Xtnbreg: Adding country dummies to factor gravity panel model leads to chi2 test statistic not being reported

[Ralf Mischkowski]

Dear Statalist users,

I have created a dataset on bilateral knowledge flows. I am currently running a factor gravity like model with random effects. To regress the model I use xtnbreg. However, when adding country fixed effects c_* the chi2 test statistic is not being reported in the output:

I use this code for the regression

xtnbreg coinv ltotpat1 ltotpat2 lgdp1 lgdp2 ldistance colony comlang contig lksisum i.year c_*, vce(bootstrap)

Here is the output. I would be very thankful on any advice on this issue:

[Joao Santos Silva]

Dear Ralf Mischkowski,

Are you sure you want to use NB regression with random effects? It is not very robust to departures from the assumed distributions.

Best wishes,

Joao

[Carlo Lazzaro]

Ralf:
just an aside to Joao's more substantive advice: see -help j_robustsingular- about the missing Wald chi2 outcome and p-value.

[Ralf Mischkowski]

Dear Joao Santos Silva , dear Carlo Lazzaro,

I appreciate your prompt response. My variable of interest is time invariant for the group variable and hence i cannot use a fixed effects method. My supervisor has already ruled out mixed effects model. Thus, random effects model is the only way out:

"Taking in consideration the fact that our independent variables are constant over time (at least for the analyzed period), the only modelling method for the panel data analysis is the random effect."

(Pelau, Pop, 2018)

If there is anything that you (or, anyone else) can recommended then it will be really helpful. Thank you for your time

[Ralf Mischkowski]

Dear @Carlo Lazzaro,

thank you for the help j_robustsingular hint. I was able to resolve the issue by increasing the reps.

[Joao Santos Silva]

Dear Ralf Mischkowski,

I would simply use Poisson regression with clustered standard errors because that is very robust.

Best wishes,

Joao

[Ralf Mischkowski]

Dear @ Joao Santos Silva,


Thank you very much for your advice. My dependent variable represents patent data which is overdispersed. To cope with that problem, I chose to use negative binomial regression (as Poisson regression assumes constant dispersion). Is there a better way to deal with overdispersion in this context? It would be nice if you could suggest any resource/paper that would help me here.


Many thanks in advance.



Best regards,

Ralf Mischkowski

[Phil Bromiley]

If you search this listserve, you will find many discussions of nbreg and poisson estimators. Some folks strongly recommend poisson over nbreg even when data are overdispersed.

[Joao Santos Silva]

Dear Ralf Mischkowski,

The first thing to note is that overdispersion is only a problem if you are working with count data and you want to compute the probability of certain counts; if you are not working with count data, overdispersion is not even well-defined. If you just want to estimate the conditional mean, overdispersion is just like heteroskedasticity and can largely be ignored. Notice also that Poisson regression will be optimal even with overdispersion as long as the variance is proportional to the mean. Finally, whether you have overdispersion or not depends on the specification you use because what is relevant is the comparison between the conditional mean and the conditional variance (that is, the fact that the unconditional variance is larger than the unconditional mean does not imply that there is overdispersion in a regression model).

Now, Poisson regression is valid as long as the functional form of the mean is correctly specified and that makes it very robust. This property in not shared by the NB regression or by estimators using random effects because in these cases the properties of the estimator depend critically on the distributional assumptions that rarely have a reasonable justification. So, this is why I would go for Poisson regression.

All of this can be found in good textbooks such as the ones by Jeff Wooldridge and Cameron and Trivedi. Unfortunately, many other textbooks give very confusing or misleading advice on these issues.

Best wishes,

Joao

[Ralf Mischkowski]

Dear @Joao Santos Silva,

I highly appreciate your valuable advice on this issue. I will consider it when choosing and running the final regressions. Furthermore, I’m very thankful for the suggested literature.

Dear @Phil Bromiley,

thanks for the hint at the discussion between nbreg and poisson.


Best regards,

Ralf Mischkowski