RE: RE: st: xtnbreg - robusteness check and model relevance

[Simon Falck <simon.falck@abe.kth.se>]

Dear Ariel and Jay,

Thanks for your nice replies on my question regarding -xtnbreg-.

The study uses longitudinal count data, why a panel approach using -xtnbreg- with fixed-effect seems a natural approach. Since I have time-varying covariates, I guess a possible route could be using a pooled negbin estimator with time effects and cluster standard errors -nbreg DV X1 X 2 Xn i.year, vce(cluster id)-

I agree that -countfit- is useful to determine the best fitting model. AIC, BIC and LL also provides useful information.

Similar to Jay´s suggestion, I also had in mind generating predicted counts and compare predictions from the NEGBIN model with the observed (actual) values. The "better model" would be the model with the smallest deviation of the predicted values in comparison to the observed values. There is a nice paper (see: Barbosa, Guimarães, and Woodward, 2004, Foreign firm entry in an open economy: the case of Portugal) where this is done (see table 3). However, their study includes time-effects but not individual-fixed effects, which a conditional fixed effect NEGBIN model seems to control for if some specific assumption are meet. Providing a similar table has been one ambition in my paper, why I am interested in knowing how to derive the predicted values after -xtnbreg-.

My question is very similar to a recent question: http://www.stata.com/statalist/archive/2012-06/msg00311.html, for which I couldn't find any replies.

I am very thankful for further information on this issue, especially how to derive the predicted values after -xtnbreg- with -fe- option.

Best,
/Simon


-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Ariel Linden, DrPH
Sent: den 16 januari 2013 18:18
To: statalist@hsphsun2.harvard.edu
Subject: re: RE: st: xtnbreg - robusteness check and model relevance

Hi Simon,

Is there a reason that you must use -xtnbreg- as opposed to -nbreg- ? If you don't have time-varying covariates, why not use a "straight-up" count modeling approach with an offset? If you go that route, you can use
-countfit- (part of the spost9_ado package written by Scott Long & Jeremy Freese (findit -countfit- or - spost9_ado-), to help determine the best fitting model. 

If you must stay with the longitudinal modeling approach (ie., -xtnbreg- or -xtpoisson-), you could review the BIC and AIC stats by running -estat ic-.
It's not a panacea, but it is certainly an option.

Ariel

Date: Tue, 15 Jan 2013 12:15:11 +0000
From: Simon Falck <simon.falck@abe.kth.se>
Subject: RE: st: xtnbreg - robusteness check and model relevance

Dear Jay, 

Many thanks for your swift reply, and sorry for my delayed thanks and fuzzy description of the problem. 

Yes, goodness of fit is the correct description.

In terms of evaluating model relevance, your suggestion of making a graphical test is also what I had in mind. However, I am not sure what is the correct approach to derive predicted number of events after -xtnbreg-? 

The convenient -prcounts- command do not work after -xtnbreg-, only standard count models. 

The instruction for -xtnbreg- postestimation is using -predict-, but it seems that it is not possible to apply -predict [name], rate- to derive predicted number of events in a fixed effect model. Nevertheless, it is possible to derive predicted number of events; assuming fixed effect is zero, using -predict name, nu0-, but I am not sure if this is the correct approach.

A less preferred solution would be to not to run the empirical model in a panel, but as a standard negative binomial regression model using -nbreg-, and then leaving out the fixed effect which may bias the results. 

Any suggestions on what is the correct approach in this situation?

Thanks in advance,
/Simon





- -----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of JVerkuilen
(Gmail)
Sent: den 10 januari 2013 15:30
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: xtnbreg - robusteness check and model relevance

On Thu, Jan 10, 2013 at 5:34 AM, Simon Falck <simon.falck@abe.kth.se> wrote:
> Dear Statalist,
>
>
> The relevance or precision of a count model seems often to be 
> described in terms of how close the predicted values are to the 
> observed values, usually by comparing the distribution of 
> probabilities of observed and predicted counts. However, from what I 
> understand, it is not possible to use the command -prcounts- after 
> -xtnbreg-, which is used after -nbreg- to derive predicted values. Any 
> suggestions on what is a reasonable strategy in this case?>

I'm not sure what you mean by "robustness test" exactly, but I'll assume you mean goodness of fit test.
More to the point that approach doesn't actually inform you much about the difference between NB and Poisson, because NB and Poisson will tend to make very similar point predictions. Where they will differ is in terms of the level of uncertainty in the model. In general NB will have wider confidence intervals than Poisson, sometimes much wider.

So a reasonable graphical test would be to generate predicted values for important cases and see how often they line up with the corresponding observed value. There are issues with this approach that I can think of but it's a start.

Jay


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/