[mbblack]

Thank you mbblack for your kind and very valuable reply.
As I understood, replicates in RNA-seq are essential. Apart from RNA-seq, however, may I ask about the replicates in whole genome sequencing of human samples. I am planning to with 3-technical replicates in WGS along with 2-biological replicates. do you think its too much for WGS?
Sincerely
jp.

What is the purpose of this sequencing project? What is the question(s) you want to answer with it? Your hypothesis should determine what sampling strategy you need, and how much data, number of replicates and so forth.

Any population (whatever your "population" is) based study needs to sample the population(s) in question.

[mbblack]

Is it correct to call the between sample differences in measurement variance? To my understanding variance is supposed to be "location-invariant".

The differences are not the variance. When you perform an RNAseq experiment with the intention of doing differential gene expression analysis, you presumable have at least two populations you wish to compare. You need to then be able to compute a mean and variance for the genes in each population.

Your individual tests are then based on the difference in the means, but it is the variance about those paired means that determines the statistical significance of that difference. Without replicates from each population, you do not have mean differences, you just have a difference between two numbers. And without a variance to go with the numbers that make up that difference, you have no basis for assessing statistical significance.

Your individual tests are basically asking the question "given the observed variation about the means, how likely is the observed difference between the means due to chance alone". If you don't know that variation, you cannot even ask the question.

[rskr]

The differences are not the variance. When you perform an RNAseq experiment with the intention of doing differential gene expression analysis, you presumable have at least two populations you wish to compare. You need to then be able to compute a mean and variance for the genes in each population.

Your individual tests are then based on the difference in the means, but it is the variance about those paired means that determines the statistical significance of that difference. Without replicates from each population, you do not have mean differences, you just have a difference between two numbers. And without a variance to go with the numbers that make up that difference, you have no basis for assessing statistical significance.

Your individual tests are basically asking the question "given the observed variation about the means, how likely is the observed difference between the means due to chance alone". If you don't know that variation, you cannot even ask the question.

That may be true for some distribution. The wiki says variance has to be the same throughout the range of measurements, and as pointed out without a variance there are no p-values. And in some places you can be put in jail for pretending they are.

[mbblack]

That may be true for some distribution. The wiki says variance has to be the same throughout the range of measurements, and as pointed out without a variance there are no p-values. And in some places you can be put in jail for pretending they are.

Sorry. I am not following what you mean by this post at all.

[rskr]

Sorry. I am not following what you mean by this post at all.

Take VOOM for example. I just think it is funny that after all the years of listening to biologists whine about those noisy low expressed genes with "log-normal" distribution that someone wrote a paper that suggested they model the variance, when its pretty easy to figure out that taking the log of the data amplifies the variance in the low range, meanwhile if you try to do statistics with some other distribution, the reason that statisticians are so biased towards the normal, is that there aren't any other tools to use besides that one ANOVA package. Furthermore, I am pointing out that it could be considered fraudulent in the context of a clinical trial to continue to produce statisics based on a normal when the variance is known to be heteroskedastic.

[SpreeFu]

q-value = 0.000148708

bump?

Anyone else getting strange q values??

Yes, I got hundreds of q-value = 0.000148708

Any one knows why?