P-value adjustment for different classes each tested individually Testing the effect of a certain gene on survival of two different but somehow related (brain cancer GBM and LGG) patients, with each cancer (GBM or LGG) having several subtypes, how should I adjust for multiple testing? Especially that survival analysis consists of disease-free and overal survival. Should I adjust for all measurements or can I consider GBM/LGG separately or overall survival/disaase-free-survival separatey (so that I get a significant adjusted p-value? )
 A: A lot depends on the nature of your data and your understanding of the subject matter. If it's reasonable to construct a single model for each of disease-free and overall survival among all these types of cancer, with the type of cancer included as a predictor in each of the models, you typically won't be expected to do multiple-testing corrections on the individual coefficient estimates returned by the model. What's critical is that the overall model is significant.
The null hypothesis from which you might be trying to protect yourself with multiple-comparison corrections among coefficients within a model is that no predictors are significantly associated with survival. If the survival model is significant overall, however, that null hypothesis presumably doesn't hold. With a model that is significant overall, readers will typically be able to judge for themselves how much credence to put in particular coefficients, based on the coeffcients' confidence limits and the readers' knowledge of the underlying subject matter.
In practice, you also probably won't be expected to do a multiple-comparison correction for separate models of disease-free and overall survival based on the same data (although I suppose that such a correction on the overall significance of the 2 models might be warranted). I'd actually worry more about the quality of the disease-free survival model, as disease-free survival is typically determined at scheduled clinic visits. The disease-free survival times are thus best thought of as interval censored, while they are often modeled as precisely defined event times, like dates of death.
