An evaluation of the Statistical Convertibility of Function Points into COSMIC Function Points

Context: Since the introduction of COSMIC Function Points, the problem of converting historical data measured using traditional Function Points into COSMIC measures have arisen. To this end, several researchers have investigated the possibility of identifying the relationship between the two measures by means of statistical methods. Objective: This paper aims at improving statistical convertibility of Function Points into COSMIC Function Points by improving previous work with respect to aspects --like outlier identification and exclusion, model non-linearity, applicability conditions, etc.-- which up to now were not adequately considered, with the purpose of confirming, correcting or enhancing current models. Method: Available datasets including software sizes measured both in Function Points and COSMIC Function Points were analyzed. The role of outliers was studied; non linear models and piecewise linear models were derived, in addition to linear models. Models based on transactions only were also derived. Confidence intervals were used throughout the paper to assess the values of the models' parameters. The dependence of the ratio between Function Points and COSMIC Function Points on size was studied. The union of all the available datasets was also studied, to overcome problems due to the relatively small size of datasets. Results: It is shown that outliers do affect the linear models, typically increasing the slope of the regression lines; however, this happens mostly in small datasets: in the union of the available datasets there is no outlier that can influence the model. Conditions for the applicability of the statistical conversion are identified, in terms of relationships that must hold among the basic functional components of Function Point measures. Non-linear models are shown to represent well the relationships between the two measures, since the ratio between COSMIC Function Points and Function Points appears to increase with size. Conclusion: In general, it is confirmed that convertibility can be modeled by different types of models. This is a problem for practitioners, who have to choose one of these models. Anyway, a few practical suggestions can be derived from the results reported here. The hypothesis that one FP is equal to one CFP causes the biggest conversion errors observed and is not generally supported. All the considered datasets are characterized by a ratio of transaction to data functions that is fairly constant throughout each dataset: this can be regarded as a condition for the applicability of current models; under this condition non-linear (log-log) models perform reasonably well. The fact that in Function Point Analysis the size of a process is limited, while it is not so in the COSMIC method seems to be the cause of non linearity of the relationship between the two measures. In general, it appears that the conversion can be successfully based on transaction functions alone, without losing in precision.