Discussion

Four main points of emphasis emerge from the results. First, for the tested data sets, the interpolation method did not significantly influence the resulting error. There are several possible explanations for this. Foremost, although the GIDS approach may be technically more appropriate than geostatistical approaches for count-based variables, the Poisson model may not have been a good fit for these data, or the data may have been approximately normal enough to remove any advantage of a Poisson-based process over geostatistical approaches. Furthermore, among weighted-average interpolation approaches—a category that includes GIDS—kriging is often the best unbiased predictor for data that are not normally distributed (Johnston and others 2003). Another count-oriented approach—Poisson kriging—has recently emerged in health geography and ecological literature, and this may be a promising future direction for count-based spatial interpolation (Goovaerts 2005, Monestiez and others 2006). In the meantime, GIDS has a number of positive characteristics. It violates fewer assumptions than geostatistical approaches—in particular, the assumption of second-order stationarity (Cressie 1993). Furthermore, the GIDS approach is transparent and easily implemented. To use more complex approaches, particularly PRISM, requires estimation of numerous parameters, so a certain degree of subjectivity is involved. The GIDS approach can easily accommodate covariates besides elevation, and, in fact, could easily be adapted for multiple covariates in order to refine the results. Finally, the GIDS approach has been implemented in R code (R Core Development Team 2006), and as such is an open source resource that may be more readily available than GIS-based interpolation approaches.

Second, the interpolations of the two-count variables appear to have an acceptable degree of error. The distribution of cross-validation errors for the GIDS interpolations are revealing in this regard. For the consecutive-day variable, across all 5 years, only 25 percent of values were exactly predicted, but nearly two-thirds of predicted values were within 1 day of the observed value. For the total-day variable, only 4 percent of values were exactly predicted, but nearly 50 percent were within 5 days and greater than 75 percent were within 10 days. This should be adequate for broad-scale ranking of areas according to their relative risk based on climatic and weather conditions.

The third and perhaps more important point is that the information provided by the constructed annual count maps is substantially different from results that can be captured using monthly climatic data sets based on 30-year normals. For P. ramorum and other currently emerging threats, it may be advantageous to identify areas that have exhibited favorable conditions in a given year and determine whether, for example, the pathogen was positively detected at any nurseries in those areas during that time period. In fact, this suggests a need for a regularly updated database, and the GIDS method may be one way to generate a regularly updated data set from the NCDC data. Recent annual maps can be used in conjunction with 30-year normal data to create a strong picture of current risk.

Fourth, if the count-based variables we calculated are reasonable representations of the level of favorable climatic conditions for P. ramorum, then this suggests that large portions of the Eastern United States—perhaps more than originally estimated—have periods during each year where they may be especially susceptible to infection. Because climate and weather may not be severely limiting factors, detailed analyses of potential pathways and potential host species distribution may be in order for much of the Eastern United States.