I have a dataset of 400 spatial points that consist of a binary response (attribute) (1, 0) and 22 continuous explanatory variables (covariates).
The variogram of the actual data exhibited signs of non-stationarity, so I have to check for the trend-surface and external drift.
Now a logistic regression fit gives quite strange results, the residuals are by no way looking normal and apparently have extreme outliers (as large as -3000) and a very large variance (~120). Whereas a multivariate linear regression model fits very well, with a reasonable R^2 (~60%) and fairly symmetric histogram/kernel density plot and a small variance of 0.3. In my Reg Kriging I used linear model to get residuals and then proceeded with variogram and kriging. In KED I again used a full models using the actual response variable rather than logit.
Am I doing it correctly? Any suggestions?
If the outliers strongly distort your results you can try a different estimator. For example you can use estimators like the Cressie-Hawkins (1980)-estimator of the Genton-Estimator (1998). They are pretty robust against outliers.
Hi Stephen, Yes, but for RK/KED it's necessary that residuals are normally distributed (or at least a symmetric distribution wth single peak to insure stationarity). I used multivariate linear model in my RK/KED and the results are quite encouraging. RK/KED gives very low kriging variance as compared to OK. To be honest, the accommodation of logistic model in kriging by itself is quite complicated.
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Writing