fitting a loess line in R

Question

I'm having some trouble using loess() and predict() in R. I have used the following code to simulate data:

Overall=0.6
RR=1.1
Noise=0.05

x=seq(from=0.01, to=10, by=0.01) 
logRR=log(RR)
logBeta0=log(Overall)


linear.pred = logBeta0 + (logRR*x) + rnorm(length(x), 0, Noise*sqrt(x))
linear.pred.high = logBeta0 + (logRR*15) + rnorm(length(x), 0, Noise/5)

PoissonRate <- function (x) ifelse(x<=9, exp(linear.pred), exp(linear.pred.high))

xyplot(PoissonRate(x)~x) #the shape of the 'raw' data


loess_fit <- loess(x~PoissonRate(x))
lines(predict(loess_fit), col = "black")

Apologies but I can't work out how to attach a picture to show what this looks like!

The last two lines of code end up just adding a random black line half off the graph, although when I have used this command on different (very similar) data before, it has seemed to work fine. What am I missing?! Any help would be great, thanks :)

Answer 1

You shouldn't call llines() (if that is what you meant by lines() ) outside of the xyplot call, at least to my understanding. ?llines has:

Description:

     These functions are intended to replace common low level
     traditional graphics functions, primarily for use in panel
     functions.

Hence, one option is to do as it suggests and build your own panel function on the fly. Here is an example:

set.seed(1)
dat <- data.frame(pr = PoissonRate(x), x = x)
loess_fit <- loess(pr ~ x, data = dat)

xyplot(PoissonRate(x) ~ x,
  panel = function(x, y, ...) {
    panel.xyplot(x, y, ...)
    llines(dat$x, predict(loess_fit), col.line = "red")
  })

which produces:

In general, I would probably do this a different way - I would generate the data outside the formula. The I would predict for new data locations spread evenly and in order over the range of x . That way, even if the input data are not in increasing order in x you get sensible predictions you can use for a line plot. Eg

## following on from the above
pred <- with(dat, data.frame(x = seq(min(x), max(x), length = 100)))
pred <- transform(pred, pr = predict(loess_fit, newdata = x))

xyplot(PoissonRate(x) ~ x,
  panel = function(x, y, ...) {
    panel.xyplot(x, y, ...)
    with(pred, llines(x, pr, col.line = "red"))
  })