5.0 DINDE and statisical strategy. In Oldford and Peters (1986a), we argue that complete software representations are not likely to be constructed for many of the analysis strategies used by practicing statisticians. Statistical strategies that show the most promise are those that both depend little upon correctly interpreting the subject-matter context surrounding the problem, and are applicable to only a relatively small number of problems. Strategies that most easily fit this description are those we call low-level low-context strategies; an example is the strategy employed to assess the number of near dependencies in a regressor matrix. "Low-level" means that the strategy typically appears as a low-level operation in some wider analysis; "Low-context" indicates that the strategy relies little upon the subject-matter context for its correct application. Typically, these strategies are applied in a narrow domain of well-defined problems yet depend on some judgement in their application. Representations of these kinds of strategies appear in DINDE as methods of the objects for which they are most relevant. If, for instance, a particular strategy to check the correctness of the functional form in a bivariate regression is desired, then it would appear on the BivariateRegression class as the method CheckFunctionalForm. The method could invoke a procedure or create a more detailed object that contained the necessary tools and information to go about applying the strategy. Higher-level strategies are of more interest; consider for example, the strategy employed by an experienced statistician using multiple regression tools to describe the relationships between two sets of variables. At this level, effective strategies typically depend critically upon correct interpretation of the background context, making it difficult to construct complete representations for them. Therefore, we propose that the context-dependent elements of the strategy be removed; the telephone conversation between the experienced and inexperienced statisticians provides a simple metaphor for what might remain (see Section 4.3). To achieve this, the strategy is first broken into its constituent parts (Pregibon, 1985, has outlined rules for doing this). These components could be connected together and provided to the user as a single overall strategy œ as in REX for bivariate regression (Pregibon and Gale, 1984) œ but we prefer that they be made directly available to the user and that the connections only be pointed out as often being appropriate. The decision to consider a component or make a connection is then left entirely to the user: at this level of strategy the analyst is more competent at interpreting the relevance of the various components than is a software representation. In DINDE, these component pieces are represented as distinct classes. Interconnections are indicated by the methods attached to the class: by their presence and by their application. Data-independent Suggestions describe a strategy to choose the next component; data-dependent analogues would be implementations of some of these small strategies. The example in Section 3 shows some of the components and interconnections of a strategy for bivariate regression analysis. For high-level strategy, visually presenting the parts and connections makes these aspects of the strategy apparent. The analyst sees the analysis develop, parts of it that are according to the canned high-level strategy, and, more importantly, parts of it that are according to the analyst's personal strategy. The analyst may then be tempted to think more about the strategy he or she employs. (Various network operations, AnalysisPaths, and user-added commentary aid the development of the latter.) Further, by collapsing pieces of the network into SubMaps, that can be named and that can have user-added commentary, the analyst can define the levels of abstraction appropriate for the problem and strategy. These various levels of abstraction help the analyst to both better manage the analysis and better illuminate the personal strategy being used. The analysis, then, is seen to be an intertwining of two strategies: that of the analyst and that of the developer. (LIST ((PAGE NIL (PAPERSIZE Letter FOLIOINFO (ARABIC ) STARTINGPAGE# 25) (0 0 612 792) ((FOLIO NIL (PARALOOKS (QUAD CENTERED) CHARLOOKS (SUPERSCRIPT 0 INVISIBLE OFF SELECTPOINT OFF PROTECTED OFF SIZE 10 FAMILY CLASSIC OVERLINE OFF STRIKEOUT OFF UNDERLINE OFF EXPANSION REGULAR SLOPE REGULAR WEIGHT MEDIUM INVERTED OFF USERINFO NIL STYLE NIL) FORMATINFO (ARABIC )) (162 36 288 36) NIL) (HEADING NIL (HEADINGTYPE DINDE) (120 3600 492 36) NIL) (TEXT NIL NIL (72 72 468 648) NIL))) (PAGE NIL (PAPERSIZE Letter FOLIOINFO (ARABIC )) (0 0 612 792) ((FOLIO NIL (PARALOOKS (QUAD CENTERED) CHARLOOKS (SUPERSCRIPT 0 INVISIBLE OFF SELECTPOINT OFF PROTECTED OFF SIZE 10 FAMILY CLASSIC OVERLINE OFF STRIKEOUT OFF UNDERLINE OFF EXPANSION REGULAR SLOPE REGULAR WEIGHT MEDIUM INVERTED OFF USERINFO NIL STYLE NIL) FORMATINFO (ARABIC )) (162 36 288 36) NIL) (HEADING NIL (HEADINGTYPE DINDE) (120 3600 492 36) NIL) (TEXT NIL NIL (72 72 468 648) NIL))) (PAGE NIL (PAPERSIZE Letter FOLIOINFO (ARABIC )) (0 0 612 792) ((FOLIO NIL (PARALOOKS (QUAD CENTERED) CHARLOOKS (SUPERSCRIPT 0 INVISIBLE OFF SELECTPOINT OFF PROTECTED OFF SIZE 10 FAMILY CLASSIC OVERLINE OFF STRIKEOUT OFF UNDERLINE OFF EXPANSION REGULAR SLOPE REGULAR WEIGHT MEDIUM INVERTED OFF USERINFO NIL STYLE NIL) FORMATINFO (ARABIC )) (162 36 288 36) NIL) (HEADING NIL (HEADINGTYPE DINDE) (120 3600 492 36) NIL) (TEXT NIL NIL (72 72 468 648) NIL)))))(MODERN CLASSIC CLASSIC #ž0žË sž«zº