[Rd] gaussian family change suggestion

[Simon Wood]

Hi,

Currently the `gaussian' family's initialization code signals an error if
any response data are zero or negative and a log link is used. Given that
zero or negative response data are perfectly legitimate under the GLM
fitted using `gaussian("log")', this seems a bit unsatisfactory. Might
it be worth changing it?

The current offending code from `gaussian' is:

initialize = expression({
            n <- rep.int(1, nobs)
            if (is.null(etastart) && is.null(start) && is.null(mustart) &&
                ((family$link == "inverse" && any(y == 0)) ||
                  (family$link == "log" && any(y <= 0)))) stop(
              "cannot find valid starting values: please specify some")
            mustart <- y
        })

A possible replacement would be ...

initialize = expression({
            n <- rep.int(1, nobs)
            if (is.null(etastart) && is.null(start) && is.null(mustart) &&
                ((family$link == "inverse" && all(y == 0)) ||
                  (family$link == "log" && all(y <= 0)))) stop(
      "cannot find valid starting values: please specify some")
        mustart <- y
      if (family$link=="log") {
        iy <- y<=0
        if (sum(iy)) mustart[iy] <- min(y[!iy])*.5
      } else if (family$link=="inverse") {
        iy <- y==0
        if (sum(iy)) mustart[iy] <- min(abs(y[!iy]))*.5
      }
        })

best,
Simon


>- Simon Wood, Mathematical Sciences, University of Bath, Bath BA2 7AY
>-             +44 (0)1225 386603         www.maths.bath.ac.uk/~sw283/