[ADMB Users] NLS example

Ben Bolker bbolker at gmail.com
Fri Feb 22 14:27:14 PST 2013

  I *think* (but am not sure) that this is a case where R will do OK --
it is a case where apparently Excel does something really silly and fits
by using linear regression on log-transformed data.

  I'll work it up if I get a chance.


On 13-02-22 04:41 PM, dave fournier wrote:
> I'm sure R users would be interested in this!
> This example runs beautifully with
>     -crit 1.e-10
> It is part of some excel bashing on the R list in
> https://stat.ethz.ch/pipermail/r-help/2013-February/347920.html
> there is na example description
> http://www.itl.nist.gov/div898/strd/nls/data/mgh09.shtml
> Wonder how nls does in R
> tpl,dat and pin files are attached.
>> On 13-02-22 02:40 PM, dave fournier wrote:
>>> On 13-02-22 11:33 AM, Ben Bolker wrote:
>>> I agree that backwards compatibility is important.
>>>   We could go back to having a robust flag which would
>>> activate the options.   I think a lot of the existing trouble with this
>>> data
>>> set is with
>>> the bounds on tmpL and u.
>>> That was probably why sometimes -noinit would crash
>>> There will always be a new data set
>>> that will break something.
>>    Of course.  I just don't want to break *old* data sets, or at least I
>> want to allow people to be able to un-break things if the default
>> settings break them ...
>>>> On 13-02-22 02:28 PM, dave fournier wrote:
>>>>> On 13-02-22 07:26 AM, Ben Bolker wrote:
>>>>> Then I reran the model and the variances got stuck near 0.
>>>>> a small penalty for a while fixes that.  So I think a model
>>>>> something like this would be good.
>>>>>>       Hans, can you put this version of the glmmADMB TPL file up on
>>>>>> the SVN
>>>>>> repository so the buildbot will build it?  It includes a
>>>>>> beta-binomial
>>>>>> likelihood option and Dave's latest NB tweak ...
>>>>>>      (I probably do have write access to the SVN repository if I
>>>>>> could
>>>>>> figure it out, but I only have to do this a few times a year so I'm
>>>>>> being lazy ...)
>>>>>>      thanks
>>>>>>        Ben
>>>>     Thanks, Dave.  I will look this over and try to understand the bits
>>>> and pieces.  If possible I would like to try to make the upper bound on
>>>> the random effects values into a user-settable parameter (controlled by
>>>> admb.opts/admbControl) ...  Now that we have code that people are
>>>> using,
>>>> I worry a lot about making backward-incompatible changes (i.e. this
>>>> code
>>>> makes some models run OK, but breaks others that are out there in the
>>>> wild somewhere).
>>>>     cheers
>>>>       Ben

More information about the Users mailing list