The data provided by Google and Apple for the purposes of tracking changes in mobility during this COVID-19 crisis have been useful to gauge, at a high level, how our physical movement habits have changed due to implemented movement controls and lock-downs. We dive deeper into these datasets, to more deeply understand how mobility habits have changed, and extract more information that could be useful to gauge the efficacy of COVID-19 related interventions.
From analysing mobility data from Google's COVID-19 Mobility reports; across residential, retail, workplace, transit station, parks and grocery & pharmacy mobilities, we see more granular differences in stringency, speed and timeliness of controls implemented across different countries
We can surmise that control of a particular type of mobility may not matter as much as the timeliness and effectiveness of its execution; after all every type of mobility can be a sizeable vector for spread given a large enough number of people. Nonetheless, there could be two broad workable strategies for their use by authorities; (1) a slow and steady approach of reducing mobility, in advance and well ahead of potential outbreaks, or (2) a rapid lock-down that has the immediate effect of stopping mobility in its tracks. Regardless, for countries where outbreaks have hit critical mass, total lock-downs are generally the only option left to avert a higher number of deaths
While more work is needed to further refine perspectives on the efficacy of different types of controls, it is possible that retail and transit mobilities would need to be controlled more stringently as they naturally converge large crowds; whereas controls on workplace, parks, grocery/pharmacy mobility could be lightened as a potential first step to progressively lift stringent lock-downs
Google, Apple constantly knows where you are, and they can possibly guess what you might be doing
Most Android users are by default wholly invested into the Google eco-system (Google owns operating system Android), so one way or another they are likely to attempt to extract as much information about the you, the user, as possible, and that includes location data. The term used by Google is location history (LH); a collection of data pertaining to a persons physical movement history which is further enriched by several attributes such as time stamps, mode of transport, name and type of location etc. See figure below for an example excerpt.
While strictly speaking, location history (LH) is switched off by default, opt-in mechanisms via web browser preferences, or app usage such as Google Maps and even weather updates often time trigger LH to be switched on permanently, until explicitly opted out from. More often than not, we have found that is indeed the case for most users. We will not go into what Google does or intends to do with this data, that is a topic for another time. Suffice to say, they collect an awful lot on each and every one of us.
Apple takes a similar approach, however from our knowledge only logs data from usage of Apple Maps, and not across general usage of the device or other apps across their suite.
Figure: Example excerpt from Google's location history tracking which is switched on by default in most instances for Android users
COVID-19 mobility data: indeed helpful to indicate impact of control measures
As we had previously discussed here, both Google and Apple have volunteered abstracts of their (anonymised) mobility data to the public for purposes of monitoring efficacy of movement control restrictions put in place during the COVID-19 crisis. This is available for limited time, although the period has yet to be defined by either party. For Google, the source of this data is an anonymised form of each Android phone user's location history.
The general format of the data provided by Google is simple enough. It denotes percentage changes over time, from baseline data 14 February 2019 across six activity areas (e.g. retail, grocery and pharmacy, parks, transit stations, workplace and residential). This is available for almost every country in the world, and further broken down for larger countries into cities or regions. The data is updated once a week and trails behind T-5 on day of reporting - the data used here in this article was published 16 April, for reporting up to 11 April 2020. A sample of the raw data is provided in the figure below for the case of Italy.
At a glance, the data is very useful to study. Firstly, it indicates that movement controls were initiated around 8th of March (denoted by the steep decline in the mobility curves), and since then time spent in Residential activities have increased some 30-35% from baseline, while other activities such as Retail, Parks, Transit station, Workplaces and Grocery have declined roughly 80%, 80%, 80%, 70% and 40% respectively. Strictly speaking these figures do not represent changes in 'time spent', as according to Google it is a composite score of frequency, duration, journeys etc, and further augmented to remove low confidence level data, it is a simplifying distinction that can be made with reasonable confidence in interpreting it.
Note/disclaimer for the reader: Google warns that the activity categorisation embedded in these reports is not necessarily definitive, and the definition of location attributes that provide indication of mobility type may vary from one country to the other
Figure: Mobility changes from baseline data for Italy
Going deeper: cross correlation analysis gives us a peak into how activities are related to one another
As there is a reasonable amount of temporal data provided in the datasets, it allows us to view each activity change day on day as time-varying signals that can be cross-correlated to one another to identify how they are possibly related (or rather cross-correlated). Cross correlation is calculated utilising day-on-day changes in the value of mobility expressed as % change from the day before; here we use a 7 trailing day window for computation of the metric.
Figure: Cross-correlation for Italy for before and after the control period
Figure: Rolling cross-correlation between different activities for Italy (window 7 days trailing)
These figures are an example of such an analysis conducted for Italy. Notes;
Before movement controls, a large portion of the population utilised transit hubs (-88% corr. to residential), and spent roughly 5 days at work (-71% corr.) Grocery and pharmacy mobility was not a high utilisation of time (only -4% corr to residential)
One third of the population (-30% corr to residential) would engage in retail, and close to -50% would visit parks, which presumably includes walking about their own neighbourhoods
Public transit is a primary mode of transportation regardless of destination for Italians (+60% corr workplace, +70% parks, +50% retail)
After movement controls, retail, grocery, parks and transit are all positively correlated to residential activity, by at least +70%, however note these are from much lower base figures given the decline from the control date. This is consistent with the notion that most are perhaps frequenting such destinations within close vicinity to their own homes given the movement controls enforced by authorities
What perhaps is slightly odd is the high negative correlation between residential and the workplace (-90% corr), particularly during weekends; which suggests people are leaving their homes to go to work during these days
Looking at the rolling correlation chart; it is clear to see how correlations for home/residential-transit, transit-workplace, home-retail, home-transit have entirely flipped through the control period. What is perhaps rather less defined and rather erratic before controls started is the correlation between mobility to parks and transit stations, as well as grocery/pharmacies - when was the last time you went to the park?
It should be noted that this is relatively short finite time series and therefore the accuracy of the computation may suffer as a result. As such the tabulated percentages are indicative and directional only
More country data to attempt to spot trends
Table below: summary of results from a few select countries.
Accompanying notes: Data is collated from 14 February to 11 April 2020; Changes from baseline are estimated to nearest 5% - detailed country tables for data up to 5 April are available in our previous article here. Cross-correlation results have been rounded to nearest 10% - we advise, particularly due to the short window period and lack of gradation granularity in the datasets from Google, that the reader take the figures as indicative and directional only
Refer to 'effect of movement controls, changes from baseline' in table above: It is quite understandable that stringent controls were implemented in Italy given how hard hit they were by the crisis. Mobility seems to have been a clamp down across the board from retail, to transit and workplaces; resulting in a marked increase of 30% time spent at home (residential). On the other hand, South Korea, without a formal lock-down policy in place, quite clearly placed little restriction to inhibit mobility; while there was some noted reduction in workplace and transit station mobilities, grocery and pharmacy, workplaces and residential barely changed
Likewise, the UK seems to have clamped down on mobility albeit not as strictly as Italy; as they too were hard hit with many deaths over the month or so leading up to early April. As we had touched on previously, some countries like Malaysia took a hard stance at locking down their populations in spite of low infection and deaths; where mobility seems to have been even more restrictive than Italy (declines of -50-85% across categories). For Indonesia, the sheer size of the country perhaps made it difficult for authorities to definitively restrict movement of its population; reductions are registered across the board are barely reaching 50% of Malaysia's. Thailand's population most likely took their own initiative to reduce physical interaction early given earliest detection of cases in January. They were perhaps already more vigilant by the time the actual locking down happened; transit and park mobility was down 10% even as early as late January (not tabulated here but figure on Thailand mobility changes are included further below)
Refer to 'before movement controls, cross-correlation': Expectedly, there is a negative correlation across the board between residential and activities such as retail, transit and workplace mobilities, ranging between 20% to 90%. This percentage expresses in simple terms, the amount of time spent on another activity versus the first.
Note that in Malaysia, due to early awareness of the pandemic, many employers had already started a combination of work from home and modular team policies within the month of February, thus possibly explaining the positive correlation between residential and workplace activity. Thai's likely took their own initiative to avoid public places even as early as February, given early detected cases there and hence show positive correlation between residential and retail mobilities
Public transport hubs seem to play a central role particularly in Italy, the UK and Indonesia (with very high correlation of -(70-90%) between residential and workplace. In addition, for the three countries, transit station mobility is high for travel to retail and workplaces, given 80% correlation to retail for both Indonesia and the UK, 50% for Italy. Transit-workplace correlation is also notably high for Indonesia, South Korea and Italy at 60%, 70% and 60% respectively. Meanwhile, residential to transit mobility is less correlated for Malaysia at -40% and Thailand -60%, likely indicating a distinctly lower usage of public transportation
Refer to 'after movement controls, cross-correlation': The reduced mobility across the board in retail, transit stations, and workplaces results in a positive correlation of some of these activities to residential mobility (the exception being South Korea). It is however notable that for the UK, Thailand and Indonesia transit activity is still significant with residential-transit station correlation of -40%, -40% and -80% respectively. Furthermore, residential-workplace correlation for these three countries are between -(80-90%) suggesting that many are still taking public transport to go to work. In Indonesia specifically, retail is still taking people out of their homes (-30% correlation); coupled with the fact that residential mobility is only up 15% from baseline, it appears their population is still actively shopping
Tying it all together: adding COVID-19 infection growth statistics, and a brief discussion
We refer in this section to the previous analysis we conducted on investigating COVID-19 growth for a range of countries (from a previous article), and where appropriate have made adjustments to reflect the latest forecasts. Do refer to that article for a background discussion on how we are seeing markedly different growth rates in infection and deaths across countries
For Phase 1, i.e. the initial growth phase, Italy and UK were categorised as high initial growth countries (2.3 and 2.8 days to double), Malaysia, Thailand and Indonesia low initial growth (5-6 days to double), whereas South Korea very low initial growth (14 days to double). For purposes of this discussion we will omit South Korea from the consideration set due to the unique and rather advanced way it has handled the virus spread by not solely leveraging on mobility controls
For Phase 2, the tapering growth phase through movement controls - see table below for updated reference data. Based on latest projections, more specifically the speed to reduce the transmission factor to 1 within the table, we would categorise countries in this list as follows: hard taper - Malaysia, South Korea, and Thailand; moderate - Spain; weak - UK, Italy, Indonesia, Belgium. It may come as a surprise that Italy and the UK are placed in the same stringency category despite our earlier deduction on their different level of mobility controls (where Italy appears more stringent). This comes down to the speed at which the those controls and their stringency are implemented. As Italy was one of the first countries where COVID-19 broke out outside of China, it was relatively late to put in place strong controls quickly, as there was literally no playbook to handle the situation. Measures were rolled out, but in a phased manner over a number of weeks and that was crucial time lost to overcome the virus
Meanwhile, Malaysia locked-down decisively and swiftly, practically reaching steady state levels of mobility within 3-4 days; Italy in comparison took some 3 weeks. This is perhaps the first important insight - speed at rolling out controls at scale is important, not just how stringent those measures are. The timeliness is also a factor, as we had described previously - i.e. when was the mobility threshold reached versus the number of infected persons there were in the population
But which mobilities are most important to restrict first? The analysis here cannot possibly address this question definitively, but we can attempt to generate a number of hypotheses and substantiate to a cursory level. We should warn the reader however; while we will attempt to glean generalised takeaways from observations made across these countries, it should be remembered that each country is complex and unique in its circumstance, so such generalisation while practical may not be the most sensible approach
We will explore these hypotheses in the following sequence; retail, workplace and transit mobilities
It is consistent across all countries that retail mobility is generally made rather restrictive under movement controls and lock-downs. Thailand in fact started very early in this regard (see figures below), with retail and park mobilities dipping as much as 10% two weeks ahead of official lock-down date; the fact that the country eventually achieved the same level of control effectiveness as Malaysia is perhaps rather telling. Indonesia on the other hand still maintains relatively higher mobility between residential and retail in spite of controls, which possibly contributes to the reason why their taper is weaker. Unfortunately what the data does not tell us is how much the controls should discriminate between large and small retailers; only Google perhaps knows the answer
For workplace mobility, we think it may be appropriate to analyse this in tandem with transit mobility. Again Thailand makes for an interesting first case study. Even after controls, workplace mobility remained relatively high (-25% from baseline vs Malaysia and UK at -70%). And yet; with a higher number of people still at the workplace, it still registers a reasonably high correlation with transit mobility at a correlation of 30% (i.e. some 30% of people still use transit hubs to get to work). Meanwhile, looking at the UK and Italy, notwithstanding the fact that both countries locked-down too late, it is noted that both countries have proportion of transit users (high correlation residential to transit stations at -70% and -90%); likewise for Indonesia at -60%. It is quite possible that transit stations are a potentially significant vector for transmission of COVID-19 and perhaps workplaces themselves are not as big an issue as the public transportation the populations take to get to work
So to summarise again, timing of controls perhaps plays as significant a role as stringency - controlling early and gradually does work (Thailand) vs. reacting too late (UK, Italy); so too does hard stringency (Malaysia).
Lastly, a brief word on residential and park mobilities. It is perhaps not intuitive to think about residential mobility as a vector for virus spread. However, given recent events in Singapore where the virus has spread rapidly within foreign worker dorms, and in Malaysia, two large clusters were found among participants in a religious event and separately, a local wedding; it may be important to control potential community spread through such social events, which may not be classified within other mobility types. On the other hand, parks if not accessed via public transportation and as long as they do not congregate large crowds could be permissible - South Korea registered an increase in park mobility of 25% after controls, whereas Thailand park mobility reduction was only 40% (vs. Malaysia and Italy at 80-85%). We should probably warn the reader that it is well beyond the ability of the data and analysis laid out here to form strong opinions on any of these; we did think it helpful however to mention these to round up our discussion on this topic
Table: country growth statistics (based on projections on 22 April)
Table: reference case detections at 50% mobility date (reference from another article) - note: Apple COVID-19 Mobility data was used here
Figure: Malaysia changes in mobility (14 Feb - 11 April)
Figure: Thailand changes in mobility (14 Feb - 11 April)
Figure: UK changes in mobility (14 Feb - 11 April)
Figure: Italy changes in mobility (14 Feb - 11 April)