Protected Characteristics and Proxy Variables
Understanding Protected Characteristics
Protected characteristics form the foundation of fair treatment legislation and define which attributes cannot be used to discriminate against individuals. In the UK, these characteristics include age, disability, gender reassignment, marriage and civil partnership, pregnancy and maternity, race, religion or belief, sex, and sexual orientation.
These categories represent legally protected groups that employers, service providers, and organisations must consider when making decisions that affect people’s lives.
When testing AI systems for bias, practitioners must identify which protected characteristics might be relevant to their specific applications. For example, a recruitment algorithm might inadvertently discriminate against candidates based on age or gender. A credit scoring model could unintentionally disadvantage applicants from certain ethnic backgrounds.
A healthcare diagnostic tool might show different performance rates across various demographic groups. These examples illustrate how protected characteristics can manifest through proxy variables or direct discrimination.
Organisations must understand that protected characteristics are not merely categories to avoid but represent fundamental human rights that must be respected. The Equality Act 2010 requires that decisions affecting protected characteristics must not result in disadvantage or discrimination. Practitioners conducting AI bias testing must identify these characteristics early in their testing framework to ensure proper evaluation methods are applied.
Recognising Proxy Variables
Proxy variables represent indirect indicators that may capture protected characteristics through seemingly neutral data points. These variables often appear in datasets as demographic information, geographic locations, educational backgrounds, or other factors that correlate with protected characteristics. For instance, postcode data might proxy racial composition of neighbourhoods, while educational institution names could indicate socioeconomic status or ethnic background.
Consider a university admissions algorithm that uses academic performance, extracurricular activities, and personal statements as inputs. The proxy variables might include school type, which could indirectly represent socioeconomic status or geographic location. A candidate from a school in a historically disadvantaged area might have different opportunities, yet these factors don’t directly indicate protected characteristics.
However, they can create outcomes that disproportionately affect protected groups through compounding effects.
- Postcode data often correlates with ethnic composition and socioeconomic factors
- Occupational titles may proxy gender or ethnic backgrounds
- Geographic location can indicate access to resources or historical disadvantage
- Historical data patterns might reflect past discriminatory practices
Identifying proxy variables requires careful data exploration and understanding of how seemingly neutral features might unintentionally capture protected characteristics. Practitioners must examine datasets for these correlations through statistical analysis and domain knowledge. The goal is to detect when neutral-looking data points might create discriminatory outcomes through indirect pathways.
Testing Approaches for Protected Characteristics
Effective testing requires practitioners to examine outcomes across protected characteristic groups rather than treating them as single categories. For example, rather than simply testing gender as a binary variable, analysts should consider intersectional approaches that examine combinations of characteristics. A woman from an ethnic minority group might experience different outcomes than a white woman or a man from the same ethnic group.
Statistical tests must focus on meaningful differences in outcomes rather than absolute disparities. Practitioners should apply techniques such as demographic parity, equalised odds, or calibration across protected groups. These approaches measure whether similar individuals from different protected characteristic groups receive similar treatment from AI systems.
The focus should always be on outcomes rather than inputs or proxy variables alone.
Organisations must develop testing frameworks that explicitly account for protected characteristics through stratified analysis. This involves creating subgroups based on protected characteristics and measuring performance metrics within each group. The testing process should identify whether any protected group experiences significantly different outcomes, regardless of whether these differences result from direct discrimination or proxy variables.
When proxy variables are identified, practitioners must determine whether these represent legitimate business requirements or potential sources of bias. The testing framework should include validation steps that examine whether proxy variables create discriminatory outcomes. This validation process involves comparing results across similar individuals who differ only in protected characteristic categories, rather than through proxy variables.
Effective bias testing requires practitioners to maintain detailed documentation of their approach to protected characteristics. This documentation should explain how proxy variables were identified, what testing methods were applied, and how results were interpreted. The goal is to create defensible processes that can withstand scrutiny and demonstrate compliance with relevant legislation.