Data Collection

Figure 1

Figure 1. Dead bird cluster surveillance system. SAS, SAS statistical package (SAS Institute, Inc., Cary, NC); ESRI, Environmental Systems Research Institute; SQL, structured query language.

Data collection procedures were the same for 2000 and 2001 and have been described in detail elsewhere (4,6). In brief, dead birds were reported by the public through an interactive voice-response telephone system or the Internet (Figure 1). The information included the date found and the location and species of the dead bird. A sample of dead birds that met selection criteria (i.e., the bird was recently dead, had little apparent decay or trauma) were submitted for necropsy and testing (4). Since pigeon (Rock Dove) deaths are common but rarely associated with WNV (7,8), they were excluded from all clustering analyses.

Mosquitoes were collected weekly from >100 traps dispersed throughout the city (4). Nearly all areas of the city were within 2 miles of a mosquito trap. Multiple mosquitoes from the same trap and of the same species were pooled, and each pool was tested for evidence of WNV; a result was considered positive if at least one mosquito was infected. The New York City Department of Health and Mental Hygiene conducted citywide active hospital-based physician and laboratory surveillance for human WNV infection (6).

Geocoding

All dead bird reports, mosquito traps, and human case-patients with address information were geocoded to a point location, where possible, with the ArcView Geographic Information System (GIS) software (ESRI, Redlands, CA). Multiple dead bird reports from the same location on the same date were counted as one. Dead bird reports were attributed to the census tracts (n=2,215; mean area = 0.13 square miles) in which they were found by using a spatial join. The latitude and longitude in decimal degrees for each census tract centroid were used in cluster analyses.

Spatial Scan Statistic 
 For early detection of localized clusters of dead birds, we used a prospective surveillance system that is based on the spatial scan statistic (9). This scan statistic uses a circular window to represent potential geographic clusters. By continuously changing the circle center and radius, the window scans the geographic area for potential localized clusters without incorporating prior assumptions about their size or location.

Even in the absence of a WNV epidemic, more birds will be reported from some areas than others. To adjust the analysis for such geographic variability, we used historical dead bird counts from a given census tract as baseline (pre-outbreak) controls; recent birds counts were used as cases. We defined cases as the dead bird reports that occurred in the 7 days before the date of analysis. A minimum 2-week buffer zone between case and control birds was also established, thereby limiting the influence of emerging clusters on the analysis. Thus, before laboratory confirmation of a WNV-attributable dead bird cluster, controls were defined as dead bird reports from April 1 until 21 days before the day the analysis was performed. April 1 was chosen as the start date for the control period because many bird species change their geographic location and habitats during the winter months. After laboratory confirmation of WNV from within a cluster area, the control file was frozen to include only birds reported before that cluster, that is, during the pre-outbreak period.

For each circle evaluated for potential clustering with the scan statistic, census tracts are classified as being inside or outside the potential cluster, depending on their centroid location. The number of recently observed dead birds (cases) inside and outside the potential cluster is compared with the expected case-count on the basis of the geographic distribution of historical controls. For example, the expected number of cases inside the circle is equal to the total number of cases citywide, multiplied by the proportion of all dead birds found within the circle during the control period. Based on the observed and the expected, a Bernoulli-based likelihood is calculated for each circle, and the circle with the maximum likelihood is defined as the most unusual cluster, that is, the cluster least likely to be due to chance. To adjust for the multiple testing inherent in the many possible cluster locations and sizes, we evaluated the statistical significance (p value) of this cluster by using Monte Carlo hypothesis testing (9–11). In this method, the likelihood of the most unusual cluster in the observed dataset is rank-ordered among the maximum likelihoods of 999 simulated (randomized) datasets.

We performed the data processing using automated SAS programs (SAS Institute, Inc., Cary, NC) to organize the dead bird reports into case and control files, invoke the spatial scanning software (SaTScan version 2.1, freeware available from: URL: http://www.satscan.org), extract cluster information from the SaTScan output file, archive the cluster data, and export a dbase file for mapping purposes. Using Arcview GIS software, we spatially joined the dbase file to a census tract layer and produced a map displaying the cumulative frequency of dead bird clusters in each census tract.

To test the method, analyses were performed retrospectively for the 2000 data through serial daily replications for every day from June 1 through October 1. To maximize the sensitivity and timeliness of this early warning system, all clusters with p<0.10 were mapped. A prospective dead bird clustering surveillance system was implemented in real-time beginning June 22, 2001; daily analyses were performed by using the same definitions of cases and controls as for the year 2000. Animated maps showing the results for 2000 and 2001 and automated SAS programs for reproducing these analyses on local dead bird reporting data are online (available from: http://www.nyc.gov/health/cluster.html).

Year 2000 Simulation

Figure 2

Figure 2. Dead bird clusters, West Nile virus (WNV)-positive dead birds, human cases, and mosquito traps, New York City, 2000. The shading represents the cumulative frequency of dead bird clusters in each census...

The first evidence of clustering in New York City in the 2000 simulation was on June 14, when a group of 26 census tracts in northern Staten Island were found to have had 14 dead bird reports in the previous 7 days (June 7–14); only 5 were expected on the basis of the baseline period (April 1–May 24) (p=0.06) (Figure 2a). By June 17, all of Staten Island and an adjacent part of Brooklyn were included in the dead bird cluster (observed = 36, expected = 19, p=0.02). The first laboratory evidence of WNV from New York City was reported from Staten Island 1 month later, from two dead birds collected on July 5 and a mosquito pool collected on July 7 (Figure 2b). Spraying insecticide for adult mosquitoes was conducted on July 19. The first human case in 2000 was a resident of Staten Island whose illness began on July 20 and was diagnosed on July 28 (Table). During the next 2 months, Staten Island would prove to be the epicenter of the WNV encephalitis human outbreak in 2000, with most of the human cases and positive mosquito pools in New York City (Figures 2c and 2d). 
 Dead bird clustering was also apparent in western Brooklyn (June 17), eastern Queens (July 6), and southern Brooklyn (August 16)—all sites that subsequently had numerous WNV-positive birds and mosquitoes and human cases (Figure 2 and Table). Only one patient’s onset of illness was not preceded by dead bird clustering near his place of residence. That patient was a construction worker living in Manhattan, who attributed his illness to mosquito bites he sustained while working on an outdoor construction project in eastern Queens.

Year 2001 Implementation

Figure 3

Figure 3. Dead bird clusters, West Nile virus (WNV)-positive dead birds, human cases, and mosquito traps, New York City, 2001. The shading represents the cumulative frequency of dead bird clusters in each census...

In 2001, data for dead bird clustering analysis were first available on June 22. A retrospective analysis of dead bird reports before this date showed repeated clusters in central Staten Island (Figure 3a). Prospective real-time surveillance beginning June 22 found repeated clustering in eastern Queens, which prompted a program of intensified larval surveillance and control as well as abatement of standing water starting June 27. Reports from this area were prioritized for dead bird pickup and testing, and additional mosquito traps were set. On July 19, the clustering in eastern Queens was shown to be likely due to WNV through laboratory confirmation in a pool of mosquitoes collected on July 3 and a hatch-year live bird sampled on July 6 (Figure 3b). 
 During the next 2 months, six areas with major dead bird clustering were identified, prompting intensified surveillance activities. Although active surveillance for human infections was implemented citywide, five of seven diagnosed human cases in 2001 were identified in residents of four cluster areas (Staten Island, eastern Queens, South Brooklyn, East Brooklyn). An additional human case occurred in a homeless person who frequented an area close to the southern Queens cluster area.

Dead bird reporting and analysis were interrupted by the September 11 attacks on the World Trade Centers. The subsequent low number of reports contributed to widespread geographic clusters, containing only a small number of dead bird reports and encompassing most of Brooklyn and Manhattan. The seventh diagnosed human case in the city was in a resident of a cluster area in lower Manhattan. In all, dead bird clusters occurred 0–40 days (median 12) before the onset of human illness and 12–45 days (median 17) before human diagnosis. In most cases, dead bird clusters also preceded time of collection of WNV-positive mosquitoes and birds (Table).