Software spending is increasing within the DoD, NASA, and other technologically advanced organizations, with significant effects on program budgets. Cost estimators must have the best tools available. However, many current models are problematic due to inaccuracy and unavailability of the input parameters, the technical expertise and expense required to operate them, and the difficulty in explaining their outputs. Two databases were analyzed: 60 NASA/JPL software projects and 116 projects from the International Software Benchmarking and Standards Group database. Models developed using ordinary least squares regression employed parameters representing the presence of project characteristics. The models' predictive characteristics were compared based on the source database, project size, model transformations, and the variable combinations. COCOMO 81 estimates were calculated for comparison for the NASA/JPL projects. Few of the models met the mean absolute percentage error (MAPE) standard of 25 percent or less; however, managers may find mean error (ME) to be a better metric for evaluating software cost models. ME results for the NASA/JPL projects suggest that managers may prefer the simpler categorical models to the more complex models for smaller programs. The best of these models had 223.4 person-months, or 86.3 percent less error on average than the COCOMO based regression. | A Hedonic Approach To Estimating Software Cost Using Ordinary Least Squares Regression And Nominal Attribute Variables by Marc D Ellis, Paperback