Accounting for seasonal patterns in syndromic surveillance data for outbreak detection

Table 1 Detection probabilities (multiplied by 100) by forecast method, length of training data, and data set.

The (x,y,z) entry in each cell denotes 100 times the DP (with a false alarm rate of one per year) for the real data, data simulated from the nonhierarchical model, and data simulated from the hierarchical model, respectively. Individual entries are each based on 1000 simulated outbreaks, so their confidence limits are approximately ± 2. In the last row, boldface entries indicate optimistic DPs attributable to using data simulated from the nonhierarchical model. The eight methods are based on applying Page's test to forecast errors based on forecasts from: (1) a fit to Eq. (1), (2) EWMA, (3) moving average with a 1-day gap, (4) moving average with a 3-day gap, (5) historical average using day d, d-1, d+1 in the training data to predict day d in the testing data, (6) historical average using three 4-day windows, (7) historical average using day d, and the (8) historical average using one 4-day window.

ISSN: 1472-6947