The problem of the identification of models defined by complex systems is frequently encountered in chemical engineering. Due to convergence difficulties and/or an excessive computational burden, simplified models are often employed. Similarly approximate models are used in a preliminary phase for the estimation of the most significant parameters. In these cases the residuals (i.e. the difference of the experimental data to the values predicted by the approximate model) do not belong to a well-defined distribution function. Thus the usual regression methods, such as those based on maximum likelihood, can sometimes lead to seriously biased estimates. A new regression technique for avoiding either under- or overestimation due to compensation or cumulation of experimental errors and model deviations is presented in this paper.