Regression in a Panel-Dataset

[Finley Voigt]

Hello there,

ich have a problem and need help. I need to conduct a regression analysis with an dependent count variable, which is heavily overdispersed, most of the datapoints are zero or close to zero. However our independent variables are binare (Dummy Variables) that we need to use as moderators. With our Dummy Variables we have extreme differences (for example not many 1 and a lot of 0).

If we know use the fix effect our variables are so highly insignificant, and we dont know why, because if we use the xtpoisson that is not correct to use in this situation, we guess, it fits perfectly. The pannel is highly unbalanced and with gaps.

This is done with the nbreg and it does not fit and is insignificant:

Doing it with the poisson it looks like that:

Do you have any tipps, or need more information, please contact me. I would highly appreciate it. :D

[Andrew Musau]

...because if we use the xtpoisson that is not correct to use in this situation

While checking the significance of coefficients is no way to determine the suitability of an estimator, the notion that Poisson regression is invalid for overdispersed data is simply incorrect. Also note that overdispersion is a property of the conditional distribution (so just graphing the outcome is not sufficient to claim that your data is overdispersed). Fixed-effects Negative Binomial relies on very strong assumptions. All you need to do is cluster your standard errors in the fixed-effects Poisson specification, and you can proceed with that. For more details on the comparison between FE Poisson and FE NegBin, see #3: https://www.statalist.org/forums/for...-poisson-model.

[Finley Voigt]

While checking the significance of coefficients is no way to determine the suitability of an estimator, the notion that Poisson regression is invalid for overdispersed data is simply incorrect. Also note that overdispersion is a property of the conditional distribution (so just graphing the outcome is not sufficient to claim that your data is overdispersed). Fixed-effects Negative Binomial relies on very strong assumptions. All you need to do is cluster your standard errors in the fixed-effects Poisson specification, and you can proceed with that. For more details on the comparison between FE Poisson and FE NegBin, see #3: https://www.statalist.org/forums/for...-poisson-model.

Thank you !! I will try my best to follow your instructions :D Thank you very much

[Finley Voigt]

While checking the significance of coefficients is no way to determine the suitability of an estimator, the notion that Poisson regression is invalid for overdispersed data is simply incorrect. Also note that overdispersion is a property of the conditional distribution (so just graphing the outcome is not sufficient to claim that your data is overdispersed). Fixed-effects Negative Binomial relies on very strong assumptions. All you need to do is cluster your standard errors in the fixed-effects Poisson specification, and you can proceed with that. For more details on the comparison between FE Poisson and FE NegBin, see #3: https://www.statalist.org/forums/for...-poisson-model.

Hi Andrew,
we tried by using vce(robust) behind the xtpoisson, fe but nothing really changed and most the variables are still very insignificant (P>z = >0.05). We tried and adjusted our model.

We try to extract the digital power ( measured in digital patents count) for different CEO attributes. Therefore we merge different datasets and extract the needed information. Than we calculate the variables and because it is done in literature we lag all independent and control variables. We also do the Hausmann-Test that tells us that we need FE. We want to in the final regression check how gender and the age of ceos influence digital patent power, with some moderators like overconfidence of ceos and digital experience.

Does one have any further tipps on why our regression is so unsignificant, can that be rooted in the merger of different datasets into one or within a calculation of a variable, but that are usually straight forward calculations or already given in the dataset (for example digital experience is already a given variable).

Would be highly appreciated :D