Comparison of approaches for estimating non-linear relationships using the SESAW study [
(a) Comparison of a simple linear model, fractional polynomial (of which the best fitting was equivalent to the simple linear model), linear splines, tertiles and a non-parametric smoother (see Table 1 for the respective AICs to assess model comparison). (b) As in Figure 2(a) with an extension to the y-axis to show the complete range of BMI and the observed data plotted (n = 1462 points). We see visually the result of comparing the AICs in Table 1 that due to the large variance in BMI scores there is no evidence for anything more complicated than a simple linear model. Further, there is nothing statistically to choose between the linear and tertile fits. However, the linear model has the benefit of not being data-dependent.