[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: convergence problems with zinb

From   "Kieran McCaul" <[email protected]>
To   <[email protected]>
Subject   st: RE: convergence problems with zinb
Date   Thu, 14 Aug 2008 08:03:37 +0800

Hi Margaret,

You might like to look at the various maximize options that you can
tinker with.  

Type - help maximize

There are four different algorithms used by ml and Stata follows a rule
for stepping though these when fitting a model.  There is an option
called - difficult - which tells the ml program to use a different
stepping rule.  You could try this.

Any algorithm used to maximize the log-likelihood has to start with some
initial coefficient values and sometimes these can be near a flat or
concave region of the likelihood function.  Consequently, the fitting
algorithm will just wander around this region, unable to get out.
Changing the initial values might therefore improve the performance of
the algorithm.

The option - trace - will give you the current coefficient estimates at
each iteration.  This might give you a clue as to what's going on. 

I don't know how many variables you have in your model, but you could
try fitting a separate model for each variable thus obtaining a
coefficient estimate for each variable and then use these as your
initial estimates in the full model.

Kieran McCaul MPH PhD
WA Centre for Health & Ageing (M573)
University of Western Australia
Level 6, Ainslie House
48 Murray St
Perth 6000
Phone: (08) 9224-2140
Phone: -61-8-9224-2140
email: [email protected] 

-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Holland,
Sent: Thursday, 14 August 2008 5:15 AM
To: [email protected]
Subject: st: convergence problems with zinb

I believe a zero-inflated negative binomial would be the best fit for a
model I am trying to run, but I've had trouble with convergence. I was
wondering if anyone else has had this problem and, if so, if there are
any tricks to helping it converge or ways to try a different algorithm.

I have tried running the same model in R and SAS with less success. I
have done some imputation for this project and found a single imputation
set that will converge on Stata, but will not converge on R or SAS.
Based on this set, zinb is a better fit than zip or nbreg.

I have found that a hurdle model (logit / zero-truncated negative
binomial) converges more easily and has only slight worse fit than the
zinb in the set that will converge. However, theoretically the zinb
model makes more sense.

Any suggestions?
Thank you,

*   For searches and help try:

*   For searches and help try:

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