A Study of the Predictive Accuracy of Alternatively Estimated Statistical Cost Models

The focus of this paper, which draws liberally from an earlier CNA paper that also employed Monte Carlo methods, is on the predictive accuracy of cost-estimating relationships whose parameters are estimated by both ordinary least squares regression and non-linear least squares, first when the error structure is known to be multiplicative and then when it is known to be additive. The multi-predictor model that served as the basis of the Monte Carlo simulations is representative of models that cost analysts frequently deal with. Results suggest that predictions based on non-linear least squares are substantially better than those based on ordinary least squares when the errors are additive, and are only slightly inferior when the error term is multiplicative.