S. Amasaki (Japan)
Project management, Software engineering, Cost estimation, Regression models
Effort estimation models are widely investigated because they have an advantage over expert judgment in terms of objectivity and repeatability. Linear regression models are the most major methods used in the past study. While these studies carefully determined predictor variables and model formulation, error distributions are fewer considered. Furthermore, characteristics of linear regression models using different error distributions have not studied with actual datasets. This study compared log-normal and Gamma regressions for effort estimation in terms of their predictive performance. Both regressions were examined with multiple datasets and two formulation approaches. As a result, it was found that log-normal and Gamma regressions have contrasting characteristics though the difference is diminished when uncertainty of effort is well explained by predictor variables. Furthermore, it was found that which error distribution is favored depends on what one wants to estimate. These results contribute some suggestions to effort estimation model construction.
Important Links:
Go Back
