Question

[Otilia]

I want to fit a bivariate 3-component normal mixture with MIXMOD, but
the estimates for the variances of each component don't make sense for
simulated data that else can be estimated quite well EMMIX. I am
trying to fit a model with

a) 3 unequal covariance matrices
b) 3 equal covariance matrices.

when I do a), I get that the covariance structure i am fitting is

.running
|-| pk_Lk_C (k=3)

Does anyone know - what does this structure above mean, or what
command i should write to get a structure a)? Could not follow either
from Statistical or Full manual for MIXMOD.

Otilia

[Florent Langrognet]

Otilia a écrit :

In Mixmod, you can give one (or several) gaussian model(s) among the 14
gaussian models.
See : http://www-math.univ-fcomte.fr/mixmod/statdoc_2_1/node30.html

If you don't specify one model, pk_Lk_C will be used (the default
gaussian model).
This model works well in more cases but not in your case because your 3
components have unequal covariances matrices.

Choose the most general model (pk_Lk_Ck) for your structure a)

See the userguide to do this.
If you want I help you, you have to tell me how do you use Mixmod ? In
Scilab environment ? In Matlab environment ?

Florent



--
Florent Langrognet
Ingénieur de recherche CNRS
Laboratoire de Mathématiques de Besançon - UMR 6623
16, route de Gray
25030 Besançon
Tel : 03 81 66 63 19
web : http://www-math.univ-fcomte.fr/pp_Annu/FLANGROGNET/
Mixmod software : http://www-math.univ-fcomte.fr/mixmod

[Otilia]

Thank you for responding, this really helps. I have trouble linking
the notation of the 14 Gaussian models (e.g. pk_Lk_C) with a certain
decomposition in http://www-math.univ-fcomte.fr/mixmod/statdoc_2_1/node30.html,
since the notation there is in terms of lk, A, D, no pk and C!.

So pk_Lk_C are equal covariance matrices and pk_Lk_Ck are unequal
covariance matrices?

Thanks
Otilia

On Sep 26, 4:32 pm, Florent Langrognet <florent.langrog...@univ-

>  florent_langrognet.vcf
> < 1KViewDownload

[Otilia]

i am using MIXMOD as a function in MATLAB.

On Sep 26, 4:32 pm, Florent Langrognet <florent.langrog...@univ-
fcomte.fr> wrote:

>  florent_langrognet.vcf
> < 1KViewDownload

[Otilia]

Fixed it and found out also what is pk / Lk /Ck from the CSDA, 2006
paper. I used:

model=struct('name','','subDimensionFree',[],'subDimensionEqual',[]);
model.name='Gaussian_pk_Lk_Ck';
out = mixmod(X,p,'tabmodality',[],'model',{model});

and it worked.

Thank you!!!

Otilia

On Sep 26, 4:32 pm, Florent Langrognet <florent.langrog...@univ-
fcomte.fr> wrote:

>  florent_langrognet.vcf
> < 1KViewDownload

[Gilles...@math.u-psud.fr]

Hi Otilia

>
> Thank you for responding, this really helps. I have trouble linking
> the notation of the 14 Gaussian models (e.g. pk_Lk_C) with a certain
> decomposition in
> http://www-math.univ-fcomte.fr/mixmod/statdoc_2_1/node30.html,
> since the notation there is in terms of lk, A, D, no pk and C!.
>
> So pk_Lk_C are equal covariance matrices and pk_Lk_Ck are unequal
> covariance matrices?
>
>

That is (almost) correct.

A feature of model Pk_Lk_C is that the volumes of the clusters
(the determinant of the variance matrices) are free.
It allows for clusters with different volumes. It leads
to parsimonious and versatile models.

However, if you want a model with strictly equal covariance matrices,
it is the model
pk_L_C.

All the best

Gilles