[ADMB Users] ADMB and hierarchical (multi-level) models

[H. Skaug]

Hi Saang-Yoon,

Let me add one point that maybe is of use to you. If you
have your prediction problem with

Ynew = beta0 + beta1*X + error, where error ~ N(0, sigma2)

Since "error" is contiunous here, you can deal properly with this in ADMB.
You simply define error to be a random effect (see RE manual).
There is no likelihood contribution associated with Ynew, except
from that comming from "error". Hence your concern about
Ynew affecting the beta estimates is solved.

This may slow down you program severly, but uncertainty in Ynew will
be dealt with properly. That is: both the parameter uncertainty
and randomness in error is accounted for. If error is discrete
this approach does not work, however.

Hans


On Mon, Aug 16, 2010 at 11:59 PM, Saang-Yoon <shyunuw at gmail.com> wrote:
> Hello, Hans.
> Thank you very much for your comments.  By the way, in my example,
> newY is continuous.   In the multinomial component, newY is a
> continuous parameter.  In the regression component (expressed as
> Prior), newY is a continuous random variable.
> Saang-Yoon
>
> On Aug 16, 9:17 am, "H. Skaug" <hsk... at gmail.com> wrote:
>> Saang-Yoon,
>>
>> Because your newY is a discrete random variable, I do not think
>> the prediction problem fits well into ADMB. I would fit the model
>> in ADMB, and write a small back-end that does Monte Carlo simulation.
>>
>> >> Y = beta0 + beta1*X + error, where error ~ N(0, sigma2)