[ADMB Users] MVN Example?

[Chris Gast]

Hello,

Does anyone have some code to compute the log-likelihood for a multivariate normal model that they're willing to share? I'm having trouble getting matrix determinants and inverses to work.

I found a link to a previous discussion, indicating such an example would be available in Roa's article "A Likelihood-Based Model of Fish Growth With Multiple Length Frequency Data" in JABES, but I'm having trouble accessing the supplementary material.

If I can get this working (a weighted LMM with autocorrelated residuals and spatially-correlated random effects), I'd be happy to add it to the examples list, since I can't seem to find a MVN likelihood example there either.

Thanks,

Chris

-----------------------------
Chris Gast
cmg...@gmail.com

[Mollie Brooks]

Hi Chris,

I'm doing an MVN model. My procedure section looks like this

PROCEDURE_SECTION

for(int i=1; i<=ncoeffs; i++)

{

coeffs(i)=ucoeffs(i)*scale(i);

}

nLL=0;

for(int p=1; p<=nponds; p++)

{

calc_predictions(p);

calc_varcov(p);

dvar_vector diff=predictions(p)-obs(p);

dvar_matrix tmp_varcov=varcov(p);

nLL+=.5*(nobs(p)*log(2*M_PI)+ln_det(tmp_varcov)+ diff*solve2(tmp_varcov,diff));

}

The first for loop scales all my coefficients to make it conditioned better.

Then I have observations of a bunch of ponds and each pond is assumed to be independent of the others.

I defined functions calc_predictions() and calc_varcov() to fill in those objects; the code in these is pretty specific to my problem. The important thing fir efficiency is that my covariates are in a design matrix created using model.matrix() in R.

Try using solve() where I have solve2(). If you have a problem with solve, I'll explain the complications of why I used this version.

I hope this helps.

cheers,

Mollie

Mollie Brooks
Ph.D. Candidate
NSF IGERT Fellow
Biology Department
University of Florida
mbr...@ufl.edu
http://people.biology.ufl.edu/mbrooks

_______________________________________________
Users mailing list
Us...@admb-project.org
http://lists.admb-project.org/mailman/listinfo/users

[Chris Gast]

Thanks, Mollie.  For some reason, ADMB can't seem to find the ln_det() function.

test_mat.obj:test_mat.cpp:(.text+0x13af): undefined reference to `ln_det(df1b2matrix&)'

I'll keep playing with it and see if I can figure it out.

Chris

-----------------------------
Chris Gast
cmg...@gmail.com

[Chris Gast]

After changing line 1978 in df1b2fun.h to match the correct function header on line 220 of  df1b2ludcmp.cpp, I got your syntax working successfully.

I'm not sure who specifically to address regarding making this correction, but hopefully they're reading this thread already....?

Mollie: I'm interested in what complications you experienced with solve(). Is this related to some speed issues which I found noted on previous admb-users posts?

Thanks Mollie (and Jim Ianelli for an off-list reply).

Chris

-----------------------------
Chris Gast
cmg...@gmail.com

[Rubén Roa]

Rubén H. Roa-Ureta, Ph. D.
AZTI Tecnalia, Txatxarramendi Ugartea z/g,
Sukarrieta, Bizkaia, SPAIN

Hi Chris,

I've told the JABES people that the link to sup. mat. is broken, to no avail it seems.
I'll send you the code later.

Ruben