# Industry (and year) fixed effects with non-panel data?

For my thesis i'm trying to estimate the effect of several variables on the recommendation level of analysts artound Mergers and Acquisitions. My dependend variable is 'recommendation' which can be (a) upgraded (b) unchanged (c) downgraded. I have several independend variables like payment, firm size ect.

Because my depended variable is ordinal and discrete, I would like to run an ordered logit model (ologit in stata). My professor wants me to add fixed effects for industry and if possible for the year.

I consider my data as non-panel data, because I'm looking on recommendation levels and i'm not comparing them over time. but, as far as I know, you can only run fixed effects on panel data, is this correct?

I'm using stata 13, and now I just add dummies for the 'fixed effect'. My code in stata looks like:

xi: ologit recommendation cash size value return i.industry i.year, cluster(industry)


is this enough to add the 'fixed effects' in my regression or is there a better way/solution to add industry effects and, if possible, year effects to my regression?

Kind regards

Fixed effects regression is not limited to panel data. You can have multiple observations within the same person (over time), which is panel data, but you can also have multiple observations within an industry and/or within a year, which is your design. It is the nesting of observations within a higher level unit that is necessary.

Using industry/year indicator (dummy) variables is a trick that can be used to get a fixed effects model in linear regression. Unfortunately, that does not extend to non-linear models like ordered logit. So even though the model can be sensible, it is not a fixed effects model. I don't think a fixed effects ordered logit has been implemented in Stata.

Asides:

Are you sure that an ordered logit is right for your case? Your variable being ordered is not enough to justify an ordered logit. The ordered logit is a parsimonious model in that it estimates just one effect for the two contrasts, but I find it easy to imagine that the effect of a variable on upgrading differs from the effect of that same variable on downgrading. This would be a violation of the parallel lines assumption. That is not just a technical inconvenience but a substantively important question.

It is now good practice to avoid the xi: prefix when possible and use the factor variable notation instead.

• Dear Maarten, Thanks for your fast reply. My choise for an ordered logistic model is based on the fact that I'm trying do do similar research like Kolanski and Kothari (2008), who examine recommendation and changes in recommendations with an ordered logit. Another study (Becher et al, 2014), claim to run logistic regressions with year and industry fixed effects, but unfortunately they don't specify their which logistic regression they run. I do share the concern that the paralel lines assumption is violated. could this be prevented by running an mlogit or a simple logit? – Ralf Jun 25 '15 at 7:54
• and is it possible to control for industry effects in these models? – Ralf Jun 25 '15 at 7:54