Dear Stata users,

I'm looking at the effect of coauthorship with a star scientist on the individual's productivity in a panel data from 1996-2017 in event-study framework. The treatment indicator switches to 1 when author coauthor's with a star scientist and it stays 1 since then. The data consist of authors from different subject fields such as Agriculture, Chemistry, Economics, etc. To create a comparable dependent variable, I use field normalized citation as the dependent variable (it is the ratio of total citation received by author i in subject field j in year t to mean number of citations in subject field j in year t), therefore the dependent variable is non-negative and continuous. The dependent variable consists 22-25% of zeros.

In my understanding, in Chapter 18 of "Econometrics of Cross section and Panel Data" Jeff Wooldridge explains that even if the dependent variable is non-negative continuous variable and estimates are robust. Also, I have read articles that have the similar context to my research topic, they also provide statements in favour of Poisson QML fixed-effect model for non-negative continuous dependent variable. However, in their research, the dependent variable is count data.
I find only this paper uses the same research context to mine and uses non-negative continuous dependent variable using the model, "https://ift.tt/3icxZDw".

This is my first research study. I need guidance on the my question :
  • Is applying Poisson QML fixed-effect model appropriate for non-negative continuous variable in event-study design? Or should I take add one to dependent variable and take log i.e. ln(1 + dependent variable) and use TWFE.
Currently I have used "xtqmlp" STATA command to run the Poisson QML fixed-effect model. The results look sensible. However, the results from TWFE regression are off.

Does anyone has any suggestion on the above query. Any help is appreciated.

Thank you so much.