I am studying West Nile Virus risk factors in a specific county using mosquito surveillance data over a ten year study period. The highest resolution of the mosquito data that could be obtained was at the city level. There were 2,040 traps set over the study period, but since the data standardization was at the city level, this resulted in my sample size being N = 57. I have performed OLS regression in R and in ArcMap 10.2. I selected my explanatory variables based on known predictors from the literature and then used the leaps function in R to select the best fit model. Due to the low number of observations, I did not want to run the risk of overfitting the model so I decided to keep the number of predictors under 6. The best subset of variables from the leaps function was y = β0 + β1x1…+ β4x4 + e; where y is the percentage of WNV positive traps in a city, and the explanatory variables are mosquito abundance, population density, percent under federal poverty level, and percent Hispanic. The OLS regression in ArcMap produced a significant Koenker statistic, indicating spatial non-stationarity. I want to run a geographically weighted regression but I suspect my sample size is too small.
I am self-taught ArcMap and this is the first time I’ve done regression using this software. I’ve done some tutorials and it is suggested that a small sample size should NOT be used for GWR; however, I wonder if it is dependent on the application. My analysis is more exploratory than anything, and I would still like to present a GWR regression but also present the limitations associated with the small sample size and encourage mosquito control programs to geocode their trap locations.
What limitations are associated with a small sample size in GWR and is there any way to quantify these limitations using the GWR output?