<dumping another draft written in April>
I'm reading the section on (predictive) margins in the Stata user's
guide. I sounds to me a lot more intuitive than the type 1, 2,3 sum of
squares and orthogonality or coding requirements in the R discussions.
the thread at
https://stat.ethz.ch/pipermail/r-help/2005-September/079305.html
looks pretty negative on them
Maybe it's the difference if you have complete control over your
design in an experiment, or if you have to worry about lots of
covariates and a sample where you cannot choose much about the
composition (surveys with unequally distributed respondents, "natural"
experiments in economics). ?
Josef