Monitoring task loading with multivariate EEG measures during complex forms of human-computer interaction.
Electroencephalographic (EEG) recordings were made while 16 participants performed versions of a personal-computer-based flight simulation task of low, moderate, or high difficulty. As task difficulty increased, frontal midline theta EEG activity increased and alpha band activity decreased. A participant-specific function that combined multiple EEG features to create a single load index was derived from a sample of each participant's data and then applied to new test data from that participant. Index values were computed for every 4 s of task data. Across participants, mean task load index values increased systematically with increasing task difficulty and differed significantly between the different task versions. Actual or potential applications of this research include the use of multivariate EEG-based methods to monitor task loading during naturalistic computer-based work.
This paper documents the development and evaluation of a neurophysiology-based method for deriving a continuous index of task loading from individuals engaged in complex computer-based work. Task loading is used here to refer to the degree to which neural resources are activated by effortful task performance. Because a human operator's functional capacity is limited and can often be exceeded in work environments that demand sustained vigilance to multiple streams of information, the likelihood of performance errors can be high (Card, Moran, & Newell, 1983). The ability to monitor task loading continuously might thus be valuable in task analysis research, in efforts to improve the usability of human-computer interfaces (Raskin, 2000), and in efforts to design appropriate adaptive automation strategies (Byme & Parasuraman, 1996; Morrison & Gluckman, 1994; Parasuraman, Sheridan, & Wickens, 2000).
Although both subjective estimates and assessment of overt behavioral performance can permit detailed inferences about task loading, other modalities might also provide convergent insights into the degree to which a task environment is demanding of neural resources. To be useful in applied contexts, a monitoring method must be robust enough to be reliably measured under relatively unstructured task conditions and sensitive enough to consistently vary with some dimension of interest. Furthermore, it should not interfere with operator performance, it should be applicable across many contexts, and it should have reasonably good time resolution. Some physiological measurements - particularly measures of central nervous system activity, such as the electroencephalogram (EEG) - appear to meet these requirements.
EEG and event-related potential measures have been shown to be fairly sensitive to variations in task difficulty (Gevins, Smith, McEvoy, & Yu, 1997; Humphrey & Kramer, 1994; McCallum, Cooper, & Pocock, 1988; Parasuraman, 1990). In a recent study, Gevins et al. (1998) demonstrated that EEG-based, participant-specific, multivariate pattern recognition methods could be used to discriminate levels of task loading under highly controlled task conditions. In that study, continuous EEG data were recorded from 8 healthy adults performing a simple task in which working memory demands were varied in order to modulate task difficulty. As task difficulty increased, performance accuracy decreased, response times increased, and a change in the power of the EEG spectra in the theta (5-7 Hz) and alpha (8-12 Hz) frequency bands occurred. In particular, during the highest task load there was an increase in theta power over frontal regions of the scalp and a decrease in alpha power over widespread scalp regions, relative to me asures taken during the lowest task load. Two-class, neural-network-based pattern recognition functions applied to EEG spectral features reliably distinguished low load from high load and moderate load from low or high load.
Such results suggest that with further development, it might become possible to use EEG-based methods for unobtrusively monitoring task loading in individuals engaged in computer-based work. However, the task used in the prior study made minimal demands on perceptual and motor systems and required that a participant's effort be focused on only a single activity. The ability to reliably measure task loading in individual participants under such constrained circumstances might in itself be useful. For example, it has been applied to the problem of assessing the effect of environmental stressors on cognitive functions (Gevins & Smith, 1999). Even so, in more naturalistic work environments, task demands are usually less structured, and mental resources often must be divided between competing activities. Thus it remains to be determined whether such methods can successfully generalize to the problem of monitoring task load under a wider range of activities.
Recent studies have demonstrated that more complicated forms of human-computer interaction, such as playing video games, produce a mental effort-related modulation of the EEG that is similar to that observed during controlled laboratory tasks (Pellouchoud, Smith, McEvoy, & Gevins, 1999; Smith, McEvoy, & Gevins, 1999). This implies that it might be possible to extend EEG-based multivariate methods for monitoring task load to such circumstances. The present study was thus performed in order to further evaluate the potential of the EEG as a modality for monitoring the degree to which an individual's mental resources are engaged by this form of computer-based work.
EEG measurements were made as participants performed the Multi-Attribute Task Battery (MATB; Comstock & Arnegard, 1992). The MATB is a computer-based multitasking environment that simulates some of the activities a pilot might be required to perform. It has been used in several prior studies of mental workload and adaptive automation (e.g., Fournier, Wilson, & Swain, 1999; Parasuraman, Molloy, & Singh, 1993; Parasuraman, Mouloua, & Molloy, 1996). The data collected during performance of the MATB were used to test whether it is possible to derive combinations of EEG features that can be used for indexing task loading during a relatively complex form of human-computer interaction.
Sixteen healthy young adults (21-32 years, mean age 26 years; 8 women, 8 men) participated in the study. All were right handed and had normal or corrected-to-normal vision and hearing. Participants received a cash honorarium in exchange for their participation. All participation was fully informed and voluntary.
Task Structure and Procedure
Study participants performed versions of the MATB that varied in difficulty. The MATB included four concurrently performed subtasks in separate windows on a computer screen (for graphic depictions of the MATB visual display, see Fournier et al., 1999, and Molloy & Parasuraman, 1996). Performance data were collected for each task separately from keyboard and joystick controls. The four subtasks were systems monitoring, resource management, communications, and tracking. The systems monitoring task required the operator to monitor and respond to simulated warning lights and gauges. In the resource management task, fuel levels in two tanks had to be maintained at a certain level. Fuel level could be controlled by pressing keys to turn on and off a series of pumps.
The communications task simulated receiving audio messages from air traffic control and required pressing keys to make frequency adjustments on navigation and communication radios. In the communications task, distractor stimuli were occasionally presented in which frequency instructions for a different aircraft (identified by a call sign different from that of the operator's own aircraft) were given. The two-dimensional compensatory tracking task simulated manual control of aircraft position using a joystick with first-order control characteristics. Participants were seated comfortably about 60 cm from the computer screen as they performed the task battery.
Prior to the test day, participants completed a training session to ensure that they were well practiced on the tasks before the EEG recording. On the practice day, participants were first given oral instructions for the four subtasks, emphasizing that effort should be divided equally among them and that response times and accuracy were equally important. Participants initially practiced the tracking component alone for 3 min. In subsequent 3-min runs, the resource management, then monitoring, and finally the communications task were added. Participants then received extensive practice for low-, medium-, and high-load versions of the task (LL, ML, and HL, respectively), first in 3-min runs and then in 5-min runs.
Manipulating the difficulty of each subtask served to vary load. Tracking was set to automatic in the LL condition, but in the ML and HL conditions a varying degree of drift in the joystick control made the tracking slightly or moderately difficult. The forcing function for tracking consisted of a sum of nonharmonic sine waves. The frequency of the forcing function was low in the ML task condition and medium in the HL task condition, according to the MATB program specifications. In the resource management task, a number of fuel pumps "failed" and subsequently reset themselves during the course of the task. Failure of the pumps inhibited the flow of fuel between tanks, rendering it more difficult to maintain a target fuel level (2500 units) in the primary tanks.
The number of pump failures increased with task difficulty (5, 10, or 20 events of each type in the LL, ML, and HL tasks, respectively). Furthermore, the resource management task was automated in the LL condition so that the events did not require a response. In the monitoring subtask, the number of critical gauge/warning light stimuli increased with load level (5, 10, or 20 events of each type in the LL, ML, and HL tasks, respectively). In the communications subtask, the number of own-ship and other-ship instructions increased with load level (3, 5, or 10 events of each type in the LL, ML, and HL tasks, respectively).
On the subsequent test day (no more than 1 week later), each participant returned for the EEG recording session. In this session participants completed a total of five blocks of the MATB task. Each block contained one 5-min run at each load level(LL, ML, and HL) and one 5-min passive-watching control (PW) condition. In this condition participants were asked simply to watch and listen to the stimuli in an LL version of the MATB task while their hand rested on the joystick but to not make any behavioral responses. The first two of the five blocks served to further habituate participants to the testing protocol, and data from those two blocks were not included in subsequent analyses. The final three blocks consisted of one baseline block, used to obtain data for deriving multivariate EEG task load indices, and two subsequent test blocks, used to validate those indices independently. Participants were given a short break between blocks. Order of task version within blocks was counterbalanced across blocks and pa rticipants.
At the end of each 5-min task run, self-report assessments of task loading were obtained with the NASA Task Load Index (TLX) rating scale (Hart & Staveland, 1988). This subjective assessment required participants to manipulate a screen cursor to rate their task experience, from low to high, on visual-analogue continua (corresponding to an integer range of 0-100) for each of six dimensions (mental demand, physical demand, performance, temporal demand, and frustration). The mean of these six scales was employed in the analyses reported herein, with equal weighting given to each subscale (Nygren, 1991).
The EEG was recorded using an elastic cap with electrodes placed at 28 scalp locations ([F.sub.p1], [F.sub.p2], Af3, [A.sub.f4], [A.sub.fz], [F.sub.19], [F.sub.7], [F.sub.3], [F.sub.z], [F.sub.4] [F.sub.8], [F.sub.t10], [T.sub.7], [C.sub.3], [C.sub.z], [C.sub.4], [T.sub.8], [P.sub.9], [P.sub.7], [P.sub.3], [P.sub.z], [P.sub.4], [P.sub.8], [P.sub.10], [O.sub.1], [O.sub.z], [O.sub.2], [I.sub.z], with an electronically linked mastoids reference (electrode positions described in Sharbrough et al., 1990). Electro-Oculogram activity was recorded from electrodes located above each eye, referenced to an electrode at the outer canthus of each eye. Physiological signals were band-pass filtered at 0.01 to 100 Hz and sampled at 256 Hz. Impedances were less than 5 Kohms. Data were recorded continuously during each 5-min run of the MATB task. EEG processing and preliminary analyses were performed with the MANSCAN software package (SAM Technology, San Francisco, CA). Adaptive filtering methods (see Du, Leong, & Gevins, 199 4) were employed to minimize eye movement contamination of the EEG. The data were then visually inspected, and data segments containing residual eye movements and blinks, instrumental noise, and movement artifacts were eliminated from subsequent analyses.
Group Analyses of Behavior and EEG
In a first set of analyses, conventional statistical procedures (using the SPSS software package) were employed to identify the ways in which behavioral and physiological variables were systematically affected by the experimental manipulations. Behavioral data consisted of mean response times (RTs) and error rates (ERs) for the monitoring and communications subtasks in the LL, ML, and HL conditions. Mean tracking deviation scores and mean deviation from target fuel level in the ML and HL conditions were also examined. (In the LL condition, participants did not perform these two subtasks.)
For EEG data, log power spectra were derived from 2-s windows and averaged over all data segments for each participant for PW, LL, ML, and HL levels. Measurements of average power were made of the frontal midline theta signal (5-7 Hz) at [F.sub.z], and of the alpha signal at [P.sub.z]. Alpha power was measured in both the lower or "slow" (8-10 Hz) and the upper or "fast" (10-12 Hz) alpha bands (see Gevins et al., 1997; Klimesch, Schmike, & Pfurtscheller, 1993).
Repeated-measures analyses of variance (ANOVAs) were performed to compare behavioral and EEG measures across task load levels. Where appropriate, degrees of freedom for each analysis were adjusted using the Greenhouse-Geisser procedure to correct for violations of the sphericity assumption in repeated-measures designs. Significance levels reported in the following paragraphs reflect this adjustment, but the unadjusted degrees of freedom are presented for simplicity.
Development and Application of Participant-Specific Index Functions
In the second set of analyses, participant-specific multivariate functions were derived and used to compute a task load index from short segments of EEG data. The participant-specific functions were developed on samples of EEG data taken from the baseline block of data for each participant from the HL and PW versions of the MATB task, and then they were tested on data samples from the last two test blocks.
To create the participant-specific functions, the EEG data were first decomposed into 4-s windows of data. In order to form a candidate feature set for function derivation, a set of spectral power estimates of activity in subbands of the theta and alpha frequency ranges was extracted from each data window (Gevins et al., 1998). The set of features in the theta band included estimates derived from anterior frontal and frontal midline electrode positions ([A.sub.fz], and [F.sub.z]) in the 5- to 6-Hz, 6- to 7-Hz, and 7- to 8-Hz sub-bands. The set of features in the alpha band included estimates derived from left, middle, and right frontal ([F.sub.3], [F.sub.z], [F.sub.4]) and left, middle, and right parietal ([P.sub.3], [P.sub.z], [P.sub.4]) electrode positions in the 8- to 10-Hz and 10- to 12-Hz sub-bands.
The decision to include this particular set of features in the list of candidate features was based on two a priori factors. First, we decided to eliminate from consideration data from electrodes placed directly over primary visual, auditory, or somatomotor cortices so that the resulting indices would not be disproportionately affected by task-related activation of these regions. Second, topographical studies have identified the frontal midline region as an optimal location for detecting load-sensitive EEG signals in the theta band (Gevins et al., 1997; Inouye et al., 1994; Ishii et al., 1999) and have shown that alpha band signals over the frontal and parietal regions tend to be relatively more sensitive to the attentional demands of tasks than do alpha band signals recorded over other regions (Gevins et al., 1997; Klimesch et al., 1993).
The candidate set of 18 potential features was reduced to a subset of four optimal features uniquely selected for each participant. This involved eliminating all but the top candidate features from the total feature set after ranking them with respect to the degree of statistical distance or dissimilarity they displayed between the set of samples obtained in the HL task version and the set obtained in the PW task version during the calibration test block. Using this subset of four features, a unique multivariate distance function was then defined for each participant that maximized the divergence (Tou & Gonzalez, 1974) between the two classes of calibration data samples. The output of this function was then smoothed with an exponentially weighted window based on the preceding four data samples. Finally, it was normalized and range restricted for participants relative to their index values in the calibration data so that it would vary continuously over a range of two standard deviations, between 1 (highest ta sk load) and 0 (lowest task load).
The resulting multivariate EEG-based task load indices were then validated on new data by applying them to the other blocks of data collected for each participant and deriving EEG task load index estimates for each 4-s data sample. A final set of analyses involved a more detailed look at the individual participant data. In particular, to estimate how well the EEG index functions distinguished between load levels, we conducted statistical tests in single participants by comparing the distributions of the EEG task load index values computed for each participant between pairs of task versions.
Subjective ratings of task loading as measured by the NASA TLX scale are summarized in Table 1. Participants reported progressively higher task demands across the LL, ML, and HL versions of the MATB. This increase produced a highly significant effect of task version, F(2, 30) = 33.0, p < .001. Post hoc pairwise comparisons of subjective load ratings indicated that the means for each task level significantly differed from the others. These results thus indicate that from the participant's perspective, the stimulus and response manipulations were effective at varying the composite difficulty of the MATB task.
Measures of overt task performance in the MATB task are also summarized in Table 1. As described, the monitoring and communications subtasks required active responding in all three task versions, whereas the tracking and resource management subtasks were automated (did not require overt responding) in the LL task version. Error rates in the monitoring and communication subtasks were very low in all task versions, most likely because of the extensive training participants had received on the task. As a result, error rates were not sensitive to the task manipulations.
In contrast, reaction times in these tasks did vary among task conditions. For example, in the monitoring subtask, RTs in the LL task were on average only about 60% of the duration of the much slower RTs measured in the HL task, F(2, 30) = 22.3, p < .001. Post hoc pairwise comparisons indicated that whereas the LL RTs were faster than those observed in either the ML or the HL task, the RT difference between the ML and LL task versions was not significant. RTs in the communications task also tended to increase in a monotonic fashion among the LL, ML, and HL tasks. When comparing all three task versions and correcting for violations of sphericity, this trend did not quite reach significance, F(2, 30) = 3.4, p = .07. However, when comparing only the extremes of the task versions (LL vs. HL), a significant effect of task version was obtained, F(1, 15) = 12.7, p < .005.
Although average scores were higher in the HL than in the ML condition of the resource management subtask (indicating greater error or deviation from target resource levels), this difference did not reach significance. In contrast, for the tracking subtask a significant effect of task version was obtained, F(1, 15) = 34.6, p < .001, with greater root mean square error in the HL task than in the ML task. In sum, measures of overt behavioral performance also tended to be consistent with the notion that the differences in demands among the task versions served to increase task loading.
A final set of analyses was performed to determine whether the pretraining on the tasks was adequate for the participants to have reached an asymptotic level of performance. In particular, subjective ratings and overt performance data were compared for the first and third blocks of the HL task condition from the final three test blocks obtained (i.e., those that are the focus of the analyses here). No significant between-block differences were observed, indicating that performance had stabilized by the time the critical recordings were made.
Visual inspection of average power spectra, collapsed across participants and test blocks, revealed task-related EEG modulation in the theta and alpha frequency bands (Figure 1). These task-related changes were similar to those observed in prior studies. That is, EEG activity at frontal electrodes was dominated by a sharp peak in the spectra in the theta (~6-7 Hz) range that was largest in amplitude in the HL task condition and smallest in the LL and PW conditions. In contrast, EEG activity in the alpha (8-12 Hz) band was most prominent at posterior sites, with largest amplitude in the PW condition and smallest amplitude in the HL condition.
These observations are further illustrated by Figure 2, which plots the values of individual features in the theta and alpha bands after normalizing the data within each participant subject across the different task versions. A repeated-measures ANOVA of the effect of varying load level (PW, LL, ML, and HL) on normalized theta power at [F.sub.z] revealed a significant effect of task difficulty, F(3, 45) = 5.5, p < .01, with a progressive increase in theta from the PW to the HL task version. Both 8- to 10 Hz, F(3, 45) = 9.7, p < .001, and 10- to 12 Hz, F(3, 45) = 25.0, p < .001, alpha band activity measured at [P.sub.z] also differed significantly in power across task conditions, with a progressive decrease in alpha band activity from the PW to the HL task condition.
EEG Task Load Index Results
The purpose of this analysis was to determine whether the index values produced by the multivariate EEG task load functions that were derived for each participant would succeed in distinguishing between MATB conditions that varied in difficulty. As noted in the Methods section, multivariate functions were first derived for each participant based on a weighted set of EEG features that best discriminated data from calibration runs of the HL and PW task versions in that participant. The resulting indices were then applied to new samples of data obtained in subsequent blocks of MATB performance. The results described here reflect the outcome of applying the indices to the new blocks of data.
Table 2 summarizes the features utilized in the participant-specific functions. Each of the 18 potential features (see "Methods") included in the candidate feature set was used in the EEG task load index of at least one participant. Across participants, the most frequently selected features for inclusion in the participant-specific functions were alpha band measures at the left parietal electrode, [P.sub.3]. Between the 8-to 10-Hz and 10- to 12-Hz bands, alpha activity over this region was included in the equations of 62.5% (10 of 16) of the participants. The least frequently selected features were the alpha band measures over the left frontal region, which between the two frequency bands were included in the equations of only 25% of the participants.
For comparison among participants, the index values computed for each segment were averaged across data segments for each participant within each task condition. For each individual, Table 3 shows his or her mean EEG task load index in each task condition (HL, ML, LL). For comparison purposes, each participant's overall subjective task load ratings are also included in Table 3, as are his or her mean RTs in the monitoring subtask, because that subtask was included in all difficulty versions of the MATB and because the RTs in that subtask were highly sensitive to the experimental manipulations. The BEG load index for the PW condition is also shown in Table 3, but, by definition, no behavioral responses were collected in that condition.
In order to examine the overall relationship between these measures, the subjective, overt behavioral, and EEG load index values were rank ordered for each participant, and the resulting rankings were correlated using nonparametric methods (Kendall's tau or [tau]; Kendall, 1963). There was a high degree of correspondence between the three types of measures. For example, the subjective ratings and EEG load index were correlated with a [tau] = .79, p < .001. The correlation between the EEG load index and the RT measure was slightly lower but still highly significant, [tau] = .61, p < .001. Thus the EEG-based task load index appeared to be influenced by changes in task difficulty in a fashion similar to the way subjective and behavioral measures were.
Indeed, although there was substantial variability between participants, on average the mean EEG task load indices across participants increased in a consistent, monotonic fashion with increasing task difficulty, F(3, 45) = 28.0, p < .001. Planned comparisons of individual means further indicated that the indices for each load level significantly differed from all other load levels. Inspection of the data from Table 3 indicates that most of the individual participants displayed a steady increase in task load index value from the PW condition to the HL condition.
For 15 of the 16 participants, the task load index was substantially higher in the HL task than in the PW task. One of the participants (#9) did not display a consistent pattern of increase across load level, and for several of the participants the EEG task load index values were very similar or even reversed from the expected direction between adjacent task difficulty levels. Some degree of such between-subjects noise variation in an experiment is to be expected; it is worth noting that similar between-subjects variations can be observed when comparing the subjective or overt behavioral measures among task versions in individual participants.
To further characterize the effectiveness of the procedure at an individual level, Table 4 presents p values from t tests comparing data from each participant at each load level with his or her data at other load levels. This exploratory analysis was intended to convey an initial impression of how well this technique might work for monitoring task load variations in individual operators. In essence, it provides an estimate of the distance among the means of the distributions of the EEG load index scores calculated during performance of each task version. The HL versus LL comparison was significant (at p < .05 or better) for all 16 participants. The HL versus PW comparison and the ML versus LL comparison were both significant for 15 out of 16 participants. Not surprisingly, the least consistently significant differences were observed in the HL versus ML and LL versus PW comparisons. Even so, data from these nearest neighbor extreme conditions were discriminable at the p < .05 level in at least a third of the participants and at the p < .10 level in half the participants.
In this study task loading was manipulated by varying the demands imposed on participants who were operating in a computer-based multitasking simulation environment. Both objective and subjective measures indicated that the task manipulations increased the mental effort required for task performance. As in prior studies, variations in task difficulty were found to affect EEG signals. Participant-specific functions that combined multiple EEG features to create task load indices were derived from a sample of each participant's data and then applied to new test data from that participant. This technique yielded sensitive indices of variations in neural resource utilization that were customized to each participant and that varied systematically with task difficulty. These results are discussed in the remainder of this paper.
Modulation of EEG Spectral Features with Changes in Task Loading
In the current study EEG theta signals at frontal recording sites increased with increasing task load. This finding is consistent with a number of prior studies that have found enhanced theta band activity in tasks requiring sustained mental effort (Gundel & Wilson, 1992; Miyata, Tanaka, & Hono, 1990; Yamamoto & Matsuoka, 1990). Topographic analyses have indicated that this task loading-related theta signal tends to have a sharply defined potential field with a focus in the anterior midline region of the scalp (Gevins et al., 1997; Inouye et al., 1994); such a restricted topography is unlikely to result from distributed generators in dorsolateral cortical regions. Instead, attempts to model the generating source of the frontal theta rhythm from both EEG (Gevins et al., 1997) and magnetoencephalographic (Ishii et al., 1999) data have implicated the anterior cingulate cortex as a likely region of origin. This cortical region is thought to be part of an anterior brain network that is critical to attention contr ol mechanisms and that is activated by the performance of complex cognitive tasks (Posner & Peterson, 1990; Posner & Rothbart, 1992). Indeed, in a review of more than 100 positron emission tomography (PET) activation studies that examined anterior cingulate cortex activity. Paus, Koski, Caramanos, and Westbury (1998) found that the major source of variance that affected activation in this region was associated with changes in task difficulty. The present results are thus consistent with these views, implying that performance of the more difficult versions of the MATB task placed high demands on frontal brain circuits involved with attention control.
Alpha activity was found to systematically decrease in power as the MATB versions increased in difficulty. Inverse relationships between alpha power and task difficulty have been found in numerous prior studies (Gevins et al, 1997, 1998; Gevins, Zeitlin, Doyle, Schaffer, & Callaway, 1979; Gevins, Zeitlin, Yingling, et al., 1979; Glass, 1966; Gundel & Wilson, 1992; Sterman, Mann, Kaiser, & Suyenobu, 1994). Because of this relationship, the magnitude of alpha activity during cognitive tasks has been hypothesized to be inversely proportional to the fraction of cortical neurons recruited into a transient functional network for purposes of task performance (Gevins & Schaffer, 1980; Mulholland, 1995; Pfurtscheller & Klimesch, 1992). This hypothesis is consistent with current understanding of the neural mechanisms underlying generation of the alpha rhythm.
More specifically, the magnitude of an EEG signal is directly related to the proportion of neuronal groups oscillating in synchrony (Elul, 1969; Lopes da Silva, 1991). Microelectrode recordings made directly at the cortex indicate that alpha is coherent over areas that are similar in size to the extent of individual cortical columns (Lopes da Silva & Storm van Leeuwen, 1978). This local coherence is probably driven through intracortical connections between nearby neurons, with individual cells acting as either resonators or oscillators and with relative independence between groups. These coherent local groups are presumed to be the generators of the alpha rhythm recorded at the scalp. When the brain is at relative rest, thalamocortical inputs serve to drive a high proportion of the alpha generators such that they come to oscillate in phase (Bullock & McClune, 1989; Lopes da Silva, Vos, Mooibreck, & Van Rotterdam, 1980). This phase locking produces sufficient summation of synaptic activity across populations of alpha generators to yield signals large enough to be recorded at the scalp. As the demands of a task increase, different regions of cortex might come to be recruited into transient functional networks through longer-range intracortical interactions, with an associated decrease in the overall proportion of local alpha generators that are passively oscillating in synchrony and a reduction in the power of alpha signals recorded at the scalp.
Convergent evidence for this view is provided by observations of a negative correlation between alpha power and regional brain activation as measured with PET (Larson et al., 1998; Sadato et al., 1998). It is also consistent with the frequent finding from neuroimaging studies of greater and more extensive brain activation during task performance when task difficulty increases (Baker et al., 1996; Bunge, Klingberg, Jacobsen, & Gabrieli, 2000; Carpenter, Just, & Reichle, 2000; Garavan, Ross, Li, & Stein, 2000). Such results suggest that as task difficulty in the MATB increased, there was increased commitment of cortical resources to task performance.
Some views of the structure of the mental resources that can be allocated to task performance posit a relative independence of the resources involved with cognitive processes and those involved with perceptual processing and motor expression (Gopher, Brickner, & Navon, 1982; Wickens, 1991). In prior studies in which the perceptual and motor requirements of tasks were kept constant while the cognitive demands of those tasks varied (Gevins et al., 1997, 1998), patterns of EEG modulation were found that were very similar to those observed here. Because of this, it is reasonable to suppose that a significant proportion of the variance observed in EEG signals in the current study was related to between-task differences in cognitive, rather than perceptuomotor, factors. However, in the experiment performed here and, more generally, in naturalistic task environments, increased task difficulty tends to be associated with greater perceptual and motor demands. Thus although the current data provide evidence for the ba sic feasibility of the analytic approach described herein, an important area for future research is the development of physiological indices that differentiate between the loading of separate resource systems.
Derivation and Testing of Participant-Specific Multivariate Load Indices
The results reviewed indicate that on average, EEG signals in the theta and alpha bands systematically index variations in task loading in a computer-based video game task. This is important information in that any potentially successful strategy for developing EEG-based indices of task loading requires building on what is known concerning the commonalities between participants in their reactions to cognitive stressors. Indeed, such knowledge is critical for defining the constellation of features that constitute the candidate feature set for deriving any multivariate indices that might be developed.
Once such a set is defined, EEG-based multivariate indices can be developed in at least two ways. The most common alternative is to define a feature subset and weighting scheme that provides a parsimonious fit to a set of training data that was composed of examples from many participants and then test how well the function generalizes to new data from other participants. The less common alternative (the one adopted here) is to fit personalized functions to examples of data from individual participants and then to apply those functions to new samples of data from the same participants.
Two studies have explicitly compared the relative effectiveness of these alternative approaches. Although the groupwise analysis strategy was found in both cases to produce significant results, it was also found that participant-specific functions were relatively more effective at discriminating degrees of task loading (Gevins et al., 1998) and at detecting intoxication or fatigue (Gevins & Smith, 1999) than were group functions. Although the two approaches were not explicitly compared in the current study, the data suggest that a similar pattern would have been observed here. That is, substantial between-participant diversity was observed in the particular subsets of 4 features (of the 18 in the candidate feature set) selected for inclusion in the participant-specific functions. Thus the participant-specific analytic strategy employed was general enough to apply to all members of the participant population, yet it did not assume an unrealistic degree of homogeneity between participants.
In preliminary studies we also found that this participant-specific approach generalizes to other interactive tasks, including text editing, Web searching, and performing computer-based aptitude tests (Smith & Gevins, 1998).
Such results suggest that an optimal strategy for developing technology for using physiological information to measure task loading will, on one hand, capitalize on the commonalities between participants in order to design efficient recording systems that simplify the collection of appropriate features for the candidate feature set. On the other hand, performance of such systems will be enhanced by employing analysis strategies that are sensitive to between-participant heterogeneity. This is the approach employed in the present study.
When using this approach, we found that across the group, the mean EEG task load indices across participants increased in a monotonic fashion with increasing task difficulty. We also found that all of the gradations of task difficulty, from passively watching the computer display to the highest active task load, could be discriminated from one another in a statistically significant fashion. Furthermore, high correlations were found between the pattern of change in the EEG task load indices among task versions and the corresponding changes observed in subjective estimates and behavioral measures of task difficulty.
For the most part, a similar pattern of results was also observed in analyses limited to individual participants, although in some participants the order of EEG task load index values strayed from monotonicity. Such between-participant variation in an experiment is not surprising, and these same types of variations can be observed when comparing the subjective or overt behavioral measures among task versions in individual participants.
At the extreme, one participant (#9 in Tables 2 and 3) failed to show the typical modal increase in task load index value with increasing task difficulty. A review of this participant's data suggests that the problem arose from the baseline data used to calibrate the index function, which was unrepresentative of the participant's subsequent data samples. In fact, the two subsequent test data samples collected for this same participant displayed a normal pattern of task-related EEG modulation. Thus this failure has important practical significance in that it indicates that any general methods based on this approach would need to incorporate control procedures for gauging the appropriateness of the calibration data used for function derivation.
This outlier case aside, for all of the remaining 15 individuals, the index values for the HL task were significantly larger than those for either the LL or PW tasks. In contrast, the index values for the HL task were significantly larger than those for the ML task for only about half the participants.
A frequent failure to differentiate adjacent load levels was also observed when comparing the LL and PW tasks. This relative insensitivity to differences between adjacent task load levels does not necessarily reflect limitations of the particular analytic approach used here; it is consistent with past studies that have employed other analytic techniques. Most past studies that have used physiological measures to index task load have attempted only to differentiate low- from high-load conditions. The few studies that have employed more gradations of task difficulty (e.g., Fournier et al., 1999; Gevins et al., 1998) have found that the relatively small physiological changes between adjacent difficulty levels were less statistically reliable than the relatively large differences between endpoints. Furthermore, in exploratory analyses performed on the current data using either discriminant analysis or neural network pattern recognition techniques, we found that the same individual cases that were difficult to dis criminate using the procedures employed here were also difficult to discriminate using alternative methods.
On some levels this relative difficulty in discriminating small variations in task loading is not surprising: It takes more statistical power to discriminate two distributions with a high degree of overlap than it does for two distributions with less overlap. On other levels the observation is more informative. For example, visual inspection of the EEG spectra in the individual participants indicated that in general, the individual cases in which two adjacent load levels could not be algorithmically discriminated also tended to be cases in which there were no visually obvious differences in signal strength between task conditions. This suggests that the problem is not just one of statistical power but, rather, that there are important individual differences in the range in which the EEG is responsive to variations in task difficulty.
Part of these variations may be attributable to between-participant differences in motivation or strategy. For example, some participants may make a large effort even in trivial task conditions. Alternatively, some participants may be incapable of making, or unwilling to make, the proportionately greater effort required for successful performance in the most difficult conditions, instead shedding subtasks and settling for low levels of performance speed or accuracy. Part of such differences may also reflect differences in cognitive capacity among individuals (Gevins & Smith, 2000). That is, for some participants, even the ML task may require a full commitment of their intellectual resources, whereas other individuals may still have resources in reserve at that difficulty level.
These possibilities present compelling and important topics for future research and point to the criticality of analytic strategies that adapt gracefully to individual differences in brain functions and cognitive abilities.
These results indicate that multivariate combinations of EEG variables can be used to track changes in task loading that occur as individual participants engage in complex computer based work under conditions of variable difficulty. This modality can provide measures of covert task loading even in circumstances in which regular samples of task performance may be lacking or when overt failures in performance are infrequent. Furthermore, such data can be obtained without restructuring tasks or imposing secondary tasks on a human operator. Thus methods such as those described in this paper might eventually evolve into a monitoring modality that could complement the strengths of more traditional methods (such as subjective ratings or measurements of overt behavior), extending the set of tools available to researchers interested in examining issues of attention and performance in naturalistic contexts.
The National Aeronautics and Space Administration and the Air Force Office of Scientific Research supported this work. We thank Linda McEvoy, Jennie Barber, Gail Chang, Georgia Rush, and Daphne Yu for their contributions to this study.
Baker, S. C., Rogers, R. D., Owen, A. M., Frith, C. D., Dolan, R. J., Frackowiak, R. S., & Robbins, T. W. (1996). Neural systems engaged by planning: A PET study of the Tower of London task. Neuropsychologia, 34, 515-526.
Bullock, T. H., & McClune, M. C. (1989). Lateral coherence of the electroencephalogram: A new measure of brain synchrony. Electroencephalography and Clinical Neurophysiology, 73, 479-498.
Bunge, S. A., Klingberg, T., Jacobsen, R. B., & Gabrieli, J. D. (2000). A resource model of the neural basis of executive working memory. Proceedings of the National Academy of Sciences, USA, 97, 3573-3578.
Byrne, E. A., & Parasuraman, R. (1996). Psychophysiology and adaptive information. Biological Psychology, 42, 249-268.
Card, S. K., Moran, T. P., & Newell, A. (1983). The psychology of human- computer interaction. Mahwah, NJ: Erlbaum.
Carpenter, P. A., Just, M. A., & Reichle, E. D. (2000). Working memory and executive function: Evidence from neuroimaging. Current Opinion in Neurobiology. 10, 195-199.
Comstock, J. R., & Amegard, R. J. (1992). The multi-attribute task battery for human operator workload and strategic behavior research (NASA Tech. Memorandum 104174). Hampton, VA: Langley Research Center.
Du, W., Leong, H., & Gevins, A. (1994, June). Ocular artifact reduction by adaptive filtering. Presented at the 7th IEEE Signal Processing Workshop on Statistical Signal and Array Processing, Quebec City, Canada.
Elul, R. (1969). The physiological interpretation of amplitude histograms of the EEG. Electroencephalography and Clinical Neurophysiology, 27, 703-704.
Fournier, L. R., Wilson, G. F., & Swain, C. R. (1999). Electrophysiological, behavioral, and subjective indexes of workload when performing multiple tasks: Manipulations of task difficulty and training. International Journal of Psychophysiology. 31, 129-145.
Garavan, H., Ross, T. J., Li, S., & Stein, E. A. (2000). A parametric manipulation of central executive functioning. Cerebral Cortex, 10, 585-592.
Gevins, A., & Smith, M. E. (1999). Detecting transient cognitive impairment with EEG pattern recognition methods. Aviation, Space and Environmental Medicine, 70, 1018-1024.
Gevins, A., & Smith, M. E. (2000). Neurophysiological measures of working memory and individual differences in cognitive ability and cognitive style. Cerebral Cortex, 10, 829-839.
Gevins, A., Smith, M. E., Leong. H., McEvoy, L., Whitefield, S., Du, R., & Rush, G. (1998). Monitoring working memory load during computer-based tasks with EEG pattern recognition methods. Human Factors, 40, 79-91.
Gevins, A., Smith, M. E., McEvoy, L., & Yu, D. (1997). High-resolution EEG mapping of cortical activation related to working memory: Effects of task difficulty, type of processing, and practice. Cerebral Cortex, 7, 374-385.
Gevins, A. S., & Schaffer, R. E. (1980). A critical review of electroencephalographic EEG correlates of higher cortical functions. CRC Critical Reviews in Bioengineering, 4, 113-164.
Gevins, A. S., Zeitlin, G. M., Doyle. J. C., Schaffer, R. E., & Callaway, E. (1979). EEG patterns during "cognitive" tasks: II. Analysis of controlled tasks. Electroencephalography and Clinical Neurophysiology, 47, 704-710.
Gevins, A. S., Zeitlin, G. M., Yingling, C. D., Doyle, J. C., Dedon, M. F., Schaffer, R. E., Roumasset, J. T., & Yeager, C. L. (1979). EEG patterns during "cognitive" tasks: I. Methodology and analysis of complex behaviors. Electroencephalography and Clinical Neurophysiology, 47, 693-703.
Glass, A. (1966). Comparison of the effect of hard and easy mental arithmetic upon blocking of the occipital alpha rhythm. Quarterly Journal of Experimental Psychology, 8, 142-152.
Gopher, D., Brickner, M., & Navon, D. (1982). Different difficulty manipulations interact differently with task emphasis: Evidence for multiple resources. Journal of Experimental Psychology: Human Perception and Performance, 8, 146-157.
Gundel, A., & Wilson, G. F. (1992). Topographical changes in the ongoing EEG related to the difficulty of mental tasks. Brain Topography, 5, 17-25.
Hart, S. G., & Staveland, L. E. (1988). Development of NASATLX (Task Load Index): Results of experimental and theoretical research. In P. A. Hancock & N. Meshakati (Eds.), Human mental workload (pp. 139-183). Amsterdam: North-Holland.
Humphrey, D., & Kramer, A. F. (1994). Toward a psychophysiological assessment of dynamic changes in mental workload. Human Factors, 36, 3-26.
Inouye, T., Shinosaki, K., Jyama, A., Matsumoto, Y., Toi, S., & Ishihara, T. (1994). Potential flow of frontal midline theta activity during a mental task in the human electroencephalogram. Neuroscience Letters, 169, 145-148.
Ishii, R., Shinosaki, K., Ukai, S., Inouye, T., Ishihara, T., Yoshimine, T., Hirabuki, N., Asada, H., Kihara, T., Robinson, S. E., & Takeda, M. (1999). Medial prefrontal cortex generates frontal midline theta rhythm. Neuroreport, 10, 675-679.
Kendall. M. G. (1963). Rank order correlation methods (3rd ed.). London: Griffin.
Klimesch, W., Schmike, H., & Pfurtscheller, G. (1993). Alpha frequency, cognitive load, and memory performance. Brain Topography, 5, 241-251.
Larson, C. L., Davidson, R. J., Abercrombie, H. c., Ward, R. T., Schaefer, S. M., Jackson, D. C., Holden, J. E., & Perlman, S. B. (1998). Relations between PET-derived measures of thalamic glucose metabolism and EEG alpha power. Psychophysiology, 35, 162-169.
Lopes da Silva, F. H. (1991). Neural mechanisms underlying brain waves: From neural membranes to networks, Electroencephalography and Clinical Neurophysiology, 79, 81-93.
Lopes da Silva, F. H., & Storm van Leeuwen, W. (1978). The cortical alpha rhythm in dog: The depth and surface profile of phase. In M. A. B. Brazier & H. Petsche (Eds.), Architectonics of the cerebral cortex (pp. 319-333). New York: Raven.
Lopes da Silva, F. H., Vos, J. E., Mooibreck, J., & Van Rotterdam, A. (1980). Relative contributions of intracortical and thalamocortical processes in the generation of alpha rhythms, revealed by partial coherence analysis. Electroencephalography amsd Clinical Neurophysiology, 50, 449-456.
McCallum, W. C., Cooper, R., & Pocock, P. V. (1988). Brain slow potential and ERP changes associated with operator load in a visual tracking task. Electroencephalography and Clinical Neurophysiology, 69, 453-468.
Miyata, Y., Tanaka, Y., & Hono, T. (1990). Long-term observation on Fm-theta during mental effort. Neuroscience, 16, 145-148.
Molloy, R., & Parasuraman. R. (1996). Monitoring an automated system for a single failure: Vigilance and task complexity effects. Human Factors, 38, 311-322.
Morrison, J. G., & Gluckman, J. P. (1994). Definitions and prospective guidelines for the application of automation. In M. Mouloua & R. Parasuraman (Eds.), Human performance in automated systems: Current research and trends (pp. 256-263). Hillsdale, NJ: Eribaum.
Mulholland, T. (1995). Human EEG, behavioral stillness and biofeedback. International Journal of Psychology, 19, 263-279.
Nygren, T. E. (1991), Psychometric properties of subjective workload measurement techniques: Implications for their use in the assessment of perceived mental workload. Human Factors. 33, 17-33.
Parasuraman, R. (1990). Event-related brain potentials and human factors research. In J. W. Rohrbaugh, R. Parasuraman, & R. J. Johnson (Eds.), Event-related brain potentials (pp. 279-300). New York: Oxford University Press.
Parasuraman, R., Molloy, R, & Singh, I. L. (1993). Performance consequences of automation-induced "complacency." International Journal of Aviation Psychology, 3, 1-23.
Parasuraman, R., Mouloua, M., & Molloy, R. (1996). Effects of adaptive task allocation on monitoring of automated systems. Human Factors. 38, 665-679.
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man and Cybernetics -- Part A: Systems and Humans, 30, 286-297.
Paus, T., Koski, L., Caramanos, Z., & Weatbury. C. (1998). Regional differences in the effects of task difficulty and motor output on blood flow response in the human anterior cingulate cortex: A review of 107 PET activation studies. Neuroreport, 9, R37-R47.
Pellouchoud, E.. Smith, M. E., McEvoy, L., & Gevins, A. (1999). Mental effort-related EEG modulation during video-game play: Comparison between juvenile subjects with epilepsy and normal control subjects. Epilepsia, 40(Suppl. 4), 38-43.
Pfurtscheller, G., & Klimesch, W (1992). Functional topography during a visuoverbal judgment task studied with event-related desynchronization mapping. Journal of Clinical Neurophysiology, 9, 120-131.
Posner, M. E., & Peterson, S. E. (1990). The attentional system of the human brain. Annual Review of Neuroscience, 13, 25-42.
Posner, M. I., & Rothbart, M. K. (1992). Attentional mechanisms and conscious experience. In A. D. Milner & M. D. Rugg (Eds.), The neuropsychology of consciousness (pp. 91-111). San Diego: Academic.
Raskin, J. (2000). Humane interface: New directions for designing interactive systems. Boston: Addison-Wesley.
Sadato, N., Nakamura, S., Oohashi, T., Nishina, E.. Fuwamoto, Y, Waki, A., & Yonekura, Y. (1998). Neural networks for generation and suppression of alpha rhythm: A PET study. Neuroreport. 30, 893-897.
Sharbrough, F., Chatrian, G., Lesser, R. P., Luders, H., Nuwer, M., & Picton, T. W. (1990). Guidelines for standard electrode position nomenclature. Bloomfield, IL: American EEC Society.
Smith, M. E., & Gevins, A. (1998, September). EEG-based monitoring of cognitive load during human-computer interaction. Presented at the Annual Meeting of the Society for Psychophysiological Research, Boulder, CO.
Smith, M. E., McEvoy, L. K., & Cevins, A. (1999). Neurophysiological indices of strategy development and skill acquisition. Cognitive Brain Research, 7, 389-404.
Sterman, M. B., Mann, C. A., Kaiser, D. A., & Suyenobu, B. Y. (1994). Multiband topographic analysis of a simulated visuomotor aviation task. International Journal of Psychophysiology, 16, 49-56.
Tou, J. T., & Gonzalez, R. C. (1974). Pattern recognition principles. Reading, MA: Addison-Wesley.
Wickens, C. D. (1991). Processing resources and attention. In D. L. Damos (Ed.), Multiple-task performance (pp. 1-34). London: Taylor & Francis.
Yamamoto, S., & Matsuoka, 5. (1990). Topographic EEG study of visual display terminal VDT performance with special reference to frontal midline theta waves. Brain Topography, 2, 257-267.
[Figure 2 omitted]
TABLE 1 Mean (SD) Subjective Load Ratings on the NASA TLX Scale for Each Task Version and Behavioral Performance Measures for Each Subtask (N = 16) Subtask/Measure Low Load Medium Load Subjective load (0-100) 14.08 (10.22) 39.72 (18.15) Monitoring: RT (s) 1.84 ( 0.52) 2.77 ( 0.87) Monitoring: Error rate 0.07 ( 0.11) 0.02 ( 0.02) Communications: RT (s) 8.72 ( 0.78) 9.19 ( 1.57) Communications: Error rate 0.01 ( 0.06) 0.05 ( 0.07) Tracking (RMS error) -- 16.57 (15.56) Resource manage (RMS error) -- 139.30 (87.47) Subtask/Measure High Load Subjective load (0-100) 53.12 (21.14) (**) Monitoring: RT (s) 3.02 ( 0.48) (**) Monitoring: Error rate 0.05 ( 0.06) Communications: RT (s) 9.40 ( 0.84) (*) Communications: Error rate 0.04 ( 0.04) Tracking (RMS error) 39.56 (21.05) (**) Resource manage (RMS error) 169.02 (70.84) (*)p < .01 when comparing low- and high-load task versions. (**)p < .001 when comparing low- and high-load task versions (medium- vs. high-load comparison for the tracking task, which was set to autopilot in the low-load condition). TABLE 2 Anatomical Regions and Frequency Bands for the EEG Features Included in Each Participant-Specific Index Anterior Left Midline Right Left Frontal Frontal Frontal Frontal Parietal Theta 5-6 Hz 9, 11, 16 -- 2, 3, 7, 16 -- -- 6-7 Hz 1, 2, 9, 10, -- 2, 4, 16 -- -- 11, 13 7-8 Hz 5, 16, 9 -- 15 -- -- Alpha 8-10 Hz -- 12, 3 6, 5, 10, 12, 3 1, 7, 8, 9, 13, 15 10, 14, 15 10-12 Hz -- 1, 12 3, 6 8, 6, 12 1, 6, 7, 8, 11, 13, 14, 15 Midline Right Parietal Parietal Theta 5-6 Hz -- -- 6-7 Hz -- -- 7-8 Hz -- -- Alpha 8-10 Hz 4, 14, 7 4, 5, 8, 13, 14 10-12 Hz 2, 3, 10, 11 4, 5, 12 Note. Interior cells list the set of participants (N = 16) who utilized the corresponding feature in their participant-specific index functions. Cells with dashes denote features that were not included in the set of candidate features used to create index functions. TABLE 3: Comparison of NASA TLX Subjective Load Ratings (Scale 0-100), MATB Monitoring Task RTs (s), and EEG Task Load Index Values (Scale 0-1) for Individual Study Participants for the High-, Medium-, and Low-Load Versions of the MATB Task and the Passive Watching Condition Subjective Monitoring Load Ratings Subtask RTs Participant HL ML LL HL ML 1 30 15 8 3.3 2.9 2 45 38 23 3.2 3.1 3 64 51 33 2.4 2.2 4 52 45 7 2.9 3.4 5 51 37 17 2.9 2.0 6 31 18 11 3.4 4.0 7 55 49 16 2.3 1.5 8 60 41 31 3.0 3.0 9 48 36 28 2.8 1.7 10 48 47 11 2.6 2.3 11 64 32 12 3.5 2.2 12 76 45 5 3.9 4.8 13 90 50 1 3.5 3.2 14 88 87 4 3.4 2.5 15 27 27 15 2.6 3.0 16 31 4 1 2.1 2.0 Mean 53 40 14 3.0 2.8 Monitoring EEG Task Subtask Loading Index RTs Participant LL HL ML LL PW 1 1.9 .68 .60 .27 .32 2 1.6 .53 .49 .14 .19 3 1.4 .70 .53 .45 .41 4 1.9 .69 .64 .4 .34 5 1.5 .62 .46 .31 .19 6 1.9 .52 .48 .41 .38 7 1.3 .67 .68 .62 .18 8 2.9 .62 .59 .34 .36 9 1.1 .62 .60 .87 .66 10 1.9 .55 .55 .42 .31 11 1.5 .61 .53 .51 .46 12 2.0 .63 .57 .27 .17 13 1.5 .66 .63 .40 .28 14 1.8 .56 .58 .49 .51 15 2.2 .82 .60 .46 .33 16 1.0 .63 .59 .50 .47 Mean 1.8 .63 .57 .43 .35 Note: Subjective and behavioral measures were not obtained during passive watching. TABLE 4: Pairwise Comparisons of Mean Task Load Index within Individual Participants and between the Various Task Difficulty Conditions Comparison Participant HV vs. ML ML vs. LL LL vs. PW HL vs. LL HL vs. PW 1 .002 .001 .350 .001 .001 2 .250 .001 .380 .001 .001 3 .001 .020 .240 .001 .001 4 .090 .001 .170 .001 .001 5 .001 .001 .005 .001 .001 6 .020 .004 .060 .001 .001 7 .910 .002 .001 .002 .001 8 .080 .001 .500 .001 .001 9 .660 .001 .001 .001 .410 10 .820 .001 .020 .001 .001 11 .040 .250 .210 .020 .001 12 .020 .001 .040 .001 .001 13 .670 .002 .030 .001 .001 14 .130 .001 .530 .001 .002 15 .001 .001 .001 .001 .001 16 .410 .001 .980 .001 .050 Comparison Participant df 1 27 2 27 3 27 4 25 5 27 6 25 7 27 8 27 9 27 10 29 11 27 12 25 13 23 14 27 15 25 16 23 Note. P values represent the outcome of a dependent t test (two-tailed) computed on subsamples of each participant's data. Degrees of freedom were the same for each pairwise comparison within each participant but differed between participants because some data were rejected because of artifacts in the EEG.
|Printer friendly Cite/link Email Feedback|
|Author:||Smith, Michael E.; Gevins, Alan; Brown, Halle; Karnik, Arati; Du, Robert|
|Article Type:||Statistical Data Included|
|Date:||Sep 22, 2001|
|Previous Article:||Display signaling in augmented reality: Effects of cue reliability and image realism on attention allocation and trust calibration.|
|Next Article:||Effect of flooring on standing comfort and fatigue.|