Regression and dimensional analysis in coastal engineering: some pitfalls Available to Purchase

Most civil engineers are familiar with the concept of regression analysis. Fewer seem aware of the importance of model specification which includes defining the correct functional form of the regression equation. Furthermore, most engineers appear to equate regression with the process of minimising the sum of the squares of the differences between the measured values and the values predicted by the equation fitted to the data. But providing a least-squares fit to the data is only one of several possibilities. While its popularity is easily understood, the process suffers from the disadvantage that it is not robust: it is sensitive to the presence of outlying data points. The alternative regression procedure of minimising the absolute deviations of the fitted equation from the data is better in this respect. The paper focuses on linear regression analysis, particularly as it is used in conjunction with dimensional analysis, in the field of coastal engineering. However, the issues raised with regard to linear regression are equally relevant to non-linear regression. The reader is also reminded of the hazards of spurious correlation and examples are given both of poor and good model specification.