Behavioural Data: Increased Positive Distractor Effect under Contralateral MIP Stimulation
A value-based decision-making task was used, in which participants were presented with two choosable options (high-value option HV or low-value option LV) and one distractor (D; Figure 1). As explained in the Methods, we looked at the impact of TMS to either MIP or a nearby control site (in the vicinity of area V5 or MT). However, both experiments also included an additional type of control condition: trials in which no TMS was applied.
First, we consider behavioural performance in the control situation in the absence of any TMS. To do this we combined data from Non-TMS trials of both MIP and MT sessions. All participants displayed above chance performance in both the main decision-making task, and the “matching” trials (additional trials that aimed to prevent participants from ignoring the identity of the distractor; see Methods), with an average accuracy of 72.84% (SD = 6.71; Figure 2a), and 70.14% (SD = 8.75) respectively, indicating that all participants followed task instructions. The average RT of the main task was 887.71ms (SD = 102.17ms; Figure 2b). There were no differences in either accuracy, t(30) = 1.14, p = 0.265, Cohen’s d (d) = 0.14, 95% confidence interval (CI) = [-0.01, 0.03], or reaction time (RT), t(30) = 0.64, p = 0.525, d = 0.11, CI = [-23.61, 45.33], between Non-TMS trials in MIP and MT sessions (Figure 2a, b).
In general larger distractor values should promote more accurate choices when decisions are hard (positive D-HV effect on trials with small HV-LV value difference) and, in contrast, they should impair choice accuracy when decisions are easy (negative D-HV effect on trials with large HV-LV value difference) 6,21. In other words, there should be a negative (HV-LV)(D-HV) interaction effect. We tested whether this was the case in the current experiment by applying the same GLM (GLM1) as in Chau and colleagues (2014, 2020). In particular, it involved the following terms: the difference in value between the two available options (HV-LV), their sum (HV+LV), the difference between the distractor value and the high-value option (D-HV), and the interaction term (HV-LV)(D-HV). On Non-TMS trials, there was a positive HV-LV effect (t(30) = 17.09, p < 0.001, d = 3.07, CI = [0.69, 0.88]; Figure 2c) and a negative HV+LV effect (t(30) = -4.35, p < 0.001, d = -0.78, CI = [-0.38, -0.14]), suggesting that more accurate choices were made on trials that were easier and consisted of options with poorer values. There was no D-HV effect (t(30) = -0.95, p = 0.350, d = -0.17, CI = [-0.17, 0.06]) but critically there was a negative (HV-LV)(D-HV) interaction effect (t(30) = -3.52, p = 0.001, d = -0.63, CI = [-0.16, -0.04]). To further examine the pattern of the negative (HV-LV)(D-HV) effect, we median split the data according to HV-LV levels and applied GLM2 to test the critical D-HV effect. On hard trials with small HV-LV, there was a positive D-HV effect (t(30) = 2.62, p = 0.014, d = 0.47, CI = [0, 0.02]; Figure 2d), whereas on easy trials with large HV-LV, there was a negative D-HV effect (t(30) = -3.15, p = 0.004, d = -0.57, CI = [-0.01, 0]).
In addition, divisive normalisation models also predict that the size of the negative distractor effect should be smaller when the total HV+LV is large 21. This is because the variance in D then makes a smaller contribution to the overall normalisation effect that depends on HV+LV+D. If present, such an effect can be demonstrated by a positive (HV+LV)D interaction effect. This was indeed the case in the data of the current study (Supplementary Figure S1). In summary, these results are broadly consistent with recent demonstrations that both positive and negative distractor effects are reliable and statistically significant but that they predominate in different parts of the decision space 6,21.
Finally, we note that, in addition to the positive and negative distractor effects on choices between HV and LV, there is a third route by which the distractor can affect decision making – salient distractors can capture attention and eventually be chosen19,20. In an additional analysis reported in Supplementary Figure S2, we showed that an attentional capture effect by the distractor was also present in our data.
Next, we examined whether MIP has any role in generating distractor effects by comparing TMS and Non-TMS trials. In addition, because neurons in the intraparietal sulcus mostly have response fields in the contralateral side of space, we took care to consider whether any impact of TMS might be particularly robust when the distractor was presented contralateral to the MIP region that was targeted with TMS. We therefore split the trials according to whether the distractor was located on the contralateral side of space to the TMS. In other words, each analysis involved approximately one-fourth of the data that was split according to Stimulation (TMS/Non-TMS) and Distractor Location (ipsilateral/ contralateral side). Ideally, the trials should be split further according to difficulty, as indexed by the HV-LV difference, in order to isolate the negative distractor effect on easy trials that may be linked to MIP. However, that would mean that the analyses would rely on approximately one-eighth of the data and run the risk of becoming under-powered due to the small number of trials. Hence, we adapted GLM1 by removing the (HV-LV)(D-HV) term and keeping the remaining terms – HV-LV, HV+LV, D-HV (GLM3). We should now expect the absence of a D-HV main effect in the control Non-TMS data because the positive and negative distractor effects cancel out one another when they are no longer captured by a negative (HV-LV)(D-HV) interaction term. However, if TMS disrupts the negative distractor effect specifically and spares the positive distractor effect, then a positive D-HV effect should be revealed in the TMS data.
The results of GLM3 confirmed once again that in the control data, on average, higher accuracy was associated with larger HV-LV differences (t(30) = 17.04, p < 0.001, d = 3.06, CI = [0.76, 0.97]; Figure 3a) and lower HV+LV sum values (t(30) = -3.50, p = 0.001, d = 0.63, CI = [-0.33, -0.09]; Figure 3b). The average impact of D-HV on accuracy was not significant (t(30) = 0.69, p = 0.496, d = -0.12, CI = [-0.08, 0.16]; Figure 3c), which, as already explained above, is consistent with the simultaneous presence of both positive and negative distractor effects on the hard and easy trials respectively that we have previously demonstrated (Figure 2d). Next, a Site (MIP/MT) x Stimulation (TMS/Non-TMS) x Distractor Location (contralateral/ipsilateral) ANOVA was applied to the beta values of each predictor (Figure 3). We focused on examining the three-way interaction relating to the distractor value, in which a significant effect would suggest a robust TMS effect when it was applied to a specific brain region and when the distractor was presented at a specific location. When examining three-way interactions of this type, we found no significant effects associated with predictors that did not incorporate the distractor value D such as the HV-LV predictor (F(1,30) = 1.50, p = 0.230, ηp2 = 0.05; Figure 3a; Table 1) or the HV+LV predictor (F(1,30) = 3.06, p = 0.090, ηp2 = 0.09; Figure 3b). This suggests that these interactions were unaffected by TMS. Critically, however, the D-HV predictor showed a statistically significant three-way interaction (Site x Stimulation x Distractor Location: F(1,30) = 4.44, p = 0.044, ηp2 = 0.13; Figure 3c). This is consistent with a relative increase in the positive distractor effect (previously associated with vmPFC6,22) at the expense of the divisive normalization effect associated with intraparietal areas such as MIP. No other main or two-way interaction effects of D-HV were observed in the ANOVAs, F < 0.87, p > 0.358 (Table 1).
In order to explore the three-way interaction on the D-HV predictor, we split the data into contralateral and ipsilateral sets, i.e. trials in which the distractor was presented contralaterally/ipsilaterally to the TMS pulse, and performed a Site x Stimulation ANOVA on each set of trials. In each ANOVA, the terms associated with the opposite side were also entered as covariates. We found no Site x Stimulation effects in the ipsilateral data set (F < 1.87, p > 0.180). Since traditional frequentist statistics are less ideal for supporting claims of null effect, we performed Bayesian tests to compare the D-HV effect between TMS and Non-TMS trials in the ipsilateral data set (when the distractor had been presented ipsilateral to the MIP TMS). The results confirmed that there was an absence of TMS effect when it was applied over ipsilateral MIP (BF10 = 0.209) or MT (BF10 = 0.271). In the contralateral data, we found a significant Site x Stimulation interaction effect, F(1,26) = 4.99, p = 0.034, ηp2 = 0.16 (no other effects were significant, F < 0.73, p > 0.400), indicating that the Site x Stimulation x Distractor Location effect was driven by the contralateral presentation condition (when the distractor was presented contralateral to the MIP TMS). In other words, this is consistent with a relative increase in the positive distractor effect (previously associated with vmPFC) at the expense of the divisive normalization effect associated with intraparietal areas such as MIP that occurs mainly when distractors are presented contralateral to the TMS site.
To clarify the nature of this effect, we split the contralateral data further into MT and MIP sets and repeated the ANOVA, entering only the Stimulation factor and including all other conditions as covariates, on each set. We found a significant effect of TMS on MIP conditions (F(1,24) = 4.32, p = 0.049, ηp2 = 0.15), with TMS trials showing a more positive D-HV effect than Non-TMS trials. The MIP-TMS effect became even clearer after the grey matter volume (GM) of the same region was also entered as a covariate (F(1,23) = 7.02, p = 0.014, ηp2 = 0.23; the next section explains the importance of considering the GM and explains how the GM indices were obtained). In contrast, we found no effect in MT conditions (F(1,24) = 1.24, p = 0.277, ηp2 = 0.05), and this lack of TMS effect was confirmed by an additional Bayesian test (BF10 = 0.225). The results remained similar even after entering the GM of MT as an additional covariate (F(1,23) = 0.19, p = 0.664, ηp2 = 0.01). These results suggest that TMS of MIP had a significant impact on promoting the positive distractor effect on accuracy at the expense of the opposing negative (divisive normalization) distractor effect dependent on intraparietal sulcus areas such as MIP 6,9. The effect was especially clear when the distractor was presented contralaterally to the MIP TMS site.
While our primary focus is on accuracy as an index of response selection, our diffusion model21, like most diffusion models, suggests that in many cases, factors that increase response selection accuracy will also increase response selection RT. This was true in the present case (Supplementary Figure S3).
Finally, as explained in Supplementary Figure S1, the negative distractor effect should have become smaller when HV+LV was large (on such trials the distractor constitutes a smaller part of the total value of the stimuli and ultimately it is this total value that determines divisive normalization). This was revealed as a positive (HV+LV)D interaction effect. In Supplementary Figure S4, we showed the (HV+LV)D effect also became marginally less positive after MIP-TMS (F(1,30) = 3.77, p = 0.062, ηp2 = 0.112).
In summary, the D-HV term indexes an important aspect of the influence of the distractor on behaviour. At baseline it is associated with two significant absolute effects; large distractor values are associated with higher accuracy when decisions are difficult (Figure 2d, left) and they are associated with lower accuracy when decisions are easy (Figure 2d, right). Thus, the balance of the distractor effect changes across the decision space defined by the choice values (such as the differences in their values). The balance of distractor effects also significantly changes with the disruption of MIP using TMS; the positive distractor effect becomes stronger at the expense of the negative distractor effect (Figure 3c).
MRI Data: VBM Confirms Link between MIP and the Impact of TMS on the Distractor Effect
So far, we have provided evidence that MIP is causally related to the negative distractor effect because the antagonistic positive distractor effect emerged prominently and to a significantly greater extent after MIP was disrupted. To investigate the importance of MIP, and the impact of its disruption further, we sought an explanation of individual variation in effects. We might expect individual variation in MIP volume to be related to the degree of TMS modulation of the distractor effect. Recent studies suggested that those with smaller GM in the target region also demonstrate stronger behavioural changes after receiving TMS 23. Thus, in individuals with larger MIPs, the negative distractor effect should be less susceptible to TMS disruption and the opposing positive distractor effect should appear weaker.
To test this, we performed a voxel-based morphometry (VBM) analysis to examine the relationship between GM and the TMS effect on participants’ decision-making. The analysis was focused on parietal and occipital cortex in the same hemisphere to which TMS had been applied. The same GLM3 (HV-LV, HV+LV, D-HV) as described in the behavioural analysis was used. In order to extract the effects of TMS, we subtracted the beta values associated with Non-TMS trials from those associated with TMS trials (in contralateral conditions). This was conducted once for MT conditions, and once for MIP conditions. Interestingly, the effect of MIP TMS on the distractor value’s (D-HV) impact on decision accuracy showed a negative relationship with MIP GM (p = 0.029, TFCE corrected, centred around MNI X(-30), Y(-52), Z(42), Figure 4a). This implies that in individuals with larger MIP GMs, there was less difference between TMS and Non-TMS trials (less relative increase in the positive distractor effect at the expense of the negative distractor effect), because TMS had a weaker impact on disrupting the MIP-related negative distractor effect. No significant GM differences were observed in any other parietal and occipital regions (p > 0.170).
We followed these tests up by extracting the GM at MIP and MT. First, we illustrate the VBM results again by showing that there was a significant partial correlation between MIP GM and MIP TMS impact on the D-HV effect (r(27) = -0.61, p < 0.001; Figure 4b, left panel), after controlling for the other explanatory variables entered in the VBM (i.e. TMS effects on HV-LV and HV+LV). In addition, the correlation remained significant even without controlling for HV-LV and HV+LV (r(29) = -0.37, p = 0.043). Since visual inspection revealed a potential outlier with small MIP GM and a large TMS effect on the D-HV predictor, we repeated the partial correlation analysis by excluding this data point. The results did not change qualitatively (r(26) = -0.55, p = 0.002).
Next, we ran three additional analyses to demonstrate that the correlation was specific to the MIP GM only when TMS was applied to MIP itself (but not when TMS was applied to the control MT region). First, there was no correlation between MIP GM and TMS effect when it was estimated in the control sessions where MT was stimulated (r(27) = -0.05, p =0.816; Figure 4b right). Second, despite finding no effect in MT in the VBM analysis, we extracted the GM in the control MT region and tested whether it was related to the TMS effect in the experimental MIP session. We found no significant correlation (r(27) = -0.19, p = 0.325; Figure 4c left) even though this analysis is less conservative than the VBM analysis illustrated in Figure 4a. Finally, there was also no relationship between the GM of the control MT region and TMS effect extracted from the control MT-TMS sessions (r(27) = -0.03, p = 0.871; Figure 4c right). The conclusion that there was a null effect was supported by supplementary Bayesian analyses (BF10 < 0.275 in all three relationships illustrated in Figure 4b right and 4c left and right).
Finally, we note that the correlation between MIP GM in each individual and the impact of MIP TMS on the D effect in behavior was more strongly negative than the correlation between MIP GM in each individual and the impact of MT TMS on the D effect in behavior (i.e. comparing the correlations in Figure 4b left and right; z = -2.47, p = 0.014).
Eye-Tracking Data: TMS Affects Gaze Shifts between D and HV
Previous work has suggested that the positive distractor effect, which is prevalent on hard trials (Figure 2d) and which becomes more prominent when MIP is disrupted (Figure 3c), is linked to particular patterns of eye movement. Chau et al. (2020) showed that larger distractor values are associated with more gaze shifts between the D and HV options and, ultimately, more accurate HV choices are made. This suggests that accumulation of evidence in favour of the HV, as opposed to the LV, option is prolonged when D captures overt attention, and this eventually leads to more accurate decision-making. Similarly, in other settings, participants who are allowed extra time to revise their initial decisions tend, ultimately, to make more accurate decisions 24,25. Therefore, we tested whether the positive relationship between distractor value and D-to-HV gaze shift was replicable in the current study. We also tested whether the positive relationship became even stronger after the TMS disrupted the negative distractor effect and spared the positive distractor effect. Hence, we again applied GLM3 (HV-LV, HV+LV, D-HV) to predict gaze shifts between D and HV.
First, we showed that larger D-HV values were related to more gaze shifts between D and HV (t(30) = 4.75, p < 0.001, d = 0.85, CI = [0.03, 0.08]). The result is similar to that reported by Chau and colleagues 21. However, in the previous study the relationship between the difference in D/HV values, D-HV, and gaze shifts was apparent for D-to-HV gaze shifts whereas in the present analysis the association was with gaze shifts between HV and D in either direction. In addition, HV-LV and HV+LV had no clearly significant effect on the bidirectional gaze shifts (t(30) < 1.55, p > 0.131). We then performed a critical test that examined whether large distractor values were more strongly related to more gaze shifts between D and HV after MIP was disrupted using TMS. This was done by comparing the effect of D-HV on gaze shifts between HV and D using a Site (MIP/MT) by Side (contralateral/ipsilateral) by Stimulation (TMS/Non-TMS) ANOVA. This was analogous to the analysis in Figure 3c that examined how TMS modulated the distractor’s influence on choice behaviour. Note that 5 participants were excluded in this analysis due to the absence of gaze shifts between HV and D in some conditions of this ANOVA. The results showed a significant Site × Stimulation interaction effect, F(1,25) = 8.85, p = 0.006, ηp2 = 0.26. Follow-up ANOVAs with factors of Side and Stimulation revealed in MIP sessions the effect of distractor value on gaze shifts was significantly higher in TMS compared to Non-TMS trials (F(1,25) = 5.30, p = 0.030, ηp2 = 0.18), while in MT sessions there was no significant difference between TMS and Non-TMS trials (F(1,25) = 1.62, p = 0.214, ηp2 = 0.06).
To avoid excluding data from 5 participants due to splitting the data multiple times into small sets, we ran a similar analysis by collapsing the conditions of Side (contralateral/ipsilateral). Then we performed a two-way Site by Stimulation ANOVA to compare the effect of D-HV on gaze shift between HV and D. Again, the results showed a significant two-way interaction effect, F(1,30) = 6.73, p = 0.014, ηp2 = 0.18 (Figure 5a, right; Table 2). No effects were found when a similar ANOVA was performed to compare the effect of HV-LV and HV+LV on the gaze shifts between HV and D (F < 2.71, p > 0.110; Figure 5a, left and middle respectively).
The impact of the distractor value, D, on gaze shifts is sometimes more apparent when D-to-HV unidirectional gaze shifts are considered (as opposed to gaze shifts in either direction) 21. We examined these gaze shifts separately from those in the opposite direction (HV-to-D shifts). When we performed a two-way ANOVA, a Site by Stimulation interaction was only found in the D-to-HV direction (F(1,30) = 5.82, p = 0.022, ηp2 = 0.16; Figure 5b, right), but not the HV-to-D direction (F(1,30) = 1.03, p = 0.319, ηp2 = 0.03; Figure 5c, right). These results show that distractors with higher values lead to increased numbers of gaze shifts from the distractor to the high-value option, and importantly, that this effect is increased by applying TMS over MIP.
In the analyses above, we used the predictors HV-LV, HV+LV, and D-HV to predict gaze shifts in order to be consistent with the previous analyses of the choices participants took (Figure 3). However, in order to predict gaze shifts, it might be argued that the effects of individual option values, instead of differences in, or sums of, values of options are more easily interpretable. We therefore repeated the same analyses, but using the predictors HV, LV, and D, instead of HV-LV, HV+LV, and D-HV. The results do not change qualitatively (Supplementary Figure S5).
Finally, to explore the effects these gaze shifts had on decision-making, we used gaze shifts in all possible directions (HV to LV, LV to HV, LV to D, D to LV, HV to D, and D to HV; Figure 5d) to predict decision accuracy. We found that increased numbers of gaze shifts towards the low-value option were associated with lower accuracy (HV to LV: t(26) = -4.6, p < 0.001, d = -0.88, CI = [-0.21, -0.08]; D to LV: t(26) = -4.74, p < 0.001, d = -0.91, CI = [-0.19, -0.07]). Importantly, we found increased accuracy to be associated with increased gaze shifts from the distractor to the high-value option (D to HV: t(26) = 5.77, p < 0.001, d = 1.11, CI = [0.13, 0.27]). No other gaze shifts revealed significant results (p > 0.116). Together, these findings suggest that after TMS disrupts the negative distractor effect mediated by MIP, large distractor values are more influential in promoting D-to-HV gaze shifts which, ultimately, are associated with more accurate choices.