05-23-2013 03:39 AM
I am using proc reg to estimate a two models where my sample size is 107 and 82. The population size is about 3000 at least, but it is impossible to get more data (public records).
I have estimated my model, and found that it contains heteroskedasticity. Furthermore, as some of my explanatory variables are on individual levels and not means, I need to log-transform to get normally distributed errors. But this introduces even more heteroskedasticity.
To account for this, I know of two different methods.
I can use the model statement in proc reg to get heteroskedasticity-consistent robust standard errors. The main problem here is that these robust s.e. are only asymptotically robust. I am not sure how well these perform in my small samples of 82/107 obs.
I can use feasible GLS to calculate the weights needed to account for the heteroskedasticity in the model - and estimate the model with a weight statement in proc reg. The question here is pretty much the same, I am not sure how well WLS perform in my small sample.
Method, for this procedure, that I know of:
-take the log of the squared residuals from intended model
-regress log(u^2) on all X-vars and get fitted value (g_hat)
-Do WLS using 1 / exp(g_hat) as the weight.
- In the SAS support page, there is a method (By McKinnon and White) in the model statement to get heteroskedasticity-consistent robust s.e. when the sample size is under 250 (HCCMETHOD=3). Is my sample to small for this method, or will these robust s.e. produce consistent and
- Using the FGLS/WLS-procedure, will it produce consistent/efficient results on my small sample?
Any input much appreciated