Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Stepwise Regression using Multiple Imputed Data - stata 11

From   Richard Williams <[email protected]>
To   [email protected], <[email protected]>
Subject   Re: st: Stepwise Regression using Multiple Imputed Data - stata 11
Date   Tue, 09 Aug 2011 00:44:52 -0500

At 09:34 PM 8/8/2011, Russell Simco wrote:

The "stepwise" model selection program is not available when using the "mi estimate:" prefix in stata-11. I would like to do this manually using "mi estimate: ologit".

Besides all the usual concerns with stepwise, with MI there would be the added problem that arbitrary or minor changes in your imputation procedure, e.g. using a different random number seed, would produce different imputed data sets and hence could lead to different variable selection. It wouldn't surprise me if there are other reasons why the use of sw with mi is invalid. You might take a look at the svy & stepwise FAQ for other ideas on how to proceed:

But, if you are bound and determined to do stepwise manually anyway, it doesn't seem that hard (other than possibly being extremely tedious). Suppose you have x1-x10. Run 10 bivariate regressions of y with each X. The X with the largest T value wins the first round (assuming it meets your criterion for entry, e.g. is significant at the .05 level). Lets say x1 wins.

In the 2nd round, run 9 regressions, e.g. y with x1 and x2, y with x1 and x3, etc. Winner (i.e. the added X with the largest significant T value) advances to the 3rd round.

Repeat process until no more x's are statistically significant.

Or, work backwards. Regress Y on x1-x10. The least significant x (assuming at least one meets your criteria for removal) gets dropped. Repeat as needed.

The manual (at least in Stata 12) also describes the logic behind stepwise.

Overall, I am not sure why you would want to do this, I think it is probably a bad idea to do it, and if I was going to do it I'd be tempted to do it with the original un-imputed data (at least then my fate isn't being decided by an arbitrary random number seed). But if you really really want to do it seems straightforward albeit tedious.

Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
HOME:   (574)289-5227
EMAIL:  [email protected]

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index