Does an event studies analysis require an OLS model?

[Mansi Jain]

Hi! I am writing to ask whether an event studies analysis can only happen with a linear regression model. I am trying to look at the effect of fire occurence on visits to National Forests. My Y variable is thus a count variable, and I have panel data (for any given site, I have data for 20 years). Since fires occur at different times in different places, I was planning to conduct an event studies analysis to answer my question. I was toying with two ideas:

-- using xtreg, fe on Stata to regress log(visits) on X. I know this is not ideal since Y is a count variable, but this is my understanding of how event studies analyses are usually done.

-- Since it is count data, I was considering doing a negative binomial regression with fixed effects. On Stata this would mean doing nbreg with dummies for all the site IDs included. My understanding is this is better than using xtnbreg, fe. I would then be able to interpret the coefficient to say that a 1 unit increase in X leads to a (beta* 100)% change in Y when beta is small (correct)?

My question is:

- I am leaning towards the second option because I know it's usually better suited for count data. Would the results from the second be considered as an event studies analysis and/or a generalized difference in difference, or does it have to be a standard OLS regression to be considered an event studies analysis?

- Which model seems better given this information?


Thank you so much! I appreciate it.

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

I would lean to the panel nb or poisson estimator. You have count data and should treat it as such. I wouldn't worry about "being able to interpret..." leaning to dummies instead of fe - just use the margins command after the estimation and you'll get the solution. I would use the panel estimator provided by Stata - particularly in non-linear models, using dummies for "fixed effects" can run into complex difficulties.

[Mansi Jain]

Phil Bromiley, thank you so much for your response! That's very helpful. One quick followup: my units of analysis are spatially autocorrelated. Do you know if I can use spatial autocorrelation commands with xtnbreg? I can't figure out how to do that. Thank you so much!