While you were busy watching the mostly shambolic performance that pass for the Euro 2012, I was unboxing my set of complementary “Quantitative Applications in the Social Sciences” volumes (you know, the series of flimsy paperbacks whose sickly green sleeves conjure memories of psychiatric wards and man in lab coats). The latest addition to my collection of these is Paul Allison’s volume on Fixed Effects Regression Models. There seems to be a good deal of confusion about what constitutes a “fixed” (as opposed to a “random”) effect, and Allison does a great job in clarifying the issue. In a panel design, assuming that effects are fixed basically implies that you have to estimate a separate, time-invariant effect for each object (e.g. person, country, party …) from the data. If you run a random effects model, you simply estimate the variance of the distribution of these object-specific effects, which is much more efficient.
Obviously, random effects models are all the rage at the moment, but Allison (who also gave us a very accessible introduction to missing data problems) has much to say in favour of fixed effects. Crucially, he demonstrates how fixed effects can be included in linear, logistic and even survival-time regression models, and how fixed and random effects can be combined to create a hybrid model. This is by far the most clear and concise introduction to this topic, and I like the book a lot. And now, back to the other four books in that case.