というNBER論文マンキューコーエンが称賛している。論文の原題は「Policy Implications of Models of the Spread of Coronavirus: Perspectives and Opportunities for Economists」で、著者はChristopher Avery(ハーバード大)、William Bossert(同)、Adam Clark(ヘルムホルツ環境研究センター*1)、Glenn Ellison(MIT)、Sara Fisher Ellison(同)。

There are two primary approaches for modeling the spread of disease: (1) “Mechanistic” and (2) “Phenomenological”. The distinction between these models is generally analogous to the distinction between structural and reduced form models in economics. Just as proponents of structural work tout the ability to extend their models to conduct counterfactual analysis, advocates for mechanistic approaches to disease modeling highlight the importance of out of sample predictions...
...models usually end up with the same general components, e.g. representing the rates at which new individuals become infected, at which infected individuals recover or die, etc. The primary distinction between phenomenological and mechanistic models in epidemiology, therefore, tends to be more directly related to how models have been parameterized than on the functional forms themselves. Models fit based on a priori biological assumptions, or boots-on-the-ground efforts to identify infected individuals and trace their contacts and resulting infections, tend to be labeled as mechanistic. In this sense, mechanistic models in this literature may be seen as analogous to macroeconomic models that fit some parameters to data and then calibrate other parameters to match external evidence. Models that are parameterized through curve-fitting based on reported case or mortality data, tend to be labeled as phenomenological.
The most common approach that has been used to model the spread of SARS-CoV-2 is the “Susceptible / Infectious / Recovered” (SIR) model. In essence, SIR models can be viewed as continuous-time Markov chain models where only a limited number of transitions between states are possible.
SARS-CoV-2の拡散をモデル化する際に最も良く用いられる手法は、「Susceptible / Infectious / Recovered」(SIR)モデルである。基本的にSIRモデルは、状態間の可能な推移の数が限られた連続時間マルコフ連鎖モデルと見做すことができる。


In addition to problems related to heterogeneity, there are three additional challenging aspects of SIR models that make their dynamics especially difficult to predict. First, because the model is nonlinear, small changes in parameter values and initial states have large effects on dynamics. ...Second, because dynamics in these models tends to be both complex and non-monotonic, classic model diagnostics and fitting tools may not be good indicators of whether a model will produce good extrapolations. For example, many models have high predictive ability when fit to the early stages of an epidemic, where growth in the number of infected individuals is approximately exponential. However, comparatively few models are able to predict the saturation point at which the number of new infections begins to decline, much less the expected number of infected individuals at the peak of the epidemic. Third, these problems are compounded by the fact that disease transmission involves substantial time lags. ...
Sadly, all three of these challenges seem to be particularly acute for SARS-CoV-2.