Panel data - DV continuous variable with many true zeros

[Mohana Priya]

I am new to this forum and know only basic econometrics(self-studied). As a part of my research, I am looking at cattle feed usage levels (kg/cattle) in dairy farms as a function of extreme weather parameters. For this study, I have identified 100-panel balanced farms for a period of 20 yrs. My research question is: Does the seasonal /Monthly weather extremes impact the annual feed usage levels (kg/cattle) on the farm. I am at looking at within farm variation over time.

Interestingly, the overall average percentage of farms that purchased feed for the study period is only 53%. It means there are a lot of true zeros in the dependent variable. I read that Fixed Effects estimator will not be appropriate for the heavily right side tailed data.

I have decided to slightly modify my research question - Does the weather influence bulky feed purchase or not? To answer the same which of the model should I choose - xtlogit (fe, re) or xtprobit (re).
Please help me with the same.

[Carlo Lazzaro]

Mohana:
welcome to this forum.
In my opinion, the issue here seems more substantive than statistical: what's the reason why a relevant percentage of your sample did not purchase cattle food during the selected timespan?
Did they stock enough cattle food to avoid further purchasing? There were merges and/or aquisitions between farms during the 20-period that you considered in your analysis? Otherwise, it is hard to believe that during a 20-year span of time so many farmers did not purchase cattle food.
That said, I do not think that changing your research question can fix this issue, nor I would sponsor changing research question after seeing the data.

[Phil Bromiley]

To add to Carlo's helpful posting, I am not sure what your mean by the 53%. Does that mean 47% never bought or that at the farm-year level 53% of the farm-year observations are zero?

If it is 47% never bought, I think you probably want to think about what differentiates the never buyers from the others. What determines who never buys is probably not the same as a zero for a farm that sometimes buys and sometimes does not buy. That is, there are probably stable farm-level factors that determine the never buyers.

If it is 53% of the farm-year observations, then you should consider a panel tobit since a farm cannot buy less than zero. Your xtlogit is a reasonable alternative, but it throws away the information you have on volume purchased.