Highlights
• Humans frequently report that they detected errors already before executing the error response.
• Early error sensations occur consistently across tasks and metacognitive measures.
• Early error sensations are not caused by an expectation bias.
Abstract: Errors in choice tasks are not only detected fast and reliably, participants often report that they knew that an error occurred already before a response was produced. These early error sensations stand in contrast with evidence suggesting that the earliest neural correlates of error awareness emerge around 300 ms after erroneous responses. The present study aimed to investigate whether anecdotal evidence for early error sensations can be corroborated in a controlled study in which participants provide metacognitive judgments on the subjective timing of error awareness. In Experiment 1, participants had to report whether they became aware of their errors before or after the response. In Experiment 2, we measured confidence in these metacognitive judgments. Our data show that participants report early error sensations with high confidence in the majority of error trials across paradigms and experiments. These results provide first evidence for early error sensations, informing theories of error awareness.
Keywords: Error awarenessError detectionMetacognition
4. General discussion
Participants in experiments on error detection frequently report that they already knew that an error has occurred before the response was executed, a phenomenon we term early error sensation. The goal of the present study was to investigate whether these anecdotally reported early error sensations exist and whether they can be reliably reported. In four experiments using two experimental approaches, we provided evidence that early error sensations indeed exist, and that they occur on the majority of error trials. When participants were asked to classify responses in a flanker task either as being correct, as early detected errors, or as late detected errors in Experiment 1a, they reported early errors in 73.7% of errors. When an additional category for detected errors with unclear timing was introduced in Experiment 1b, early errors were reported in 59.1% of trials. When participants had to wager on the feeling of early error detection, they placed high bets on 62.4% (Exp. 2a) and 70.9% (Exp. 2b). These data demonstrate that early error sensations are reported very consistently across different primary tasks (flanker task vs. number/let discrimination) and secondary tasks (error classification vs. post-decision wagering).
Crucial, however, is the question whether these introspective reports indeed reflect that errors were detected before the response, or whether participants were unable to discriminate between early and late errors and simply guessed that early errors must occasionally occur. A challenging problem for measuring early error sensations is that we cannot objectively determine whether a given error was detected early or late. To deal with this problem, we introduced a reference for the metacognitive reports of early error sensations. In Experiment 2, we used a Visual Awareness task in which participants had to wager on the accuracy of their responses. In the subsequent Error Awareness task, we instructed participants to place high bets on early error sensations only if they were similarly confident as for the high bets in the Visual Awareness task. We argued that this induces a common metric for judging confidence of the two tasks, which allowed us to interpret the metacognitive reports of early error detection with respect to the metacognitive judgments of visual awareness. This reasoning receives support from previous findings showing that humans represent confidence in a task-unspecific format which allows them to compare confidence across tasks with a similarly high precision as confidence within tasks (de Gardelle & Mamassian, 2014). Moreover, it has recently been suggested that integrating information from different sources into a common metric might even be the major purpose of metacognition (Shea & Frith, 2019). In Experiment 2a, the frequencies of high bets were coincidentally similar in both tasks. We can thus infer that the average confidence by which participants reported early error sensations in this experiment corresponded to the average confidence by which they were aware of the visual stimuli in the Visual Awareness task. This confidence level ought to be rather high given that the objective performance in the Visual Awareness task was far above chance level.
We found no evidence that metacognitive reports of early error sensations were subject to an expectation bias. If participants simply guessed that early error sensations must occasionally occur, these guesses should be influenced by expectations about the frequency of early error sensations. To investigate whether such an expectation bias exists, we manipulated the difficulty of the Visual Awareness task, and thus the frequency of high bets in this task. However, whereas the frequency of high bets in the Visual Awareness task varied between Experiments 2a and 2b, the frequency of high bets in the Error Awareness task remained constant across the two experiments. This suggests that metacognitive judgments about early error sensations are not influenced by a specific expectation bias induced by the frequency of high bets in the Visual Awareness task. While we cannot fully exclude a general bias towards instruction-driven expectations about early error sensations, our results strongly suggest that metacognitive judgments on early error sensations are very consistent and reliable across experimental procedures.
We found no evidence that early and late detected errors differ with respect to any objective features. It has been reported that uncertainty or conflict during response selection can influence post-response decision process and metacognitive judgments about errors (Steinhauser et al., 2008, Yeung and Summerfield, 2012). As a consequence, variables like stimulus congruency or RT could potentially influence subjective judgments about early error sensations. However, we found no robust evidence that this was the case in the present study. Participants reported early error sensations in a similar proportion for congruent and incongruent errors in Experiment 1. Moreover, RTs were similar across all error types. A small RT difference between early and late detected errors in Experiment 1a disappeared when we controlled for errors with unclear timing in Experiment 1b. This suggests that the emergence of early error sensations is not related to specific features of task processing like stimulus congruency or RTs. Thus, our data provide little evidence that early error sensations reflect the objective latency of error detection, which has been found to correlate with RT when response speed was directly manipulated (Steinhauser et al., 2008).
An important question is why early error sensations occurred on the majority of trials whereas the neural correlates of error awareness emerge not until 300 ms after an error (e.g., Steinhauser & Yeung, 2010). There are at least two possible explanations. A first explanation is that conclusions about the timing of error awareness from EEG measures like the Pe are incorrect. The Pe is often considered the earliest neural correlate of error awareness and the role of the Pe for the emergence of error awareness has been described within an evidence accumulation account (Steinhauser and Yeung, 2010, Ullsperger et al., 2010). It is assumed that the Pe reflects the accumulated evidence that an error has occurred, and that error awareness emerges when this evidence exceeds a threshold. The evidence is provided by cognitive, autonomous, motor and sensory processing (Bode and Stahl, 2014, Wessel et al., 2012, Wessel et al., 2011), but does not necessarily rely on early error processing represented by the Ne/ERN (Di Gregorio et al., 2018). One possibility is that the feeling of error awareness emerges already before the Pe, for instance, at the time point of the Ne/ERN or even earlier (Bode & Stahl, 2014). The Pe could represent a later stage of metacognitive processing, perhaps related to the emergence of confidence about response accuracy (Boldt & Yeung, 2015).
A second explanation is that early error sensations are a metacognitive illusion. Error awareness could emerge at the time of the Pe but the illusion is created that the error has been detected already before the response. This mechanism could serve to subjectively synchronize error awareness with the timing of the objective error in the same way as visual awareness is subjectively aligned with the onset of a visual stimulus. In the context of visual awareness, expectations and other top-down variables can influence the accumulation of sensory evidence and consequentially metacognitive judgments about stimulus awareness (de Lange et al., 2010, Kouider et al., 2010). Moreover, a backward referral process has been assumed to synchronize the subjective time point of visual awareness with the objective stimulus to create a coherent perception in the stream of consciousness (Libet et al., 1979, Libet et al., 1983). A similar process could align the subjective time point of error awareness with the emergence of the objective error. This temporal alignment of actions (i.e., a response) and their effects (i.e., the feeling of being incorrect) could further serve to evoke a sense of agency, i.e., the feeling of having caused an effect. Indeed, previous studies have shown that action-effect contingencies are influenced by their temporal contiguity and vice versa. Humans tend to perceive two events more causally related the closer they occur in time (Greville & Buehner, 2010), and causality judgments correlate with the perceived temporal contiguity between actions and their sensory effects (Haering & Kiesel, 2016). In other words, these metacognitive illusions on early error sensations could serve to reconstruct temporal contiguity between perception, action and metacognitive contents (Kouider et al., 2010).
While we obtained clear and robust results across several experiments, the present method has also some limitations. A first limitation is that using a categorical measure for the timing of error detection implies a loss of information as time is a continuous phenomenon. However, differentiating only between errors detected before and after the response has the advantage of imposing considerably lower cognitive load than using a continuous measure. For instance, in the classical Libet studies (Libet et al., 1983), participants had to indicate the time of voluntary action initiation on a visual clock. However, in addition to considerable methodological weaknesses (Trevena & Miller, 2002), monitoring a clock represents a difficult secondary task that presumably interferes with both, the primary task and the task to detect errors. In contrast, our categorical measure uses the response as a reference rather than a continuous timer. As error detection already involves response monitoring (Steinhauser et al., 2008), only minimal additional load should be imposed.
As already discussed, a second limitation is that we have no objective measure that verifies the existence of early error sensations. Future studies could solve this problem by measuring neural correlates of early error sensations. Strong evidence for the existence of early error sensations would be provided if not only the Pe but also the earlier Ne/ERN would correlate with early error sensations. If only the Pe differed between early and late detected errors, this would suggest that early error sensations emerge during the later stage of conscious error processing. However, if such a difference was found also for the Ne/ERN, this would point to early error signals such as response conflict (Yeung et al., 2004) or prediction errors (Holroyd & Coles, 2002) as the origin of early error sensations. It is even possible that brain activity preceding the response can affect metacognitive judgments on early error sensations. ERP differences between errors and correct responses have been found prior to the response (Bode & Stahl, 2014) or even on the previous trial of simple tasks (Hajcak et al., 2005, Hoonakker et al., 2016, Ridderinkhof et al., 2003), as well as in tasks involving complex sequences of motor programs such as piano playing (Maidhof, Rieger, Prinz, & Kloesh, 2009) . In a similar vein, a study using self-report measures has revealed that internal error prediction occurs before responses in skilled typing (Rieger & Bart, 2016). Here, the question arises whether this activity serves as a cue for metacognitive judgments, or whether metacognition relies on direct access to the timing of these neural events.
A further question is whether early error sensations are related to early incorrect response activation. On correct trials, early incorrect response activation leads to a phenomenon called partial errors (Burle et al., 2002, Coles et al., 1995, Endrass et al., 2008), which can be consciously reported by participants (Rochet, Spieser, Casini, Hasbroucq, & Burle, 2014). Future studies could investigate whether such early incorrect response activation on error trials is responsible for early error sensations. Indeed, lower response force for errors than correct responses has been shown in skilled typing (Rabbitt, 1978). As this phenomenon has been interpreted as resulting from inhibition of the error response before actual response execution, it could be taken as indirect evidence for early error sensations. Future studies could examine whether errors accompanied by early error sensations are executed with lower response force than late errors.
The present study provides first evidence that participants have the subjective feeling of detecting errors already before they occurred. We show that these early error sensations can be robustly measured across different tasks and metacognitive judgments. Our results add to the broad body of evidence that humans have metacognitive access to a multitude of performance parameters. Previous studies could show that participants are able to report whether an error has occurred or not (Rabbitt, 1968, Rabbitt, 2002), to provide graded confidence judgments on the accuracy of their response (Boldt & Yeung, 2015), to classify the type of error they committed (i.e., to which distractor stimulus they responded; Di Gregorio et al., 2016), and to estimate their RTs in choice tasks (Bryce & Bratzke, 2014). These metacognitive contents are used for optimizing decision processes (Desender et al., 2018, Desender et al., 2014). Metacognitive representations on the timing of error detection could form another piece of information to support this optimization.
Posts
No comments:
Post a Comment