In this note we discuss in general how to test the implementation code of statistical tests, and then we treat in detail the case of the Kolmogorov-Smirnov test. It will be shown that some "obvious" expected properties, like the flatness distributions of p-values from repeating drawings from the same parent distribution, are not indeed reproduced even in absence of bugs in the code, due to either asymptotic approximations in the formulas used to compute the p-value, or to the discreteness of the distance distribution in the case of direct Monte Carlo evaluation of the p-value. This makes the code-testing more complicated.
Code-testing of statistical test implementations / F., James; A., Pfeiffer; A., Ribon; P., Cirrone; S., Donadio; S., Guatelli; A., Mantero; B., Mascialino; L., Pandola; S., Parlati; M. G., Pia; Viarengo, Paolo. - ELETTRONICO. - (2003), pp. 100-113. (Intervento presentato al convegno Phystat2003 SLAC, Statistical Problems in Particle Physics, Astrophysics and Cosmology tenutosi a Stanford (CA), USA nel 8-11/9/2003).
Code-testing of statistical test implementations
VIARENGO, PAOLO
2003
Abstract
In this note we discuss in general how to test the implementation code of statistical tests, and then we treat in detail the case of the Kolmogorov-Smirnov test. It will be shown that some "obvious" expected properties, like the flatness distributions of p-values from repeating drawings from the same parent distribution, are not indeed reproduced even in absence of bugs in the code, due to either asymptotic approximations in the formulas used to compute the p-value, or to the discreteness of the distance distribution in the case of direct Monte Carlo evaluation of the p-value. This makes the code-testing more complicated.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.