Statistics and Social Sciences

Negative Life Events
Statistics and Software Selection

Sanford Weinstein

[Ed: Links to web pages and/or e-mail addresses which have become inactive since the publication of this article have been enclosed in curly brackets { }. Replacement links have been provided where possible.]

Researchers usually separate these events from any larger continuum of which they may be a part. That is, they think of negative events happening so that the zero-point of an event (its non-occurrence) and infinity (or the highest possible occurrence) are opposite poles on a continuum. This arbitrarily places a floor or ceiling on measurements.

In such cases, especially those involving health and safety, most respondents report information near the non-occurring end of the scale, though some will report higher levels. In schools, studies of at-risk behaviors often produce the "half-of-a-bell-curve" look of a truncated distribution. What has been cut off is the positive part of the larger picture.

For example, measures of drug use usually ask about how often, if ever, one has used drugs or the last date of drug use. Although school-age drug use is on the rise, most students have limited experience, if any. Therefore, in examining the proportion of students using drugs (other than alcohol or tobacco), most report little if any use and some may be actively anti-drug in their behavior. However, researchers rarely measure positive actions such as admonishing drug-using peers; confronting parents and other adult family members about drinking, smoking or drug use; refusing to go to parties where drugs and alcohol are present; or anti-drug activism in school.

In these instances, the problem-oriented perspective imposes a floor at an arbitrary neutral point, and generates distributions that seriously violate assumptions of the most powerful, useful and commonly understood statistical methods of analysis. Such violations distort parameters, lessen the robustness of findings, and obscure the meaning of statistical results.

To cope with these often unpredictable violations, researchers use arcane and complex procedures to transform data or manage distribution problems. Unfortunately, many of the most popular, easy to use and readily available statistical software packages cannot perform such analyses. For example, SPSS and Statistica are popular programs because they are easy to learn and provide excellent data management and graphics capabilities. However, they do not offer methods for addressing these distribution problems.

More specialized programs such as LIMDEP do offer these analytic methods, but can be difficult to learn and do not offer good data management procedures or graphics generators. What is more, most users will only turn to these programs on rare occasions, when dealing with data distributions that don't meet standard statistical assumptions. Therefore, they are likely to forget how to use the program between instances of usage. SAS has all the desired features—wonderful capabilities for analysis, data management and graphics—but it has a steep learning curve, so people shy away from using it unless absolutely necessary.

Researchers therefore rely upon professional statisticians who are knowledgeable about unusual analyses and the often obscure software needed to conduct them. Unfortunately, gaps frequently exist between the subject matter expertise of the researcher and the technical knowledge of the statistician. Opportunities for novel examination or creative interpretation fall through these gaps, and readers of the research literature often find published reports of such work confusing.

For example, one recent evaluation of the Leadership Program, a school-based violence prevention program tested in 1997 in 53 New York City public schools, used "Frequency of hostile acts in the classroom" as a criterion measure. Clearly, this addressed only the negative side of peer relations.

Though student ratings of classroom hostility adequately served the evaluation and provided good evidence of the program's success, ratings were distributed from high to low in a way that differed dramatically from the normal bell curve. Most ratings were at the very bottom of the scale, and the remainder fell above these in a half-bell curve like the one described above. Consequently, the evaluator relied heavily on less powerful distribution-free statistical analysis, and was frustrated in attempts to construct a useful explanatory model, despite assistance from a skilled statistician.

This case illuminates one way for researchers to avoid such entrapment. At the planning stage, researchers should seek to identify the broader context of negative events, beyond the focus of the problem, and incorporate it into measurement of those events.

After the evaluation, the researcher recognized that peer relations have many positive elements too, and that the measurement objective should not be limited to hostility, but extended to peer relations more generally. Future attempts to evaluate school violence prevention should assess positive behavior such as respect, courtesy, generosity and benevolence, and not only hostility and aggression.

From this broader perspective, peer relations form a continuum with friendly and hostile acts at opposite ends, and neutral or benign behavior at the center. In theory, measures of peer relations across this continuum assume a normal distribution around that center.

The bell curve generated by this approach is amenable to standard analyses that can be readily performed by any statistical software package, including more user-friendly ones such as SPSS and Statistica. NYU has licenses for several packages, including the new release of SPSS 8.0 (see the related article on SPSS 8.0).

Though the approach has yet to be tested, it seems a viable alternative to the problem-oriented focus. There is a need for careful thought before data gathering if one is to reduce frustrating, tedious and labor-intensive scrambling afterward. Good old-fashioned variance and covariance analysis, or ordinary least squares regression analysis, may then be useful, conveniently available in user-friendly software, and comfortably familiar in straightforward examinations of negative event data.

Posted May 18, 1998. Last reviewed December 6, 2005.