The Intelligent Design of Fixed Effects

Among Malcolm Gladwell and Steven Levitt, there is a dialog on the crime drop in the late 90s. Gladwell, in the Tipping point, attributes the drop to "broken window"-based police work. Levitt & Dubner in Freakonomics aren't convinced and come up with Roe V. Wade a the significant event. One element of their difference is the assumption of fixed effects. In this case L&D argue that controlling for State and year effects, broken windows theory doesn't sufficiently explain the 1999 drop in crime.

In our project and many statistical models, we create fixed effects, aka-dummy variables. The idea is that there will be variation that will occur between years, or states, or other categories of variables that aren't among your chosen variables. Thus the model assumes that a good economy, or an employment rate of Y will always accounts for X coefficient, every year, in every state. Stepping back though, this isn't a given, the economy or employment could certainly affect people more one year than another. But in econometrics, those variations are lumped together with all the other factors associated with that year or state.

Fixed effects are related to another element of a statistical model, the error term. The error term assumes, that yes, there maybe other variables at work, but they cancel each other out, and that only the important variables are included. This is the "intelligent design" Achilles heel of econometrics. The "model/evolution" explains only so much until you can't explain it "error term=designer". The difference is that a judge isn't about to throw out econometrics as methodologically soft.