All incident cases of WNV in humans were obtained from the New York State Department of Health by county for each of the years from 2000 through 2010 
. In addition, surveillance of WNV in dead birds was also recorded in each county from 2000 through 2008. NYS DOH identifies human WNV cases by an ongoing surveillance system, which was initially adopted during and after the 1999 epidemic in New York City 
. Briefly, New York State's Department of Health relied on the Health Information Network during the initial outbreak to identify cases of WNV disease. Subsequently, in preparation for sustained transmission in the following season, the Department developed a surveillance program that combined the monitoring of mosquitoes, bird populations, and human cases. The coordinated actions of the Health Information Network and the newly developed species monitoring served as the foundation for the subsequent decade of WNV surveillance in New York State 
Climate data were obtained from a monthly surface data product from the National Climate Data Center, which manages meteorologic data products for the National Oceanic and Atmospheric Administration 
. Daily recordings of precipitation and temperature were not available from this agency, thus monthly recordings provide the finest resolution available. Total monthly precipitation for each county for each year in New York State was obtained and summed over the six month period from May through October representing the mosquito and WNV transmission seasons. Seasonal precipitation appeared to demonstrate greater relevance for WNV occurrence than did annual precipitation. Using the total accumulated precipitation over 12 months, there was no crude association between annual precipitation and human infection (IRR
1.0, 95% C.I. 0.99–1.0, p
0.46) and no crude association between annual precipitation and bird infection (IRR
1.0, 95% C.I. 0.99–1.0, p
0.39). Moreover, replacing seasonal precipitation with annual precipitation in the multiple logistic regression model did not quantitatively or qualitatively change the association between hydrogeographic area and human infection, or between temperature and human infection. Moreover, this was similar to another report that failed to identify any association between annual precipitation and human WNV infection 
. Based on the data under current investigation, as well as their agreement with previously published work, it was decided to use seasonal precipitation rather than annual precipitation.
Mean monthly temperatures were also recorded for each of the 11 years from 2000 through 2010. Given the obvious problem of temperature variation by month, but the potential residual impact of higher temperatures overall throughout the WNV season, this study took a unique approach to quantifying temperature. Rather than creating discrete panel temperature data for each month, a monthly-weighted cumulative measure of temperature was used to represent the mosquito and WNV seasons as a whole from May through October. The mean temperature for each month from May through October was multiplied by its weighting factor, which consisted of the proportion (ranging from 0 to 1) of that month's temperature's estimated contribution to the overall occurrence of WNV in humans throughout the season. The relative importance of each month's temperature to the collective season's WNV occurrence was determined by the average distribution of heat across each month from May through October. Hotter months were weighted higher because of their relative importance for mosquito development and for WNV transmission. The specific weighting is as follows: the mean temperature for May was multiplied by 0.05, the mean temperature for June was multiplied by 0.15, and July, August, September, and October mean temperatures were multiplied by 0.3, 0.3, 0.15, and 0.05, respectively, and are depicted in the equation below:
These weights reflect the relative importance considered for each month's mean temperature to the overall occurrence of WNV for each year. The weighted mean temperatures for each month were then summed over all six months to create a single seasonal temperature for each county for each year under study, which then has the same unit interpretation in degrees Celsius. To test the validity of this seasonal temperature measure, sensitivity analyses were conducted wherein the analyses were repeated using the mean temperature for each month alone, again from May through October. The associations between temperature and WNV, and the temperature-adjusted associations between precipitation and hydrogeographic area and WNV, were very similar between the seasonal temperature measure and the individual monthly mean temperatures. Therefore this investigation proceeded with the use of the monthly-weighted seasonal temperature, which is a more comprehensive measure of temperature, as well as being more easily interpretable on a year to year basis.
Hydrogeography data were obtained from the United States Geological Survey National Hydrology Dataset 
. The area, in square kilometers, of all lakes, ponds, swamps, and marshes was used to designate the total surface water per county. The area per county was obtained by calculating the geometry in a geographic information system (GIS). There were no differences observed across types of hydrogeographic features, i.e. lakes and ponds versus swamps and marshes, so all freshwater bodies were combined for easier presentation and interpretation of results. Hereafter, these combined water bodies will be referred to as hydrogeographic features, and the area they represent will be referred to as the hydrogeographic area. While both precipitation and these hydrogeographic features contribute to the overall surface water present in the landscape, the distinction between them follows the assumption that the former, i.e. precipitation, defines the more transient water features in the landscape, while the latter, i.e. hydrogeography, defines the more fixed water features in the landscape.
Random effects models were fit using Poisson regression for panel data and the results used to identify the associations between the hydrogeography and climate factors and human WNV infection per county across the 11 years of WNV surveillance. These panel data are longitudinal data, wherein instances of observations can be conceptualized as discrete time points at each time, i. Moreover, panel data allow for both a temporal and spatial designation of the observations, thus they are fundamentally multidimensional data. With the application of random effects models, each case at time i and space j can be considered as the outcome of a random process. Poisson regression models count data, which are assumed to follow a Poisson distribution. This is a probability distribution of events observable in discrete intervals of time and space. In the current study, the Poisson regression is modeling the annual counts of human infections per the total county population, and therefore estimates the incidence rate. As such, incident rate ratios are reported in the results. Total seasonal precipitation, seasonal temperature, and hydrogeographic area all demonstrated bivariate associations with WNV cases and were subsequently included in an adjusted Poisson model as independent variables. The number of WNV-infected dead birds each year was also associated with the number of human infections per county. An additional aim of this study was to explore whether the associations between climate and hydrogeography factors and human WNV cases may be mediated by the extent of bird infections per year. Mediation cannot be adequately assessed simply by adding the purported mediator to a regression model and examining the change in regression coefficients of the model predictors whose effects are hypothesized to be mediated. To more appropriately assess the potential mediation of climate and hydrogeography by bird infection, instrumental variable regression was applied with hydrogeographic area and precipitation as the specified instruments, once again in regression analysis that accounted for the county and time panels inherent in the data.
Finally a map was constructed to show the spatial fit of hydrogeography and climate factors as predictors of WNV cases. A simple approach was adopted by adding just two layers to the map. The first layer is a monochrome choropleth of the predicted cumulative incidence of WNV from 2000 to 2010. The second layer overlays the choropleth with graduated symbols representing the actual number of observed cases per county over the same 11 year period.
ArcGIS was used for all mapping procedures and for calculating the Moran's Index to identify whether spatial clustering of the residuals was present (ESRI), and STATA v.11 was used for the Poisson and instrumental variable regression modeling, both specific to panel data, with the xtpoisson and xtivreg command procedures, respectively (STATA Corp., College Stattion, TX).