Dear R mailing list,
I'm trying to fit a censored regression model to a large (dimension of the
design matrix is 2e5 by 7) right truncated data by means of the
survreg(Surv()) function, as suggested by Paul Johnson on his "R Tips"
web
page.
Possibly due to the sensitivity to the initial values of the Newton-Raphson
routine in use by survreg(), resulting regression outputs turn out to be
unreliable. Should the "init" field be left blank, the maximisation
barely
starts (2 iterations), and ends up yielding estimates close to 0 or +/-
infinity.
I guess the problem mainly lies in reasonably specifying starting values for
the optimisation embedded in survreg(). Even by inputing OLS estimates (the
default in the SAS lifereg method) ensuing inferences are remarkably
different, and with significantly lower log-likelihood, than the ones
returned by the SAS proc lifereg.
Has anybody encountered a similar problem? If so, any suggestion/advise would
be grately appreciated. Many thanks in advance, all best,
--
Dr. Stefano Conti
Department of Probability and Statistics
The Hicks Building
University of Sheffield
Sheffield S3 7RH
UK
Phone # +44 (0)114 2223854
Fax # +44 (0)114 2223759
