Ping Zhang wrote:
I wonder if anyone out there could either your wisdom or experience
about the following:
If I had to calculate MLE using generic optimization algorithms, what
does one usually do with calculating standard errors? I know we could do
bootstrap or jackknife. Or if one is brave enough, one could derive the
2nd order derivatives and start from there, i.e. SE = ( -L2 )^{-1},
where L2 is the 2nd order derivative. Can't we use (L'L)^{-1} instead,
where L is the 1st order derivative? Any general suggestions?
Which "generic optimization algorithms" are you using? Both nlminb and
optim from MASS will provide the hessian. The latter will even provide
the hessian when the analytical gradients are not provided. Then the
inverse of the hessian gives you the covariance matrix provided you are
minimising a -logLik.
--sundar
|