Internal and external information in error processing

To adapt ongoing behavior in a changing world human beings have to compare performed actions against their intended outcome. This comparison might be based on internally as well as on externally represented information. Using event related brain potentials (ERP, see [1]) time locked to motor responses, performance monitoring processes based on internal information are reflected in a sharp negative deflection with a peak latency of 70–100 ms after erroneous responses and a fronto-central maximum. Several functional interpretations have been advanced for this error-(related) negativity (ERN or Ne [2, 3]) of which the following are pertinent here: The conflict-monitoring approach holds that the amplitude of the ERN varies as a function of the degree of conflict between two or more prepotent response tendencies. In contrast the error-monitoring theory postulates that the ERN is related to the detection of the erroneous response. If the ERN indeed signals the detection of an error, the question arises in which way its amplitude is related to the consciousness of having made an error. Two different, but partially overlapping concepts addressing this topic: error awareness and error certainty. By using antisaccade [4, 5] or GoNogo inhibition tasks [6] it was shown that the ERN could occur without conscious error awareness. Imaging studies supporting this view showing no activation differences between aware and unaware errors in the corresponding error processing anatomical structures (anterior cingulated cortex [7], rostral cingulate zone [8]). In contrast to error awareness there is some evidence that the subjective certainty of having made an error influences the ERN amplitude. Scheffers and Coles [9] reported varying ERN amplitudes according to the participants' rating of being correct or not. Errors for which subjects were uncertain about their incorrect response resulted in a decreased ERN amplitude compared to "sure" incorrect responses. Pailing and Segalowitz [10] reported a similar pattern of results by inducing uncertainty about the subjects' task performance. In their Flanker task a decreased ERN was observed when subjects had to perform a secondary task at the same time. The linear relationship of error certainty and ERN amplitude was shown by results of Luu et al. [11] who defined two criteria for making a correct response, (i) pressing the correct button and (ii) responding faster than an individually adjusted response deadline. While the ERN to incorrect button presses was largest, the amplitude after correct but too slow reactions depended on response speed: the slower the response, the larger the ERN-amplitude and, presumably, the stronger the subject's certainty of a time-out error.

The use of externally provided information on performance is reflected in the feedback related negativity (FRN, or feedback ERN). Miltner and colleagues [12] were the first to describe a negativity peaking around 250 ms after feedback presentation with a maximum over the midfrontal scalp (see also [13]). The functional meaning of the FRN is assumed to be very similar to the ERN: non-satisfying outcomes of relevant events are leading to a negative deflection compared to positive outcomes. Beyond this there is some disagreement concerning FRN amplitude variations: While some authors reported FRN amplitude variations according to different amounts of reward or punishment [14–18] others showed that the FRN's amplitude reflects reward and punishment in a more binary fashion [19–24]: non-satisfying, punishing or non-rewarding events resulted in a pronounced negative amplitude irrespective of the indicated amount of punishment or non-reward. In line with this view, neutral feedback has to be considered as a non-satisfying event leading to an FRN, which can only be differentiated from positive, but not from negative feedback ([21], but see [25] for a different interpretation).

Most studies investigating the functional meaning of the FRN did this without taking its relationship to the ERN into account (for a review see [26]). By contrast, the groups around Holroyd, Coles and Nieuwenhuis conducted a series of studies using probabilistic learning tasks to determine the relationship between ERN and FRN [13, 27, 28]. This paradigm required to learn the correct stimulus response mapping via trial-to-trial feedback. Holroyd and Coles [27] and Nieuwenhuis et al. [13, 28] used different levels of feedback validity to manipulate the predictive value of a given response. According to these authors the predictive value refers to the subjects' ability to infer the outcome of an action – in the studies cited above the resulting feedback – on the basis of the given response. If the predictive value of a response is high, internally available information enables the organisms' action monitoring processes to indicate an error. These processes fail to detect an error in situations with low predictive values. Therefore externally provided information has to be used for adapting ongoing behavior. As predicted, Holroyd and Coles [27] and Nieuwenhuis et al. [13] reported increased ERN, but nearly absent FRN amplitudes in conditions with high predictive values. In conditions with non-informative feedback the predictive value was low, therefore decreased ERN, but increased FRN amplitudes were shown. This inverse relationship in the appearance of ERN and FRN indicated that a non-rewarding event per se is not sufficient to release an FRN, the feedback has also to contain relevant information not available before. More generally within the stimulus – response-feedback sequence only the first error indicating event results in a negative deflection within the ERPs.

In the present study the predictive value of a given response was varied by the apparentness of having made an error. Similar to the paradigm used by Luu and colleagues [11] we applied an Eriksen-Flanker task [29] with an additional response deadline procedure. Participants had to fulfill two criteria for making a correct response: (1) pressing the correct button and (2) performing this response with adequate speed. While the first criterion, correctness of the current button press, had a high predictive value, the predictive value of the second criterion, performing the button press with adequate speed, varied according to the subjects' ability to detect a passing of the valid response deadline. Since responses with correct button presses, but very slow reaction times could be detected by the subjects they had a high predictive value. In contrast the predictive value of reactions with response times passing the valid deadline slightly was low. While Luu and colleagues [11] defined the deadline based on the subjects' performance during the training session only, we adjusted the individual response deadline after each block. Thus, our subjects were informed more precisely about their actual performance by receiving three kinds of symbolic feedback: (1) correct button press and response speed faster than the valid deadline, (2) correct button press, but response speed slower than the valid deadline and (3) incorrect button press with no information regarding the response speed.

Based on previous findings [13, 27] we predicted for responses with correct button selections but an obvious violation of the response deadline increased response-locked negativities, but no variations for the feedback-locked component. When subjects made a correct button press missing the response deadline narrowly the given response had no predictive value. In this case we expected an inverted ERN-FRN relationship: no variation for the response-locked component, but an increase in the feedback-related negativity.

The data presented here indicate a functional interrelation of action monitoring processes which are assumed to underlie error and feedback related negativities: In cases in which self-generated information was sufficient to recognize a time-out error for sure an ERN was released, but no FRN was observed. This was true for the LL condition. By contrast, in the EL condition no ERN was seen but solely an FRN was observed. This suggests that action monitoring processes had to rely on externally provided information in this case. This pattern of results indicates that whenever feedback information is redundant, as in the case of the late late responses, it does not trigger a FRN.

This is in contrast to most studies investigating the nature of the FRN: typically, subjects performed different types of gambling tasks where the response indicates a choice between alternatives only. In this kind of paradigms the feedbacks' outcome, which provided crucial information about winning or losing, is not under control of the subjects' behavior [e.g. [15, 18, 20, 21, 24, 33]]. Some studies also reported the appearance of a FRN without any related action [34]. The general reasoning of these studies was that the FRN reflects whether the outcome of an action was worse than expected. Importantly, the feedback stimulus was the only source of information. It is obvious that in these studies the robust release of an FRN only takes place, because subjects were unaware or unsure regarding the adequacy of their response. In one other study [25], however, an associative learning task was employed that demanded learning for a set of pictures whether a given picture required a left or a right button press. As over the course of an experimental run participants established stimulus-response mappings, the ERN to performance errors increased in amplitude, while at the same time the FRN to negative feedback decreased in amplitude. Importantly, this study already hinted at the fact, that FRN amplitude variations are inversely related to the predictability of an action's outcome. In the present study the expectation of an action's outcome was manipulated by the ability to detect an error. Whenever the error detection, indicated by the appearance of an ERN, was enabled based on self-generated information and according to this detected before the feedback presentation, the expectation corresponded with the real action's outcome, in the present study the feedback. No additional performance monitoring process was necessary and therefore, no FRN was released. If subjects were unsure about their performance or failed to detect an error, this condition was very similar to the experimental situation in the studies mentioned before. In such a case feedback stimuli indicated reliably an action's outcome worse than expected and a FRN was released. It is noteworthy that the only difference between the feedback in the EL and LL condition was the subjects' ability to detect a violation of the response deadline criterion before. Physically the feedback stimuli were identical.

A necessary prerequisite for showing this relationship was to establish a condition where a relevant criterion was violated without being noticed by the subject. Based on the known limitations of human beings to estimate their own reaction time precisely [12, 35] this was achieved in the EL condition in which subjects made the correct button press but exceeded the response deadline criterion slightly. The absence of an ERN can be seen as a clear indication that action monitoring processes failed to detect this kind of errors. In contrast, if subjects detected a response speed error, like in the LL condition, a negative deflection is seen with the typical latency and topography of an ERN despite the correct button selection. The most pronounced ERN was related to incorrect, but in time button press responses. This pattern of results is in line with data reported by Luu and colleagues [11] and replicates their finding that action monitoring processes are able to consider more than one dimension concurrently. Luu et al. argued that ERN amplitude varies according to the degree of self-monitoring processes: the later the response, the stronger the demands for an attentional self-monitoring system and the larger the ERN amplitude. This argumentation is in line with a "certainty" account: the increase of the ERN is related to the subjective certainty of having made an error and is able to explain the pronounced ERN amplitude difference between the LL RT-bin and the incorrect button press shown in the present and the Luu et al. ([11] LATE condition) study: the incorrect button selection is obviously the most distinct type of error, while exceeding the response deadline is less obvious and therefore harder to detect by the subject [9, 36].

Several studies [4] have shown that the ERN and concomitant error-related activations of the medial frontal cortex in fMRI can occur without subjective awareness of the error. With regard to the current set of data we therefore cannot assume that subjects are consciously aware of errors in the LL condition. This would need to be tested by a modified experiment. What we have shown is that "late late" time-out errors are internally detected by the action monitoring system resulting in a reduced impact of external feedback information in this kind of error.

We have shown that the FRN is released only in cases where the feedback stimulus contains non-redundant information. This implies the release of a FRN depends also on results of those action monitoring processes, which are based on self-generated information and are executed before the appearance of feedback information.