We are doing miRNA expression analysis using Taqman OpenArray panels. We are analyzing aproximately 500 miRNAs. I wanted to know if we must do Bonferroni or Benjamini-Hochberg in order to adjust p-values and avoid type I errors. May these multiple comparison corrections be substitutedwith qPCR validation of candidate miRNAs (those differentially expressed in OpenArray panels)? I have used two technical replicates and removed those that differs in more than 0.5 Cts.
If you are looking at all possible comparisons, then it seems to me that you need Bonferroni or B-H. See S. Knudsen, Cancer Diagnostics with DNA Microarrays, Wiley, sect 3.6.3. I don't understand the second question. Best wishes.
If you are not testing two substantial alternative hypotheses then there is no such thing like a "type-I error". It seems you do significance tests, that is, you calculate p-values with reference to some (null-)hypothesis (like: "the expected miRNA expression is the same under both experimental conditions"). This hypothesis is an abstract mathematical construct and there is not question if this is right or wrong. Logically, it is wrong anyway - but that's not the point. The relevant question is to find out how (in-)compatible the observed data is with this hypothesis (and the model). There is no way to control error-rates in such a context, because there is no sensible definition of errors. It is therefore not applicable to control family-wise error rates or any other kinds of compound error rates.
You can use p-values as measure to order the miRNAs and investigate the k miRNAs with the lowest p-values (with k as a number you can handle).
If you use p as a "surprise index" you need to consider that "at least one 'strange' result among many attemps" is not as surprising as a 'strange' result in one single attempt. You can use corrections like Bonferroni to make the "surprise-interpretation of p-values" in a screening more similar to that of single, planned comparisons. I don't see any particular advantage of this procedure, but some people like it.
Instead of a "validation" by qPCR you might consider to use the qPCRs to refine (improve) the estimates, using the estimates from the OpenArray panels as Bayesian priors.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling