Citizens DCI Audit Summary Provided by Citizens

Citizens is required under New York City Law (Local Law 144) to conduct what is known as a “bias audit” in connection with automated employment decision tools (AEDTs) Citizens uses to screen applicants for hire. The bias audit is conducted by a third-party auditor. Citizens uses the results of the bias audit to ensure that the AEDTs do not discriminate based on any protected characteristics (e.g., gender, ethnicity).

A bias audit produces “bias audit tables” for the AEDTs used by Citizens. The bias audit itself considers things like historical applicant demographics and how the AEDTs used by Citizens score job-related competencies. A bias audit table must account for how various groups were scored by the AI tools, and the bias audit tables are sorted by gender, ethnicity, the intersection of gender and race, the number of applicants, the “passing rate” of each group, and the “impact ratio” for each demographic group.

Note, the “passing rate” refers to what percentage of applicants passed the assessment at a given tier. The “impact ratio” takes the group with the highest passing rate and uses it to divide all other groups’ passing rates to create a ratio. That ratio expresses the difference in passing rates between demographics (e.g., between male and female applicants). These figures are provided for each classification of scores the AEDT provides.

The AEDT Citizens uses classifies scores into “bottom”, “middle”, and “top” tiers by percentile-ranking scores and grouping the scores into thirds. The bias audit table then compares the passing rates of these scores between these three tiers. Each AEDT scoring algorithm can have up to 40 comparisons across gender, ethnicity, and the intersection of gender and race across the score comparisons.

Although Local Law 144 does not define “bias” or what is “good” or “bad” in terms of passing rate differences, to gain a rough estimate of where a passing rate might be separated by a wide margin, Citizens uses what is known as “the EEOC’s 4/5ths rule.” The 4/5ths rule considers impact ratios below 0.80 as being practically significant. This 4/5th rule is not a definitive indication of bias, but it is a helpful benchmark for understanding the results.

When Citizens uses AEDTs for hiring decisions, Citizens uses 16 different competency-based scoring algorithms, allowing for up to 640 total comparisons. Also note that apparently large differences in passing results can come from instability in the underlying data, due to small sample sizes. (A small sample size is typically a group which is less than 2% of the overall applicant population, or less than 30 individuals.) Small sample sizes often lead to misleading results. Differences can also come from the grouping or “aggregation” of data from multiple sources The bias audit table below aggregates data.

In total, 57 comparisons (out of 640 comparisons or 8.9%) resulted in an impact ratio below 0.80. These results were mainly concentrated in the top tier scores (40 impact ratios below 0.80) across all groups, as compared to middle and bottom tier scores.

Although NYC Local Law 144 requires these top tier to other tier comparisons be charted in the bias audit table (due to how the AEDT classifies scores), employment decisions are not solely made by simply selecting individuals grouped into the top tier. Most impact ratios were also concentrated in the intersection of gender and ethnicity (42 impact ratios, with 30 coming from top-tier comparisons), which have smaller group sizes than gender or ethnicity alone. 15 impact ratios below 0.80 were found in ethnicity, with 10 of them coming from top tier comparisons. Finally, there were no impact ratios below 0.80 for gender.

This summary is provided for informational purposes only. Please review the bias audit tables below, which provide the full and accurate results from the bias audit. For more information, please feel free to contact us at citizenscandidatesupport@citizensbank.com.