more institutionalization, less homicide?
bernard harcourt is guest-blogging at volokh this week, offering engaging posts on deinstitutionalization, incarceration, and homicide. he nicely explicates a new state-level analysis of the national evidence discussed in his january times op-ed and texas law review piece.the pattern appears to hold up under a more stringent state-level panel specification: aggregate institutionalization (prisons plus mental institutions) bears a strong inverse relationship to homicide rates over a long historical period. moreover, the correlation between homicide and aggregate institutionalization is far stronger than the correlation between homicide and imprisonment. my sense is that the individual-level literature shows rather modest associations between violence and mental illness. what might account for the strong aggregate relationship?
posted by christopher uggen @ 6:34 PM
4 comments
4 Comments:
What is your view of the time period included in the dataset? Mass incarceration is large a result from the 1980's forward, while the data from the paper is from the 1930's onward. I would need to read the analysis more carefully, but my initial assessment would be that there may be an issue of applicability to present times.
Text is easy to read, flows nicely.
The time-series modeling strategy is somewhat old-fashioned.
The author does not appear to test for non-stationarity. The highly-trended series indicate shifting mean and variance across different portions of the series. He is estimating things in level form (e.g., Table II.1). Same holds true in the dense state-level panel. This can lead to incorrect statistical inference. Also needs to test for cointegration.
A simple first step would be to first difference dep. and all ind. variables and apply OLS. Better strategy would be to use error correction model (ECM). In the ECM you can test for short-run and long-run effects of all ind. variables so you fd the dep. and then add in lagged fd and level forms of all regressors. This is typically estimated with OLS.
Author should have said more about experimentation with lags. While he did correct for AR(1), is there a distributed lag?
What about the direction of causality? Does the homicide rate respond to changes in institutionalization or does institutionalization respond to changes in homicide? What is the behavioral model that the statistical model is supposed to represent?
The %urban variable is kind of silly. It is the percent of a state living in an area w/ 2,500 or more people (or 500 persons per square mile). It increases linearly over time even though many places the Census Bureau would not count as a metro area have a pop. of more than 2,500. Definition changes over the time period of his study.
Paper ends with some bizarre notion of "social physics" as an interpretation. I am thoroughly baffled at how the author is interpreting the findings. He pays some homage to "social construction of deviance", but points out that his definition of deviance (homicide) is pretty much a fixed definition.
Nonetheless, he has unearthed some very complex and rich data that can be used to further elaborate upon what he has done in this study.
thanks for the thoughtful comments, mike and tom. a provocative piece, to be sure...
nonetheless, the data will seriously make a lot of people re-think the incarceration revolution
thanks for sharing this scholar's work with us, i have downloaded a lot of his SSRN papers
Post a Comment
<< Home