Available at: http://web.mit.edu/cjpalmer/www/AIIP-Bankruptcy-Bias.pdf
Abstract:
We document substantial racial disparities in consumer bankruptcy outcomes and investigate the role of racial bias in contributing to these disparities. Using data on the near universe of US bankruptcy cases and deep-learning imputed measures of race, we show that Black filers are 16 and 3 percentage points (pp) more likely to have their bankruptcy cases dismissed without any debt relief in Chapters 13 and 7, respectively. We uncover strong evidence of racial homophily in Chapter 13: Black filers are 7 pp more likely to be dismissed when randomly assigned to a White bankruptcy trustee. To interpret our findings, we develop a general decision model and new identification results relating homophily to bias. Our homophily approach is particularly useful in settings where traditional outcomes tests for bias are not feasible because the decision-maker’s objective is not well defined or the decision-relevant outcome is unobserved. Using this framework and our homophily estimate, we conclude that at least 37% of the 16 pp dismissal gap is due to either taste-based or inaccurate statistical racial discrimination.
Commentary:
To start with, it is necessary to bear in mind that these findings do not assert that bankruptcy judges or Chapter 13 Trustees are systematically and overtly racist, but subject to the same implicit biases, racial and otherwise, as the rest of us humans. The authors instead identify at least three possible implicit and often subconscious bases for these statistical gaps:
- Taste-based Racial Bias: This can arise if the race of the filer alters the utility she receives when a particular outcome is realized, e.g., a trustee dislikes fraud more when committed by a Black filer).
- Inaccurate Statistical Discrimination: This can result in dismissals if decision makers have inaccurate beliefs about how a filer’s race predicts the outcomes they care about.
- Accurate Statistical Discrimination: Here, while a filer's race may accurately predict differential likelihoods of outcomes, decision makers value any statistical discrimination remains potentially problematic given that the non-race characteristics used in statistical discrimination models are themselves often the product of historical discrimination.
This is a long need supplement to the earlier finding by Jean Braucher, Dov Cohen, and Robert M Lawless in “Race, attorney influence, and bankruptcy chapter choice,” Journal of Empirical Legal Studies, 2012, 9 (3), 393–429, which found that the the implicit racial biases of consumer debtor attorneys influence the chapter choices made by debtors, with a key difference being that this paper "focus[es] on documenting disparities after an individual has entered bankruptcy while holding constant all filer characteristics that existed at the time of the bankruptcy filing." As such, the racial differences found "arise mostly due to the bankruptcy system itself rather than choices that consumers make prior to filing." These differences can include, for example, White trustees exercising their discretion in part by allowing lower expenses by Black filers, leading to requiring higher required payments to creditors.
It is, however, worth noting that the authors, in assessing when any implicit racial biases might come into play in a Chapter 13 case, do inaccurately assert that Chapter 13 cases "require a confirmation hearing to approve the Chapter 13 plan", when, in fact the vast majority of Chapter 13 cases are confirmed without an actual hearing, especially one where the debtor appears personally before the bankruptcy judge. Additionally, the article assumes that it is the Means Test or structural barriers that deter debtors from filing Chapter 7 instead of Chapter 13, when in most cases Chapter 13 cases are filed either to cure delinquencies on home and vehicle loans (which cannot easily be done in Chapter 7) or because property exemptions are insufficient. Repeating my regular refrain, the academic authors of this article would certainly have benefited from direct engagement with the various boots on the ground actors in the bankruptcy system.
Also, as always with my posts regarding math-heavy law review articles such as this, I plead complete ignorance regarding the statistical analyses, especially as in this article, deep-learning AI was extensively used to both identify the race of debtors, judges and trustees, but also to evaluate outcome. I leave to others more competent to parse and dissect that most important aspect of this research.
Ultimately, while the authors do conclude with several policy recommendations, including increasing diversity in among trustees and judges and collecting more direct and clear debtor demographics (a long sought dream in the Ivory Tower) to better diagnose disparities of all types, the vital first real step would be the recognition by those trustees and judges that they are not immune to implicit biases and must confront and struggle against these, just as we all must.
With proper attribution, please share this post.
To read a copy of the transcript, please see:
Blog comments