AN INTELLIGENT STATISTICAL PROBLEM-SOLVING INTERFACE

Abstract

Use of currently available statistical analysis software by naive users is susceptible to errors because such software only provide the functions to perform the numerical calculations but do not advise users on pitfalls if any of the assumptions underlying an analysis is violated. Moreover, users need to learn the software's language. A language independent intelligent statistical tool interface "shell" is proposed that will, given knowledge about statistical analysis strategies and about writing programs in the underlying statistical tool, guide users on the correct conduct of statistical analyses as well as automatically generating programs in the tool language to do so. The thesis looks into the design of the "shell", how it is used to capture statistical knowledge and how it will generate programs automatically. Intelligent statistical tool interfaces have been studied by various researchers. This thesis is a close parallel to work done in the AT&T Bell Laboratories, namely, Student [Gale86b], although the idea was conceptualized separately. REX [GaPr82, Gale86a] from which Student developed was a source of inspiration. The work in the thesis differs from Student on the following points: 1. diagnostic tests to check assumption violations are treated at the same level as the main analyses, that is, they are treated as calls to other analyses from the main analysis; 2. acquisition of statistical analysis knowledge is separated into two stages, the acquisition of language-independent analysis strategies from an expert analyst and the acquisition of program segments to generate programs in an underlying statistical analysis tool from -a tool expert; and 3. an object-oriented programming method is used to capture knowledge about and to generate programs in the underlying statistical analysis tool language. The interface generator may be run in three modes: LS 1, to capture statistical analysis strategies; LS2, to capture programming language knowledge; and ESSA, the resulting intelligent statistical problem-solving interface. The term "analysis strategy" is used in a restricted sense, as in Student, and is interpreted to be the complete set of tests and calculations to conduct a safe analysis and calculate the model parameters for a statistical analysis, as well as the order to carry them out and the actions to take when assumptions are violated. In this sense, each analysis has only one strategy that is relatively independent of application domain influences and is dependent only on features in the dataset(s). Each strategy may be conceived as having five stages: identification of the analysis, checking constraints on the dataset(s), performing the analysis computations, checking for assumption violations and fixing violations, if any. Knowledge about strategies is imparted by describing them one at a time to LS 1. Internally the strategy is stored in an analysis frame with pointers to input data frames, computation frames, and assumption violation checking frames. An expert in a statistical tool then runs through each analysis and provides program segments corresponding to the computations required by the strategy. The result is an intelligent statistical analysis assistant, ESSA, that guides users through the correct conduct of analyses. ESSA obtains from a user specifications of the desired analysis and the datasets to be analyzed which are passed in a computation frame to a program generator. The program generator takes the computation frame and the program segments provided by the tool expert, and outputs a program instance in the target tool language.