pval and padj

[oomvel]

Dear all,

I am working on the differential expression of miRNAs using DeSeq program. Most of the miRNAs showed pval more than 0.05 with padj value 1.

What does the padj value 1 really mean? Can I consider the candidates with pval less than 0.01.

Thank you,

Velmurugan

[blanco]

Hi oomvel,
padj means adjusted p-value in the sense it is adjusted for multiple testing.

I am perhaps not the most experienced user here to answer this but, when you do a lot of statistical tests and get a p-value for each of them the p-value sort of loses its original meaning and it is necessary to make adjustments to make it more stringent.

One method to do this is called the Bonferroni correction and it simply multiplies the p-value with the number of tests (if you do 50000 tests and get a p-value of 0.0000004 then its adjusted p-value would be 0.02). This method was/is considered too conservative by most and the most popular adjustment method used now is called benjamini-hochberg correction. I don't know the details of that but that is the one which is used in DESeq and Cuffdiff.

The rule of thumb is to take padj <= 0.05 as meaning statistically significant.

[mbblack]

Dear all,

I am working on the differential expression of miRNAs using DeSeq program. Most of the miRNAs showed pval more than 0.05 with padj value 1.

What does the padj value 1 really mean? Can I consider the candidates with pval less than 0.01.

Thank you,

Velmurugan

Whenever you do a series of multiple, simultaneous statistic inferences, the number of false positives (or Type I errors, rejecting the null hypothesis when it is in fact true)) will be inflated when comparing all the individual test statistics as a whole. In other words, if you do enough tests, by chance alone, you will find some that appear to support your test hypothesis. In genomics, since we are often doing many thousands or tens of thousands of simultaneous statistical tests, the problem is not trivial.

The adjusted P-value is an attempt to correct for that affect. The most commonly used method is to correct for False Discovery Rate. Instead of focusing on the individual test statistic's p-Value, FDR attempts to control the expected proportion of incorrectly rejected null hypotheses amongst all the rejected hypotheses.

There are numerous ways of trying to compute an FDR, the two mentioned in the first reply being common, and the BH method (also referred to as a step-up p-Value in some software) is one of the most common used in genomic analyses.

So when looking at FDR, you need to decide on what proportion of false positives you are willing to accept (your "alpha" cutoff value), then reject your null hypothesis based on that, which will ensure that your actual type I error amongst all your significant tests is < "alpha". So if you use a significance cutoff of FDR < 0.05, less than 5% of all your significant results would be expected to be false positives.

P.S. an alternative approach is the FWER (Family-Wise error rate) which computes the actual probability of at least one Type I error amongst a set of simultaneous comparisons.