Multiplicative regression models are characterized by stochastic errors that operate multiplicatively, rather than additively, on the model predictions. Both the lot-midpoint representation of the learning curve model and many cost-estimating relationships are examples of multiplicative regression models.

Several strategies are available for estimating multiplicative regression models, including minimization of the sum-of-squared errors in predicting the logarithm of cost, and minimization of the sum-of-squared percentage errors in predicting the level (not logarithm) of cost. In our view, the choice of estimation method should be guided by the statistical properties of the resulting estimators, not the intuitive appeal of the fitting criterion being optimized. In particular, we recommend that serious consideration be given to iteratively reweighted least squares (IRLS), which gives both consistent estimates and a tractable covariance matrix under minimal distributional assumptions. Moreover, IRLS is equivalent to minimizing the quasi-likelihood function.