For my thesis I am using as dependent variable the fraction of cash as part of the total price offered by the bidder. So, it's a fractional response that lies between [0,1]. I am not sure which regression should I use in Stata.

Also, my sample comprises 500 acquisitions in Europe announced in the period 2002-2016 from companies in different sectors (some companies have multiple acquisitions). The professor told me I should "control for year and industry (Fama French 12 - ffinds) fixed effects and adjust heteroskedasticity-robust standard errors for bidder clustering".

Here is what I have tried to do:

  1. fracreg logit Y X1 X2 i.Year i.ffinds, vce(cluster ID)
  2. glm Y X1 X2 i.Year i.ffinds, family(binomial) link(logit) robust nolog , but I cannot cluster for ID
  3. I have tried to use xtlogit, but I am not able to apply it to multiple fixed effects

Which one is the correct approach? Also, using i.Year and i.ffinds I have too many dummies in the output. Is there a way to suppress them (like the option absorb used with reg)?

  • $\begingroup$ I would have thought from the details you give that beta regression was the way forward. I have no idea how to do that in Stata and anyway asking for code is off-topic on this site. $\endgroup$ – mdewey Apr 14 '17 at 11:01

The first example is exactly how I would have done it.

  • Modeling proportions is what fracreg is for (although it's not the only way, with beta regression being the obvious alternative).
  • You're controlling for year and industry.
  • You're adjusting the standard errors in the way he requested.

The second example, even if you could get it to work right (offhand, I'm surprised you can't use a cluster VCE here), would give you the same answer as the first. That's how fractional logistic regression used to be done in Stata, using glm with certain options.

I strongly suspect the third example wouldn't work even if you could get the specification right; I don't know for sure, but I've never seen any research on estimating fixed-effect fractional logit models, let alone research that suggests you can just call the likelihood a quasi-likelihood and charge ahead.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.