Chapter 5. Integrating Prospective and Retrospective Cues to the Sense of
7.2. Integrating Prospective and Other Cues to Sense of Agency
In line with previous proposals (Moore & Fletcher, 2012; Synofzik et al., 2013), our work shows that the SoA involves integration of a variety of cues. These cues may be of different types, e.g. sensorimotor, or conceptual, and may become available at different times, e.g. at the time of the action, or of the outcome. Critically, the weighting of these cues, and whether they have independent or interrelated effects
on SoA is dynamically updated, and highly dependent on context and availability of cues. SoA is highly flexible and adaptable to current context and task demands.
7.2.1. Role of Awareness of Biases and Choice
Our work shows that the effects of action selection on SoA are independent of whether one is aware, or not, of the stimuli that influence action selection. We showed a consistent reduction in SoA due to disruptions to action selection induced subliminally (Chapter 4), and supraliminally (Chapters 2 & 3). Previous studies used subliminal priming of actions to ensure that any effects found could not be attributed to participants knowing that their actions were manipulated. Moreover, it was unclear whether selection fluency might interact with awareness. We demonstrate that unawareness of biases is not necessary to investigate the effects of action selection to SoA. [But see Damen et al. (2014), and section 7.3 below.]
The effects of selection fluency on agency were also independent of choosing freely between action alternatives, or following instructions (Chapter 4 & Experiment 2 in Chapter 2; cf. Wenke et al., 2010). Moreover, we found that free choice trials were associated with a higher SoA than forced choice, in the subliminal priming study (Chapter 4). However, there was no robust effect of choice when using supraliminal flankers (Chapter 2, Experiment 2). The combination of free and forced choice trials, combined with awareness of the flankers might have led participants to never feel very free, even when they could choose between a left or right key press. Freely choosing our actions, relative to following instructions, is thought to be central to our SoA. Indeed, it has been linked to SoA (Barlas & Obhi, 2013; Caspar et al., 2016; Wenke et al., 2010). However, it has been suggested that our sense of freedom is only one aspect of SoA (Pacherie, 2008). Therefore, it might serve as another independent cue to SoA.
7.2.2. Lack of Relation to Outcome Predictions
The EEG study in Chapter 4 further showed that the influence of selection fluency on SoA is not dependent on changes in outcome processing. We found that outcome monitoring, indexed by the outcome-locked feedback-related negativity (FRN)
component, was associated with agency ratings (Figure 4.8.b, p. 134). However, we did not find any direct effects of selection fluency on this measure of outcome processing. Additionally, we found no relation between the action-locked CRN and the outcome-locked FRN, even though both were associated with agency ratings. Therefore, the CRN and FRN components reflect two independent processes that, respectively, make prospective and retrospective contributions to SoA (see Figure 4.9, p. 136).
In line with this, the multi-study analysis described in Chapter 5 showed that selection fluency can have a general effect on SoA throughout instrumental learning, when participants rely on action-outcome interval to guide their agency judgements (experiments in Group 2, see Figure 5.5, p. 159). This general effect of selection fluency suggests that we learn to use it as a heuristic cue to SoA. Typically, dysfluent selection will be predictive of unsuccessful or unexpected outcomes. These effects may also be associated with, or the heuristics mediated by, the affective consequences of conflict, or dysfluency. This would be in line with other fluency effects on judgements (Alter & Oppenheimer, 2009; Winkielman et al., 2015).
7.2.3. Interactions with Outcome Monitoring
When faced with a new environment, an optimal cue integration account might have predicted that the poor reliability of action-outcome knowledge in a new environment could be compensated by a greater influence of selection fluency to SoA. If so, as reliable action-outcome contingencies were learnt, less reliable cues based on selection fluency should have decreasing effects. Our multi-study analysis (Chapter 5), in fact, showed that any interactions between instrumental learning and selection fluency were in the opposite direction. That is, the effects of selection fluency on SoA
increased during instrumental learning (see Figure 5.5, p. 159). Interestingly, this
occurred only when participants were focused on learning action-outcome contingencies (the experiments from Group 1, described in Chapters 2 and 4).
We ruled out the possibility that people learned, in a context-specific manner, to use action selection processes as a proxy for true action-outcome contingency. This
account would not be consistent with the aforementioned finding of a general, heuristic effect of selection fluency on SoA, even in a new environment (Group 2 experiments). Instead, we propose the novel suggestion that this interaction arises because action selection processes directly influence instrumental learning rates. That is, conflict during action selection could disrupt the learning of action-outcome associations, which would resulting in a slower learning rate (in linking dysfluent action - outcome), compared to fluency (i.e. linking fluent action - outcome). This proposal would be compatible with a heuristic use of selection fluency as a cue to SoA, and may perhaps even underlie the acquisition of this heuristic. Yet, it remains speculative, since we only measured agency ratings, rather than directly test knowledge of action-outcome contingencies. Notwithstanding that limitation, we believe agency ratings could reflect outcome knowledge indirectly, since participants were instructed to attend to the outcomes, and we observed an increase in agency ratings across the trials (for Group 1 studies, in which the interaction was found).
This account could also explain how conflict experiences become associated with specific outcomes, as seen when collecting agency ratings at the end of a block (half of participants in Experiment 1, Chapter 2; Wenke et al., 2010). From this perspective, lower agency ratings for outcomes that follow dysfluent selection would be related to poorer knowledge about action-outcome contingency. Interestingly, once fully predictable action-outcome contingencies have been well-learned, subliminal priming of actions can influence sensory attenuation (Stenner, Bauer, Sidarus, et al., 2014), and sensory predictions in sensory regions, even before an action was made (Stenner, Bauer, Heinze, et al., 2014). This supports the idea that selection fluency can affect the linkage between action and outcome. Also worth noting, studies on metacognition have shown that high confidence errors, e.g. on general knowledge questions, are associated with better error correction, by learning from feedback, than low confidence errors (Metcalfe, Butterfield, Habeck, & Stern, 2012). This suggests that the link between (erroneous) responses and their outcomes (corrective feedback) was weakened for low confidence responses.
More consistent with an optimal cue integration account, one study found that action selection had a stronger influence on SoA when outcomes were unexpected (Sidarus et al., 2013). This study used partial action-outcome contingency (67%), which meant there was some “expected uncertainty” (Yu & Dayan, 2005) regarding the outcomes. That is, the occasional violation of outcome predictions was expected, and not diagnostic to SoA. Therefore, this may have led to a reduction of the contribution of outcomes to SoA when predictions were violated, and a relative increase in the weighting of action selection processes on SoA. Moreover, results also showed that when action selection was fluent, outcome expectation had a reduced effect on SoA. This suggested a reciprocal influence between the two cues, depending on their signal, and not only their general reliability.
Those studies involved distal and abstract action-outcome contingencies. Chapter 3 used a videogame-like task to manipulate action selection, and additionally manipulated the proximal outcomes of action, i.e. the means to achieve a goal. Adding a discrepancy between participants’ mouse movements and the movements of a cursor on the screen, led to a large reduction in SoA, which could overshadow the effects of action selection (Exps. 1 & 2). These findings are also be consistent with optimal cue integration, as accurate motor control would be a highly reliable and salient indicator of agency over the game.
Therefore, the type of outcomes (proximal means vs. distal ends), and whether action-outcome associations are still being learnt, or are already known, may impact on the effects of action selection on SoA. Additionally, knowledge about the outcomes may affect the mechanisms through which action selection influences SoA, that is, whether it changes: the learning process itself; outcome predictions; or the weighting of outcome information.