|
2. Thin Plate Splines: Tps
2.1 The Model where f(X) is a d dimensional surface, and the object is to fit a thin plate spline surface to irregularly spaced data, e are random errors with zero mean, uncorrelated and with variances /W_i. A mathematical summary of this type of spline is given in the last section of this manual. 2.2 Cross-Validation where N = (1/n) * Y' (I - A())' W (I - A())Y and D = (1 - trA() /n)^2. Here, Y' denotes Y transposed, etc... It is also possible to include a cost parameter that can give more (or less) weight to the effective number of parameters beyond the base polynomial model. An alternate estimate is by maximum likelihood (see mathematical section). 2.3 Tps Example: ozone ozone.tps <- Tps( ozone$x, ozone$y) summary( ozone.tps)
Residuals: min 1st Q median 3rd Q max -6.801 -1.434 -0.5055 1.439 7.791 REMARKS Covariance function: rad.cov
2. the total number of unique points. 3. the maximum degree of polynomial (m - 1) included in the spline model. 4. total number of polynomial terms up to the maximum degree. This is the total number of conventional parameters for the spline. 5. total number of effective parameters associated with the spline curve including those of the polynomial. This number is controlled by the parameter and is the trace of the smoothing matrix, A(). 6. spline degrees of freedom subtracted from the total number of observations. 7. estimate of by maximum likelihood. 8. estimate of by using residual sum of squares and effective d.f. 9. - 11. these are explained in the discussion of the Krig function and the spatial model. 12. value of smoothing parameter used to find spline estimate. If no value has been specified, the smoothing parameter estimated by GCV is reported. 13. value of cost parameter. 14. value of GCV function at . This is an estimate of the average squared prediction error for a given value of . If has been estimated, then the minimum value of the GCV function is reported. plot( ozone.tps) Finally, here is the simplest example of using the predict function. The default is to predict at the observed data locations. ozone.tps.pred <- predict( ozone.tps) ozone.tps.pred [1] 38.35702 38.62821 38.44499 39.01395 38.79481 39.72864 38.30791 39.35719 [9] 39.19664 39.63307 40.98843 40.05890 41.11455 40.87318 41.42214 40.35936 [17] 40.03565 39.10773 41.34989 40.83722 |