Strict Bonferroni - looking for p value I run 18 exploratory regressions and would like to know what the strict Bonferroni p-value would be. 
Nine regressions are negative binomial regressions with 4 predictor variables, the other 9 are WLS regressions with 6 predictor variables. These associations are yet unexplored (my dependent variables are clinical questionnaire items instead of clinical sum scores, we are looking for heterogeneity). 
I am sure more information is needed, and would be glad if you could either help me directly (I will gladly provide all information that is necessary) or hint me at literature that enables a non-stats person to do the calculation himself. 
I know that Bonferroni is very strict and do not plan to actually apply it to my analyses - I plan to use .01 as type I error to detect significant effects (and would appreciate feedback). My dependent variables are substantially correlated (between .3 and .6), but I at least want to provide what the p-value would be in the paper and discuss why it would be too conservative. 
 A: If your analyses are at this stage exploratory then I strongly advise that you do not do any 'corrections' for multiple comparisons at all. They are mostly important where hypotheses are being tested within the Neyman-Pearson error-decision framework which implicitly precludes the re-testing of hypotheses. (That aspect of N-P hypothesis testing is usually not emphasised, but it is exactly what 'inductive behaviour' requires.) If you are at the exploratory stage then there is no reason at all to decrease the sensitivity of your analyses by 'correcting' via a Bonferroni or other method.
For hypothesis formation there is also no good reason to use a low cutoff for 'significance', which will lower the power of the tests. Instead, consider examining (and presenting, if appropriate) the actual P-values. The comparisons with the lowest P-values are the most convincing and should be focussed on in new experiments. (You cannot, of course, use the same data to form and test a hypothesis.)
You might find a recent paper of mine in the British Journal of Pharmacology helpful for understanding the differences between N-P hypothesis testing and the more commonly useful Fisherian 'significance testing'. http://www.ncbi.nlm.nih.gov/pubmed/22394284 
Another approach—probably better, but certainly more complex—is to use a multi-level Bayesian model that has the many comparisons built into it as model components. For many different types of problem such a model seems to do away with any reason to consider multiple comparison corrections.http://www.stat.columbia.edu/~gelman/research/published/multiple2f.pdf
This blog item might be relevant too: http://andrewgelman.com/2011/01/data_exploratio/ 
(I have to admit here that I do not completely understand the multi-level modelling, and I have never used it myself.)
