Re: st: xtnbreg question, part 2

[David Jacobs <jacobs.184@sociology.osu.edu>]

In theory there's no reason why you can't use the Stata NBREG count 
estimator for time-series data.  One potential problem concerns 
autocorrelation.  As you probably already know, OLS assumes independence 
between cases, but this assumption often is violated when the data is 
longitudinal.  Corrections are available for least squares estimators , but 
I am only aware one article that attempted to correct negative binomial 
estimates for autocorrelation and that correction evidently was an ad hoc 
method supplied to the author by a colleague.  Another possibility is the 
Newey West which corrects for both autocorrelation and heteroskedasticity 
and is available for NBREG in Stata (or at least this was the case in 
version 7)

Often, however, there are alternatives.  First, if your counts are large 
the number of zero limit scores is modest, your counts probably won't be 
poisson distributed.  If they are normal or if a transformation will make 
them normal you should use least squares and this estimator can readily be 
corrected for autocorrelation.  Use normality tests on your counts and see 
if they reject normality.  Or try transformations like loging or square 
rooting and then test for normality.

If your counts are low and probably poisson distributed, you might get a 
break and not have autocorrelation.

Dave Jacobs

At 11:15 AM 3/15/2004 -0800, you wrote:

> Hi, I posted the following message on Friday and it was suggested by a 
> list member that I should be using the time series modules, not the panel 
> modules. I was also referred to a Stata volume on time series models.
>
> I am still uncertain about the answer though: my dependent variable is 
> count data, which is not appropriately analyzed by linear models. I looked 
> at the table of contents and index for the time series volume before I 
> sent my original message and I did not see any commands for count data 
> (such as negative binomial models). However, the xt series does have xtnbreg.
>
> So, my questions still remain: (1) is there a time series (but not cross 
> sectional time series) command in stata for negative binomial models of 
> count data (and if so, what is the command); and (2) if there is not, can 
> I approximate this by manipulating how i and t are set in the xtnbreg command?
>
> Cheers,
>
> Jennifer
>
>
> > ------------------------------
> >
> > Date: Fri, 12 Mar 2004 16:44:18 -0800
> > From: "Jennifer S. Earl" <jearl@soc.ucsb.edu>
> > Subject: st: xtnbreg question
> >
> > I have time series data on a set of variables for one nation for a period
> > of about 40 years. My dependent variable is a count variable. After
> > looking over the Stata manuals for xtnbreg, I am not sure how to
> > accommodate the fact that while I have time series data, these data are
> > not composed of cross-sections with multiple units in each cross section
> > (e.g. I do not have data for multiple nations in each year, just data on
> > one nation for multiple years).
> >
> > The manuals seems to suggest that Stata wants to handle data where i=panel
> > identifier (e.g., an id number for a person who contributes data for
> > multiple years, or a country id when multiple countries are present in the
> > time series) and t=time identifier (e.g., year) in xtnbreg. Further
> > xtnbreg *requires* that "i" is set either in the xtnbreg command or using
> > "iis (varname)", even though it does not require "t" to be set.
> >
> > I have two questions: (1) should I be using a command other than xtnbreg
> > to analyze this time-series count data; and (2) if xtnbreg is appropriate,
> > will using the following settings:
> >
> > iis year
> > tis year
> > xtnbreg dep_count_var ind_var1 ind_var2...
> >
> > produce appropriate results?
> >
> > Thanks in advance,
> >
> > Jennifer
> >
> >
> > *
> > *   For searches and help try:
> > *   http://www.stata.com/support/faqs/res/findit.html
> > *   http://www.stata.com/support/statalist/faq
> > *   http://www.ats.ucla.edu/stat/stata/
> >
> > ------------------------------
> >
> > Date: Fri, 12 Mar 2004 20:03:13 -0500
> > From: David Greenberg <dg4@nyu.edu>
> > Subject: Re: st: xtnbreg question
> >
> > If you have data for a singlenation for multiple years, you should be
> > estimating time series models, not models for panel data. Stata Press
> > publishes an entire manual devoted to such models.
>
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/