In April 2020, the Centre for Policy Research (CPR), began a collaboration with the Government of Punjab (GoP) to revise the State’s COVID-19 testing strategy. This is the first in a series of research notes based on findings of this collaboration.
Traditional disease surveillance techniques (and the one’s recommended by WHO for COVID 19) mandate surveillance based on symptoms. The challenge with COVID 19, however, is that there are a large number of patients who are asymptomatic and can infect people even though they do not have any recognisable symptoms. Addressing the problem of asymptomatic patients requires policy to identify additional models that enable the surveillance system to reach asymptomatic patients within the community and test those at risk of infection.
To make this case, this note draws on available data from the State of Punjab to better understand the role of asymptomatic patients in disease spread. Its arguments are three-fold:
- The likelihood of whether an individual will display symptoms dramatically affects the policy of who to test, and how cost-effective that testing is;
- Asymptomatic patients are critical to disease spread. Thus testing strategies will thus need to shift from only contact tracing infected cases to, in addition, active case finding through population surveillance;
- There are large potential benefits to research that improves the likelihood of finding a positive patient when they are tested (or predictive value of testing). This cannot be limited to those with symptoms because asymptomatic patients can and do spread infection.
A Simple Example
To illustrate the importance of asymptomatic carriers, we begin with a simple example. Consider a country where people are connected to one another, but some have more connections than others. A number of them are infected, and a proportion of those infected are asymptomatic. The country has a limited number of RT-PCR tests that can be deployed in one of two ways:
- Policy 1: only testing those who show symptoms (symptomatic testing).
- Policy 2: entering the community and testing those individuals who interact with the most people, e.g., shopkeepers (community testing). At first glance, this seems like a highly ineffective strategy because of the low likelihood of identifying positive cases with a RT-PCR test.
Under both of these policies, those that test positive are isolated, blocking further infections. Once this is done, the government counts the number of people who came into contact with the infected people before they were detected. This number is the ‘at risk’ population: the number of infections that could have been prevented had the infected people been tested on time.
Figure 1 reveals the testing problem. The country’s healthy individuals are represented by blue and infected in red. Some people who are infected show symptoms, while others will show some symptoms for other reasons, even if they’re not infected. Symptomatic individuals are denoted by “+”.
Presented in this way, it becomes clear that when there are more asymptomatic than symptomatic cases (more red circles than red +’s), and we test only those who show symptoms, we are less likely to identify positive cases.
We note that purely randomly testing in the population will be significantly less efficient in detecting infection if the likelihood of infection among the symptomatic population is much greater than the likelihood of infection among the asymptomatic population. But as symptoms become less reliable in signaling infection (that is, a small proportion of those displaying symptoms actually have the infection), the relative efficiency gains to symptomatic testing drops.
In this note we show that a simple adjustment to the random testing procedure may actually induce a more efficient protocol than symptomatic testing. In our “community testing strategy” we propose testing the individuals who interact with most people within the community, who we refer to as the most “central” people in the community network.
Figure 2 compares the results of our two candidate policies (symptomatic testing and community testing) under a simulation in which 20% of the population is tested. Here, we assume that about two-thirds of infected people show symptoms. In the top of Figure 2, we see that we have 151 total risky connections. That is, people who are connected directly to someone who is infected, whether the person shows symptoms or not. With symptomatic testing, people who are designated with at “+” are tested (whether they’re infected or not, since we don’t know that until they’re tested). Once a person tests positive they’re quarantined, eliminating their risky connections.
Those who interact with many others are simultaneously most susceptible to getting infected and may plausibly infect many others. For this reason, identifying and testing those with high levels of interaction can be effective in detecting infection in the population and isolation for those testing positive among this high interaction group can significantly decrease subsequent spread of infection. This is the basis of the ‘community testing’ strategy.
Among those infected, if a high proportion of those infected show symptoms (like in figure 2), we still expect symptomatic testing to be more efficient. But as this proportion drops (and we have increasingly many asymptomatic individuals among the infected), community-based testing should become more efficient. Indeed, there is evidence that the fraction of infected people showing symptoms could be quite low during the current outbreak.
Figure 3 shows the results of simulations in which we varied the probability of symptoms if infected (on the vertical axis), and the baseline infection i.e. the fraction of people infected within the population (on the vertical axis). Our metric for testing effectiveness was the difference in ‘at-risk’ individuals from community versus symptomatic testing — meaning that if fewer people are marked ‘at-risk’ using community testing, then this kind of testing is more effective.
The blue color in Figure 3 means that community testing is more effective, in that it eliminates more risky connections. Yellow/orange colors indicate that symptomatic testing is more effective. A first striking find is that symptomatic testing is only more effective in a narrow range of circumstances. For symptomatic testing to be effective we need
- the prevalence overall to be relatively low and
- the likelihood of being infected to be high if you show symptoms.
Without both of these characteristics, symptomatic testing is about the same or less effective than community based testing. The intuition behind the second benchmark was described above--if there are not many infected people among the symptomatic group, then it is an inefficient group to target. The first benchmark (low prevalence) is necessary because if the prevalence gets very high (above about one in five), the chances of finding an infected person go up dramatically regardless of testing strategy. Based on data from GOP, we estimate that 23% of detected cases report symptoms and 11% of the population tested is symptomatic. The proportion of symptomatic people in whom infection is detected is 3.7% while the same proportion for asymptomatic persons is 1.6%.
The horizontal line in Figure 3 represents an estimate for the fraction of infected who show symptoms from an article in the New England Journal of Medicine by Sutton et al. who tested a population of women in a labor and delivery ward. After testing all of the women in the ward, only about 13 percent of those who were infected showed symptoms. Other estimates have been presented from various settings, but they consistently are well below half.
Figure 3 Community versus Symptomatic Testing: Simulation Results
What is the likelihood of testing positive when people have symptoms versus when they do not?
Answering this question requires a head-to-head comparison of randomly selected asymptomatic individuals versus those who are tested after screening for common symptoms. As Punjab has not yet tried RT-PCR based tests among random samples in the community, this data is not available. However, the existing testing data clearly suggests that such an exercise can yield significant benefits.
In Figure 4, we have plotted the likelihood of showing symptoms for those who have tested positive/negative over time. The red curve shows the probability of someone with a positive test showing symptoms, while the black curve shows the probability of someone with a negative test showing symptoms. From mid-May onward, having symptoms becomes predictive of a positive test (as the red curve is higher than the black curve). However, it is still the case only about a quarter of those testing positive show symptoms -- meaning that there is approximately a 3-to-1 ratio of asymptomatic positive cases to symptomatic positive cases in this data. Those who test positive are, thus, significantly less likely to be symptomatic.
Figure 4: The Likelihood of Showing Symptoms Given the Test Result
Do asymptomatic people pass on the condition to others?
The short answer is yes, they do.
We analysed the contact data that was provided by Integrated Disease Surveillance Programme (IDSP), GoP. This data made it possible to reconstruct the route of infection across various generations. A detailed analysis of this data is presented in a separate note.
For this note, we linked the contact database provided by IDSP to the ICMR database provided to us. We were able to match 1989 out of 2183 observations in the contact database, i.e., a match of 91%. While it is important to examine why 195 observations could not be matched, we consider this to be quite good quality, given that the data generating processes are very different. This matching allows us to look at the categorisation of the people infecting and the people infected, as presented in Table 1 below.
In this database, there is no history of infection of 670 patients, of which 299 are related to a cluster of patients who returned from Nanded. From this group, 155 patients infect other people, and from the group of people they infect another 45 people infect other people, making a total of 200 patients who infect other people. The nature of symptoms on those infecting and those being infected is shown in Table 1.
We see here that, given that the conditions are recorded properly, asymptomatic patients account for 81% of all transmissions in Punjab, similar to the share of infecting patients (76%). Indeed, the average of 6.9 infections per infected patient is the same as the average for Punjab in this data and more than the average of 5.4 for symptomatic patients.
What are the implications of this data?
The dominance of infections from asymptomatic patients implies that testing strategies therefore need to shift from contact tracing of infected cases to active case-finding through population surveillance. This active case-finding is based on screening for the symptoms of COVID-19, but when there are many asymptomatic carriers, the model is incomplete.
Although India has not formally accepted that there is community transmission in many States, it is increasingly likely that this is the case. Testing strategies therefore need to shift from contact tracing of infected cases to active case-finding through population surveillance. In the current symptomatic testing model, this active case-finding is based on screening for the symptoms of COVID-19, but when there are many asymptomatic carriers, the model may fail badly; not only in terms of identifying those who are infected, but also in terms of budgetary requirements for testing.
To see this, note that the critical cost we are interested in is the cost per positive patient detected. If we only focus on the cost of the COVID-19 test and conservatively assume a lower cost to the government of Rs.4000, the positivity rate in Punjab of 3% implies that 33 samples are being tested in order to detect a single positive case. Therefore, the cost per positive patient detected runs at Rs.1.33 lakh. Even if the cost of the test is brought down to half this sum, it will still remain at Rs.65,000. Since patients will have to be tested multiple times throughout the months of the epidemic, these budgetary implications can quickly become onerous.
However, the cost per patient detected is not the only cost of testing. Once community testing is required, we must also factor in the cost of visiting a person in the community to obtain the sample and transport it to the testing center. To see how this affects the testing strategy, assume each sample costs Rs.T to acquire. With random sampling, if p% of the population is positive, for every 100 patients tested, approximately p will be positive, so the cost per positive is (100/p)*(4000 + T).
What about symptomatic screening? Under symptomatic screening, households are first visited and a test is administered only if a person is symptomatic, so the cost implications now depend both on what fraction of households report infections (the fewer the fraction, the higher the cost of acquiring samples) and the likelihood that people are positive conditional on reporting symptoms (the higher the likelihood, the lower the testing cost per positive patient detected). If, for example, out of 1000 people visited, only 50 report symptoms and out of the 50, 10 are actually infected, this implies that the total cost will be 50*4000 + 1000T and the cost per positive will be 20,000 + 100T.
From large-scale screening, it appears that sample acquisition costs of symptomatic surveillance are going to be very high (only a very small fraction of households report SARI or ILI symptoms) and that symptoms are not very predictive of infection (test positivity of symptomatic individuals has only recently eclipsed that of asymptomatic individuals and only marginally so). All of these would imply that for detecting positive cases in the community, symptomatic testing may be extremely expensive.
Note that this is not an argument for or against community testing. It is, however, an argument for an active agenda that helps us discover better markers of the infection among the population in Punjab. With current test positivity rates of 3%, finding better markers that can push up the positivity rate even to 5% can dramatically reduce the cost per positive patient detected. Furthermore, early detection can allow disease transmission to be effectively halted, keeping the infection rates low throughout the state.
Postscript: Since June 2020, Punjab has identified a set of high-risk populations that are being tested as part of their overall testing strategy. This is now being refined, in collaboration with CPR, to design a risk stratifed sampling strategy based on an effort to identify better test predictors.
Other research notes as part of the series can be accessed below: