Demographic Parity

Choosing to select an equal proportion of applicants from every group is the core idea behind Demographic Parity. This fairness metric requires an algorithm's positive outcomes to be equal across different groups, regardless of other factors. For example, if a university's automated admissions tool selects 15% of female applicants, it must also select 15% of male applicants to satisfy this standard.

Ask this: Is the outcome rate the same for every group this system evaluates?

Demographic Parity, also known as Statistical Parity, defines fairness as statistical independence between a decision and a protected attribute like race or gender. The core principle is that the proportion of individuals receiving a positive outcome should be the same for all defined groups.

This means if an algorithm screens candidates for a job, the acceptance rate for men, women, and non-binary individuals should be roughly equal. Under Demographic Parity, the model's outputs must be statistically independent of gender on average, though a model may sometimes use gender as an input to enforce this very outcome.

Consider a different domain: targeted advertising. An ad platform achieving demographic parity would show a high-paying job advertisement to an equal percentage of its users across different age groups, rather than showing it predominantly to younger users who might have a higher click-through rate.

Another example is in pretrial risk assessment. An algorithm with demographic parity would flag the same proportion of defendants from different racial backgrounds as 'high risk' of re-offending, irrespective of historical crime data for those groups.

The major challenge with this approach is that it ignores whether the actual underlying rates of qualification or risk (the 'base rates') differ between groups. Forcing equal outcomes can lead to selecting less-qualified candidates from one group or overlooking qualified ones from another to meet a statistical target.

This makes Demographic Parity a powerful but controversial tool. It's often used to correct for historical biases or in situations where qualification is subjective, but it can conflict with merit-based selection when objective measures differ across groups.

Demographic Parity is one of the earliest and most intuitive formal definitions of algorithmic fairness. Research focuses on its statistical properties, its relationship to other fairness definitions, and the societal implications of enforcing it, especially when it conflicts with individual fairness or model accuracy.

• Dwork et al. (2012) contrasted individual fairness with group-level metrics, discussing statistical parity as a condition where outcomes are independent of protected attributes. [1]
• Feldman et al. (2015) analyzed how to measure and remove disparate impact, a legal concept closely related to Demographic Parity, and proposed methods to transform data to satisfy this fairness constraint. [2]
• Kleinberg, Mullainathan, and Raghavan (2016) demonstrated a key impossibility result in fairness: when base rates differ between groups, it is impossible for a model to simultaneously achieve calibration within groups and equal error rates (such as those required by Equalized Odds). [3]
• Verma and Rubin (2018) surveyed over 20 definitions of fairness, categorizing Demographic Parity as a 'group fairness' metric that focuses on equalizing outcomes, in contrast to metrics that focus on equalizing error rates. [4]

The primary criticism of Demographic Parity is that it can lead to perverse outcomes when the underlying base rates of the trait being measured (e.g., qualification for a job, risk of loan default) are not equal across demographic groups. Forcing equal outcomes can mean denying opportunities to more qualified individuals in one group to meet a quota for another, which can be seen as unfair at an individual level. It can also reduce the overall utility or accuracy of the model, a tension often called the 'fairness-accuracy tradeoff'.