Predictive technologies are big business these days - massive, in fact. Taking data about the behaviour of groups and using it to determine is something that is proving popular for everything from supermarket shopping to governmental decisions, from your listening habits to traffic calming.
One of the article's in today's Technology Guardian explores the predictive industry, and highlights some of the problems that a number of experts have said are arising. In essence, they are worried that taking statistical indications of group behaviour will always ignore the outlying results.
Author Christine Evans-Pughe points out an example regarding sex offenders. Standard estimates produced using one leading piece of software suggest that 36% of offenders will re-offend within 15 years (based on 95% confidence of results falling within the range of 28% and 45%) - but across the same confidence spread the individual range is from 3% to 91%. Imagine the consequences if, for example, you are that 3% offender but treated like a 36% offender. It might actually increase your chance of reoffending.
"The statistical issue of the difference between precision of estimates in a group against that for an individual is not peculiar to psychology and it is not because psychological variables are less reliable. It's to do with inherent variability in human beings," says Cooke, professor of forensic clinical psychology at the Douglas Inch Centre and Glasgow Caledonian University.
Cooke has also looked at medical literature on predicting the probability of heart attacks, cancer and other conditions that rely on physical measurements and found the same large error margins. "You can get a good model statistically that fits well for a group but it doesn't predict well for individuals, " he says.
The problem isn't really with the numbers per se, but the way their indications are applied. This was a subject picked up by our Bad Science columnist Ben Goldacre in a column last December: It's not so easy to predict murder - do the maths. But what's the solution? Less prediction? More intelligent mathematical modelling? Or just better statistical education?
