Dear Yiling:
I have heard back from the LIFEREG developer regarding the FPZD error
with a null model on your data set. If you recall I was able to
overcome the FPZD error and as it turns out the sorting I did (by upper
lower) was helpful with your data. Below, I have edited the text of the
LIFEREG developer's reply slightly for clarity:
+ - - - - - - -
The reason the model converges after some data manipulation is the PROC
SORT step, which rearranges the data. Sometimes order of observations
can influence numerical computations, and it happens in this case that
the rearranged data converges with no error, but the original data
causes a FPZD. This can't be guaranteed to work for any data; it just
happens to work in this case.
The user might try different initial values; e.g. -
model (lower,upper)= / itprint dist=Weibull intercept=1 scale=2;
which works for this data.
The FPZD has been eliminated in the SAS9.2 (Phase 2) release due to some
numerical improvements in the function that computes the Weibull
likelihood. Note that the problem still exists in the current SAS9.2
(Phase 1) release which is available now, but is fixed in the SAS9.2
(Phase 2) release.
+- - - - - - -
Note that specifying initial values can be helpful as indicated in
Example 39.3: Overcoming Convergence Problems by Specifying Initial
Values - -
http://support.sas.com/onlinedoc/913/getDoc/en/statug.hlp/lifereg_sect30
.htm
I hope that helps.
Paul
