PATHOGEN DYNAMICS GROUP COVID BLOGS
BLOG 15: Daily lateral flow testing of COVID-19 contacts can be as effective as quarantine and carries lower social/economic costs
8 August 2021
In our group’s latest paper, we investigated the costs and benefits of two strategies for reducing transmission from individuals identified by contact tracing after close contact with someone with COVID-19. The first strategy is requiring all such individuals to isolate at home for 10 days; the second is requiring the individuals to self-test every day for 7 days, and only isolate if they test positive. Our main finding was that the two strategies are expected to reduce transmission by a similar amount, but much less isolation is required with daily testing. Read on, or read the paper itself, for more details. The paper – Ferretti et al. 2021, available here – is an unpublished preprint, and has not yet undergone peer review.
First, some background. Preventing the spread of COVID-19 and other infectious diseases is often a question of identifying who is at risk of passing the disease on to other people. Being specific about risk allows public health resources to be concentrated where they are most effective in stopping transmission, and minimise disruption. For COVID-19, one of the most effective interventions is isolation, i.e. staying at home*. However, isolation is certainly disruptive, as is now all too familiar from national lockdowns. It would be better by far if we could only isolate those individuals currently at risk of infecting others, and not the whole country. This is the purpose of contact tracing. Contact tracing starts from individuals known to currently have the disease – ‘cases’ – and identifies individuals they had close contact with, focusing on the times when the case was most likely to be infectious. These identified individuals – ‘contacts’ – are at risk of having the disease, and their isolation can prevent them from spreading it further.
An alternative to requiring contacts to isolate is to require them to take tests frequently, and to only isolate them if and when they test positive: ‘daily contact testing’, or DCT. By design, DCT means contacts who are not infected do not need to isolate, which is appealing. However, the success of DCT in reducing transmission depends on several factors: what tests are available in sufficient quantities, giving sufficiently quick results? How accurate are those tests? Is the test accuracy related to how infectious the person is, and if so, to what extent? How careful will individuals be about contacting others, before and after getting a positive result? Currently in the UK, lateral flow antigen tests (LFA tests, also called ‘rapid’ tests or LFDs) are widely available, can be self-administered, and give results within 30 minutes. PCR tests have lower availability, must be sent to laboratories for processing, and return results more slowly, hence they are not a feasible option here. So DCT, were it to be implemented in the UK in 2021, would be based on LFA tests.
The other questions above do not have such easy answers. How well LFA tests are able to detect an infection (the test ‘sensitivity’) depends on how much virus there is in the person’s nose and throat (the ‘viral load’). The risk of an infected person infecting someone else also depends on their viral load. The viral load varies over the course of an infection – starting low, increasing until a peak, then declining – and varies between individuals. Vaccination affects how likely it is someone gets infected in the first place, and affects their viral load if they do get infected. People developing symptoms are expected to change their behaviour, and symptoms can happen at different stages of the infection, or sometimes not at all. To understand how all of these factors interact, we developed a mathematical model† of how an infection varies over time and between individuals. Our model is informed by measurements of viral load, symptoms, infectiousness, and test accuracy from multiple sources.
Strategies should not be evaluated in a vacuum, of course: the real question is less about how successful DCT is expected to be, and more about its success relative to the alternative. And success is measured not only by reduction in transmission, but by what is required to achieve that reduction. We compared DCT with a ‘quarantine’ strategy, i.e. current guidelines in the UK. In this strategy all traced contacts are required to isolate, even if they independently seek a test and obtain a negative test result. (UK guidelines are scheduled to change on August 16th 2021, not from quarantine to DCT, but to exclude fully vaccinated individuals from the requirement to isolate if contact traced.)
Our findings are quite intuitive:
if most traced contacts would adhere very strictly to staying at home, but few would bother taking their daily tests properly or isolating if they got a positive result, then the quarantine strategy would do a better job reducing transmission;
if most traced contacts would diligently take their daily tests and strictly stay at home on a positive result, but few would bother staying at home under the quarantine strategy, then the DCT strategy would do a better job reducing transmission;
in the more balanced scenario that people would show an intermediate level of compliance with both strategies, then both strategies would be similarly effective at reducing transmission;
DCT always results in contacts spending much less time in isolation than the quarantine strategy. This is even more true for vaccinated individuals, who are much less likely to be infected and test positive.
Though the principles might seem obvious, quantifying the precise numbers involved is not. In our recent preprint we provide estimates. Below we reproduce one of the figures from the paper showing how much the two strategies are expected to reduce transmission, and how many days an average contact spends in isolation to achieve this. It also shows how much this benefit and its cost depend on how strictly contacts adhere to isolation in the quarantine strategy, and what fraction of them would agree to test daily in the DCT strategy. The top two graphs show that DCT, under our set of assumptions, would reduce transmission as well as the quarantine strategy with between 50 and 70% adherence. The bottom two graphs show that DCT requires traced contacts to isolate for much less time on average.
In all four graphs above, the x axis shows the amount of time since a contact was exposed to a case. We assume it takes three days from the exposure until the contact is reached by contact tracing. In the top two graphs, the y axis shows how many transmissions were averted thanks to one of the strategies considered, for an average contact (who may or may not be infected). In the bottom two graphs, the y axis is how many days of isolation were needed for an average contact. Blue lines show the DCT strategy, black lines show the quarantine strategy, and different styles of black line show the effect of varying adherence to quarantine. The two graphs on the left are specific to unvaccinated contacts, the two on the right to fully vaccinated contacts (whom we assume are four times less likely to be infected).
Another important reason to consider the DCT strategy is that it can start a second round of contact tracing from those contacts who test positive, and reach the people these contacts have infected. Normally this would only happen if and when infected contacts develop symptoms, decide to request a PCR test, and then obtain the result. Daily testing means the positive result can be obtained more reliably and much more quickly, and so contact tracing can prevent more transmissions from occurring. (Speeding up contact tracing is particularly important for keeping up with the fast spread of COVID-19, and was a key motivation for our proposal in March 2020 to introduce app-based contact tracing.)
Our interpretation of these results is that an intermediate level of engagement with either either strategy is most likely, and so switching from a quarantine strategy to a DCT strategy is a good idea in the UK and similar settings (where vaccination levels are high, LFA test are widely available, and the aim is mitigation of the epidemic rather than elimination). We expect DCT would be roughly as effective at stopping transmission, but at far lower social and economic cost.
Some further thoughts on public perception.
Contact tracing has been used successfully for a very long time for many different infectious diseases. At the start of the pandemic it was a new idea for most people, fortunately, because not for 100 years have we had a virus spread so easily and with such damaging health effects. Contact tracing is unfamiliar to most people, critically important, and taking place at scale across the nation and globally. This means we need to consider carefully how people react to it.
The psychology and behavioural aspects of contact tracing are tricky issues. Thinking in general terms, we can all recognise that a small fraction of the population isolating is much better than everyone isolating. In practise, when it happens to us, it can feel unfair that we are asked to stay at home when our friends are out enjoying themselves, or especially when staying at home brings financial strain. A sense of unfairness can easily lead to a sense of it being unnecessary: I feel fine, I wasn’t close to that infected person for long, contact tracing most likely got it wrong, I really want to go out, vaccines mean things aren’t as bad as they were anyway, etc. Reduced adherence to isolation can result in contact tracing not achieving enough: by trying to avoid isolating people who aren’t at risk of spreading the disease, we can end up failing to isolate those who are. This brings us back to the virus rapidly spreading everywhere, either requiring more drastic control measures, or resulting in much illness and many lives lost, and a health system struggling to cope.
There are several avenues for addressing this problem that are out of the scope of infectious disease science. Good communication with those who are contact traced could increase their commitment to isolate properly, as could a high level of trust in the institutions perceived to be involved in the process: health systems, and perhaps the government. Providing financial aid to those struggling to stay at home for work reasons could increase their ability to isolate properly. There is also scope for infectious disease science to contribute, however, which we explored in our latest study. DCT, in theory at least, targets the potential problem of poor adherence to quarantine most directly by making contact tracing even more specific in identifying who needs to isolate, and giving them evidence for their need to isolate.
Our paper is one piece of evidence for public health authorities, who can balance this with other pieces of evidence and further considerations, to maintain or update the overall strategy that works best. It is not a recommendation directly to individuals who have been contact traced about what they should do next: we encourage such individuals to always follow the guidance provided by their public health authority.
Blog post by Chris Wymant
* Epidemiologists say “self-isolation” for when people known to have the disease are required to stay at home, and “quarantine” for when people at risk of having the disease are required to stay at home. The distinction in terminology has often been lost in more general discussion, and it often isn’t important anyway. That said, it is of course important whether a particular person actually has the disease or not.
† For those unfamiliar with modelling, a model is just a simplified picture of complex reality. Creating and using models is essentially the only way we understand anything, because the full set of complexities present in reality mean any two situations always differ from each other somehow, and we must choose what to focus on as relevant and what to ignore. “Where there’s smoke there’s fire” is an example of a model. Mathematical models are models that are explicit about questions of the form “how much” or “to what extent”, by using equations. For example, the more nuanced question of “Given that there’s smoke, how likely is there to be fire?” would require a mathematical model instead of a proverb.
Explore our Articles
August 2021
No 15: Daily lateral flow testing of COVID-19 contacts can be as effective as quarantine and carries lower social/economic costs by Chris Wymant
No 14: Evaluating epidemiological impacts of the NHS COVID-19 app: an August 2021 update by Michelle Kendall and Christophe Fraser
March 2021
No 13: Characterising SARS-CoV-2 genetic diversity and transmission by Katrina Lythgoe, Matthew Hall, Tanya Golubchik and Ana Bulas Cruz
February 2021
No 12:Evaluation of the NHS COVID-19 contact tracing app by Lucie Abeler-Dörner
October 2020
No 11: Epidemiological changes on the Isle of Wight after the launch of the NHS Test and Trace programme: a preliminary analysis by Michelle Kendall
September 2020
No 10: How to simulate the COVID-19 epidemic and public health interventions like lockdown, testing and contact tracing by Lucie Abeler-Dörner
No 9: COVID-19 is most frequently transmitted shortly before and after symptom onset by Luca Ferretti and Ana Bulas Cruz
No 8: Exposure notifications have a positive impact at all levels of uptake by Lucie Abeler-Dörner
July 2020
No 7: Lessons to be learned from the Isle of Wight’s success in controlling its COVID-19 epidemic by Michelle Kendall
May 2020
No 6: Balancing functionality with privacy for effective contact tracing apps by David Bonsall and Michelle Kendall
April 2020
No 5: Epidemiological model available open source to drive innovative strategies to stop coronavirus epidemic by Christophe Fraser, David Bonsall and Robert Hinch
No 4: Simulating the use of digital contact tracing to control COVID-19 by Michelle Kendall
No 3: User Acceptance of Mobile Contact Tracing App by Ana Bulas Cruz
March 2020
No 2: Ethics and Social acceptance by Ana Bulas Cruz
No 1: Maths tells us the epidemic can be stopped by Lucie Abeler-Dörner
BLOG 14: Evaluating epidemiological impacts of the NHS COVID-19 app: an August 2021 update
2nd August 2021
“You need to self-isolate” are words which none of us want to hear, and being “pinged” by the NHS COVID-19 app can be annoying and disruptive. Does contact tracing via an app actually work? Can we really be confident that following the advice will help to protect ourselves and others? The app has been sending out record numbers of self-isolation notifications through June and July 2021 so it is more important than ever to try to quantify any positive impact it is having.
The good news is that the app is behaving as it should. Notifications have risen precisely in line with growing case numbers and increasing contact rates as people move around and interact more, as shown in the graph below (note the log scale on the y-axis).
We know from the theory that quick, large-scale contact tracing helps to reduce the spread of coronavirus, so we have been checking to see whether the results bear out in practise.
Although we can see that the app is operating as expected and notifying people in a timely manner so that they can reduce their chances of passing the virus on, it is hard to quantify exactly how many infections the app is preventing. Partly that is because it’s always hard to count things which haven’t happened! And particularly it’s because the app is private by design, very little data is collected, and what data we do have is completely anonymised. There is no linkage in the data to address who infected whom, or when, or where, and so some estimates of the app’s effectiveness naturally come with quite a range of uncertainty.
Here are some things we do know. From publicly available data we know that the proportion of people entering their positive test results into the app has been growing through 2021, reaching around 50% from mid June.
When a positive test result was entered and the user consented to contact tracing, the number of notifications sent out to close contacts has been changing as contact rates have changed with the steps of the roadmap, and in July 2021 was averaging 4-5 notifications per positive case.
We also find that 15% of people who entered a positive test result into the app had previously been advised to self-isolate by the app, so that if they self-isolated successfully (or even nearly successfully!) they will have reduced their chances of passing on their infection.
From these and similar measures we are able to estimate the number of people per day who received a notification and later were confirmed to be infected, which peaked at nearly 6000 per day in mid July. These people would likely have gone on to infect a number of others, where that number is based on the national R rate but takes into account imperfect adherence to quarantine and the fact that delays in tracing individuals can mean they were only traced halfway through their infectious period. This calculation gets a bit complicated: the fact that it relies on the national R rate means that the app averts more cases per day when the epidemic is growing, but the more cases the app averts, the more it reduces the observed R itself. Surveys estimate at least 60% effective adherence to a contact tracing self-isolation notice so we use this figure whilst noting that it is likely to change over time in line with public feeling about the importance of limiting the spread, and trust in the app.
Despite all these caveats - and there are more to come! - using this approach we estimate that the app directly prevented between 15,000 and 30,000 cases (depending on how well individuals self-isolated) thanks to notifications sent out during the first three weeks of July.
Taking into account a conservative calculation of chains of transmission averted, we find that the app may easily have prevented as many as 50,000 cases as a result of these 1-21 July notifications, equating to 1,400-1,600 hospitalisations. We will soon be publishing data and the full details of our current methods so that our results can be properly scrutinised and replicated. We continue to develop our methods each week as the epidemic progresses.
We do not know what proportion of these contacts may have also been traced by other methods (word of mouth or standard contact tracing), or would soon have experienced symptoms so self-isolated anyway. This means that our results could be over-estimates, although we do know that app contact tracing is typically faster than standard contact tracing, meaning that contacts receive their app tracing notification earlier in their infectious period and - when test result turnarounds are quick - before they develop symptoms.
On the other hand, our results could be under-estimates because not everyone who is infected gets tested, nor do all positive tests get registered through the app. Further, we tend to consider cases directly averted, that is, cases which would have been caused by an app user if they had not self-isolated when notified. Of course, these people who were directly protected may have gone on to infect others, meaning that we significantly under-estimate the number of cases averted in instances where there would have been a long transmission chain stemming from them, and/or where one of those infected would have been a superspreader. When we do consider this transmission chain effect we conservatively limit our estimates to “two-step” chains, that is, the individuals directly protected by the app, and the people who they would have infected.
There is also a peculiar “triangular effect” at play. This works as follows: imagine that you and I regularly take the same bus ride to work, and one day we both sit near to a person who is infectious with covid-19. It so happens that you and this infectious person both use the app, and I don’t. When the person tests positive you receive a contact tracing notification and self-isolate for up to 10 days but I don’t, and I keep using the bus.
Now, if it happened that this person had infected you, then the contact tracing process would happen to have worked as traditionally expected, notifying you so that you reduced your chances of passing the virus on. But if it happened that the infectious person infected me and not you, perhaps just because of the way the air was flowing, the quality of your mask, or some other random factors, it might appear that the contact tracing had gone wrong and that you self-isolated unnecessarily. But, by self-isolating for those days, you avoided using the bus and getting infected by me. In this way, the app not only prevents people from passing on their own infections but can even prevent them getting infected in the first place when infections are circulating in their regular contact networks.
This triangle effect is a social network effect, and network effects are expected to become increasingly pronounced during periods of intense transmission, since many people mix in dense social and work networks. The app will ask people to self-isolate who socialise or work in contact networks where the virus is circulating. Network theory states that ‘my contacts have more contacts than I do’ and therefore contact tracing may have a disproportionate effect in preventing transmissions in environments where there is a high risk of transmission. Each step of the roadmap for reopening has added more daily connections on contact networks and introduced more opportunities for superspreader events, and with high prevalence in the recent Delta wave it is possible that contact tracing has become more relevant than ever before in this pandemic. We are currently investigating the contribution of the so-called ‘pingdemic’ to the subsequent drop in cases in recent weeks.
Finally, we have not evaluated pre-emptive cautious behavioural effects (“I really want to go to the family gathering next weekend so I will do fewer risky things now to reduce the chances of getting pinged”, thereby also reducing one’s chances of getting infected) nor have we included in our analysis the effect of QR code check-ins and subsequent advice to self-isolate.
A published analysis of the effectiveness of the app in England and Wales during the winter Alpha wave found that the app helped to reduce the size of the epidemic by roughly 14-24%. The lower of these estimates came from a modelling analysis somewhat similar to the method described here, whereas the larger estimate came from a statistical analysis comparing epidemics in local authorities with varying app uptake. A similar statistical analysis is in progress for the Delta wave.
The app’s primary purposes are to help the user: accessing information, checking symptoms, booking tests, receiving test results, and - via contact tracing - reducing the user’s chances of becoming infected and infecting others. Those "others" may or may not be app users, and so the app helps everyone in reducing the size of the overall epidemic.
A secondary benefit of the app in summer 2021 has been in providing epidemiological information about the state of the epidemic in England and Wales. The number of notifications per person testing positive gives an indication of national contact rates. Much like Google mobility or similar measures, it indicates when people are interacting more; uniquely among mobility measures it gives an insight into the contact rates of people who are currently or very recently infectious. From this we are able to estimate the R rate amongst app users before they test positive, which serves as a leading indicator for the national R rate as shown below in red and blue respectively. The app-based estimate is a lagging indicator, meaning that it only tells us about the R rate in the general population one or two weeks beforehand, but it is a leading indicator in the sense that its results are available before they can be reliably estimated from national case data. In the plot below it can be seen that data available on the same day (23 July) reliably showed a sharp decrease in R from app data which was not visible from national case data at the time but started to become apparent within the following week; it took a full week’s more data for the blue line to show the same extent of decline as in the red line here. These measures have been providing valuable insights into the progression of the epidemic.
The NHS COVID-19 app is an important part of the toolkit in limiting the spread of covid-19 and helping protect the NHS in England and Wales. Thanks are due to everyone who has been using the app, particularly to those who have entered positive test results and agreed to contact tracing, and those who have endured periods of self-isolation. We encourage eligible readers to install the app if they haven’t already and to keep actively using it, which will of course become a little easier for the fully vaccinated after 16 August when the advice will change from self-isolation to daily testing. Previous analysis showed that every 1% increase in users led to a 1-2% reduction in national cases so that every new user makes a positive difference, and the network effect means that if you and many of those you meet are using the app, you will be less likely to get infected or to infect others.
By Dr Michelle Kendall, University of Warwick and Professor Christophe Fraser, Big Data Institute, University of Oxford, independent scientific advisors for the NHS COVID-19 App.
BLOG 13: Characterising SARS-CoV-2 genetic diversity and transmission
9th March 2021
Over the last year there has been an unprecedented effort to sample and sequence SARS-CoV-2 in order to monitor its spread and identify new variants. Most sequencing analyses rely on a single consensus (average) sequence for each infected person, but, in practice, individual infections consist of a large population of viral particles, and these particles can have different mutations. In this study, we used deep sequencing to obtain multiple SARS-CoV-2 genomic sequences per individual to characterise the genetic variation of their viral population. This allows us to investigate key aspects of SARS-CoV-2 evolution: how many viral variants are present in a typical infection (answer - not many), how does this diversity change as infection progresses (answer - a lot), and how many viral particles are transmitted when one individual infects another (answer - usually one).
What did we do
We sequenced 1313 nasopharyngeal swabs from over 1000 individuals, including 41 with multiple samples taken at 2 to 4 timepoints. The samples were collected from symptomatic individuals on admission to hospital, and from healthcare workers and some of their close contact, by Oxford University Hospitals and Basingstoke and North Hampshire Hospital, between March and June 2020. This spanned the substantial part of the UK’s first wave.
Characterising diversity - a cautionary tale
The SARS-CoV-2 genome has nearly 30,000 nucleotide positions. We classed these sites as variant if more than 3% of ‘reads’ at that site in a given sample differed from the consensus - following convention, we call these iSNVs (intrahost single nucleotide variants). Characterising diversity is not just a matter of counting mutations. Observed variants can also be artefacts - for example caused by the chemicals used to store and process the samples. We took care to mask positions in the genome prone to this - not doing so can lead to vastly overestimating how much diversity is transmitted among individuals and the transmission bottleneck size (how many viral particles are transmitted). The viral load (how many viral particles there are in a sample) is also important - at low viral loads random sampling effects can result in overestimating how many iSNVs are present.
There is a consistent pattern of low within host diversity
To confidently characterise intra host diversity, we selected 563 iSNVs that were observed in at least one individual with a high viral load. Most samples had less than five of these iSNVs. Considering the subset of samples with a high viral load, 66% of them had at least one iSNV, with a mean of 1.4 per sample. This indicates that SARS-CoV-2 within-host diversity is low during acute infection, when there is a high viral burden and transmission is most likely. Characterising diversity later in infection is harder because viral loads tend to be lower, but considering all iSNVs, our results indicate a dynamic landscape of variants coming and going during infection, and increasing diversity as infection progresses.
A narrow transmission bottleneck hinders the spread of new within host variants
Understanding how likely it is that new SARS-CoV-2 mutations will emerge and spread requires knowledge of how many viral particles are transmitted - if a new variant emerges in an individual. When the transmission bottleneck size is small, this new mutation is unlikely to be transmitted. However, if by chance it is transmitted, it could dominante the viral population in the recipient making onward spread in the population much more likely. We used epidemiological data to select households where two people tested positive within two weeks of each other. Direct transmission was assumed if the consensus sequences had fewer than three differences, and the infecting (source) individual had to have at least two clearly distinct viruses. 14 household pairs met these criteria, with estimated bottleneck sizes varying between between 1 and 8 viral particles - in the majority of cases it was 1. This is a narrow margin for transmitting mutations and inevitably most variants will be lost when the virus changes host.
Evolution in Action
Even though transmitting a new SARS-CoV-2 is a relatively rare event, new mutations can become widespread. We hypothesized that this could be due to a variant emerging within an infected person, becoming dominant in that host, and then transmitting, or due to being transmitted as a minority variant and then becoming the most common in the new host. To investigate, we built a phylogeny based on consensus sequences, where the branching indicates differences between the consensus sequences of different samples. When a new variant becomes dominant, there will be a change in the consensus of the phylogeny. We observed a strong correlation between the presence of iSNVs and a consensus change, which is consistent with our hypothesis. The figure below represents a cluster associated with healthcare where there was onward transmission to close contacts. Here the red branches represent a change from consensus G to A, but where prior to this we can see A as an iSNV, (represented by the red dot) and after the change G remains present but now as a minority variant (green dot). We are witnessing ‘evolution in action’.
To conclude
Our results indicate that acute SARS-CoV-2 infections tend to have few, if any, genetic variants. Even when mutations do arise, it will be hard for them to spread because of the narrow transmission bottleneck. Most of the time only the most common variant will be passed on and the others will be lost. In rarer cases, a minority variant might get through the bottleneck and become the consensus sequence, or there might be co-transmission of multiple variants, leading to a mixed infection.
These findings imply that within-host emergence of vaccine- and therapeutic-escape mutations is likely to be relatively rare, at least during early infection when viral loads are high. However, when there are more than 100 million people infected worldwide, rare events can still occur. We have observed immune escape variants in high viral load samples, and as population immunity increases and vaccines and therapeutics are rolled out, there will be an increasing selective pressure on the virus. This underlines the need to drive down the number of infections, making these rare events even rarer, and continued monitoring of SARS-CoV-2, so we can detect concerning new variants and act before they become dominant in a population.
by Katrina Lythgoe, Matthew Hall, Tanya Golubchik and Ana Bulas Cruz
BLOG 12: Evaluation of the NHS COVID-19 contact tracing app
It’s been over four months now since the NHS COVID-19 contact tracing app was launched in England and Wales on 24th September 2020. There was a lot of excitement around the launch and millions of people downloaded the app in the first week. So did it prevent anybody from catching COVID-19? A new analysis by our team at the University of Oxford and Mark Briers' team at the Alan Turing Institute suggests it averted between 200,000 and 900,000 infections in the period between 1 October and 31 December, when 1.9 million people in England and Wales were infected with COVID-19.
The app sent out over 1.5 million quarantine notifications
Evaluating the effect of the contact tracing app is not easy, as the privacy-preserving approach means that hardly any data is collected from individual users. The information available is summary information, like the number of daily notifications and the percentage of notified users who declare a positive test. Out of 33.7 million eligible people with compatible smartphones, the app has been downloaded on 21 million unique devices, and is regularly used by at least 16.5 million users in England and Wales. This corresponds to 49% of the eligible population with compatible smartphones and 28% of the total population including children. Between 1 October and 31 December, the app sent out over 1.5 million quarantine notifications, most of them in the second half of December when cases were rising rapidly across the UK due to the new B117 viral variant. This corresponds to 4.5 notifications for each user who shared their positive test result though the app.
App usage differs by area
Users of the NHS app enter a part of their postcode, which makes it possible to map active app usage to the 338 administrative areas in England and Wales. The map below shows that app usage varies considerably across areas of England and Wales.
We used this geographical variability in app usage to ask if areas with high app usage saw fewer infections than areas of low usage. From earlier work, we expect the number of infections to drop steadily with each new user.
We started by comparing app usage with the number of infections, and found that areas with high app usage indeed had fewer cases between 1 October and 31 December than areas of low usage. However, there are two problems with this simple comparison:
We know very little about the people who use the app, and it is likely that they are not representative of the population as a whole.
Many other interventions have taken place, including a national lockdown and different local restrictions introduced through the government’s tier system.
We therefore used two independent approaches to control for these confounding factors. None of these approaches can fully control for all unknown factors, but each controls for some of the key factors we expect to have an effect.
Predicting averted infections from app notifications
The app provides data on how many people use it in each postcode area, how many notifications are sent out to contacts of infected individuals, and how many of those later register a positive test with the app. Combining these data with data from two surveys of how well people adhere to quarantine, we can model how many infections were averted by the app. By this approach, the app led to around 284,000 fewer infections between 1 October and 31 December. The graph below shows that in areas of high app usage, cases could have been almost 30% higher without the app. For each 1% increase in app usage, an area saw on average a reduction in cases of 0.8%.
Predicting averted infections from cases in “comparable” areas
App users are very likely not representative of the population as a whole, both in their behaviour as well as in their demographics. Indeed, we found that app users live in areas which are more rural and have lower levels of poverty, compared to the national average. Areas with a high fraction of app users also already had fewer infections in August and September, before the launch of the app. This suggests that app usage is associated with multiple measured and unmeasured factors that are also associated with lower case numbers.
To control for measured factors like areas being more rural or having lower levels of poverty, we directly included these variables into the analysis. To control for unmeasured behavioural factors like being more cautious, wearing masks, avoiding close contacts, being more likely to seek a test, adhering more closely to quarantine etc, we divided all 338 areas into five groups based on cases per capita in August/September. Comparing areas within these quintiles showed that these demographic and behavioural factors were very good predictors of the number of cases in each area before the launch of the app.
To control for local restrictions, we then compared the 338 regions to neighbouring regions within the same quintiles. For each percent increase in app usage, an area saw on average a reduction in cases of 2.3%. This corresponds to approximately 594,000 infections averted in the period between 1 October and 31 December (with a 95% confidence interval of 317,000 to 914,000 infections).
Improving the app increases the number of infections averted
The app was continuously updated over the study period. These were mostly small changes but in early November the risk scoring algorithm got a major overhaul, making the identification of individuals at risk of exposure from an infected individual much more accurate. The number of notifications per infected individual doubled as the app got better at detecting high-risk contacts. Since this change occurred only within the app (government policy and app users were the same directly before and after) we were able to analyse the impact of this update on the number of infections averted by the app.
We repeated both analyses, splitting them up into periods before and after the app upgrade. Overall, we found that for a 1% increase in app usage, cases were reduced by 0.8%. Considering the two periods separately, we found that 1% increased app usage reduced cases by 0.3% before the upgrade and by 0.9% after the upgrade. The matching neighbour analysis found an overall reduction in cases by 2.3% for 1% increase in app usage, split into 1.1% before and 2.7% after the upgrade. This is encouraging, as future updates could lead to further reductions in the number of cases.
To conclude
Taking all analyses together, the data suggests that from 1 October to 31 December, between 200,000 and 900,000 COVID-19 infections were averted by the app. This is a significant number compared to the 1.9 million infections which occurred during this period. Owing to the nature of the available data, it is not possible to perfectly control for confounding factors. However, we have used different approaches to control for the most important confounders. Even with this adjustment, the effect of the app remains very strong. A dedicated study conducting surveys on app users and their contacts would be needed to firmly establish causation.
The data furthermore suggests that if all areas had reached the 90th percentile of usage in the areas with the highest uptake (36%), a further 200,000 to 400,000 infections could have been averted on top of those averted already.
The public health message is very clear: Areas with high app usage suffered significantly fewer cases of COVID-19 than comparable areas with lower app usage, and there are strong reasons to believe that part of this effect is due to the app. The NHS COVID-19 contact tracing app is a powerful and sustainable public health intervention which will continue to reduce the number of infections in the future, especially if further resources are invested to increase usage.
For another blog article by Mark Briers, Christophe Fraser and Chris Holmes, visit this page.
BLOG 11: Epidemiological changes on the Isle of Wight after the launch of the NHS Test and Trace programme: a preliminary analysis
15 October 2020
The NHS Test and Trace programme was rolled out in May 2020, first on the Isle of Wight (5th) and then across England (18-28th). We looked to see whether there were any changes in the course of the epidemic which could be attributable to the Test and Trace programme, and whether there was any difference in the programme’s effect on the Isle of Wight, where it also included version 1 of the NHS contact tracing app. Here we provide a brief summary of our findings; the full details are now published in The Lancet Digital Health.
Using Public Health England’s daily “case counts” (positive swabs) we developed a method to estimate the number of new infections per day - there is a lag because people rarely get swabbed the day they are infected. The graph below shows Pillar 1 (hospital) case counts on the Isle of Wight in green, and our estimated number of new infections per day in red. We see that new infections were decreasing from mid-April. There was a slight increase at the end of April which might be an artefact of increased testing from 5th May, but then there was a noticeable decrease in infections despite the increased testing shortly after the Test and Trace launch.
Similarly, the reproduction number R - the average number of further infections each infected individual causes - decreased rapidly on the island following the Test and Trace launch. These early results support the idea that the Test and Trace programme helped to bring the epidemic under control on the Isle of Wight.
But how did the Isle of Wight’s epidemic compare to those in other areas? We used a variety of methods which all led to the same conclusion: something quite different happened on the Isle of Wight. The incidence and R rate dropped more rapidly than in other areas over comparable time periods. The graph below is a “nowcast” of the expected number of new hospital cases in the near future. The island had quite a large epidemic in mid-April, positioned “middle of the pack” when compared with other areas until May; by June-July it led the way with case counts down to less than one a week, even when we included Pillar 2 (community) cases. Using another method we found that in March-May the Isle of Wight was ranked amongst the worst R rates in England (147th out of 150) but by mid-June it was positioned 12th best.
The success on the Isle of Wight is striking, and the course of its epidemic was different from other areas - a statistically significant difference. To establish that the Test and Trace programme was the cause of the success would require ruling out all other possible causes, for which we would need more than the publicly available data. The island’s epidemic control certainly warrants further investigation because it might translate to other local and national strategies.
If Test and Trace did have an impact we will also need to disentangle which aspects had the greatest effect. Was it the huge advertising campaign at the launch? Did people self-isolate more carefully after a positive test result? Was the contact tracing programme getting ahead of the virus, advising people to quarantine before they infected others? And if so, was that primarily human contact tracing or via the app? A key piece of the puzzle would be: of the people who tested positive, how many were already self-isolating because they had been contact-traced? We hope that more data will soon be made available so that we can answer these important questions.
We have made our analyses available in LocalCovidTracker where you can explore them in more detail, and see our ongoing tracker which updates daily. It enables you to view the epidemic trend in individual local authorities in England and Wales to identify outbreaks and see the effects of interventions like local lockdowns.
by Michelle Kendall (Research Fellow at University of Warwick, research conducted while Senior Researcher at Oxford University’s Nuffield Department of Medicine)
BLOG 10: How to simulate the COVID-19 epidemic and public health interventions like lockdown, testing and contact tracing
22 September 2020
Mathematical models are needed to tell us how much one thing depends on another. They can be used for understanding and forecasting various phenomena, and we regularly see them applied to model climate change, election results or financial indicators. Throughout the COVID-19 epidemic, mathematical modelling has been used to predict the number of cases and deaths, hospital capacity, requirements for PPE, tests and contact tracers, and many more parameters. These models need to be sufficiently complex to yield realistic predictions, but sufficiently simple and fast enough to use. They can also be used to simulate different interventions, like testing, contact tracing and mask wearing, and how they influence each other when implemented at the same time, and design strategies that are likely to work. In our view, this is the most powerful use of mathematical models in the area of infectious diseases.
Our latest study presents OpenABM-Covid19, the mathematical model of COVID-19 we have used to simulate the epidemic and test different interventions. What can this particular model add to the body of modelling on the COVID-19 epidemic? We think it has five important advantages compared to many of the other models: it is
1) able to model the spread of the disease dependent on the different networks in which people interact
2) granular enough to model individual behaviours and interventions,
3) computationally fast,
4) open-source and modular, which enables others to add to it, and
5) well documented and widely tested, further encouraging adoption and providing validation.
Let’s have a look at these one by one.
1) Networks
In our model, individuals interact in three networks: their home network where they interact daily with their household; an occupation network in which they meet a subset of their colleagues every day; and a random network which corresponds to random encounters e.g. in restaurants or on public transport. The network structure makes it possible to implement interventions differently between the networks. During lockdown, for example, contacts in the occupation and random networks decrease for most people, whereas contacts within the household increase. Manual contact tracing reaches contacts in the home and occupation network, whereas digital contact tracing also reaches the random network. The networks are also different across age groups: children interact in an “occupation” network which is different from the adult network and represents schools. Retired people interact mainly with each other in social networks. The demographics of these networks are based on UK data but can easily be adjusted to fit any other country for which data are available.
2) Granularity
Mathematical models can be split into two very broad categories: those which model groups of people (e.g. susceptibles, infected, and recovered) and those which model individuals. Our model is an agent-based model (ABM) which models individuals. Our model population is set at 1 million people, with an age, gender and household size based on data from the UK's Office for National Statistics. This granularity allows the simulation of interventions that act on an individual-level like testing, quarantine, and digital and manual contact tracing.
3) Speed
Agent-based models tend to be computationally intensive, as they model the behaviour of individual people and all their interactions over a period of time. Most agent-based models therefore tend to be limited by the number of individuals that can be simulated, - often in the hundreds rather than the thousands. In coding our model, we have prioritised computational efficiency. The model takes approximately 3 seconds to run for each day of the simulation on a standard laptop computer. This makes it possible to simulate realistic and meaningful numbers of individuals, e.g. a city of 1 million inhabitants.
4) Collaboration
The model is freely available on Github and coded in a modular way which encourages collaboration. So far, a modeller from the contact tracing team in Singapore has added a module on manual contact tracing, researchers at the Francis Crick Institute have contributed a module on infections in hospitals, and a team from Goldman Sachs has integrated an economics module. A team from Google Research added a module on splitting up the work network by industry, re-coded the model to run on distributed computing units, and analysed the efficacy of the Apple/Google contact tracing app in collaboration with our team at Oxford University. The model also has a Python interface which allows a broader group of people to interact with it, thanks to the efforts of a team at IBM UK
5) Documentation and testing framework
For wide adoption of a model, good documentation and a formal testing framework are essential. An automated testing framework means that after each addition to the code, internal tests can be run to check if the model still follows the basic principles that were chosen in the beginning. The tests will check whether the model behaves as expected, e.g. if the fraction of people dying from COVID is set to 0 in all age groups, there mustn’t be any deaths at the end of the simulation. Tests can also check if the input of the user clashes with other parameters, e.g. if a user entered the demographic structure of Nigeria, for example, which is very different from the UK structure, the test would prompt the user to also update the data for the household structure. These tests are of particular importance if a model is open-source and therefore improved and expanded by developers from different teams. Currently, over 200 of these tests are run routinely. Extensive documentation and exhaustive testing frameworks take time to develop which is often outside the possibilities of an academic group. We have received a lot of help in this area from industry collaborators for which we are very grateful.
In conclusion, we present a model which is able to quickly test the impact of key policies in the fight against COVID-19 like lockdowns, testing, quarantine and contact tracing, while simulating a sufficiently large number of individuals interacting in different networks. Its modularity, documentation, testing framework, and accessibility via Python invite contributions to this open-source model. We are aiming to create a community of developers and users to supply policy makers with a tool to simulate the likely impact of different non-pharmaceutical interventions tailored to individual countries. The model is already being used by NHS England and Wales to predict hospital capacity. Future plans include the addition of a geographical structure in which individuals live and work in one region but can interact with individuals from neighbouring regions.
BLOG 9: COVID-19 is most frequently transmitted shortly before and after symptom onset
11 September 2020
Understanding the timing of transmission dynamics of Covid-19 is crucial for optimising public health interventions and reducing the burden on communities. Knowledge of the period of high infectiousness is particularly useful in the context of contact tracing as it allows to identify the contacts at highest risk of having contracted the disease, maximising the number of infectious individuals in quarantine while minimising the number of people unnecessarily isolated. The timing of Covid-19 transmission is centered around the question “When is someone who has been infected with COVID-19 most likely to infect others?”.
The analysis focuses on individuals who eventually develop symptoms, studying 191 transmission pairs from multiple data sets. Each transmission pair includes an index case and a secondary case, infected by the index case. Datasets were selected because they contained information not only about the dates of onset of symptoms for both cases, but also intervals of exposure for the index and/or secondary case. For these pairs and individuals, four key timings related to Covid-19 transmission were known or inferred:
the incubation period - the time from someone becoming infected until they develop of symptoms;
the serial interval - the time between the index case and the secondary case developing symptoms;
the generation time - the interval between an individual getting infected and then infecting somebody else;
TOST - the time from onset of symptoms of the index case to when they transmit the virus to someone else.
As shown below, the TOST distribution is centered around 0 days, regardless of the heterogeneities among datasets, as shown in the figure below. The majority of transmissions occur within a few days before or after symptom onset. This means that transmission events peak at the time of onset of symptoms of the index case. From an individual perspective, if I am infected, I will be most infectious close to the time when I first develop symptoms.
This could mean that the infectiousness of an individual happens to peak as they develop symptoms or that their infectiousness is actually dependent on the development of symptoms. Fitting the data to appropriate models to test these hypotheses reveals that infectiousness is in fact determined by symptom onset, rather than coincidentally synchronised with it.
It is worth noting that the proportion of transmissions that occur after the onset of symptoms is dependent on several factors such as local interventions and individual behaviour. In most countries affected by the pandemic, individuals were advised and/or required to self-isolate if they experienced symptoms, reducing the chance of spreading the virus when symptomatic. Hence, the lack of significant transmission a few days after the onset of symptoms might not be a biological characteristic of the virus but rather a reflection of the epidemiological biases of the dataset.
It was also hypothesised that the incubation time could affect the length of the infectious period. The effect of the incubation time on the generation time and the TOST was modelled and inferred to follow the pattern illustrated below.
These results indicate that infectiousness increases gradually from the time of infection until the onset of symptoms. In other words, individuals who take longer to develop symptoms will have a proportionally longer period when they are likely to infect others.
A key result from this paper is the quantification of the relative amount of transmissions that occurs before symptoms develop (pre-symptomatic), in the day of onset of symptoms and the day after (early symptomatic) and late symptomatic periods. Approximately 40% of transmissions were found to be pre-symptomatic. This emphasises the need for mass testing , effective contact tracing and continued good practices, such as physical/social distancing, increased hygiene and wearing a face mask.
Crucially, it was found that 35% of transmissions occur on the same day or the day following the onset of symptoms. This heavily emphasises the need to impose immediate self-isolation on even mild symptoms. It also highlights the need for testing systems that are accessible and have a quick turnaround time for the results.
The late symptomatic transmission corresponded to only 24% of all transmissions. However, as previously mentioned, this should be taken with a grain of salt as this value is affected by self-isolation and public health interventions.
This characterisation of the timing of transmission of Covid-19 has some direct implications for public health policy, in particular related to traditional or digital contact tracing.
Whenever data regarding the date of infection is available, it should be used to enable a more effective contact tracing by applying longer tracing intervals when longer incubation times (> 4-5 days) are observed. This avoids missing potential transmission from individuals that have been infectious for a longer time.
Data regarding onset of symptoms should be recorded where possible, preferably within the app for digital contact tracing. This information is valuable when building risk scoring algorithms that help to minimise both the people unnecessarily quarantining and the infectious individuals not in isolation.
Overall, these results strongly emphasise the difficulties in detecting infectious individuals and stopping onward transmissions, and consequently the importance of investing in preventative measures such as community testing and contact tracing, as well as continued good practices such as increased (hand) hygiene, wearing face masks and physical distancing, all of which can stop transmission before and near the symptom onset with relatively low social and economic costs.
Our results also highlight how self-isolation of cases is much less effective if it starts 24-48h after symptoms appear, due to the high risk of transmission around symptom onset. Individuals with ambiguous symptoms should take extra precautions immediately, e.g. following good practices and reducing their social contacts as much as possible for 2-3 days after feeling under the weather. Immediate self-isolation of all suspected cases would be advisable, but could carry a significant economic cost; introduction and widespread availability of rapid home tests could be a game-changer.
by Luca Ferretti and Ana Bulas Cruz
BLOG 8: Exposure notifications have a positive impact at all levels of uptake
4 September 2020
Contact tracing has been a successful method for containing the spread of infectious diseases for centuries. The global COVID-19 epidemic has added a new tool to the toolbox of many public health authorities: digital contact tracing. Apps have now been launched in many different countries and states. This is a new technology - how much can they contribute to containing the epidemic? And how many people need to adopt the app to make a difference? Can exposure notifications usefully interact with traditional contact tracing? Can they help regions to maintain low enough case counts to start opening up again?
We set out to answer these questions in collaboration with Google’s Research team by modelling the introduction of a digital contact tracing app based on Apple’s and Google’s Exposure Notifications System (ENS) in the three largest counties in Washington state - King, Pierce and Snohomish - which include Seattle (in King County) and the regions to the north and the south.
We used US census data to create virtual populations of the three US counties that match the real populations in age, gender, and housing type. The virtual inhabitants of the three counties interacted at home, work, socially and in “random” situations such as travelling on a train or eating in a restaurant. The workplaces were modelled using employment numbers from the 20 most common industries in the three counties e.g. retail industry and health care. COVID-19 was then introduced into the simulation and the virtual inhabitants transmitted the virus and died at rates which matched the real epidemics in King, Pierce and Snohomish Counties.
On 11 July 2020, the simulation starts to differ from reality: we introduce exposure notifications and forecast how the epidemic will develop until the end of the year depending on how many people adopt the app. The graph below shows the predicted number of daily new infections in King County - the graphs for Pierce and Snohomish look very similar, see the full results. Our key finding is that the success of exposure notifications doesn't depend on everybody downloading and using it. Uptakes of 15% of the population already reduce the number of infections by 4-12% in the different counties, and consequently the number of deaths by 2-15%. At 30% uptake, the reduction in infections and deaths is already quite substantial, averting 400 new infections a day at the peak of the simulated epidemic. These levels of uptake are well within reach. Most countries that have released tracing apps based on opt-in installation have at this stage reached uptakes between 10% and 35% of the population. But importantly, the app doesn’t depend on everybody downloading and using it.
We next asked how the app could be combined with manual contact tracing. The office of the Governor of Washington State recommends a minimum of 15 contact tracers per 100,000 people. Based on data from King County, we assumed that tracing would take 24 hours, with one contact tracer contacting one infected person and their contacts (three on average) per day. The graph below shows the epidemic in King County with and without manual contact tracing (CT) at 15 tracers per 100,000 inhabitants, and with and without exposure notifications used by 30% of the population. Both interventions decrease the number of newly infected people, the app more than the manual contact tracing in this simulation. Of note, the two interventions have an “additive” effect which means they each reduce the number of infections on their own but especially when used together. In reality this effect could be even greater if the two types of tracing were part of a coordinated response, with manual contact tracing focusing on population groups which are less likely to adopt the app. Other interventions like social distancing, mask usage, personal protective equipment, etc. can also be modelled and reduce the predicted number of infected people even further.
So what does that mean for the epidemic? Is it safe to open back up or will this inevitably lead to a second wave? To simulate the reopening, we used Google mobility data from 11 July 2020 (0% reopen) compared to pre-epidemic mobility on 1 March 2020 (100% reopen). The simulation shows how much reopening is possible with different combinations of manual and digital contact tracing. For example, if the total percentage of infected people in the population should be kept below 25%, the simulation tells us that only 10% of re-opening is possible without an app, while 100% re-opening is possible if everybody owning a smartphone used the app. Intermediate levels of app uptake allow a re-opening of 20-30%. This stresses the importance of using all options that we have to suppress the resurgence of COVID-19. Social distancing and mask wearing are likely to remain part of our lives for the foreseeable future.
As part of the Washington State Department of Health’s “Safe Start” plan, a key target metric to reopen Washington is to reach fewer than 25 new cases per 100,000 inhabitants over a course of two weeks (less than on average 2 cases per day). The graph below shows how a combination of manual contact tracing and exposure notifications can reduce the time to achieve this goal against a background of social distancing (mobility levels of May 2020). Without digital and manual contact tracing the goal is not reached within the time limits of the simulation (161+ days), whereas high levels of app uptake and a doubling of the contact tracing workforce push the time to reach this key metric to below two months. Even at currently recommended levels of manual contact tracing staffing and moderate levels of app uptake, dropping below 25 cases per 100,000 inhabitants per fortnight can be brought forward by more than 40 days.
We conclude that
● The use of exposure notifications is useful at all levels of uptake
● Exposure notifications can and should be combined with other interventions like manual contact tracing and social distancing to reduce as many infections and save as many lives as possible
● A combination of interventions can help keep the counties and states to control the epidemic and, in the case of the three counties modelled here, reach consistently low levels of infections within months.
While we have put care into using the best possible data available, the simulation still represents a simplification of the real world. The results should therefore be viewed as an exploration of possible outcomes and the relative impacts of different interventions, and not as precise predictions. For example, we have modelled each county separately and have not modelled movement across county borders. This is something we’re working to include. We have also assumed that it takes two days from symptom onset to receiving a positive test result, which requires rapid testing systems and an awareness in the population that speed is of the essence. Contact tracing should be part of an integrated strategy needed and should also include widespread testing with rapid turnaround of results, and consistent messaging, so that people know to request tests when appropriate, and to isolate when sick, infected, or traced by contact tracers and exposure notifications. These measures together may reduce the burden of social distancing measures that accumulates over time.
In future work, we will share results on the impact of cross-border movement, and how exposure notifications can be strategically and cost-effectively integrated into local, regional and national policy frameworks.
BLOG 7: Lessons to be learned from the Isle of Wight’s success in controlling its COVID-19 epidemic
15 July 2020
Public health authorities across the globe have been responding to the COVID-19 pandemic in different ways, with a variety of results. It is important to try to learn from each other’s successes, but sometimes there are challenges in translating effective strategies between settings with different demographics and cultures. For epidemiologists it can be particularly informative if different areas within the same country adopt different strategies because it allows for a more direct comparison. This is exactly what happened in the UK in May: a Test and Trace (TT) programme was launched first on the Isle of Wight on 5th May before being rolled out to the rest of the UK on the 18-28th May. The Isle of Wight’s roll-out also included version 1 of the NHS contact tracing app. We looked to see if the Isle of Wight’s epidemic changed after the TT launch and if the Isle of Wight fared any differently from comparable areas of the UK.
We used publicly available “case counts” - the numbers of new COVID positive swabs which were taken each day, as reported by Public Health England. From these we estimated the number of new infections per day (among people who eventually became recorded as a case). To do this we note that people rarely get swabbed the day they are infected; typically there is an incubation period between infection and symptoms of about 5 days, and then a delay whilst arranging a test (shorter for walk-in and drive-through tests, longer for home testing kits which are posted out). Such tests get recorded as “Pillar 2” or “community” data. Another category of tests are “Pillar 1” tests conducted in hospitals on patients and health care workers. For hospitalised patients the delay between infection and swab test encompasses the incubation period and a period of worsening symptoms, typically another 5 days, before they reach hospital. We adapted mathematical methods to handle these delays and tried to estimate the timings of new infections as accurately as possible.
The graph below shows daily Pillar 1 case counts on the Isle of Wight in green, and our estimated number of new infections per day in red. We see that numbers of infections were decreasing from mid-April. There was a slight increase in the number of infections at the end of April which might be an artefact of the increased testing which occurred on 5th May and onwards when TT was launched. We then see a noticeable decrease in infections, despite the increased testing, shortly after 5th May.
Similarly, when we used the infection counts to estimate the reproduction number R - the average number of further infections each infected individual causes - we found that it decreased rapidly on the Isle of Wight following the TT launch. In the graph below the red line is our best estimate of R. The grey bands show possible values that R could have taken; they get wider towards the end when there were very few Pillar 1 cases which makes it difficult to estimate R accurately.
The next question we asked was how does the Isle of Wight compare to other areas? Did everywhere see a reduction in infections around the 5th of May, or after their TT launch? We looked at this question in various ways and each of our methods showed that something quite different happened on the Isle of Wight: it fared significantly better than other areas over comparable periods of time. For example, the graph below is a simplistic “nowcast”: we take average numbers of new infections and multiply by R to get a measure of the expected number of new hospital cases in the near future, assuming no changes in interventions. We present the results scaled per capita, so that more or less populous areas are easily comparable. We see that the Isle of Wight had quite a large epidemic in mid-April and was “middle of the pack” when compared with other areas until May, when it suddenly fared considerably better than other areas. By another measure (a Maximum-Likelihood method which worked straight from the case count data) we found that the Isle of Wight was ranked third worst out of the 150 areas of England in March-May, but ended up tenth best by the end of our analysis in June. Full details are available in our pre-print.
We conclude that:
The Isle of Wight saw a significant improvement in infection numbers and R after the TT launch
The reduction in infections on the Isle of Wight was significantly better than those of comparable areas in England and the wider UK
To prove that the TT programme was the cause of the success in the Isle of Wight would require us to rule out all other possible causes, and we have not done that in this early analysis - we would need more data. What we can say is that the Isle of Wight controlled their epidemic particularly well and that this warrants further investigation in case the causes might translate to other local and national control strategies. If it can be shown that TT had an impact, it will also be important to disentangle the aspects which had the greatest effect. Was it, for example, the huge advertising campaign and positive community spirit at the launch which meant people were more alert and followed guidance more than elsewhere, when lockdown measures were becoming ever more tiresome? How much impact did testing have: did people self-isolate more carefully when they had a positive test result as well as their symptoms? Was the contact tracing programme getting ahead of the virus, advising people to quarantine before they developed symptoms and infected others? And if so, was that primarily human contact tracing or via the app? A really important piece of data which would help us to determine that would be: of the people who tested positive, how many were already self-isolating because they had been contact-traced? We hope that more data will soon be made available so that we can better answer these questions.
In the meantime, working with the publicly available data, we have made our analyses available in this interactive EpiNow-C19 tool which updates daily. It allows you to zoom in (literally!) on our results and to view the epidemic trend in all 150 local authorities using a drop-down menu on the left. We hope to add more data and functionality very soon.
BLOG 6: Balancing functionality with privacy for effective contact tracing apps
8 May 2020
Contact tracing smartphone apps are a key component in many countries’ strategies for stopping the spread of coronavirus, exiting lockdown and returning to a new normal. The aim of contact tracing is straightforward: to quarantine only infectious individuals, allowing freedom of movement for everyone else and removing the need for a widespread lockdown. Performing contact tracing via a smartphone app allows the process to happen at a speed and scale capable of keeping up with the fast spread of COVID-19.
All contact tracing, both traditional and digital, involves some people sharing information about their recent interactions, and some people temporarily giving up their social freedoms according to the advice they receive. This can only be justified when the system works. At the same time, a voluntary digital system will only work if users feel confident to share their data and to trust its instructions. Functionality and privacy are therefore closely entwined and must be considered together in order to design an effective contact tracing app.
For a contact tracing app to be effective in its function of reducing the number of coronavirus infections it must satisfy key epidemiological requirements, which we describe in this supporting document. An app’s primary purpose is to quickly notify people that they may be infected so that they can take precautions before they infect others. Achieving this accurately requires minimising the number of notifications sent to people who are not infectious, as well as minimising the number of people who are infectious and not quarantined. This will be most effective with high user uptake and adherence to the app’s instructions, rapid notifications, integration with local health policy, and a transparent system for evaluation.
At the extremes, it would be possible to design an extremely accurate app with no privacy (which would probably have to be compulsory to achieve high user uptake) or a perfectly private app with low accuracy (which would quickly see users uninstalling it if they were frequently mistakenly quarantined and / or the disease was rapidly spreading). Both would fail on principles of ethics and epidemiological effectiveness. Treading the balance between these extremes are two main approaches known as the centralised and decentralised architectures. We present a detailed comparison of these approaches in this latest paper. Here we take a brief look at the main difference between them and provide examples of how this affects their function. Whilst noting that the details are evolving and subject to change, a current description of the main difference is as follows.
In the centralised architecture, the process of passing anonymised information from the sick individual to their contacts is mediated by a central server. When an individual self-reports symptoms or is diagnosed as a confirmed case, their contact history is uploaded to the server which performs risk assessment computations and sends notifications to some of their contacts accordingly.
In the decentralised architecture, the process of passing information from the sick individual to their contacts happens through direct broadcasts of anonymised lists of sick individuals over the phone network. Each phone then regularly performs a computation to determine whether it has been associated with a risky contact with one of the sick individuals.
The decentralised approach is widely, though not unanimously, thought to be better from a privacy perspective. However, we identify three important epidemiological benefits of the centralised approach. Firstly, it allows contacts to be alerted according to a calculation about their risk. This can take into account approximate measures of:
- how close the contact was (simply on the same bus, or neighbouring seats)
- how long the contact lasted (a brief encounter or a day spent in the same office)
- at what stage of the infection the contact occurred (many days ago when the sick person was unlikely to be infectious, or near the time that symptoms started when they were highly likely to be infectious). The only way currently known to use this in the decentralised approach reveals the timing of the onset of symptoms, undermining the privacy of the sick person.
This leads to greater accuracy than a decentralised approach which has to either issue blanket notifications to all such contacts, likely quarantining many people unnecessarily, or reduce notifications using a blunt threshold. An example of the latter might be “contacts within the last two days”, which would miss a full-day contact from three days ago but unnecessarily quarantine based on a brief passing contact yesterday.
It is important to note that the risk calculation concerns the likelihood of the person who came into contact with a sick person being infectious. Some approaches have only considered collecting data to estimate the risk of disease for the app user, rather than the risk of the app user infecting others. This misses the point of contact tracing, whether digital or manual. The reason to notify individual B that they may have been infected by individual A is to allow B to self-quarantine and reduce their potential for infecting others. It does not help B to avoid making contact with A because that contact has already taken place.
The second benefit of the centralised approach is that it enables evaluation and improvement. Digital contact tracing is a new tool devised for a rapidly evolving epidemic. As such it requires constant fine-tuning before, during and after launch to monitor its effectiveness, keep track of any unexpected behaviours, identify malicious use, and where necessary adjust it as the epidemic changes and algorithmic improvements are developed. All of this requires aggregated data to be collected and analysed (then deleted). This second benefit is related to the first in that it allows improvement to the algorithm to quarantine only those individuals who are most likely to be infectious. This creates a positive feedback loop of increasing user trust and confidence. One possible solution to this problem in the decentralised model is to ask many users to ‘donate’ their data on a regular basis. However, it is not yet clear whether this may result in more privacy concerns than a centralised approach.
The third benefit surrounds early warnings and early release from quarantine. Our modelling has shown that the delay involved in waiting for a positive test result seriously hampers the effectiveness of contact tracing apps. Although contact tracing could be performed more accurately if it were dependent on test results, in practice it will be ineffective if it is initiated after the 48-72 hours it currently takes for a person to receive their test result - by this time most of their infectious contacts will have occurred. Speed is therefore essential: for contact tracing to work at all it must find infected contacts before they infect others. A much faster and more effective system is to send out “amber” warnings to alert people that they have been in contact with someone with symptoms. With that knowledge, a contact may decide to take extra precautions such as working from home and not visiting a vulnerable relative, and they can make preparations to fully self-isolate soon if necessary. When the person with symptoms later receives a test result, the notification can be updated to a “red” self-isolation instruction if the test was positive or release the person if the test was negative.
For an app to effectively suppress the COVID-19 epidemic whilst maximising freedom of movement and maintaining privacy therefore requires a finely balanced design, and the optimal strategy may vary between countries. Wherever a design feature introduces a trade-off between privacy, functionality, uptake and adherence, the relative impacts of these factors should be quantitatively compared using mathematical models.
Modern public health practice already involves digital storage of identifiable information of individuals, where identifiers link information which people have consented to share based upon trust that systems are in place to protect their data, and because they understand their disclosure will benefit themselves and other people. It is possible that public opinion in some countries is focused more on the effectiveness of an app - whether it can help to end the epidemic, avoid lockdowns, and protect individuals and the people they meet - than whether the app is truly privacy-preserving.
Contact tracing apps are being developed across the world at incredible pace. In Oxford we’re committed to evaluating and optimising the epidemiological effectiveness of all contact tracing approaches: manual and digital, centralised and decentralised, with the common aim of stopping the COVID-19 pandemic. We strongly support communication and collaboration between the developers of all architectures so that all may continue to improve, and as-yet unanticipated solutions may be found which achieve an optimum balance between functionality and privacy.
by David Bonsall and Michelle Kendall
BLOG 5: Epidemiological model available open source to drive innovative strategies to stop coronavirus epidemic
30 April 2020
We’re sharing the code behind OpenABM-Covid19, our state-of-the-art agent-based epidemiological model for coronavirus. We developed the simulation to refine our understanding of the coronavirus epidemic and support the configuration of digital contact tracing designed to slow or even stop the spread of the virus. Now in the final stages of refinement, we’re releasing open-source epidemiological model code to provide an opportunity for modellers, scientists and policy makers to explore innovative strategies to help stop the epidemic.
We developed our simulations to model the epidemic dynamics of coronavirus and specifically the effect of interventions centred on digital contact tracing, testing and patient isolation. We’re now openly sharing our code and parameter configurations to provide the epidemiological tools to develop more novel strategies for controlling the epidemic. Our simulations of coronavirus demonstrate that public health measures can stop epidemic resurgence and avoid another lockdown: it’s in all our interest to stop coronavirus. Together we can save lives, resume work and social interactions and revive the economy by helping to suppress the epidemic.
We’ve simulated coronavirus in a model city of 1 million inhabitants with a wide range of realistic epidemiological configurations to explore options for controlling transmission. We’re keen to work with others to refine our simulation approaches and provide the best evidence possible to inform public health strategies.
Currently, the simulations are based on UK demographics, the model can be easily adjusted to simulate coronavirus epidemics in other countries. Other parameters such as the age structure of the severity of symptoms can be adjusted to model coronavirus under different scenarios, especially where there is still uncertainty in our knowledge of this epidemic. This flexibility will enable health, financial and other sectors to robustly consider strategies to respond to this rapidly evolving epidemic.
The simulation code is not the app development code, rather the code used to model the epidemiological scenarios that help to define and refine the configuration of a digital contact tracing approach.
Releasing the code open source will allow the NHSx and other countries to explore the configuration options, to optimise their digital contact tracing programmes for release, and to help evaluate dynamic policy strategies.
To prepare the model code to be made open source we were supported by Faculty and IBM UK. Our Oxford team developed and coded the model in C and used it to simulate the use of different configurations for a contact tracing app. IBM UK and Faculty worked alongside each other and built the Python interface for the model which enables it to be used more widely by partners, such as NHSx.
Dr Nicole Mather, IBM UK’s Services Life Sciences Lead, mobilised and led a team that worked with us at breakneck speed to develop the Python interface. The Python interface facilitates greater transparency into the output of the model and allows for dynamic policy interventions to be easily tested. The IBM UK team is currently working with the Oxford team on an optimisation engine for the model to compare the policy options and evaluate their effectiveness.
A team led by Dr Ilya Feige, Director of AI at Faculty also supported the development of the Python interface so that the model can easily be calibrated for different outbreaks. The Faculty team also prepared the repository for open sourcing, such as arranging the open source licence and helping to test the Python interface.
The Oxford team is currently focused on understanding the effect that different interventions can have on containing the epidemic. This will allow the potential users of the model, including the NHS, to make productive use of the tool when they need to plan for real-world scenarios.
We believe that a successful lockdown exit strategy will require a range of public health interventions, including a contact tracing app, ongoing physical distancing and traditional contact tracing. This offers us the chance to suppress the epidemic until we have vaccines and treatments widely available. By sharing our model open source we hope to encourage partnership and innovation amongst expert modellers, scientists and policy-makers to build-upon and innovate from this epidemiological foundation.
Find out more about our simulations, and download the epidemiological model code.
by Christophe Fraser, David Bonsall and Robert Hinch
BLOG 4: Simulating the use of digital contact tracing to control COVID-19
Update 16 April 2020
Previously, we showed in our Science paper how a digital contact tracing smartphone app could be an effective method for achieving COVID-19 epidemic control. We demonstrated that it can provide a practical, feasible and ethical method for scaling-up traditional contact tracing to match the fast pace of COVID-19 spread, while targeting quarantine measures to those at risk in order to stop the spread of the disease. We have now developed an individual-based epidemic simulation which enables epidemiologists, app designers and policy makers to compare a variety of algorithm configurations for digital contact tracing under a range of assumptions about the epidemic, the technology, a country’s demographics, and user engagement (see our report). We identified a safe and effective starting configuration for the UK that can be subsequently optimised to prevent transmission while minimising the number of people in isolation. Here we give a brief overview of the simulation and our key results.
The simulation considers a “model city” of 1 million people whose ages and contact patterns have been calibrated to UK demographics. All of our parameters are openly documented and modifiable so that they can be adapted to fit other countries’ data, and refined to match our understanding of COVID-19 as the epidemic progresses. The computer model simulates people moving around between their homes, workplaces, schools, and random social gatherings. It allows us to fast-forward through an epidemic to consider what happens when some of these people are infected. Crucially, it should not be thought of as a precise forecast but as a means of comparing the effectiveness of interventions. For example, we might find that interventions A and B result in similarly low numbers of infections but B involves quarantining fewer people.
The model should not be thought of as a precise forecast but as a means of comparing the effectiveness of interventions.
The spread of COVID-19 in the simulation is determined by a collection of parameters including the type of interaction (household interactions are more likely to result in spreading an infection than workplace interactions), the infectiousness of the transmitter, and the susceptibility of the recipient. An individual’s susceptibility, severity of infection and infectiousness are dependent on their age, as are their probabilities of being hospitalised, requiring critical care, and either dying or recovering.
A brief summary of current UK government advice for COVID-19 is 7 days of self-isolation for symptomatic individuals, 14 days for their household, and persisting isolation for individuals over the age of 70. We apply these instructions in our simulation, taking into account realistic levels of adherence and a daily dropout rate for people giving up on self-isolating. Against this backdrop we test the effectiveness of a national lockdown and of different algorithms for digital contact tracing being introduced at the end of the lockdown.
When we model app-based interventions we do not assume that everyone would use a digital contact tracing smartphone app. Although our user acceptance survey indicated broad support and uptake, we also tested the impact of low uptake. Using OFCOM data on UK smartphone ownership in each age group, we ran our simulation with varying uptake of 0-80% amongst those who actually own a suitable smartphone: uptake of 80% is shown by the blue bars in the plot below.
We consider six key variations on digital contact tracing which are summarised in the diagram below. These range from the scenario where there is no contact tracing app through scenarios with app implementations which vary in their levels of contact tracing applied, their “release mechanisms” which allow individuals to end their self-isolation earlier, and their use of testing, clinical diagnosis or self-reporting for the identification of cases. The full details can be found in our report.
Our results show that digital contact tracing can have a profound, life-saving impact on the progression of COVID-19. We quantify its impact on numbers of new infections, hospitalisations, ICU admissions and deaths, as well as the number of people in quarantine and the number of tests required each day. We consider six scenarios and test 20 variations on our key assumptions, plus varying levels of app-uptake - our full set of results can be found in our report. Here we’ll focus on describing the impacts of the six scenarios on numbers of new infections when using our best-estimate parameters and varying the percentage of app uptake.
The graphs below correspond to the five variations on app implementations. In each graph, the first vertical dashed line corresponds to the start of the UK lockdown and the second line is for the end of lockdown in our model, 35 days later. The height of the coloured lines shows the size of the peak of new infections if there are no further lockdowns, with different levels of app uptake (dark blue for 0%, yellow for 80%). This demonstrates the power of digital contact tracing in limiting the number of infections. We show that any use of digital contact tracing has an overall protective effect on the population, including those without smartphones, and that this effect is stronger the more people use the app. With high app uptake it is plausible that a second lockdown would not be required. This is based on our estimate that the first UK lockdown happened when around 1% of the population were infected: some of the interventions we model prevent infection levels getting that high again.
Our model is open-source so that everyone can use it. Policy-makers can calibrate it to their country’s demographics and decide which scenario produces the most tolerable trade-off between numbers of infected people and number quarantined, taking into account factors like the availability of tests, anticipated app uptake, and existing interventions. It is worth repeating that these results should not be thought of as a precise forecast but as a means of comparing the effectiveness and impact of interventions. Different scenarios may be suitable within the same country at different stages of the epidemic or as post-intervention data emerges or if the situation changes, for example if widespread testing becomes available. It is therefore essential to design digital contact tracing apps so that their underlying algorithms can be updated at any time to better respond to current epidemiological analysis.
The icons used in our diagram are taken from Font Awesome and licensed under the Creative Commons Attribution 4.0 International license, which can be found here.
BLOG 3: User Acceptance of Mobile Contact Tracing App
As outlined in 'Science', our mathematical analysis indicates that app-based contact tracing, with the general functionalities described in 'Mobile App', could control the epidemic if the app is used by enough smartphone users, and will continue to have an impact on transmission even at lower levels. We believe it will help contribute towards avoiding repeated lockdowns.
Given these potential benefits and the uncertainties around user acceptance, we conducted an online survey to assess the acceptability of app-based contact tracing in several countries back in March and April 2020, led by Johannes Abeler from the Department of Economics at Oxford University. These were the central questions:
- Overall support: Given the privacy concerns, would enough people install the app for it to be useful?
- Barriers to adoption: What are people’s main reasons for, and against, installing the app?
- Installation regime: How could voluntary installation vs. automatic installation by mobile phone providers affect app adoption?
In the UK survey, 1055 residents, selected to be representative of the UK population on gender, age, region of residence and employment status, responded to the survey between the 20th and 22nd March 2020. Read the full UK report, if you would like to see the surveys for other countries, click on the relevant report on the right panel.
Below are our key findings for the UK, we found similarly encouraging results in Italy, France, Germany and the US.
There is wide support for app-based contact tracing
Vast majority would comply with advice to self-isolate
More than 90% of respondents said they would definitely or probably comply with the recommendation of self-isolating for 14 days if they had been in close contact with a person infected. Among those who would not definitely comply, over 60% said they would be more likely to comply if they could be tested for the virus by the NHS within two days from the start of their self-isolation (a negative test allowing them to stop self-isolating).
Main reasons for and against installing the app
We asked respondents about their main reasons for and against installing the app. They could select up to 5 reasons from a menu of options or name their own reasons. Respondents are most concerned about “government using the app as an excuse to increase surveillance after the epidemic”. They are also concerned that the app would “make them feel more anxious” and that their “phone might get hacked”.
The main reason given for installing the app was to “protect family and friends”, but “responsibility to the community”, “might stop the epidemic” and “lets me know the risk of infection” are also often named.
Older respondents worry more about their phone being hacked and less about government surveillance. Otherwise, the reasons given do not appear to depend on age, gender, or willingness to install the app.
A majority supports automatic installation with an opt-out possibility
Respondents were asked to consider an automatic installation policy: the government would require mobile phone providers to automatically install the app on all phones, but users would be able to uninstall it at any time. We explained that this would maximise the chance of stopping the epidemic.
Approximately 72% of respondents said they would probably or definitely keep the app installed if it were automatically installed. Close to 70% agreed that the government should ask mobile phone providers to automatically install the app to maximise the chances of stopping the epidemic. This share rises slightly if we ask those who did not agree completely about their stance once someone in their community or someone they know has been infected. Most respondents said their opinion of the government would improve if such a policy were in place, independent of their political leaning.
In conclusion
Our findings were encouraging news during the early phase of the crisis and we have since seen strong uptake in Germany, Ireland and France. The survey suggests that people were willing to install the app and to comply with the self-isolation advice. Our results provided an early indication on the acceptability of app-based contact tracing. If the design of the app in any country and the messaging around its launch can alleviate fears about future surveillance and hacking, and generally reduce the anxiety around the epidemic, we would expect the number of app downloads to increase.
At the time of deploying these surveys, we could only ask hypothetical questions about future behaviour. The real decisions about installing the app are turning out differently in every context and country. Our survey does not address the legal and ethical implications of using such an app.
If you would like to learn more about this survey please find our publication here or read our country-specific reports for the UK, USA, Germany, France, and Italy.
by Ana Bulas Cruz
BLOG 2:Ethics and Social acceptance
30 March 2020
Scientific and epidemiological evidence suggest that a contact tracing app has the potential to contribute to reducing the suffering caused by the pandemic and minimise the harms caused by long periods of lockdown. These benefits and the avoidance of harms are clearly of great moral significance. If they are to be realised, however, several other ethical requirements need to be met.
The team's paper Ethics of instantaneous contact tracing using mobile phone apps in the control of the COVID-19 pandemic, published in The Journal of Medical Ethics, summarises how the successful and appropriate use of the app as an intervention relies on the ability to command well-founded public trust and confidence. This applies to the use of the app itself and of the data.
"There are well-founded public concerns on the implications of digital tracing and these have been included in our consideration and conceptualisation of the app's configuration since inception," explains Professor Christophe Fraser, co-lead on the mobile contact tracing app team at Oxford's Nuffield Department of Medicine.
There are well-established ethical arguments recognising the importance of achieving health benefits and avoiding harm. These arguments are particularly powerful in the context of a pandemic with the characteristics of COVID-19.
The contact tracing app-based system offers the possibility of both reducing the number of cases and enabling people to continue their lives in an informed, safe, and socially responsible way. It offers the potential to achieve important public benefits whilst maximising autonomy.
What are the requirements for the intervention to be ethical and command public trust?
- Oversight by an inclusive and transparent advisory board, which includes members of the public.
- Agreement and publication of ethical principles which will guide the intervention.
- Guarantees of equity of access and treatment.
- Use of a transparent and auditable algorithm.
- Embedding of evaluation and research to ensure learning for future outbreaks.
- Careful oversight of and effective protections around the uses of data.
- Knowledge sharing with other countries, especially low and middle-income countries.
It will be essential to ensure that the intervention involves the minimal imposition possible and that decisions in policy and practice are compatible with respect for three moral values: equal moral respect, fairness, and the importance of reducing suffering.
Prof Mike Parker, Director of the Wellcome Centre for Ethics & Humanities and Ethox's Centre, in Oxford's Nuffield Department of Population Health, who has led the ethics for this research, summarises his key recommendations:
"The appropriate use of a coronavirus mobile application will depend on building public trust. This requires high ethical standards throughout the intervention including: guaranteeing equal access and treatment; addressing privacy and data usage concerns; adopting a transparent and auditable algorithm; adapting smartphone deployment strategies to support specific groups, such as health care workers, the elderly and the young; and proceeding on the basis of individual consent."
Read: Parker M. et al. Ethics of instantaneous contact tracing using mobile phone apps in the control of the COVID-19 pandemic. Journal of Medical Ethics. Online May 6. 2020.
The Wellcome Centre for Ethics and Humanities is funded by a Wellcome Centre Grant (203132/Z/16/Z).
by Ana Bulas Cruz
Resources
Parker M. et al. Ethics of instantaneous contact tracing using mobile phone apps in the control of the COVID-19 pandemic. Journal of Medical Ethics. Online May 6. 2020.
Schaefer G and Ballantyne A. Downloading contact tracing apps is a moral obligation. Journal of Medical Ethics. May 4 2020.
Research in global health emergencies, report by Nuffield Council on Bioethics.
Data protection and coronavirus, Information Commissioner's Office.
Processing of personal data in the context of the COVID-19 outbreak, statement by the European Data Protection Board.
BLOG 1: Maths tells us the epidemic can be stopped
28 March 2020
There is one core mathematical truth in infectious disease epidemiology: If each infected individual infects on average less than one other person (reproductive number R<1), the epidemic declines. If each infected person infects on average more than one person (R>1), the epidemic grows exponentially. Interventions designed to stop the spread of the epidemic will therefore be successful if they achieve a reduction of R to below 1.
We re-estimated key parameters of the Covid-19 epidemic, built a mathematical model of transmission, and came to the conclusion that the epidemic can be stopped if isolation of infected individuals and quarantining of their contacts is sufficiently fast, sufficiently effective, and happens at scale. We published a paper in Science describing this work.
Manual contact tracing is a valuable component of test and trace strategies. However, when an epidemic is growing exponentially, this method alone rapidly stops being able to cope with the increase in the number of cases, lacking speed, efficiency and scalability. Our calculations show that effective tracing can still be achieved with the help of a contact tracing mobile app if used by a sufficiently large proportion of the population.
Using a mobile app to register if the user has been in close contact with other app users poses many ethical challenges which need to be resolved before an app can be used. We briefly address this in our Science paper and we discuss this more extensively in our ethics paper.
Below is a graphical overview of the Science study and the results.
