Question: DESeq2 modelling and Wald's test
gravatar for Uday Rangaswamy
19 months ago by
Indian Institute of Technology, Madras, India
Uday Rangaswamy120 wrote:

Hi all,

Can you please explain to me the relation between Wald's test and Negative binomial generalized linear model? As for my understanding, the count data is modeled using negative binomial generalized linear model after which Wald's test is applied to figure out whether a particular gene is significant or not. Please correct me if I'm wrong.

rna-seq deseq2 • 1.9k views
ADD COMMENTlink modified 19 months ago by Kevin Blighe51k • written 19 months ago by Uday Rangaswamy120
gravatar for Kevin Blighe
19 months ago by
Kevin Blighe51k
Kevin Blighe51k wrote:

RNA-seq raw count data 'naturally' follows a negative binomial distribution (Poisson-like), so, the DESeq2 authors model the data as such. By 'model the data', we merely imply that we build a regression model of the data such that we can make statistical inferences from it [the data].

So, after normalising the raw counts, the following occurs:

For each gene, a logistic regression model with the negative binomial as family is fit:

gene1.model <- glm.nb(gene1 ~ CaseControl + ..., data=MyData)
gene2.model <- glm.nb(gene2 ~ CaseControl + ..., data=MyData)
*et cetera*

Once we have modeled each gene, a simple way to derive a P value for each model coefficient (i.e. CaseControl, etc) is by applying the Wald Test and selecting the coefficient of interest:

wald.test(b=coef(gene1.model), Sigma=vcov(gene1.model), Terms=c(2)) #term '2' would be CaseControl

The Wald test is a standard way to extract a P value from a regression fit.


NB - this is not the exact code used by DESeq2, of course. This is just giving you a broad overview with some simple R functions. For one, DESeq2 models dispersion in addition to everything that I have mentioned above, and the Wald test is not used in each case to derive p-values in DESeq2.

ADD COMMENTlink modified 7 months ago • written 19 months ago by Kevin Blighe51k

Thank you so much for that.

ADD REPLYlink written 19 months ago by Uday Rangaswamy120

Kevin, thanks for your explanation. Let me one naive question please? Why we need to make a GLM model before performing a Wald test itself (as i can understand it's just a simple t-test in rough approximation?)? Why not just perform a Wald test on count data?

ADD REPLYlink written 5 months ago by Denis140

A Wald test requires a coefficient and its standard deviation, which are tested for difference from 0. Yes, in a way that's sort of like a single group T-test, but you'd still need to perform a fit first in order to derive the coefficient.

ADD REPLYlink written 5 months ago by Devon Ryan92k

Sorry, but it's stiil not quite clear for me. I thought that T-test for two (in majority experiments) independent groups would be more intuitive and obvious. So, why a single group T-test? In addition, when i perform T-test for example in R i don't need anything except trait observations in two groups. In particular, no coefficiets are required for that. So why in DESeq2 i need to estimate some coefficient before the test itself? Could you elaborate on this please (just for understanding)? Finally, as i understand the GLM output already contains p-values. So why just not use these ones to test for DE genes?

ADD REPLYlink modified 5 months ago • written 5 months ago by Denis140

We're talking about different things with the single-group T-test, just ignore that.

You can do a standard T-test for RNAseq data, your power will just be terrible. Packages like limma were developed to get around this and can be used with RNAseq. You estimate a coefficient with a T-test too (a T-test is a kind of GLM), you're just not aware of it. The p-values from the GLM (in particular, summary()) are from a Wald test. You don't have to explicitly run that separately, but it can be convenient to do so to more easily extract the information you want (it also allows more flexibility, wherein you can use contrasts).

ADD REPLYlink written 5 months ago by Devon Ryan92k
Please log in to add an answer.


Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 2.3.0
Traffic: 1969 users visited in the last hour