Researchers have analyzedÂ data about mobile phone use in Rwanda to figure out how wealthy a phone’s user isÂ â and they say they might be able to do the same kind of analysis for any other country.
The study, published today in the journal Science, applies big-data models toÂ look atÂ much more than income. The Rwandan data, for example, could be massaged to predict which phone users owned a motorcycle or a TV.
Joshua Blumenstock, the study’s lead author and an information scientist at the University of Washington, isÂ now working on a follow-up project to see how easily the computer models can be applied to places beyond Rwanda.
“In every country, we hypothesize that there’s a relationship between how people use their phone and how wealthy they are,” he said in a Science podcast. “The exact nature of that relationship is going to change from one country to another, and it might even change from one year to the next within a country. But fundamentally, you’d think that there are these relationships that exist.”
Why Rwanda?Â Blumenstock and his colleaguesÂ chose that African country because they wereÂ interested in finding new approaches to demographic profiling in developing countries, where it’s difficult and expensive to get good survey data.
“Our focus is on modeling poverty in a developing country because that’s the sort of application where I think the need is greatest, and it is in these contexts where alternative sources of demographic data are often unreliable, out of date or non-existent,” he told GeekWire in an email.
The researchers took advantage of the fact that the mobile phone penetration rate isÂ more than 70 percent in Rwanda. That’s typical for developing countries, where land lines are scarce and mobile rates are much lower than they are in the United States. “You can get by with 50 cents or a dollar a month, and send several SMSâs and make several phone calls,” Blumenstock explained.
The researchers were able to get detailed, anonymized metadata on a year’s worth of interactions involving 1.5 million subscribers in the country’s biggest mobile phone network. The data didn’t reveal the content of the calls; rather, they dealt with metrics such as call duration, time of call and location.
The team also surveyed 856 subscribers to develop detailed profiles of their economic status, including whether they had electricity and whether they owned goods such as refrigerators, TVs, bicycles, scooters or radios.
The phone statistics and the survey data for individual users were blended together in a series of computer models to come up withÂ the best fit. The models identified patterns of phone use that were particularly predictive of wealth.
For example, people whose calls were concentrated during the 9-to-5 work hours tended to be wealthier than those whose calls didn’t follow that pattern. People who bought $10 worth of pay-as-you-go service at a time tended to be richer than those who bought their service in increments of 50 cents to $1. Rwandan mobile subscribers have to pay when they make a call, but not when they receive a call â so people who made proportionately fewer calls tended to be poorer.
The performance of the model varied, depending on what the researchers were trying to predict.Â The model did a goodÂ job of predicting which users were in the bottom 25 percent of the wealth spectrum (with a score of 0.81 on aÂ statistical scaleÂ known asÂ “area under the curve”). It was also pretty good at predicting who had a refrigeratorÂ (0.88) or a television (0.84). But it wasÂ no better than random picks when it came to predicting who ownedÂ or didn’tÂ ownÂ a radio (0.50). Part of the reason may be that so many Rwandans haveÂ radios, it’s hard to pick someone without one.
The second part of the experiment was to see how well the model did with the larger sample of 1.5 million subscribers. When the researchers compared their numbers withÂ Rwandan census data, they found that the model was surprisingly accurate for mapping the country’s distribution of rich and poor. “The correlation was more than 90 percent,” Blumenstock said.
He said the computerized analysis took four weeks and cost just $12,000, compared with the millions of dollars it would have cost to do a nationwide survey in Rwanda. The study concluded that if statistics on mobile phone usage are analyzed in the right way, they’re good enough to guide economic policy in developing countries.
But whatÂ about privacy concerns? It’s a little scary to think that someoneÂ could figure out how wealthy you are, based on your mobile phone usage. And it’s not just mobile phones. “There’s nothing special about phone data. … You could imagine doing something similar with a different data stream, be it Twitter data or Facebook data,” Blumenstock said.
Here’s what Blumenstock said about the dark side of big data in his email:
“Once outside of the controlled research environment, I do think there’s legitimate reason to be concerned that derivative methods could be used in ways that many people âÂ including myself â would find objectionable. To a certain extent, this is already happening. In fact, part of the inspiration for this project was the idea that we might be able to target poverty as effectively as Google targets advertisements, or Amazon targets product recommendations; the idea that we could take state-of-the-art methods developed primarily for commercial purposes, and translate them into humanitarian tools. But to the extent that this paper helps advance the state of the art, it might enable other applications by commercial entities. …
“So more generally, I think this study helps highlight the need to put in place careful protections and regulations over what happens to the data we generate. And not just our mobile phone data, but data generated in our day to day use of social media, e-commerce, credit cards, mobile apps, and so on. Exactly what those protections would look like is an area of active debate â some of my colleagues at UW and Berkeley are thought leaders in this space â but at this point there’s still definitely something of a ‘Wild West’ feel to how such data is being collected and used.”
When it comes to big data, we’re all in the developing world. So what should we do to avoid being exploited? Or is it already too late?
In addition to Blumenstock, the authors of “Predicting Poverty and Wealth From Mobile Phone Metadata” include Gabriel Cadamuro and Robert On.