Are you sure what you found is not just Simpson's paradox?
When I look at the 2 KM plots for white/blacks, they are mostly the same. It's pretty clear that the model is not prejudiced against blacks, in fact it's somewhat prejudiced against whites. [1]
Your main editorial claim is that whites tend to be misclassified as "good" and blacks as "bad."
But I think what's actually happening is that algorithm is more likely to misclassify low_risk as "good", and high_risk as "bad".[2] Combine that with vastly more whites than blacks being low_risk (as you show earlier) and you get the observed "injustice".
I'll also note that the KM for whites flatten out at 2 years, unlike for blacks. This is actually a big deal if statistically significant. But that's a separate conversation.
Footnotes:
1 - this is acknowledged in methodology page "black defendants who scored higher did recidivate slightly more often than white defendants (63 percent vs. 59 percent)."
2 - why that is I don't yet fully understand (and I'd like to) but it looks's to be simple math that follows from low risk mostly not recidivating, and high risk mostly yes recidivating
Does this sentence "Northpointe does offer a custom test for women, but it is not in use in Broward County.
" imply that the base COMPAS model does not take gender into account?
When I look at the 2 KM plots for white/blacks, they are mostly the same. It's pretty clear that the model is not prejudiced against blacks, in fact it's somewhat prejudiced against whites. [1]
Your main editorial claim is that whites tend to be misclassified as "good" and blacks as "bad."
But I think what's actually happening is that algorithm is more likely to misclassify low_risk as "good", and high_risk as "bad".[2] Combine that with vastly more whites than blacks being low_risk (as you show earlier) and you get the observed "injustice".
I'll also note that the KM for whites flatten out at 2 years, unlike for blacks. This is actually a big deal if statistically significant. But that's a separate conversation.
Footnotes:
1 - this is acknowledged in methodology page "black defendants who scored higher did recidivate slightly more often than white defendants (63 percent vs. 59 percent)."
2 - why that is I don't yet fully understand (and I'd like to) but it looks's to be simple math that follows from low risk mostly not recidivating, and high risk mostly yes recidivating