Standard Biases? the Use of Race in Selection Devices for Decision-Making

Background and Purpose

Racial inequality persists despite major advances in formal, legal racial equality. Scholars argue that implicit bias and other forms of “new prejudice” combined with subjective organizational decision-making practices perpetuate racial inequality. The standardization of decision-making promises to eliminate the subjectivity that allows implicit bias to creep into consequential decisions. We compare “selection devices” – simple quantified tools for decision-making – in four distinct contexts: criminal justice, child welfare, consumer credit, and undergraduate admissions in order to challenge the idea that standardized decision-making tools will always serve to eliminate bias.

Methods:

We used comparative historical sociological methods in order to compare the use of selection devices across fields. In selecting cases, we sought variation in both the explicit agendas of the decision-making tools and the kinds of variables used in the devices, while holding constant the form of the tool itself (simple additive models). Data is drawn from primary sources such as historical records, documents related to the development of the tools, and interviews with people who utilize them. Additionally, secondary sources, such as debates about the selection devices themselves and other scholarship related to each case provide context to the data. Content analysis was employed to map the different dimensions of how race was approached both within and across tools. 

Findings: 

Models in each field vary in their approach to race within selection devices. Early parole models encoded race as a risk factor for recidivism. The overt racism of this model reflected the overt racism of the times. Credit scorers were interested in predicting defaults in order to improve loan decisions and would have been delighted to use race in their credit scoring systems. However, banks believed, accurately, that the explicit colorblindness of credit scoring models would insulate them against charges of overt racism. Child welfare agencies, in contrast, had an explicit antiracist agenda of bias elimination. Structured decision-making purported to improve the accuracy and reliability of child welfare decisions as well as eliminate implicit biases and in the system. The differences in intentions between credit scoring and child welfare tools did not result in differences of the tools themselves. The University of Michigan’s affirmative action system explicitly incorporated and valued race in its scoring system in the interest of promoting diversity. Because racial minorities suffer from so many cross-cutting disadvantages, Michigan’s point system put a positive value on the race variable – in sharp contrast to the use of race in early parole models.

Conclusion and Implications:

The data demonstrates that standardization must be understood as a heterogeneous practice capable of producing very different outcomes. We argue that standardization efforts that produce colorblind selection devices – those that explicitly avoid using race as a variable – function similarly whether they are driven by antiracist agendas (as in the case of child welfare) or status quo agendas (as in credit scoring). This structural similarity should make us skeptical of the capacity of colorblind devices to serve antiracist agendas.