Pseudo Poisson Maximum Likelihood

[kristina009]

Hello,

I need to do this PPML estimation and as I haven't done that one before, I was wondering if there is a possibility to run this in eviews or not at all?

thanks,
Kristina

[EViews Glenn]

You should be able to use the count estimator or the GLM estimator to estimate PML models.

[kristina009]

The thing is, because this is a pseudo PML it should also be used with all variables, not just count. It's rather specific which is why I was wondering...

[EViews Glenn]

The count module does require an integer dependent variable. But the GLM estimator allows you to estimate using the log link and normal errors, which is the linear exponential family equivalent of the Poisson. The options for computing the dispersion allows for over/under dispersion.

[kristina009]

Hi Glenn, I know this was a while ago but I'm actually only now getting around to doing this analysis however I encountered a bizarre problem in the loading part of the data. For the first time ever in using eviews, it's stopped me from identifying panel data. I get an error message which says NAs found in id series. I don't know why this is because I checked and the data are clean and I've been using this same dataset in the exact same format for the past 5 years. Any ideas why?

[EViews Glenn]

I'd double check the ID series. Do a descriptive statistics on it and see if the number of observations matches the length of the workfile.

[kristina009]

I will have to check it - I was able to open it when eviews suggested dropping certain observations which I did - now I will have to find out what the problem is with those. Anyway, thanks for that.
Going back to my original problem, the ppml. I followed your instructions and I managed to get an estimation but I am just not comfortable with all the options and I'm wondering if you can expand a bit on your last PPML explanation or recommend if there is a section of the helpfiles that I should read. Also, could I run fixed effects, I don't see the option as in the OLS and GMM sections?

[EViews Glenn]

The count estimator allows for QML estimation, but requires integer dependent variables. There is an example in the documentation of the negative binomial QML estimation method. The GLM offers QML estimation without the integer requirement. There's an exponential regression example in the docs.

Briefly, the exponential regression example in the GLM chapter is close to what you want. The log link will let you do a pseudo-poisson model. If you choose a normal family as in the example, it will estimate the dispersion for you. If you choose thePoisson family, you'll be estimating the standard Poisson model with mean-variance equality. You could also choose the negative binomial family for a specified dispersion parameter. In all cases, a sandwich estimator for the variance option will give you robustness along some dimension. If this is unclear, you really should look at your references for PPML as there is a lot to this literature. I also highly recommend the Wooldridge chapter that we cite in our references.

Wooldridge, Jeffrey M. (1997). “Quasi-Likelihood Methods for Count Data,” Chapter 8 in M. Hashem Pesaran and P. Schmidt (eds.) Handbook of Applied Econometrics, Volume 2, Malden, MA: Blackwell, 352–406.

Neither the count nor the GLM estimators in EViews have been extended to panel data. If you want individual effects, you'll have to include them yourself (probably best done using the @expand keyword).

[kristina009]

Glenn, thanks - will be going over your comments and recommendations and will come back if I still have any quires (hopefully not though).

[kristina009]

Hello again,
Thanks for all your help so far, I was able to manage to estimate a PPML and I manually added dummies for time in order to get a FE. I have a much, much larger cross section so time FE is the only one I can estimate. I am puzzled by the results because the dummies are all significant if I don't include a constant while they're insignificant if I do. I've been scratching my head why. Then I remembered that I only know that I was meant to use FE when I used to do an OLS with DV>0 whereas now I have a DV that includes zeros. Is there by any chance a possibility to run a Hausman test in this circumstance? I wasn't able to find that option but I thought I'd check with you. Also, I read somewhere that there can be a sort of R-sq. estimated from this regression, is there a code I can use in eviews?
Sorry to keep bothering you with this issue - I don't have much time and this is a completely new estimation procedure for me.
thanks,
Kristina

[EViews Glenn]

Run a Wald test on the coefficients of the time dummies.