Poisson regression with fixed effects and time fixed effects creates no values

[Jana Schue]

Hello all,

I work with panel data and use a nonnegative count variable as dependent variable (patent count) and a binary variable with two values (purpose) as independent variable.
As regression I use poisson with fixed effects. I tested for overdispersion, autocorrelation, multicolinearity and time-fixed effects. Everything except multicolinearity is present. My command looks like this:
xtpoisson pcount purpose `controls' i.fyear, fe vce(robust)

When I run the regression without i.fyear (which I want to use for the time-fixed effects), I get a significant result. However, when I run i.fyear as well, the regression table is simply empty and spits out no values (see table below).

Does anyone have an idea if I can fix time-fixed effects differently at -xtpoisson-, or why else it doesn't work?

If there are any other relevant tests for -xtpoisson- besides the ones mentioned above, I would appreciate info here as well.

Many thanks in advance,
Jana
------------------------------------------------------------------------------
| Robust
pcount | Coefficient std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
purpose | -.0519881 . . . . .
cf | -3.47e-07 . . . . .
growth | -.1231627 . . . . .
lev | .0001373 . . . . .
rdi | -.0099032 . . . . .
capex | .0000163 . . . . .
adex | .0000519 . . . . .
os | -.0107794 . . . . .
|
fyear |
2007 | .0279703 . . . . .
2008 | .0257085 . . . . .
2009 | 18.35652 . . . . .
2010 | 21.86815 . . . . .
2011 | 23.40286 . . . . .
2012 | 24.13357 . . . . .
2013 | 24.62137 . . . . .
2014 | 24.82756 . . . . .
2015 | 24.96078 . . . . .
2016 | 25.02995 . . . . .
2017 | 25.12183 . . . . .
2018 | 25.07121 . . . . .
2019 | 24.84527 . . . . .
2020 | .2126817 . . . . .
------------------------------------------------------------------------------

[George Ford]

With only two dummies as regressors, it may be the year/fe eat the two dummies (colinear).

[Jana Schue]

Hi George,
thank you very much for your answer. In fact, only the independent variable is binary (purpose), the dependent variable (patent count) is a count variable (0,1,2,n).

[George Ford]

Even worse. Likely the 2 fixed effects account for the variation in X.

[Jana Schue]

That doesn't sound good. Do you have a tip for me on how to fix this?

[George Ford]

Hard to say, and it's a guess that you have collinearity. You might contemplate why you need either or both FE from a conceptual perspective.

You're getting significant year fixed effects, so that probably needs to be there, but the coefficients are very similar after the first couple of years and may be not be significantly different. Big change after first couple of years--I'd investigate that if you don't know why.

vce(robust) deals with overdispersion so I wouldn't worry about that. I believe that vce(cluster <fixed effect variable>) addresses autocorr and hetero and overdispersion.

[Jeff Wooldridge]

By any chance do your explanatory variables only change across t and not with i? Including the dummy variable which perhaps indicates a pre-post period? Then you’d have perfect collinearity.

Also, how did you xtset your data?

[Jana Schue]

Hello,

thank you very much for the answers.
Since I am still very new to Stata, I am unfortunately not completely sure. My independent variable (purpose), which is the binary one, shows for each year whether voluntarily invested in carbon offsetting credits (1), or not (0). The dependent variable patent count shows the number of patents for each year.
I have two other variables in my master thesis, Tobin's q and emissions (both with -xtreg-), and the regression has no problems with fe and i.fyear.
However, yesterday I noticed that if I delete the last part, i.e. vce(robust) from the command xtpoisson pcount purpose `controls' i.fyear, fe vce(robust) and run the command, that significant results also come out. The problem with the empty table is only present if I have included all three parts in the command. Can i.fyear or vce(robust) possibly be substituted here?

Or might it still be an option to switch to xtnbreg if that fits better?

I have set my data as panel data with this command: xtset gvkey fyear

Thanks in advance,
Jana

[George Ford]

what is in `controls' ?

[Carlo Lazzaro]

Jana:
the issue is probably related to -i.fyear-.

[Jana Schue]

My controls are the following: cf growth lev rdi capex adex os (which are cash flow, growth, leverage, r&d intensity, capital expenditures, advertising expenditures and organizational slack).

Is there maybe any other way to fix time-fixed effects here? Or would that also be a basis then to argue that I let those out?

[Carlo Lazzaro]

Jana:
you could start off with keeping -i.year- in and add one control at a time and check when the issue you complained about creeps up.

[Jana Schue]

Thanks for the tip. In fact, the problem is there with every single control variable.

[Jeff Wooldridge]

You should show us a sample of your data, as I can't see what's causing the problem if the controls vary across both i and t.

[Jana Schue]

Here is a snippet of my final dataset. purpose, b2c and dirty are binary variables. purpose is my independent variable and for the -xtpoisson- regression I use pcount (patent count) as dependent variable. Maybe I don't understand correctly what it means that the controls vary across t and i.

gvkey fyear capex adex sic tq os cf growth lev rdi pcount ccit purpose emissions emissions_estimated b2c dirty
1045 2015 6151 110 4512 .9710574 3.558618 9217 3.648802 0 0 0 0 42038000 42038000 0 0
1045 2016 5731 116 4512 .9367234 3.696223 4502 -.0199588 6.431704 0 0 0 0 39279000 39279000 0 0
1045 2017 5971 135 4512 .9690576 3.434643 3936 .0492167 6.384361 0 0 0 0 39388000 39388000 0 0
1045 2018 3745 128 4512 .8058636 3.347701 3569 .053824 -201.355 0 0 0 0 40604000 40604000 0 0
1045 2019 4268 129 4512 .7621446 3.276446 4004 .027175 -283.4237 0 0 0 0 41439600 41439600 0 0
1045 2020 1958 50 4512 .8195997 3.74241 -6520 -.9707422 -5.973642 0 0 0 0 20089000 20089000 0 0
1075 2005 645.55 0 4911 .6275108 4.983398 570.815 0 0 0 0 1 1
1075 2006 758.769 0 4911 .7277359 7.854283 675.787 .1297002 0 0 0 0 1 1
1075 2007 941.644 0 4911 .7020075 8.118397 672.216 .0351995 0 0 0 0 1 1
1075 2008 954.397 0 4911 .6128393 7.716234 603.915 -.0454442 0 0 0 0 16401625 16401625 1 1
1075 2009 775.354 0 4911 .6362064 10.90158 486.337 -.0210011 1.136371 0 0 0 0 15741639 15741639 1 1