Abstract
Some nations have the option of pursuing a policy of complete elimination of covid-19 instead of a policy of “flatttening the curve” so that the load the disease places on medical facilities is bearable. A policy of elimination requires a rather onerous “lockdown”. As the goal of elimination is approached, it is therefore important that there be an informed trade-off between the risk of disease re-emergence and the duration of the “lockdown”. Here it is shown that an important factor in assessing this trade-off, is the distribution of secondary cases, not just the expected value of this distribution, R0. It is shown that a distribution in which “super-spreaders” are more prominent in the epidemiology, allows for an earlier release from “lockdown” with reasonable safety despite some probability of asymptomatic cases. There is some evidence to support such a distribution for covid-19. Analytical calculations and simulations show that once there is only one recognised case in some subregion, release from “lockdown” will be reasonably safe after just one or two further incubation periods.
Introduction
The experience of China indicates that covid-19 can be eliminated. Elimination requires lowering the reproductive rate of the infection by “lockdown” measures until there are very few active cases and then waiting for several incubation periods without new cases before ending the “lockdown”. However, it will always remain possible that, because of asymptomatic cases, the infection will re-emerge after lifting of “lockdown”. This paper shows that the risk of such re-emergence depends quite markedly on the distribution of the number of secondary cases, not just its average value, R0.
The importance of the distribution of secondary cases, is illustrated by several contrived probability distributions for the policy where restrictions are lifted once there is just a single symptomatic case. More realistic probability distributions and simulation are used to estimate the probability of re-emergence with three different policies - lifting restrictions when there is just one symptomatic patient, waiting one incubation period after this time before lifting restrictions and waiting two incubation periods before lifting restrictions.
This paper assumes that an attempt has been made to eliminate covid-19 from a region/province/nation. It assumes no appreciable herd immunity. It also assumes that there is no quarantine free cross border human travel into the region considered and this restriction will be maintained. It is assumed that the attempt has been successful in that numbers of symptomatic covid cases have dwindled to a single case which is then isolated.
However, it is known that a proportion of cases are asymptomatic. Here this proportion pa, is taken to be 50%. It is further assumed that the symptomatic cases up to their time of isolation and the asymptomatic cases, both have the same distribution for the number of secondary infections they generate. It is assumed that all symptomatic covid cases are recognised by testing but there is no testing of asymptomatic people. It is also assumed that R0 will resume its pre-covid value when restrictions are lifted. For the purposes of this paper, the average number of secondary cases R0, in an unrestricted population is assumed to be 2.5.
Methods
It is known from the theory of branching processes in discrete time, that the probability of extinction of a branching process is given by the (smallest) solution to the equation ϕ(u) = u where ϕ(x) is the probability generating function for the number of secondary cases of a single index case [1]. For example, if there is a probability of 0.99 that an index case has no secondary cases and probability 0.01 that the case has 250 secondary cases - giving R0 = 2.5, then ϕ(x) = 0.99x0 + 0.01x250. The solution to the equation ϕ(u) = u is found by Newton’s method. It is denoted by pe ≡ pextinction. The probability of extinction is calculated similarly for a number of other contrived distributions for the number of secondary cases generated by a single infected person, all with R0 = 2.5. The coefficient of variation is calculated for each of these distributions. The probability of extinction for two standard probability distributions are also calculated. The calculations here assume that once a closed region is down to a single symptomatic case, there may or may not be secondary cases from this case prior to their isolation and there may also be secondary cases from asymptomatic cases present at the time of isolation of this last symptomatic case. However, even where there are such secondary cases, the branching processes from however many secondary cases there are, may go extinct.
We assume as an uninformative Bayesian prior, that the number of asymptomatic cases at the time when there is only 1 symptomatic case, is given by a geometric distribution, so that the probability that there are k asymtomatic cases is . The probability that none of these cases give rise to an ongoing chain of secondary, tertiary, …, cases is denoted by pT 0, the probability of Total elimination. It is found by summing the probability that there are k asymtomatic
for all k from 0 to ∞. This gives
Two less contrived, standard, distributions are considered further. These are the Poisson distribution corresponding to an unstructured randomness where any infected person is equally likely as any other to seed many secondary cases. The second distribution is the negative binomial distribution. There is now some evidence from observation of the initiation of the epidemic in various countries, that the initial seeding cases gave rise to offspring cases in a pattern suggestive of a negative binomial distribution with parameters specified below [2]. This is a distribution with a thin but very long right tail, indicating an important role for super-spreaders. A distribution for the number of secondary cases where very occasional superspreaders are prominent in the epidemic, is also suggested by variation in the time to community transmission in various nations, despite likely repeated introduction by travellers. If most infected travellers seldom spread the disease, the onset of community transmission may be delayed until the rather rare super-spreader arrives.
For the Poisson distribution and this Negative Binomial distribution, two further policy scenarios are considered - continuing restrictions for 1 incubation period after the last symptomatic person has been isolated, and continuing retrictions for 2 incubation periods. It is difficult to obtain an analytic equation for the probability of extinction and so simulation is used. Each simulation employs 10 million runs. A fortran program corresponding to these simulations is given as an attachment.
Results
The probability distributions considered, all give an expected number of secondary cases, R0 = 2.5 They are described below, with P (X = k) representing the probability of k secondary cases :-
P (X = 0) = 0.99 and P (X = 250) = 0.01 as in the section above
P (X = 0) = 0.9 and P (X = 25) = 0.1
P (X = 0) = 0.5 and P (X = 5) = 0.5
and
and
(as close as possible to deterministic R0 = 2.5)
P (X = j) is Poisson with parameter λ = 2.5
with
and r = 0.1 This is a Negative Binomial. Its parameters are inferred from details in reference and correspond to R0 = 2.5 with dispersion measure r = 0.1
The table below gives the probability, pe, that a single branching process with the given numbered distribution becomes extinct in that it does not result in an indefinitely large number of descendants. A column giving the coefficient of variation is given for each disribution. Also given is pT 0, that is the probability that overall there will be no re-emergence of the epidemic, despite releasing restrictions without waiting an incubation period to check for symptomatic cases to arise either from previous unrecognised asymptomatic cases or from the symptomatic case before isolation.
The following table considers distributions 7) and 8) in more detail. As well as the policy of releasing restrictions zero incubation periods after the last symptomatic case is isolated, simulations are used to calculate the probability of no epidemic re-emergence if there is a policy of releasing restrictions after 1 incubation period and after 2 incubation periods. These probabilities are denoted by pT 1 and pT 2 respectively. For comparison, information given in the previous table on the probability of extinction if restrictions are lifted on isolation of the last symptomatic case, is repeated.
Unfortunately, in implementing policies whose outcome are denoted pT 1 and pT 2, there will be occasions where relief from “lockdown” will be further delayed by re-emergence of symptomatic cases before the one or two incubation periods are completed. In the case of the Poisson disstribution labelled 7), this happens 50% and 58% of the time. In the case of the negative binomial distribution labelled 8), this happens 27% and 28% of the time respectively.
Discussion
A policy of elimination requires a more intense “lockdown” than a policy of “flattening the curve” so that the epidemic peak is within the capacity of a region’s ICU beds. However, a study by the Imperial College [3] shows that there are two rather fine lines between flattening the curve sufficiently and on the one hand, elimination, or on the other hand, an insufficiently flattened epidemic and a huge death rate. Even if “optimal” curve flattening can be achieved there will still be many deaths. In addition, whilst slightly less intense lockdown is required for curve flattening rather than elimination, the period of “lockdown” for curve flattening, will have to be continued for far longer. An advantage of the curve flattening approach is that curve flattening could lead to herd immunity and thus allow less restricted international travel to parts of the world that have not achieved elimination, though this prospect is uncertain. On the other hand, international travel without quarantine will not be possible between a country that has achieved elimination and a country which has not, at least until a vaccine is available. For this author, the arguments in favour of a policy of elimination are ovewhelming, for nations/regions which have the resources to implement it and who have secure borders. New Zealand is following a policy of elimination and there seems to be emerging realisation that this is a real option for Australia. Since the intensity of “lockdown” required for elimination will be quite onerous, it is important that the length of “lockdown” required is balanced against the probability that this lockdown period will be successful in eliminating the virus.
It is clear from the results, that simple knowledge of R0 = 2.5 is not sufficient for calculations that will enable decision makers to balance risk of re-emergence of the epidemic against length of lockdown. In general, it can be seen that distributions with a large coefficient of variation, that is, distributions which in this context give a greater role to occasional super-spreaders, are those where elimination can be achieved more quickly after the last symptomatic case is diagnosed. The negative binomial distribution used here gives a large role in the epidemic for occasional superspreaders, in contrast to the Poisson distribution. There is also some evidence for the negative binomial distribution with the parameters given. Commonsense would also suggest that people are unlikely to be uniform in their propensity to spread the virus, either because socially they are a more central node in a large network, or perhaps because some may have more biological propensity to infect others. However, even with the less likely, more pessimistic choice, the Poisson distribution, the calculations show that there is better than 89% chance of elimination if two incubation periods pass without re-emergence of symptomatic cases. With the Negative Binomial distribution, elimination would then be almost certain. It would seem that, when numbers of symptomatic cases have been reduced to rather low levels, lockdown periods, could be decreased further by subdividing a region. If a region implementing an elimination strategy can be subdivided into subregions which can be temporarily disconnected from each other in terms of human travel, such that each subregion has only one case, relief from restrictions can commence quite quickly.
There are some clear limitations to this study. No sensitivity analysis has been done beyond calculations for the specific distributions described. In particular, there is no sensitivity analysis for likely different values for the proportion asymptomatic pa, or R0 under lockdown. However, the fortran programs below could facilitate such a sensitivity analysis if there was more knowledge about the range of plausible values. The calculations and simulations also assume discrete time steps and a different analysis will be required to account for branching in continuous time. There are some assumptions that are too pessimistic, for example, some asymptomatic cases will be detected by contact tracing. There is also implicitly some assumptions that may be too optimistic. For example, it is assumed that there is no animal or human long term carrier of the virus in the region under consideration. Nevertheless this study gives cause for optimism about the length of lockdown required for any region attempting to eliminate this disease.
Data Availability
There is no relevant data as such - just computer programs developed for this project. The main programs are included as an attachment at the end of the paper. Programs which are near duplicates of each other or that are almost trivial are omitted, but the whole suit of programs used here are available from the author.
Attachments
Fortran program using Newton’s method to calculate the probability of extinction of the branching process for various probability distributions
The fortran program below calculates by simulation, the probability of elimination when there are two incubation periods in which, under “lockdown” conditions, number of secondary cases are determined by a negative binomial distribution, with R0 reduced from 2.5 to 0.8, but with the original coefficient of variation unchanged. This is achieved by changing r from 0.1 to and changing p from
to
. Since this negative binomial is used repeatedly in the simulation, it is useful to calculate its cumulative distribution using a separate program. Because of numerical instability the distribution is truncated at 170 secondary cases with some arbitrary provision for the the tiny but not entirely negligible tail beyond 170. It will be clear to fortran programmers how the program can be simplified to deal with a policy of lifting lockdown after one instead of two incubation periods. It will similarly be clear how to modify it for a Poisson distribution with λ = 0.8 during the lockdown period and with pe changed so that it is the probability of extinction for a Poisson branching process given in the first table. A few other very short, almost trivial, programs are also used in this project. They are not copied here but are available on request to the author.