When we try to predict the behaviour of an individual, a worker or a company, we often only have a handful of time periods for any one person or organisation. For this reason, researchers have traditionally used a statistical tool called shrinkage estimation. If we don’t have enough information to make an accurate individual forecast, we look at the group average and pull that person’s forecast towards the middle.
In a recent study, we show why this approach, while common, often leads to unfair and inaccurate results for the very people who stand out the most (Giacomini et al, 2025).
Overcoming the tyranny of the majority
Our research highlights a major flaw in standard statistical methods, which is called the tyranny of the majority. Traditional methods are designed to be as accurate as possible for a whole group on average. To achieve this, they use ‘shrinkage’ – a process that takes an individual’s noisy prediction and moves it towards the group mean. The logic is that if we have very little data on you, it is safer to assume that you are just like everyone else.
But we find that this logic creates a serious problem for anyone who is actually different. If you are an outlier – perhaps a teacher who is exceptionally effective or a company with unique hiring practices – the standard method will almost certainly misjudge you. It effectively punishes you for being different by assuming that your unique results are just random noise. This makes the system look accurate on paper for the majority, but it fails the individuals who matter most.
We have developed a new method to fix this, which we call individual shrinkage. Rather than borrowing strength from the crowd, our approach looks at a person’s own individual history to decide how much to trust their data.
If your past record is consistent, our method gives your current data more weight. If your record is very messy, it leans more on the overall average, but it does so in a way that is tailored specifically to you.
We find that this method is especially powerful in groups where many people are significantly different from the average. By focusing on the individual rather than the crowd, we can provide much more accurate predictions without forcing everyone to look the same.
Accuracy at the individual level is what counts
We argue that in the real world, accuracy at the individual level is often more important than being right for the group as a whole.
Think about a school deciding which teachers to keep, a hospital receiving rewards based on patient outcomes, or a counsellor giving personalised financial advice. In these cases, a mistake for one person isn’t cancelled out by a correct guess for someone else. We need the mathematics to be right for that specific person because the stakes – jobs, health and livelihoods – are incredibly high.
To prove the value of our method, we revisit a famous study of how large US employers discriminate against job applicants based on gender and race (Kline et al, 2022). When we use our new approach, we discover that the level of systemic discrimination was actually higher than previously thought.
The older methods were ‘shrinking’ the results of discriminatory firms towards the average so much that their bad behaviour was being masked. Our method identifies more biased firms because it doesn’t try to average away their behaviour.
A key part of why our method works is that it is cautious. We use a decision-making framework that minimises the ‘worst-case’ mistake. Instead of making a risky bet that everyone fits a certain statistical mould, we choose the prediction that would cause the least amount of regret if our assumptions were wrong. This makes our method much safer to use in the real world.
Changing how we use data to judge people
Our research represents a call for a shift in how we use data to judge people. We want to see this ‘individual-first’ approach used in more policy settings, from education to healthcare. It is a fairer way to use statistics because it respects human variety rather than trying to erase it. When a computer model can potentially decide a person’s future, that individual deserves to be judged on their own merits, not just by how they compare to a majority group of which they might not be part.
Ultimately, we want to move towards a world where data help us to understand people’s unique stories. By using individual shrinkage, we can make sure that our predictions are as unique as the people that they are meant to describe. This is not just a matter of better mathematical analysis: it is a matter of making fairer decisions for everyone.
