[Joao Santos Silva]

Thanks, Michael. These actually look like fractional data (Jeff also has a couple of important papers on that!); I wonder whether there is an upper bound for these ratios.

Anyway, the histograms suggest that you may actually be OK with a Poisson regression because most of the observations are concentrated near the lower bound. The fact that you do not have counts is not an issue at all because of the robustness of the Poisson regression (with or without fixed effects).

Cheers,

Joao

[Michael Hellwig]

Thanks for the reply, Joao. That's exactly why I rather opt for the Poisson regression.
Anyhow, there is no upper bound for these ratios. It's perfectly possible that a firm massively invests in their low stock of equipment yielding investment ratios above 1. I even had to drop firms who started from zero as this would imply a ratio of infinity... However, that's ok, as these firms are not of interest in my study.

[Jeff Wooldridge]

Michael: You figured it out: No restrictions on under- or overdispersion, and you can have it vary by i. That variable seems to be a fraction restricted to be between zero and one, right? Is it ever zero? Either way, you might want to try fractional response using correlated random effects. See my 2008 Journal of Econometrics paper with Leslie Papke, "Panel data methods for fractional response variables with an application to test pass rates."

[Michael Hellwig]

Thanks, Jeff. My variable is not really a [0;1]-fraction. The investment ratio could also take values greater than one or even become negative. Yet, in my sample it's in the interval [0.0003; 0.3806].
Still, I'll take a look at your paper. But I'm not so sure whether this random effects approach will help me in my diff-in-diff analysis where I'd rather use fixed-effects.