I have been reading about Cuffdiff, and this is what I understand of its variance calculation. It seems to fit a negative binomial distribution, calculating the variance across samples by using the read counts with similar expression levels. If there are multiple isoforms, it also includes a parameter representing its own uncertainty in the assignment of isoforms. Is this correct?
What I am having more trouble understanding is when the importance sampling comes in, or if indeed the program uses importance sampling anymore (I read somewhere that it did not). Can anyone shed some light on this for me? When is this incorporated?
I am concerned because I read somewhere that this used to cause frequent failure of significance tests, and I will be using a data set with a lot of variability.
What I am having more trouble understanding is when the importance sampling comes in, or if indeed the program uses importance sampling anymore (I read somewhere that it did not). Can anyone shed some light on this for me? When is this incorporated?
I am concerned because I read somewhere that this used to cause frequent failure of significance tests, and I will be using a data set with a lot of variability.