2. A Short Description of Triogram Modeling
Fitting Triogram ModelsIn the PostScript version of our paper, we derive some simple properties of a basis for the space G of continuous, piecewise linear functions defined over a given conforming triangulation of a polygonal region in the plane. In this case, the term ``piecewise" refers to the fact that our estimates are linear within each triangle, and are allowed to break (or hinge) along edges in the triangulation. Given G we estimate an unknown, bivariate function as a member of G. To be more precise, let be a random sample from the distribution of a random vector W, and let l(g, W), denote the log-likelihood linking the distribution of W to functions in G. Using this notation, the Triogram estimate is given by
For concrete examples in both the regression and the density estimation contexts, the reader is referred to our PostScript paper. In general, the Triogram methodology can be applied to a large class of extended linear models, which also includes generalized regression, polychotomous regression and hazard regression (Stone et al. 1995). So far in this section, we have considered applying maximum likelihood to fit a Triogram model only for a fixed triangulation (and hence a fixed space G). In the remainder of the section, we describe a stepwise approach to Triogram model building that at each step alters an existing triangulation by adding or deleting a single vertex. A Stepwise Algorithm
The adaptive Triogram procedure starts with an initial triangulation and a maximum likelihood estimate in the associated space G. In many applications a natural initial configuration may be determined from a priori knowledge about the function we are estimating. For situations in which the initial triangulation is not so clearly defined, we provide several choices in our Triogram software: the user can choose between the smallest triangle, the smallest equilateral triangle, and the smallest axis-oriented rectangle that contain all the data with a possible magnification factor to avoid boundary problems. Note that only the procedures for determining the first two of these triangulations are invariant under affine transformations of the data. In general, if the initial model is not sufficiently flexible to capture the major features of the data, we enrich G by stepwise refinements to the triangles in the starting mesh.
The most natural way to proceed from one step to the next in the stepwise addition procedure is to introduce a new vertex into the existing triangulation, thereby adding one new basis function to the existing spline space. This operation requires a rule for connecting this point to the vertices in the current triangulation so that the new mesh is also a conforming triangulation. Essentially, we have three options for vertex addition: we can place a new vertex on either a boundary or an interior edge, splitting the edge, or we an add a point to the interior of one of the existing triangles (see Figure 4 and equation 11 in the PostScript version of our paper for more details.) At each step in the addition process, we select from the set of candidate vertices the point that maximizes the Rao (score) statistic. Stability considerations may dictate that we do not consider for addition vertices in areas where there is little data. Moreover, we have found it useful to avoid creating triangles having one or two very small angles. Restrictions such as these are easily incorporated into the stepwise addition procedure.
There are two possible strategies for reducing the dimension of an existing piecewise linear spline space. In each case, we enforce the condition that functions in the space be continuously differentiable across a given edge in the existing triangulation. Observe that a continuous, piecewise linear function has continuous partial derivatives across an edge if and only if the function is linear on the union of the two triangles that share the edge. Using the correspondence between vertices and basis functions alluded to earlier, we can show that the subspace of spline functions satisfying this condition is characterized by a simple linear constraint on the coefficients of the basis functions in G. In many cases, enforcing continuity of the first partial derivatives across an edge is equivalent to removing a single vertex from the triangulation (again, the reader is referred to our PostScript paper for details). Therefore, one procedure for stepwise deletion in the bivariate context involves using the Wald statistic to choose between continuity constraints across edges that simplify the underlying triangulation. An alternative deletion procedure is somewhat more aggressive and involves choosing from among all the continuity constraints, regardless of how the edge is positioned relative to the other edges in the partition. The important distinction between these two procedures is that only in the first case are we actually guaranteed that the structure of the existing triangulation is simplified at each step.
During the combination of stepwise addition and stepwise deletion, we generate a sequence of models. When the unknown function we are estimating is a log-density function or a generalized regression function, we choose between these models using BIC (Bayesian information criterion). In the regression context we use a slightly modified version of GCV (generalized cross validation) to identify the best model from our chain. In the interactive Java applets on the next few pages, we open a separate window to track either the BIC or GCV score for the various models we encounter during addition and deletion.
You are currently reading Section 2, a short review of Triogram models.
- 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 .