Simultaneous Confidence Bands in GLM

• Software was called Parfit , which computes the pointwise and 4 other simultaneous confidence bands (SCB) in Table 3 of Sun, Loader and McCormick (2000).

  The new version of the software has been incorporated into Locfit. If you have a Locfit that has a date newer than Feb 24, 2001, you already have the "Parfit" for computing the SCB.

  To compile in Linux, for example, in the $SHOME/library directory or your own locfit directory, type

  ./install.S5
  

  To make the SCB's, use e.g.

  Splus
  library(locfit,first=T)
  fit4 <- scb(deaths ~ lconc, type=4, w=nins, data=insect,
    deg=1, family="binomial", kern="parm")
  plot(fit4)
  

  Arguments are type (= 0 to 4; 0=pointwise, 1 to 4 as in table 3), w (= a vector of prior weights; e.g. When the responses are binomial, nins indicates `n' for each observation, default is a vector of 1's, assuming a 0-1 response; When the responses are Gaussian, it is 1/sigma^2 for non-homogeneous data) and kern="parm" (to force the parametric fit), as well as other typical Locfit or Splus arguments.

  Simulations are also in the source code, but it has to be compiled in stand-alone mode, use

  make clean    # removes object files from above (different format!)
  make          # builds locfit
  ./locfit
  locfit> x=runif(100) y=rnorm(100)
  locfit> scbmax y~x kern=parm corr=2
  (e.t.c. other arguments as before.)
  

• Reference: Sun, Loader and McCormick (2000). Confidence Bands in Generalized Linear Models Annals of Statistics. Vol 28, No 2, 429-460.