st: RE: rvfplot and heteroskedasticity

 From "Verkuilen, Jay" To Subject st: RE: rvfplot and heteroskedasticity Date Sun, 25 Nov 2007 09:24:25 -0500

[delurk]

>>>My estimation has heteroskedasticity problems. After
correcting the standard erros with -robust-, the graph
after -rvfplot- shows the same pattern. How can I see
the heteroskedasticity problem has been solved?<<<

As was pointed out, "robust" doesn't actually *fix* the mean structure of your model. It corrects the VCE matrix and lets you do inference in the presence of heteroscedasticity in the model, +++provided you got the mean structure right+++. If you did not, then the situation can be quite different depending on how you went wrong.

First, see if you can figure out what's going on. Heteroscedasticity can be caused by a lot of factors ranging for "eureka" moments to, more typically, a mistake on your part, failure to do an appropriate transformation (though be careful since you lose the original DV, which presumably meant something to you), or an untransformable source such as a floor/ceiling effect, a coarsely discrete variable, etc.

Once you've thought about it, you may want to consider:

-It's a nuisance. If so, consider using regression based on Student's t, or one of the many robust regression methods out there. One is found in Stata but I don't know what it is. These will downplay the outlying observations.

-If you believe the heteroscedasticity is a finding (and it may well be), check out quantile regresssion (qreg in Stata). This lets you model the conditional quantiles of your distribution. See the book by Roger Koenker or Google for the various tutorials available on the web.

-If the DV is really NOT normal and not transformable to normal, e.g., it is L-shaped, you need to look at another model. Quantreg is (I believe) still a choice. Other choices include using the various survival analysis programs found in Stata, which will let you do regression using other distributions.

JV

<<winmail.dat>>