New Help For Researchers With """"Rusty"""" Statistical Skills

Statistical Navigator from The Idea Works Inc., Columbia, Mo., was designed not to perform analyses but to work as a supplement to existing statistical packages. It uses a type of artificial intelligence called expert systems to determine, from users' input, which statistical analyses are most appropriate for their data. (Statistica from StatSoft, Tulsa, Okla., uses similar technology in its module "Statistical Advisor.)"Expert systems are so named, explains Edward Brent, president of The Idea Works, "because they are designed to help a user with a specific, complex task in much the same way an expert consultant would work, if one were available.Actually, Brent prefers the term "decision support system over "expert systems to describe Statistical Navigator. The latter phrase might convey the wrong impression, he fears--that users answer a few questions, then blindly accept answers that the program spits out."In each case, Statistical Navigator explains what each analysis does and how it does, or does not, fit the user's research objectives and assumptions. The researcher makes the final decision about which one to use, Brent says. "We believe that the appropriate role for Statistical Navigator is to help researchers think through their analysis problems, not to help them avoid thinking.Statistical Navigator resembles a human consultant in the way it interacts with a user. It starts out with general questions, then moves to ones that are more specific. For example, if the user is knowledgeable enough to select a broad category of analysis from a list (causal analysis, measures of association, scaling and classification, and so forth), the package works at that level.If researchers are less sure of what they are doing, the program gets more specific. The next step is a list of common research questions, asking the user to select the one that best describes his or her own objective. If this doesn't work, the list of questions becomes more detailed and comprehensive. Finally, users have the option of using Statistical Navigator's natural language interface to enter a brief phrase describing their research. Unlike a human, Statistical Navigator doesn't grow tired or impatient, no matter how long this process takes.Once Statistical Navigator has the information it needs, it pre- sents the four analysis techniques that best fit the user's objectives and assumptions. A final report gives a detailed description of each of the four analyses, a numerical score representing how well each matches the user's requirements, a discussion of possible significance tests where appropriate, references to statistical literature in which the user can find out more about the techniques, and, whenever possible, a list of the commonly available statistics packages that perform the analyses. Like InStat, Statistical Navigator makes on-screen help available throughout the process. Also like InStat, Statistical Navigator's help screens are written for an audience that is not expert in statistics.Trevor Hastie, chairman-elect of the American Statistical Association's Statistical Computing section and a scientist at AT&T Bell Laboratories in Murray Hill, N.J., cautions that programs like these "can be useful for someone who really doesn't know what they're doing, but you only get out what has been put in to the package. Software based on expert systems assumes that whoever made the rules made the right decisions, he says."It would seem that these packages would be the ones to recommend to nonexperts, Hastie adds. "But they might be giving a wrong answer, and a nonexpert might not realize it."We do not intend Statistical Navigator to replace advice from a competent statistician, says Brent. "The program may suggest the researcher consider some form of analysis which is quite new to them. They may be well advised to seek out competent statistical advice.--C.D.P. One-way ANOVA or chi-square test? Two-sample t test or Wilcoxon matched pairs?Statisticians have no trouble with questions like these. Determining the appropriate analysis for a given set of data is what they are trained to do. Research scientists, on the other hand, typically have had only a course or two in statistics and lack a thorough understanding of the subtleties of the different statistical analyses."Since [researchers] are trying to generalize something about an entire population based on a small sample, it's very important to use the correct test," says Mary Newman, a toxicologist with Liposome Technology Inc. in Menlo Park, Calif. "But statistics is kind of a black box, something we aren't really trained in.This casual acquaintance with statistics is only part of the problem for researchers when it comes to analyzing data. An abundance of statistics software, available now on every type of computer from mainframe to Macintosh, makes the task even more difficult because many stat packages resemble a Thanksgiving dinner--too much of a good thing."Powerful packages such as SPSS, SAS, SYSTAT, BMDP, and others provide an embarrassment of riches to the user in the possible forms of statistical analysis, says Edward Brent, president of The Idea Works Inc., Columbia, Mo. "There are literally hundreds of forms of analysis available today.Today's crop of statistics software springs from two sources: mainframe-based programs (most have migrated "down to personal computers and workstations, as well) and newer packages written specifically for personal computers. Packages in the former group were originally created for statisticians and assume a familiarity with both statistics and computing that many research scientists do not have. Also, in terms of the analyses they offer, these packages are extremely comprehensive. One pharmacological researcher, who asks not to be identified, likens them to a car with 43 gears. "Either you just don't drive it because it's too hard to figure out, or you putter along, never taking it out of first gear, he says.Software written specifically for personal computers, on the other hand, often provides less functionality than mainframe packages (fewer forms of analysis, for example, or the ability to handle fewer cases and variables), but the trade-off is easier use. These programs--StatView from Abacus Concepts, Berkeley, Calif., and Statistica from StatSoft, Tulsa, Okla., for instance- -make good use of icons, on-line help, and mouse manipulation to greatly simplify data analysis for scientists.Two other packages, InStat from San Diego-based GraphPad Software Inc. and Statistical Navigator from The Idea Works, go even further, to the point of acting like a combination of a favorite stat book and a patient consultant, helping scientists understand the different analysis options so that they choose the appropriate ones. (Statistical Navigator does not perform actual statistical analyses; see accompanying story.)This second group of software packages represents an approach to statistics that has been a long time in coming, according to Harvey Motulsky, president of GraphPad. "I'm surprised the personal computer had been out so long before people made more usable statistics programs, he says.Part of their charm is that most of these newer packages were written not for statisticians, but for scientists. "My experience with mainframe stat software was back in graduate school, says John May, a pharmacologist at Whitby Research in Richmond, Va. "Now, I try to avoid those programs at all costs. Even then, if I could, I used my programmable calculator instead of the mainframe stat package.InStat, in particular, was designed to meet the needs of scientists who are infrequent users of statistics software and somewhat unfamiliar with statistics in general. Two attributes enable it to do this. First, InStat was deliberately designed to lead a user logically through each step of the data analysis process, from entering data to viewing (and understanding) the results. In fact, although most statistics packages do not give users any guidance about entering data, getting data into the computer in the correct format goes a long way toward simplifying the analysis process to follow.InStat uses dialog screens to ask users questions, then builds on their answers to lead them to the next step. For example, an early screen asks which general type of analysis is needed (column comparisons, linear regression, or chi-square and related analyses of contingency tables). InStat uses the answer to create a custom spreadsheet into which data are entered.Once the data are in, the program presents the user with its choices for analysis. With two columns of data, for instance, InStat defaults to the unpaired t test, reporting the two-tailed P value. If this is satisfactory, the user hits a button marked "Do It, and those calculations are made. If other analyses are preferred, the user selects them from a menu.The second feature that makes InStat a blessing to the nonexpert in statistics is its use of plain English--a distressingly rare occurrence in both software and statistics. At most points in the data-analysis process, if a user is confused about a particular term or analysis, he or she can call up a help screen that clarifies it in easy-to-understand language. Here is an example of such a help screen:"Recommendation: Select Fisher's test. The corrected chi-square test is more popular, but only because the calculations are much simpler to perform by hand. Fisher's test yields the exact P value; the chi-square calculations yield an estimate...InStat doesn't try to be all statistical analyses to all researchers, which accounts in part for its user-friendliness. It provides the more commonly needed tests in the realm of biostatistics. Concerns that, in their effort to pare down and simplify InStat, the program's developers might have omitted a number of important analyses appear to be groundless. Diana Gabaldon, writing in Laboratory Microcomputer (10:2[30-33], 1991), was "entirely charmed by the simple thoroughness of this package. She adds that the program includes about 85 percent of the statistical tests normally needed. Pharmacologist May uses InStat and says it lacks only one post hoc test he uses. The company promises this test in the next release.Although InStat is very good at making statistics understandable, it does have limitations. It cannot read most spreadsheet files directly, for example. Its graphs are very simple, and the user is not able to modify or enhance the graphs. Still, at $110, InStat seems well worth its price.Many scientists, although experts in one field, have to venture into the very different, very complex field of statistics to communicate their research to others. User-friendly statistical software helps structure their selection of the appropriate statistical tests and ease the burden of running those tests.Caren D. Potter is a freelance writer based in McKinleyville, Calif.

New Help For Researchers With """"Rusty"""" Statistical Skills

Interested in reading more?

Become a Member of