:2BComputers and the Social Sciences. :2B(Comput. Soc. Sci.) [ Special double issue on statistical packages] :3BParadigm Press, P. O. Box 1057, Osprey, FL 33559-1057, 2, 1/2 (Jan.-June 1986), ISSN 0748-9269. In this issue, four statistical packages for microcomputers are reviewed, and criteria for selection of a package for use in social science research and teaching are discussed. The goals of the issue are (1) to offer a new methodology for reviews of statistical software and (2) to inform researchers and instructors about the packages.
The editors reviewed only those few packages that could handle a wide variety of data analysis tasks: BMDPC 1985, P-STAT-8 1.18, SPSS/PC 1.1, and SYSTAT 2.0. A full-featured version of SAS was not yet available for microcomputers, and so was not included in the ratings. (Note:Since the reviews were written, SAS and SPSS/PC+ were released, and SYSTAT 2.0 became version 3.0. In addition, other full-featured packages, such as STATA, have also become available.)
A major objective was to avoid both the superficiality of reviews that cover a wide range of techniques and the narrowness of in-depth treatment of only one or two packages. Instead, four professional social scientists used the following question as their key review criterion: “How well could I use this package for my research and instruction?” Reviewers learned all four packages, and then used them to perform typical research tasks. Each reviewer then evaluated all four packages with respect to a single, assigned functional area. The areas were (1) regression and time series, (2) scaling and classification, (3) file handling and data management, and (4) instruction.
The issue has three sections: (1) the four featured area reviews, (2) responses of the software vendors to the reviews, and (3) comments by others on issues in statistical package design and selection.
Grant Blank’s lead paper, “Reviewing statistical software on microcomputers: methods and vision,” sets forth the issue’s review strategy. Statistical capabilities should include not only the usual descriptive techniques, but also those appropriate for good prediction, time series, and classification. He points out the need for being able to handle combinations of continuous and categorical independent and dependent variables. Other desirable characteristics include extensibility, flexible data management capabilities, integrated input and output between package modules, adequate speed and numerical accuracy, good vendor support, a simple user interface, and both interactive and batch capabilities. He concludes with a plea for a continuing dialogue between social science users of software and designers and vendors.
The first of the featured reviews, “A review of regression, correlation, and [time series?] capabilities in four full-featured microcomputer statistical packages,” by Robert Nash Parker, employs a benchmark approach. Packages were compared in terms of how long (in “real time”) it took to accomplish each task; disk space taken up by internal system files; and in terms of user friendliness, adequacy of documentation, and general usefulness. Tables are presented showing time required for each package to accomplish each task on several data sets. Another table presents the strong and weak features of each package. According to Parker, “SYSTAT appears to have a significant edge overall on the other packages.”
“Scaling and classification with microcomputers,” by Edward E. Brent, Jr. and James D. Campbell, examines the programs with respect to “how well they perform three common analysis procedures: cluster, factor, and smallest space analysis.” These authors chose to focus on the packages’ ability to perform publication-quality work (summary indices and graphic and tabular displays).
Brent and Campbell describe the computation and display features required for each of the three procedures, give examples of desirable features, and describe their benchmarking procedure. Then, they review each package and offer guidelines for selecting among packages. Suggestions for the improvement of scaling and classification features in general are included. Detailed feature comparison tables are presented. The authors found advantages to each program, depending on the type of processing desired by the user.
Andrew A. Norton’s paper, “Data management capabilities of full-featured microcomputer statistical systems” sought to “. . . identify those operations that are crucial to many statistical analyses but not supported by some of those systems.” The paper includes a features comparison of the data management capabilities of the four systems and a glossary of the terminology used in the paper. Each system is then evaluated as a whole. The comparisons include: structured flow of control; arrays of variables; within-record statistical functions; file merging; file restructuring; input and output; handling of missing values; macro code inclusion; and adding, deleting, and modifying individual records. The author found advantages and limitations for each of the systems.
The fourth review is Richard J. Peterson’s paper, “Teaching introductory survey analysis with full-featured microcomputer statistical packages.” The teacher’s evaluation criteria differ somewhat from those of a research user. The criteria used were: appropriateness of statistical procedures, control language “transparency,” existence of simple data entry and data modification capabilities, good error messages and help files, and a readable and useful manual. The paper explains the evaluation criteria and then applies each criterion to the four programs. SPSS/PC is clearly preferred. Peterson correctly asserts that full-feature research packages are the ones that should be used in education, not special cut-down packages, a question which needs more public discussion.
The vendors’ responses include explanations for design decisions, planned enhancements, and information about relevant features ignored by the reviewers.
The final eight papers in the volume discuss packages and package selection in general.
“Selection of statistical software for micros,” by Harry V. Roberts, compares interactive and batch program architectures, points out the need for various nonstatistical program features, and lists desirable statistical capabilities.
“Comments on what researchers need in microcomputer statistical software packages,” by William J. Kennedy, emphasizes numerically sound algorithms, adequate interfaces between various programs, and high-quality interactive graphics.
“The need for exploratory data analysis software,” by Neil W. Polhemus, compares the reviewed packages with several others. He points out the need to be able to read commercial DBMS and spreadsheet files, emphasizes the need for flexibility and speed if programs are to be interactive, and lists desirable graphics capabilities.
“Discriminating Statistical Software,” by Leland Wilkinson, provides interesting insights into the statistical reasoning behind the design decisions in SYSTAT.
“Statistical packages for microcomputers: what next?,” by Lutz Erbring, is critical of current packages for using mainframe (batch) software architecture. He argues that even the reviews in this issue miss this. What is needed, he says, is to make use of the types of architectures exemplified by state-of-the-art commercial programs. He is correct.
Paul Velleman discusses, in “On evaluating statistics software,” how to make informed decisions about the acquisition and use of a statistical package. He provides a checklist of questions to be answered when considering a package, useful to anyone contemplating acquisition.
“Statistical micro-computing. Running Light. . . ,” by Gary M. Grandon, presents the perspective of the developer of another package.
“The micro revolution in social science computing,” by Jacob Cohen, asserts that “Social scientists are a heterogeneous group, particularly so in regard to the amount and variety of quantitative methods they employ. It follows that no single package is best for all researchers. . . .”
The issue’s biggest limitation is that some important topics have been given too little attention by the area reviewers: wild-code and consistency check (read data integrity]) capabilities; provisions for variable names, variable labels, and value labels; variable generation; and ability to compute statistics based on case weights. Yet, the papers still provide an excellent review of most of the features of these four important statistical packages.
Do these packages really provide adequate statistical analysis and data management capabilities on a microcomputer? Yes, provided it has a 20 megabyte hard disk, and provided the proposed uses involve only relatively simple rectangular data-sets and the statistics available from the chosen package. Researchers with complex data should still plan to augment their package with a good relational database system. Those who wish to use statistics other than those provided will need to turn to a matrix package, such as GAUSS.
A statistical package’s limitations strongly constrain the variety of research designs its user can employ. This is due simply to the high costs of trying to use data structures the package cannot handle easily, or statistics it doesn’t produce. The issue’s editor has done a valuable service in initiating discussion of these packages’ absolute level of adequacy, as well as relative comparisons between them. Package developers should pay close attention, and prospective package users of any statistical package should certainly include the entire issue on their reading list.
