Thanks for the reply. I agree that’s a natural tentative interpretation of Table 26, taken in isolation. But note that table doesn’t give any indication of confidence intervals for the relevant column.
Have a look at Table 11 (below). We see the same numbers (up to rounding) in the final row, with 90% confidence intervals. Note the negative point estimates for rape and assault are very close to the center of their confidence intervals and thus not distinguishable from zero. Basically, there was a natural experiment in which California reduced the prisoner population and for those two categories, on a relative basis to (most of) the rest of the country, crime happened to decrease very slightly, but only to an extent well within the statistical noise. (In fact of the seven categories, only motor vehicle theft is significant—and barely so—at the p=0.1 level and none is significant at the conventional p=0.05 level.)
Note the crime numbers being used here are inferred from official crimes reported, rescaled using national estimates of the reporting rate (which helps to put e.g. murder and larceny on the same footing, despite their very different reporting rates). Since only a tiny fraction of crimes in prison are reported (at least that’s my sense), that means that crimes in prisons and jails are essentially being ignored (as Roodman states in his definition of incapacitation I initially cited).
The bottom line as I see it: If someone wants to do a CEA analysis of an intervention in this space, they should think carefully about the incapacitation term as sources (Roodman and I suspect the underlying literature) will tend to exclude crimes in prisons and jails. In fairness, it’s likely hard to get good estimates, but things don’t fail to be real because they are hard to estimate.
Thanks for the reply. I agree that’s a natural tentative interpretation of Table 26, taken in isolation. But note that table doesn’t give any indication of confidence intervals for the relevant column.
Have a look at Table 11 (below). We see the same numbers (up to rounding) in the final row, with 90% confidence intervals. Note the negative point estimates for rape and assault are very close to the center of their confidence intervals and thus not distinguishable from zero. Basically, there was a natural experiment in which California reduced the prisoner population and for those two categories, on a relative basis to (most of) the rest of the country, crime happened to decrease very slightly, but only to an extent well within the statistical noise. (In fact of the seven categories, only motor vehicle theft is significant—and barely so—at the p=0.1 level and none is significant at the conventional p=0.05 level.)
Note the crime numbers being used here are inferred from official crimes reported, rescaled using national estimates of the reporting rate (which helps to put e.g. murder and larceny on the same footing, despite their very different reporting rates). Since only a tiny fraction of crimes in prison are reported (at least that’s my sense), that means that crimes in prisons and jails are essentially being ignored (as Roodman states in his definition of incapacitation I initially cited).
The bottom line as I see it: If someone wants to do a CEA analysis of an intervention in this space, they should think carefully about the incapacitation term as sources (Roodman and I suspect the underlying literature) will tend to exclude crimes in prisons and jails. In fairness, it’s likely hard to get good estimates, but things don’t fail to be real because they are hard to estimate.