[Joao Santos Silva]

Dear Kenneth,

Including or excluding the constant should not make a difference if you do not exclude the base category of the dummies.

About the coefficients, that may indicate that your model is not correctly specified or that there is something wrong with your data; it is very unlikely to mean that PPML is not suitable.

Best wishes,

Joao

[Ng Yongwen]

I would like to thank you for your help so far.

I also tried running both panel poisson fixed effect and panel negative binomial fixed effect model. I am very perplexed as to why in the former model, time invariant variables like border, distance, language and colonial relationship were dropped however for the latter model those variables still remain.

I was not able to find any possible explanation for the discrepancy so far, do you happen to know why?

Regards,
Kenneth

[Joao Santos Silva]

Dear Kenneth,

I believe that happens because the NB FE model is not a real FE model in the usual sense; that is another reason not to use it!

All the best,

Joao

[Ng Yongwen]

I would like to thank you for your advice that you have given so far, the progress of my thesis is going well.

I also realised when panel negative binomial fixed effect was used, huge number of observations were dropped due to ''having one obs per group'' and " all zero outcomes''

I tried googling to understand why there were dropped observations it but I still do not get the intuition as to why the observations were dropped.

Do you know why observations were dropped for those particular reasons?

Regards,
Kenneth

[Joao Santos Silva]

That's because those observations are not informative about the parameters of interest.

Best wishes,

Joao

[Maxence Morlet]

Dear Kenneth,

I agree with your professor: PPML with FE is the way to go. That is what I recommended initially: Poisson regression with FE.

Probably the literature uses the other approaches that you mentioned because of frequent misconceptions about overdispersion and zero-inflation. For example you say that Poisson regression would not be appropriate because the variance is larger than the mean. There are two problems with your statement: 1) to have overdispersion you need the conditional variance to be larger than the conditional mean, so you cannot conclude that Poisson regression is not appropriate just because the variance is larger than the mean; 2) even if indeed there is overdispersion, that is not a serious problem unless you want to compute probabilities of particular counts; if you just want to estimate the conditional mean, overdispersion is irrelevant.

The important thing is that the only robust count data model for panel data is PPML with FE (which is Poisson regressions with FE and robust standard errors); that is the one I would recommend.

Best wishes,

Joao

Dear Joao,

That is very useful information!

Might you perhaps have a published source I could reference in a paper conveying these two points?

[Joao Santos Silva]

Dear Maxence Morlet,

I am working mostly from home and do not have my books with me, but I believe the Cameron and Trivedi books (the one on count data and the one on microeconometrics) will cover those points. You can also check the books by Jeff Wooldridge. On the specific use of Poisson with FE, please check Jeff Wooldridge's 1999 paper.

Let me know if these references are not enough and I'll check my library next time I go to the office.

Best wishes,

Joao

[Hoda Ashari]

Dear Prof Joao Santos Silva

Thanks for all the information. I would like to ask a question about using interaction within PPML command. I am working on the impact of long Covid on healthcare resource utilization witha longitudinal database. Having a lot of zero in my outcome variable, I decided to go with zero-inflated fixed-effect poisson regression. But when I add the interaction between variable long Covid and variable "days after getting Covid" Stata says the interaction is not allowed. I checked this interaction in other models loke negative binominal and it works there, but with ppml the intection is not allowed.
Can you please help me how can I add intercation to the model?

Kind regards,

Hoda

[Joao Santos Silva]

Dear Hoda Ashari,

I am not sure if I understand your question, but please note that ppml does not estimate a zero-inflated Poisson regression; in any case, I would advise against using a zero inflated model in this context. If you decide to use PPML with fixed effects, I suggest you use the user-contributed ppmlhdfe command.

Best wishes,

Joao

[Hoda Ashari]

Dear Prof. Joao Santos Silva,

Thanks for your help and time. My outcome is 90-days healthcare resource cost that is reported in 4 waves. I have a huge amount of zero in my outcome and at first I decided to run a model with zero-inflated negative binominal model. After reading this forum I changed my mind to zero-inflated poisson regression. May I ask why don't you recommend zero-inflated model?
And regarding my previous question, I put longcovid##c.timeaftercovid in my model, but Stata says that interaction is not allowed. I think it is related to ppml command, because this interaction was allowed in the other commands.

Kind regards,

Hoda

[Joao Santos Silva]

Dear Hoda Ashari,

A zero inflated model assumes that there is a sub-population that cannot have positive costs, which I do not think is realistic. Note that a Poisson model is compatible with any proportion of zeros i the sample.

The ppml command does not accept that kind of notation; would would have to create the interactions manually and include them in the model. However, I think you can use the interactions if you use the ppmlhdfe command I recommended above.

Best wishes,

Joao

[Hoda Ashari]

Dear Prof. Joao Santos Silva,

Once again, thanks for your help. I would like to ask another question. I runned two different command for fixed-effect poisson regression as below;

xtpoisson y x z i.wave, robust fe
ppmlhdfe y x z i.wave
i.wave is a dummy variable for the wave that outcome(y) and independent variables(x z) were measured.

surprisingly, coefficient and Robust Std. Err of the wave variable in two command are almost similar to each other (for instance 0.075 and 0.078). But there is a huge differenc betwen the coefficient and Robust Std. Err of x and z in the result of two command. While there is a significant p value for ppmldhfe command's result, the p vlaue of Xtpoisson is not significant.
Also in xtpoisson around 20,000observation were dropped beacuse of either only observation per group or all zero outcomes. In ppmldhfe even the group with 1 observation have not been dropped. (By group I mean the individual)
I would like to ask if I did anything incorrectly in ppmldhfe command and if it make sense that ppmldhfe keep the group with one observation?

Thanks for your time,

Kind regards,

Hoda

[Hoda Ashari]

To add to previous post;

I runned the below command and and the final result of it was similar to xtpoisson command, but wave variable was ommited because of colliniarity, which didn't happen with xtpoisson command;

ppmlhdfe y x z i.wave, a(wave individual-id)

Kind regards,

Hoda

[Joao Santos Silva]

Dear Hoda Ashari

What you did in #28 is the correct way to absorb the fixed effects. Wave is dropped because you are also absorbing it, so you should not include it as a regressor.

Best wishes.

Joao

[Dmitrii Petrukhin]

Dear Kenneth,

I agree with your professor: PPML with FE is the way to go. That is what I recommended initially: Poisson regression with FE.

Probably the literature uses the other approaches that you mentioned because of frequent misconceptions about overdispersion and zero-inflation. For example you say that Poisson regression would not be appropriate because the variance is larger than the mean. There are two problems with your statement: 1) I ; 2) even if indeed there is overdispersion, that is not a serious problem unless you want to compute probabilities of particular counts; if you just want to estimate the conditional mean, overdispersion is irrelevant.

The important thing is that the only robust count data model for panel data is PPML with FE (which is Poisson regressions with FE and robust standard errors); that is the one I would recommend.

Best wishes,

Joao

Dear Prof Santos,

I was reading Microeconometrics: Methods and Applications ‐ by A. Colin Cameron and Pravin K. Trivedi, and they say "A second and more obvious deficiency of the Poisson model is that for count data the variance usually exceeds the mean, a feature called overdispersion. The Poisson model instead implies equality of the variance and the mean (see (20.2)), a property called equidispersion." in p.670. Could you please explain why we should be worried about the conditional rather than unconditional moments?

Kind regarrds,
Dmitrii