Evaluation of several Efron bootstrap methods to estimate error measures for software metrics

A narrow confidence interval of a sample statistic or a model parameter implies low variability of that statistic, and permits a strong conclusion to be made about the underlying population. Conversely, the analysis should be considered inconclusive if the confidence interval is wide. Efron´s (1992) bootstrap statistical analysis appears to address the fact that many statistics used in software metrics analysis do not come with theoretical formulas to allow accuracy assessment. In this paper we will present preliminary results on an empirical analysis of the reliability of several Efron nonparametric bootstrap methods in assessing the accuracy of sample statistics in the context of software metrics. In particular, we focus on the standard errors and 90% confidence intervals of five basic statistics as a tool to evaluate the Bootstrap. It was found confidence intervals for mean and median were accurately estimated, those for variance grossly under-estimated with skewness and kurtosis grossly over-estimated.