I'm trying to run a rather complicated model in DESeq2 with 16S microbiome data. More specifically, the model is "~ confounder + diseaseStatus:subject + timepoint*diseaseStatus" (a two-timepoint case-control comparison, with the same subjects at both timepoints; all variables are factors, except for the confounder, which is numeric). Probably because of how complicated this is, I keep running into the "rows did not converge in beta, labelled in mcols(object)$betaConv. Use larger maxit argument with nbinomWaldTest" issue. Now, this wouldn't be such a big problem otherwise, but many of the rows that don't converge in beta represent microbial taxa that I am very much interested in, and would like to have results for.
First of all, the output still includes a full set of results, complete with p-values, for the rows that were labeled as non-converging. Are these good for anything, or should I just ignore them completely? Some old discussions I found on this topic suggest deleting all the non-converging rows from the output. The DESeq2 documentation mentions a useOptim parameter, "whether to use the native optim function on rows which do not converge within maxit", for nbinomWaldTest, which is by default TRUE, and I assume is relevant to what's going on here, but I don't really understand what this means.
Secondly, is there any way to get around this issue? I've looked at everything google/earlier conversations suggest as solutions for this, which mostly comes down to trimming the data and increasing the maxit value for nbinomWaldTest. I've tried both, and neither trimming the data aggressively nor increasing the maxit several orders of magnitude (from the default 100 all the way to 1 000 000) help lower the number of non-converging rows. I'd be extremely thankful for any additional ideas that might help.
(I guess it might come down to "run a simpler model". I'm just fond of running it as one comparison using all the samples at once, instead of subsetting it by timepoint or some such.)
First of all, the output still includes a full set of results, complete with p-values, for the rows that were labeled as non-converging. Are these good for anything, or should I just ignore them completely? Some old discussions I found on this topic suggest deleting all the non-converging rows from the output. The DESeq2 documentation mentions a useOptim parameter, "whether to use the native optim function on rows which do not converge within maxit", for nbinomWaldTest, which is by default TRUE, and I assume is relevant to what's going on here, but I don't really understand what this means.
Secondly, is there any way to get around this issue? I've looked at everything google/earlier conversations suggest as solutions for this, which mostly comes down to trimming the data and increasing the maxit value for nbinomWaldTest. I've tried both, and neither trimming the data aggressively nor increasing the maxit several orders of magnitude (from the default 100 all the way to 1 000 000) help lower the number of non-converging rows. I'd be extremely thankful for any additional ideas that might help.
(I guess it might come down to "run a simpler model". I'm just fond of running it as one comparison using all the samples at once, instead of subsetting it by timepoint or some such.)