R help - Jun 2017 - Changing the order of the factors in a cumulative logit model

Hi all,

I am using the clm function from the ordinal package to fit a cumulative
logit model. If I run the regression without formatting my dependent
variable (which is a factor), it works fine, however I need to change the
order of the factors to be able to interpret the estimates. I have tried
doing this using the relevel function and the factors are now in the
correct order, but when running the regression now, I get this:

Warning message:(1) Hessian is numerically singular: parameters are
not uniquely determined
In addition: Absolute convergence criterion was met, but relative
criterion was not met

Why is this and what can I do about it?

	[[alternative HTML version deleted]]

> On Jun 26, 2017, at 7:52 AM, Eva Wanek <eva.maria.wanek at gmail.com>
wrote:
> 
> Hi all,
> 
> I am using the clm function from the ordinal package to fit a cumulative
> logit model. If I run the regression without formatting my dependent
> variable (which is a factor), it works fine, however I need to change the
> order of the factors to be able to interpret the estimates. I have tried
> doing this using the relevel function and the factors are now in the
> correct order, but when running the regression now, I get this:
> 
> Warning message:(1) Hessian is numerically singular: parameters are
> not uniquely determined
> In addition: Absolute convergence criterion was met, but relative
> criterion was not met
> 
> Why is this and what can I do about it?

Without seeing your data or a small data sample that replicates the problem:

I would check for:

1. Structural zeros in the cross tabulation of your dependent variable (DV) and
your independent variable[s] (IVs).

2. Perfect prediction between your DV and IVs.

3. "Small" sample size in one or more of your DV levels.


If changing the ordering of your DV levels results in the error, you may have
one or more of the above going on in one or more of your DV levels, especially
if you shift one or more levels to the extremes (e.g first or last level).

How you handle this may depend upon domain subject matter knowledge and how to
deal with possibly collapsing zero or low sample size categories and how that
affects the relevance/applicability of the model.

Regards,

Marc Schwartz