Any epidemiological model that make projections about the coronavirus pandemic requires mountains of data, especially if it has to account for human behaviour. Especially in India, these models can go very wrong because we just don’t have enough locally relevant information to feed the machine. This uncertainty requires a fundamentally different approach: It does not mean that decisions cannot be made, but it does imply that decision makers need to account for our lack of certitude.
So far, the fact that we know little about the virus and how it interacts within India’s complex economic and social mileu, has played only a small role in our decision making. The poor, who live day-to-day and have no option but to depend on public toilets and supplied water have been told to save themselves. We understand it’s your money and your life, sorry. The justification seems ironclad: Humare pas option hi kya hai (what option do we have?).
After all, leading epidemiologists and public-health experts have affirmed through complicated agent-based models that if we don’t impose a lockdown, India will have 300 million infections. Our hospitals will be overwhelmed and millions will die. But how certain are they really? And how should we account for the potential uncertainty in their estimates?
Complex models need complex data
In the classic epidemiological, or epi model, sick people infect others; some recover, others die; those who recover can’t be infected again. The epidemic rises when there are many people who have not yet fallen sick and then falls as more people gain immunity after falling sick.
Social-distancing spreads out the infection and reduces the number of infected at any given time. It also decreases the total number of people who eventually become infected through the course of the epidemic. This basic insight, derived more than 100 years ago, has withstood both the test of time and the use of increasingly complex dynamics.
But increasing the complexity of models comes at the cost of far greater data requirements. Simpler models assume that people interact randomly with each other, and each person has the same number of interactions. That is a far cry from reality, which is why the models in Europe increasingly rely on data they have gathered from contact diary studies that ask people to list all the contacts they have had in the last 24 hours.
These studies allow modelers to account for the distribution of contacts among each age-group (and with different ages). India does not have any such contact survey, and the best we can rely on is extrapolations from Europe and apply them to the age and household structure in India.
This matters. In models where people have a different number of contacts and a different position in the social network, epidemics can die out, exhibit chaotic behaviour and other dynamic patterns far beyond that of a simpler model. This is not just a theoretical curiosity. The number of contacts that people have range widely and we have found from a quick survey that is currently ongoing that even among people with the same occupation, the number of contacts range from 3 to 4,000 per person.
Current models in India either do not account for the contact distribution, or end up assuming that this distribution can be extrapolated from Western countries, which are markedly different in their social and living arrangements. Extrapolations such as these belie the advertised claim to certainty.
No accounting for human behaviour
Equally worrying is the fact that first-gen epi models predict human outcomes without accounting for human behaviour. This deep problem led to the creation of a new field in the 1990s that tried to incorporate human behaviour into epidemiological models.
In these models, humans maximise their well-being, taking into account the likelihood of infection in their behaviour. How this affects the spread of the epidemic depends on their beliefs. Humans may be fatalistic (I am going to get it anyway, so why bother), altruistic (if I go out, I may infect my family) or selfish (I can stay home while others work, and soon we will have herd immunity). Again, multiple dynamics can emerge. However, insufficient data on how humans behave during epidemics has precluded the emergence of a ‘standard model’ that convincingly marries epidemiology and economics.
Current epi models assume that if a pedestrian is crossing a street and a car is also passing at the same time, they are destined to collide. It will not consider scenarios where the driver would apply the brakes or changes the direction of the vehicle to avoid collision, or the pedestrian stops. Instead of taking seriously the idea that the pedestrian does not want to be hit by the car and the car does not want to hit the pedestrian, our response is to ensure that no cars and no pedestrians are allowed on the streets at all.
Does lack of certitude imply that we should not act?
For many problems, we don’t have a good solution. Even the simple problem of how should we test people to determine if Covid-19 is spreading remains unsolved. A positive test for a healthcare worker who sees thousands of patients a week has very different implications for determining population prevalence, or the fraction of the population that is currently infected, compared to a positive test for a person who spends most of his/her day at home. Worse, what positive tests imply for population prevalence among these two types of people will change from day to day as the epidemic spreads.
This pandemic of ignorance is not an excuse for policy paralysis. Instead, there is a well-developed framework to make policy decisions under uncertainty, and we should use it.
What policymakers can do
First, an exponential epidemic needs an exponential method of crowdsourcing expertise. India is fortunate with thousands of Indian researchers in top universities throughout the world who are willing and ready to help. Thus far, there is no public document from the government on what it plans to do once the lockdown is lifted.
There is no public data that these people can work with, and there is no structure for them to interact closely with the states. The first priority should be to ensure that we can bring in the expertise we so desperately need—and already have access to.
Second, some actions are beneficial regardless of how the epidemic plays out. There will be outbreaks and these outbreaks will surge. Ensuring that testing is widely available and free through our extensive networks of private and public labs is a prudent strategy. So is training the large number of semi-trained providers who form the bulk of our primary healthcare force to use personal protective equipment (PPE) and triage patients to appropriate care.
Third, some decisions should become better as information increases. So, every decision has to take into account not only the population impact, but also provide valuable information on the testing strategy itself. For instance, a well-designed testing strategy can tell us not only whether someone is infected, but also help us infer the structure of social networks. By choosing who to test carefully today, we can generate better information that allows us to design even better testing strategies tomorrow.
Fourth, the payoff from some decisions can increase dramatically with just a bit more information—and in these cases it is better to gather that information before acting. The cruelty of the lockdown for the poor was matched only by the ineptitude of its execution.
A simple survey that would have taken two days would have allowed us to learn that the millions of migrants who power our cities would leave in case of a lockdown. We could have put in place multiple measures to mitigate the impact of this massive exodus. Taking the decision to lockdown without first understanding its impacts on population movements has left states scrambling to do the best they can and millions of Indians are facing a very uncertain future.
Epi-models just a small part
Epi-models are just one part of a very complex puzzle, but to be crystal clear: what we have had to work with was wrong. The lockdown has now distributed the virus to every part of the country (because of the migrant movement) instead of keeping it locally constrained.
For the modelers to argue that this was poor implementation is incorrect. A model that makes a prediction can be judged only on the basis of the veracity of its prediction. That the uncertainty over population movements was not modelled and propagated through to the estimates makes it bad science. Period.
We need to rethink our approach. But we can do so only if we realise that the knowledge needed to fight this virus requires us to first admit our ignorance. Doing so will allow us to move away from the world of ultracrepidarians offering armchair advice with certitude towards a serious rethinking of how our policies should incorporate our limited understanding of this virus and its interactions with human behaviour.
The author is a professor at Georgetown University, and is a senior visiting fellow at the Centre for Policy Research.