I am analyzing a population-representative cross sectional survey to identify associations between several factors and cardiovascular disease. I am performing binary logistic regression analysis with the model having coronary heart disease as the dependent variable (0=no, 1=yes). From all the independent variables included in the model, I get 17 p-values. Many of them are <0.0001, but some are at p<0.05 but >0.01. The question is: should I use some kind of correction for all the P values? Bonferroni correction is usually too conservative, but i was thinking about the false discovery rate method. What do you think? Would that be appropriate?
You definitely should apply multiple testing. There are many options available to you. The choice depends on what kind of conclusions you want to make and what risk you want to take for the wrong conclusions. The traditional approach is to try to keep the probability of one wrong conclusion low. Bonferroni is conservative but may be adequate if all the individual hypothesis tests have very low p-values. Other options include Tukey's method bootstrap and permutation methods among other. Specific books on multiple testing go into details. I like Jason Hsu's book for an understanding of many of these traditional methods. Westphall and Young explain how the bootstrap and permutation methods work. Regarding FDR that involves allowing the probability of more than one test error (false discoveries) to occur but with small probability of more. For example you can specify the number of false discoveries. This could be say 3 in a total of 100 tests. FDR is usual done when the number of hypotheses to test are large. This is often the case in genetic testing. In such situations the tradition approaches could be hopeless because of the very high number of tests.