NaNs in standard errors of LFC

[emdann]

Hey @const-ae, I'm using the compute_lfc_se parameter in test_de to get the standard error of the log fold changes to run ash/mash.

In my results I see quite a few genes where the output lfc_se is equal to NaN. Is this expected behavior? In what cases does this happens? I couldn't quite figure it out from the code or documentation.

[const-ae]

Hey, thanks for reporting the issue. The compute_lfc_se is one of the less used features of glmGamPoi, so it is particularly useful to get feedback on this :)

The standard error is calculated in https://github.com/const-ae/glmGamPoi/blob/devel/R/predict.R#L204. As you can see, I am catching any type of error and just return NA's if anything goes wrong. Can you say a bit more about your input? Are you using the ridge penalty? Is your design potentially colinear? Anyways, I would suggest to debugonce(predict.glmGamPoi ) before you call test_de pick the index of a gene where the lfc_se is NA and inspect what error leads to the NA's.