# NOT RUN {
#==============================
# pb() and ps() functions
data(aids)
# fitting a smoothing cubic spline with 7 degrees of freedom
# plus the a quarterly effect
aids1<-gamlss(y~ps(x,df=7)+qrt,data=aids,family=PO) # fix df's
aids2<-gamlss(y~pb(x,df=7)+qrt,data=aids,family=PO) # fix df's
aids3<-gamlss(y~pb(x)+qrt,data=aids,family=PO) # estimate lambda
with(aids, plot(x,y))
with(aids, lines(x,fitted(aids1),col="red"))
with(aids, lines(x,fitted(aids2),col="green"))
with(aids, lines(x,fitted(aids1),col="yellow"))
rm(aids1, aids2, aids3)
#=============================
# }
# NOT RUN {
# pbc()
# simulate data
set.seed(555)
x = seq(0, 1, length = 100)
y = sign(cos(1 * x * 2 * pi + pi / 4)) + rnorm(length(x)) * 0.2
plot(y~x)
m1<-gamlss(y~pbc(x))
lines(fitted(m1)~x)
rm(y,x,m1)
#=============================
# the pvc() function
# function to generate data
genData <- function(n=200)
{
f1 <- function(x)-60+15*x-0.10*x^2
f2 <- function(x)-120+10*x+0.08*x^2
set.seed(1441)
x1 <- runif(n/2, min=0, max=55)
x2 <- runif(n/2, min=0, max=55)
y1 <- f1(x1)+rNO(n=n/2,mu=0,sigma=20)
y2 <- f2(x2)+rNO(n=n/2,mu=0,sigma=30)
y <- c(y1,y2)
x <- c(x1,x2)
f <- gl(2,n/2)
da<-data.frame(y,x,f)
da
}
da<-genData(500)
plot(y~x, data=da, pch=21,bg=c("gray","yellow3")[unclass(f)])
# fitting models
# smoothing x
m1 <- gamlss(y~pb(x), data=da)
# parallel smoothing lines
m2 <- gamlss(y~pb(x)+f, data=da)
# linear interaction
m3 <- gamlss(y~pb(x)+f*x, data=da)
# varying coefficient model
m4 <- gamlss(y~pvc(x, by=f), data=da)
GAIC(m1,m2,m3,m4)
# plotting the fit
lines(fitted(m4)[da$f==1][order(da$x[da$f==1])]~da$x[da$f==1]
[order(da$x[da$f==1])], col="blue", lwd=2)
lines(fitted(m4)[da$f==2][order(da$x[da$f==2])]~da$x[da$f==2]
[order(da$x[da$f==2])], col="red", lwd=2)
rm(da,m1,m2,m3,m4)
#=================================
# the rent data
# first with a factor
data(rent)
plot(R~Fl, data=rent, pch=21,bg=c("gray","blue")[unclass(rent$B)])
r1 <- gamlss(R~pb(Fl), data=rent)
# identical to model
r11 <- gamlss(R~pvc(Fl), data=rent)
# now with the factor
r2 <- gamlss(R~pvc(Fl, by=B), data=rent)
lines(fitted(r2)[rent$B==1][order(rent$Fl[rent$B==1])]~rent$Fl[rent$B==1]
[order(rent$Fl[rent$B==1])], col="blue", lwd=2)
lines(fitted(r2)[rent$B==0][order(rent$Fl[rent$B==0])]~rent$Fl[rent$B==0]
[order(rent$Fl[rent$B==0])], col="red", lwd=2)
# probably not very sensible model
rm(r1,r11,r2)
#-----------
# now with a continuous variable
# additive model
h1 <-gamlss(R~pb(Fl)+pb(A), data=rent)
# varying-coefficient model
h2 <-gamlss(R~pb(Fl)+pb(A)+pvc(A,by=Fl), data=rent)
AIC(h1,h2)
rm(h1,h2)
#-----------
# monotone function
set.seed(1334)
x = seq(0, 1, length = 100)
p = 0.4
y = sin(2 * pi * p * x) + rnorm(100) * 0.1
plot(y~x)
m1 <- gamlss(y~pbm(x))
points(fitted(m1)~x, col="red")
yy <- -y
plot(yy~x)
m2 <- gamlss(yy~pbm(x, mono="down"))
points(fitted(m2)~x, col="red")
#==========================================
# the pbz() function
# creating uncorrelated data
set.seed(123)
y<-rNO(100)
x<-1:100
plot(y~x)
#----------------------
# ML estimation
m1<-gamlss(y~pbz(x))
m2 <-gamlss(y~pb(x))
AIC(m1,m2)
op <- par( mfrow=c(1,2))
term.plot(m1, partial=T)
term.plot(m2, partial=T)
par(op)
# GAIC estimation
m11<-gamlss(y~pbz(x, method="GAIC", k=2))
m21 <-gamlss(y~pb(x, method="GAIC", k=2))
AIC(m11,m21)
op <- par( mfrow=c(1,2))
term.plot(m11, partial=T)
term.plot(m21, partial=T)
par(op)
# GCV estimation
m12<-gamlss(y~pbz(x, method="GCV"))
m22 <-gamlss(y~pb(x, method="GCV"))
AIC(m12,m22)
op <- par( mfrow=c(1,2))
term.plot(m12, partial=T)
term.plot(m22, partial=T)
par(op)
# fixing df is more trycky since df are the extra df
m13<-gamlss(y~pbz(x, df=0))
m23 <-gamlss(y~pb(x, df=0))
AIC(m13,m23)
# here the second penalty is not take effect therefore identical results
m14<-gamlss(y~pbz(x, df=1))
m24 <-gamlss(y~pb(x, df=1))
AIC(m14,m24)
# fixing lambda
m15<-gamlss(y~pbz(x, lambda=1000))
m25 <-gamlss(y~pb(x, lambda=1000))
AIC(m15,m25)
#--------------------------------------------------
# prediction
m1<-gamlss(y~pbz(x), data=data.frame(y,x))
m2 <-gamlss(y~pb(x), data=data.frame(y,x))
AIC(m1,m2)
predict(m1, newdata=data.frame(x=c(80, 90, 100, 110)))
predict(m2, newdata=data.frame(x=c(80, 90, 100, 110)))
#---------------------------------------------------
# }
Run the code above in your browser using DataLab