My study is all about evaluating the association between dengue incidences and weather parameters, land use/cover, and demographic characteristics of the study area. I am using ArcGIS regression tools OLS and GWR. The study area is divided in 150 polygons for which average values of explanatory variables are calculated using zonal statistics. There are 20 polygons where no dengue case is reported. My first question is; would it be okay if I remove these polygons from analysis since the predictive model is giving negative results in GWR model using predicted values of explanatory variables. Otherwise, is there any justification for negative dengue cases?
Also, the histograms of the variables show a non-normal distribution. When I use natural logarithm transformation, the variables look near to normal distribution (although some are still skewed). But the regression results of OLS deteriorated badly (very low squared r value). The GWR model failed to run as well. My second question is that can I still go for non-transformed variables and would that be fine if I discuss in the paper all these trails or simply omit the transformation exercise since it is not giving any result (and I think not a requirement for regression analysis).
One of my reviewers suggested to use Poisson's regression which is considered a better model for count variables as dengue cases. The scope of my study was to evaluate the spatial association between dengue and other parameters using ArcGIS. Can I perform the suggested analysis in ArcGIS (I could not find it)? If not, would stating the scope as a reason of not employing Poisson's model be a reasonable response?
