A Comparison of LAD and OLS Regression for Effort Prediction of Software Projects

Accurate effort prediction of software projects is of concern to portfolio managers, customers, vendors as well as project managers. Ordinary least squares (OLS) regression is widely used to create software prediction models, and it seems to perform just as well or better than most other, non-regression, prediction models. Software data sets may however exhibit certain characteristics that do not always comply with the requirements of OLS. In particular, we may suspect industrial software data sets to be messy due to the cost and difficulty of gathering data. Thus, we should probably be prepared to expect the distribution of residuals to exhibit heavy tails, i.e. high kurtosis, and to be skewed. Under these circumstances, OLS regression may be less efficient than certain robust regression methods. In particular, it can be shown that the Least Absolute Deviation (LAD) regression method is more efficient than OLS when the distribution of the residuals exhibits high kurtosis. In general, it is also useful to apply more than one method to confirm that the initial model specification and results are reasonable. In this paper, we investigate empirically if LAD is more efficient than OLS on a data set of 87 ERP projects. The results suggest that: i) for this particular data set, LAD does not seem more efficient than OLS, ii) LAD does, however, confirm the initial OLS prediction model.