Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: pgmhaz convergence (hazard functions)

From   Jenkins S P <[email protected]>
To   [email protected]
Subject   Re: st: pgmhaz convergence (hazard functions)
Date   Fri, 10 Oct 2003 23:14:44 +0100 (BST)

On Fri, 10 Oct 2003, Monica L. Parra Torrado wrote:

> Dear Stata List,
>  I am running the Jenkins porcedure pgmhaz with non parametric estimation of
> the baseline hazard, i.e., I am including weekly dummies instead of
> "log(seqvar)"
> However, it has run for 120 hours and 35 iterations and has not come up with
> any result yet.  After the last ten iterations it has been displaying the
> message "unproductive step attempted".  Does this mean it will not converge?
> I have run this model with diferent specifications and although this message
> appeared sometimes, it did converge.
> How can I know that the maximization will not converge? Does Stata give up
> at any moment?  Shall I wait till then? Or does stata keep trying forever?
> What should I do?

Be patient, perhaps.

The program often runs (very) slowly, given use of numerical derivatives,
and the relatively large size of social science data sets.  Moreover the
likelihood surface is not globally concave, so convergence problems can be
common, especially when one specifies a non-parametric baseline hazard.
(Ensure too that you have events at each and every interval.) It is often
a good idea to run -pgmhaz- with the -trace- option on so that you can see
if there are any problematic parameters. (E.g. the program trying to force
the gamma variance towards zero, i.e. log variance towards a large
negative number?)  ... hence maybe experiment with different starting
values for the (log) gamma variance.  All this is mentioned in the
original STB article, a version of which is available (in MS Word)  from
my Survival Analysis Using Stata webpages.

good luck Stephen

PS -pgmhaz- stands for Prentice-Gloeckler-Meyer HAZard regression, after
the original proposers of the model -- there is no J in it. (PG for the
grouped data proportional hazards model;  M for adding in gamma
distributed unobserved heterogeneity.)
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index