Breaking the Bias – Lessons from Bayesian Statistical Perspective

Equitable and fair institutions are the foundation of modern democracies. Bias, as referring to “inclination or prejudice against one person or group, especially in a way considered to be unfair” is naturally anathema to such ideals. Unfortunately, efforts to reduce impact of bias have been slow and difficult. This is because it is scientifically acknowledged that humans are innately inclined to bias. A significant share of those biases isn’t even known to persons who are supposed to hold them: hence called ‘unconscious bias’. It is also established in many scientific studies that members of a group which was supposed to be biased against also hold the same biases towards their own group (e.g., women being biased against women). So why is bias so inherent part of our biology?

Consider the early man in Savannah. He is suspicious of – as in, biased against – people who don’t look like him, bcause people who don’t look like him are usually from different tribes and people from different tribes are usually at war with his tribe. It’s s survival heuristic his brain has evolved. This is no different from other heuristics we are still encumbered with in the modern world which have roots in the early environment, such as ‘fat is tasty’, ‘sweet s edible’, etc. One way we can model bias is by borrowing from one of the canonical formulae in statistics: Bayes’ Law.

In English, and in the context of this discussion, this means that my magnitude of bias against a certain group, after observing an incident related to that group, is the output of multiplication of two numbers: what’s the magnitude of bias I held against that group before the incident (“prior”), and how directionally aligned that incident was to direction of bias (“evidence”). This change in magnitude of bias is called Bayesian update. In other words, if I was highly biased against the group, but recently observed an incident where a person/group behaved in the way opposite of what bias would have expected, then my bias against that group will reduce.

How can we leverage this Bayesian perspective to understand our fight against bias? One way focuses not on reducing bias, but on reducing the impact of bias: by reducing reliance on human decision making by better use of data; and by tracking the impact of bias on outcomes, and aiming for equity in outcomes by giving preferential treatments to protected groups. However, the majority of sociological and psychological methods of reducing bias typically focus on making people aware of biases, reiterating the need for reducing bias, and counting on the moral compass for people to behave differently (first two articles on the Google search). In Bayesian formulation, these efforts focus on reducing prior and are valid means to combat bias. Yet, prior is often formed in early childhood and adolescent experiences, and is heavily influenced by family, upbringing, and peer group. Hence, changing the prior when the employee joins the corporate workforce may be a tad too late.

However, little effort is done on exposing more favourable evidence. Occasionally there is a conversation to have biased people interact with people from the protected groups so that they realize that their biases are not fair. This largely occurs in a context where people are not just biased but actively antagonist to each other, but rarely in a corporate context. Nevertheless, the value of this in reducing bias cannot be underemphasized. Some ways evidence can be introduced in the workplace setting to counteract against prior is by designing teams and interactions which forces different people to interact with each other. Multiple repeated interactions will help move prior sufficiently to reduce bias over time. Reshuffling teams or seating to make them mix-gendered teams, for an instance, could achieve this outcome.

There is a different Bayes’ Law in mind of each person holding bias for each of the protected groups. While we talked about affecting prior and evidence, we shouldn’t forget the concept of ‘group’. Workplace efforts to reduce biases are competing from Bayesian update of bias from evidence/incidents from the external world as well, which makes this a challenging task. Hence reframing the group into smaller groups actively can also reduce bias, at least in the workplace setting. For instance, group of “women” can be reframed as group of “working women”. Generating positive evidence for working women can be easier in business settings – and will have less competition from the external world – and so will help in reducing bias faster, even as we may not be able to reduce bias against all women”. How to do that in practice? We will have to modify our language, word choices, and reframe the context. While this is not ideal – none of this really is – but we are working with what we can manage effectively, and hopefully eventually reach the ideal outcome in long term.

In the end, we must remember that tendency to be biased was a useful primordial heuristic, thus innate in human nature, and can only be reduced to the extent of the world we live in, but continuous focus on bettering prior, positive evidence, reduced level of human involvement in decision making, and reframing the group context will surely make a dent. It already has.