Comments on study of SKR/T versus T2 formula.

摘要

In their article on the SRK/T formula, Sheard et al. used a variety of statistical techniques to bolster their claim that their alternative model (the T2 formula) significantly improves prediction accuracy. One claim worth examining is that the prediction error of the T2 formula was 9.7% less than the prediction error of the SRK/T formula, based on the slope of a regression line between the 2 errors. When performing linear regression, it is critical to analyze the distribution of residuals for normality. If there is a pattern to the residuals, the basis on which linear regression is done is not present. The scatterplot in Figure 5 of the article shows clear evidence of a violation of this assumption; particularly for points with greater than ?00 diopter (D) error measurements, many more points are on one side of the regression line than the other. Most of these outlier points appear to have very similar errors using either model. Thus, the basis of their claim of better predictions by T2 based on the slope of the regression line is invalidated by the analysis of the residuals. Furthermore, 1 outlier, which appears to be more than -2.00 D error for SRK/T and less than -1.00 D error for T2, might be unduly influencing the regression slope.