Standard errors in gam(), spm(), geoR in R

• The standard errors produced as the se.fit output component from predict.gam() are standard errors for the function estimate, not prediction standard errors. I.e., the noise uncertainty (variability in observations about the function) is not part of the standard errors.
  • Also note that the output of predict.gam() is an array. You can freeze R if you try to use this output as an input to lm(). Instead use the numeric version of the output, e.g., c(predict.gam(mod)).
• The standard errors produced by the predict.spm() function in SemiPar are also just standard errors for the function estimate and not prediction standard errors.
• The standard errors produced by kriging in the geoR package and output as the $krige.var list item (i.e. as variances, the square of the standard error) are prediction standard errors, but function standard errors can be obtained using output=output.control(signal=TRUE)) as an arigument to krige.conv().
• The standard errors produced by kriging in the geoR package and output as the list item (i.e. as variances, the square of the standard error) also appear to be just standard errors for the function uncertainty.

Keywords: gam, mgcv, geoR, R, standard errors, predict.gam, prediction, predict.spm, krige.var, kriging

Last modified 12/22/06.