Likelihood ratio or wald test for meta-analysis

[statisticalgenetics]

Hi there!

This is my first time posting but I've always been a frequent lurker on previous posts that have been super helpful, so apologies if I'm not following any appropriate etiquette. I am working on an RNA sequencing project and looking have different outputs, which provides either a likehood ratio test statistic and p-value or wald test statistic and p-value. I prefer the LRT over Wald due to some subsets having smaller sample size. If it matters, this is an output from DESeq2.

I am currently looking into meta-analyzing different datasets and was wondering whether I would need to Z-score normalize my LRT statistic to do this? I recall learning it is similar to a chi-square distribution, so would it be okay to square it? Or should I use scale() in R? Or maybe there is a better way of meta-analyzing RNAseq datasets that produce LRT / Wald statistics, but I'm not too familiar.

Thank you all for your help! I truly appreciate it. Let me know if there is any clarification you may need.

 

[hlsmith]

I cant recall the rationale, but i believe i have heard Frank Harrel refence the preference of LR over Wald tests, at least in logistic regression.

Not sure of particulars for RNA seq stuff, is there risk of false discovery if too many associations explored. So later on you want to do meta-analytics with these results? I am guessing they are unitless, but unsure. I havent doneMA with anything but effect estimates like risks or odds ratios.

 

[GretaGarbo]

but i believe i have heard Frank Harrel refence the preference of LR over Wald tests, at least in logistic regression.

Yes, and in the book by Pawitan "In all likelihood" it says that a likelihood ratio test is better than a Wald test, since it gives a better approximation. (But in the normald distribution case they will give the same answer.)