DEXSeq fitDispersionFunction error

[id0]

I am trying out DEXSeq. I've used it successfully a couple of times, so my seth should be okay. However, for one of my projects, I get the following error:

> ecs <- fitDispersionFunction(ecs)
Warning message:
In glmgam.fit(mm, disps[good], coef.start = coefs) :
  Too much damping - convergence tolerance not achievable

What exactly does this mean? Is it okay to ignore the warning and proceed? Is there a way to fix it?

[Simon Anders]

It means that the function glmgam.fit, which DEXSeq uses internally to perform the fit, is unsure whether the fit is accurate. Hence, use the 'plotDispEst' function to manually check the fit quality. If it looks good (red line goes through the upper part of the top cloud of dots), you can proceed.

[id0]

The plotDispEsts graph looks good (red line through the upper cloud).

I assume that sometimes it doesn't. Wouldn't you be stuck then?

[areyes]

Never heard it happen, but if it does I would go back and double check data quality.

[Simon Anders]

I assume that sometimes it doesn't. Wouldn't you be stuck then?

You could still ask for help here in that case.

This is not that unusual in statistics: No analysis method will ever be perfectly general, and there will always be cases where it fails. This is why diagnostic plots are important: They help you determine whether your chosen analysis method seems to work for your data. If it does not, you are stuck, of course, but at least you know that you should not trust the results. And that you need to look for work-arounds or alternatives. Be wary of methods that claim to always work.