The modern world runs on data and statistics. In AI chatbots, to test scientific results, to predict fraud. Rock-solid numbers look neutral, provide guidance in the complex world and make patterns recognizable. But they are absolutely not value-free mathematical tools, sees science historian Iris Clever. “We place so much trust in algorithms and data, without really understanding what human choices and biases they contain.”
Take Karl Pearson (1857-1936), one of the forefathers of modern statistics. His most famous invention is the correlation coefficient, also known as Pearson’s rwhich tests relationships between different variables. What always surprises Clevers students at the University of Chicago: “Pearson was a racial scientist. He had an enormous skull collection, about seven thousand pieces, from all over the world. Often collected in a colonial context. His ultimate goal was to classify them by race using the latest statistical methods. And scientists still use those methods.”
Clever discovered Pearson through her fascination with the medical and cultural history of the body. This turned out to be strongly linked to colonialism, as she discovered during her master’s degree in history in Utrecht and her PhD in Los Angeles. This is how she learned about the Nias masks, plaster casts of the faces of residents of the Indonesian island of Nias. These were made by Dutch scientists on a colonial expedition, part of which now hangs in the Rijksmuseum. “They form a bizarre snapshot of how colonial science worked. I really wanted to know more about it.”
So she delved into the archives, looking for descriptions of scientific encounters during colonial times. Unfortunately, the colonists appeared to write little about this, both in scientific publications and in diaries and letters between racial scientists. What Clever did find: “Data, data and more data. Boxes of tables of body measurements, eighty pages of nonsense about statistics.”
Was that amazing?
“It was striking, but also logical: racial science is by definition a classification problem. They compared skull measurements with each other to divide people into different groups. That was a very common idea at the time, in the early twentieth century. And to make such comparisons, you need a lot of data and new statistics. That is exactly where Pearson is Coefficient of Racial Likeness developed: he combined all kinds of skull measurements into one number.
“A skull measurement consists of as many as 32 different measured distances that are mutually dependent on each other, and Pearson thought that his contemporaries were sloppy with this. So he initiated multivariate statistics, with the aim of classifying breeds at the very best.”
Should we overhaul modern statistics?
“Not that either. We owe a lot to Pearson’s statistics, and I wouldn’t say that statistics themselves are racist. But we have to realize where the origins of these methods lie. I want to get rid of the idea that many of my students have when they walk into my classroom: that statistics is by definition a neutral tool. It was not invented by an unsuspecting scientist and then used by racial scientists, it was developed specifically for racial classification. That seems to me to be an interesting and important historical fact.”
The further the data gets from its origin, the faster that origin is forgotten
And what about data?
“Collecting, publishing and, above all, properly preserving data was central to racial science. When classifying a skull, you benefit from a rich collection of data to compare your measurements with, and every new skull measurement adds something to that collection. Well, every new measurement… skulls that looked too strange or fell outside the existing groups were not always included in the collections. In this way, racial scientists built an artificial homogeneity in their collections. Precisely the idea behind racial science, namely that there is a strong defined varieties are then built into the methodology and data.”
“And those built-in, human values can easily travel through time and space. Even in Pearson’s time, it was a matter of sending it by post and transcribing it. And the further the data gets from its origin, the faster that origin is forgotten. This continues even today, for example in forensic software such as Fordisc, which compares unidentified bodies with reference databases to track down missing people. These databases are partly based on the old skull collections of racial scientists, including Pearson’s collection. The ideas of a century ago therefore influences the conclusions of the software.”
The scientific consensus now is that biological race does not exist. Does this mean we get rid of the problem of racialized data?
“Unfortunately not. Take for example my colleagues in medical research, here in Chicago. They train an algorithm to predict the development of tumors, based on medical scans of the tissue. It works incredibly well, and it is all based on visual data from patients. But although they do not use any data about race when training the model, and doctors could not read it from the photos, the machine can predict race with a great deal of certainty. Without them asking, without wanting to, they can their algorithm to classify people by race.
“And they were unable to figure out exactly how the model does that. Presumably with a combination of all kinds of separate properties of such a tumor, which makes it extremely difficult to isolate. In itself, the ‘race skill’ is not a problem in this case, but it shows that despite the best intentions and measures, something like this can still show up in a model. And that entails risks. Such a model could use that race skill to treat different groups of people unequally. For example, prescribing more medical supplies to white patients, higher insurance premiums for non-white patients, or referring fewer non-white patients for follow-up screenings.”
While in this case the researchers have done everything possible not to include ethnicity in the training of the model.
“Exactly. Good intentions are beautiful and terribly important, but not enough to solve the problem. Mathematician Cathy O’Neil describes it aptly in her book Weapons of Math Destruction: ‘Big data processes do not invent the future; they codify the past.’ And to understand the consequences of this, we need historical research into the origins of data. Only then can we leave behind the outdated norms and values attached to the data.”
Books that Iris Clever has used in her work.
Photo Laura McDermott/The Globe and Mail