Re: st: mmregress question

[Steve Samuels <sjsamuels@gmail.com>]

Bruno, perhaps you don't need -mmreg-  or -msreg-  and you can take another approach for your data.

1. Identify multivariate outliers in the predictor variables with the authors -mcd- command ("findit").  It also depends on p-subsets but there's no harm trying.

2. If you find observations clearly separate from the main group, they are potential high leverage points.  Take them out. 

3. Now run -rreg- and -qreg- or, better, -bsqreg- on the reduced data set.  These do not depend on subset identification and without likely high-leverage points should have decent resistance properties

Without knowing more about your data, seeing your commands, and exactly where iteration fails (in the estimation of the scale parameter or of the regression coefficients, it is difficult to say more. 

Good luck,

Steve






I should have pointed out, as you did not, that -mmregress- is a contributed package written by Verardi and Croux. It can be found by -findit-.

I already stated how to get the identical results from run to run; reread my post.

I  also mentioned that dummy variables will give -mmregress- and -sregress- trouble and stated that -msregress-, another part of the package, must be used.  Since you don't mention trying -mmregress-, I guess that you have no dummy variables.


Other issues that could problems:

* Too many predictors for your number of observations)

In that case, try a reduced model

* The model is misspecified

I can think up data configurations where the wrong model, even for a single predictor, could cause -mmregress- and -ssregress- to be unstable. So you might, for example, try a fractional polynomial, instead of linear terms; or consider interactions (I know this conflicts with the parsimony recommendation).


Lower on the list of causes of your problems:

* Using too many or too few subsets.  The -help- shows how to vary the number.

* Not using the latest version


I've run -mmregress- repeatedly on several data sets of the same size as yours and have never encountered the instability you've seen.  So I suspect the problem is unique to your set-up, not "how to select appropriate starting subsample".


Steve







On Dec 2, 2011, at 2:56 AM, bcoric@efst.hr wrote:

Than you Steve.

I was using sregress and mmregress respectively as it stressed at the p.
444. However, results change significantly in each case when I repeat
these two commands.

You suggest that the reason is that the algorithm implemented in stata
starts by randomly picking N subset of p observation.
Could you please tell me how to set same N subsets before implement
mmregress.

Could you also tell me please do you have any recommendation how to select
appropriate results. Namely, selection of different starting subset will
probably result with different results again. If this is the case than,
question is, how to select appropriate starting subsample?

Bruno

> 
> The FAQ ask that you give precise references; in this case you are
> presumably referring to: Verardi, V., & Croux, C. (2009). Robust
> regression in Stata. Stata Journal, 9(3), 439-453.
> 
> The article states that -mmregress- starts with an initial S-estimator and
> (page 444) that
> 
> "The algorithm implemented in Stata for computing the S-estimator starts
> by randomly picking N subsets of p observation (defined as p-subset) where
> p is the number of regression parameters to estimate."
> 
> In other words, unless the same seed is -set- before each instance of
> -mmregress-,  one would expect the results to differ from run to run.
> 
> Also, if there are indicator ("dummy") variables in your model,
> -mmregress- will have problems (p 445). If you have dummies,  you should
> instead run -msregress-, part of the same package, and list them only with
> the dummy() option.
> 
> If you still are having difficulties, I suggest that you contact the
> package's authors.
> 
> Steve
> 
> 
> 
> 
> 
> 
> 
> 
> 
> On Dec 1, 2011, at 12:59 PM, bcoric@efst.hr wrote:
> 
> I would like to ask for help with implementation of stata command
> mmregress.
> I am doing simple cross section analysis. Since I have just 48
> observations I am very much concerned about the possible influence of
> outliers.
> Previously I was using stata command rreg for such chases. However, I came
> across the paper, Robust regression in Stata (2009), which argues that
> rreg command does not have expected robustness properties and recommend
> mmregress instead.
> However, I faced some problems in its implementation. Namely, its
> subsequent implementation to the same model leads to different values of
> regression coefficients. Moreover, detected outliers also change.
> I presume, that algorithm use iterative procedure taking previous
> estimates as starting values. However, results are very different in
> respect to the sign, size and significance of coefficients, and do not
> converge after subsequent applications.
> Does anyone can tell me is there anything, some procedure that I should
> follow, to get robust (consistent) results.
> Bruno Coric
> 
> 
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
> 
> 
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
> 


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/