Confidence interval per slice

[delcenjo]

A small slice can show a large regression by chance. Reporting a bootstrap interval on each slice's delta, or borrowing the significance test from evalgate, would separate real slice regressions from noise.