ppmlhdfe for poisson panel-time fixed effects

[Eran Itskovich]

Dear all,

I'm working on a project in which I run a fixed-effects Poisson regression with dummy variables for panel and time. I came across the -ppmlhdfe- command, and I wonder if it's what i'm looking for. I have two questions:

1. What is the difference between running:

ppmlhdfe depvar indepvar1... , absorb(i.panel#i.time) vce(robust)

and running:

egen panel_time = group (panel time)
xtset panel_time
xtpoisson depvar indepvar1... , fe vce(robust)

(besides the fact that -ppmlhdfe- is much faster?)

2. For the time variable I have 2 options: year or month.
Sholud I use just month for the fixed-effects (i.e. absorb(i.panel#i.month)), or should I use both year and month (i.e. absorb?(i.panel#i.year#.i.month))?

Thanks!
Eran

[Joao Santos Silva]

Dear Eran Itskovich,

I do not think you are using these commands correctly; if you do, they should produce the same result. For example, for xtpoisson with panel and time fixed effects you need something like

xtset panel time
xtpoisson depvar indepvar1... i.time , fe vce(robust)

Please check the help files of these commands for examples.

Best wishes,

Joao

[Eran Itskovich]

Dear Joao Santos Silva,

Thank you for your reply.
Actually, I do get the same results when running -xtpoisson- and -ppmlhdfe-, just like you suggested. I wonder, however, when I should use one of them and not the other. In other words - what is the differences between these approaches?

Eran

[Jeff Wooldridge]

It seems to me that you're not separately including time and panel fixed effects but panel x time fixed effects. If you didn't have monthly data -- so your time variable was annual -- this would eliminate all variation in your data and you'd get no estimates. You can only do it because your have monthly data. But I doubt that's what you mean. By generating the group dummies the way you did, the same thing happens with xtpoisson.

If you absorb by i.panel#i.year#.i.month then you will have nothing left to explain. How many total time periods do you have? How big is N?

[Eran Itskovich]

Dear Jeff Wooldridge ,

I have a total of 190 locations, across 13 years. Each observation is a day in a specific location, which means I have a total of 902,120 obsevations (190*4,748). I first tried to run the -ppmlhdfe- using absorb(i.location#i.day), but it drops a lot of singletons so the model can't run properly.
That's why I tried to use month as the time fixed-effect. Does it make sense?
Anyway, I ran the following command:

ppmlhdfe depvar indepvar1..., absorb(i.location#i.year#i.month) vce(robust)

and it produced the results I expected. The question is whether I got the fixed-effects right.
Can you please advice me on this?

Thanks!
Eran

[Arianna Gentilini]

Hi all,

Hopefully this is the best place where to post my question but please feel free to direct me if otherwise.
I have a slightly different issue than Eran, as when I run ppmlhdfe and xtpoisson I get different results, which makes me think that I am doing something wrong with the fixed effects.
My panel comprises 1,646,910 unique observations, nested within 29 countries,1,893 diseases and across 30 years. My dependent variable is a nonengative count with a large share of 0s (90%) that indicates whether a clinical trial took place in a certain country-disease-year.

I set my panel as follows:

egen country_disease= group(country disease), label
xtset country_disease year

Then, I run my FE Poisson model:

xtpoisson depvar i.x1, fe eform nolog

However, when I run ppmlhdfe as follows, I get considerably different results:

 ppmlhdfe depvar i.x1, a(year country_disease) eform nolog

On a related note, when I try interacting the FE - as I noticed earlier in this thread, I get the following result:

ppmlhdfe depvar i.x1, a(i.year#i.country_disease) eform nolog
(dropped 1646910 observations that are either singletons or separated by a fixed effect)
insufficient observations

Any help would be tremendously appreciated,

Thanks!

[Joao Santos Silva]

Dear Arianna Gentilini,

Your xtpoisson regression does not include the fixed effects that you include when using ppmlhdfe. On the interaction, I do not think you need the prefix i., but check the help file.

Best wishes,

Joao

[Arianna Gentilini]

Thanks Joao Santos Silva for the promt reply.
In the helpfile is says I should use i.var1#i.var2 to absorb pairwise combinations of two or more categorical variables (e.g. country-time fixed effects), which is what I would like to do.
However, I get an error message as below:

ppmlhdfe depvar i.x1, a(i.year#i.country_disease) eform nolog
(dropped 1646910 observations that are either singletons or separated by a fixed effect)
insufficient observations

I am unsure what insufficient observations means as I should have enough observations for the analysis (xtpoisson runs without problems).
Have you ever came across such issue? I have observations for 54,897 country-disease pairs across 30 years.

Best wishes,
Arianna

[Joao Santos Silva]

Dear Arianna Gentilini,

The following command is from the help file

 ppmlhdfe trade fta, a(imp#year exp#year imp#exp) cluster(imp#exp)

As you see, they do not use the i. prefix.

Best wishes,

Joao

[Arianna Gentilini]

Thanks Joao.
Unfortunately, even without the i., i.e. by using the following code, I still get the same result of insufficient observations.

ppmlhfe depvar i.x1, a( year#country_disease) eform nolog

I was wondering if you ever came across such message and if you knew how I could address it.

Many thanks,
Arianna

[Joao Santos Silva]

I suggest you contact the authors of the command.

Best wishes,

Joao