"Didn't converge" error

[Nick]

Can anyone shed light on a "Didn't converge" error received when
running qtl scans (scantwo, fitqtl, etc.) using method="hk"?

[Karl Broman]

With method="hk" this would only come up with model="binary" (that is, for binary traits). In that case, we're fitting a logistic regression model, which requires an iterative algorithm. I would call the "Didn't converge" message is a warning rather than an error. It indicates that at some position or pair of positions (or for the given model in fitqtl), the parameter estimates or likelihood hadn't reached the chosen level of tolerance for convergence (default is tol=1e-4) within the maximum number of iterations (default maxit=1000).

You can try increasing maxit, say to 10000. And this would be particularly important for fitqtl, in which you are interested in a specific multiple-QTL model. But in my experience, the failure to converge mostly comes up in scanone or scantwo and seems to be at positions without good evidence for a QTL, and so lack of convergence is not really a big deal.

karl

On Apr 5, 2011, at 9:30 AM, Nick wrote:

> Can anyone shed light on a "Didn't converge" error received when
> running qtl scans (scantwo, fitqtl, etc.) using method="hk"?
>

> --
> You received this message because you are subscribed to the Google Groups "R/qtl discussion" group.
> To post to this group, send email to rqtl...@googlegroups.com.
> To unsubscribe from this group, send email to rqtl-disc+...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/rqtl-disc?hl=en.
>