Re: st: "alternative" stepwise approaches: Hierachical and Mixed

[Richard Williams <Richard.A.Williams.5@ND.edu>]

At 02:09 PM 1/26/2007, Diego Bellavia wrote:
> 1) Mixed approach: I have read some paper (clinical research) using 
> thi s approach and I tried to get it in STATA
> using :
>
>                sw outcome var1 var2 var3 var4, pe (0.1) pr (p 0.2)
>
> The command works fine only when pe is less than pr. I sayd it works 
> but I do not understand if it is a "mixed" stepwise approach
> and actually, what is a "mixed" approach. Anyone can explain what 
> are the different steps in such an approach ? Finally, when the 
> "mixed" approach is better
> than the standard backward (forward) one ?
I don't know how the papers you read defined "mixed", but what you 
are doing here is alternating between removing variables and adding 
them.  The reference manual discusses this in more detail than the 
online help does.  If I understand it correctly, the estimation 
begins with all variables included and then backwards selection is 
used to eliminate vars.  Then, it switches to forward selection to 
add vars back in.  Then, it goes back to backwards.  The process 
stops once no more vars are either added or removed.  (Unless maybe 
it gets into some sort of infinite loop!).

I believe the idea is that a var might be insignificant and get 
dropped; but once other vars have been dropped, it might become 
significant and qualify for re-inclusion.  Or conversely, something 
might be significant and make it in, but once other vars are added it 
can become insignificant and should be dropped.

If you add the forward option, it will start with forward selection 
and then switch to backward.

When should you use the "mixed" approach?  I guess if you want to be 
as mindlessly atheoretical as possible, "mixed" is even better than 
straight forward or backwards selection.  stepwise approaches in 
general don't attract high praise, although i think they sometimes 
can have a little merit (e.g. it can be nice to see if a mindless 
atheoretical model comes out the same as or close to the model you 
derived theoretically).


> 2) Hierarchical stepwise: using the "hier" option, it works fine, 
> but, again, it is not really clear to me what is the meaning of a hierarchical
> model building, in what it is different, for example, to a standard 
> backward/forward approach, and what STATA does during the different steps.
> I have read something about the hierarchical model building 
> technique, but the material I have found on-line were not that clear.
In your first example it wouldn't matter whether you list the vars as 
x1 x2 x3 x4  or if you listed them as x3 x4 x1 x2, i.e. order 
wouldn't matter.  It would matter if using hierarchical.  So yes, I 
suppose you could have your interaction terms listed last, and if 
using backwards hierarchical, it would begin by testing whether the 
interactions belong in the model.  If they are significant, the 
process stops.  If not significant, then it would go to your next 
term or set of terms (in the order you listed them), and test 
it/them.  Process stops once the term/terms test is/are significant.

Or, if you use forward hierarchical, it tests to see if the first 
term should make it in.  If no, the process stops and it goes to the 
next term and repeats the process.

Note that you can group vars, if the var list is x1 x2 (x3 x4 x5) 
then x3 x4 x5 will be tested simultaneously, e.g. you might group 
your interaction terms that way.

The hierarchical approach is, or can be, fairly theoretical.  You can 
get much the same thing using -nestreg- though.


-------------------------------------------
Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
FAX:    (574)288-4373
HOME:   (574)289-5227
EMAIL:  Richard.A.Williams.5@ND.Edu
WWW (personal):    http://www.nd.edu/~rwilliam
WWW (department):    http://www.nd.edu/~soc 
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/