[ADMB Users] Integrating R and ADMB to generalize TPL files

[H. Skaug]

Hi,

Dave is right that the it is hard for the average user of glmmADMB to
understand the implication of the independence assumption. Maybe
the default should instead be to calculate a full covariance matrix,
i.e. we should change the default which currently is corStruct="diag".
Then the user must actively choose independence, and hopefully will think
through what it means.

Hans

On Wed, Nov 7, 2012 at 6:49 PM, dave fournier <davef at otter-rsch.com> wrote:
> On 12-11-07 09:29 AM, Ben Bolker wrote:
>>
>> On 12-11-07 11:58 AM, dave fournier wrote:
>>>
>>> Here is an example of why I dislike the R approach
>>> so much.
>>>
>>> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2012q4/019349.html
>>>
>>> I was curious to find out why introducing to additional parameters for
>>> the
>>> variance of the random effects would produce no improvement in the
>>> model.  So first I wanted to run glmmadmb as seth had done.
>>>
>>> I put his data in a file and read it into R.
>>>
>>> Then I ran his model by grabbing his call to glmmadmb.
>>> I was surprised to see that I got a different result from Seth.
>>>
>>> Examination of the design matrix and I realized that one of the columns
>>> of
>>> the data needed to be changed to a factor.  If I had not got the
>>> different result I would have thought that the analysis was correct.
>>> That was the first gotcha.  With R you never really know what
>>> you are doing.   This was for the simpler model with paramerter for
>>> the variance of the RE's.
>>
>>    I did see that, and I agree there is a price to pay for convenience.
>
>
> It is not convenient to effortless get to the wrong answer,
> just a bit of a waste of time.
>>>
>>> Then I used Seth's R script to run the more complicated model.
>>> He was right that the likelihood did not change.  Strange!
>>>
>>> So I examined the design matrix for the RE's.
>>> The problem was obvious.  rather than a different
>>> parameter for each time period the design matrix had
>>> an overall mean and differences from it.  While this
>>> would be the same for fixed effects it is not equivalent
>>> for random effects because it assumes a different covariance
>>> structure.
>>>
>>> I hacked together the correct model by hand and ran it on his data.
>>
>>    I agree that the model Seth was specifying was that there is variation
>> among trees in their time-1 response; variation in the difference
>> between time 1 and time 2; and variation in the difference between time
>> 1 and time 3 (unless I'm wrong, which has been known to happen).
>
> All I know is that he said
>
> The data suggest that variance may differ
> sharply before and after treatment (factor variable 'time', values 1/2/3),
>
> I don't see anything about differences just that the variance should depend
> on time.