreting β1 and β2

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I We would know how to interpret β1 if the interaction

term was not there.

I Since in that case would just have an ordinary

multivariate logistic model.

I This happens when an observation corresponds to a

non-smoker, for then smoke=0.

logit P(Y = 1) = β0 + β1 × sex + β2 × 0 + β3 (sex × 0)

= β0 + β1 × sex

7
 Interpreting β1 and β2

I Amongst non-smokers

logit P(Y = 1) = β0 + β1 × sex

I We know how to interpret β1 in this case as its a

univariate logistic model.

I β1 is the log-odds ratio comparing males and females

amongst non-smokers.

I exp(β1 ) is the odds ratio comparing males and females

amongst non-smokers.

8
 Interpreting β1 and β2

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I To interpret β2 we need to get rid of the interaction term

without getting rid of the β2 smoke term.

I Same argument as before but now set sex=0 (female):

logit P(Y = 1) = β0 + β1 × 0 + β2 × smoke + β3 (0 × smoke)

= β0 + β2 × smoke

I β2 is the log-odds ratio comparing smokers with

non-smokers amongst females.

9
 Interpreting β3

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I To interpret β3 rewrite the regression equation:

logit P(Y = 1) = β0 + [β1 + β3 smoke]sex + β2 smoke

I This looks like a multivariate regression model with sex
and smoke as predictors where:
I β1 + β3 smoke is the log-odds ratio for males vs. females;
I β2 is the log odds ratio for smokers vs. non-smokers.

I β3 is the difference between the log-odds ratio comparing

males vs females in smokers and the log-odds ratio
comparing males vs. females in non-smokers.

10
 Interpreting β3

logit P(Y = 1) = β0 + β1 sex + β2 smoke + β3 (sex × smoke)

I We could just as well have rewritten the equation this way:

logit P(Y = 1) = β0 + β1 sex + [β2 + β3 sex]smoke

I β3 is the difference between the log-odds ratio comparing

smokers vs non-smokers in males and the log-odds ratio
comparing smokers vs. non-smokers in females.
I So we have two ways of thinking about β3 :
1 either as modification of the effect of smoke by sex
2 or the modification of the effect of sex by smoke.

11
 Quick Lookup Table

We can draw up a table for the 4 types of observation:

sex smoke logit(P(Y = 1))

1 Male Yes β0 + β1 + β2 + β3
2 Male No β0 + β1
3 Female Yes β0 + β2
4 Female No β0
I This allows us to find the function of the parameters
corresponding to a log-odds ratio and vice versa.
I e.g. 3 - 4 shows us that the log-odds ratio for smokers
vs. non-smokers amongst females is β2
I e.g. 1 - 2 shows us that the log-odds ratio for smokers
12 vs. non-smokers amongst males is β + β
 Interaction Between a Dummy Variable and a
Continuous Variable

I Consider a logistic model where the main predictors are

sex (a dummy coded as before) and age (in years)

logit P(Y = 1) = β0 + β1 sex + β2 age + β3 (sex × age)

I β0 is the log-odds in favour of a female age 0 suffering

from a heart attack.

13
 Interaction Between a Dummy Variable and a
Continuous Variable

I Consider a logistic model where the main predictors are

sex (a dummy coded as before) and age (in years)

logit P(Y = 1) = β0 + β1 sex + β2 age + β3 (sex × age)

I β1 is the log-odds ratio for males vs. females amongst

people of age 0.

14
 Interaction Between a Dummy Variable and a
Continuous Variable

I Consider a logistic model where the main predictors are

sex (a dummy coded as before) and age (in years)

logit P(Y = 1) = β0 + β1 sex + β2 age + β3 (sex × age)

I β2 is the log-odds ratio corresponding to an increase in

age by 1 year amongst females.

15
 Interaction Between a Dummy Variable and a
Continuous Variable

I Consider a logistic model where the main predictors are

sex (a dummy coded as before) and age (in years)

logit P(Y = 1) = β0 + β1 sex + β2 age + β3 (sex × age)

I β3 is the difference between the log-odds ratio

corresponding to a change in age by 1 year amongst males
and the the log-odds ratio corresponding to an increase in
age by 1 year amongst females.

I β3 is also difference between the log-odds ratios for males

vs. females in two age homogenous groups which differ
by 1 year.

16
 Quick Lookup Table

Again we can draw up a table, this time considering groups of

individuals aged z and z + 1

sex age logit(P(Y = 1))

1 Male z+1 β0 + β1 + β2 (z + 1) + β3 (z + 1)
2 Male z β0 + β1 + β2 z + β3 z
3 Female z+1 β0 + β2 (z + 1)
4 Female z β0 + β2 z

I e.g. 3 - 4 shows us that the log-odds ratio

corresponding to an increase in age by 1 year amongst
females is β2
I e.g. 2 - 4 shows us that the log-odds ratio for males vs.
females amongst people aged z is β1 + β3 z
17
 Interaction Between 2 Continuous Variables

I Consider a logistic model where the main predictors are

BP (blood pressure in mmHg) and age (in years)

logit P(Y = 1) = β0 + β1 BP + β2 age + β3 (BP × age)

I β0 is the log-odds in favour of a person with a BP of

0mmHg and age 0 suffering from a heart attack.

I Ridiculous interpretation (model can’t apply when age or

BP are close to 0, but we hope it is good for the ranges we
are interested in.)

18
 Interaction Between 2 Continuous Variables

I Consider a logistic model where the main predictors are

BP (blood pressure in mmHg) and age (in years)

logit P(Y = 1) = β0 + β1 BP + β2 age + β3 (BP × age)

I β1 is the log-odds ratio corresponding to an increase in BP

by 1mmHg amongst people aged 0.

19
 Interaction Between 2 Continuous Variables

I Consider a logistic model where the main predictors are

BP (blood pressure in mmHg) and age (in years)

logit P(Y = 1) = β0 + β1 BP + β2 age + β3 (BP × age)

I β2 is the log-odds ratio corresponding to an increase in

age by 1 year amongst people with a BP of 0mmHg.

20
 Interaction Between 2 Continuous Variables

I Consider a logistic model where the main predictors are

BP (blood pressure in mmHg) and age (in years)

logit P(Y = 1) = β0 + β1 BP + β2 age + β3 (BP × age)

I β3 is the difference between the log-odds ratios

corresponding to an increase in age of 1 year for two BP
homogenous groups which differ by 1 mmHg.

I β3 is also difference between the difference between the

log-odds ratios corresponding to an increase in BP of 1
mmHg for two age homogenous groups which differ by 1
year.

21
 Quick Lookup Table

Again we can draw up a table, this time considering

individuals with BP w and w + 1 and aged z and z + 1

BP age logit(P(Y = 1))

1 w+1 z+1 β0 + β1 (w + 1) + β2 (z + 1) + β3 (w + 1)(z + 1)
2 w+1 z β0 + β1 (w + 1) + β2 z + β3 (w + 1)z
3 w z+1 β0 + β1 w + β2 (z + 1) + β3 w(z + 1)
4 w z β0 + β1 w + β2 z + β3 wz

I e.g. 3 - 4 shows us that the log-odds ratio

corresponding to an increase in age by 1 year amongst
those of BP w is β2 + β3 w.
I e.g. 2 - 4 shows us that the log-odds ratio
22
 Final Comment on Interpretation

I Remember whenever you give an interpretation of a

quantity γ in terms of a log-odds ratio there is always an
equivalent interpretation of exp(γ) as an odds-ratio.
I Whenever you give an interpretation of a quantity γ as
the log-odds in favour of an event you can always give
two equivalent interpretations
1 of exp(γ) as the odds in favour of the event,
exp(γ)
2 of 1+exp(γ) as the probability of the event.

23