Week 49 Influenza Forecast for the 2012-2013 U.S. Season

We present results of a forecast initiated Week 49 (beginning December 9, 2012) of the 2012-2013 influenza season for municipalities in the United States. The forecast was made on December 14, 2012. Results from forecasts initiated the two previous weeks (Weeks 47 and 48) are also presented. Also results from the forecast generated with the SIRS model without absolute humidity forcing (no AH) are shown. 1 Retrospective Forecast Retrospective forecast skill is calculated for individual cities, as well as for census region in aggregate, and all cities (excluding the pandemic years 2008-2009 and 2009-2010, which will need to be handled separately in the future). The forecast methods are similar to those described in Shaman and Karspeck (2012). Based on the relationship between prediction accuracy and ensemble spread of these retrospective forecasts we can assign calibrated confidences to our current predictions. Some cities work well in isolation (St. Louis, NYC–not shown), others do not. The question is whether the good and bad cities should simply be aggregated by region, which would suggest that the predictability is really the same among them, but the sample size for an individual city is too small (too few years). Or is it that the statistics are robust and that predictability varies among cities due to differences in local population size, population age structure, geography and connectivity among individuals, etc. We don’t know the answer to this question yet, so for now will give certainty estimates based on the local/municipal record, the regional aggregate and the national aggregate. ∗Corresponding author address: Jeffrey Shaman, Department of Environmental Health Sciences, Mailman School of Public Health, Columbia University, 722 West 168th Street, Rosenfield Building, Room 1104C, New York, NY 10032. E-mail: [email protected] 1 ar X iv :1 21 2. 46 78 v1 [ qbi o. PE ] 1 9 D ec 2 01 2 Figure 1 shows the results for all cities in aggregate using climatological AH and a factor of 5 mapping. Overall the relationship is informative; however, for all lead times there is a basic plateau of skill once the ensemble log variance drops below 2.5 to 3 weeks2. Figure 1: Plot of ensemble mode forecast accuracy versus ensemble spread measured as log ensemble variance +1. Left) 114 cities in aggregate. The runs are binned in increments of 0.25 units and stratified by forecast lead time: 1-3 weeks (red), 4-6 weeks (cyan), 7-9 weeks (blue), 10+ weeks (magenta). Right) Same as left, but the 114 cities grouped by census region. When the cities are grouped by region, there is some heterogeneity. Some regions (e.g. West North Central) show marked improvement of forecast accuracy/skill with decreasing spread across all lead times. Other regions show much more limited skill–the Mountain region only has skill at 1-3 weeks, and the East North Central has problems at 1-3 weeks. 2 2012-2013 Forecast 2.1 Week 49 Forecast The Week 49 forecast (initiated December 9, 2012) basically stays on track with predictions made in prior weeks. Atlanta and Chicago are all predicted to be at peak (±1 week) during week 49