4

Possible Duplicate:
Threshold for correlation coefficient to indicate statistical significance of a correlation in a correlation matrix

Context

I am doing an exploratory study to investigate the relationship between a drug (actually measured in two ways - by direct and indirect methods) and 15 various parameters. There are three groups of individuals, on which I compute correlation statistics separately ($n=11$, $n=11$ and $n=24$ - the first two are paired). Obviously, the number of statistical tests is large, and I would like to control for multiple testing.

I am using Spearman correlation as data are limited and do not (always) follow a normal distribution. I am working in R.

Questions

  • How should I control for multiplicity of testing?

This is an hypothesis generating study, and I would rather prefer to use less conservative methods.

I have been looking into Benjamini–Hochberg's procedure (known under R as "BH" or "fdr") which allows to control the false discovery rate. However, I am not sure if I would violate dependence assumptions. And I am not sure if this could be used in correlation statistics.

Perhaps I should not adjust for multiple testing at all. In my mind, if both direct and indirect methods give the same association, false positives are highly unlikely.

Community
  • 1
user4229
  • 91
  • 5
  • 2
    If you don't find an answer in the existing threads on multiple testing for correlations, could you please explain how this question is different? To find them, link on the "multiple-comparisons" tab and then add "correlation" to the search terms in the textbox at the upper right corner of the page. You should get 8 results. – whuber Apr 19 '11 at 16:37
  • Thanks! I am new at this forum, and do as you suggest! Anyway, again...thanks for humble answer! – user4229 Apr 19 '11 at 22:30
  • 3
    @user4229 I'll second @whuber's remark and think you'll find useful clues about MT issues with correlation measures, if not an exact answer to most of your questions here. Now, if you're really interested in discussing FDR vs. single-step methods (such as the Bonferroni method), FP vs. FN issues in semi-exploratory studies, or multiple testing when dealing with correlated predictors, then I would suggest you rewrite your question in that particular direction, making this question a less obvious duplicate of some of the earlier threads. – chl Apr 20 '11 at 14:20

0 Answers0