[R] Bayesian stepwise (was: Forward Stepwise regression based on partial F test)

Spencer Graves spencer.graves at pdf.com
Thu Feb 24 18:29:22 CET 2005


      Does anyone know of research fully Bayesian stepwise procedures 
assuming that models not considered by the stepwise would essentially 
have zero posterior probability? 

      I need to analyze the results of ad hoc experiments run in 
manufacturing with crazy confounding and possible supersaturation (i.e., 
more potentially explanatory variables than runs), when each run is very 
expensive in both time and money.   There have to be ways to summarize 
concisely and intelligently what the data can tell us and what remains 
uncertain, including the level of partial confounding between 
alternative explanations.  I think I've gotten reasonable results with 
my own modification of Venables & Ripley's stepAIC to compute an 
approximate posterior over tested models using the AICc criterion 
described, e.g., by Burnham and Anderson (2002) Model Selection and 
Multi-Model Inference (Springer).  Preliminary simulations showed that 
when I used the naive prior (that all models are equally likely, 
including the null model), the null model is usually rejected when 
true.  What a surprise!  I think I can fix that using a more intelligent 
prior.  I also think I can evaluate the partial confounding between 
alternative models by studying the correlation matrix between the 
predictions of alternative models. 

      Comments?
      Thanks,
      Spencer Graves

Frank E Harrell Jr wrote:

> Smit, Robin wrote:
>
>> I am hoping to get some advise on the following:
>>  
>> I am looking for an automatic variable selection procedure to reduce the
>> number of potential predictor variables (~ 50) in a multiple regression
>> model.
>>  
>> I would be interested to use the forward stepwise regression using the
>> partial F test. I have looked into possible R-functions but could not 
>> find this
>> particular approach.  
>> There is a function (stepAIC) that uses the Akaike criterion or Mallow's
>> Cp criterion. In addition, the drop1 and add1 functions came closest 
>> to what I want
>> but with them I cannot perform the required procedure. Do you have 
>> any ideas?  
>> Kind regards,
>> Robin Smit
>> --------------------------------------------
>> Business Unit TNO Automotive
>> Environmental Studies & Testing
>> PO Box 6033, 2600 JA Delft
>> THE NETHERLANDS
>
>
> Robin,
>
> If you are looking for a method that does not offer the best 
> predictive accuracy and that violates every aspect of statistical 
> inference, you are on the right track.  See 
> http://www.stata.com/support/faqs/stat/stepwise.html for details.
>




More information about the R-help mailing list