THIS ARTICLE APPEARED IN AUG

THIS ARTICLE APPEARED IN AUG. 9-15, 1997 ISSUE OF "THE ECONOMIST"

SCIENCE AND TECHNOLOGY

Social research

Asking for trouble

SPARE a brief thought for journalists, forced to extract salacious gossip on the seamier sides of life from people unwilling to reveal it. Only a brief thought, mind-but now consider how much worse is the plight of a more socially useful group, those who have to compile not just anecdotes, but meaningful statistics on the prevalence of the grimmer social and physical ills, such as rape or HIV infection.

People are often scared to report these things, so the official sources for such statistics (e.g., police or hospital reports) are likely to be underestimates-possibly big ones. Indeed, for some activities, such as heroin use, no useful numbers exist at all.

The solution, according to Peter Killworth at the University of Southampton, in England, Russell Bernard at the University of Florida, Gainesville, and their colleagues, is simple. Instead of asking people, ask their friends.

The researchers reason that although a lot of information never makes it to the ears of officialdom, gossip gets everywhere. And though people do not like to reveal embarrassing facts about themselves, even when promised anonymity, they are quite happy to spill the beans about others-so long as they do not (unlike in their dealings with journalists) have to name names.

What is needed is a technique to turn this gossip into scientific information. In a paper recently accepted by Evaluation Review, Dr Killworth and his colleagues present their latest attempt to do just that.

The theory is straightforward. Find out, for each member of a representative sample of a population, how many individuals that person knows altogether, and how many he knows in particular groups of interest (eg, drug users, or women called Olivia). From the answers, you can calculate how big the group of interest is as a proportion of the overall population.

The first attempt to do this was led by Dr Bernard, when he tried to improve on official estimates of the number of deaths in the Mexican earthquake of 1985. As he discovered then, the practice is more complex than the theory. For one thing, the basic theory assumes that all people have an equal chance of knowing members of a certain group. Clearly this is false. Someone who is himself HIV-positive, for example, is likely to know many more people with the virus than someone who isn't. This kind of effect might average out-some people have more HIV-positive acquaintances, some fewer-but that is untested.

The theory also assumes that people know everything about their acquaintances. This, too, is obviously untrue, especially when the subject of the enquiry is shady or stigmatised. But more knowledge about social behaviour can help compensate for this. For example, a study led by another member of the collaboration, Gene Shelley of Georgia State University, suggested that when people test positive for HIV they cut down their circle of contacts by, on average, a factor of three.

A more immediate problem, however, is that people are not very good at working out how many other people they know. So Dr Killworth and his colleagues had to do it for them. They asked each respondent in their survey not only about the three groups whose sizes they wanted to estimate (women who had been raped, people with HIV and homeless people), but also about 29 reliably known quantities, such as diabetics, jailbirds and men named David. The numbers of people a respondent knew in those groups could be used to infer how big his circle of acquaintances was.

This method has other advantages. The researchers could check its performance by seeing how well it predicted the known group sizes. In future, they might work out how good an informant each person in a survey is by seeing how well he estimates the sizes of all the known groups. This would allow them to give more weight to better informants when totting up the estimates for the three unknown groups.

Their current results, taken from a survey of 1,554 respondents conducted in late 1994, were that America then had 800,000 HIV-positive people, that 194,000 women had been raped in the previous 12 months, and that 526,000 people were on the streets (give or take 5-10% in each case). These fitted within the range of estimates arrived at by other methods. In 1990, America's Centres for Disease Control collected estimates of between 300,000 and 1.5m cases of HIV infection nationally; estimated rape figures in the same year varied from 130,000 to 683,000; and official estimates for the number of homeless people have varied between 200,000 and 600,000.

So the method appears to work, and may thus be able to point out just where in the range of previous estimates the truth lies. But there is more to be done. For instance, the model tends to exaggerate the size of small groups and diminish that of large ones. The researchers are now hoping to learn more about how information spreads through social networks (or rather, why it sometimes does not)-data which they hope will correct this anomaly. Using them they can then refine their model to compensate for gossip's imperfections.