Does temporal spatial procedure have negative binomial for panel data?

[Art Kendall]

Learning about the procedure for dealing simultaneously with temporal and spatial autocorrelation has been on my TO DO list since it came out.  

There is a dataset for a number of countries over 30 years. The predicting variables are volatilities of some series of continuous variables. The predicted variables are series of count variables. Negative binomial regression was used. Some of the activity counts are  "transnational" and some are "domestic".  

Would it be reasonable to suggest the author look at the temporal spatial procedure?  Or am I misunderstanding the general idea of the procedure?

Does the procedure have a negative binomial regression  approach for the temporal part that also looks at spatial autocorrelation?

[jkpeck]

The STM procedure is integrally tied to map data, with or without time series data.  It can handle events, but requires maps.

The TCM procedure handles time series and different source, which might be countries, divisions, etc, but AFAIK, the target variables are continuous or min/max/mode/sum aggregates, so negative binomial would not be appropriate.

[Andy W]

It is hard to give good advice for panel data modelling since there are so many variants (both the nature of the data matters, as well as the type of inferences you wish to make).

There are several reasons why even with count data it may not make sense to go with count models for panel data. It may be normal models fit just fine, (https://andrewpwheeler.com/2017/09/03/graphs-and-interrupted-time-series-analysis-trends-in-major-crimes-in-baltimore/), or the volatility is so high even negative binomial is not a good fit. (And how to model the dispersion, a single dispersion, dispersion for every panel unit?)

Count models are easy to mess up and produce illogical inferences, such as explosive in auto-regressive effects (https://andrewpwheeler.com/2017/07/05/dont-include-temporal-lags-of-crime-in-cross-sectional-crime-models/), or just effect estimates that don't make sense in general (https://andrewpwheeler.com/2014/06/17/poisson-regression-and-crazy-predictions/), so on their face the estimates will not generalize out of sample.

You can of course mess up linear panel data models as well, but it is more tricky for count data IMO, so I would prefer those not wary to just default to linear models.

[Art Kendall]

Thank you for your feedback.

To the author I suggested looking into the SPSS temporal and spatial procedures to see it they could be used to provide additional insight.

To the editor I suggested that if the article is re-submitted that I would approach AMSTAT leadership to find people I could talk to and perhaps even have them look at the article. AMSTAT has both Statistics Without Borders and a committee on volunteering (I don't recall the committee name).

I also suggested I could try to find out from IBM who was responsible for those procedures.