I am using Splus 3.4 under SGI IRIX 5.3.
My problem is poisson regression, where my data y_i
are counts assumed to be distributed as a Poisson
r.v with mean
\sum_{j=1}^{n} \lambda_j p_i,j
where p_i are vectors of independent variables which sum to 1,
and \lambda is vector of parameters to be estimated.
With simulated data, I find using nlminb to directly minimise the
log-likelihood produces unbiased parameter estimates, compared
to the results of glm(family=poisson).
My problem is in determining the standard errors of the parameter
estimates, which I need for prediction.
To be precise, here are several ways of obtaining the variables p_i,
and one of the parameters is crucial for the intended application.
Thus, my criteria for the best model is that which minimises the
prediction error associated with that one parameter. Hence, the
need for standard errors.
Is there some way of computing these directly from the value returned
by nlminb, or must I resort to bootstapping?
Cheers
Steve Cumming
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|