Generalized Linear Models

We have previously worked with regression models where the response variable
is quantitative and normally distributed. Now we turn our attention to two types of
models where the response variable is discrete and the error terms do not follow
a normal distribution, namely logistic regression and Poisson regression. Both
belong to a family of regression models called generalized linear models.

Generalized linear models are extensions of traditional regression models that

allow the mean to depend on the explanatory variables through a link function,
and the response variable to be any member of a set of distributions called the
exponential family (e.g., Normal, Poisson, Binomial).

We can use the function glm() to work with generalized linear models in R. It’s
usage is similar to that of the function lm() which we previously used for multiple
linear regression. The main difference is that we need to include an additional
argument family to describe the error distribution and link function to be used in
the model. In this tutorial we show how glm() can be used to fit logistic
regression and Poisson regression models.

A. Logistic Regression

Logistic regression is appropriate when the response variable is categorical with

two possible outcomes (i.e., binary outcomes). Binary variables can be
represented using an indicator variable Yi, taking on values 0 or 1, and modeled
using a binomial distribution with probability P(Yi=1) = i. Logistic regression
models this probability as a function of one or more explanatory variables.

To perform logistic regression in R, use the command:

> glm(response ~ explanantory_variables, family=binomial)

Note that the option family is set to binomial, which tells R to perform logistic
regression.

Ex. A car manufacturer was interested in creating a model for determining the
probability that families will purchase a new car during the next year. A random
sample of 33 suburban families was selected. Data on annual income (in
thousands of dollars) and the current age of the oldest family car (in years) was
obtained. A follow-up interview was conducted a year later to determine whether
or not the family actually purchased a new car during the year (Y=1 if the family
purchased a car and 0 otherwise).

We are interested in determining the probability that a family purchases a new

car given their income and the age of their oldest car.
To read in the data set and fit a logistic regression model we type:

> dat = read.table("Purchase.txt",header=TRUE)

> results = glm(new ~ income + age, family=binomial)
> results

Call: glm(formula = new ~ income + age, family = binomial)

Coefficients:
(Intercept) income age
-4.73931 0.06773 0.59863

Degrees of Freedom: 32 Total (i.e. Null); 30 Residual

Null Deviance: 44.99
Residual Deviance: 36.69 AIC: 42.69

According to the output, the model is logit( i) = -4.74 + 0.068*income + 0.60*age.

After fitting the model, we can test the overall model fit and hypothesis regarding
a subset of regression parameters using a likelihood ratio test (LRT). Likelihood
ratio tests are similar to partial F-tests in the sense that they compare the full
model with a restricted model where the explanatory variables of interest are
omitted. The p-values of the tests are calculated using the 2 distribution.

To test the hypothesis H0: 1= 2=0 we can compare our model with a reduced
model that only contains an intercept term. A likelihood ratio test comparing the
full and reduced models can be performed using the anova() function with the
additional option test="Chisq".

> results.reduced =glm(new ~ 1, family=binomial)

> anova(results.reduced,results, test="Chisq")
Analysis of Deviance Table
Model 1: new ~ 1
Model 2: new ~ income + age
Resid. Df Resid. Dev Df Deviance P(>|Chi|)
1 32 44.987
2 30 36.690 2 8.298 0.016

The likelihood ratio test statistic is 2=8.298 with a p-value=0.016. Hence, we

have relatively strong evidence in favor of rejecting H0.
As a next step, we perform tests on the individual regression parameters.

> summary(results)

Call:
glm(formula = new ~ income + age, family = binomial)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.6189 -0.8949 -0.5880 0.9653 2.0846

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -4.73931 2.10195 -2.255 0.0242 *
income 0.06773 0.02806 2.414 0.0158 *
age 0.59863 0.39007 1.535 0.1249
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Null deviance: 44.987 on 32 degrees of freedom

Residual deviance: 36.690 on 30 degrees of freedom
AIC: 42.69

Number of Fisher Scoring iterations: 4

To test H0: 1=0, we use z = 2.414 (p-value=0.0158). Hence, the family’s income
appears to have a significant impact on the probability of purchasing a new car,
while controlling for the age of the families oldest car.

To test H0: 2=0, we use z = 1.535 (p-value=0.1249). Hence, the age of a family’s
oldest car does not appear to have a significant impact on the probability of
purchasing a new car, once income is included in the model.

To compute how the odds of purchasing a car changes as a function of income

use the commands:

> exp(coef(results))
(Intercept) income age
0.008744682 1.070079093 1.819627221

To create a 95% confidence interval for the estimate, type:

> exp(confint.default(results))
2.5 % 97.5 %
(Intercept) 0.0001420897 0.5381773
income 1.0128238514 1.1305710
age 0.8471457285 3.9084695
We see that the odds ratio corresponding to income is 1.070 (95% CI: (1.013,
1.131)). This implies that if we fix the age of the oldest car, increasing family
income by one thousand dollars will increase the odds of purchasing a new car
by 0.07.

We are often interested in using the fitted logistic regression curve to estimate
probabilities and construct confidence intervals for these estimates. We can do
this using the function predict.glm. The usage is similar to that of the function
predict which we previously used when working on multiple linear regression
problems. The main difference is the option type, which tells R which type of
prediction is required. The default predictions are given on the logit scale (i.e.
predictions are made in terms of the log odds), while using type = "response"
gives the predicted probabilities.

To predict the probability that a family with an annual income of $53 thousand
and whose oldest car is 1 year old will purchase a new car in the next year, type:

> pi.hat = predict.glm(results, data.frame(income=53, age=1),

type="response", se.fit=TRUE)
> pi.hat$fit
[1] 0.3656668

This tells us that the predicted probability is 0.37. In order to obtain confidence
intervals we instead need to work on the logit scale and thereafter transform the
results into probabilities. To create a 95% confidence interval for the estimate,
type:

> l.hat = predict.glm(results, data.frame(income=53, age=1), se.fit=TRUE)

> ci = c(l.hat$fit - 1.96*l.hat$se.fit, l.hat$fit + 1.96*l.hat$se.fit)

To transform the results to probabilities type:

> exp(ci)/(1+exp(ci))
[1] 0.1145063 0.7198689

For a family with an annual income of $53 thousand and whose oldest car is 1
year old, the estimated probability of purchasing a new car is 0.366. A 95% CI is
given by (0.115, 0.720).
B. Poisson Regression

Data is often collected in counts (e.g. the number of heads in 12 flips of a coin or
the number of car thefts in a city during a year). Many discrete response
variables have counts as possible outcomes. Binomial counts are the number of
successes in a fixed number of trials, n. Poisson counts are the number of
occurrences of some event in a certain interval of time (or space). While Binomial
counts only take values between 0 and n, Poisson counts have no upper bound.

We now consider a nonlinear regression model where the response outcomes

are discrete counts that follow a Poisson distribution. Poisson regression
provides a model that describes how the mean response , changes as a
function of one or more explanatory variables. To perform logistic regression in
R, we use the command:

> glm(response ~ explanantory_variables, family=poisson)

Note that we specified the family to be poisson, which tells R to perform Poisson
regression.

Ex. Researchers studied 41 male African elephants over a period of 8 years.

The age of the elephant at the beginning of the study and the number of
successful matings during the 8 years were recorded. We assume the number of
matings follows a Poisson distribution, where the mean depends on the age of
the elephant in question.

We can fit a Poisson regression model using the following code:

> dat = read.table("elephants.txt", header=TRUE)

> attach(dat)
> results = glm(mating ~ age, family=poisson)
> summary(results)

Call:
glm(formula = mating ~ age, family = poisson)

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.58201 0.54462 -2.905 0.00368 **
age 0.06869 0.01375 4.997 5.81e-07 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

Null deviance: 75.372 on 40 degrees of freedom

 Residual deviance: 51.012 on 39 degrees of freedom
AIC: 156.46

Number of Fisher Scoring iterations: 5

To determine whether there is a significant relationship between the mean

number of matings and the age of the elephants we test H0: 1=0. The test
statistic is z=4.997 (p-value<0.0001). Hence, it appears that age does impact the
mean number of matings.

To see how the mean number of matings increases per year use the commands:

> beta =coef(results)

> beta
(Intercept) age
-1.58200796 0.06869281
> exp(beta[2])
age
1.071107

To create a 95% confidence interval for the estimate, type:

> exp(confint.default(results))
2.5 % 97.5 %
(Intercept) 0.07069036 0.5977577
age 1.04263544 1.1003563

Hence, each additional year is associated with a 7.1% increase in the mean
number of matings. A 95% confidence interval is given by (1.043, 1.100), which
represents a 4.3 - 10.0% increase.