Re: mulitivariate linear regression and Bonferroni

The one key concept Bob pointed out is practical significance, which
was what I wanted to bring up in my original post.

It is fine and dandy for us statisticians to prescribe Bonferroni when
we are not actually analyzing the data. In practice, with many
variables, if you use Bonferroni, you might as well throw the entire
data set out because it will tell you that nothing is significant.
Sometimes that is the right answer; the entire data set is indeed all
noise and no signal. But most of the time, if we use a more finely
tuned method, we can detect a small signal.

The lesson of multiple comparisons is that we need to be careful when
we are simultaneously testing many things. It is the mentality that is
important so that the investigator will make mental note of this risk.
Using Bonferroni or similar tests blindly to shove the whole issue
under the table is dangerous.

DM

.