Thursday, November 12, 2015
Thursday, November 5, 2015
Beating the Stats on Reoffending
Somewhere - never to be binned - there exists a risk assessment completed ahead of my parole hearing. OASys - Offender Assessment System - is the standard risk assessment tool used by Probation and Prison services. And it is utter cobblers.
It rests on an algorythm that analyses 8 pieces of biographical information. With a statistically sufficiently large group, it can be seen which biographical factors are best able to predict reoffending rates. For instance, the age of first imprisonment is a strong indicator.
The perceptive will have already noted two major problems. Firstly, the group who shares my particular biography must be tiny, if it exists at all. And secondly, resting on actuarial data means that OASys doesn't measure change. Attempts are made to address this by including clinical data - ie, staff opinions - but this subjectivity is itself fraught.
My OASys proffered the odds of my reoffending in the first two years at 53.4%. This number has always tickled me, because of that decimal point. Here we have two criminal justice agencies checking out their collective wits in favour of a mathematical model which claims to be able to predict human behaviour to two decimal places. That is risible on its face.
More problematically, the OASys Manual makes it crystal clear that these assessments are scores for the GROUP of people who share characteristics. It may well be that people with teenage convictions have higher reoffending rates. As a group. But to extrapolate from these group scores to individual risk is both statistical nonsense and very unethical. Because 53% of a GROUP with a shared characteristic reoffends does not mean each INDIVIDUAL in that group has a 53% reoffending risk.
So I have outperformed a flawed risk assessment tool that is improperly used to assess individual risk. I could feel prouder.
It rests on an algorythm that analyses 8 pieces of biographical information. With a statistically sufficiently large group, it can be seen which biographical factors are best able to predict reoffending rates. For instance, the age of first imprisonment is a strong indicator.
The perceptive will have already noted two major problems. Firstly, the group who shares my particular biography must be tiny, if it exists at all. And secondly, resting on actuarial data means that OASys doesn't measure change. Attempts are made to address this by including clinical data - ie, staff opinions - but this subjectivity is itself fraught.
My OASys proffered the odds of my reoffending in the first two years at 53.4%. This number has always tickled me, because of that decimal point. Here we have two criminal justice agencies checking out their collective wits in favour of a mathematical model which claims to be able to predict human behaviour to two decimal places. That is risible on its face.
More problematically, the OASys Manual makes it crystal clear that these assessments are scores for the GROUP of people who share characteristics. It may well be that people with teenage convictions have higher reoffending rates. As a group. But to extrapolate from these group scores to individual risk is both statistical nonsense and very unethical. Because 53% of a GROUP with a shared characteristic reoffends does not mean each INDIVIDUAL in that group has a 53% reoffending risk.
So I have outperformed a flawed risk assessment tool that is improperly used to assess individual risk. I could feel prouder.
Subscribe to:
Posts (Atom)
