A software toolbox for data analysis and regression, considering data precision and numerical error propagation

摘要

An algorithm for data analysis and regression by orthogonalized-variable-based stepwise regression (SROV) has been developed and was implemented as a MATLAB toolbox. The program uses QR decomposition based on Gram-Schmidth orthogonalization, which is highly resilient to numerical error propagation, for regression. Variables are selected to enter the regression model according to their level of correlation with the dependent variable and they are removed from further consideration when their residual information gets below noise level. The use and benefits of SROV are demonstrated by two examples. The first one involves removing non-influential dimensionless groups from a regression model. In the second one the nonlinear terms that should be included in an optimal thermodynamic property correlation are selected.