Timely monitoring and prediction of the trajectory of seasonal influenza epidemics allows hospitals and medical centers to prepare for, and provide better service to, patients with influenza. The U.S. Outpatient Influenza-like Illness Surveillance Network, or ILINet, collects data on influenza-like illnesses from over 3,300 health care providers, and uses these data to produce indicators of current influenza epidemic severity.

ILINet data provide an unbiased estimate of the severity of a season's influenza epidemic, and are typically reported at a lag of about two weeks.

Other sources of influenza severity, such as indices calculated from search engine query data from Google, Twitter, and Wikipeida, are provided in near-real time. However, these sources of data are less direct measurements of influenza severity than ILINet indicators, and are likely to suffer from bias.

We begin by describing general methods for inference on state space models implemented in the NIMBLE R package, and demonstrate these inferential methods as applied to influenza outbreak forecasting. We then examine model specifications to estimate epidemic severity which incorporate data from both ILINet

and other real-time, possibly biased sources. We fit these models using Google Flu Trends data, which uses the number of Google searches for influenza related keywords to calculate an estimate of epidemic severity.

We explicitly model the possible bias of the Google Flu Trends data, which allows us to make epidemic severity predictions which take advantage of the recency of Google Flu Trends data and the accuracy of ILINet data, and we preform estimation using Bayesian methods. Models with and without explicit bias modeling are compared to models using only ILINet data, and it is found that including GFT data significantly improves forecasting accuracy of epidemic severity. We also propose hierarchical models which incorporate multiple seasons of influenza data, and evaluate the forecasting benefits that hierarchical modeling confers.