re: the impact of using parametric tests when population distributions fail to approximate the normal distribution. I believe the direction of error can go either way (type I or type II), and depends on the magnitude of departure from good fit. For example, run Pearson and Spearman correlation coefficients on an exponential function: Spearman's will give a correlation coefficient of 1.0 while Pearson's will (misleadingly) be lower. Bartlett's test for homogeneity of variances among groups in ANOVA reportedly fails to distinguish between non-normal distribution vs. unequal variances. The Anscombe quartet (Anscombe FJ. Graphs in Statistical Analysis. AMERICAN STATISTICIAN 1973;27:17-21) nicely illustrates what various distributions can do to linear regression if one doesn't pay attention to regression diagnostics... A few textbooks make claims about robustness of specific methods, and some seem to contradict others, but not all that many offer that with much cited in the way of empirical evidence. David Birnbaum, PhD, MPH Clinical Assistant Professor Dept. of Health Care & Epidemiology University of British Columbia, Canada