5. A Regression Surface with a Ridge

Cleveland and Fuentes (1996) discuss data collected during an experiment on the processing of liquid crystal mixtures. The response is the voltage V necessary to turn the material from opaque to clear. In our analysis we use two predictors: the percentage P of liquid crystal in the mixture and the temperature T of the mixture, measured in degrees Celsius. The experiment originally contained a third factor (the intensity of the light used in the processing) that was dropped half way into the experiment. Cleveland and Fuentes (1996) fit a model consisting of two half planes that join along a line in T and P space. Triograms are a natural approach for automatically fitting a piecewise planar model.

Figure 3. Triogram fits to the experimental data of Cleveland and Fuentes. As in the previous dynamic figures, 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 complete experiment involves just 47 data points. Since we have so little data, we do not want to use too many basis functions in our initial model. Therefore, we took as our initial triangulation a 15% enlargement of the smallest triangle that contained all the data. We obtained the 15% expansion by positioning the barycenter of the original triangle at the origin, multiplying the shifted coordinates by 1.15 and then moving the triangle back to its original position. In Figure 3 we present an interactive display of both the data and our sequence of triogram fits. As in the previous example we required the minimum number of data points in each triangle to be four. Since the data set is so small, it seemed reasonable to consider a somewhat smaller number of possible new vertices than in the simulated example above, and so we set K=4. The largest model that was fitted had only 9 vertices, since none of the triangles could be further subdivided without violating the requirement on the minimum number of data points. The GCV criterion with penalty parameter 4 selected a Triogram model with six vertices.


You are currently reading Section 5, an application of Triograms to some experimental data collected at Bell Labs.

  1. Introduction and acknowledgements
  2. A short description of Triogram modeling
  3. A regression example with simulated data
  4. A bivariate density estimation example
  5. A regression surface with a ridge
  6. References
  7. A PostScript version of Triogram Models
  8. Send us comments !

You can also bail out and head back to my home page .