Fault localization is a major activity in software debugging. Many existing statistical fault localization techniques compare feature spectra of successful and failed runs. Some approaches, such as SOBER, test the similarity of the feature spectra through parametric self-proposed hypothesis testing models. Our finding shows, however, that the assumption on feature spectra forming known distributions is not well-supported by empirical data. Instead, having a simple, robust, and explanatory model is an essential move toward establishing a debugging theory. This paper proposes a non-parametric approach to measuring the similarity of the feature spectra of successful and failed runs, and picks a general hypothesis testing model, namely the Mann-Whitney test, as the core. The empirical results on the Siemens suite show that our technique can outperform existing predicate-based statistical fault localization techniques in locating faulty statements.