White by Default: Bias in Criminal Racial Assignment

Misclassification Inflates White Rates 4-6%, Deflates Hispanic Rates
20-31%: Evidence from 1.5 Million Records

- We trained a multinomial logistic regression model (k = 18, n = 1.5
million) using racial probability classes extracted from DeepFace's racial
classifier and first and last name racial summary statistics from the US
census and (Rosenman et al., 2023), achieving 92.76% accuracy in three-race
classification (Black, White, Hispanic).

- A sufficiently accurate linear model trained on biased data learns the
true signal from noise. Systematic deviations indicate mislabeling by
authorities rather than model error.

- 29% of individuals predicted to be Hispanic were officially classified as
White by Department of Corrections authorities.
This pattern persisted at high model confidence (95-100%), where 22.4% of
predicted Hispanics were still assigned as White.

- Correcting for misclassification increases Hispanic criminal record rates
by 31%, decreases White rates by 6%, and decreases Black rates by 1%.

- Bias between other racial pairings was minimal and symmetrical (equal
numbers of Blacks misclassified as White and vice versa).

- State-level analysis showed no correlation with political ideology (r =
0.21, 95% CI: -0.36 to 0.67, p = 0.472), indicating random administrative
error rather than deliberate bias.

- The proportion of predicted Hispanics assigned White (r = -0.80, 95% CI:
-0.95 to -0.38, p = 0.003, n=11) and the proportion of predicted Whites
assigned Hispanic both correlated with Native American ancestry among
Latinos (r = 0.74, 95% CI: 0.26 to 0.93, p = 0.009, n=11).

…

<https://www.uncorrelated.xyz/p/white-by-default-systematic-bias>

<https://archive.ph/O7pld>