Because nonlinear regression is an iterative method, computational errors in a single cycle can be propagated into the final results. Although such propagation of errors is rarely a problem, it sometimes crops up when infinite series functions or built-in computer library functions are used to construct models , The sine function on some computers sometimes causes propagation of errors, as can iterative or finite difference algorithms within the model. Error propagation can usually be eliminated or minimized by doing the computation in double precision.
In spite of their rather infrequent occurrence, we must make sure that computational errors do not bias the results of regression analyses. This can be done by analyzing sets of data simulated from the model that include various amounts of random noise (see Section 3.A), approximating the noise levels to be expected in the experiment. The parameters from such data sets are known, and the accuracy of the nonlinear regression method in determining them can be assessed before moving on to real experimental data [1, 8],
Was this article helpful?