Thank you Gilbert! I'll be on the lookout for your email.
Best,
Dan
On Thu, Feb 9, 2012 at 3:11 PM, gilbert.mackenzie
<[log in to unmask]> wrote:
> Dear Dan
>
> You can do it it R already - I do not know what the upper
> limit on the number of variable is - I will send you some code tomorrow
> when I tidy
> the R script file up.
>
> I gave a simple course on it at the ASSESS meeting in York in
> October last and Susana gave a related talk.
>
> Best
>
> Gilbert
>
>> Hi Susana,
>>
>> Thanks for your response. A few points of clarification (I'm afraid my
>> initial posting may not have been clear).
>>
>> 1) There are only 3 variables in the example I provided. HOWEVER, I am
>> interested in developing code to do this for any number of variables
>> for future applications.
>>
>> 2) The variable list will be a mixture of continuous and categorical
>> explanatory variables.
>>
>> 3) I was planning on using hit rates as my primary criterion for model
>> selection instead of traditional model fit statistics (although I am
>> currently gathering and tracking model fit stats too).
>>
>> 4) I already have SAS code to do all this for all possible main
>> effects models. I just need to generalize it to consider all possible
>> models including the interaction terms (apparently, I should have
>> taken more computer science and combinatorics classes...).
>>
>> Any ideas?
>>
>> Thanks!
>>
>> Dan
>>
>>
>>
>>
>> On Thu, Feb 9, 2012 at 10:53 AM, Susana Conde<[log in to unmask]> wrote:
>>>
>>> Dear Abner,
>>>
>>> In non-hierarchical models, the Wald test statistics of the parameters
>>> depend on the design matrix coding (i.e. one may find that they are
>>> non-zero
>>> with some code, and then some of them happen to be zero with a different
>>> code, for example).
>>>
>>> As it seems that you have "only" three variables, there are 19 possible
>>> hierarchical models (provided that, if any interaction is added in the
>>> model, all the interactions that include exactly the same variables but
>>> considering different categories are also added): the 18 that are written
>>> in
>>> your list below plus the null model).
>>>
>>> Gilbert and I would recommend to work only in the hierarchical class, if
>>> possible.
>>>
>>> Best wishes,
>>>
>>> Susana
>>>
>>>
>>>
>>> On 09/02/2012 12:04, gilbert.mackenzie wrote:
>>>>
>>>> Dear Abner
>>>>
>>>> The technology is in R
>>>>
>>>> I will send you some details.
>>>> Going out now.
>>>>
>>>> Recall that non-hierarchical models are of NO
>>>> scientific interest as the results are not invariant
>>>> to the choice of design matrix..
>>>>
>>>> Susana is expert.
>>>>
>>>> Best
>>>>
>>>> Gilbert
>>>>
>>>>
>>>>
>>>> The best way forward
>>>>>
>>>>> Hello everyone,
>>>>>
>>>>> I need to write a program (either in SAS or R) that generates all
>>>>> possible regression models WITH all possible combinations of
>>>>> interactions (essentially the All Possible Regressions variable
>>>>> selection method for OLS extended to binary logistic regression and
>>>>> also extended to include all possible comibnations of interactions).
>>>>>
>>>>> If the input explanatory variable list was:
>>>>>
>>>>> VAR = GENDER ETHNICITY COUNTRY
>>>>>
>>>>> then I want a SAS data set or R data frame generated that looks the
>>>>> table below. (BTW, these are all possible HIERARCHICAL models, there
>>>>> are additional non-hierarchical combinations possible. Ideally the
>>>>> program would be able to generate ALL possible combinations, but have
>>>>> a macro parameter or function argument (e.g., hierarchical = TRUE)
>>>>> which would identify and remove the non-hierarchical models from the
>>>>> results).
>>>>>
>>>>> Model Number RHS
>>>>>
>>>>> 1 GENDER
>>>>> 2 ETHNICITY
>>>>> 3 COUNTRY
>>>>> 4 GENDER ETHNICITY
>>>>> 5 GENDER COUNTRY
>>>>> 6 ETHNICITY COUNTRY
>>>>> 7 GENDER ETHNICITY GENDER*ETHNICITY
>>>>> 8 GENDER COUNTRY GENDER*COUNTRY
>>>>> 9 ETHNICITY COUNTRY ETHNICITY*COUNTRY
>>>>> 10 GENDER ETHNICITY COUNTRY
>>>>> 11 GENDER ETHNICITY COUNTRY GENDER*ETHNICITY
>>>>> 12 GENDER ETHNICITY COUNTRY GENDER*COUNTRY
>>>>> 13 GENDER ETHNICITY COUNTRY ETHNICITY*COUNTRY
>>>>> 14 GENDER ETHNICITY COUNTRY GENDER*ETHNICITY GENDER*COUNTRY
>>>>> 15 GENDER ETHNICITY COUNTRY GENDER*ETHNICITY ETHNICITY*COUNTRY
>>>>> 16 GENDER ETHNICITY COUNTRY GENDER*COUNTRY GENDER*COUNTRY
>>>>> 17 GENDER ETHNICITY COUNTRY GENDER*ETHNICITY GENDER*COUNTRY
>>>>> ETHNICITY*COUNTRY
>>>>> 18 GENDER ETHNICITY COUNTRY GENDER*ETHNICITY GENDER*COUNTRY
>>>>> ETHNICITY*COUNTRY GENDER*ETHNICITY*COUNTRY
>>>>>
>>>>> Any ideas?
>>>>>
>>>>> Thanks!
>>>>>
>>>>> Dan
>>>>>
>>>>> You may leave the list at any time by sending the command
>>>>>
>>>>> SIGNOFF allstat
>>>>>
>>>>> to [log in to unmask], leaving the subject line blank.
>>>>>
>>>>>
>>>>
>>>
>>> --
>>>
>>>
>>>
>>>
>>>
>>> ------------------------------------------------------------------
>>>
>>> Post-Doctoral Researcher in Statistics,
>>> ELDERMET Project
>>>
>>> Department of Statistics, Roinn na Staidhrimh,
>>> School of Mathematical Sciences, Scoil na nEolaiochta Matamaitice,
>>> Room 1.65, Seomra 1.65,
>>> Western Gateway Building, Aras an Geata Thiar,
>>> Western Road, Bothar Iarthar,
>>> University College Cork, Coláiste na hOllscoile Corcaigh,
>>> Cork, Ireland Corcaigh, Eire
>>>
>>> Tel. (+353 21 420) 5813 Fón. (+353 21 420) 5813
>>> http://euclid.ucc.ie/pages/staff/Susana/Susana.htm
>>>
>>> -------------------------------------------------------------------
>>>
>>
>
>
> --
>
>
>
> _____________________________
>
> Professor Gilbert MacKenzie
> Centre of Biostatistics,
> Room B2034
> Dept. of Mathematics& Statistics,
> University of Limerick,
> Limerick
> Ireland
>
> CBS ~ http://www.ul.ie/biostatistics
>
> BIO-SI ~ http://www.ul.ie/bio-si
>
> Gilbert ~ http://www.staff.ul.ie/mackenzieg
>
> Email: [log in to unmask]
>
> Tel: 00 353 (0)61 213499
> Fax: 00 353 (0)61 334927
>
> Belfast (Home) Contact: 00 44 2890 682358
>
> Professor MacKenzie was Visiting Professor of Statistics at ENSAI,
> Rennes, France from January 2010 until May 2011.
>
> ENSAI contact: [log in to unmask]
>
>
>
> _________________________
>
>
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|
