Implementing the strategy

The nominal entry point for regression analysis is the topmost node, the Regression Hub. Entry means that a dataset has been identified and that the intention to pursue a regression analysis on it declared. This produces the interactive window seen in Figure 2.

  

Figure 2: Regression Hub

It displays some information on the data (variate names, case names, dataset name) and some buttons at the bottom. Each of the first three buttons lead to one of the next nodes in the strategy -- here to the Linear Modelling Hub, to Summary Stats, or to Regression plots respectively. The plotting buttons prompt the user to choose among a wide variety of plots available for any dataset. Each column refers to one- two- and multiple-variate plots; the second row (which would on a screen read ``Plot by'') produces a table of plot whose cells are arranged according to the values of one or two other variates.

The construction of the regression-hub illustrates the modularity of the code. Two major pieces of the display are available directly from the underlying dataset. Calling the generic function display on the dataset without :signposts? returns a compound-view displaying the dataset information shown in Figure 2. A separate call to the generic function signposts on the dataset produces the collection of plot buttons -- this is why no separate display is produced corresponding to the stratefic node Plots of Data in Figure 1. With the data display and the data signposts in hand, the designer of the regression-hub need only construct the three buttons leading to new nodes and lay out all of these in an appropriate display as in Figure 2. The viewed-object of the resulting layout is a regression-hub object and conversely the display of Figure 2 is the result of calling display on a regression-hub object -- the collection of buttons are its signposts.

Upon entry, each node in the strategy network produces a new window containing an interactive compound view that displays relevant information. For most nodes, the compound view offers the user the possibility of changing the display to show other relevant features of the underlying information. Figure 3

  

Figure 3: A ``regression plot'' showing a cubic added to the plot. Selecting the curve (with ``CTRL'' pressed as well) pops up a menu that allows the user to display the underlying fit-object with signposts.

shows the display associated with the Regression plots node. This would appear if, as shown in Figure 2, the ``Regression Plot'' button was selected from a regression-hub display. Various polynomial fits can be added to the display (through the buttons) and the variates displayed can be changed (through pop-up menus available by direct mouse interaction with the plot itself).

Note again the loose and coarse nature of the strategy. The node that displays a fit and which leads to diagnostics and inference can be reached from a signost in the Linear Modelling Hub. But it can also be reached through any view that has a fit-object as its viewed-object. For example selecting the curve in Figure , the user can choose to have the underlying fitted cubic displayed as if it had been reached through a Linear Modelling Hub. Viewed objects, the strategic functions display and signposts are critically important to designiong and creating a flexible strategy.

Some node displays, like the Regression Hub, contain buttons which lead directly to the next nodes in the strategy -- hence the names hub for the display, and signposts for buttons of this type. Nodes which have no signposts are said to be terminal in the strategy. Terminal nodes may be extremely interactive as, for example, in the case of the Regression plots node; they just do not provide signposts that lead to new nodes. Any number of nodes may be displayed in separate windows at the same time.

The structure of each node breaks into three distinct descriptive pieces: initial information displayed, activities available, and signposts. Aside from the Regression Hub and the Regression Plots node already discussed, the characteristics of the remaining nodes from Figure 1, as implemented by the students, are described in the appendix.