Predicted values in gestatictical methods or Root-Mean-Square Error

[Anonymous User]

Original User: Nicholina

Hello, i need yr help very much!
I try to use Geostatistical Analist with Local Polynomial Interpolation method.

When i put parameters to model, click Next, and i get window with table of measured, predicted and error values (first picture), i click finish and get model.
[ATTACH=CONFIG]31279[/ATTACH]

After i click on my Local Polynomial Interpolation layer and pick "Validation/Prediction", i get new layer with points, i open its table (picture second) and get the same measured values (which i used for interpolation), new predicted values (that r soooo different from that which r on first picture!!!).
[ATTACH=CONFIG]31280[/ATTACH]

Can some one explaine me, because of what is so big difference in predicted values?

I worry because i have so big Root-Mean-Square Error on first picture, and so small and good after counting data (alone, without using arcgis) on second picture.

I will be glad to hear any yr suggestions and variets! Thank you!

[Anonymous User]

Original User: Eric6346

The reason these two diagnostics are giving different predictions/errors is that they're doing two different things.

The first graphic is showing crossvalidation statistics.  Crossvalidation works by throwing out one input point.  It then uses all the remaining points to predict back to that same location.  It does this procedure for every input point and calculates various diagnostics.  The logic here is that if your model is good, it should be able to accurately predict the value of one input point location using all the other points.  This procedure allows you to determine if your model is good at predicting values in unmeasured locations.

The second graphic is about prediction error.  It predicts back to each input point location without throwing out the input point.  Obviously, the predictions should be much more accurate when it is allowed to use the measured point in the prediction. 

If you were to do this same workflow with IDW, you would see that the second graphic would all contain perfect predictions because IDW is an exact interpolation method (meaning that the surfaces passes through the input points exactly).  Local Polynomial Interpolation is not an exact interpolator, so you will still see some prediction errors even when you use the input point in the calculation.

Let me know if any of this isn't clear.

[VeronikaLem]

Thank you for yr explanation!

I have tried model IDW, after Radial Basis Function and Kriging.

I tried to change all values in different ways to make Root-Mean-Square error smaller. But from start error i get just -0,2 maximum lower... so if i had error 1,1 i get 0,8. Even "Optimize model" gives worst error near 0,9 (Local polynomial interpolation). But its still too big. I wait for maximum 0,2 error.
I use real data about groundwater level, not from my head. Is very important to get real results with lower error. Do u have any ideas what shell i do?