Inference for partially observed epidemic dynamics guided by Kalman filtering techniques

Despite the recent development of methods dealing with partially observed epidemics (unobserved model coordinates, discrete and noisy outbreak data), some limitations remain in practice, mainly related to the amount of augmented data and the adjustment of numerous tuning parameters. In particular, coordinates of dynamic epidemic models being coupled, the presence of unobserved ones leads to a statistically difficult problem. Our aim is to propose a generic inference method easily practicable and able to tackle these issues. Using the properties of epidemics in large populations, we first build a two-layer model. Through a diffusion based approach, we obtain a Gaussian approximation of the epidemic density-dependent Markovian jump process, which represents the state model. The observational model consists in noisy observations of the observed coordinates and is approximated by Gaussian distributions. Then, we develop an inference method based on an approximate likelihood using Kalman filter recursions to estimate parameters of both state and observational models. Performances of estimators of key model parameters are assessed on simulated data of SIR epidemic dynamics for different scenarios with respect to the population size and the number of observations, and compared with those obtained by the currently largely used method of maximum iterated filtering (MIF). Finally, we apply our method on a real data set of influenza outbreak in a North England boarding school in 1978.