[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: RE: st: Interesting numerical accuracy/collinearity issue
Mark Schaffer <M.E.Schaffer@hw.ac.uk> has a follow-up question about -ovtest-:
> My follow-up question is simple: why does the shifting and scaling used by
> Stata's -ovtest- introduce greater accuracy rather than, say,
> greater rounding error? (Either accuracy or error would remove the
> numerical collinearity.) The algebra doesn't help me here, since all three
> methods are algebraically equivalent. I'm guessing that there's probably a
> general principle about how best to maintain numerical precision, but I
> don't know what it might be.
Actually, the three methods you describe are not all algebraically equivalent
according to -_rmcoll- and -coldiag2-. The algebra I mentioned only shows us
that the regression models yield a statistically equivalent F test.
The direct approach and your center/rescale method after taking powers are
algebraically equivalent to each other, but -ovtest-'s center/rescale then
take powers is not.
Let's just look at x^2 and x^3, if the values of x are not near zero (say they
are all positive), then it is easy to see how x^2 and x^3 can become
numerically collinear--even if you center/rescale them after taking the
Now generate z from the centered/rescaled values of x; this results in z^2
always being positive whereas z^3 is negative where z is negative. There is
no mistaking them to be collinear in this case.
Incidentally, I do not think of this as an accuracy or numerical precision
issue. To me it is more like shifting x into regions where we are better
equipped to numerically distinguish between powers of x.
* For searches and help try: