Figure 3

Illustration that within a meta-analysis the tertile approach will tend to a linear model. The SESAW data [3] were split into 20 sub-samples (A-T), each with n = 75. This plot shows four meta-analyses which combine an increasing number of the sub-samples (A, A-G, A-M, and A-T). The linear model approach is consistent and approaches the equivalent analysis using the full data. Conversely, the tertile approach becomes increasingly bumpy, as each sub-sample has data-dependent tertile cut-points. In the limit, as illustrated, the tertile combined analysis will tend towards the linear model approach.