The subscriber lists for CAAN are being reorganized. If you want to continue to receive this newsletter please mail in your name and address. A form is provided at the end of this issue.
We regret the delay in getting this issue of CAAN to you. It was necessitated by uncertainties in the finances, the work arrangements, and the work loads of your editors. The organization of CAAN has changed somewhat. Lee Sailer has withdrawn as co- editor. I and many others thank him for his pioneering contributions to CAAN which include its name. His present location is at the School of Business, Behrend College, Erie, PA.
With support from the College of Arts and Sciences of Oakland University we will be able to continue some free subscriptions. Subscription moneys are voluntary, but urgently needed. We will mail free issues only when the operating funds are above a certain critical level. See the subscription information at the end of this issue.
This issue contains a long, informative article on statistical packages for IBM-PC and compatible microcomputers, so news items have been held to a minimum. More news will appear in the next issue. Please continue to send in news and articles about your computer and methods activities. So far we have been able to publish almost everything, and would rather enlarge the magazine than leave out something of value.
Since the computer world seems to be headed toward an era of throw-away hardware and increasingly valuable software, and since methodological algorithms are the basis of computer use in all branches of anthropology, we would like CAAN to be a place where methods can be discussed, whether or not they have already been implemented on computers. The potential for computer implementation -- and it is a wide one -- is all that should count. Write us with designs for programs or the parameters that you think should be important in designing them. Refer to the section "How to Submit Materials to CAAN at the end of this issue.
Last winter I embarked on a search for a microcomputer statistics package that would serve research and teaching purposes. The programs selected were to be used in conjunction with a soon- to-be completed microcomputer laboratory stocked with ten Zenith 150 microcomputers, each equipped with two floppy disk drives and 320 kilobytes of RAM. This turned out to be a formidable task since there are over 175 commercial statistics packages available for personal computers. These programs perform a wide variety of statistical functions and run on many different types of small computers.
In choosing packages to review, I adhered, insofar as possible, to the following criteria: (a) the programs had to run on an IBM-PC or close compatible; (b) the packages had to be floppy disk based (i.e. no hard disk required); (c) they could not require more than 320K of RAM; (d) they had to perform all standard univariate and bivariate statistics; (e) some non-parametric statistics (especially chi-square and contingency tables) were required; (f) support of multivariate statistics including ordinary multiple regression was essential. I imposed one additional requirement. The evaluation had to be based on a full working copy of the program complete with its manual, not on a self- serving demonstrator put together by the software manufacturer.
I reviewed most of the published literature on microcomputer statistical packages and I poured through a great deal of advertising literature before selecting the packages for actual review. I eliminated many packages from consideration because they were not for the right machine, they were too limited in what they did, too focused on a narrow audience (e.g. econometric forecasting), too hardware specific (e.g. might only run on an IBM-PC with BASIC in ROM), or too hardware extensive (e.g. 512 kilobytes of RAM and a 10 megabyte hard disk).
My choice narrowed down to twelve packages. These were AbStat, Sigstat, MSUSTAT, Wallonic Statpac, Statpro, Statgraphics, SYSTAT, Number Cruncher, ELF, Northwest Analytical Statpak, BIOM- PC, and Crisp. I added three other programs (BMDP- PC, SPSS-PC, and Epistat) even though they did not meet one or more of the conditions I imposed when I began. I was able to obtain review copies of all but MSUSTAT, Abstat, Crisp, BMD-PC, and SPSS-PC; Wallonic Statpac arrived too late to be included in this review. I felt that SPSS-PC was important enough to try to acquire a copy by other means. I was able to get access to a copy that another Department here at Portland State had already purchased. I did not review the remaining programs. Table 1 lists the names and addresses of the publishers of all software reviewed. Specific details about the capabilities of each package are contained in Table 2. This latter table indicates a wide variance in price, range of functions, and utility. I concluded very quickly that no one package could simultaneously satisfy both research and teaching needs.
Programs that were best for research proved to be too complicated for beginning students; those best for students proved too constraining for research.My evaluation of the programs resulted in a letter "grade" based on a set of key factors. Any user's final evaluation depends on a complex interplay of wants and needs. Those factors I considered to be most critical were: (a) the programming language; (b) ease of use (including how well-written, accurate, and informative the manual was); (c) range of statistical functions supported; (d) numerical accuracy of the algorithms as measured by the Longley (1967) data set and benchmark described in Lachenbuch (1983) and Pease and LePage(1984); (e) visual presence and quality of graphics; (f) number of variables, cases and groups usable within various programs; (g) the flexibility of the programs (e.g., ability to create command files, ability to choose between menus and individual programs, availability of source code); (h) ease of importing data from a variety of foreign sources (e.g. mainframes, spreadsheets, database programs, other statistics programs); and finally (i) the presence or absence of copy protection. Protected programs prevent the legitimate user from making archival copies of a program already purchased.
I found that programs which supported the Intel 8087 Math Co-Processor were generally better than programs not able to access this chip. The math processor speeds up all real number calculations by as much as a factor of 100, and, because of its 80- bit temporary real format, it is capable of 16 to 18 digits of precision. I also found that programs written in FORTRAN or compiled BASIC were faster, more accurate, and more trouble-free than programs written in Pascal (either IBM or Turbo), or APL.
Table 2 summarizes the essential features of all the programs I reviewed. The information came mainly from the program manuals; occasionally I used other sources. When the manuals were poorly written and without indexes, the only way to determine a program's capabilities was by trial and error. I used question marks to indicate areas where I could not obtain verifiable information by any means.
The tables should not be used as the sole basis for making a decision about purchasing, or not purchasing, a program. No program should be purchased without first using it with real data (preferably data already analyzed with a program known to give "correct" results). In this way the potential buyer can insure that the program does what he wants, does what he expects, and gives reasonably accurate answers. A potential buyer may also wish to consult other reviews of these packages (see Lee and Cohen, 1985; Vanderbok, 1985; Fridlund, 1985; Carpenter, Deloria, and Morganstein, 1984; Lachenbruch, 1983; Pease and Lepage, 1984; Schervish, 1985; and Lubowsky, 1985). I have not been able to represent many of the idiosyncrasies, bugs, and annoyances of the various programs in tabular form. The following comments reflect my thoughts on the strengths and weaknesses, as well as the unique features, of each program I reviewed.
SYSTAT is extraordinarily fast and accurate. It comes on 6 diskettes and contains 11 different modules. The modules are not integrated in any special way; however each program operates on a common data set created under the DATA module.
SYSTAT is not a program for the statistically uninitiated. It requires a sophisticated background in statistics to be able to take advantage of its power. The Multivariate General Linear Hypothesis (MGLH) module is the workhorse of SYSTAT and will be the most oft-accessed program. All regression analyses, ANoVA, t- and F-tests, analysis of covariance, factor analysis, discriminant analysis, and canonical correlation analysis are done within the MGLH.A user without a solid grounding in research design, matrix algebra, and general linear models will find this module very difficult to use. The manual gives examples of common statistical designs using this module. Even so, the statistically naive will not be able to make much sense from the output, or be able to tap this module's raw power. By contrast, a worker with the above skills who is comfortable with statistical techniques described in Winer (1971) will be able to analyze almost any type of linear model with the MGLH program.SYSTAT has several features that make it desirable for research uses. First, the author of the program has spent his professional life as a statistician with a major research interest in accuracy of computational statistical algorithms. As a result, SYSTAT sports the most accurate results of any program tested (including mainframe SPSS). Second, the program incorporates a mini-BASIC interpreter (written in FORTRAN) in its DATA module. This permits the user to code virtually any transformation, restructure any data set, or do many different kinds of sampling, standardization, and ranking. If what you want done is codeable in BASIC, then you ought to be able to code it in SYSTAT. Third, the program accepts command files to perform repetitive tasks. Thus a researcher who wishes to do analysis of covariance to compare regression slopes for 20 pairs of variables across 30 groups, for example, could simply code the necessary model statements into a command file, "submit" the command file to the statistics procedure, have the output directed to a disk file or line printer, and leave the program unattended for several hours.
The program is not without faults: (1) it is not yet possible to perform a true discriminant function analysis with computation of discriminant scores, structure matrices, etc., such as is obtained in BMDP or SPSS (or SIGSTAT, see below); (2) while the number of cases is disk limited, there are no indications of limits on variables, or numbers of groups for certain procedures; by trial and error I determined that with grouped data no more than 12 dependent variables are allowed with 6 groups; with 23 groups, only 1 dependent variable is allowed (to get around this limitation there is a "large MGLH" available for an extra $25.00); (3) any analysis that uses grouped data forces the user to create a numerical grouping variable because alpha grouping variables are not permitted; groups are selected sequentially from the file and one cannot, without restructuring the data set, select a numerically non-adjacent subset of the groups for analysis; (4) the principal components routine based on covariance matrices yields eigenvectors that have not been standardized to unit length unless the components are rotated (but Johnson and Wichern, 1982 provide a simple way of going from unstandardized to standardized eigenvectors); (5) the clustering module provides no agglomeration schedule; instead it provides only the dendrogram.The dendrogram is done in character graphics and lacks sufficient resolution to establish whether ties are ties or merely a lack of resolution in the output. It is also not evident from the program or the manual whether the clusters are variable- or pair-group clusters. The manual doesn't specify this and without the agglomeration schedule there is no way to determine this directly. Finally, (6) there is no simple one-sample or two-sample t-test available. The only way to get simple t-scores is to run a one-way ANOVA. The t-value can be obtained as the square root of the reported F value.SYSTAT's manual is well-written but very terse. It contains examples of the commands necessary to invoke all essential statistics. It is also liberally sprinkled with the author's justified warnings about the misuse of statistics. Each chapter also includes an excellent bibliography of essential references for all procedures used in SYSTAT.
I reviewed a version (2.0) of SYSTAT that was copy-protected with Prolok. The company wisely has decided to discontinue copy-protection beginning with version 2.1 (due out in 1985). The copy-protection scheme provoked many complaints, including one from me. It also may have been responsible for some interesting problems when I used SYSTAT in conjunction with Borland's Sidekick and/or IBM's TopView.
All things considered, SYSTAT has no real competition. Its range of functions, speed, accuracy, and even its price make it, in my opinion, the best statistics package currently available for IBM-type microcomputers.
Number Cruncher Statistical System is a very respectable, and modestly priced, alternative to SYSTAT. The program comes on 3 diskettes and consists of 32 programs woven together in a tightly bound system. NCSS performs all standard univariate, bivariate, and multivariate statistical procedures in a well-structured menu-driven environment. The program is outstanding for the availability of all common statistical procedures. It also offers users most of the new "robust" and "exploratory" data analysis techniques such as box plots, stem and leaf diagrams, robust regression using a variety of weighting schemes, and trimmed and biweighted means (Wonnacott and Wonnacott, 1984). It also offers a rich set of built-in data transformations. Users who do not require sophisticated control over the program, and who do not need Multidimensional scaling, cluster analysis, non-linear regression, ridge regression or analysis of covariance will find this program quite adequate.
ELF STATPAK EPISTAT NCSS SYSTAT
1 VERSION 5.2C 3.1 3.0 4.1 2.0
2 LANGUAGE PASCAL BASIC-I BASIC-I BASIC-C FORTRAN
3 COPYABLE Yes Yes Yes Yes PROLOK
4 RETAIL PRICE $395 $495 $25 $199 $495
5 ACAD DISCOUNT $350 $350 $25 $199 $495
6 TCHG VERSION No Yes No No Yes
7 MULT COPY DISC Yes Yes No >20 >10
8 SOURCE CODE No Yes Yes No No
9 SYSTEM NEEDS 192K 64K 64K 196K 256K
2 DRIVES 2 DRIVES 1 DRIVE 2 DRIVES 2 DRIVES
BASICA BASICA GRPH. BR
10 8087 SUPPORT Yes No No No Yes
11 ACCURACY NONE 4.73 N/A 6.79 8.41
12 ASCII FILES
Reads Yes Yes Yes Clumsy Yes
Write ? Yes ? Yes Yes
13 EDITOR No Yes No Yes Yes
14 IMPORT FILES
Lotus .WKS No No No No No
Dbase II .DBF No No No No No
Visicalc .DIF Yes No No No No
15 SEARCH CASES Yes Yes Yes Yes Yes
16 SORT CASES/VARS Yes Yes Yes Yes Yes
17 TRANSFORMATIONS Yes Yes Yes Yes Yes
18 COMPUTED VARS ? Yes No Yes Yes
19 VAR LABELS No No No No Yes
GRAPHICS
20 Character No Yes No No Yes
21 HiResolution No No Yes Yes No
22 Pie No No No No No
23 Stem-Leaf No No No No Yes
24 Histogram No Yes No Yes Yes
25 Box&Whisker No No No Yes Yes
26 Scatter No Yes Yes Yes Yes
27 Bar No No No No No
28 Contour No No No No Yes
29 Dendrogram No No No No Yes
30 Fourier No No No No Yes
31 Star Plot No No No No No
32 Function Plot No Yes No No No
33 3-D Scatter No No No No No
34 Hidden-Line No No No No No
35 Probability No No Yes Yes Yes
STATISTICS
36 RxC Tables 3-WAY N-WAY 3-WAY N-WA N-WAY
37 Chi-square Yes Yes Yes Yes Yes
38 Other tests No Yes Yes Yes Yes
BIOM-PC STAT- STATPRO SIGSTAT SPSS-PC
GRAPHICS
1 VERSION 1984 1.0 2.0 3.0 1.0
2 LANGUAGE FORTRAN APL PASCAL FORTRAN FORTRAN
3 COPYABLE Yes Yes Yes Yes PROLOK
4 RETAIL PRICE $75 $695 $795 $595 $795
5 ACAD DISCOUNT $75 $400 $795 $395 $795
6 TCHG VERSION No No No No No
7 MULT COPY DISC No >4 >4 Yes >2
8 SOURCE CODE Yes No No No No
9 SYSTEM NEEDS 256K 320K 192K 256K 320K
2 DRV 2 DRV 2 DRV 2DRV 10 MB
GRPH. BRD GRPH. BRD HARDDISK
10 8087 SUPPORT Yes Yes Yes Yes Yes
11 ACCURACY 6.15 7.04 5.23 6.62 8.05
12 ASCII FILES
Reads Yes Very Clumsy Yes Yes Yes
Write Yes Yes Yes Yes Yes
13 EDITOR No Yes Yes No Yes
14 IMPORT FILES
Lotus .WKS No Yes Yes No No
Dbase II.DBF No No Yes No No
Visicalc.DIF No Yes Yes No No
15 SEARCH CASES No Yes Yes No Yes
16 SORT CASES/V No Yes Yes No Yes
17 TRANSFOR- Yes Yes Yes Yes Yes
MATIONS
18 COMPUTED No Yes Yes No Yes
VARS
19 VAR LABELS Yes Yes Yes Yes Yes
GRAPHICS
20 Character Yes No No Yes Yes
21 HiResolution No Yes Yes No No
22 Pie No Yes Yes No No
23 Stem-Leaf No Yes Yes No No
24 Histogram Yes Yes Yes Yes Yes
25 Box&Whisker No Yes Yes Yes No
26 Scatter Yes Yes Yes No Yes
27 Bar No Yes No No Yes
28 Contour No Yes Yes No Yes
29 Dendrogram No No Yes No No
30 Fourier No Yes Yes No No
31 Star Plot No Yes Yes No Yes
32 Function Plot No Yes No No No
33 3-D Scatter No Yes No No No
34 Hidden-Line No Yes No No No
35 Probability No Yes Yes No Yes
STATISTICS
36 RxC Tables 3-WAY N-WAY N-WAY N-WAY 10-WAY
37 Chi-square Yes Yes Yes Yes Yes
38 Other tests Yes Yes Yes Yes Yes
ELF STATPAK EPISTAT NCSS SYSTAT
NON PARAMETRIC
39 Sign Test No No No Yes Yes
40 NP Correlation No No No Yes No
41 Runs Tests No No No Yes No
42 Mann-Whitney U No Yes No Yes Yes
43 Kruskal Wallace No Yes No Yes Yes
44 Signed Ranks No Yes No No Yes
45 2-Way ANoVA No Yes No Yes Yes
46 Kolm.-Smirnov No Yes No No Yes
47 Rank Correl No Yes Yes Yes Yes
48 Log Linear Mod No No No No Yes
49 Other No Yes Yes No No
CORRELATION
50 Pearson Yes Yes Yes Yes Yes
51 Gamma No No No No Yes
52 Tau No No No Yes Yes
53 MU2 No No No No Yes
54 Missing Values Yes Yes No Yes Yes
55 Autocorrelation Yes Yes No Yes Yes
T-TESTS
56 Paired Yes Yes No Yes Yes
57 Groups Yes Yes No Yes Yes
REGRESSION
58 Linear Yes Yes Yes Yes Yes
59 Non-Linear No No No No Yes
60 Polynomial No Yes No No Yes
61 Multiple No Yes No Yes Yes
62 Stepwise Yes Yes No Yes Yes
63 Multiple Forw. Yes Yes No No No
64 Multiple Back. Yes Yes No No No
65 Harm.& Periodic No Yes No No Yes
66 Ridge No No No No POSSIBLE
67 Leaps&Bounds No No No No No
68 Residuals Yes Yes No Yes Yes
69 Wt. Least Sqrs. No Yes No Yes Yes
70 Robust Regress. Yes No No Yes Yes
71 Logistic No No No No POSSIBLE
CLUSTER ANALYSIS
72 Single Linkage No No No No Yes
73 Average Link. No No No No Yes
74 Complete Link. No No No No Yes
75 K-Means No No No No Yes
ANOVA
76 N-way 3-WAY 2--WAY 4-WAY 30 9-WAY
77 Contrasts Yes No No Yes Yes
78 Unequal Cell Sz No 1-WAY ONLY No Yes Yes
79 Random Effects No Yes No Yes Yes
80 Ancova No No No Yes Yes
BIOM-PC STAT- STATPRO SIGSTAT SPSS-PC
GRAPHICS
NON-PARAMETRIC
39 Sign Test No Yes No No Yes
40 NP Correlation No No No No Yes
41 Runs Tests No Yes No No Yes
42 Mann-Whitney No Yes No No Yes
43 Kruskal Wallace Yes Yes No No Yes
44 Signed Ranks No Yes No No Yes
45 2-Way ANoVA No No No No Yes
46 Kolm.-Smirnov No Yes No No Yes
47 Rank Correlation No Yes No No Yes
48 Log Linear Mod Yes Yes No Yes Yes
49 Other Yes Yes Yes No Yes
CORRELATION
50 Pearson Yes Yes Yes Yes Yes
51 Gamma No ? No No Yes
52 Tau No ? No No Yes
53 MU2 No ? No No No
54 Missing Values No APPARENT Yes Yes Yes
55 Autocorrelation No Yes Yes Yes No
T-TESTS
56 Paired Yes NOT APPARENT Yes Yes Yes
57 Groups Yes NOT APPARENT Yes Yes Yes
REGRESSION
58 Linear Yes Yes Yes Yes Yes
59 Non-Linear No Yes Yes Yes No
60 Polynomial Yes Yes Yes Yes No
61 Multiple Yes Yes Yes Yes Yes
62 Stepwise No Yes Yes Yes Yes
63 Multiple Forw. No No No Yes Yes
64 Multiple Back. No No No No Yes
65 Harm.& Periodic No Yes No Yes No
66 Ridge No Yes No Yes No
67 Leaps&Bounds No No No Yes No
68 Residuals Yes Yes Yes Yes Yes
69 Wt. Least Sq No Yes No Yes No
70 Robust Regression No No No No No
71 Logistic No Yes No PLANNED No
CLUSTER ANALYSIS
72 Single Linkage No Yes No PLANNED Yes
73 Average Link. No No Yes PLANNED Yes
74 Complete Link. No Yes No PLANNED Yes
75 K-Means No SEED No No No
METHOD
ANOVA
76 N-way 9-WAY >3-WAY 3-WAY 8-WAY 5-WAY
77 Contrasts No No No Yes Yes
78 Unequal Cell Yes Yes Yes Yes Yes
79 Random Effects No Yes Yes Yes Yes
80 Ancova Yes No No Yes Yes
ELF STATPAK EPISTAT NCSS SYSTAT
MULTIVARIATE PROCEDURES
81 Manova No No No No Yes
82 Disc. Functions Yes No No Yes Limited
83 Pr. Components Yes No No Yes Yes
84 Factor Analysis Yes No No No Limited
85 Canonical corr. No No No No Yes
86 Gen. Lin. Mod. No No No No Yes
MULTIDIMENSIONAL SCALING
87 Non-Metric No No No No Yes
88 Metric No No No No No
89 TIME SERIES Yes Yes No Yes Yes
90 USER INTERFACE (1) (1) (1) (2) (3)
91 ERROR HANDLING POOR AVER AVER GOOD XLNT
92 RATING F B B+ A- A-
BIOM-PC STAT- STATPRO SIGSTAT SPSS-PC
GRAPHICS
MULTIVARIATE PROCEDURES
81 Manova No No No Yes PLANNED
82 Disc. Functions No Yes Yes Yes PLANNED
83 Pr. Components No Yes Yes Yes Yes
84 Factor Analysis No Yes Yes Yes Yes
85 Canonical corr. No Yes Yes Yes No
86 Gen. Lin. Mod. No No No No No
MULTIDIMENSIONAL SCALING
87 Non-Metric No No No Yes No
88 Metric No No No Yes No
89 TIME SERIES No Yes Yes Yes No
90 USER INTERFACE (4) (5) (5) (4) (6)
91 ERROR HANDLING POOR POOR AVER GOOD XLNT
92 RATING C+ D- C A- B
NOTES
(1) Menu driven program that will run with or without a menu.
(2) Menu driven program. Experienced user can run without menus but not
recommended.
(3) Command Driven Program. Will Accept Submit files for repetitive tasks.
(4) Separate Programs. No integration. Uses rigid syntax control card files except for
interactive programs.
(5) Rigid Menus. No access to individual programs.
(6) Like SPSS-X. Command Driven
I was pleasantly surprised by NCSS's general ease of use and sophistication. For example, Principal Components Analysis is available with or without rotation and will work with either a correlation or covariance matrix. This makes it particularly suitable for work in multivariate allometry, for example, where it is necessary to use the covariance matrix. The regression procedures are a pleasure to use. The interactive robust regression analysis offered in NCSS is not found in any other package I reviewed. It gives the user a choice of several different weighting schemes (e.g. Andrew's sine, Tukey's Biweight), and all essential information is available to the user throughout the robust regression session. Perhaps most surprising of all is the program's remarkable accuracy and speed considering it is written in BASIC and does not take advantage of the 8087. The manual is exceptionally well-written and the program's author was extremely knowledgeable and helpful in the one instance where I needed help.
The program has some bugs that are worth mentioning: the ASCII file import routine is sloppy and frustrating to use; it has no provision for remedying an error midway through the process - it aborts and you must begin anew. It also allows you only to import data in fixed (columnar) format; no provision has been made for free-form data import. The discriminant routine failed on two counts: it does not allow a variable numbered higher than 10 to be entered; it also does not exit from a zero determinant gracefully (by the time you read this, these latter two bugs will have been fixed).
Aside from these minor complaints, I think the program is excellent, and especially worth considering for student use.
EPISTAT is the only public domain software examined. It does a variety of simple univariate and bivariate statistics in a loose menu-driven environment. The menu system is largely unobtrusive; it is there if you need it but individual programs can be run from outside the menu system.
An Epidemiologist wrote the programs; consequently EPISTAT includes some rather unusual non-parametric statistics, including Mantel-Haanzel Chi-Square tests, useful in epidemiology. Epistat has graphing functions to handle plotting bivariate linear regression scattergrams and the predicted regression line. There is no formal manual for the program, although a short "read-me" type file is included on the disk. Most pertinent information is contained within the programs.
Epistat performs all the routine statistics one might need. It is powerful enough for simple statistical problems and most researchers may not need any more sophisticated techniques.
While beggars should not be choosy, I do have a few complaints about EPISTAT. First, the programs are not written to handle double-precision arithmetic. This could cause problems in limited circumstances. Since the programs are in interpreted BASIC, the source code is available and a knowledgeable user could easily modify the programs. Second, the file transfer routine is clumsy to use. The program handles only FORTRAN- type data. Output from spreadsheet programs or Dbase II-III cannot be easily integrated into EPISTAT. Third, the program allows only a very limited set of data transformations. The user who needs non- standard but common transformations will have to transform the data external to EPISTAT. This demands proficiency in another program (e.g. Dbase II or III, Lotus 1-2-3) in order to effect thenecessary data transformations. Finally, the absence of a multiple regression program is a significant omission, but at $25 one shouldn't be too greedy. These problems are more annoying than crippling. They do not mar otherwise excellent programs.
SIGSTAT is a slightly modified version of the mainframe BMD package originally developed by UCLAin the early 1960's. The terms of UCLA's grant to develop BMD required that the programs be placed in the public domain. Significant Statistics has faithfully ported most of the important BMD programs to the IBM-PC, rescaled them to fit this environment, and added some important new programs to make the package more like BMD-P statistically. The current version of Sigstat contains 43 programs spanning 10 diskettes.
If you didn't like mainframe BMD, you won't like SIGSTAT. Each of the programs is separate, and little or no communication exists among them. Nevertheless, every program is bug-free, very fast, accurate, and is a faithful copy of its mainframe counterpart. I analyzed my dissertation data using BMD07M on an IBM 360. I ran the identical data set through SIGSTAT's BMD07M; the results were identical (down to typographical errors that appear in the original BMD output) to those obtained 13 years ago on a mainframe. It took 4 minutes on the micro to complete an analysis involving 375 cases, 21 variables, and 23 groups.
SIGSTAT supports a wider array of functions than most of the other programs I reviewed. It performs most types of multidimensional scaling (e.g. MDSCAL, INDSCAL, etc.), and it does a wide variety of uncommon procedures such as polynomial regression, non-linear regression, harmonic regression, ridge regression, Furnival's leaps and bounds regression (a variant of all subsetregression), Box-Jenkins time series, log-linear models, Probit analysis, and life-table construction. Its regression programs are without peer. It does not support Cluster analysis or high- resolution graphics (these are planned for future versions of the program).
Sigstat is an excellent set of programs, but it is not problem-free. Its most significant problem is its lack of a decent manual. The present version's manual seems to be a photocopied collage of many different pages from the original BMD manual (Dixon, 1977). Modifications to the original BMD instructions are included as afterthoughts. Documentation for new programs, such as RIDGE, is written in a very different style. The manual has a very amateur quality to it. There is no excuse for this poor a manual in a program that retails for $595.00. The manual also lacks an index.
SIGSTAT's other significant difficulty arises from the retention of the original BMD format. BMD forced users to follow a very rigid and structured format for control cards. If the user made one error in typing the control cards the program unceremoniously aborted. Sigstat preserves this unfortunate aspect of BMD. The user is still locked into the rigid control card structure for most programs (although some of the programs are now interactive), and the error handling is poor. If you don't have access to the Microsoft FORTRAN compiler manual, the error messages generally will be cryptic and assuredly meaningless. The manual does have explanations of most common error messages, but I still encountered several errors not listed in the SIGSTAT manual.
Despite these criticisms, SIGSTAT is still a very good collection of programs. The UCLA BMD project has long been known for the high quality and accuracy of its statistical algorithms. That quality carries over into SIGSTAT. If you need avant-garde regression programs, and also want to be able to do different types of Multidimensional scaling, SIGSTAT is your only choice. (As I was finishing this review, Significant Statistics announced that a new version would be released in early Summer, 1985. The new version adds 12 new programs, including Cluster Analysis, makes interactive all of the presently control card oriented programs, and adds a completely rewritten manual).
The Winchendon Group's program ELF (Economics Linear Forecasting) program ostensibly performs a wide variety of common statistical functions. Unfortunately, I found that it did not perform much of anything very well. File import routines were poorly implemented, the discriminant function routine provided no useful statistical information and I encountered bug after bug in the various other modules. Errors inexplicably would bounce me out into the operating system and the program hung my system on several occasions. This program does not merit your consideration.
STATPRO, marketed by Wadsworth Professional Software, is the most extensively advertised package evaluated. The program has been a staple in the microcomputer market for several years; it was first released for Apple computers. After I spent several frustrating weeks reviewing the latest version of Statpro, ITOI, Wadsworth's parent company, capped my exasperation by announcing divestiture of Wadsworth Professional Software. The industry trade papers seem to feel that it is unlikely that the line will be picked by anotherpublisher. In light of this development I see noneed to belabor my negative review of this program. Relevant information can be gleaned from Table 2.Suffice it to say that the program is cumbersome (14 diskettes to make use of all programs andassociated data files), agonizingly slow (even withthe 8087 chip), and wasteful of disk space. In my opinion it offers nothing that is not available inother, better-implemented programs.
On paper, STATGRAPHICS seems to have every statistic and math function any researcher might ever want or need. After reading the pre-release literature I couldn't wait to try it out. Regrettably, the program never measured up to its hype.
I received version 1.0 (2/85) to review. This version is poorly designed, riddled with uncountable software "bugs," and saddled with a comprehensive but nearly incomprehensible manual.
STATGRAPHICS is written in APL. This language is reputed to have a rich set of mathematical primitives that exceed anything in any other language. Thus, in principle APL is an excellent language for mathematical and, hence, statistical programming (see Pease and LePage, 1984 for a particularly strong APL advocacy; however, Statgraphics fails to justify their enthusiasm). This program uses the STSC APL interpreter and will exploit the Intel 8087 math chip if one is available.
The APL interpreter and STATGRAPHICS consume 320K of RAM. The interpreter and the program take
almost 2 minutes to load. Once inside the program, the user is confronted with a rigid menu structure, and an endless array of sub-menus. The program has "fast-tracks" that allow an experienced user to bypass some of the more annoying menus. Even so, the manual is so poorly put together, it took me some time to figure out what fast-tracks were and how to use them. This was a typical problem whenever I used the manual. Information is scattered throughout the manual, some in logical places and some in not-so-logical locations.
The manual offers no clues about the program's data limits. Apparently data limits are related to available "work space" and the type of problem. However, whoever wrote the program didn't figure that mortals worried about such mundane problems. Consequently little time was spent considering what might happen if a user presented the system with an analysis that required more memory than was available. On three different occasions I gave the program a problem that turned out to require more memory than was available. Each time the program began its work oblivious to the size of my data set. Only after considerable grinding away (once for 32 minutes) did it inform me that there wasn't enough memory to complete the problem. Since memory requirements for multivariate procedures can be determined easily from the size of the cross-product matrices (and these don't take very long to compute), it is inexplicable to me why STATGRAPHICS can't ascertain its own memory needs quickly. Surely it doesn't take 32 minutes to compute the size of a cross-product matrix (unless, of course, APL isn't what it's cracked up to be). Alternatively, size limits should be stated explicitly in the manual. That way it becomes the user's responsibility to insure that he/she doesn't overload the system. Regardless of the choice, the present situation is intolerable.
Statgraphics has definite problems with the way it uses memory; as best I can figure, the program inhales it. My test machine was equipped with 384K of RAM. Once the system was loaded, 188 kilobytes of workspace remained. After running out of memory three times, I increased the memory to 512K; the workspace only increased to 233 K. Only one-third of the added RAM was converted to usable workspace. I haven't a clue where the other 83K of added RAM went.
This program has major problems with the data entry and data import routines. It took me over 7 hours to figure out how to import a simple ASCII file. The manual seemed clear enough but after spending 7 hours I began to think that there might be a bug in the program. Customer support at STSC never was able to help. Eventually I quit trying to import the file as a "formatted ASCII file". I was only able to import the data as an "unformatted ASCII file". The resulting APL file was 150K long. The original ASCII data file was only 43K. Either APL file structures are incredibly inefficient, or Statgraphic's implementation is poorly executed.
The full-screen data editor and all menus use a cursor extensively. Neither the editor nor the menus run properly on either a Compaq or a Columbia despite an installation routine that allows the user to select the Compaq directly and the Columbia indirectly. I tested the program on a Compaq; the cursor was invisible during most of the test phases, although occasionally it would surface. I spent much of my time trying to figure out where, on the screen, the cursor was. Occasionally I got lucky and guessed the cursor's location. Other times I guessed wrong with results ranging from ludicrous to disastrous.
Invisible cursors are common problems for software houses that fail to do the research necessary to verify that that most IBM "compatible" computers do not use precisely the same screen drivers as IBM. Statpro made the same incorrect assumption.
Statgraphic's strengths are its outstanding graphics. It does almost all statistical graphics (e.g. box and whisker, scattergrams, regression fits with confidence limits, factor plots, etc. ) as well as two and three dimensional mathematical function plots. All the graphics interface with a graphics editor that gives the user full control over labeling, scale, and perspective.
As with so much of the program the graphics have numerous problems. For example, Statgraphics supports cluster analysis, but it does not plot dendrograms. Hard copy of the screen images can only be obtained with an EPSON or IBM graphics printer, or a Hewlett-Packard plotter. As with the screen drivers, the company has not done its homework. Few printers claiming to be Epson compatible really support the Epson standard in graphics mode. Consequently, none of the Epson compatible printers I tried would yield usable graphics output. (In fairness, STSC doesn't claim the program will work with printers other than Epson's, but the program isn't very useful if it doesn't support a wider variety of printers).
Statgraphics has far too many deficiencies to recommend it. While its graphics are good, they alone don't justify its purchase. Many other stand- alone graphics programs (e.g. Energraphics, Microsoft Chart) perform as well as STATGRAPHICS while supporting a wider array of printers. The non-functioning file import routines, the lack of clear cut data limits, the cumbersome menu system, the poor manual, and profligate use of memory, to list a few significant problems, conspire to make this a program worth avoiding right now. STSC's heart is in the right place, but the program needs a lot more work before it is released.
Northwest Analytical's STATPAK is a collection of interpreted BASIC programs that perform most common univariate and bivariate analyses. The package includes a competent data editor and nice transformation routines, but there is nothing particularly special about STATPAK. Its greatest virtue is also its worst vice. The program is distributed as modifiable BASIC source code. This format permits the skilled user to modify any program to suit any particular need. Some users need this ability, but the price paid for this is an extremely slow, and occasionally inaccurate, set of programs. Having the source code allows the user to compile the programs to speed them up, but this necessitates some extensive rewriting of the code to remove those features of interpreted BASIC that are incompatible with compiled BASIC (e.g. variable array sizes). The amount of work required to rewrite sections of the program is considerable, the resulting programs will be less flexible, and once done, the user will lose the customer support of NorthWest Analytical.
The package offers few surprises. Its routines are fairly standard, although the inclusion of Fourier Analysis is somewhat unusual. The programs are conservatively written using standard and well- tested algorithms (one exception to this is detailed below). The programs are very easy to use. The manual is well-written and information is easy to find. Statpak has the most extensive index of any program I reviewed.
Problems may exist with the multiple linear regression program in Statpak. This program gave the most discrepant results of any of the packages tested. Statpak's MLR program gave only 4.7 digits of accuracy on the Longley test data; other programs gave significantly higher accuracy. This suggests that Statpak's algorithm may not be sensitive to roundoff error. This, in turn, may mean that results obtained using data sets moderately ill- conditioned could be questionable. It is not likely that this error arises from using BASIC. Number Cruncher, also written in BASIC (albeit compiled), yields far better accuracy on the Longley data. This makes STATPAK's regression algorithm suspect.
Statpak lacks multivariate procedures, and the graphics, such as they are, are extremely limited. This package is expensive for what the user gets. NCSS, which has a much wider spread of common statistics at less than half the retail price, is a far better buy.
While I would not recommend this package, it might merit attention. My only reason for even considering this package would be for access to the source code. I suspect there are cheaper ways to get this kind of information.
BIOM-PC is a package keyed to Sokal and Rohlf's (1981) textbook Biometry. It computes all statistical tests described therein.
The BIOM-PC programs are primitive compared to others I reviewed, but they use well-tested algorithms, support the 8087 math chip (if present), and give accurate results. The output conforms to the textbook, which makes the textbook an essential adjunct to the manual. The manual is very spartan; it has no index, and explanations are kept to a minimum.
I experienced only one problem with BIOM-PC. The Longley data taxed the output format of the multiple regression program so that it overflowed the fields. Users without access to the source code would not be able to fix this problem; however it is unlikely that users interested in this package would have data as nasty as the Longley set.
For $60 you can buy the source FORTRAN code for BIOM-PC. The source code has to be compiled with the Microsoft FORTRAN compiler (version 3.13 or higher). Plan on spending at least 3 unpleasant hours compiling and linking the programs. Users without access to the compiler can buy a precompiled version for $75; however you won't get the source code.
If you use Biometry, you should get the package. If you have no use for the textbook, I doubt whether BIOM-PC offers you much.
Most readers of this paper have used mainframe SPSS. SPSS-PC is the microcomputer version of this mainframe behemoth.
SPSS-PC requires about 3 megabytes of mass storage. Since the program, like its mainframe counterpart, is heavily overlayed, you must have at least a 10 megabyte hard disk to run it. SPSS-PC requires 320 K of RAM but it won't make use of more than 384K. The 8087 math chip is supported and its presence in the computer is strongly recommended.
SPSS-PC performs all common univariate and bivariate statistical analyses as well as multiple regression, factor analysis, cluster analysis, and log-linear models. All of the statistical modules look and act like their counterparts in SPSS-X.
The manual, like all SPSS documentation, is voluminous, comprehensive, and very well written. It gives almost all computational formulae, and is replete with examples of command sequences for all common statistical procedures. SPSS-PC's command structure is similar to that used in mainframe SPSS- X. Where there are significant (or even minor) differences, these are clearly stated in the manual. Mainframe SPSS users who don't have access to SPSS-X may find the transition from SPSS-9.0 to SPSS-PC difficult since the syntax is different and no provision exists for translating earlier SPSS command files into equivalent SPSS-PC files; the command files will require manual recoding.
SPSS-PC is the most expensive program I reviewed. It is also the only program that is copy- protected (now that SYSTAT has removed its protection). At retail it is 60% more expensive than SYSTAT and a third more costly than SIGSTAT. The user gains nothing for the price difference. In fact, both of the latter programs give the user considerably more statistical power.
The decision to purchase SPSS-PC will depend on many factors. Obviously you need a hard disk. If you don't have one and want SPSS-PC, then SPSS-PC will cost closer to $2000. Given that you have a hard disk, SPSS-PC might be a wise investment if you are already using mainframe SPSS-X, want to have portability between the mainframe and microcomputer, and have no need for such procedures as discriminant analysis, MANOVA, and Multidimensional scaling (but see note below). If you are using SPSS-9.0 (not X), you may find the transition to the PC version clumsier than you would like and file portability more limited.
Users who plan to transfer files between SPSS-X and SPSS-PC may find that communication between mainframes and PC's is more complicated than anticipated. While the microcomputer version comes with KERMIT (a file transfer program for micro- to mainframe communication), many mainframe computers do not support KERMIT protocol. This means that the user will have to rely on the local computer center to figure out how to import SPSS-PC files into SPSS-X and the converse. Conclusions
In the final analysis, several programs proved suitable for our research uses. SYSTAT was my first choice; SIGSTAT was a distant second. Both programs are fairly difficult to use, although a persistent user can master both programs quickly. SYSTAT has a far better user interface and gives the user more control. Its error handling is excellent, it is blindingly fast if the 8087 is installed , and it can perform virtually any statistical analysis the researcher is likely to need.
SIGSTAT contains a large array of important statistical routines. It is best used if needs include full discriminant analysis, a wide range of regression procedures, and a rich set of metric and non-metric multidimensional scaling routines. Under those circumstances this program should be your first choice.
SPSS-PC has received uniformly good reviews, but it requires a hard-disk, is rather expensive, is copy-protected, is missing key multivariate procedures, has primitive graphics, and does not take advantage of the full memory capability of the PC. If you have a hard disk and are considering SPSS-PC, I would encourage you to wait until SAS-PC is released. Alternatively, you might want to wait until the just announced new additions to SPSS-PC appear. These new additions include MANOVA and discriminant analysis (see Keywords, SPSS's semi- annual newsletter, Spring 1985 issue for more information).
For teaching purposes, NCSS would be my first choice. It is tightly menu-driven, easy to use, and has all the functions a student is likely to need. Although the program is written in compiled BASIC, it is surprisingly fast and astonishingly accurate.
Everyone should have EPISTAT. It is the best bargain available. It includes most commonly-used statistics and would easily suffice for those areas typically covered in the first semester of a statistics course. Most researchers need little more than this.
It is interesting that none of the highly recommended programs have particularly good graphics. For users who want a rich array of graphics, I would recommend purchasing a stand-alone graphics program. Those statistics programs that include a wide selection of graphics routines are also burdened by the overhead these routines require. This overhead consumes memory, adds to the clumsiness of the program, and degrades program performance. I have yet to find a competent statistics program that also includes decent graphics capability. Perhaps the two are mutually exclusive on today's personal computers.
In the final analysis you must determine what is best for your needs. My experience suggests that reading reviews, reading product literature, or talking to computer store salespeople is not adequate; indeed these sources may be misleading. Use this information to guide your initial hunt for programs that might be worth reviewing. But remember that nothing substitutes for actual use of the program. This is the only way to ensure that the program does what you expect.
Since completing this review several manufacturers have introduced updates to their products. The most significant ones are upgrades to SYSTAT (version 2.1) and to STATGRAPHICS (version 1.1). The previously announced upgrade to SIGSTAT has not yet materialized.
The changes to SYSTAT merely add frosting to a generally delicious cake. In addition to removing copy protection, SYSTAT now permits access to MS-DOS commands from inside the programs. The cluster routines have been enhanced to allow the cluster schedule to be printed along with the dendrogram. The MGLH routine now computes discriminant scores and allows them to be saved to disk. All of the graphics routines now use either IBM graphics characters or ASCII character graphics. Users with IBM graphics printers may now get high resolution hard copy of the screens. Users without IBM graphics printers have access to alternate drivers that continue to print character graphics. Other changes include a typeset manual, and large versions (512K RAM required) of all routines rather than just MGLH. The package now comes on 5 diskettes rather than 6.
STATGRAPHICS 1.1 corrects many of the problems noted above. The invisible cursor problem has been corrected and boot time has been speeded up dramatically. The file import routines have been clarified, and it is now possible to import a data file in less than 7 hours. Unfortunately, there is no new manual, and the menu structure of the program is unchanged. Consequently, information is still difficult to find and the program still very cumbersome to use. With support of the IBM Enhanced Graphics Adapter, STATGRAPHICS produces exquisite on-screen graphics; however the program still will only support IBM graphics printers and their exact clones. Finally, the memory usage problem has not been resolved. Since reviewing version 1.0 I have brought my machine up to 640K. I gained only 91K more of workspace (512K to 640K --> 233K to 324K) for the addition of 128K of RAM. I still cannot understand why so much RAM is lost to the system.
Finally, the long awaited SAS-PC is in production and will be out by the time you read this. Unfortunately, SAS-PC will be available only in Corporate and University Licenses. The complete set of SAS modules in PC form will be licensed for something in the neighborhood of $7000 with a sizable annual renewal fee. Not only does this preclude individual purchase, it also puts it out of reach of many colleges as well. There is some push to get SAS to offer individual licenses; SAS will consider this in early January.
Carpenter, James, Dennis Deloria, and David Morganstein. 1984 Statistical software for microcomputers. Byte 9(4): 234-264.
Dixon, W.J. 1977 BMD: Biomedical Computer Programs. Los Angeles: University of California Press.
Fridlund, Alan J. 1985 Taking the numbers bull by the horn. Infoworld, February 11, pp. 42-50.
Johnson, R. A. and Wichern, D. W. 1982 Applied Multivariate Analysis. New York: Prentice-Hall.
Lachenbuch, P. A. 1983 Statistical programs for microcomputers. Byte 8(11): 560-570.
Lee, R. S. and Cohen, J. 1985 Good with stats: SYSTAT and EPISTAT. PC Magazine 4(5): 179-186.
Longley, J. W. 1967 An appraisal of least-squares programs from the point of view of the user. J. Am. Stat. Assoc. 62: 819-841.
Lubowsky, B. 1985 Choosing the right variable. Digital Review. March 1985, pp. 57-66.
Pease, J. W. and Lepage, R. 1984 An evaluation of selected microcomputer statistical programs. Working Paper #15, Michigan State University, Department of Agricultural Economics.
Schervish, M. J. 1985 Statistical computing software reviews: SYSTAT. Am. Statist. 39: 67-70.
Sokal, R. R. and Rohlf, F. J. 1981 Biometry. Second Edition. San Francisco: W. H. Freeman.
Vanderbok, W. G. 1985 SPSS arrives on the PC. PC Magazine 4(5): 191-195.
Winer, B. J. 1971 Statistical Principles in Experimental Design. Second Edition. New York: McGraw- Hill. Wonnacott, T. H. and
Wonnacott, R. J. 1984 Introductory Statistics for Business and Economics. Third Edition. New York: Wiley.
This section of the CAANews is devoted to reports of what anthropologists are doing with computers. Please send the information on what you are doing, your successes and failures, and what you are interested in to the editors. We will publish your reports with your name and address so our other readers can get in touch with you if there is a possibility of mutual assistance. The reports are going into a data-base that can be searched in the future if someone needs to find a person to help with a CAA problem. We will also publish comments here.
Please let us know if you would also like to participate in on-line conference to exchange this information.
Applied anthropologists now using one of the computer packages SAS, SPSS, or SPSSX in conjunction with qualitative techniques are invited to contact Margaret S. Boone (MSBUHT@GWUVM.BITNET) or write her at the U.S. GAO, Room 5741, Washington, D.C. 20548, to participate in a possible session at one or more upcoming meetings. Those using spreadsheets, like LOTUS, for analysis are also invited to make contact.
Through a grant from the Fund for the Improvement of Post-Secondary Education, the Social Science Research and Instructional Computing Laboratory of North Carolina State University announces the availability of honoraria awards for microcomputer software authors in the social sciences.
Authors of Apple or IBM microcomputer programs can be paid $250, $500, or even $1000 in honoraria if they contribute social science-related software and documentation to the public domain, or allow its distribution on a non-exclusive basis.
If you are interested in developing and contributing software to the social science academic community in this manner, or if you wish to obtain information on presently-available low-cost social science microcomputer programs, please contact G. David Garson, Director, Social Science Research and Instructional Computing Laboratory, North Carolina State University, Raleigh, N.C. 27695-8101, or call 919-737-2468 or 919-737-3067.
The Social Science Microcomputer Review, now in its third year of publication, invites papers on social science applications of microcomputing to research and teaching. Starting with Volume III, Spring 1985, the Social Science Microcomputer Review will be a peer-reviewed journal published by Duke University Press on a quarterly basis.
The SSMR has a national editorial board of scholars in the field, and invites the submission of articles by interested colleagues. The journal's editor is G. David Garson, Assistant Dean, School of Humanities and Social Sciences, North Carolina State University at Raleigh. The associate editors are Michael L. Vasu and Frank M. Howell, both of whom are Associate Directors of the Social Science Research and Instructional Computing Laboratory, North Carolina State University at Raleigh.
In addition to articles, SSMR also invites shorter notes on microcomputer experiences, institutional news, syllabi, bibliographies, resource notes, programming tips, software evaluations, applications reactions, and microcomputer-related questions. The SSMR invites interested colleagues to submit their names for consideration as peer reviewers of manuscripts, books or software.
Please write for our style sheet, which calls for submission of three non-returnable double-spaced manuscripts, preferably accompanied by an IBM or Apple word processor diskette.
Subscriptions are $24.00 individual, $40.00 libraries, add $6.00 for subscriptions outside the United States. Please make checks payable to Duke University Press and mail to Periodicals Dept., 6697 College Station, Durham, NC 27708. Social Science Microcomputer Review P.O. Box 8101 NCSU Raleigh, N.C. 27695-8101 (919) 737-2467: Editor (919) 737-3791: Associate Editors (919) 737-3067: Software Department (919) 684-2173:
Subscription requests should be sent to James Dow, Editor, CAAN, at any of the addresses which follow. Subscriptions are $6.00 US for four issues. The previous volume of CAAN, Vol. 1, consisting of four issues is available on a MS/PC-DOS 5 1/4 diskette for $10.00 US. For other diskette formats add $5.00 US.
Material should be submitted to James Dow, Editor, CAAN, at any of the addresses which follow. CAAN has a magazine style in which news, letters, notes, and articles are welcome. We prefer material submitted electronically, on diskettes or via network mail.
The Computer Assisted Anthropology News, Dept. of Soc. and Anth., Oakland University, Rochester, Mich. 48063, USA
James_Dow%UMich-MTS.Mailnet@MIT-Multics.ARPA MAILNET James_Dow@UMich-MTS.Mailnet FIDOMAIL James Dow, Fido #5 Net 120 COMPUSERVE 70150,266 UUCP umich.uucp MCI Mail 281-2421 TELEX 6502812421 TELEPHONE (313) 370-2430, 375-1258