- Thread starter Franci_
- Start date

Personally, I perform Bonferroni corrections very rarely. If there are only few comparisons to be made,

then usually in my work there is one primary comparison, and some secondary comparisons, and I do

not mind making the secondary comparisons without Bonferroni (except a reviewer would demand it).

If there are many comparisons, Bonferroni destroys power and would therefore lead to a meaningless

analysis. But there certainly are circumstances, e.g. with a very large sample and/or the absolute priority

to prevent false-positive results, where Bonferroni (or preferably Bonferroni-Holm) can be justified.

I do not know whether you'd be better off with multiple/multivariate approaches such as multiple regression.

If you need to make 30 distinct comparisons, and all null hypotheses are true, then the probability of* at least*

one false-positive finding is about 80%. Instead of Bonferroni (0.05/30=0.0017) I would be inclined to choose

a conservative, but not so crazy significance level, such as 0.01. If all 30 null hypotheses were true, the chance

of at least one false-positve result would be 26% then. But if some null hypotheses were not true, you would

retain enough power with an alpha of 0.01.

Just my 2pence

Karabiner

then usually in my work there is one primary comparison, and some secondary comparisons, and I do

not mind making the secondary comparisons without Bonferroni (except a reviewer would demand it).

If there are many comparisons, Bonferroni destroys power and would therefore lead to a meaningless

analysis. But there certainly are circumstances, e.g. with a very large sample and/or the absolute priority

to prevent false-positive results, where Bonferroni (or preferably Bonferroni-Holm) can be justified.

I do not know whether you'd be better off with multiple/multivariate approaches such as multiple regression.

If you need to make 30 distinct comparisons, and all null hypotheses are true, then the probability of

one false-positive finding is about 80%. Instead of Bonferroni (0.05/30=0.0017) I would be inclined to choose

a conservative, but not so crazy significance level, such as 0.01. If all 30 null hypotheses were true, the chance

of at least one false-positve result would be 26% then. But if some null hypotheses were not true, you would

retain enough power with an alpha of 0.01.

Just my 2pence

Karabiner

Last edited: