A replication of the use of regression towards the mean (R2M) as an adjustment to effort estimation models

Abstract

The paper performs an independent replication of the Jorgensen et al. study that advocates exploiting a phenomenon known as regression to the mean for software project productivity when predicting software project effort. We used two further industrial data sets in which we compare accuracy levels with and without this adjustment. Our results were broadly consistent with those from the Jorgensen study. Using the R2M resulted in a small increase in predictive accuracy. For one data set it was necessary to first partition it into more homogeneous subsets. Also when there was very weak correlation between predicted and actual productivity using the sample mean was the least bad strategy. We believe that independent validation of results is an important activity. Specifically our results add further support for the R2M approach in that there is a small, but positive, effect upon prediction accuracy. By combining results from both studies we observe a consistency across all 7 data sets