error: glm.fit: algorithm did not converge

[kde...@gmail.com]

I have 96 samples I am trying to perform analyses on but get this error when I try to run in methylkit:

Warning messages:

1: glm.fit: algorithm did not converge 

2: glm.fit: algorithm did not converge 

Is there a limit to the sample size we can use in methylkit?  or do you have another suggestion?

Thanks

[Altuna Akalin]

it is not an error but a warning, I think it is related to perfect separation problem that has been asked and answered before in the forum, my suggestion is to ignore it.

--
You received this message because you are subscribed to the Google Groups "methylkit_discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to methylkit_discus...@googlegroups.com.
To post to this group, send email to methylkit_...@googlegroups.com.
Visit this group at http://groups.google.com/group/methylkit_discussion.
For more options, visit https://groups.google.com/groups/opt_out.

[Kerry (Deere) Machemer]

I guess I thought the convergence might be a problem as it was the only error I received.  I'm wondering why there were 0 entries from my Diff25p command.  I would have expected that just by chance I might have gotten at least 1 entry.  I have run similar analyses on other sample sets with no problems.  So, I'm stumped.  I've even tried a few different ways of looking at cases and controls (a few different phenotypes) all with the same result:

Getting my filtered (at least 10x coverage and present in all samples)

OSBMeth <- unite(filtered.OSBobj, destrand = FALSE) gives 53,339 entries

Performing the Diff calculation results in the same number of entries (as expected):

OSBDiff = calculateDiffMeth(OSBMeth) gives 53,339 entries

However, the Diff25p gives 0 entries:

OSBDiff25p <- get.methylDiff(OSBDiff, difference = 25, qvalue = 0.01)

results in 0 entries.  I only get the headers for an output:

"chr" "start" "end" "strand" "pvalue" "qvalue" "meth.diff"

--
You received this message because you are subscribed to a topic in the Google Groups "methylkit_discussion" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/methylkit_discussion/m2k7MD2vZ-4/unsubscribe.
To unsubscribe from this group and all its topics, send an email to methylkit_discus...@googlegroups.com.

To post to this group, send email to methylkit_...@googlegroups.com.
Visit this group at http://groups.google.com/group/methylkit_discussion.
For more options, visit https://groups.google.com/groups/opt_out.

--

Kerry (Deere) Machemer, Ph.D.
Department of Ecology and Evolutionary Biology
M154 Guyot
Princeton University
Princeton, NJ  08544-2016
609-258-3792
mach...@princeton.edu

[Altuna Akalin]

there are couple of things you can check

1) check the clustering of your samples

2) plot the histogram of methylation differences on the unfiltered methylDiff object sth like: hist(OSBDiff$diff.meth) to see the general effect size

3) use normalizeCoverage() and try recalculating the differential methylation stats.

These are some suggestions only, I haven't seen your data but there are people who used the package with similar amount of samples, I don't think it is a sample size issue.

Best,

Altuna