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

RE: st: xtmepoisson and xtmelogit error messages

From   David Jacobs <>
Subject   RE: st: xtmepoisson and xtmelogit error messages
Date   Mon, 04 Feb 2008 14:02:42 -0500

Tony has a good idea, but unfortunately there are no panel zero inflated models in Stata as yet.

One way around this problem is to use the zero inflated models that are available in Stata and cluster on your case ids. Another possibility is to use the zero inflated panel routines available in Limdep. Although this stat. package is not nearly as convenient or comprehensive as Stata, it does have more panel routines. Still another idea is to use the hurdle models in Limdep or consult the Long and Freeze book (in the Stata book store) on how to estimate these models in Stata

Dave Jacobs

At 01:44 PM 2/4/2008, you wrote:

Another option might be to consider this as a zero-inflated poisson
model or zero inflated binomial model - I'm not sure if this is
available directly in the xt series.  If the zeros are identifiable
(e.g., you know which subjects in the Poisson must be zero), you can
model the zero vs. non-zero with stmelogit, and the non-zero with
xtmepoisson, but with the distribution here, it looks like this could be
problematic.  If you have some structural zeros, these could be modeled

The main thing would be to make the model understandable to people who
weren't intimately familiar with the data.


Peter A. Lachenbruch
Department of Public Health
Oregon State University
Corvallis, OR 97330
Phone: 541-737-3832
FAX: 541-737-4001

-----Original Message-----
[] On Behalf Of Maarten buis
Sent: Monday, February 04, 2008 12:27 AM
Subject: Re: st: xtmepoisson and xtmelogit error messages

--- Margit Averdijk <> wrote:
> I am estimating a growth curve model based on unbalanced panel-data
> with 6 waves of data. The dependent variable is victimization, and
> age is the sole predictor variable. There are 17 different age
> cohorts and models are estimated for each age cohort separately.
> Some information about the data: the victimization variable counts
> the number of victimization events and is highly skewed with about
> 96-98% being zero, 1-3% being one, and less than 1% being 2 till 9.
> Every cohort contains around 1000 persons. I would like the model to
> contain a random intercept and random slope.

Controling for each seems definately necesary. Though that does depend
a bit on the range of age you are studying: if the range is 30-47, than
it might not even be necesary, if it is 16-90 than age is definately
going to have a non-linear effect. However, breaking the data up in 17
chuncks is probably too wasteful. Given the large number of zeros you
have precious little data. One possible solution is not to break up
your data but control for age using a cubic spline, as you probably
want the effect of age to remain non-linear (going from 30 to 50 will
probably have less effect than going from 50 to 70).

Hope this helps,

Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434

+31 20 5986715

Sent from Yahoo! Mail - a smarter inbox

*   For searches and help try:

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index