I think this misses the main point. On the whole, a response that is
nearer normal in marginal distribution will be easier to model than one
that isn't, but as is emphasised in every good regression text, the
assumption that is being made in modelling is that of conditional
normality given predictors and moreover that is probably the least
important assumption you could make in any case. -qladder- does not
address that.
This point has been much laboured on the list!
The main reason I can see for transformation here is that the
relationship involving original and possibly transformed variables is
more nearly linear than it was before. If you have just one x, the issue
is best explored by a series of scatter plots. If you have several x,
then commands like -mrunning- from the SJ can help get a handle on how
successful the transformations have been. Whether you have one predictor
or several, a basic criterion for a good model is that the residuals are
essentially patternless. The standard postestimation graphs help here.
Nick
[email protected]
Martin Weiss
Well, if you want ideas for a whole range of transformations, try
-qladder-.
Other than that, you have provided the measure of a better
transformation by
employing the R-square. I do not think that there is a procedure in
Stata
that explicitly endorses such a model selection strategy. In terms of
covariates, -stepwise- comes to mind...
wangxin
I have a linear regression model which is not very
good since the R-square is very low.I test the
Log-transformed dependent also, which turns to be not
good also. I am wondering is there a way in Stata to
test which transformation for variable is better?
*
* 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/
