The methodological description consists of four parts: (i) abstraction and aggregation of the outcome data, (ii) abstraction, aggregation, and mapping of the exposure data, (iii) static mapping of the outcome data, and (iv) composition of the dynamic map. Details of these four steps are as follows.
Health outcome data--abstraction and aggregation
Hospitalization records for Salmonella infections were abstracted from the Centers for Medicare and Medicaid Services (CMS) for all Medicare recipients aged 65 or above in the contiguous U.S. for 2002 (Alaska, Hawaii, Virgin Islands and Puerto Rico were excluded from the analysis). In the U.S., more than 95% of the population aged 65 or above are covered by Medicare; hence the CMS records are nationally representative. Each hospitalization record contains information on the date of admission, ZIP code and state of residence, age and up to 10 diagnostic codes based on the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM). Records containing any diagnosis starting with "003", indicating "other Salmonella infections", were included in this analysis. Conditions starting with "002", indicating "typhoid and paratyphoid fevers" caused by Salmonella typhi infection, were excluded due to their low frequencies.
Temporally, all records were aggregated to a monthly level; spatially, to a county level. Monthly hospitalization rates were computed using county population aged 65 or above as the denominator. Population counts were obtained from the Census 2000, Summary File 1. To minimize spurious high rates caused by extremely low denominators, a spatial re-aggregation scheme [see Figure ] was applied to incorporate counties with a low number of elderly into the adjacent counties until the total number of elderly exceeded 1,000, a parameter that was selected based upon the population of elderly per county and the frequency of diseases. The re-aggregation collapsed the number of counties within the continental U.S. from 3,109 to 2,794. On the maps, the boundaries between two aggregated counties were removed to form one larger area. For each aggregated county, the elderly population and monthly Salmonella
hospitalization counts were summed together. Monthly hospitalization rates per 10,000 elderly, Rij
, were calculated for each aggregated county i (i = 1:2,794) and month j (j = 1:12) as follows:
The re-aggregation scheme. The chart explains the chain of decisions made when aggregating counties with low elderly population in this study. Arrows with "Y" indicates "Yes"; "N" indicates "No".
Exposure data--abstraction, aggregation and mapping
Ambient temperature data were obtained from the PRISM Group at Oregon State University for average maximum monthly temperature, average maximum annual temperature, and average maximum temperature anomaly for 2002. Data were downloaded in ESRI ArcInfo Ascii Grid format with a spatial resolution of 4 kilometer grid cells and imported into ESRI's ArcMap software. The aggregated county polygons were overlaid onto each temperature grid and the mean temperature or temperature anomaly was calculated for each county. Choropleth maps for the average monthly and annual maximum temperature were created by considering the full range of county-level temperature values with the classification emphasizing temperatures above freezing shown in shades of yellow to dark orange. The monthly temperature anomaly maps show the deviation in temperature on a monthly basis from the 30-year norm for each month. The choropleth maps for the temperature anomaly emphasize temperature deviations greater than +3°C in pink and less than -2°C in purple.
Exposure data on livestock were obtained from the 2002 quinquennial Census of Agriculture gathered by the U.S. Department of Agriculture's National Agricultural Statistics Service. County-level data on the number of broiler and other meat-type chickens sold from livestock farms to food distributors were downloaded and mapped by matching the county FIPS codes with county-level polygons. A choropleth map of the number of broiler chickens sold in 2002 was created by first performing a logarithmic transformation on the data to alleviate the problematic skewness, followed by assigning the data into five categories using a natural breaks classification.
Mapping of the health outcome
Twelve monthly static maps for 2002 were created. To facilitate visual comparison between months, the bin sizes adopted in the legend were unified across the 12 maps. Countywide Salmonella hospitalization data for the elderly were extremely skewed due to a disproportionately high amount of low rates and a small amount of high rate outliers. Seasonal changes further exaggerated the problem since there was a two to three-fold increase in salmonellosis during the summer time. To alleviate the problems caused by the skewness, a logarithmic transformation was applied. The logarithmic transformation helps to optimally assign categories into high and low brackets (Figure ). The unified bin sizes were derived by assigning a natural break classification scheme to the month with the highest amount of hospitalization cases. The resulting classification scheme was then applied to the other maps.
Natural breaks classification using logarithmic transformation. A. Natural breaks classification on untransformed Salmonella hospitalization rate data, B. Natural breaks classification on log transformed Salmonella hospitalization rate data.
Hospitalization rates were represented on the map with graduated dot symbols. The graduated colour and symbol size of the dots allow for a quick visual comparison between high and low rates; larger and darker dots indicate higher rates and smaller and lighter dots indicate lower rates. The sizes of the graduated dot symbols are proportional to the outcome rates and hence are preferred over polygons with graduating shades, which are harder to interpret due to the wide range of county sizes. The size of the largest graduated dot symbol was determined using the average county size so that overlapping of dots and county boundaries were minimized.
The monthly maps of county-level Salmonella-related hospitalizations were superimposed onto maps of environmental exposures including: average monthly maximum temperature, the number of broiler chickens sold in 2002, and monthly temperature deviation from the 30-year norm. Maps were exported from ArcMap and imported into Adobe Flash software as an individual key frame. The frames were sequentially incorporated into a movie.
The speed for playing the map movie was optimized so that that the viewers can have enough time to identify clusters and commit the frame to short memory, allowing them to distinguish changes when the next frame appears. Herein, the dynamic maps are shown at a frame rate of one frame per second and the Salmonella infection data span two key frames. An interactive interface consisting of control buttons for stop, play, move forward or back one frame, and replay are added to help the viewers investigate the dynamic map at their own pace.
To indicate the calendar time, a label with the year and month was incorporated into the maps. For maps that span more than one year, a sliding calendar bar was added to help the viewer track time duration. The sliding calendar bar could also be interactive so that the viewer could easily navigate to a specific point in time.
Narrations were incorporated into the movies to help the viewers appreciate the dynamics of Salmonella hospitalizations with respect to the exposures.