Source: United Kingdom – Executive Government & Departments
A preprint, an unpublished non-peer reviewed study, reports on the latest data from the REACT-1 study on COVID-19 spread across England.
This Roundup accompanied an SMC Briefing.
Prof Kevin McConway, Emeritus Professor of Applied Statistics, The Open University, said:
“The latest preprint from the REACT-1 survey, run by Imperial College working with Ipsos MORI, covers the whole of round 6 of that survey. As in previous rounds, a reasonably representative sample of the population of England was chosen and asked to provide swabs that are tested to detect the presence of the virus that causes Covid-19. Because these people are tested only to determine the pattern of infection across the country, because they are tested regardless of whether they are showing any symptoms, and because the people tested are representative of the whole population, this can produce a better estimate of how many people in the country are infected than can be obtained from the daily numbers of new confirmed cases of Covid-19. Those daily numbers depend on which people are being tested, how easy it is to get a test, on the numbers of tests being performed, and so on, and could change if any of those factors change, even if the underlying number of infected people does not change. We’re lucky enough in England to have two such infection surveys being carried out regularly, REACT-1 and the ONS Infection Survey. I don’t know of any other country that has even one such survey running on a regular basis.
“An interim report on Round 6 of REACT-1 was published two weeks ago (29 October), based on the results from the first half of the sampling for this round, using swabs taken between 16 and 25 October. This new report covers the whole of round 6, so adds information from swabs collected between 26 October and 2 November. The previous interim report said that the rate of infection appeared to be increasing fast. I did express some doubts at the time about the reliability of that conclusion, which was largely based on data from just 8 days of tests. Does the new report throw any more light on that aspect, and generally how does it compare?
“The result that is most straightforward, from this round of REACT-1 and indeed all the previous rounds, is the estimate of the percentage of people in England who would test positive for the virus (if the whole country could be tested). For the whole of round 6, the estimate is that 1.3% of the population would test positive for the virus, on average, on each day of the round – that is, 13 people in every thousand. Because this estimate is based on a sample, there’s some statistical uncertainty, though not all that much uncertainty because the sample of people tested is very large. The true percentage could plausibly be somewhere between about 1.2% and about 1.4% (that is, about 12 and about 14 per thousand people). This estimated percentage is considerably higher than for round 5 of REACT-1 (carried out between 18 September and 5 October), when the estimate was less than half as big, at 6 people in every thousand. And, as in the interim report on round 6, this increase isn’t confined just to some parts of the country or some groups of people. The report also gives estimates of the prevalence of infection for more than 40 groups within the population, defined by gender, age, the regions where they live, their type of employment, ethnicity, household size, and more. In all but one of these groups, the infection rate went up between round 5 and round 6. (The one group where the rate fell rather than increasing was in the largest households, with 7 or more people, but there are few people in the REACT-1 samples that live in households that big, so that their infection rate could not be estimated very accurately.) There are differences between the groups on how much it went up. Notably, it increased proportionally less in people aged 18-24 than in other age groups (where the infection rate was particularly high in round 5), and in the North East compared to other English regions – but it still has increased in these groups too. None of this is good news. Prevalence of infection remains higher in most of the North than in the rest of England, but it seems to be increasing at a slower rate in some of the North, and in London, than elsewhere. That is slightly encouraging, but it doesn’t so clearly apply right across the North.
“Several of the details of these patterns are different from those based just on the first part of round 6 (as reported in the interim report). The researchers consider that this might well be because of changes in the pattern of results between the first and second half of round 6, and indeed they report some evidence of changes within the two halves, particularly a reduction in the infection rate in Yorkshire & the Humber. They report that there seems to have been something of a dip in the number of infections, across the country, just during the second half of the round, with the lowest point around 30 October. They mention some possible reasons that might lie behind this pattern, but say they can’t be sure of the reasons, and I’m not sure either. They suggest that it might have something to do with school half-term, or to the poor weather in October, or to behaviour changes from the tier system of restrictions or from people becoming more cautious as a result of the publication of data showing increasing infection rates and speculation about a possible ‘circuit-breaker’. But it could also be something to do with anomalies in the sampling process, with people tested on different days during round 6 not being sufficiently similar to one another. The researchers say that they did check for some differences like that, but feel that they did not make enough impact to be important – but, given that the number of people tested on any one day is not so high, I’m not sure how accurate these estimates of very short-term changes in infection rates can be.
“The numbers I have given so far are estimates of the percentage of people who would test positive. The researchers also make a calculation that allows for the estimated number of false negatives (people who are actually infected but have a negative test result, perhaps because of a failure to take the swab correctly), and they estimate the number of infected people in the English population as around 1 million. They also estimate that there could have been roughly 100,000 new infections a day, at around the end of October.
“REACT-1 also give estimates of the R number and the growth rate in new infections. Just based on the first part of round 6, these estimates in their interim report were alarmingly high (and higher than most, if not all, other estimates covering roughly the same period). In this new report, they do not give estimates for the whole of round 6 – they argue that this would be inappropriate because the pattern of change in numbers of infected people are too different in the two halves of the round. I think that’s wise – and I’m not sure how appropriate are the R estimates they give for the two halves of round 6 separately, given that each is based on the changes of infections, in the survey, over just a few days. But, for what it’s worth, their overall estimate for the second half of the round is quite encouraging, with a range for R from about 0.7 to about 1.0, and an estimate that the change in the number of infected people each day is between zero change and a decrease of about 2% a day.
“Since there is another infection survey running in England, the ONS infection survey, we can look at how the results of the two surveys compare. The latest main results of the ONS survey, published last Friday, relate to the period 25-31 October. That’s quite similar to the period covered by the second half of round 6 of REACT-1 (26 October-2 November). The main results do differ quite substantially. REACT-1 estimate the percentage testing positive as 1.32%, with an interval (showing the statistical uncertainty) from 1.20% to 1.45%. The ONS estimate is 1.13%, with an interval from 1.07% to 1.20%. The intervals don’t quite even overlap. I have no clear evidence-based reason for saying that one estimate is more likely to be accurate than the other. There are differences in the lists that are used by the two surveys to obtain the people that they invite to join their studies. ONS use a standard list of addresses, and REACT-1 uses lists of people registered with GP practices. Neither list is perfect – there’s no such thing as a perfect list of people or households in the country, and in any case not every one who is invited will agree to take part, by any means. Both surveys make adjustments to their results to allow for the fact that the people who do take part will not be perfectly representative of the whole population. But the results do still differ, and there isn’t a more accurate estimate to compare them with. I think we just have to bear in mind that there may be rather more uncertainty in the estimates than is fully described by the margins of error that the reports give. What is somewhat encouraging, though, is that both surveys give at least an indication that the very rapid rates of increase of infection, that we saw in early October, may well be slackening off a bit.
“In broad terms, the regional pattern of infection rates from the whole of REACT-1 Round 6 does roughly match the regional pattern in the latest ONS survey results. That’s still roughly the case if we just look at the REACT-1 results for the second half of round 6, where the dates are closer to those for the most recent ONS results, but there are some differences between the two surveys in the regions in the North.
“The two surveys differ much more, though, in their estimates of the daily rate of new infections. REACT-1 estimates somewhere between 90,000 and 104,000 new infections a day, around the end of October. ONS estimated between 37,700 and 59,600 new infections, or more precisely people having a positive test result for the first time, each day between 25 and 31 October, and more generally they wrote in last week’s report that “incidence appears to have stabilised at around 50,000 new infections per day”, or roughly half the REACT-1 estimate. This is only partly explained by the higher prevalence of people testing positive in the REACT-1 study. But the ONS survey can directly estimate the rate of new infections, because it tests people more than once, and so knows within a few days when someone in their survey is newly infected. REACT-1 can’t do that because it tests people just once, and instead they get their estimates by assuming that people remain infected for 10 days on average, so they simply divide their estimate for the total number of infected people by 10 to get the daily rate of new infections. I think that’s a bit over-simple, and on this I’d prefer the ONS estimates.
“Either way, though, even if it is indeed about 50,000 new infections each day rather than about 100,000, that is a lot. The REACT-1 survey points out, correctly of course, that their results can’t tell us anything at all about the impact of the second national lockdown in England on infection rates, because their sampling for this round finished before the lockdown started on 5 November. The next set of ONS infection survey figures, due tomorrow, will cover the very first few days of the second lockdown, but that’s probably not enough to show a clear effect (if there is one). We’ll have to wait longer before patterns become clear – but with total amounts of infection continuing to increase according to both surveys, we’ll also have to hope that something will cause those increases to be reversed.”
Prof Sheila Bird, Formerly Programme Leader, MRC Biostatistics Unit, University of Cambridge, said:
“REACT-1 has released its report on the completion of its 6th round of swab-testing (75% sensitivity assumed). In total, 86,000 persons aged 5+ years participated during 16th to 25th October 2020 (round 6a) and 74,000 during 26th October to 2nd November (round 6b), prior to England’s second lock-down which came into effect on 4th November 2020.
“Measured message: Round 6’s overall weighted prevalence of SARS-CoV-2 infection in the community in England was 1.30% (1.21% to 1.39%), or 13.0 per 1000 during 16 October to 2 November: up from 0.6% (0.55% to 0.71%), or 6 per 1000, during the previous round (18 September to 5 October). However, the weighted prevalence was substantially the same in rounds 6a and 6b, which is a measure of good news.
“During the 16 to 25th October, R was estimated at 1.6 (1.3 to 1.9) with swab-test-positive infections doubling every 9 days (95% uncertainty: 18 days to 6 days); during 26th October to 2 November, R was estimated at 0.85 (0.7 to 1.0).
“Intriguing verdict: Round 6 reports some additional temporal structure, the strength and consistency of which was checked rather thoroughly and persisted, namely: during round 6b, a fall then rise in prevalence was observed with the lowest point around October 30th. Having driven south from Scotland on Sunday 1st November (ahead of lockdown in England), should I look to my laurels? I’ll plea for the Scottish verdict of “not proven”.
“Reason for weighting: The most deprived two quintiles contributed only 25% (40,690, nearest 10) of REACT-1’s 160,175 participants – not the 40% that should be expected; the middle quintile accounted for 34,380 (21%); while the least deprived two quintiles accounted for 85,100 participants (53% versus 40% expected). Re-weighting does not necessarily correct for systematic bias within-deprivation-quintile between the quintile’s participants and non-responders.
“Re-weighting did, however, increase the estimated prevalence in the most deprived quintile from its unweighted estimate of 1.6% (uncertainty: 1.4% to 1.8%) to 2.1% (uncertainty: 1.7% to 2.5%) versus around 1% in the middle and least deprived pair
“COVID case contact: Between Round 5 and Round 6, weighted prevalence doubled from 0.5% to 1% (.09% to 1.1%) for persons who had no known contact with a COVID-case; remained around 9% for those in recent contact with a confirmed COVID case (such as, for example, members of the household of a confirmed cases whom Test & Trace asks to self-isolate); and more than doubled from 1.1% (uncertainty: 0.6% to 1.8%) to 3.6% (uncertainty: 2.3% to 5.5%) for those in contact with a suspected COVID case. The latter is probably explained by “suspect” being now more likely to herald imminent confirmation that was the case in Round 5.
“Asymptomatic infections: Across all rounds of REACT-1, at least half the people with detectable virus: will not report symptoms on the day of testing or in the week prior. The strong recommendation on 23 July 2020 by the Royal Statistical Society’s COVID-19 Taskforce was that persons whom Test, Trace and Isolate asks to self-isolate should be offered swab-testing on a random pair of days during their quarantine to learn efficiently about asymptomatic infections. Shamefully – in terms of infection control – Test, Trace and Isolate’s high-risk cohorts have been offered no such swab-testing. Instead, resources have directed towards mass screening, rather than testing where infections, including asymptomatic, are most likely to be located.”
Preprint (not a paper): ‘REACT-1 round 6 updated report: high prevalence of SARS-CoV-2 swab positivity with reduced rate of growth in England at the start of November 2020’ by Steven Riley et al. was posted on Thursday 12 November 2020. This work is not peer-reviewed.
All our previous output on this subject can be seen at this weblink:
Prof Kevin McConway: “I am a Trustee of the SMC and a member of the Advisory Committee, but my quote above is in my capacity as a professional statistician.”
Prof Sheila Bird: “SMB is a member of the Royal Statistical Society’s COVID-19 Taskforce.”