Our goal is twofold: one, to understand people’s ideas of fairness about data and risk assessment; two, to explore the history of risk assessment with a focus on banks and insurance companies.
Earlier1 See Class Notes Week #3. we discussed how data can be biased: reliance on proxy variables that weakly track the outcome of interest; differences in sample size for data about socially relevant groups; data mirroring existing social biases; data not equally representative; etc.
We now ask a different question. What sort of data would it be (un)fair to use in predictive algorithms, focusing on lending and insurance policy decisions? This question can be examined from different disciplinary angles. In the interest of interdisciplinarity, we examine this question by looking at Barbara Kiviat’s research in economic sociology.2 Kiviat (2021), Which Data Fairly Differentiate? American Views on the Use of Personal Data in Two Market Settings, Sociological Science 8: 26-47
Kiviat constructed a representative sample of the US population and asked participants questions about what types of data they would consider fair to use in lending and insurance policy decisions. Data included things such as: accident history, speeding tickets, driving style, credit score, when and where a person drives, Zip code, rent payment on time, utility and TV bill payment on time, income. Check out Kiviat’s research findings on pp. 34-35 of the paper.
Making decisions about individuals based on group characteristics, averages, long-term tendencies seems objectionable in its own way:
“predictions are about what is true on average for a group of people, and yet those who receive treatment are specific individuals.” (p. 39)
Survey respondents, however, had a more discerning view. They distinguished between morally acceptable averages and morally unacceptable ones. Respondents said:
“A person might have to work 3rd shift or be a drug dealer. Both work at night and shouldn’t be judged on the same criteria.”
So they made a distinction between groupings:
“The problem is not that people have been grouped but that they have been inappropriately grouped. When the boundaries of a group capture too many different sorts of people and behavior, holding each individual responsible for the average is likely to assign blame where it does not belong, a sort of moral ecological fallacy.” (p. 39)
Current debates about algorithmic fairness happened before in different ways under different terminologies in 1970s. Instead of algorithmic fairness, the more common terminology back then was actuarial fairness. Following the work of Ochigame3 Ochigame (2020), The Long History of Algorithmic Fairness, Phenomenal World. Check out Ochigame’s alternative search engine at searchatlas.org, it is instructive to sketch a history of fair risk assessment, reaching back to the 17th century when the concepts of probability and chance were first theorized mathematically.
Here is Pascal in a letter to Fermat (1654): This is the way I go about it to know the value of each of the shares when two gamblers play, for example, in three throws, and when each has put 32 pistoles at stake. Let us suppose that the first of them has two (points) and the other one. They now play one throw of which the chances are such that if the first wins, he will win the entire wager that is at stake, that is to say 64 pistoles. If the other wins, they will be two to two and in consequence, if they wish to separate, it follows that each will take back his wager that is to say 32 pistoles. Consider then, Monsieur, that if the first wins, 64 will belong to him. If he loses, 32 will belong to him. Then if they do not wish to play this point, and separate without doing it, the first should say–I am sure of 32 pistoles, for even a loss gives them to me. As for the 32 others, perhaps I will have them and perhaps you will have them, the risk is equal. Therefore let us divide the 32 pistoles in half, and give me the 32 of which I am certain besides. He will then have 48 pistoles and the other will have 16.
Notions of fairness were closely connected with questions of chance, uncertainty and the future. For example, consider the problem of points that prompted early developments in probability theory. Suppose we play a game in which if I get three heads first, I win and get all the money on the table. If you get three tails first, you win and get all the money. Suppose the game is interrupted. Say I got two heads, and you got one tail. How should we fairly distribute the money on the table? Each of us has contributed equal amounts of money.
Algorithms and mathematization were also seen as the solution to interminable disputes, war, social divisions and diplomatic intractability as in Leibniz’ calculemus!
Condorcet (1743-1794) and Laplace (1749-1827) set to mathematize the moral sciences, especially jurisprudence and legal decision-making. For example, Condorcet’s jury theorem shows that the accuracy of jury decisions improves monotonically the more people are involved, assuming decisions are by majority voting and each juror is even slightly better than chance at tracking the truth. Hence, juries should decide by majority voting and consist of several people.
This quantitative and probabilistic approach to the justice system fell into disrepute in the 19th century, partly because of the fall of mechanistic assumptions about human nature. But this did not mean that probability itself feel into disrepute.
Ochigame writes:
From Hacking (1990), The Taming of Chance, Cambridge UP: The avalanche of numbers, the erosion of determinism, and the invention of normalcy are embedded in the grander topics of the Industrial Revolution.
“At the end of the Napoleonic era, the sudden publication of large amounts of printed numbers, especially of crimes, suicides, and other phenomena of deviancy, revealed what philosopher Ian Hacking calls”law-like statistical regularities in large populations." These regularities in averages and dispersions gradually undermined previously-dominant beliefs in determinism"
Life insurances in the 19th century relied on risk assessment. They were soon charged with racial bias:
“Risk classification soon surfaced controversies over racial discrimination: in 1881, life insurance corporations started to charge differential rates on the basis of race … When civil rights activists challenged this policy, the corporations claimed differences in average mortality rates across races as justification.”
Two views about “fair” pricing emerged, one closely tied to the accuracy of the prediction based on current available data, and another more aspirational:
Can this disagreement about fair pricing of insurance policy be instructive to navigate current debates about algorithmic fairness?
“…in 1884, Massachusetts state representative Julius C. Chappelle—an African American man born in antebellum South Carolina—challenged the fairness of the policy and proposed a bill to forbid it. The bill’s opponents invoked statistics of deaths, but Chappelle and his allies reframed the issue in terms of the future prospects of African Americans, emphasizing their potential for achieving equality.”
Actuarial risk assessment became widespread in criminal justice (parole, sentencing) in 1980s, although the idea had been around for a while, at least as early as 1920s.4 See Harcourt (2006), Against Prediction: Profiling, Policing, and Punishing in an Actuarial Age, Chicago UP Credit scoring started to be widely adopted by financial institutions in 1960s. Standardized testing became widely adopted about the same time.5 See e.g. Elementary and Secondary Education Act (1965) as well debates about the fairness of standardized tests in Hutchinson and Mitchell (2019), Translation Tutorial: A History of Quantitative Fairness in Testing
Here is how the concept of “actuarial fairness” came about, in response to charges of racial bias. Ochigame notes:
“In the 1970s, at the height of controversies surrounding redlining, U.S. civil rights and feminist activists argued that risk classification in the pricing of insurance was unfair and discriminatory … the insurance industry disseminated the concept of “actuarial fairness”: the idea that each person should pay for her own risk." What conception of algorithmic fairness is Actuarial Fairness more closely related to among the ones we examined in earlier classes?
Ochigame describes the difference between the debates about actuarial fairness in the 1970s and today’s debates about algorithmic fairness, as follows:
“In the 1970s, proponents of”actuarial fairness" simply equated it with predictive accuracy; they posed fairness as equivalent to the optimization of utility in risk classification. Today, proponents of “algorithmic fairness” tend to define fairness and utility as distinct, often competing, considerations. Fairness is generally considered a complementary consideration or constraint to the optimization of utility, and proponents often speak of “trade-offs” between fairness and utility. This distinction responds to a widespread recognition that the conventional optimization of utility in actuarial systems—typically the maximization of profit or the minimization of risk—can be inherently unfair or discriminatory."