[ADMB Users] Examples of multilevel multiprocess model

[H. Skaug]

Dear Shige,

There does not exist an example that is similar to your model. The model
http://www.otter-rsch.com/admbre/examples/socatt/socatt.html
can serve as a template, and the ADMB-RE manual describes
how to set up correlated random effects (2.3.1 Correlated random effects).

However, I am concerned about your sample sizes. If you have individual
specific covariates, I think you will be limited to maybe 5000 mothers.
If you do not have covariates, or are to stratify your covariates,
you may be able to use the trick described in the section
"2.4.4 Frequency weighting for multinomial likelihoods" in the manual.
(This feature is not tested very well).

Hans

On Tue, Jun 2, 2009 at 1:17 AM, Shige Song <shigesong at gmail.com> wrote:
> Dear Hans,
>
> Many thanks for the response. I have about 40,000 mothers and about 150,000
> children. For each child, we know whether s/he has survived to age 10 and,
> if not, age at death. Now I want to estimate a joint model with three random
> intercept logistic regression on (1) if a child died at infancy (age at
> death <1), (2) if a child died in age 1-5, and (3) if a child died in age
> 6-10. It is easy to estimate these models separately, assuming Gaussian,
> non-Gaussian, or a non-parametric heterogeneity at mother-level. The
> difficulty arises when I try to estimate these equations jointly. For
> multivariate Gaussian heterogeneity, there is aML and Sabre (in theory
> GLLAMM can also do it but it is too slow to be practical); but none of them
> handles multivariate non-Gaussian heterogeneity. ADMB seems to be the only
> hope at this moment (besides programming the whole thing in Fortran).
>
> In this particular case, the joint model is important because the
> correlations between the heterogeneity terms are of substantive interest.
>
> Best,
> Shige
>
> On Mon, Jun 1, 2009 at 11:10 PM, H. Skaug <hskaug at gmail.com> wrote:
>>
>> Shige,
>>
>> multivariate log-normal for the unobserved hetegeineity is OK, but you
>> need to give us a brief summary of your model, including the number of
>> "individuals", if you want more detailed feedback
>>
>> The largest collection of latent variable models in ADMB is found here:
>>
>> http://www.otter-rsch.com/admbre/examples.html
>>
>> Hans
>>
>>
>> On Sun, May 31, 2009 at 12:50 AM, Shige Song <shigesong at gmail.com> wrote:
>> > Dear All,
>> >
>> > I am trying to port an analysis from aML to ADMB. This is a multilevel
>> > multiprocess model (some people call it multivariate multilevel model)
>> > of
>> > child mortality. I am jointly modeling infant mortality (death before
>> > age
>> > one), mortality between age 1-10, and mortality between age 11-20.
>> > Because
>> > aML assumes a multivariate normal distribution for the unobserved
>> > heterogeneity terms, some reviwers think my results are minly driven by
>> > this
>> > distributional assumption. I would like to try alternative distribution
>> > assumption for these unobserved hetegeineity terms such as multivariate
>> > log-normal or multivariate t. ADMB seems to be the only softare choice
>> > for
>> > this.
>> >
>> > It will be really great if there is an example of multilevel
>> > multiprocess
>> > model using ADMB so that I can tailor it to suite my own research needs.
>> > Can
>> > anybody help me with this?
>> >
>> > Many thanks.
>> >
>> > Best,
>> > Shige
>> >
>> > _______________________________________________
>> > Users mailing list
>> > Users at admb-project.org
>> > http://lists.admb-project.org/mailman/listinfo/users
>> >
>> >
>
>
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>
>