Emil O W Kirkegaard

I was supposed to be doing other work right now, but I just discovered there's a new USSC study on racial sentencing disparities with regression output at the end. So a couple thoughts after a brief skim! https://www.ussc.gov/sites/default/files/pdf/research-and-publications/research-publications/2017/20171114_Demographics.pdf …

So the actual history is converted to a score. Which is mapped onto six categories. Which then determines the range for a given offense. Which is what actually made it into the model.

However, given how vague this is, I'd have guessed that if you added a simple piece of extra info -- whether the person had any violence in their criminal history -- that would give a significant result. It didn't. So there's that.

Technical Q that may make no difference; Why is age included as a dummy variable (25 and under vs. >25) instead of, say, a polynomial? Just seems weird to throw out so much detail when age + age^2 is so common.

Protip: use a spline for age, not age + age². That misses a lot of nonlinear patterns and suggests curvelinear interpretations that may be wrong. Easy to do with the rms package for R. https://web.as.uky.edu/statistics/users/pbreheny/764-F11/notes/9-27.pdf …