Models serve many functions. They improve our abilities to reason, explain, design, communicate, act, predict, and to explore. Models take on particular importance when we confront a challenge characterized by complexity and uncertainty, such as how to reduce the spread and mitigate the impact of COVID 19. In these context, we use models to reason through how the disease spreads, to explain data as it arrives, to communicate information to individuals, in particular to explain why experts advocate drastic actions such as cancelling all university classes, festivals, and (potentially) entire industires, when only a few hundred people have been effected, they help us predict how the epidemic will spread, and most importantly, they help us design policies and guide actions in real time.
Models can reveal what would be requied to prevent the spread of COVID-19 and what might decrease the rate of spread. Slowing the rate increases the possibility that the virus mutates to a less lethal form, creates more time for discovery of a vaccine, reduces peak demand for hospital beds, and allows for more time for experimentation of policies that reduce or prevent the spread.
When spread of disease is both complex and uncertain. Disease spread is complex because it consists of diverse populations (people and the many strains of COVID-19), who interact within a contact structure (in this case, through physical interaction or by touching commmon objects), who adapt based on their information and state (people wash their hands, avoid large crowds, etc... and the virus evolves). Disease spread is uncertain both because we lack information (we do not know who has the disease and how those people behave nor do we know the the likelihood of transmission) and because we cannot predict future adaptations by people and evoluitions of the virus. The virus could, by chance, mutate to a less lethal variant. Or, scientists could produce a vaccine.
The complexity of the system, the gaps in current information and the significants of random events in the future, all combine to make accurate predictions of how COVID-19 will spread impossible. Any one model will be wrong. Thus, in trying to predict the future, evaluate a potential policy action, we should employ multiple models of each type, what is know as an ensemble. The various models in an ensemble make different assumptions about how the virus spreads and how people interact and adapt. They thus lead to different predictions. The variants within a class of model also often differ in how they categorize people. This too leads to different outcomes.
In addition to applying multiple variants of a single model, we also want to construct diverse models. Model diversity serves two functions. First, as models serve different functions, they differ in their granularity (the number of assumptions, parameters, and so on) and in their focus (the variables they consider). For example, our first class of models will compare linear and exponential growth. Those models include no details about geography, social networks, or adaptation. Their two purposes are two help people reason and then to communicate effects. In a linear model changine the coefficient from 0.9 to 1.1 has a modest effect. In an exponential model, a similar change in an exponent transforms a non event to a catastrophe.
Second, models that make different assumptions and include different dimensions and features will almost certainly lead to different insights about policy actions. Considering multiple models that bring different perspectives, that highlight different features such as social networks, transportation systems, and behavioral adaptions, enables us to create a dialogue across models that results in more robust understandings and wiser policy actions.
Contributors: Michael Lin