Lecture 13: Introduction to

Logistic Regression

Sandy Eckel
seckel@jhsph.edu

13 May 2008

1
Logistic Regression

Basic Idea:

Logistic regression is the type of
regression we use for a response variable
(Y) that follows a binomial distribution

Linear regression is the type of regression
we use for a continuous, normally
distributed response (Y) variable

Remember the Binomial Distribution?

2
Review of the Binomial Model

Y ~ Binomial(n,p)

n independent trials

(e.g., coin tosses)

p = probability of success on each trial

(e.g., p = ½ = Pr of heads)

Y = number of successes out of n trials

(e.g., Y= number of heads)

3
Binomial Distribution Example

Binomial probability density function (pdf):

n y
P(Y = y ) =   p (1 − p )
n− y

 y
Example:

4
Why can’t we use Linear Regression
to model binary responses?

The response (Y) is NOT normally distributed

The variability of Y is NOT constant

Variance of Y depends on the expected value of Y

For a Y~Binomial(n,p) we have Var(Y)=pq which
depends on the expected response, E(Y)=p

The model must produce predicted/fitted
probabilities that are between 0 and 1

Linear models produce fitted responses that vary
from -∞ to ∞

5
Binomial Y example

Consider a phase I clinical trial in which

35 independent patients are given a
new medication for pain relief. Of the 35
patients, 22 report “significant” relief
one hour after medication

Question: How effective is the drug?

6
Model

Y = # patients who get relief

n = 35 patients (trials)

p = probability of relief for any patient

The truth we seek in the population

How effective is the drug? What is p?

Want a method to

Get best estimate of p given data

Determine range of plausible values for p

7
How do we estimate p?
Maximum Likelihood Method
The method of maximum likelihood estimation chooses
values for parameter estimates which make the observed
data “maximally likely” under the specified model

Likelihood Function: Pr(22 of 35)

1.0e-10
Max Likelihood
Likelihood

5.0e-11

MLE: p=0.63

0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
p=Prob(Event) 8
Maximum Likelihood
Clinical trial example

Under the binomial model, ‘likelihood’ for observed Y=y
n y
P(Y = y ) =   p (1 − p )
n− y

 y

So for this example the likelihood function is:
 35  22
P(Y = y ) =   p (1 − p )
13

 22 

So, estimate p by choosing the value for p which makes
observed data “maximally likely”

i.e., choose p that makes the value of Pr (Y=22) maximal

The ML estimate of p is y/n
= 22/35
= 0.63
The estimated proportion of patients who will experience
relief is 0.63 9
Confidence Interval (CI) for p

Recall the general form of any CI:

Estimate ± (something near 2) x SE(estimate)

p (1 − p ) pq

Variance of p̂ : Var( p̂)= =
n n

“Standard Error” of p̂ : pq
n

Estimate of “Standard Error” of p̂ : pˆ qˆ
n
10
Confidence Interval for p

95% Confidence Interval for the ‘true’

proportion, p:

pˆ ± 1.96
pˆ qˆ
= 0.63 ± 1.96
(0.63)(0.37)
n 35
 LB: 0.63-1.96(.082)
UB: 0.63+1.96(.082)
=(0.47, 0.79)
11
Conclusion

Based upon our clinical trial in which 22 of 35

patients experience relief, we estimate that 63%
of persons who receive the new drug experience
relief within 1 hour (95% CI: 47% to 79%)

Whether 63% (47% to 79%) represents an
‘effective’ drug will depend many things,
especially on the science of the problem.

Sore throat pain?

Arthritis pain?

Childbirth pain?

12
Aside: Review of Probabilities and Odds

The odds of an event are defined as:

P(Y = 1) P(Y = 1)
odds(Y=1) = =
P(Y = 0) 1 - P(Y = 1)
=
p
1-p

We can go back and forth between odds and
probabilities:
p
Odds =
1-p
p = odds/(odds+1)
13
Aside: Review of Odds Ratio

We saw that an odds ratio (OR) can be

helpful for comparisons.

Recall the Vitamin A trial where we
looked at the odds ratio of death
comparing the vitamin A group to the
no vitamin A group:

odds(Death | Vit. A)

OR =
odds(Death | No Vit A.)
14
Aside: Review of Odds Ratio Interpretation

The OR here describes the benefits of

Vitamin A therapy. We saw for this
example that:

OR = 0.59

The Vitamin A group had 0.60 times the
odds of death of the no Vitamin A group; or

An estimated 40% reduction in mortality

OR is a building block for logistic
regression

15
Logistic Regression

Suppose we want to ask whether new

drug is better than a placebo and have
the following observed data:

Relief? Drug Placebo

No 13 20
Yes 22 15
Total 35 35
16
Confidence Intervals for p

Placebo
( )

( ) Drug

0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
p

17
Odds Ratio

odds(Relief | Drug)
OR =
odds(Relief | Placebo)
P(Relief | Drug) / [1 - P(Relief | Drug)]
=
P(Relief | Placebo) / [1 - P(Relief | Placebo)]
0.63/(1 - 0.63)
= = 2.26
0.45/(1 - 0.45)

18
Confidence Interval for OR

CI used Woolf’s method for the standard

error of log(Oˆ R ) (from lecture 6)

1 1 1 1

se(log(Oˆ R)) = + + + = 0.489
22 13 15 20

find log( ˆ
O R ) ± 1 . 96 se (log( ˆ
O R ))

Then (eL,eU)

19
Interpretation

OR = 2.26

95% CI: (0.86 , 5.90)

The Drug is an estimated 2 ¼ times

better than the placebo.

But could the difference be due to
chance alone?

YES ! 1 is a ‘plausible’ true population OR

20
Logistic Regression

Can we set up a model for this binomial

outcome similar to what we’ve done in
regression?

Idea: model the log odds of the event,

(in this example, relief) as a function of
predictor variables

21
A regression model for the log odds
 P(relief | Tx) 
log[ odds(Relief | Tx) ] = log  = β + β Tx
 0 1
 P(no relief | Tx) 
where: Tx = 0 if Placebo
1 if Drug

log( odds(Relief|Drug) ) = β0 + β1

log( odds(Relief|Placebo) ) = β0

log( odds(Relief|D)) – log( odds(Relief|P)) = β1

22
And…

Because of the basic property of logs:

log( odds(Relief|D)) – log( odds(Relief|P)) = β1

 odds(R | D) 
 log  =
 β1
 odds(R | P) 

And: OR = exp(β1) = eβ1 !!

So: exp(β1) = odds ratio of relief for patients

taking the Drug-vs-patients taking the Placebo.
23
Logistic Regression
Logit estimates Number of obs = 70
LR chi2(1) = 2.83
Prob > chi2 = 0.0926
Log likelihood = -46.99169 Pseudo R2 = 0.0292

------------------------------------------------------------------------------
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Tx | .8137752 .4889211 1.66 0.096 -.1444926 1.772043
(Intercept) | -.2876821 .341565 -0.84 0.400 -.9571372 .3817731
------------------------------------------------------------------------------

Estimates:

log( odds(relief|Tx) ) = βˆ0 + βˆ1Tx

= -0.288 + 0.814(Tx)

Therefore: OR = exp(0.814) = 2.26 !

So 2.26 is the odds ratio of relief for patients taking the Drug
compared to patients taking the Placebo 24
It’s the same as the OR we got before!

So, why go to all the trouble of setting up a

linear model?

What if there is a biologic reason to expect
that the rate of relief (and perhaps drug
efficacy) is age dependent?

What if
Pr(relief) = function of Drug or Placebo AND Age

We could easily include age in a model such
as:
log( odds(relief) ) = β0 + β1Drug + β2Age

25
Logistic Regression

As in MLR, we can include many

additional covariates

For a Logistic Regression model with
r number of predictors:
log ( odds(Y=1)) = β0 + β1X1 + ... + βrXr
Pr(Y = 1) Pr(Y = 1)
where: odds(Y=1) = =
1 − Pr(Y = 1) Pr(Y = 0)

26
Logistic Regression
Thus:
 Pr(Y = 1) 
log   = β0 + β1X1 + ... + βrXr
 Pr(Y = 0) 

But, why use log(odds)?

Linear regression might estimate anything

(-∞, +∞), not just a proportion in the range of
0 to 1

Logistic regression is a way to estimate a
proportion (between 0 and 1) as well as some
related items
27
Another way to motivate using log(OR) for the
lefthand side of logistic regression

We would like to use something like

what we know from linear regression:

Continuous outcome = β0 + β1X1 + β2X2+…

How can we turn a proportion into a

continuous outcome?

28
Transforming a proportion…

A proportion is a value between 0 and 1

The odds are always positive:
 p 
odds=   ⇒ [0,+∞)
 1− p 

The log odds is continuous:

 p 
Logodds= ln  ⇒ (−∞,+∞)
 1− p 

29
“Logit” transformation of the probability

Measure Min Max Name

Pr(Y = 1) 0 1 “probability”

Pr(Y = 1)
0 ∞ “odds”
1 − Pr(Y = 1)

 Pr(Y = 1) 
log  -∞ ∞ “log-odds” or “logit”
 1 − Pr(Y = 1) 

30
Logit Function

Relates log-odds (logit) to p = Pr(Y=1)

logit function
10

5
log-odds

-5

-10
0 .5 1
Probability of Success
31
Key Relationships

Relating log-odds, probabilities, and

parameters in logistic regression:

Suppose we have the model:
logit(p) = β0 + β1X
 p 
i.e. log   = β0 + β1X
 1-p 

Take “anti-logs” to get back to OR scale
 p 
  = exp(β0 + β1X)
 1-p 

32
Solve for p as a function of the coefficients

p/(1-p) = exp(β0 + β1X)

p = (1 – p)⋅exp(β0 + β1X)

p = exp(β0 + β1X) – p ⋅ exp(β0 + β1X)

p + p ⋅exp(β0 + β1X) = exp(β0 + β1X)

p ⋅{1+ exp(β0 + β1X)} = exp(β0 + β1X)
exp(β + β X )

p= 0 1
1 + exp(β + β X )
0 1
33
What’s the point of all that algebra?

Now we can determine the estimated

probability of success for a specific set
of covariates, X, after running a logistic
regression model

34
Example
Dependence of Blindness on Age

The following data concern the Aegean

island of Kalytos where inhabitants
suffer from a congenital eye disease
whose effects become more marked
with age.

Samples of 50 people were taken at five
different ages and the numbers of blind
people were counted

35
Example: Data

Age Number blind / 50

20 6 / 50

35 7 / 50

45 26 / 50

55 37 / 50

70 44 / 50

36
Question

The scientific question of interest is to

determine how the probability of
blindness is related to age in this
population

Let pi = Pr(a person in age classi is blind)

37
Model 1 – Intercept only model

logit(pi) = β0*

β0*= log-odds of blindness for all ages

exp(β0*) = odds of blindness for all ages

No age dependence in this model

38
Model 2 – Intercept and age

logit(pi) = β0 + β1(agei – 45)

β0 = log-odds of blindness among 45 year olds

exp(β0) = odds of blindness among 45 year olds

β1 = difference in log-odds of blindness

comparing a group that is one year older than
another

exp(β1) = odds ratio of blindness comparing a

group that is one year older than another
39
Results

Model 1: logit(pi) = β0*

Logit estimates Number of obs = 250
LR chi2(0) = 0.00
Prob > chi2 = .
Log likelihood = -173.08674 Pseudo R2 = 0.0000

------------------------------------------------------------------------------
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(Intercept) | -.0800427 .1265924 -0.63 0.527 -.3281593 .1680739
------------------------------------------------------------------------------

exp( −.08)

logit( p̂i) = -0.08 or pˆ i = = 0.48
1 + exp( −.08)

40
Results

Model 2: logit(pi) = β0 + β1(agei – 45)

Logit estimates Number of obs = 250
LR chi2(1) = 99.30
Prob > chi2 = 0.0000
Log likelihood = -123.43444 Pseudo R2 = 0.2869

------------------------------------------------------------------------------
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | .0940683 .0119755 7.86 0.000 .0705967 .1175399
(Intercept) | -4.356181 .5700966 -7.64 0.000 -5.473549 -3.238812
------------------------------------------------------------------------------

logit( p̂i) = -4.4 + .094(agei - 45)

or exp(− 4.4 + 0.094(age − 45)
pˆi = i
( (
1 + exp − 4.4 + 0.094 age − 45
i
)
41
Test of significance

Is the addition of the age variable in the

model important?

Maximum likelihood estimates:
β̂1 =0.094 s.e.(β̂1 )=0.012

z-test: H0: β1 = 0

z=7.855; p-val=0.000

95% C.I. (0.07, 0.12)

42
What about the Odds Ratio?

Maximum likelihood estimates:

OR = exp( β̂1)= exp(0.094)= 1.10

SE( β̂1 ) =SE(log(OR) ) = 0.013

Same z-test, reworded for OR scale:
Ho: exp(β1) = 1

z = 7.86 p-val = 0.000

95% C.I. for β1 (1.07, 1.13)
*(calculated on log scale, then exponentiated!!)
e(0.094 – 1.96*0.013), e(0.094 + 1.96*0.013)

It appears that blindness is age dependent

Note: exp(0) = 1, where is this fact useful?
43
Model 1 fit

Plot of observed proportion -vs-

predicted proportions using an intercept
only model

1 Observed
Prob Blindness

.5

Predicted

0
20 40 60 80
Age
44
Model 2 fit

Plot of observed proportion -vs-

predicted proportions with age in the
model
1
Observed
Prob Blindness

.5
Predicted

0
20 40 60 80
Age 45
Conclusion

Model 2 clearly fits better than Model 1!

Including age in our model is better

than intercept alone.

46
Lecture 13 Summary

Logistic regression gives us a framework in which

to model binary outcomes

Uses the structure of linear models, with
outcomes modelled as a function of covariates

As we’ll see, many concepts carry over from
linear regression

Interactions

Linear splines

Tests of significance for coefficients

All coefficients will have different
interpretations in logistic regression

Log odds or Log odds ratios!
47
HW 3 Hint

General logistic model specification:

Systematic:
logit( P(Yi = 1)) = log(odds(Yi = 1)) = β 0 + β1 x1 + β 2 x2

Random:
Yi ~ Binomial (n, p ) = Binomial (1, pi )

where pi depends on the covariates for person i

48