3. A Simulated Regression Example
Our first example involves data simulated from a bivariate regression model proposed by Gu, Bates, Chen and Wahba (1990). The design consists of 300 ``semi-random" points in the unit square. At each point our response is , where the true regression function f is given byand , are independent, standard normal random variables. This problem has been considered by a number of authors for evaluating the performance of various schemes based on tensor-product splines (Breiman 1991, Friedman 1991).
Figure 1. Triogram fits to the simulated regression data. In this dynamic figure, you can rotate the perspective plot by clicking in the panel and dragging the mouse. You can scroll through the model selection process by positioning your mouse in the perspective plot window (not the GCV window) and hitting:
Note that the model corresponding to the minimum GCV value is the one selected by the Triogram procedure.
- the 'n' key or the spacebar for the next step,
- the 'p' or 'b' keys for the previous step,
- the '0' (zero) key to return to the initial model, or
- the 'd' key to toggle between our fits and the underlying data .
In the computations reported here we used the same design points as Gu et al. (1990). For our initial triangulation we divide the unit square into four triangles by drawing in both diagonals, yielding an initial model with five degrees of freedom. Since this data set is fairly small, it is computationally feasible to fit models with many triangles, and to consider many possible candidate vertices. With this in mind, we set K=5 in equation (11) of our paper and entertain new vertices at the points given in the right hand panel of Figure 5 in the same paper. For this presentation, we took the maximum number of vertices to be 20 and we required a minimum number of four data points in each triangle. The penalty parameter a in the GCV criterion was set equal to 4 (see equation number 10 in the Triogram paper). While this choice seemed to result in the smallest mean integrated squared error across our simulations, taking a in a neighborhood of 4 yielded very similar results.
In Figure 1 above, we present a simple interactive display to give you a feel for what the Triogram procedure is doing from one step to the next. You are able to step through the chain of fits (first the addition phase and then the deletion phase) and examine more closely a perspective plot of the fit. This type of display is by far the best for understanding our Triogram models.
You are currently reading Section 3, an application of Triograms to some simulated data.
- Introduction and acknowledgements
- A short description of Triogram modeling
- A regression example with simulated data
- A bivariate density estimation example
- A regression surface with a ridge
- References
- A PostScript version of Triogram Models
- Send us comments !
You can also bail out and head back to my home page .