Gregoire The correlation between actual value and error of estimation is always present to some extent and is simply due to the estimation process. High values will b eunderestimated from neighbouring samples. Low values will be overestimated from neighbouring samples. The only way you can remove this is by using a more complex estimator than a weighted average. Have you plotted the actual value versus the estimates? This will tell you whether you are getting any meaningful prediction or not. Generally, the stronger the correlation here the less you'll get with the errors. FYI: we use (actual - estimate) in our discussions. Not sure why, just a personal preference. Isobel http://www.kriging.com
Monica Palaseanu-Lovejoy <[EMAIL PROTECTED]> wrote: Hi, If there is a very high nugget effect i would expect that the predictions are very close to the mean of the data, with very little variation. In this case you would get a very high correlation (either close to 1 or -1 - depending on how you calculated the residuals). Did you check for local outliers??? If you have a high percentage of local outliers kriging is not a good choice - in my experience - stationarity is usually violated, and the predictions are very poor indeed. Maybe you should investigate other methods of interpolations ..... one of my favorite is multiquadric radial basis function which in many cases can be compared with kriging, performs better when a high percentage of local outliers exist, and does not require stationarity. Monica ==================================== Monica Palaseanu-Lovejoy, PhD Jacobs Technology US Geological Survey Florida Integrated Science Center 600 4th Street South St. Petersburg, FL 33701 Ph: 727-803-8747 x 3068 Fx: 727-803-2031 email: [EMAIL PROTECTED] ==================================== "Gregoire Dubois" <[EMAIL PROTECTED]> Sent by: [EMAIL PROTECTED] 01/30/2008 06:59 AM Please respond to "Gregoire Dubois" <[EMAIL PROTECTED]> To <ai-geostats@jrc.it> cc Subject AI-GEOSTATS: Correlation between kriging residuals and input data Dear list, Having fit a variogram to a dataset (indoor radon measurements) and applied cross-validations, I noticed the perfect negative correlation (-0.95) between my kriging residuals and my input data. This means that I am overestimating as much the low values as I am underestimating the high values, something I am expecting since the mean of the residuals -> 0, a property of kriging. Fine so far. What I am puzzled about is of the possible reasons of getting such a strong slope (close to -1) of the plot of my residuals against my input data? This, I understand, highlights that I am doing a systematic error somewhere which I want to avoid obviously. I thought I extracted properly the spatially correlated component of my dataset (the variogram of my residuals seems to show a pure nugget effect) but I still can't find any reasonable explanation for the systematic errors. Any hints? I must have missed something obvious here. Many thanks for any feedback. Best regards, Gregoire