> The predictor is probably not biased against any particular race - the race_factorAfrican-American:score_factorHigh term is not statistically significant. Or, as ProPublica puts it, it's "almost statistically significant".<p>This is not how you interpret p-values. Not attaining statistical significance (insofar as you think that that's a good measurement to start with) means there isn't sufficient evidence to confirm that a factor truly makes a difference. It is not proof that no racial bias is present -- to prove a negative with the same 95% confidence we expect elsewhere, the analysis would need to have 95% power to detect the bias if it did exist... and unfortunately it just so happens that interaction effects always have lower power for the same sample size than first-order effects.<p>The author also accuses ProPublica of cherry-picking numbers but then picks the one number from the entire article that is least convincing (one of the numbers on predictive accuracy of the algorithm for black vs. white people) and ignores the other statistics about predictive accuracy that are mentioned in the article.<p>Do I necessarily want to defend ProPublica here? I dunno, I haven't gone through the analysis they did for the article in depth. I also like the author's note that statistical models are often wonderful tools for reducing human biases and (in this case) reducing racism. But ultimately I do feel the author is being very hostile when ultimately his rebuttal is pretty feeble.
I'm still confused if race is a feature in the original model. More over, the offense and re-offense rate of black people is also tied to racial bias by the police.<p>If the police let more speeding white guys off with a warning than black (as a rate) then the bias is applied at the data level.
yummy: I'm commenting here because I can't figure out how to comment on your blog. :(<p>As a potentially useful but minor correction: Sorelle, who you mention in your previous post, is a "she".