###
#  Application of Generalized Estimating
#  Equations to predict US presidential 
#  election outcomes based on the age
#  of the candidate.
###

# Load in data, turn days into years, set contrast 
# for incumbency
dat = read.csv("usaElect.csv")
dat = cbind( dat, demAgeYear = dat$demAge/365 )
dat = cbind( dat, repAgeYear = dat$repAge/365 )

contrasts( dat$incumb ) <- contr.treatment(n=3,base=2)

# Load in package for fitting gee models
library(geepack)

# First consider elections from 1960 to 2016
ss=4

md1 = geeglm( 
  demWin ~ demAgeYear + repAgeYear + incumb, 
  id=demName, data=dat, 
  family=binomial(), 
  corstr = "exchangeable",
  subset = ss:18
);

summary(md1)

# Then consider elections from 1948 to 2016
ss=1

md2 = geeglm( 
  demWin ~ demAgeYear + repAgeYear + incumb, 
  id=demName, data=dat, 
  family=binomial(), 
  corstr = "exchangeable",
  subset = ss:18
);

summary(md2)

# Plot fitted models

prd1 = predict( md1 )
prd1 = 1/(1+exp(-prd1))
prd2 = predict( md2 )
prd2 = 1/(1+exp(-prd2))

plot( dat$demAgeYear[1:18], dat$demWin[1:18], las=1 )
lines( 
  sort(dat$demAgeYear[4:18]),sort(prd1,decreasing = T),
  col='blue',lwd=2
)
lines( 
  sort(dat$demAgeYear[1:18]),sort(prd2,decreasing = T),
  col='red',lwd=2
)