Re: st: RE: variance in glm

[Argyn Kuketayev <akuketayev@mail.primaticsfinancial.com>]

this is what i've done.

1. got residuals of linear regression
regress Y X
predict res, r

2. tabulated for each X, and got frequencies and st devs
table X, c(N res sd res) replace
g w=table1-1

3. computed relative errors
g re=table2/X

4. regressed relative errors over inverse squares of X, with weight
proportional to freq of X
g Xin2=1/X
regress re Xin2 [aweight=w]

R-squared     =  0.8316
Adj R-squared =  0.8281

finally, i have an estimate of st dev of errors:
2.75 / X + 0.5 X

---
now i want to plug this into Stata, so that it assumes that
Var[Ei] = 2.75 / Xi + 0.5 Xi
in Yi = Xi + Ei

any idea how?

thanks




On Wed, Mar 2, 2011 at 7:55 PM, Keith Dear <keithdear4@gmail.com> wrote:
> Argyn,
> You might consider calculating squared residuals and modelling those as as
> function of X using gamma errors (glm).  If the original errors Ei are
> normal, their squares will be distributed proportional to chi-squared
> (special case of gamma). Then you can test for constant variance, which
> would be rejected if the mean of the squared residuals turns out to depend
> on X.
> Keith
>

thanks
-- 
Argyn Kuketayev

*
*   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/