How Does the Dependence between Inspectors Impact Capture-Recapture Estimation? An Empirical Study

Abstract

Studies on Capture-recapture (CRC) method in software engineering suggested that the accuracy of the estimates were still far from satisfying. Although debates exist, researchers believe the dependence between different inspectors is one of the factors impacting the accuracy of CRC estimates. In this study, we apply linear correlation to measure the degree of dependence (measured by correlation coefficient, i.e., COV) between inspectors, based on which, we investigate the relation between COV and estimation accuracy (measure by Relative Error, i.e., RE) using the Unconditional Maximum Likelihood Estimator (Mt-UMLE) through nine sets of data from five papers. Besides, the confidence interval (CI)-(which is also an important indicator to reflect CRC estimating accuracy), may also be impacted by COV, to this end, we also investigated the relation between COV and CI. The major results show that: a) It is clear that there is an obviously inverse proportion on COV and CI, and when the COV=0, the radius of CI is almost in interval [1, 6] (exclude the outliers). b) There has also an inverse proportion with COV and RE, and the absolute value of RE increased with the absolute value of COV increased. So we can conclude that the accuracy of CRC estimates had been impacted by the dependence of inspectors. Our study suggested that the correlation cannot be ignored in the study of CRC.