[ADMB Users] Beta deviate

[Coilin Minto]

Hi Dave,

Thanks very much for the code and explanation of the truncation.

Running beta_deviate(x,a,b,eps) returns:

Illegal value in inv_cumd_beta

beta_deviate(a,b,x,eps) runs.

Doing a search for "beta_deviate" in the source returns:

df1b2-separable/df1b2betdev.cpp:df1b2variable beta_deviate(const
df1b2variable& _a,const df1b2variable& _b,
df1b2-separable/df1b2fun.h:df1b2variable beta_deviate(const
df1b2variable& _x,const df1b2variable& _a,
linad99/cbetdev.cpp:double beta_deviate(double a,double b,double x,double eps)
linad99/fvar.hpp:dvariable beta_deviate(const prevariable& _x,const
prevariable& _a,
linad99/fvar.hpp:double beta_deviate(double x,double a,double b,double
eps=1.e-7);
linad99/vbetdev.cpp:dvariable beta_deviate(const prevariable& _a,const
prevariable& _b,

So a mixture of positioning. inv_cumd_beta takes (a,b,x) in all cases
in the source.

Apologies for the back an forth, its a really useful general method
and I hope this contributes to the code.

Best regards,
Coilin

On Wed, Aug 1, 2012 at 1:01 AM, dave fournier <davef at otter-rsch.com> wrote:
> On 12-07-31 11:02 AM, Coilin Minto wrote:
>
> One last thing. I checked the docs you referenced on beta_deviate.  They are
> not quite accurate.
>
> The function is
>
>   beta_deviate(x,a,b,eps)
>
> where eps has a default value of 1.e-7
>
> The reason for eps is that for some values of a or b  the function and its
> derivatives
> go to infinity as x goes to 0 or 1. This can cause instability in the
> estimation procedure.
>
> To avoid this I added to option to fit a truncated beta which runs from eps
> to 1-eps
> The general ideas is that to avoid instability if it occurs one does the
> fitting in a number of stages
> and shrinks eps toward 0 after starting with something like eps=.001 at the
> first stage.
>
>
>
>> Hi Dave,
>>
>> Thank you very much for the code.
>>
>> I included the globals section:
>>
>> GLOBALS_SECTION
>>   #include <betai.cpp>
>>   #include <dfbetai.cpp>
>>   #include <df1b2invcumdbeta.cpp>
>>
>> which gave the linking error:
>>
>> g++ -s -L"/usr/local/admb"/lib -L"/usr/local/admb"/contrib fmi.o
>> -ldf1b2o -ladmod -lcontribo -ladt -lado -ldf1b2o -ladmod -ladt -lado
>> -lcontribo -o fmi
>> fmi.o: In function `betacf(prevariable const&, prevariable const&,
>> double, int)':
>> fmi.cpp:(.text+0x4a16): undefined reference to
>> `dvariable::operator=(df1_two_variable const&)'
>> collect2: ld returned 1 exit status
>>
>> Commenting out:
>>
>> //#include <dfbetai.cpp>
>>
>> works, with the minor change of placing the random effect in the third
>> position
>>
>> beta_deviate(a,b,u);
>>
>> in contrast with the ADMB-RE manual.
>>
>> It might be interesting to to get an example/benchmark together based
>> on the the binomial-beta model in Lee and Nelder (1996) using the data
>> in Crowder (1978).
>>
>> Thanks very much for your time and help.
>>
>> Best regards,
>> Coilin
>>
>> References:
>>
>> Crowder, M.J. (1978). Beta-binomial anova for proportions. Applied
>> Statistics, 27(1), 34-37.
>>
>> Lee, Y. and Nelder, J.A. (1996). Hierarchical generalized linear
>> models. Journal of the Royal Statistical Society B, 58(4), 619-678.
>>
>>
>> On Mon, Jul 30, 2012 at 9:16 PM, dave fournier <davef at otter-rsch.com>
>> wrote:
>>>
>>> Its really frustrating. I put a hell of a lot of work into that function
>>> and
>>> it got wrecked.
>>> in the open sourcing fiasco.   I have managed to get it to work again by
>>> purging a lot of the
>>> cephes code and reintroducing the old free world code.
>>>
>>> You can try compiling and linking these three files into your code and
>>> see
>>> what happens.
>>>
>>> _______________________________________________
>>> Users mailing list
>>> Users at admb-project.org
>>> http://lists.admb-project.org/mailman/listinfo/users
>>>
>