h.wickham at gmail.com
Sun Jun 8 21:48:27 CEST 2008
> Well put Doug. I would add another condition, which I don't know how to
> state precisely. The settings for the other terms, which are usually
> marginal medians, modes, or means, must make sense when considered jointly.
> Frequently when all adjustment covariates are set to overall marginal means
> the resulting "subject" is very atypical.
> To me much of the problem is solved one one develops a liking for predicted
> values and differences in them.
Maybe I'm still misunderstanding, but isn't that exactly what effects
displays are? They're just some way to allow you to say, I'm
interested in variables x, y and z, and I don't really care about the
other variables in the model - what are some typical predictions?
The effects package implements this idea for categorical x, y, and z,
but the basic idea remains the same for continuous variables - except
instead of using all the levels of the factor, you'd use a grid within
the range of the data.
More information about the R-help