% \VignetteIndexEntry{Medical Care - Zero-Inflated and Zero-Hurdle-Model}
%\VignetteEngine{knitr::knitr}
%\VignetteEncoding{UTF-8}

\documentclass[a4paper]{article}

\title{Medical Care - Zero-Inflated and Zero-Hurdle-Model}

\begin{document}

\maketitle
First  the medcare data are loaded:
<<results='hide',eval=TRUE>>=
library(catdata)
data(medcare)
attach(medcare)
@

The dependent variable "ofp" (numbers of physician visits) is a count variable, so a poisson-family glm seems to be a good choice.
<<eval=TRUE>>=
med1=glm(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school,
         family=poisson,data=medcare[male==1 & ofp<=30,])
summary(med1)
@
In many real-world datasets the variance of count-data is higher than predicted by the Poisson distribution, so we fit a quasi-Poisson model with dispersion parameter.
<<eval=TRUE>>=
med2=glm(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school,
         family=quasipoisson,data=medcare[male==1 & ofp<=30,])
summary(med2)
@
With an estimated dispersion parameter of 4.69 the standard errors are much bigger now.
An alternative to a quasi-poisson model is to use the negative binomial distribution.
<<eval=TRUE>>=
library(MASS)
med3=glm.nb(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school,
            data=medcare[male==1 & ofp<=30,])
summary(med3)
@
In this model the standard errors are slightly lower with the result that "healthexcellent" and "married" are now significant. (level=0.05)
<<eval=TRUE>>=
@
In count data there are often much more zeros than expected. Therefore one can fit a "zero-inflated" model using the pscl package. In the first "zero-inflated" model one assumes that the occurence of zeros does depend on covariates:
<<results='hide',eval=TRUE>>=
library(pscl)
@
<<eval=TRUE>>=
med4=zeroinfl(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school|1,
              data=medcare[male==1 & ofp<=30,])
summary(med4)
@
In the second "zero-inflated" model  the occurence of zeros can depend on covariates:
<<eval=TRUE>>=
med5=zeroinfl(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school,
              data=medcare[male==1 & ofp<=30,])
summary(med5)
@
An alternative to "zero-inflation" is the "zero-hurdle" model. In the following similar models as above are fitted.
<<eval=TRUE>>=
med6=hurdle(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school|1
            ,data=medcare[male==1 & ofp<=30,])
summary(med6)
@
<<eval=TRUE>>=
med7=hurdle(ofp ~ hosp+healthpoor+healthexcellent+numchron+age+married+school,
            data=medcare[male==1 & ofp<=30,])
summary(med7)
@
\end{document}