how to get adjusted r square in survey regression

[crystal feng]

I'm building a model with prespecified groups of predictors but have no ideal on how to derive (adjusted) r2 in survey regression (using svy: reg...). Is there any package available for the model fit test?


Thank you for your suggestion in advance.

[Sergiy Radyakin]

clear
version 8
sysuse auto
svyset [pw=weight]
svyreg price length
display "The r2 is: " %6.4f `e(r2)'

[Richard Williams]

She wants adjusted r^2 though. That is usually something that only matters in little dinky samples though, so I wonder how important it could be for a data set that was svyset.

[Sergiy Radyakin]

See here for adjusted r2: http://www.stata.com/statalist/archi.../msg00850.html

[Steve Samuels]

I wrote the 2008 post, but I'm not sure that I would recommend the calculation now. In version 13, svy:reg does not estimate e(r2_a) Plain regress with probability weights does estimate e(r2_a) with probability weights, but it uses t uses the sample size \(n\) in the formulas for adjusted R^2).

In 2008, I was trying to create the population analog of adjusted \(R^2\)
,
\[
aR^2 = 1 - \frac{\widehat{\sigma}_e^2}{\widehat{\sigma}_Y^2}
\]
where the numerator variance is the mean square residual and the denominator is the sample variance of the outcome.

I now agree with Richard.

Crystal, if your choice of model depends on the difference between \(aR^2\) and \(R^2\), then you are using the wrong criteria. There are better ways of making a decision, for example testing the significance of the added variables or linktest or even building a model on one part of the data set and validating it on another.

[crystal feng]

Thank everyone for sharing your valuable thoughts. Richard and Steve, you are making very good points. I'm thinking R square might be quite enough and it isn't a quite important criterion for model selection though.