on 25-Nov-2019 (Mon)

Our (Osthus et al, 2019) modeling approach extends the data assimilation/dynamic modeling approach and can be viewed as a combination of the machine learning/regression approach and the data assimilation/dynamic modeling approach. Our model, referred to as a dynamic Bayesian (DB) model, explicitly accounts for systematic deviations between the mechanistic model and the data that are unable to be explained by pure observational noise. This systematic deviation is referred to as model discrepancy and is modeled with a flexible, statistical model.

Discrepancy modeling is an often used and effective modeling approach in the field of computer experiments, where systematic deviations between mechanistic models and data can be common (e.g., Kennedy and O’Hagan, 2001;Bayarri et al., 2007; Higdon et al., 2008;Brynjarsd´ottir and O’Hagan, 2014).

Including a discrepancy model is an appealing way to account for the systematic inadequacy of the mechanistic model. The basic insight leading to the inclusion of a discrepancy model in our DB model is that the disease transmission model and the data-generating model are not equivalent. Disease transmission is merely a component of the data-generating process. Thus, even if a mechanistic model were able to accurately identify the disease transmission process, there may still be a systematic discrepancy between the disease transmission model and the data, introducing a source of uncertainty unable to be accounted for by observational noise alone.

We (Osthus et al., 2019) address the discrepancy/forecasting issue by modeling the discrepancy hierarchically across all flu seasons. This allows us to borrow common discrepancy structure observed in past seasons in the forecasting of the flu for the current season.

This paper (Osthus et al., 2019) makes contributions and advances in the following ways.

• We introduce and demonstrate the importance of discrepancy modeling to the growing and consequential field of flu forecasting. Discrepancy modeling is done hierarchically, allowing information to be shared across available flu seasons.
• We demonstrate the superiority of our approach relative to all models that competed in the CDC’s 2015–2016 and 2016–2017 flu forecasting challenges, providing yet another instance where discrepancy modeling is not only conceptually appealing but also practically effective.
• In an effort to advance flu forecasting capabilities, much effort has been spent identifying possibly useful, nontraditional data sources such as Google (Ginsberg et al., 2009) and Wikipedia (Generous et al., 2014). Alternatively, as we demonstrate, flu forecasting can be improved through carefully made modeling choices, making use of the available traditional data hierarchically.

The basic insight and motivation leading to Osthus et al. (2019)'s development of the DB model is that the disease transmission model (e.g., the SIR model) and the data-generating model are not equivalent; disease transmission is a component of but not equal to the data-generating process. The data-generating model is non-exhaustively comprised of a disease transmission process, a healthcare provider visitation process, an influenza-like illness determination process, and a reporting process. Thus, even if a disease transmission model more sophisticated than the SIR model were used, of which there are numerous (e.g., the SIRS model, the SEIR model), there might still be a disagreement between the best version of the disease transmission model and the data. Rather than attempt to model each component of the data-generating model, we acknowledge there will likely be a systematic disagreement between the best version of the disease transmission model and the data. We then model the commonalities of the discrepancy across flu seasons with a flexible, hierarchical model. The hierarchical discrepancy model allows us to leverage patterns in the data the disease transmission model is incapable of capturing and not simply model the discrepancy as white noise.

The DB model assumes future flu seasons will exhibit similar trajectories to past flu seasons. We showed that the more dissimilar a flu season was as compared to the other considered flu seasons, the worse forecasts were. Because of this underlying assumption, the DB model would be inappropriate for forecasting pandemic influenza.

You may wonder why we are using the name MANUFACTURER in the ApplyMap() function, and not the name [Aircraft Manufacturer] that we renamed it to. This is because renamed fields only become known by the name after the entire LOAD statement has been executed