The typically exploratory and experimental nature of interactive statistical analysis ensures that its structure is far wider and deeper. It is difficult to imagine that we could a priori determine the length of the longest possible path in an analysis let alone identify all conceivable actions an analyst might wish to take at any given step. This open-ended nature means that any computational environment which presupposes a closed structure (like that of tic-tac-toe) will eventually overly constrain the analyst.
Designers of statistical software often make the error that the closed world approach is sufficient. Examination of the Interface proceedings from the 1970s will reveal that command line systems began this way but soon had statisticians hitting the edges of the system's possibilities. To quote but one source:
``Even the statistician's operating language, context and syntax, became formed from the names of available programs and functions. In order to regain his individuality, it became necessary for the thinking statistician to teach computers to do his wishes ...That is, he had to learn to program or hire a programmer.'' ...Guthrie (1975, pp 8-9)In fairly short order, `macro' capabilities which permitted programming were added to most command language systems.
Similarly, direct manipulation interfaces are often first built for either wide and shallow structures or narrow and deep ones. A simple example of the former would be a pop-up menu from which to choose a singe item. A more complex example would be a stand-alone interactive graphic - such as a brushable scatterplot matrix or a grand tour type dynamic scatterplot. An example of a narrow and deep structure would be the sequential prompting for information preparatory to proper execution of some task.
Direct manipulation interfaces can be made to work very well for either kind of structure (wide and shallow or narrow and deep). But, as with the early command line interfaces, if this is all that the user can do then the software soon becomes overly restrictive.
When it becomes apparent that the interface does not meet perceived needs, or perhaps in response to users requests for certain functionality, designers often give into creeping featurism, the first of D.A. Norman's (1988) ``two deadly temptations''. The result is that `features' are added to the existing design which, if not carefully done, quickly add to its complexity (Norman, 1988, believes that the complexity grows with the square of the number of features). The resulting more complicated design can be rendered less useful than the original simpler design.
Creeping featurism is to be avoided. If it cannot, the only solution is in a careful organization of the design: to modularize, to divide and conquer. Gratuitous addition of features adds unneeded complexity.
Sometimes complexity is introduced intentionally - as if it were desirable itself. This is Norman's second deadly temptation - the worshipping of false images. In a graphical user interface we might see all features made available at once with gratuitous use of multiple colours, buttons, knobs, etc. The false image is that a complex interface implies technical sophistication, when the reality is that complexity leads to confusion. Again the best advice is avoidance - keep things as simple as possible.
Interactive statistical analysis is substantively more difficult than the game presented at the beginning of Section 2. Yet it is hoped that software can be designed which, like the tic-tac-toe representation of Section 2, will considerably simplify the interaction between user and system. Complexity is to be managed by the system so that the user is freed to concentrate on the analysis. Achieving this is no easy task.
A mixed strategy of direct manipulation and keyboard commands is desirable. Moreover, because some tasks will always be outside the existing design, programmatic control needs to be available to the user. Programmatic control can come from either the keyboard or from direct manipulation - an early statistical example of some programming functionality in a graphical user interface can be seen in Desvignes and Oldford (1988) - although the keyboard is likely to remain the programming input device of choice.
Having the user able to concentrate on the analysis implies that interaction with the system needs to be in terms which are familiar to an analyst as opposed to a programmer. Statistical analysis concepts must form the fabric by which direct manipulation interfaces and text-based programming are seamlessly integrated. This is the fundamental design challenge for interactive statistical systems.