Facebook movement data could help find new Covid-19 locations, study finds

Anonymised Facebook data on people’s travels could be used to identify the spread of Covid-19 in locations where health officials are not yet aware of it, a new Australian study has found.

Published in the Journal of the Royal Society Interface on Wednesday, University of Melbourne researchers analysed anonymised population mobility data provided by Facebook as part of its Data for Good program to determine whether it could be a useful predictor in determining the spread of Covid outbreaks based on where people were travelling.

The research analysed three outbreaks in Australia: the Cedar Meats outbreak in Melbourne’s west, the larger second wave in Victoria, and the Crossroads Hotel outbreak in New South Wales.

Pete Evans’ Facebook page removed for breaching misinformation policy with Covid posts

Read more

The research found locations where people had predictable and periodic movement – such as travelling to and from work – provided more useful indicators of virus spread than social settings. In the case studies, the data was therefore more useful in predicting virus spread in the Cedar Meats outbreak than the Crossroads Hotel outbreak.

When it came to analysing Victoria’s second wave, which started with the confined suburb lockdowns in late June and early July, the analysis found mobility data could have alerted the government the spread had already moved beyond the suburbs initially confined to lockdown.

“Our examination of the second wave of community transmission in Victoria showed that several weeks before it was recognised, the spatial distribution of a small number of active cases was indicative of the outbreak distribution more than 30 days later when interventions were introduced,” the researchers said in the paper.

“This observation indicates that even when case numbers were small, low-level community transmission may have already been taking place throughout the region of metropolitan Melbourne. This suggests that earlier selective lockdown measures, extending beyond the borders of regions in which cases had been identified, may have been more effective at containing transmission.”

Lead researcher from the University of Melbourne, Cameron Zachreson, told Guardian Australia it was too hard to say whether the data could have changed the Victorian government’s decision-making around when to lock down Melbourne during the second wave.

“We analysed the mobility data and we lined that up against postcodes that had been locked down and it was very clear from looking at that, that locking down these particular postcodes probably wasn’t going to have the desired effect but at the same time that postcode lockdown didn’t last very long,” he said.

“That was expanded to the greater Melbourne region in a couple of days. I think [it] became very clear that that approach wasn’t going to be sufficient.”

Facebook to remove false claims about Covid vaccines

Read more

Zachreson said the data would be useful in cases where not a lot is known about an outbreak.

“Looking at mobility information like this can can give you some kind of a decent idea. There’s definitely a signal there. It’s not going to give you everything you want.

“There’ll be places that happened to be high risk that aren’t in the data, like in the Crossroads outbreak, there were cases in the Blue Mountains. That was because those people travelled a long way on a trip that they wouldn’t normally take. And that’s the kind of thing that the aggregate mobility patterns we have don’t capture well.”

Zachreson said the data could also be used by governments to determine where to declare potential hotspots, rather than focusing on arbitrary local government areas, or whole cities, like Sydney currently.

“It can give you a less arbitrary determination for where the high risk zones are,” he said.

Zachreson stressed the data would not allow the researchers to identify someone, since Facebook would have already anonymised it. Governments also couldn’t access the data from the researchers unprocessed, meaning they would not be able to identify people through the dataset either.