Categorical Response Variables

Examples:
ì Non - smoker
Whether or not a person Y =í
smokes Binary Response îSmoker
ìSurvives
Success of a medical Y =í
treatment îDies

Opinion poll responses ìAgree

ï
Y = í Neutral
Ordinal Response ïDisagree
î
 Example: Height predicts Gender
Y = Gender (0=Male 1=Female)
X = Height (inches)
Try an ordinary linear regression
> regmodel=lm(Gender~Hgt,data=Pulse)
> summary(regmodel)
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 7.343647 0.397563 18.47 <2e-16 ***
Hgt -0.100658 0.005817 -17.30 <2e-16 ***
 Gender

0.0 0.2 0.4 0.6 0.8 1.0

60
65

Hgt
70
75
Ordinary linear regression is used a lot, and is
taught in every intro stat class. Logistic regression
is rarely taught or even mentioned in intro stats,
but mostly because of inertia.
We now have the computing power and
software to implement logistic regression.
 π = Proportion of “Success”
In ordinary regression the model predicts the
mean Y for any combination of predictors.
What’s the “mean” of a 0/1 indicator variable?
å yi # of 1' s
y= = = Proportion of " success"
n # of trials
Goal of logistic regression: Predict the “true”
proportion of success, π, at any value of the
predictor.
 Binary Logistic Regression Model
Y = Binary response X = Quantitative predictor
π = proportion of 1’s (yes,success) at any X
Equivalent forms of the logistic regression model:
Logit form Probability form
æ p ö e
β0 + β1 X
logç ÷ = b 0 + b1 X π =
è1- p ø β0 + β1 X
1+ e
What does this look like?
N.B.: This is natural log (aka “ln”)
 Logit Function
no data Function Plot
1.0

0.8

0.6
y

0.4

0.2

-10 -8 -6 -4 -2 0 2 4 6 8 10 12
x
exp (bo + b1• x )
y=
1 + exp (bo + b1• x )
 Binary Logistic Regression via R
> logitmodel=glm(Gender~Hgt,family=binomial,
data=Pulse)
> summary(logitmodel)
Call:
glm(formula = Gender ~ Hgt, family = binomial)

Deviance Residuals:
Min 1Q Median 3Q Max
-2.77443 -0.34870 -0.05375 0.32973 2.37928

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 64.1416 8.3694 7.664 1.81e-14 ***
Hgt -0.9424 0.1227 -7.680 1.60e-14***
---
Call:
glm(formula = Gender ~ Hgt, family = binomial, data = Pulse)

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 64.1416 8.3694 7.664 1.81e-14 ***
Hgt -0.9424 0.1227 -7.680 1.60e-14***
---

64.14 −0.9424 Ht
e
πˆ = 64.14 −.9424 Ht
1+ e
proportion of females at that
Hgt
> plot(fitted(logitmodel)~Pulse$Hgt)
> with(Pulse,plot(Hgt,jitter(Gender,amount=0.05)))
> curve(exp(64.1-0.94*x)/(1+exp(64.1-0.94*x)), add=TRUE)

1.0
jitter(Gender, amount = 0.05)

0.8
0.6
0.4
0.2
0.0

60 65 70 75

Hgt
 Example: Golf Putts

Length 3 4 5 6 7
Made 84 88 61 61 44
Missed 17 31 47 64 90
Total 101 119 108 125 134

Build a model to predict the proportion of

putts made (success) based on length (in feet).
 Logistic Regression for Putting
Call:
glm(formula = Made ~ Length, family = binomial, data =
Putts1)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.8705 -1.1186 0.6181 1.0026 1.4882

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.25684 0.36893 8.828 <2e-16 ***
Length -0.56614 0.06747 -8.391 <2e-16 ***
---
 ⎛ pˆ ⎞
log⎜ ⎟ vs. Length
⎝ 1− p ⎠
ˆ

1.5
1.0
Linear part of
logistic fit
logitPropMade

0.5
0.0
-0.5

3 4 5 6 7

PuttLength
Probability Form of Putting Model

1.0
e3.257 -0.566 Length
pˆ =
1 + e3.257 -0.566 Length
0.8
Probability Made

0.6
0.4
0.2
0.0

2 4 6 8 10 12

PuttLength
 Odds
Definition:
p P(Yes)
= is the odds of Yes.
1 - p P( No)

p odds
odds = Ûp =
1- p 1 + odds
 π odds
odds = ⇔π=
1− π 1+ odds
Fair die
Event Prob Odds
even # 1/2 1 [or 1:1]
X>2 2/3 2 [or 2:1]
roll a 2 1/6 1/5 [or 1/5:1 or 1:5]
 1.0
e 0+1*x
π=
1 + e 0+1*x

0.8
π increases
0.6

by .231
Prob

x increases
0.4

by 1

x increases the odds increase by

0.2

by 1 a factor of 2.718

π increases by .072
0.0

-3 -2 -1 0 1 2 3

x
 Odds
Logit form of the model:
⎛ π ⎞
log ⎜ ⎟ = β0 + β1 X
⎝ 1− π ⎠

⇒ The logistic model assumes a linear

relationship between the predictors
and log(odds).
π β0 + β1 X
odds = =e
1− π
 Odds Ratio
A common way to compare two groups
is to look at the ratio of their odds
Odds1
Odds Ratio = OR =
Odds 2
Note: Odds ratio (OR) is similar to relative risk (RR).

p1 1− p 2
RR = OR = RR *
p2 1− p1
So when p is small, OR ≈ RR.
X is replaced by X + 1:
β0 + β1 X
odds = e
is replaced by
β0 + β1 ( X +1)
odds = e
So the ratio is
β0 + β1 ( X +1)
e β0 + β1 ( X +1)−( β0 + β1 X ) β1
β0 + β1 X
=e =e
e
 Example: TMS for Migraines
Transcranial Magnetic Stimulation vs. Placebo
Pain Free? TMS Placebo
YES 39 22
NO 61 78
Total 100 100

πˆ TMS = 0.39 odds = 39 / 100 = 39 = 0.639 πˆ = 0.639 = 0.39

TMS
61 / 100 61 1+ 0.639
pˆ Placebo = 0.22 22
odds Placebo = = 0.282
78

= 2.27 Odds are 2.27 times higher of getting

0.639
Odds ratio =
0.282 relief using TMS than placebo
 Logistic Regression for TMS data
> lmod=glm(cbind(Yes,No)~Group,family=binomial,data=TMS)
> summary(lmod)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.2657 0.2414 -5.243 1.58e-07 ***
GroupTMS 0.8184 0.3167 2.584 0.00977 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 6.8854 on 1 degrees of freedom

Residual deviance: 0.0000 on 0 degrees of freedom
AIC: 13.701

Note: e0.8184 = 2.27 = odds ratio

> datatable=rbind(c(39,22),c(61,78))
> datatable
[,1] [,2]
Chi-Square Test for
[1,]
[2,]
39
61
22
78
2-way table
> chisq.test(datatable,correct=FALSE)
Pearson's Chi-squared test

data: datatable
X-squared = 6.8168, df = 1, p-value = 0.00903

> lmod=glm(cbind(Yes,No)~Group,family=binomial,data=TMS)
> summary(lmod)

Call:
Binary Logistic Regression
glm(formula = cbind(Yes, No) ~ Group, family = binomial)Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.2657 0.2414 -5.243 1.58e-07 ***
GroupTMS 0.8184 0.3167 2.584 0.00977 **
 A Single Binary Predictor for a Binary Response

Response variable: Y = Success/Failure

Predictor variable: X = Group #1 / Group #2
• Method #1: Binary logistic regression
• Method #2: Z- test, compare two proportions
• Method #3: Chi-square test for 2-way table
All three “tests” are essentially equivalent, but the
logistic regression approach allows us to mix other
categorical and quantitative predictors in the model.
 Putting Data
Odds using data from 6 feet = 0.953
Odds using data from 5 feet = 1.298

è Odds ratio (6 ft to 5 ft) = 0.953/1.298 = 0.73

The odds of making a putt from 6 feet are

73% of the odds of making from 5 feet.
 Golf Putts Data
Length 3 4 5 6 7
Made 84 88 61 61 44
Missed 17 31 47 64 90
Total 101 119 108 125 134
pˆ .8317 .7394 .5648 .4880 .3284
Odds 4.941 2.839 1.298 0.953 0.489
.5648 61
E.g., 5 feet: Odds = = = 1.298
1− .5648 47
.4880 61
E.g., 6 feet: Odds = = = 0.953
1− .4880 64
 Golf Putts Data
Length 3 4 5 6 7
Made 84 88 61 61 44
Missed 17 31 47 64 90
Total 101 119 108 125 134
pˆ .8317 .7394 .5648 .4880 .3284
Odds 4.941 2.839 1.298 .953 .489

OR .575 .457 .734 .513

.8317 84
E.g., Odds = = = 4.941
1− .8317 17
 Interpreting “Slope” using Odds Ratio

æ p ö
logç ÷ = b 0 + b1 X
è1- p ø

⇒ odds = e b 0 + b1 X

When we increase X by 1, the ratio of the

b1
new odds to the old odds is e .
b1
i.e. odds are multiplied by e .
 Odds Ratios for Putts
From samples at each distance:
4 to 3 feet 5 to 4 feet 6 to 5 feet 7 to 6 feet
0.575 0.457 0.734 0.513
From fitted logistic:
4 to 3 feet 5 to 4 feet 6 to 5 feet 7 to 6 feet
0.568 0.568 0.568 0.568
In a logistic model, the odds ratio is constant when
changing the predictor by one.
 Example: 2012 vs 2014 congressional elections

How does %vote won by Obama relate to a

Democrat winning a House seat?

See the script elections 12, 14.R

Example: 2012 vs 2014 congressional elections

How does %vote won by Obama relate to a

Democrat winning a House seat?

In 2012 a Democrat had a decent chance even

if Obama got only 50% of the vote in the
district. In 2014 that was less true.
In 2012 a Democrat had a decent chance even if Obama
got only 50% of the vote in the district. In 2014 that was
less true.
 There is an easy way to graph logistic curves in R.
> library(TeachingDemos)
> with(elect, plot(Obama12,jitter(Dem12,amount=.05)))
> logitmod14=glm(Dem14~Obama12,family=binomial,data=elect)
> Predict.Plot(logitmod14, pred.var="Obama12”,add=TRUE,
plot.args = list(lwd=3,col="black"))
 R Logistic Output
> PuttModel=glm(Made~Length, family=binomial,data=Putts1)
> anova(PuttModel)
Analysis of Deviance Table
Df Deviance Resid. Df Resid. Dev
NULL 586 800.21
Length 1 80.317 585 719.89

> summary(PuttModel)
Call:
glm(formula = Made ~ Length, family = binomial)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.25684 0.36893 8.828 <2e-16 ***
Length -0.56614 0.06747 -8.391 <2e-16 ***
---
Null deviance: 800.21 on 586 degrees of freedom
Residual deviance: 719.89 on 585 degrees of freedom
 Two forms of logistic data
1. Response variable Y = Success/Failure or 1/0: “long
form” in which each case is a row in a spreadsheet
(e.g., Putts1 has 587 cases). This is often called
“binary response” or “Bernoulli” logistic regression.
2. Response variable Y = Number of Successes for a
group of data with a common X value: “short form”
(e.g., Putts2 has 5 cases – putts of 3 ft, 4 ft, … 7 ft).
This is often called “Binomial counts” logistic
regression.
> str(Putts1)
'data.frame': 587 obs. of 2 variables:
$ Length: int 3 3 3 3 3 3 3 3 3 3 ...
$ Made : int 1 1 1 1 1 1 1 1 1 1 ...
> longmodel=glm(Made~Length,family=binomial,data=Putts1)
> summary(longmodel)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.25684 0.36893 8.828 <2e-16 ***
Length -0.56614 0.06747 -8.391 <2e-16 ***
---
Null deviance: 800.21 on 586 degrees of freedom
Residual deviance: 719.89 on 585 degrees of freedom
> str(Putts2)
'data.frame': 5 obs. of 4 variables:
$ Length: int 3 4 5 6 7
$ Made : int 84 88 61 61 44
$ Missed: int 17 31 47 64 90
$ Trials: int 101 119 108 125 134
>
shortmodel=glm(cbind(Made,Missed)~Length,family=binomial,data=Putts2)
> summary(shortmodel)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.25684 0.36893 8.828 <2e-16 ***
Lengths -0.56614 0.06747 -8.391 <2e-16 ***
---
Null deviance: 81.3865 on 4 degrees of freedom
Residual deviance: 1.0692 on 3 degrees of freedom
 Binary Logistic Regression Model
Y = Binary X = Single predictor
response
π = proportion of 1’s (yes, success) at any x
Equivalent forms of the logistic regression model:
æ p ö
Logit form logç ÷ = b 0 + b1 X
è1- p ø
b o + b1 X
Probability form e
p= b o + b1 X
1+ e
 Binary Logistic Regression Model

Y = Binary X1,X2,…,X
X = Single
k = Multiple
predictor
response predictors
π = proportion of 1’s (yes,
at anysuccess)
x , x , …,
at xany x1 2 k

Equivalent forms of the logistic regression model:

æ p ö
Logit form logç ÷ = b 0 + b1 X 1 + b 2 X 2 + ! + b k X k
è1- p ø

b o + b1 X 1 + b 2 X 2 +!+ b k X k
Probability form p = e
b o + b1 X 1 + b 2 X 2 +!+ b k X k
1+ e
 Interactions in logistic regression
Consider Survival in an ICU as a function of
SysBP -- BP for short – and Sex
> intermodel=glm(Survive~BP*Sex, family=binomial, data=ICU)
> summary(intermodel)

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.439304 1.021042 -1.410 0.15865
BP 0.022994 0.008325 2.762 0.00575 **
Sex 1.455166 1.525558 0.954 0.34016
BP:Sex -0.013020 0.011965 -1.088 0.27653

Null deviance: 200.16 on 199 degrees of freedom

Residual deviance: 189.99 on 196 degrees of freedom
 1.0 Rep = red,
Dem = blue
0.8

Lines are
very close
Prob of voting Yes

0.6

to parallel;
0.4

not a
significant
0.2

interaction
0.0

0 1 10 100 1,000 10,000 1,000,000

Auto industry contributions (lifetime)

 Generalized Linear Model
(1) What is the link between Y and b0 + b1X?
(a) Regular reg: indentity
(b) Logistic reg: logit
(c) Poisson reg: log
(2) What is the distribution of Y given X?
(a) Regular reg: Normal (Gaussian)
(b) Logistic reg: Binomial
(c) Poisson reg: Poisson
 C-index, a measure of concordance

Med school acceptance: predicted by MCAT

and GPA?

Med school acceptance: predicted by coin toss??

> library(Stat2Data)
> data(MedGPA)
> str(MedGPA)
> GPA10=MedGPA$GPA*10
> Med.glm3=glm(Acceptance~MCAT+GPA10, family=binomial,
data=MedGPA)
> summary(Med.glm3)
> Accept.hat <- Med.glm3$fitted > .5
> with(MedGPA, table(Acceptance,Accept.hat))

Accept.hat
Acceptance FALSE TRUE
0 18 7
1 7 23

18 + 23 = 41 correct out of 55
> with(MedGPA, table(Acceptance,Accept.hat))

Accept.hat
Acceptance FALSE TRUE
0 18 7
1 7 23

Now consider that there were 30 successes and

25 failures. There are 30*25=750 possible pairs.
We hope that the predicted Pr(success) is greater
for the success than for the failure in a pair! If yes
then the pair is “concordant”.
C-index = % concordant pairs
 The R package rms has a command, lrm, that
does logistic regression and gives the C-index.
> #C-index work using the MedGPA data
> library(rms) #after installing the rms package
> m3=lrm(Acceptance~MCAT+GPA10, data=MedGPA)
> m3
lrm(formula = Acceptance~ MCAT + GPA10)
Model Likelihood Discrimination Rank Discrim.
Ratio Test Indexes Indexes
Obs 55 LR chi2 21.78 R2 0.437 C 0.834
0 25 d.f. 2 g 2.081 Dxy 0.668
1 30 Pr(> chi2) <0.0001 gr 8.015 gamma 0.669
max |deriv| 2e-07 gp 0.342 tau-a 0.337
Brier 0. 167

Coef S.E. Wald Z Pr(>|Z|)

Intercept -22.373 6.454 -3.47 0.0005
MCAT 0.1645 0.1032 1.59 0.1108
GPA10 0.4678 0.1642 2.85 0.0044
 Suppose we scramble the cases..
Then the C-index should be ½, like coin tossing
> newAccept=sample(MedGPA$Acceptance) #scramble the acceptances
> m1new=lrm(newAccept~MCAT+GPA10,data=MedGPA)
> m1new
lrm(formula = newAccept ~ MCAT + GPA10)
Model Likelihood Discrimination Rank Discrim.
Ratio Test Indexes Indexes
Obs 55 LR chi2 0.24 R2 0.006 C 0.520
0 25 d.f. 2 g 0.150 Dxy 0.040
1 30 Pr(> chi2) 0.8876 gr 1.162 gamma 0.041
max |deriv| 1e-13 gp 0.037 tau-a 0.020
Brier 0.247

Coef S.E. Wald Z Pr(>|Z|)

Intercept -1.4763 3.4196 -0.43 0.6659
MCAT 0.0007 0.0677 0.01 0.9912
GPA10 0.0459 0.1137 0.40 0.6862