[Stephen Jenkins]

At risk of prolonging discussion that shouldn't be, I'll say that I don't particularly disagree with what Maarten or Marcos are saying in their most recent posts. My main point is that one should not conflate two things: (i) a hierarchical data structure (of which a panel is an example), and (ii) the model which one applies to those data. Hierarchical data can be analysed by both "random effects" (= mixed models to me) and "fixed effects" models. ... as Stata shows: cf. xtreg, re and xtreg, fe. As it happens, the FE approach doesn't generalise well to some cases, notably those with binary or categorical dependent variables -- in which special approaches are required. Which I think is where this thread started.