3
$\begingroup$

I am an analyst on a paper and, in writing up methods and results, noted that one of the proposed (logistic regression) models did not converge due to separation. I noted this in the results section of the paper, but the authors have asked that this be moved to the methods section.

It doesn't seem right to me, as their only justification for doing so is that it is a somewhat technical concept. But it turns out that several of the proposed regression variables were highly correlated. We wouldn't have known that before looking at the data. By that reasoning, it seems intuitive to report this in the results. But I think their issue is that even making mention of the technical aspects detracts from the results (to me, it enhances them).

Is there a resource to which I can defer that would argue in favor of presenting this evidence one way versus another? Either a methods type article or even an applied article where the authors did an exceptionally good job at describing a proposed analysis that failed to converge?

$\endgroup$
1
$\begingroup$

I found two articles in public health which approach this either way:

In results: http://jama.jamanetwork.com/article.aspx?articleid=193490

Even larger differences were observed for degree of breastfeeding, although extremely low rates of exclusive and, at 6 months, predominant breastfeeding in the control group led to GLIMMIX models that did not adequately converge and hence to unreliable estimates of the adjusted ORs.

In methods: http://nutrition.highwire.org/content/140/10/1832.full

The reduced 2-part model was used in all 3 strata for orange vegetables, because the full models did not converge.

I think the answer really is that it is a matter of style. From a strictly statistical perspective, I can only see reasons to report convergence failures in results.

$\endgroup$
  • 1
    $\begingroup$ My gut feeling reading this question was that convergence failure should only go in methods IF it helps clarify/justify the use of other methods or techniques. Your answer confirms that this seems to be what others do as well. $\endgroup$ – HFBrowning Jun 4 '14 at 20:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.