Backward Selection with Newey-West standard errors

[Mateo Bergman]

Hello,

Im using an OLS regression and have decided to use the backward selection method.

Im not really sure how to conduct the backward selection if i have autocorrelation & heteroskedasticity and use Newey-West.

Should i first runt the regression --> Test BG/BP --> Run Newey-West. and then remove the least significant varaible. Then start over, run the regression again with out the variable that was least significant --> Run "estat ic" to see if AIC/BIC decreased or if asjusted R^2 increased.

[Nick Cox]

Cross-posted at https://stats.stackexchange.com/ques...y-west-standar

Please note our longstanding request in the FAQ to tell us about cross-posting. That is good etiquette in any forum.

[Felix Bittmann]

The regular Stata command for stepwise selection is stepwise. However, this does not work with newey. Since I am not finding a suitable dataset to run with command at all in your setting, maybe you can just try this ado:

ssc install gvselect, replace
help gvselect

[Bruce Weaver]

I had not heard of gvselect (SSC) before, so I looked at the description.

. ssc describe gvselect

-----------------------------------------------------------------------------------------------
package gvselect from http://fmwww.bc.edu/repec/bocode/g
-----------------------------------------------------------------------------------------------

TITLE
      'GVSELECT': module to perform best subsets variable selection

DESCRIPTION/AUTHOR(S)
      
       gvselect performs best subsets variable selection.  The
      Furnival-Wilson (Technometrics, 1974) leaps-and-bounds algorithm
      is applied using the log likelihoods of candidate models,
      allowing variable selection to be performed on a wide family of
      normal and non-normal regression models.  This method is
      described in Lawless and Singhal (Biometrics, 1978). The log
      likelihood, Akaike's information criterion, and the Bayesian
      information criterion are reported for the best regressions at
      each predictor quantity.
      
--- snip the rest ---

In the Stata FAQ on stepwise regression, both Frank Harrell and Ronán Conroy state that "all possible subsets" regression does not address any of the well-known problems with stepwise regression. I'm curious to know how the method(s) used in -gvselect- (SSC) address those concerns. Thanks.

[Felix Bittmann]

I assume that gvselect does not solve these problems at all, it is just an alternative to stepwise. Also: the problems mentioned concern theory building / testing but there might be other uses for stepwise regressions (maybe prediction?), so there are probably some applications where it can be useful.
EDIT: there is apparently more information available here: https://journals.sagepub.com/doi/pdf...867X1501500406