Multivariable Logistic Model:

Interaction

Professor Tuan V. Nguyen

Garvan Institute of Medical Research
University of New South Wales
Sydney – Australia

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Contents

Topics
– Binary anova
– Variable selection
– Under/over dispersion

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Binary Anova

• Categorical variables are included in logistic

regressions in just the same way as in linear
regression.
• Done by means of “dummy variables”.
• Interpretation is similar, but in terms of log-
odds rather than means.
• A model which fits a separate probability to
every possible combination of factor levels
is a maximal model, with zero deviance

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Example

• The plum tree data: see the coursebook,

• For another example, see Tutorial 8)
• Data concerns survival of plum tree cuttings.
Two categorical explanatory variables, each at
2 levels: planting time (spring, autumn) and
cutting length (long, short). For each of these 4
combinations 240 cuttings were planted, and
the number surviving recorded.

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Data

length time r n
1 long autumn 156 240
2 long spring 84 240
3 short autumn 107 240
4 short spring 31 240

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Fitting

> plum.glm<-glm(cbind(r,n-r)~length*time, family=binomial,

data=plum.df)
> summary(plum.glm)
Call:
glm(formula = cbind(r, n - r) ~ length * time, family =
binomial, data = plum.df)
Deviance Residuals: Zero residuals!
[1] 0 0 0 0
Coefficients: Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.6190 0.1353 4.574 4.78e-06 ***
lengthshort -0.8366 0.1876 -4.460 8.19e-06 ***
timespring -1.2381 0.1914 -6.469 9.87e-11 ***
lengthshort:timespring -0.4527 0.3009 -1.505 0.132
Null deviance: 1.5102e+02 on 3 degrees of freedom
Residual deviance: 1.7683e-14 on 0 degrees of freedom
AIC: 30.742
Zero deviance and df!

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Points to note

• The model length*time fits a separate

probability to each of the 4 covariate
patterns
• Thus, it is fitting the maximal model, which
has zero deviance by definition
• This causes all the deviance residuals to be
zero
• The fitted probabilities are just the ratios r/n

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Fitted logits

Logits Length=l Length= Coefficients: Estimate

ong (Intercept) 0.6190
short lengthshort -0.8366
timespring -1.2381
Time= .6190 -.8366 = lengthshort:timespring -0.4527
autumn .6190 -.2176

Time= .6190 - .6190 -1.2381

1.281=
-.8366-.4527 =
spring -.6191 -1.9084

> predict(plum.glm)
[1] 0.6190392 -0.6190392 -0.2175203 -1.9083470

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Interaction plot

attach(plum.df)
interaction.plot(length,time,log((r+0.5)/(n-r+0.5)))

Lines almost parallel,

indicating no interaction

0.5
time
on the log-odds scale
autumn
spring

0.0
-0.5
logit of s/n

-1.0
-1.5
-2.0

long short

length
Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Anova
> anova(plum.glm,test="Chisq")
Analysis of Deviance Table
Model: binomial, link: logit
Response: cbind(r, n - r)
Terms added sequentially (first to last)

Df Deviance Resid.Df Resid.Dev P(>|Chi|)

NULL 3 151.019
length 1 45.837 2 105.182 1.285e-11
time 1 102.889 1 2.294 3.545e-24
length:time 1 2.294 0 7.727e-14 0.130
> 1-pchisq(2.294,1)
[1] 0.1298748

Interaction not significant

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Final model: interpretation and fitted
probabilities

> plum2.glm<-glm(cbind(r,n-r)~length + time,

family=binomial, data=plum.df)
> summary(plum2.glm)
Call:
glm(formula = cbind(r, n - r) ~ length + time, family =
binomial, data = plum.df)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.7138 0.1217 5.867 4.45e-09 ***
lengthshort -1.0177 0.1455 -6.995 2.64e-12 ***
timespring -1.4275 0.1465 -9.747 < 2e-16 ***
Null deviance: 151.0193 on 3 degrees of freedom
Residual deviance: 2.2938 on 1 degrees of freedom
AIC: 31.036
> 1-pchisq(2.2938,1)
[1] 0.1298916

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Final model: interpretation and fitted
probabilities

Estimate Std. Error z value Pr(>|z|)

(Intercept) 0.7138 0.1217 5.867 4.45e-09 ***
lengthshort -1.0177 0.1455 -6.995 2.64e-12 ***
timespring -1.4275 0.1465 -9.747 < 2e-16 ***

Prob of survival less for short cuttings (coeff<0)

Prob of survival less for spring planting (coeff<0)

Null deviance: 151.0193 on 3 degrees of freedom

Residual deviance: 2.2938 on 1 degrees of freedom
AIC: 31.036

Deviance of 2.2938 on 1 df: pvalue is 0.1299

evidence is that no-interaction model fits well.

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Fitted Probabilities

Length= Length=
long short

Time =
0.6712 0.4246
autumn

Time =
0.3288 0.1504
spring

> predict(plum2.glm,type="response")
[1] 0.6712339 0.3287661 0.4245994 0.1504006

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Variable selection

• Variable selection proceeds as in ordinary

regression
• Use anova and stepwise
• AIC also defined for logistic regression
AIC = Deviance + 2 ´(number of parameters)
• Pick model with smallest AIC

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Example: lizard data

• Site preferences of 2 species of lizard, grahami

and opalinus
• Want to investigate the effect of
– Perch height
– Perch diameter
– Time of day
on the probability that a lizard caught at a site
will be grahami

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Data

length height time r n

1 short low early 54 67
2 short high early 44 49
3 long low early 25 43
4 long high early 18 19
5 short low mid 77 98
6 short high mid 63 67
7 long low mid 64 97
8 long high mid 21 26
9 short low late 22 36
10 short high late 25 38
11 long low late 13 24
12 long high late 5 10

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Eyeball analysis
> plot.design(lizard.df, y=log((lizard.df$r+0.5)
/(lizard.df$n-lizard.df$r+0.5)), ylab="mean of logits")
1.6

early
high
mid
1.4

short
1.2

Proportion of
mean of logits

grahami lizards
1.0

long higher when

0.8

low
perches are short
0.6

and high, and in the

0.4

late earlier part of the

length height time
day
Factors

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Model selection

• Full model is
cbind(r,n-r)~time*length*height
so fit this first.

• Then use anova and stepwise to select a

simpler model if appropriate

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 anova

> lizard.glm<-glm(cbind(r,n-r)~time*length*height,
+ family=binomial,data=lizard.df)
> anova(lizard.glm, test="Chisq")

Df Deviance Resid. Df Resid. Dev P(>|Chi|)

NULL 11 54.043
time 2 14.711 9 39.332 0.001
length 1 15.680 8 23.652 7.503e-05
height 1 13.771 7 9.882 2.065e-04
time:length 2 1.170 5 8.711 0.557
time:height 2 5.017 3 3.694 0.081
length:height 1 0.001 2 3.693 0.971
time:length:height 2 3.693 0 -1.354e-14 0.158

Both approaches suggest model

cbind(s,n-s) ~ time + length + height

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 stepwise

>null.model<-glm(cbind(r,n-r)~1, family=binomial,
data=lizard.df)
> step(null.model, formula(lizard.glm), direction="both")

Call: glm(formula = cbind(r, n - r) ~ height + time + length,

family = binomial, data = lizard.df)

Coefficients:
(Intercept) heightlow timelate timemid lengthshort
1.49466 -0.83011 -1.05278 0.04003 0.67630

Degrees of Freedom: 11 Total (i.e. Null); 7 Residual

Null Deviance: 54.04
Residual Deviance: 9.882 AIC: 64.09

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Summary

> summary(model2)
Call:
glm(formula = cbind(r, n - r) ~ time + length + height,
family = binomial, data = lizard.df)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.49466 0.28809 5.188 2.12e-07 ***
timelate -1.05278 0.28026 -3.756 0.000172 ***
timemid 0.04003 0.23971 0.167 0.867384
lengthshort 0.67630 0.20588 3.285 0.001020 **
heightlow -0.83011 0.23204 -3.578 0.000347 ***
---
Signif. codes: 0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 `
' 1
(Dispersion parameter for binomial family taken to be 1)

Null deviance: 54.0430 on 11 degrees of freedom

Residual deviance: 9.8815 on 7 degrees of freedom

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Diagnostics

Residuals vs Fitted Normal Q-Q

4 6

1.5

10
11

Std. deviance resid.

5
Residuals

0.5
> par(mfrow=c(2,2))

0
-0.5
> plot(model2,

-5
-10
which=1:4)

-1.5
10 3
10

0.0 0.5 1.0 1.5 2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Predicted values Theoretical Quantiles

Scale-Location Cook's distance

0.6
No major 3
10
6
10
3.0

0.5
Std. deviance resid.

problems

Cook's distance

0.4
11
2.0

0.3
0.2
1.0

0.1
0.0

0.0
0.0 0.5 1.0 1.5 2.0 2 4 6 8 10 12

Predicted values Obs. number

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Conclusions

• Weak suggestion that Grahami relatively more

numerous in mornings/midday
• Strong suggestion Grahami relatively more
numerous on short perches
• Strong suggestion Grahami relatively more
numerous on high perches

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Over/under dispersion

• The variance of the binomial B(n,p) distribution

is np(1-p), which is always less than the mean
np.
• Sometimes the individuals having the same
covariate pattern in a logistic regression may
be correlated.
• This will result in the variance being greater
than np(1-p) (if the correlation is +ve) or less
than np(1-p) (if the correlation is - ve)

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Over/under-dispersion

• If this is the case, we say the data are over-

dispersed (if the variance is greater) or under-
dispersed (if the variance is less)
• Consequence: standard errors will be wrong.
• Quick and dirty remedy: analyse as a binomial, but
allow the “scale factor” to be arbitrary: this models
the variance as
ynp(1-p) where y is the “scale factor”
(for the binomial, the scale factor is always 1)

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Over-dispersed model
> model3<-glm(cbind(r,n-r)~time+length+height,
family=quasibinomial,data=lizard.df)
> summary(model3)
Call:
glm(formula = cbind(r, n - r) ~ time + length + height,
family = quasibinomial, data = lizard.df)
> Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.49466 0.33128 4.512 0.00276 **
timelate -1.05278 0.32228 -3.267 0.01374 *
timemid 0.04003 0.27565 0.145 0.88864
lengthshort 0.67630 0.23675 2.857 0.02446 *
heightlow -0.83011 0.26683 -3.111 0.01706 *
---
(Dispersion parameter for quasibinomial family taken to be
1.322352)
Null deviance: 54.0430 on 11 degrees of freedom
Residual deviance: 9.8815 on 7 degrees of freedom

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012
 Comparison

Binomial
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.49466 0.28809 5.188 2.12e-07 ***
timelate -1.05278 0.28026 -3.756 0.000172 ***
timemid 0.04003 0.23971 0.167 0.867384
lengthshort 0.67630 0.20588 3.285 0.001020 **
heightlow -0.83011 0.23204 -3.578 0.000347 ***

Quasibinomial
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.49466 0.33128 4.512 0.00276 **
timelate -1.05278 0.32228 -3.267 0.01374 *
timemid 0.04003 0.27565 0.145 0.88864
lengthshort 0.67630 0.23675 2.857 0.02446 *
heightlow -0.83011 0.26683 -3.111 0.01706 *

Workshop on Analysis of Clinical Studies – Can Tho University of Medicine and Pharmacy – April 2012