Epidemiology models need a middle way, between SEIR (with no forward-looking expectations) and agent-based simulation (too many free parameters). SEIR’s ultimately need to pin down a value for R0, but it is never a fixed value, especially in a global pandemic. The R0 value is only applicable at the very beginning of a pandemic where the parameters of the virus are basically unknown. Even in very small outbreaks there is learning that happens, after SARS-1, hospitals in effected nations established at least rudimentary protocols – this forever alters R0. This fact is already acknowledged by epidemiologists; it is rarely possible to compare the R0 between diseases beyond a rank ordering. The best use of R is as an estimate. Each day contact tracing and testing can inform whether the number of people being infected from each parent case is increasing or decreasing, this is the best leading indicator of future cases and deaths.
Agent-based simulation seems like it could offer the most tailored advice to policy makers. These models could suggest temporarily closing restaurants while leaving hotels open, based the transmission mechanics of a novel disease. These models could also help understand the heterogeneity of disease, which groups could be disproportionately infected, and resources could be deployed to mitigate this. This model isn’t without challenges though. Early in a pandemic there is so little known about the underlying transmission mechanics that quantifying them comes with huge error bands. The actual computation of these models is error-prone – such as the critiques of the IHME model which had non-deterministic output with using the same seed in single threaded execution. Finally, the target of this model is the disease itself rather than the people who will or will not be infected, based largely on the decisions that they make.
These models fall victim to the Lucas critique, agents form expectations about the future, and they make decisions in the face of uncertainty. In economics, the Lucas critique tells us that policy cannot be imposed without considering how agents will respond to the policy. For instance, if a government cuts taxes to boost consumer spending it will not work if consumers expect taxes to raise later to offset the budgetary shortfall. Health policy needs to heed the Lucas critique and any model should explicitly model the cost benefit analysis that is underlying the decisions that people make.
From first principals, what are important stylized facts for COVID-19 modelling:
- Initial uncertainty about all of the parameters of transmission (“I think there is a lot more asymptomatic cases’ – driven mostly by mood affliction). OpenTable data showed people cancelling their dinner reservations before the wide spread of COVID was even discovered in America.
- Better treatments brought down the case fatality rate, people should take more risks as a result.
- There is persistence in behavior that isn’t helpful, an early focus on transmission through surface contact was incorrect but we continue to invest in mitigating this risk
- Availability of a vaccine makes people takes less risk because the uncertainty of an end date is resolved, this assumes time constituency. A counter-vailing force is consumption persistence which limits the degree to which people are willing to change their consumption patterns.
- Mask wearing is a coordination problem, there are areas which have imposed strict mask mandates and see near universal mask use. There are other jurisdictions where despite mask mandates or because of lax enforcement almost no one wears masks.
DeepForest.io 