Using Box-Jenkins modeling techniques to forecast future disease burden and identify aberrations in public health surveillance reports

Objective: This study examines the potential to predict future disease burden based upon the historical record within specific jurisdictions using Box-Jenkins statistical models. It also examines what influence, if any, jurisdictional size and rate of disease has upon the accuracy of a forecast.

Methods: Box–Jenkins forecast models were constructed by stratifying 30,104 disease reports by year, disease type and jurisdiction. Three holdout samples were used to compare forecast accuracy against the actual values for the same time periods. Accuracy was determined by comparing the projected data against the holdout sample using Mean Absolute Percentage Error calculations. Forecast predication intervals were also used to explore the relationships between a holdout sample actual value and the predication interval associated with each forecast at one, two and three standard deviations.

Results: Forecasts have an absolute accuracy between 77 – 89% with the accuracy being independent of jurisdictional size (r (3) = .593, p > .05) or rate of disease (r (3) = .455, p > .05). Larger jurisdictions are significantly (rho (3) = .9, p < .05) more likely to have forecast values located within the forecast interval associated with one standard deviation.

Conclusions: This study demonstrates that it is possible to predict future disease burden within jurisdictions of varying size using Box-Jenkins forecasting techniques. It shows promise as a tool to monitor disease trends, predict disease investigation resource allocation and potentially identify outbreaks earlier. Each forecast establishes a baseline that may be used to compare observed vs. expected values.