The personalised recommendations and special offers that pop up when you order books or groceries online, and even the specific sequence of questions an insurance call centre asks about your claim, are all generated by computerised predictive algorithms derived from analysing patterns, links, and associations in large sets of data. By classifying types of people and their behaviours on this basis, shops try to increase their profits by automatically targeting those of us in their databases that seem most likely to buy certain items.

But the Government is adopting such techniques for more serious matters. Software at the Department of Work and Pensions, for instance, is beginning to try to detect fraudsters by analysing the voices of people who ring its call centres — so if you ask the wrong kind of questions, or perhaps ask the right kind of questions in the wrong way, the software could decide you’re not strictly kosher.

The British Government’s Action Plan on Social Exclusion has risk prediction as its first guiding principle. The idea is to predict life outcomes and trigger early human interventions before things go wrong — in the case of the Nurse Family Partnership scheme, even before birth. In this scheme, the unborn child of a pregnant mother might be categorised as at high risk of future criminality based on factors such as the mother’s age, her poor educational achievements, drug use, and her own family background. The mother is then visited regularly at home by a nurse and helped with parenting.

In the criminal justice system too, risk prediction instruments assess the probability of adults and young people re-offending, along with a battery of other actuarial tests for predicting future sexual and violent crime. Such techniques, which are not automated in these cases, also play a central role in evaluations to determine whether a person should be committed indefinitely as a dangerous person with severe personality disorder or whether these people, once committed, are ready for release.

The U.K. Department of Health has even developed a series of predictive algorithms for scoring those patients with long-term conditions who are at most risk of re-hospitalisation. The idea is to intervene early to minimise admissions.

The Surveillance Society report from the Information Commissioner’s Office outlined worries about predictive social sorting on the grounds that it could amount to discrimination, create new underclasses and that by the totting up of negative indicators from health, school and other records, a predictive model could make its own worst predictions come true.

Now a team of British and American researchers has flagged up a more fundamental danger with these predictive models. When applied to individuals the margins of error are so high as to render any results meaningless.

The study published in May in the British Journal of Psychiatry (bjp.rcpsych.org) by forensic clinical psychologists Stephen Hart and David Cooke and statistician Christine Michie takes as its example two popular actuarial risk asse ssment tools used to predict violence (STATIC-99 and VRAG). But the team has studied many other widely used tools such as Risk Matrix-2000 (for predicting sexual offending), and found the same high margins of error across the board.

For groups of people flagged up as high risk by Risk Matrix-2000, for example, the standard estimate is that 36 per cent will re-offend sexually in the next 15 years. The team found that the true value of that estimate for a group lies between 28 per cent and 45 per cent, 95 per cent of the time — that is, the 95 per cent confidence interval. For an individual, they found the true value of the estimate was between 3 per cent and 91 per cent, 95 per cent of the time. For STATIC-99, VRAG and other tools, the results were much the same.

No one can argue with statistically based procedures for making complex decisions under conditions of uncertainty, so long as successes and failures are aggregated across cases and the cost of errors is low, says Professor Hart, who belongs to the Department of Psychology at Simon Fraser University in Canada and is a leading authority on assessing the risk of violent offending.

“A life insurance company doesn’t care whether it makes a ‘mistake’ estimating the lifespan of a given individual — it could be wrong about every single person it insures — but as long as the pattern of life spans is predictable on average, then it can still make good money.”

But if one is interested in individual cases or if the cost of decision errors is high, then these techniques are problematic. “Families of victims who are killed by patients and offenders released improperly, and those whose civil rights are infringed when they are held improperly, are not satisfied knowing that despite the errors I made in their cases, I am still right more often than I am wrong,” says Professor Hart, who has considerable experience as an expert witness in U.S. courts defending people who have been incarcerated on the basis of actuarial risk predictions.

Professor Hart highlights a further problem with predictive assessments that compounds the effects of these error margins. In cases of uncertainty, humans will tend to anchor on the first substantial piece of information they get and any new information that contradicts this initial idea is given less attention than it merits. This is the theory of anchoring bias.

If for example a predictive model says that a frail 85-year-old man with heart problems is a high-risk sex-offender and 52 per cent likely to re-offend over the next 15 years, anchoring bias means that if the assessor is told the score is wrong, they will simply adjust it to 50 per cent or 48 per cent. “It is probably zero. But the 52 per cent poisons the assessor’s judgment,” says Professor Hart.

Predictive models are attractive because they represent an apparently scientific and rigorous yet simple approach to targeting resources and making decisions about complex human problems.

But this latest study adds to concerns that such techniques are insufficiently accurate to make important decisions about individuals. —

Printer friendly page  
Send this article to Friends by E-Mail