[R] Problems with clmm2 (ordinal data fit)

[Rune Haubo]

Dear Caroline,

Yes, it seems you have complete separation for the 'Timepoint'
variable. This means that the likelihood is unbounded for that
parameter and the optimizer just terminates when it gets far enough
out on an asymptote and improvements are below a threshold. This is
also the reason the variance-covariance matrix of the parameters i
singular, and so standard errors and summary-p-values are not
available. This doesn't mean that you cannot perform the likelihood
ratio test though, so the anova() comparisons should not be
misleading.

I would recommend, though, that (1) if you really want to fit the
mixed-effects models, use the newer implementation in clmm() instead
of clmm2(), and that you (2) fit your models without the
random-effects term, since the variance of that is clearly zero (and
hence use clm()) - this is going to add to the instability of the
optimization and singularity of the variance-covariance matrix of the
parameters.

Hope this helps,
Rune

On 16 March 2014 19:45, Caroline Lustenberger
<caroline.lustenberger at hotmail.com> wrote:
> Dear all
>
> I have ordinal data (from a questionnaire with 4 levels) for 2 groups (30 subjects each) and 2 timepoints. So I used a cumulative link mixed model to fit the data (nr = subject number).
>
> mod_FV<-clmm2(FV~GruppeVerbelendung+Timepoint,random=nr,data=data,Hess=TRUE,nAGQ=10,na.action=na.omit)
> mod_FV1<-clmm2(FV~GruppeVerbelendung,random=nr,data=data,Hess=TRUE,nAGQ=10,na.action=na.omit)
>
>
> For some questions (e.g. FV) I had the following output with NANs instead of p-values
>
> Call:
> clmm2(location = FV ~ GruppeVerbelendung * Timepoint, random = nr,
>     data = data, na.action = na.omit, Hess = TRUE, nAGQ = 10)
>
> Random effects:
>             Var      Std.Dev
> nr 5.881958e-07 0.0007669392
>
> Location coefficients:
>                                Estimate Std. Error z value Pr(>|z|)
> GruppeVerbelendung1            -0.0178      NaN        NaN NA
> Timepoint1                     18.6316      NaN        NaN NA
> GruppeVerbelendung1:Timepoint1  0.3505      NaN        NaN NA
>
> No scale coefficients
>
> Threshold coefficients:
>     Estimate Std. Error z value
> 1|2 20.7741      NaN        NaN
> 2|3 20.9486      NaN        NaN
>
> log-likelihood: -24.01783
> AIC: 60.03566
> Condition number of Hessian: 5479162.40
> (4 observations deleted due to missingness)
> Warnmeldung:
> In sqrt(diag(vc)) : NaNs wurden erzeugt
>
> However when comparing the models (mod_FV, mod_FV1) I obtain a usable result (p-val) even though the models seem not to be a good fit (high AIC and condition number of Hessian).
>
>  anova(mod_FV,mod_FV1)
> Likelihood ratio tests of cumulative link models
>
> Response: FV
>                                  Model Resid. df -2logLik   Test    Df LR stat.      Pr(Chi)
> 1             GruppeVerbelendung |  |        112 58.44384
> 2 GruppeVerbelendung + Timepoint |  |        111 27.78259 1 vs 2     1 30.66125 3.072406e-08
>
> What could be the problem? Timepoint 1 was always scored with level 1 (60 of 60 subjects had  scored 1) and during timepoint 2 subjects scored level 1-3. Might it be a problem that I only have 1 type of scoring level for timepoint 1?
> Can I use the results obtained by Likelihod ratio test for clm even if the model was not a good fit? What could I do instead?
>
>
> Thank you so much for your help and all the best
> Caroline
>
>
>
>
>
>         [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.