Ambiguity in the use of SIR models to fit epidemic incidence data

원문 초록

When fitting a multi-parameter model to a data set, computer algorithms may suggest that a range of parameters provide equally reasonable fits, making the parameter estimation difficult. Here, we prove this fact for an SIR model. We say a set of parameter values is a good fit to outbreak data if the solution has the data's three most significant characteristics: the standard deviation, the mean time, and the total number of cases. In our model, in addition to the "basic reproduction number" $R_0$, three other parameters need to be estimated to fit a solution to outbreak data. We will show that those parameters can be chosen so that each gives a linear transformation of a solution's incidence data. As a result, we show that for every choice of $R_0>1$, there is a good fit for each outbreak. We also illustrate our results by providing the least square best fits of the New York City and London data sets of the Omicron variant of COVID-19. Furthermore, we show how versions of the SIR model with $N$ compartments have far more good fits- - indeed a high dimensional set of good fits -- for each target -- showing that more complicated models may have an even greater problem in overparametrizing outbreak characteristics.