It was almost a year ago that I started playing around with Google’s mobility data. On August bank holiday last year, I travelled to Devon in southern England for some fresh air after the galloping workathon of the early period of the OxCGRT project. I had not stopped in months.
OxCGRT—the Oxford COVID-19 Government Response Tracker—was born in a brainstorming session during a Friday-morning seminar in Oxford University’s Blavatnik School of Government, where Tom Hale (OxCGRT’s PI) and I teach the core graduate politics course. Thanks to the ongoing efforts of an army of volunteers, this project records and makes freely available in real time data about the relative strength of government policies against Covid-19. It has swelled to gigantic proportions, with now more than 850 volunteers having passed through and a stable-state core team of 22 people who take care of day-to-day operations. The data cover almost every country, including many subnational jurisdictions, and have been used by countless researchers and policymakers around the world.
OxCGRT’s initial ramp up from that last Friday of Spring term 2020 was aggressive, to say the least. Within the first fortnight, Sam, a lovely former theoretical physicist-turned-financial-risk-analyst on gardening leave, quickly created an online database. Public-spirited Blavatnik School students, with barely the weekend after term to rest, jumped on board as our first volunteers. I quickly went from thinking of OxCGRT as a fun little exercise to offer focus for students who were struggling with uncertainty, to realising that we could not put it down. Our graphs were suddenly in briefings to the UN Secretary General. The Financial Times needed our API humming smoothly because our data was automatically spilling into its live charts. Soon, I was waking up at 5am to go into battle with my inbox. Meanwhile, there was an uneasy sense, growing exponentially more palpable by the day, that the public health emergency that had already unfolded in a few countries was going global. We kept at this pace through the Easter break and summer term.
Hence, by August bank holiday, fatigue loomed large in my head. But as much as I tried to put my laptop down, I was fascinated by the graphs popping up on the screen. Before leaving Oxford, I’d merged OxCGRT’s policy data with Google’s aggregate measures of how much populations had been moving around since the start of the year—how much, for example, people had been visiting places like retail parks, and for how long each day they had stayed in residences. Virtually every country in the world had imposed previously unthinkable closure and containment rules on its citizens in the proceeding months, and I wanted to know for how long people were sustaining the behaviours that their governments had demanded.
After running a few regressions, I could quite clearly see that in country after country, following an initial, dramatic reduction in mobility when lockdowns were first introduced, mobility had gradually started to sneak up. Aggregate population movements had been softly rising while stringency levels were, very often, unflinchingly high. Curiously, though, mobility didn’t keep on rising. It plateaued. It was as if societies, after a brief period of stunned obedience in response to the novelty and drama of early lockdown, had found a threshold level of physical distancing behaviour that they could actually stick to. Importantly, they appeared to have settled on a degree of movement that was much more limited than their pre-Covid averages—the data certainly did not suggest that closure policies were becoming ineffective—but there had been an undeniable shift in adherence since lockdown began.
I emailed some early graphs to the UK Cabinet Office, where Tom and I are members of the International Best Practice Advisory Group. Then I called Rafa in Brazil. Rafa is one of my best friends and an assistant professor of social science research methods at FGV-EBAPE in Rio de Janeiro. To do this properly, his joint leadership would be essential. Adherence to protective behaviour policies was a live political issue, so we needed to leave no statistical stone unturned. Indeed, concerns about whether citizens would or could continue to stick to the rules over time had been cited by some governments that had been slow to respond, and that had seen especially large numbers of deaths. This included the UK. Yet, the main epidemiological models all assumed that observance of protective behaviours wouldn’t budge over time. “We need Eduardo,” came Rafa’s immediate response. Eduardo, a brilliant behavioural scientist and generally sensible soul, would help us to navigate the fraught exchanges among scholars around the popular concept of pandemic fatigue—the idea that (depending on your definition) people get bored of, sick of, or in some ways exhausted from, practising protective behaviours over time.
Figure: Change in policy strength alongside percentage change in time spent in places of residence (relative to the average amount of time clocked each day over the first 5 weeks of 2020).
For 124 countries, the chart shows how OxCGRT's Stringency Index varied alongside Google's residential mobility measure from the first day that each country introduced a closure requirement. This stay-at-home behavioural measure closely tracked the strength of closure and containment policies when lockdown began, only to loosen a little faster than policies before the two lines continue with a roughly consistent gap. Credit: Rodrigo Furst.
The team ran analysis after analysis in the months leading up to Christmas. We added survey data to assess adherence to mask use, a protective behaviour that doesn’t have any direct correspondence to mobile-phone metrics of movements in space. And we compared differences in physical-distancing observance as measured by both surveys and phones. The advantages and limitations these two kinds of behavioural data are neatly symmetrical. Surveys are self-reported, with all of the implied potential biases baked in. Whereas the opinionless gadgets that generate mobility data make it helpfully objective. And while mobility data might only give you grand patterns exhibited by smart phone-owning populations, surveys provide details about every individual respondent—their age, gender, job, and plenty about their ideas, too. Adding survey data to our analyses thus also allowed us to compare the changing observance of different groups of people.
Rafa and I—and Rodrigo, our fast-coding research assistant—worked through the Christmas break. (With new waves of disease rising in both of our countries, it wasn’t as though we were missing much anyway.) Parts of the analysis were proving to be heavy weather. Modelling decisions had to be made and there were endless checks to work through. Our Supplementary Information section hit 500 pages. By the time we submitted the paper, in January, the asymptotic threshold patterns that had sparked my interest on a warm bank holiday weekend had turned into banana-shaped curves: after initially ebbing and then levelling off at a threshold, adherence to physical-distancing behaviours appeared to be rebounding slightly. (In the figure this is equivalent to the modest narrowing of the gap between policy strength and behaviour seen in the final couple of months.) We included controls in our models for the most likely explanations. We added weather data (thinking that the northern hemisphere winter might keep people at home), and records of disease progression and news articles about COVID-19 (thinking that people might be responding to self-assessed risk). But the “bananas” were stubbornly still there. They appeared in high and low-income countries—although not in every world region—and were evident among the employed and unemployed, the young and the old, people with and without chronic diseases, among men and among women.
One of the chief challenges in our analysis was the close correlation between the passage of time since closure rules were first imposed, and the changing strength of policies. When you conduct statistics of any kind, near-simultaneous change among variables makes it hard to pinpoint which of them—in our case, time or policy strength—is the better predictor of the variation you are trying to explain. Rafa tackled this puzzle by being as generous as possible to the opposing argument to our own, i.e. that flagging behavioural observance was entirely due to the loosening of policies, and not at all due to the ongoing trudge of living through a pandemic. He executed this, in some of our models, by assigning to the changing policies 100% of the behavioural variation that could have been due to either the passage of time or to governments altering policy strength. Then he took the remaining, unexplained variation, and found that time predicted even these “leftovers”.
Photo: An accidental WhatsApp screen-grab. Anna and Rafa on one of hundreds of video calls about the paper.
Our findings offer guidance for policymakers working on pandemic management, especially where many people are immunologically vulnerable. When we homed in on physical-distancing behaviours, we found that countries that went into the pandemic with above-median levels of social trust—that is, trust between strangers—appeared better able to sustain the adherence levels that they had enjoyed in the first month of lockdown. Initially, these countries had started with slightly lower observance (something another academic put down to the extra incongruousness of physical distancing in these societies), but they were then better able to hang on. This seems logical. After all, what’s the point of paying the high cost of staying at home, of doing your part to push Rt below 1, if you don’t trust others to do the same? Building and reinforcing social trust appears to earn countries a compliance dividend.
Crucially, we also found no thresholds (or bananas) in our analyses of adherence to mask use. Instead, people seem to have unwaveringly increased their observance of mask wearing throughout 2020, over and above strengthening policies in this area. In the paper we suggest this may be because mask use is something you get habituated to, rather like wearing a seat belt or a helmet. Following a sustained period of frequently getting into a car and popping on a seat belt, you feel a bit naked if you drive around without one.
In this way, mask use appears to have a different, perhaps semi-conscious, logic to physical distancing. Staying at home and avoiding meeting others are not only more psychologically (as well as more financially) costly to you than mask wearing, but these behaviours may also get harder the longer that you keep them up. We cannot rigorously prove this using the data that we have, but up-close studies of individuals experiencing quarantine suggest it may be true. Certainly, the fact that mask-wearing observance has gone up and up at the same time that physical-distancing observance has taken a different path, adds weight to the fatigue argument. If it were, instead, disease risk that people were primarily responding to, then physical distancing adherence would have moved in sync with mask-wearing tendencies: when things looked bad, people would have doubled down on everything protective.
None of these insights would have been possible without the efforts of hundreds of OxCGRT volunteers, coding around the world and around the clock. I would also like to wholeheartedly thank my other coauthors, Tom (a wonderful PI), Annalena and Andrew, for their hard work and respective contributions, as well as our reviewers and editor, Jamie, whose patient feedback substantially raised the calibre of the paper.