In the statistical analysis (unlike the verbiage) she is completely unable to hide the lack of bias and the accuracy of the algorithm, all of which are clearly on display in line [36]. In contrast, her verbiage somehow conveys the exact opposite impression.
Uh... it's all right there in your link, across several sections that analyze specific parts of the data.
> Black defendants are 45% more likely than white defendants to receive a higher score correcting for the seriousness of their crime, previous arrests, and future criminal behavior.
> Women are 19.4% more likely than men to get a higher score.
> Most surprisingly, people under 25 are 2.5 times as likely to get a higher score as middle aged defendants.
> The violent score overpredicts recidivism for black defendants by 77.3% compared to white defendants.
> Defendands under 25 are 7.4 times as likely to get a higher score as middle aged defendants.
> [U]nder COMPAS black defendants are 91% more likely to get a higher score and not go on to commit more crimes than white defendants after two year.
> COMPAS scores misclassify white reoffenders as low risk at 70.4% more often than black reoffenders.
> Black defendants are twice as likely to be false positives for a Higher violent score than white defendants.
> White defendants are 63% more likely to get a lower score and commit another crime than Black defendants.
Calling out one specific section that doesn't show bias doesn't magically exonerate the rest.
The algorithm is biased if it's giving the wrong score due to race or redundantly encoded race. To show that the algorithm is biased, you need to show that (score, race) pairs are more predictive than (score, ) singletons.
Line [36] and [46] both attempt to address this question. The only one of these which is statistically significant is "race_factorOther:score_factorHigh" in line [46].
The other things you bring up are interesting, but do not show bias. At best they show disparate impact which isn't remotely the same thing.
"Oh sure, this algorithm is much more likely to have false positives on blacks, and much more likely to have false negatives on whites, and the results are that blacks are more likely to treated more harshly by the system. But it's not biased because of the definition of bias I'm using!"
Orwell would have loved "disparate impact isn't bias" :)
Clearly statistics terminology is confusing you. The definition of bias is E[\hat{\theta} - \theta]. The definition of disparate impact is a predictor computing different means/quantiles for different protected classes.
To understand this intuitively, here's a simple thought experiment.
Consider Captain Hindsight, a predictor which returns the right answer 100% of the time. By definition, E[\hat{theta} - \theta] = 0, i.e. zero bias. (Also zero variance.)
Now suppose that blacks have a higher recidivism rate (hardly implausible, ProPublica's analysis suggests they do with p < 0.01).
Captain Hindsight - being 100% accurate and having no bias - must predict that blacks have a higher recidivism rate. Yet because Captain Hindsight predicts a higher recidivism rate for blacks, he now has disparate impact.
Seriously, you are calling standard mathematical terminology Orwellian? What's your angle here?
Your thought experiment here is incorrect, given that the analysis compares COMPAS results to actual recidivism rates and shows over- and under-prediction in comparison to them.
Clearly the idea that words have meanings beyond statistics terminology is confusing you.
And no, I'm not calling standard mathematical terminology Orwellian. What I'm calling Orwellian is your describing a biased system as unbiased (by attempting to reframe the discussion around a specific statistical definition, chosen by you).
It is consistently giving incorrectly low scores to white subjects and consistently giving incorrectly high scores to black subjects. That is clearly bias, at least in the colloquial sense.
The real question though is whether this is because the model is biased, or whether it is because the two populations have different levels of risk. If group A had a higher risk than group B, than I would expect the model to have a higher rate of false negatives for group A than for group B. This is just because the model is more likely to (correctly) assign members of group A as high risk, and some of these classifications will be wrong. To check for bias you have to control for this base rate difference.
The degree to which it does this cannot be distinguished from random chance (p > 0.05).
If the predictor were biased then you could build a more accurate score based on both the original scores and race_factorBlack:score_factorHigh (and other interaction terms). I.e. you'd be building a new bias in to cancel the old bias, leaving an accurate predictor.
Their analysis doesn't show that this is possible.
p > 0.05 is the type of cutoff you would see to get published in a peer-reviewed paper. Such a high bar of evidence is not necessary in this situation. To prevail in a civil suit, a person harmed by this algorithm would only have to prove that is more likely than not that the algorithm is biased.
If you are going to argue a single dictionary definition, you should immediately stop holding a mouse when using a computer. Rodents don't like being held for long times, be connected electrically to a machine, nor do they like being rubbed on a pad.
The data analysis you link to is by Jeff Larson, while the primary author of the article is Julia Angwin.
Larson is still the second author so it is certainly a big question how he can present data showing no statistical correlation between race and score then have his name on an article saying the exact opposite that is clearly pushing an agenda. And as noted, one where the owners of the publication are also involved in a competing risk assessment product.
It's not quite right that he shows no correlation between race and score. There is a strong correlation between race and score. This correlation is caused by the fact that blacks have a high recidivism rate (p = 4.52e-6).
What the analysis shows is that once you know the predicted score of the algorithm, using race doesn't give you extra information. If the scores were biased then you could correct them by using racial information to undo the bias.
https://github.com/propublica/compas-analysis/blob/master/Co...
In the statistical analysis (unlike the verbiage) she is completely unable to hide the lack of bias and the accuracy of the algorithm, all of which are clearly on display in line [36]. In contrast, her verbiage somehow conveys the exact opposite impression.