Concluding remarks

Recall the game with which we began and its ultimate representation as the more familiar tic-tac-toe. This design of the game was successful because it followed good design principles. It coded the constraints into the representation following a natural mapping of constraints to layouts which took advantage of our natural perceptual abilities. It considerably reduced the mental burden on the players and forced a reduction in the errors a player could make. The players were now free to focus on the essence of the game.

Interactive statistical analysis is also hard, considerably harder than the game of tic-tac-toe (whatever the representation). It is desirable to develop and to implement a representation which, like the game, allowed the user more time to concentrate on achieving the objectives of the analysis.

This requires attention to the design and implementation of software models of the statistical concepts which form our common knowledge. Both direct manipulation and programmatic interfaces would benefit substantially from such modelling and the common modelling would allow graphical user interfaces and text-based interfaces to be freely mixed.

To some degree, Quail is designed to provide a programming environment where ideas on modelling statistical concepts and direct manipulation interfaces can be explored. It provides one common foundation for this exploration and has been used with some success by senior undergraduate and graduate students in the development of models and interfaces (e.g. see Oldford, 1997).

So the debate is not one between command language and direct manipulation interfaces, but rather one of appropriate design of the underlying software. The seeming gulf between the two approaches is there because of the lack of a common foundation.

The goal should now be to determine the best models for the selected statistical concepts. Because design is evolutionary this will likely require much debate over the relative merits of different models. The more researchers there are involved in this, the more likely that good models will result.

Prototyping each design will be essential and we will need environments which permit rapid prototyping (Common Lisp and Quail provide one such environment). Nevertheless the model designs will need to be communicated and assessed in an implementation independent way.

The well known principles of design illustrated above provide useful starting points by which designs can be assessed but they are not sufficient. Assessment will also need to depend on how well the model represents the statistical concept and so far there is little direct experience in making that assessment - largely because it has not often been a stated goal.