It is common knowledge that experimental data are prone to have occasional gross errors. Such errors may have substantial effects on the result of the analysis. A large body of statistical literature exists on methods that are not only less sensitive to such errors, but which are also close to being optimal - resistant and robust methods. For example, instead of the non-robust mean, a trimmed mean or the biweight measure may be used; instead of the moving average, a robust LOWESS procedure may be used. Careful data analysis may allow researchers to identify the gross errors and thus take out much of their bite. This may partially explain why such methods have traditionally not been used in behavior analysis. The more recent ability to use automated data gathering systems, with very large data sets being automatically generated and summarized, inhibits the careful manual manipulation and inspection of the initial data, and increases the importance of the robust approaches.

In this talk I shall review the concepts of resistance and robustness, and give a few examples of simple robust statistics that are alternatives to the familiar summaries of center and spread. I shall demonstrate their importance on examples from rat and mouse exploratory behavior. I shall then present a robust smoothing procedure of the digitized path of a rat, and demonstrate the importance of using such robust procedures for that purpose.

© 2000 Noldus Information Technology b.v.