Re: CLT and regression

Old Mac User wrote:
You wrote...

"Any kind of nonlinear transformation will destroy a
putative linear relation. So, no transformations
without an outstanding reason, not involving the data."

I respectively suggest this this makes no sense at all.
If the data (scores) is bounded by 0 - 100... and if you
fit those data without any transformation of the scores
(such as a logistic transformation)... then your finished
model will be capable of predicting outcomes less than
0 and/or greater than 100. That alone is sufficient to
say "use a constraining transformation that will prevent
silly predictions." This is another way of saying that if
you don't do this, your model will not pass the red-face test.


Professor Rubin's "out" is "without an outstanding reason". Fitting a model that doesn't describe the bulk of the data might be consider and outstanding reason.

I would also take issue with "not involving the data". I have seen some analysts taking Tukey's ladder of transformations a bit *too* much to heart and use things like X^0.34, but since I work primarily with data that come from biological units, I'd have no problem with logging if the improvement satisfied the interocular traumatic test. Better to have fit an appropriate model after having looked at the data than stick with a model that clearly doesn't work.
.