rootogram for normal distributions

On Sun, Jan 16, 2011 at 11:58 AM, Hugo Mildenberger

Thank you very much for your qualified answers, and also for the
link to the Tukey paper. I appreciate Tukey's writings very much.

Yes, thanks to Hadley for the nice reference, I hadn't seen it before.

Looking at the lattice code (below), a possible implementation might
involve ?binning, not so?

I see a problematic part here:

? xx <- sort(unique(x))

Unique certainly works well with Poisson distributed data, but is
essentially a no-op when confronted with continous floating-point
numbers.

True, but as Achim said, rootogram() is intended to work with data
arising from discrete distributions, not continuous ones. I see now
that this is not as explicit as it could be in the help page (although
"frequency distribution" gives a hint), which I will try to improve.

I don't think automatic handling of continuous distributions is simple
(because it is not clear how you would specify the reference
distribution). However, a little preliminary work will get you close
with the current implementation:

xnorm <- rnorm(1000)

## 'discretize' by binning and replacing data by bin midpoints
h <- hist(xnorm, plot = FALSE) # add arguments for more control
xdisc <- with(h, rep(mids, counts))

## Option 1: Assume bin probabilities proportional to dnorm()
norm.factor <- sum(dnorm(h$mids, mean(xnorm), sd(xnorm)))

rootogram(~ xdisc,
          dfun = function(x) {
              dnorm(x, mean(xnorm), sd(xnorm)) / norm.factor
          })

## Option 2: Compute probabilities explicitly using pnorm()

## pdisc <- diff(pnorm(h$breaks)) ## or estimated:
pdisc <- diff(pnorm(h$breaks, mean = mean(xnorm), sd = sd(xnorm)))
pdisc <- pdisc / sum(pdisc)

rootogram(~ xdisc,
          dfun = function(x) {
              f <- factor(x, levels = h$mids)
              pdisc[f]
          })

-Deepayan