Fitting dynamic models to epidemic outbreaks with quantified uncertainty: A primer for parameter uncertainty, identifiability, and forecasts

Chowell, Gerardo

Structural parameter non-identifiability is often the most difficult to remedy as it requires appropriately modifying the model to eliminate the structural non-identifiability issue. On the other hand, practical parameter non-identifiability issues could be fixed by 1) employing an alternative model of lower complexity when possible, 2) collecting more data about other states in the system to better characterize the system dynamical features, 3) increasing the spatial-temporal resolution of the data to better constrain the model parameters and/or 4) reducing the number of parameters that are jointly estimated, perhaps by constraining a subset of the unknown parameters based on estimates previously reported in similar studies and conducting extensive sensitivity analyses on those parameters (Arriola et al., 2009 ). Finally, specific approaches have been adapted to address parameter identifiability including regularization techniques that aim for stable parameter reconstruction (Smirnova & Chowell, 2017) .