P-value of 1 for seemingly distinct PSI distributions

[Ben Carr]

Hello all,

Many thanks for creating such a useful tool! My lab makes quite heavy use of rMATS, but I've recently encountered a confusing piece of behavior regarding the calculation of p-values that I would love to get some clarification on. 

I have an event with an IncLevelDifference of -0.192, with 6 replicates in each sample group. The distribution of samples looks distinct when plotted, but rMATS reports a p-value of exactly 1. I understand it is possible to have a large effect size without statistical significance, but for this distribution to have a p-value of 1 seems rather mysterious.

I have uploaded a snippet (Groups won't let me attach it, so this is a Google Drive link) of the raw SE.MATS.JCEC.txt file, with the event in question (47518) alongside two other events from the same gene, which have more reasonable-looking p-values. I ran a Fisher's exact test as a quick sanity check on these events -- although this is of course not the exact calculation rMATS uses, the resulting value of p = 0.0002 for this event furthers my suspicion that the p = 1 finding is due to unusual behavior in the algorithm that calculates p-values. I have added this as a column to my minimal JCEC file for these three events. 

rMATS was run under default parameters with a paired design on mouse data. I attempted to run with the parameter --cstat 0 after reading this GitHub thread indicating that this allows for more significant events, but still got p = 1 for this event. Any insight into why this is occurring or suggested alternative parameters would be greatly appreciated!

Many thanks,

Ben Carr

[kutsc...@gmail.com]

https://github.com/Xinglab/rmats-turbo/issues/554