Decolonizing Recidivism Models
It has been a while since I posted because I have been going deep in my research about the so-called “science” of recidivism modeling (ie predicting the risk that a person convicted of a crime re-offends). There are many types of recidivism prediction algorithms which are used and misused on all levels of the American criminal justice system (including sentencing) supposedly to help states better allocate resources to people that need it. However these models have been shown to be highly biased against people of color, women and people living in poverty. There are lots of great people writing about this topic and how to conceptualize the problem notably the mathbabe blog and propublica. In another post I will attempt to outline why these biased algorithms are so dangerous but first I wanted to zoom out back to the reason I started writing this blog.
Statistics prides itself on being comprised of perfect, unchanging ‘laws of nature’ written in the language of mathematics. This is how people argue that it is impossible to decolonize statistics. They argue that if statistics is co-opted for an unjust cause, such as justifying the biased recidivism models above, then this is the fault of the designers and users of the models rather than the underlying statistics. The cure then is simply to produce more woke statisticians but leave the techniques alone.
I clearly disagree with this view (otherwise this blog wouldn’t exist) but even I found myself falling into this mental trap as I began to think about how to write about recidivism modelling. I started by grouping the problems I found with the algorithms into three distinct categories: emotional problems, philosophical problems and statistical problems. As I started to write, though, my thoughts felt disjoint and poorly organized. I couldn’t really bring forward the extent of my philosophical fear of these models. Obsession with categorization was a colonial era mistake European scientists made and it obviously still plagues our thinking to this day. There is no reason that philosophical and emotional arguments should hold a lower place in statistical discussion than purely statistical arguments. The real world is not constructed of purely statistical arguments and needs to be conceptualized as interdisciplinary.
This is why recidivism risk modelling is such an important example to take up because despite my misgivings about the algorithms’ use and effectiveness, such models fulfil many statistical criteria for being significantly predictive.
That is why we need to expect more from a decolonized statistics than simply developing statistically significant models.
These recidivism algorithms are proving inaccurate based on patterns of structural bias which mirror racial and socio-economic biases already present in the criminal justice system and yet people are not questioning them because of the implicit “perfection” of mathematics and mathematical modelling by extension.
It is therefore up to a decolonized statistics to welcome the emotional and philosophical discomfort that comes with using a computer program to essentially determine prison sentences for vulnerable individuals.
We need to expect as standard the development of values driven, ethically sound models. In 2017 we need to be able to expect imagination and forward thinking to help create the world we want, rather than thinking which further cements the colonial foundations of our capitalist society. Learning about decolonization is an ongoing process and it took a while before I diagnosed the problems in my own thinking.
With that said, my detailed misgivings about current recidivism models and some explanations about how to improve and decolonize them will follow in the next post.