C-19: How transmission changes under shifts in mobility: country comparisons

Updated: May 13, 2020

In this analysis, we utilise mobility data from a range of countries, specifically the onset of reduced mobilities from movement controls, and compare this to death and case detection statistics in order to quantify their effect on infection spread. Correlating the change in mobility type, with the case and death statistics enable us to gauge the efficacy of different control measures in terms of their possible ability to reduce spread of the virus. From there we postulate potential exit strategy options that could be possible for different countries by maintaining controls on the type of mobility with strongest correlation to COVID-19 spread, while partially relenting on others.

TL;DR takeaways

  • Agreement between mobility reduction and the decline in transmission derived from reported death statistics provide an estimate of the median time between infection/contraction of COVID-19 and death, at 14 days. The same method applied to case detection statistics yields a range of COVID-19 case detection delays (time to detect case after infection) at a range of 4-22 days depending on country. The wide range further confirms our suspicion that some countries are falling well behind in testing; particularly where deaths and therefore infections are high.

  • We have quantified the correlative link between transmission factor and mobility type (retail, transit, walking etc) for a range of countries, which could be crucial inputs for decision makers to help develop viable exit strategies out of lock-downs. While it is not possible to fully disassociate the effects of any one mobility type from another, we use the strength of correlation to determine which has the strongest effect in a particular country. Based on our observations, it may be possible to maintain mobility within a range of 40-60% from baseline levels yet achieve transmission factor <1.0. This hopefully is a promising insight that could help policy makers to maintain acceptable levels of economic activity despite movement controls being in place

  • Fundamentally, we believe that aggressive and broad brushed movement controls are not appropriate to be deployed in situations where numbers of the infected and deaths are low. They do not adequately discriminate towards containing the infected, while creating destructive consequences on the economy. Arguably more targeted measures such as detection, tracking and tracing, as well as good personal hygiene and interaction practices are more appropriate in such situations. At the same time, there should be adequate readiness in the country to deploy a lock-down when is indeed necessary. Hopefully this may provide some useful basis to develop a workable exit strategy from lock-downs that balances the objective of saving lives, while also preserving livelihoods

Results & discussion from analysis - see figures in section (A) to (C) below

  • Definitive estimates for the median time taken from contracting COVID-19 to death (median time for infect-death), have been difficult to estimate due to lack of certainty as to when a person actually contracts the virus. This is largely due to asymptomatic and pre-symptomatic transmission - in other words, a person is uncertain when exactly they contract the virus until they start showing symptoms. In this analysis, we utilise mobility data from a range of countries, specifically the onset of reduced mobility from movement controls that should translate to a distinct reduction in infection spread. We then compare this with transmission factor calculations from actual death and case detection statistics to estimate time for infect-death to a reasonably high confidence level. See the figure below for one such as example of this tabulation

  • All countries analysed show a consistent 14 day lag between the onset of decline in mobility (from movement controls and lock-downs), with the computed time-based transmission factor based on death statistics. We refer to this as the median time between contracting COVID-19 and death - i.e. the true time taken between contracting the virus and eventual death; estimated at 14 days, and thus replaces our previous estimate on this figure of 10 days. At 14 days, the time from contracting COVID-19 to death is still relatively short; and to add, as death statistics are possibly delayed in their reporting, this time could be shorter still by 1-2 days, attributable to the time between actual death and formal reporting of it

  • The mobility with highest correlation to transmission seems to be unique from country to country. We observed that in the cases of Austria, Germany, Switzerland and Spain, there is notable correlation between reduction in virus transmission with changes in driving mobility; suggesting inter- and intra-city travel by car could be a significant vector for transmission. In other countries studied, transit and retail mobilities show good correlation instead. However based on this data alone it is impossible to fully and objectively disaggregate the contributions from changes across a range of different mobility types

  • We apply the same method to case detection statistics, using mobility data to indicate the onset of controls; and yield an estimate of case detection delay ranging between 4 to 22 days - see table below. It is very possible that delays are due to the high volumes of cases or infected persons present in the population, making it a lengthy and tedious process for authorities to process the volume of tests necessary to detect every infected person. Summarised in the table below are detected cases for each country up to present day; note a generally longer test delay when a higher the number of cases are detected

  • Certain countries show a strong correlative fit between changes in mobility and reduction in transmission factor; i.e. R^2 fits in excess of 0.7 when utilising transmission factors derived from death statistics - for example, Switzerland, Turkey, UK, Germany, US and Thailand. For these countries it is clear how reduced mobility from movement controls and lock-downs has contributed to a distinct reduction in transmission of the virus across their populations. Generally, we see that the regression fit improves with the greater number of deaths (R^2 above 0.7), which intuitively makes sense given the larger number of data points from which to draw the regression. We do not discount the possibility that some countries could indeed exhibit weak correlation between mobility and transmission, for example where movement controls have not featured heavily such as South Korea, or where infected figures are so low that minimal community spread is taking place

  • Most countries show noticeably better correlation towards certain types of mobility, for example; retail (France, Turkey); transit (US, UK); driving (Switzerland, Germany); and walking (Thailand, India). Likewise, by utilising case detection statistics instead of deaths we find higher correlation towards transit & retail (Austria), walking & driving (Malaysia), and transit & walking (Thailand). This insight could be useful for policy makers to determine which mobilities to moderate more than others

  • Meanwhile a number of countries show poorer correlation across different mobilities (~0.3-0.6 R^2 regression fit). These include countries such as France, India, Egypt and Belgium; while Indonesia, Sweden and Japan show very poor regression fit of <0.2.

  • This is fairly expected. Japan, in earlier waves of the virus spread, eschewed any use of movement controls and opted to focus on measures such as testing, contact tracing, and good personal hygiene. Likewise Sweden also refused any stringent lock-down measures, and instead has focused on advising its population to socially distance and limits on size of gathering in groups, but otherwise does not inhibit movement. As such, weak correlation between mobility and transmission is expected. In the case of India and Indonesia however, it is perhaps a worrying sign to see such populous countries struggle to use movement controls to more impactfully drive down transmission; however the lower efficacy is expected when considering the relatively low number of cases and deaths in these countries. It is more likely that community spread at localised levels is causing continued infections in absence of mobility at the macro level

  • While the correlative link is clear, beyond mobility, there are a multitude of other factors that also contribute to COVID-19 transmission. And those measures are not captured within these datasets - cases in point are Japan and Sweden where mobility has been maintained at relatively high levels, yet transmission of the virus is kept relatively at bay, with transmission factors bouncing between 0.95-1.05. For example, measures such as timely and sufficient testing, rigorous contact tracing; strong personal hygiene practices such as wearing masks and hand washing, as well as broad crowd control measures such as prohibiting public gatherings; all equally contribute to reducing transmission, perhaps no less so than movement restriction would

  • It is perhaps in the countries where correlation is in fact low that make for interesting case examples to learn from. Thailand and Turkey appear interesting to explore as they have reached sub-1.0 transmission at mobilities still maintained between 40-60%; while Austria and Belgium continue to reduce transmission in spite of mobility staying constant at low, but not lowest, levels

  • Tabulated below are the regression fit results between mobilities and transmission factor for a range of countries. In general, death statistics were used to compute the transmission factors unless stated otherwise;

  • Another important point to note. In scenarios where there are a large number of infections and deaths, mobility at the macro level plays a crucial role in allowing or inhibiting transmission. Meanwhile when infections are low, reducing mobility may reduce movement at the macro level, however community spread at the local level may still persist. Put another way, inhibiting mobility, which is already a very blunt tool, is made even blunter tool when deployed in situations of low infections and deaths. The majority of movement being inhibited is focused on healthy people, while local level mobility within communities and families is harder to control using such measures. The data is telling: compare these two countries - the correlative fit between mobility and transmission factor for the UK is .88 compared to .45 for Malaysia; and their respective COVID-19 statistics stood at 110 deaths and 2,700 cases versus to 2 deaths and 900 cases at point of onset of the mobility decline (18 March). It is clear that the efficacy of the movement control was higher for the UK. Meanwhile, in spite of the hard lock-downs, the reduction in transmission in Malaysia was slower than expected, taking some 3 weeks to reach its transmission floor compared to 3-4 days for mobility to reduce to its lowest figure

  • The issue here, again, as we have highlighted many times is a continued reliance on case detection statistics as a barometer for infection of COVID-19. Had the UK relied instead on death statistics and inferred the computation of infected persons, they may have expedited movement controls sooner and taken a more aggressive stance in its implementation

  • To further explore the results of the regression, we set up a normalised computation experiment using the factors A and B for each country, for the equation Transmission factor = A x Mobility + B. We prioritised datasets where R^2's are within range of ~0.6-0.9 (confirming a relatively good correlative fit)

  • Each country was normalised to begin with the number of infected persons (or deaths) at Day 0 with n=1; and each computed through 3 distinct phases; Phase 1 (10 days): mobility 100%, Phase 2 (10 days): mobility linearly decline to floor level, Phase 3 (10 days): mobility held at floor level. The results are shows as follows for the countries in question, for three different mobility floors 40%, 25% and 15%;

Mobility floor = 40% at day 21 (note log plot)

Mobility floor = 25% at day 21

Mobility floor = 15% at day 21

  • Observe the shape of the curves between day 10-20 and note how the height and location of the peak coincides with the point where the transmission factor breaches 1.0. It is fairly visible how an even faster taper through lock-downs would limit deaths and infections tremendously; for example what was done in Malaysia where mobility was reduced to below 25% within just 3-4 days. Belgium and Philippines stand out as being poorly performing based on this chart, however these two have poor R^2 fits at 0.4-0.6 and therefore should be discarded from study. Otherwise, almost all countries show the ability to reach transmission factor <1 even for mobilities as high as 40%; that is, even for countries such as UK, Austria, Turkey and US where deaths and recorded initial growth rates have been higher than Asian countries

  • So, how should a country devise its exit strategy from stringent control put in place during lock-down? Let's take the example of Malaysia. The country took decisive action and a rather aggressive stance with their movement controls, locking down within a short time of less than a week to reach sub-25% mobility levels (see Malaysia data in figure section), and has held those levels ever since

  • It is possible that Malaysia could relent on some of these controls, allowing mobilities to rise up from current sub 25% levels to around 40-50%; focusing on transit and retail mobilities to begin with, as these tend to be where large crowds congregate. Regional inter-city transmission would still need to be controlled, and as such some restriction against inter-city travel should remain. The country should gather learnings about the tactics employed in Thailand and Singapore who seem to have achieved greater control of transmission with less reduction in mobility. Lastly, it should maintain a high level of readiness at all times to lock-down the country again at a moments notice if necessary. Ideally however, one could argue that the country (given low levels of infections and deaths) is in a position to design a regime that switches such controls on and off at predefined junctures over next 12 months - we covered this scenario in a previous article here

Figures section

For figures below (section A, B through C): Mobility data was sourced from Google and Apple, while Transmission Factors were computed either from reported death or case detection statistics.

  • Transmission factor is computed via the formula Tr-factor (day N) = Death (day N)/Death (day N-1). Averaging/smoothing of 7-15 days is applied to the death statistics to eradicate spurious data points. Period windowing was also applied to remove erratic data due to lower death/case detection figures

  • Mobility data may not all start at 100(%) as different baselines are used across Google and Apple sources

  • Agreement is sought between 'shape' of transmission factor curve with mobility (note different axes are used for each term)

(A) Figures: matching death statistics with changes in mobility during mobility reduction phase (14-day shift in transmission data reflecting mean time for infection-death)

Switzerland: driving mobility has highest correlation

France: retail mobility has highest correlation

Thailand: walking mobility has highest correlation

Turkey: transit mobility has highest correlation

US: transit mobility has highest correlation

UK: transit mobility has highest correlation

Germany: driving mobility has highest correlation

India: walking mobility has highest correlation

Japan: poor fit, suggesting mobility is not driving factor

Indonesia: poor fit, narrow band of transmission change + other non-mobility factors at play

Egypt: poor fit, mildly higher correlation to retail and transit

Belgium: poor fit, mildly higher correlation to transit

Sweden: very poor fit across the board (top- death stats, bottom- infection)

Thailand: higher correlation to transit & walking (infection data)

Malaysia: higher correlation to walking & driving (infection data)

Austria: good correlation to all mobility types (infection data)

Philippines: reasonably good correlation to all mobility types (infection data)

(B) Figures: additional reference data; matching death statistics to changes in mobility

(C) Figures: additional reference data; matching case detection statistics to changes in mobility


Recent Posts

See All