For a discussion of statistical tests and confidence intervals with a
parameter that might be on a boundary, see Pinhiero and Bates (2000)
Mixed-Effects Models in S and S-PLUS (Springer, Sect. 2.4).
hope this helps. spencer graves
Prof Brian Ripley wrote:
On Thu, 28 Aug 2003 Nicola.Beghetto@bancaimi.it wrote:
Thank you prof.Brian ripley for you suggestion.
Once I get the parameters from nnls.fit, I'd like to have also the
significance of the estimators, but the vcov() command (which I tried to
obtain the var-covar of the estimators,because nnls.fit doesn't give it)
doesn't give any results.
Did anyone have a suggestion on how to calculate the standard error of such
estimators?
Potentially there is no valid theory for this, as the true parameters
might be on the boundary. But in practice you may do fine by just fitting
(with lm) a model omitting all the variables whose coefs are fitted as
zero. Just check that the confidence intervals for the remaining coefs do
not go near zero.
|