New Software For PCs Helps Take Anxiety Out Of Statistics

In the last few years, developers of statistical software have tried hard to reach the nonstatistician. User-friendly, menu-driven programs, with high-resolution graphics to visualize the results of statistical analyses, are becoming increasingly available. And while software packages may not take the place of a fundamental understanding of statistics, they can help improve the statistical literacy of a scientist who understands the basics. Many of the software manuals include detailed explanations of procedures, and the ease of using some packages may encourage brave researchers to experiment with an unfamiliar routine. These newer programs are a far cry from the statistics programs many scientists cut their teeth on in graduate school--programs with little or no graphics, driven by batch files of nonintuitive commands. Take, for example, a drug company researcher who is running an experiment comparing the effects of two drug formulations on two groups of rats, a control group and a group whose immunity has been compromised. Traditionally, the scientist running statistics on this experiment had to write complicated programming instructions telling the computer what the experimental designs were; which treatments to compare; what, if any, data were to be left out; and any other information necessary to complete the run. The researcher had to construct a data file to the exact specifications of the program being used, and, finally, send the whole mess to a remote mainframe computer. The mainframe would then follow the instructions in the batch file, chew on the data--often overnight--and, provided everything worked correctly, present the researcher with a huge pile of printouts to wade through to find the relevant information. If a typo or incorrect command found its way into this batch file, the frustrated researcher had to write a new one and begin the process again. Today, while mainframe programs are still indispensable for complex statistics and huge data sets, for most scientists, the statistics programs that have been developed for personal computers are more than adequate for the job. These newer software packages have a lower frustration factor and shallower learning curve than the earlier generation of statistical programs. Many new statistics programs are interactive. Instead of writing a batch file and waiting for the results, the scientist can enter commands or data from the keyboard, look at the results, and then run a different type of analysis, without having to wait. For example, after running an initial set of statistics, the drug company researcher could separate out and compare replications, or look at interactions between, say, carcinogenicity and toxicity. The predominant vendors of mainframe statistical programs, including the SAS Institute of Cary, N.C.; Chicago-based SPSS Inc.; and BMDP Statistical Software Inc. of Los Angeles, have come out with personal computer versions of their programs over the last decade or so. When they first appeared on the market, these large and comprehensive programs tended to retain the feel of their mainframe big brothers. The user wrote a batch file that told the computer what to do with the data. The computer cogitated for a time, then spit out the results. Today's versions of some of these programs, however, are increasingly of the interactive variety. While the mainframe-like mode of operation is cumbersome and time-consuming for many nonstatisticians, others, particularly those who learned these programs in college or graduate school, find the inconvenience well worth the statistical power they wield. Besides, as biostatistician Estelle Russek-Cohen of the University of Maryland at College Park says, "If you were originally trained on a mainframe, and know SAS, it's quite natural that you would want to use it on the PC." However, she adds, SAS on the PC--like other programs originally designed for mainframes--"is a memory hog," and researchers may be hard-pressed to keep their computer hardware up to the task of running these programs. In the mid-1980s, software developers began coming out with statistics programs designed especially for personal computers. These packages, such as Systat, from Evanston, Ill.-based Systat Inc., and Statistix, from Analytical Software of St. Paul, Minn., tended to be easier to use than the mainframe-type programs, but more statistically limited. In today's market, however, the prospective buyer has a wide range of choices. Programs originally ported from mainframes have become increasingly user-friendly and graphical, while the packages originally designed for PCs have become more powerful, as well as easier to use. Zenaida Otero Keil, an assistant professor in the department of mathematics at Glassboro State College in New Jersey, teaches short courses in statistics at scientific meetings. She advises scientists in the market for a package to consider their needs carefully before making a purchase. "If you don't want to do nonlinear regressions, you don't need that option," she says. "You can use something smaller, more friendly." In fact, most packages "offer all the technical aspects that are important"--the most frequently used statistical routines, accurately executed. "What distinguishes one package from another," she says, are a host of other qualities that contribute to the user-friendliness of a package--whether the output is easy to read, the terminology is understandable, and the data are easy to input and manage. Also important are the diagnostics in a program, says Keil. Diagnostics use calculated values to determine whether the results of an analysis make sense. If the results don't pass the test, either the wrong statistical techniques were used or something was wrong with the data. Graphics are another important selling point of the various statistics programs available. Many offer high-resolution graphics with a broad choice of graphs and tables for visualizing the data. A few programs, such as Statgraphics, from STSC Inc. of Rockville, Md., and CSS:Statistica, from Tulsa, Okla.'s StatSoft Inc., boast of having graphics fully integrated with statistics, allowing the user to rapidly move from a statistical analysis to a graph and back again. Sean Gibson, an assistant products manager at STSC, says that Statgraphics goes a step further, with "interactive graphics." Consider a researcher who is graphing a comparison of plant growth to exposure to light. If a data point is lying way off the regression line, the researcher can click on the point with the mouse and find out detailed information associated with that point that may account for the unexpected result. And while statisticians cringe at the thought of "messing with the data," says Gibson, Statgraphics includes a "what-iffing tool": The researcher can use the mouse to exclude some data points, and instantly rerun the regression. The old regression line stays in place, so that the researcher doesn't confuse the real data from the "what if" data, but a new line is drawn as well. A few companies are beginning to offer a new generation of packages that take advantage of the capabilities of the newest IBM-compatibles, the 386 and 486 computers. These use microprocessors with "32-bit architecture," which basically means they can process information much faster than the older, 16-bit ones. BMDP began selling BMDP/386 last summer, and STSC launched Statgraphics Plus in December. These new packages have two important advantages, according to representatives of both companies. They process information three to five times faster than other programs, and the size of data set the program can work with is constrained only by the amount of memory (RAM) in the computer, not by the program. A scientist who once had to go back to the mainframe for very large data sets can now stick with the PC, provided he or she has sufficient RAM. But for all the integrated graphics, user-friendly interfaces, and diagnostics that statistical software packages provide, there's one thing these programs don't do: They don't substitute for a basic understanding of statistics--when to use what procedure and what the results mean. A statistics package, according to Russek-Cohen, assumes users already know what they want to do with the data, and are "looking for the easiest way to get that done." Keil says that while it is unreasonable to expect scientists to be highly literate in statistics--"their main job is not to do statistics, but to do science"--every scientist and engineer knows enough math to be capable of understanding some basic techniques. And for scientists with a good grounding in statistics, the options available on a good statistics package may help them broaden their statistical horizons. But all is not lost for scientists who barely scraped through their remedial statistics classes. A few "expert system" programs exist that actually help scientists decide which statistical procedure is appropriate to a particular data set or experiment. One such program, Statistical Navigator Professional, is part of a bundle of packages offered by the Idea Works Inc. of Columbia, Mo., called the Methodologist's Toolchest Professional. Other packages offered by the Idea Works and contained in the Methodologist's Toolchest include Hyper-Stat, a hypertext dictionary of statistical terms; WhichGraph, an expert system to help a researcher choose the best graph to use to display data; and Ex-Sample+, for determining optimum sample size. The company also produces a quarterly electronic (on computer disk) journal of reviews of statistical and other research-oriented software. Many statisticians, however, are wary of recommending an expert system. "I don't think they are far enough along to substitute for a working knowledge of statistics," says Russek-Cohen. Keil says that "they can be very helpful for a lot of common data, to point you in the right direction," but warns that they can be misused: "Expert systems may not be able to deal with some of the real intricacies of research data. They may give a false sense of security that they will tell you exactly what technique you should use." Keil likens using statistics in research to driving a car. "Nearly everyone can drive a car, but you don't need to know how to take the fuel injection system apart." Likewise, she says, a statistics package allows researchers to "use the power of statistics without the tedious calculations." As software continues to eliminate statistical drudgery, statistics should become as common in research as driving to the lab every morning--and maybe even less stressful.

New Software For PCs Helps Take Anxiety Out Of Statistics

Interested in reading more?

Become a Member of