Appendix 3—figure 2. The distractor effect was modulated as a function of difficulty and the total value of the chooseable options.

(a) Here the data have been analysed with a more complete GLM (GLM5) including all main effects and two- and three-way interactions. Across all experiments the (HV-LV)D interaction is negative – the distractor has a positive effect on decision accuracy when decisions are difficult but a negative effect when decisions are easy. This resembles the effects illustrated in Figures 3–5. Moreover, across all experiments the (HV+LV)D interaction is positive – when the total value of the choosable options changed from large to small the distractor effect also changed from positive to negative. The divisive normalization model predicts that the negative impact of D is greatest when HV+LV is small (Figure 1) but its impact is reduced when HV+LV is large and so the positive distractor effect predicted by the interacting diffusion process model predominates when HV+LV is large. (b) The same two key interaction terms, (HV-LV)D and (HV+LV)D, indexing the positive distractor and divisive normalization effects are illustrated for each component experiment that was included in panel a. ***p<0.001. Error bars indicate standard error.