Error Propagation in Software Measurement & Estimation

The real challenge in any activity is to minimize as much as possible the error between an estimate and an actual value, whatever the phenomenon to be evaluated. When dealing with software, the number of proxies can be quite high: the application of an algorithm including one or more independent variables (measures) is finalized to provide one or more output variables (estimates) for a series of measures typically about effort, cost, time, quality or other aspects of the software being developed. Recently ISO proposed also a specific standard on Measurement (ISO/IEC 15939), with a glossary aligned to the Metrology field and to the International Vocabulary of Metrology (VIM). Estimation could be seem as a “black art”, but error is intrinsic in estimates and must be managed. Thus, regardless of the estimation model (algorithm) being used, practitioners must face the uncertainty aspects of such process: errors in initial measures do affect the derived metrics (or estimated values for indirect variables). Measurement theory does provide an accurate way to evaluate such error propagation for algorithmic derivation of variable values from direct measures, as in the GUM (Guide to the expression of Uncertainty in Measurement). Although some software estimation models already propose confidence ranges on their results, the formal application of error propagation can yield some surprising results, depending on the mathematical functional form underlying the model being examined. Moving from a previous paper, this one will discuss the propagation of errors in software measurement and shows with some application and examples based on some of the most common software measurement methods and estimation models as Function Point Analysis (FPA) for product sizing issues and COCOMO (Cost Construction Model) for effort and/or duration issues as well as other ones, updating also the discussion to new advancement in the Software Quality field, in particular about product NFRs (Non-Functional Requirements). Few cases and examples will be shown, in order to stimulate a critical analysis for methods and models being examined from a possibly new perspective, with regards to the accuracy they can offer in practice.