Comparison of statistical power of classifiers when feature distributions within class are non-Gaussian. Each dataset included 1000 features per subject, where features were distributed according to a non-Gaussian distribution within each class. Results shown in Panels (a) - (c) were based on k= 1% and n= 100, 150 or 200. Results shown in Panels (d) - (f) were based on n= 150 and k= 0.5%, 1% or 5%. Results are based on 100 simulated datasets. See Additional File 1: Supplemental Figure S1 (b) for similar results when n= 50 and k= 1%.