Biological Psychology 33 (1992) 37-49 0 1992 Elsevier Science Publishers B.V. All rights reserved
Operant conditioning Werner
Sommer
and Stefan
37 0301-0511/92/$05.00
of P300 * Schweinberger
Department of Psychology, University of Konstanz, Germany
This study demonstrates that operant conditioning may increase the P300 component of the event-related potential above a level obtained without contingent training. An experimental group of subjects was rewarded for producing large P300 amplitudes and was compared with a yoked control group which was rewarded on a random basis. During training the experimental subjects increased both the amplitude of the P300 and of a subsequent frontal negative slow wave relative to the control group. These training effects were independent of prestimulus potential shifts and occurred likewise for target and nontarget stimuli. Keywords: ERPs,
P300, operant
conditioning,
biofeedback
1. Introduction Conditioning techniques are an alternative tool to the more common task-manipulative approach in studying the functional significance of eventrelated brain potentials (ERPs) (Rosenfeld, Stamm, Elbert, Rockstroh, Birbaumer, & Roger, 1984). One candidate for this technique is the P300 component which is elicited after task-relevant stimuli and whose theoretical interpretation is still controversial (e.g. Donchin & Coles, 1988; Verleger, 1988). Roger and Galand (1981) were the first to operantly condition a positive ERP component in the latency range of the P300. They reported increases and decreases in the amplitude of the visual ERP to checkerboard stimuli in the time segments between 280 to 300 ms. However, only one type of stimulus was used, which is unfavourable for eliciting the P300, and only occipital activity was recorded where P300 amplitude is small. Therefore, it is impossible to decide whether the reported component can be regarded as a P300. Shabsin (1982) rewarded five children with learning problems for large Correspondence to: Dr. W. Sommer, Fachgruppe Psychologie, Universitat Konstanz, Postfach 5560, 7750 Konstanz, Germany. * This research was supported by the Deutsche Forschungsgemeinschaft. We thank Ursula Lommen for assisting in data acquisition and R.B. Freeman, Jr., A.W.K. Gaillard, and two anonymous reviewers for helpful comments on earlier versions of this paper.
3x
W Sommer and S.’ Schweinberger
/Conditioning
P300
P300 amplitudes elicited by visual stimuli. However, no clear increases in amplitude as a function of training were found. Miltner, Larbig, and Braun (1986) showed differential effects of training for increases and decreases (up- vs. downtraining) the amplitude of ERPs between 200 and 600 ms after the onset of visual target stimuli. Recently, these authors have extended their findings to the somatosensory P260 component, elicited by pain stimuli (Miltner, Braun, & Larbig, 1987). However, as no data were avaiIable from a baseline or a random feedback condition, the differential effects of up- and downtraining might have been due to P300 enhancement in uptraining, P300 reduction in downtraining, or a combination of these. Using an intrasubject design, Sommer (1987) confirmed the reports of Miltner and coworkers of differential effects of up- and downtraining P300 for auditory stimuli, but P300 amplitude during training was not larger than in an initial condition with random feedback. The training effects seemed to be superimposed on a general amplitude decline. If downtraining the P300 is in fact easier than uptraining, as suggested by these data, differential effects of up- and downtraining may be predominantly related to the contributions of downtraining. The present experiment was designed to investigate whether P300 amplitude can be increased above the course it would take without training, regardless of whether this course is declining, stable, or increasing. Therefore, an experimental group was rewarded for large P300 amplitudes and compared with a control group receiving random reinforcement that was not contingent on P300 amplitude. This design allows for a direct comparison of the course of P300 with and without uptraining, instead of relying on the comparison of two opposite training conditions or on comparing the level at the start with that at the end of the uptraining, which may be confounded by time-dependent changes. In addition, a long interstimulus interval (ISI) of 3 s was chosen to allow the recording of slow prestimulus shifts. Also prestimulus shifts may become subject to voluntary control (Rockstroh, Elbert, Canavan, Lutzenberger, & Birbaumer, 1989) and may contribute to P300 amphtudes. 2. Method 2. I. Subjects Thirty-four subjects who reported normal or corrected-to-normal visual acuity, normal colour vision, and hearing, were recruited for this study. Most of them were psychology students and had previously participated in other experiments, but not in ERP studies. The data of six subjects were excluded because of technical problems or large amounts of ocular artifacts. The 28
W. Sommer and S. Schweinberger / Conditioning P300
remaining subjects were divided into an experimental and a control balanced for age (M = 23.2 and 22.4 years) and gender (women/men and 8/6).
39
group, = 7/7
2.2. Stimuli Auditory stimuli were 65 dB (A) sinusoidal tones of either 1000 or 1500 Hz and 65 ms duration, including 15 ms rise/fall times. The tones were presented at equal probabilities in random order through a loudspeaker 1m in front of and slightly below the eye level of the subjects. Ventilation provided a masking noise of 42 dB (A). Visual stimuli were dim (1.5 cd rn-‘) green, red, and white lights presented for 65 ms behind an opaque disk of 1.5” visual angle at eye level; ambient light was 0.9 cd m-*. 2.3. Procedure A total of eight blocks of 125 stimuli each were presented, consisting of a random feedback condition, succeeded by seven blocks of training. Blocks were separated by short breaks. In all conditions the high-pitched tones were assigned as targets. The interval between stimuli was 3 s. Subjects were given feedback about the amplitude of the P300 to each target tone, by presenting one of the light stimuli instead of the next tone. A white light indicated that an ocular or muscular artifact had occurred. Otherwise, the target was followed by a green or a red light, depending on the experimental group, condition, and the magnitude of P300. During the random feedback block, the green and red lights were presented randomly at a ratio of 2: 1, to obtain for each subject a feedback criterion for the first training block. During training the experimental group received a green light after a target if the corresponding ERP was artifact free and if the P300 amplitude exceeded the criterion. An amplitude below this criterion was followed by a red light. The criterion was defined as the average P300 amplitude during the preceding block minus 0.4 standard deviations. With constant response characteristics this algorithm provides a reward-punishment ratio of about 2: 1. As changes in response magnitude or distribution due to learning or other processes affect this proportion, each control subject was yoked blockwise to the reinforcement ratio of green to red lights of a particular experimental subject. Thus, experimental and control subjects should be comparable with respect to feedback quantities, differing only in feedback contingency. 2.4. Task Prior to the experiment subjects received and procedures of the study. They were
written information on the goal told to find out how to deal
40
W Sommer and S. Schweinberger / Conditioning P300
“mentally” with the target tones in order to produce a particular kind of electrical brain response. Whether or not the desired response was achieved would be indicated by the presentation of the green or the red light. In addition to the basic bonus of DM 15.00, subjects with above- and belowcriterion responses were respectively rewarded with or fined 0.10 DM. Information about the current balance was given after each block. Subjects were also advised to avoid muscular activity and eye movements or blinks, in which case the white light would be presented without monetary consequences. 2.5. Questionnaire At the end of the experimental session, a questionnaire was presented concerning the strategies used and considered as suited for achieving rewards: attending, valuing, expecting, counting the targets, comparing them with other events or memories, or any other. Subjects also indicated on visual analogue scales from 0 to 100 the experienced strength of association between their “behaviour” and the feedback, of the training success, and how much interest they had taken in the study. The latter questions were aimed at possible group differences in motivation and at whether the control subjects had noticed the noncontingency of the feedback stimuli. 2.6. Recording and averaging The electroencephalogram (EEG) was recorded from Fz, Cz, and Pz with linked earlobes as common reference. In addition, recordings were made of the vertical electrooculogram (EOG) from above and below the right eye and of the electromyogram (EMG) from two submandibular points. Ag/AgCl electrodes (Grass ESSH) were used, affixed with adhesive collars and filled with Beckman Electrode Electrolyte paste. The time constants for all signals were 5.5 s and low pass filters were set to 40 Hz (-3 dB attenuation, 12 dB roll-off/octave). The sampling rate for all signals was 100 Hz. As the EMG was only screened for excessive excursions, it does not matter that this sampling rate cannot faithfully represent its high frequency activity. All signals were digitized for 2 s, starting 1 s prior to stimulus onset, and stored on magnetic tape for offline processing. For the online measurement of P300 amplitudes in single-trial ERPs, the EEG from Pz was digitally low pass filtered at 3.4 Hz (- 3 dB attenuation; Ruchkin & Glaser, 1978). P300 amplitude was quantified as the most positive value within 250 and 390 ms after stimulus onset (criterion segment) referred to the average in the 100 ms period prior to stimulus onset. For feedback purposes a response was considered artifact free if a criterion of 100 PV was not exceeded by the EOG or EMG activity within a 700 ms period starting 100 ms before stimulus onset.
W. Sommer
and S. Schweinberger
/ Conditioning
41
P300
When the ERPs were averaged, these artifact criteria were applied to the whole sampling epoch of 2 s. Averaging was performed separately for the random feedback, each of the seven blocks of training, and separately for target and nontarget tones. Prior to analysis the averaged ERPs were digitally low pass filtered at 8.8 Hz (- 3 dB).
3. Results Table 1 shows the percentage of positive feedback trials in each block for the two groups. The results show that the matching procedure had worked well for the training blocks; the difference in rewards between the experimental and control group did not exceed 6.5%. Figure 1 depicts the grand average ERPs for the random feedback and the seven training blocks. Preceding the stimulus there is a gradually rising negativity. After the stimulus, a vertex potential is followed by an N200, a parietally distributed P300 around 330 ms, and finally a broad fronto-central, negative component, henceforth termed negative slow wave (NSW). 3.1. Amplitude measures The amplitude of the P300 was quantified as the average voltage between 260 and 400 ms referred either to the 100 ms before stimulus onset (prestimulus baseline) or to the time point at the beginning of the recording epoch, 990 ms before stimulus onset (early baseline). The two baseline measures were used in order to assess contributions of prestimulus activity to the P300 amplitude. In addition, we measured the NSW amplitude between 360 and 740 ms referred to the early baseline. Each amplitude measure was submitted to an analysis of variance (ANOVA) with a group factor and repeated measures for electrode site, stimulus (target vs. nontarget) and the seven blocks of training. Degrees of freedom were adjusted for violations of the sphericity assumption with the Huynh and Feldt correction method. Both measures of P300 amplitude confirmed the parietal scalp distributions, for the prestimulus baseline CM (Fz to Pz) = - 1.2, 0.6, and 3.3 pV; F
Table 1 The percentage
of positive
feedback
trials
Group
Random Feedback
Block of Training 1
2
3
4
5
6
I
Experimental Control
75 87
57 64
65 61
64 65
68 65
61 58
68 66
67 67
Pz
Fz Random
Feedback
Block
6
7
Fig
I. ERPs
averaged
for the random
feedback
over target and nontarget
and control
group (broken
condition
and seven blocks of training.
tones and superimposed
lines). Tick marks
denote
intewals
Waveshapes
for the experimental
of 500 ms; stimulus
are
(solid lines)
onset is at
I s.
W. Sommer
and S. Schweinberger
/ Conditioning
P300
43
(2/52) = 60.3; p < 0.001) and the early baseline (M (Fz to Pz) = -0.3, 2.0, and 5.6 pV; F (2/52) = 100.2; p < O.OOl), and were strongly affected by the
stimulus as a main effect (F (l/26) = 14.5 and 22.8, p < 0.001) and in interaction with the electrode site (F (2/52) = 50.6 and 45.0, p < 0.001). Both effects reflect differences in the P300 between target and nontarget stimuli; while differences were small at Pz, the amplitude reduction over Cz to Fz was more pronounced for nontargets than for targets. The interaction between the groups and training blocks was significant for both amplitude measures, but only as an interaction with electrode site (F (12/312) = 2.38 and 3.32, p < 0.05 and 0.01, prestimulus and early baseline, respectively). Figure 1 shows that during the initial blocks of training both groups displayed similar P300 amplitudes, but during the final blocks the central and parietal P300 was markedly larger for the experimental than for the control group; at Fz, P300 amplitude was negligible throughout. An additional ANOVA of the prestimulus baseline amplitude, referred to the early baseline, confirmed that there were no training effects on the prestimulus activity (Fs < 1.3; ps > 0.20). As the training effects in P300 amplitude were present with both baseline measures and absent for the prestimulus baseline alone, we may safely exclude significant prestimulus slow potential contributions. There were, however, training effects in interaction with group and electrode site, F (12/312) = 2.83, p < 0.01, in the NSW following P300. This component was of fronto-central scalp topography (M (Fz to Pz) = -3.6, -3.1, and -0.5 pV>, F (2/52) = 39.0, p < 0.001, and considerably larger for targets (M = - 3.4 PV) than nontargets (M = - 1.1 pV1, F (l/26) = 24.6, p < 0.001. The training effects for the NSW were complex with more negative values for experimental than control subjects during the final blocks at Fz (e.g. Training Block 7: -4.6 vs. -3.0 PV) but more positive values at Pz (1.5 vs. - 1.4 pV>. As the NSW was generally small at Pz, this training effect at Pz may stem from overlap with the P300 component. In order to obtain measures free of component overlap we performed a principal components analysis (PCA) with subsequent varimax rotation. 3.2. Principal components analysis The PCA was based on the covariance matrix of the ERPs, referred to the first data point and with every other sampling point omitted. Figure 2 shows the loadings of the seven-factor solution, accounting for 85.4% of the data variance. The scores of each PCA-derived factor were submitted to the same type of ANOVA as the amplitude measures. The only two factors showing significant interactions involving groups and training blocks were Factor 1 (40.1% of data variance) and Factor 5 (5.6%), which represented most of the ERP
44
W. Sommer
and 5’. Schweinberger / Conditioning
P300
45
W. Sommcr and S. Schweinberger / Conditioning P300
NSW
P300 1 0.5-
Fz
o
:
-0,5-1
1
Pz -0,5-
-0,5I
-I R’Fil
1 d
I 5
,
I 7
Blocks Fig. 3. P300 and NSW factor scores for the random blocks of training, averaged over target and nontarget mental (solid lines) and control group (broken lines).
-1
I RF1
1
I
I
,
3
I 5
I
1 7
Blocks feedback condition (RF) and the seven tones, and superimposed for the experi-
activity during the segment that was used as a criterion for feedback (see Fig. 3). The loadings of Factor 1 increased at about 100 ms after stimulus onset towards a plateau, lasting from around 250 ms until the end of the recording epoch. This factor obviously represents the fronto-central NSW (M (Fz to Pz) = - 0.32, - 0.23, 0.51, F (2/52) = 44.47, p < 0.001). It was more pro-
46
W. Sommer and S. Schweinberger
/ Conditioning
P300
nounced after targets (M = -0.25) than after nontargets (M = 0.21, F (l/26) = 19.3, p < O.OOl>, and showed a three-way interaction between the groups, training blocks and electrode site (F (12/312) = 2.88, p < 0.01) which was even clearer for the linear trend over blocks (F (linear: l/26) = 7.23, p < 0.05). Post-hoc comparisons indicated that the differential linear trends between the groups were restricted to the Fz electrode (F (l/26) = 5.91, p < O.lO>, and absent at Cz and Pz (F < 1). The absence of training effects on the parietal scores indicates that the PCA had been successful in disentangling the P300 and the NSW components. At Fz the NSW of the controls progressively declined over blocks while it was constant or even slightly increased in the experimental subjects (Fig. 3). During the initial random feedback condition there were no group differences either in the P300 or in the NSW (ps > 0.10). ERP activity almost exclusively restricted to the time segment of the P300 component is represented by Factor 5 peaking around 310 ms. For the training blocks the scores showed a parietal scalp distribution (M (Fz to Pz> = -0.05, -0.14, 0.241, F (2/52) = 6.76, p < 0.01, which was more pronounced for targets (M (Fz to Pz) = -0.11, -0.14, and 0.42) than for nontargets (M (Fz to Pz) = 0.01, -0.13, and 0.061, F (2/52) = 19.46, p < 0.001). The interaction between the groups and the training blocks approached significance in the P310 factor scores, (F (6/156) = 2.36, p = 0.06); moreover, the groups differed in the linear trends of the P300 over blocks, (F (linear: l/26) = 4.77, p < 0.05). As in the amplitude measures, the training effects in P300 factor scores were independent of the target/nontarget factor (p > 0.27). Factor 2, which covers the prestimulus time segment, did not show any group-related effects (p > 0.35). This converges with the independence of the P300 amplitude measures from the choice of baseline. To provide an idea of the intersubject variability of the training effects, we calculated for each subject the linear regression slopes of the P300 factor scores for targets over the seven blocks of training. Table 2 shows the slopes of these regression lines, indicating that 9 out of 14 experimentals, but only 3 out of 14 controls, had been able to increase their P310, x2 = 5.25, df = 1, p < 0.05. 3.3. Questionnaire data Six subjects of the experimental group and five of the controls named “waiting for the target, preparing for its arrival” as the strategy primarily used for achieving rewards. This strategy was also experienced as the most successful one by five experimental and three control subjects. The strategy used next most frequently (three subjects in each group) was to “indifferently accept the stimulus”. ‘Both groups had taken much interest in the experi-
W. Sommer Table 2 Slopes of the linear individual subjects SubjectPair
Mean SD
regression
and S. Schweinberger
of target
P300 factor
/ Conditioning
scores
47
P300
over the 7 blocks
of training
for
Group Experimental
Control
-0.13 + 0.03 + 0.01 +0.15 + 0.02 - 0.08 + 0.01 - 0.02 +0.16 + 0.06 + 0.02 - 0.09 + 0.06 -0.18
- 0.30 -0.13 - 0.32 -0.16 + 0.02 - 0.04 0.00 - 0.28 +0.12 - 0.06 -0.12 +0.16 - 0.08 - 0.03
0.00 0.10
- 0.09 0.15
merit, the controls even more so than the experimentals (M = 67.5 and 84.2 scale units, respectively; Kruskal-Wallis test statistic = 5.39, p < 0.05). The association between “behaviour” and feedback had been considered as fairly strong by experimental (M = 65.7, SD = 21.5) as well as control subjects (A4 = 64.6, SD = 14.0) and both groups had been about equally satisfied with their training success (it4 = 68.2, SD = 14.5, and A4 = 58.0, SD = 23.2, respectively).
4. Discussion In line with the hypothesis, a differential trend of the P300 amplitude was observed for the two groups across the seven blocks of training, with a larger amplitude increase in the experimental group than in the control group. This finding demonstrates genuine effects of operant conditioning on P300 amplitude, which could not be found by prior investigators who had not separated the effects of up- and downtraining. One might argue that this result merely represents the restoration of the ERP differences in the random feedback block to be seen in Figs. 1 and 3, which might have vanished somehow during the first few blocks of training. However, these group differences were not significant. Even if one takes
48
W. Sommer and S. Schweinberger / Conditioning
P300
these subtle differences seriously, they do not invalidate the conditioning results. Specifically in the random feedback block the control group had received more rewards (87%) than the experimental group (75%) - because of a programming error, as noticed later. This may have temporarily affected factors such as task involvement, leading to group differences in P300 amplitude. For the training conditions reinforcement ratios had been kept almost identical for the two groups and, initially, also the P300 amplitudes were very similar; it was only during the course of training that group differences in P300 amplitude emerged. Are the training effects specific for P300 or are they related to other ERP components? Firstly, no changes in prestimulus negativity were found which might explain the effects of training on P300. Of the other components, a group-specific trend over blocks was only to be seen in the NSW component overlapping P300 and covering the last part of the recorded epoch. The frontal negativity of the NSW in the control group decreased during the training. However, it is hard to conceive how the reduction of a frontal negativity might diminish a preceding parietal positivity (P300) - in fact, the opposite would have to be expected. While both the NSW and P300 changes appear to be related to the training, their mutual connection, if there is any, seems to be more indirect. It is possible that the decline of the P300 and NSW in the controls are both related to a gradual reduction in attention or motivation (Ruchkin & Sutton, 1983). The lack of stimulus specificity of the training effects for both P300 and NSW is in line with this view. While both components were more pronounced for targets than for nontargets, the training effect was equally present for both stimuli. This may indicate that uptraining did not so much affect stimulus-specific processes but rather helped to maintain a general responsiveness in the experimental group. On the other hand, the questionnaire reports do not suggest a decline in motivation or attention in the control group. The controls had not experienced the rewards as less related to their “behaviour” than the experimentals, nor had they been less satisfied with their training success. Moreover, both groups, and particularly the controls, had claimed much interest in the study. The other evidence available was similarly inconclusive on the issue of how control over P300 might have been achieved. The most often cited strategies were equally present in both groups. In conclusion, the present study shows that operant conditioning is able to increase the P300 component above the level observed in a randomly rewarded control group and therefore demonstrates a genuine uptraining effect. References Donchin, E., & Coles, M.G.H. (1988). Is the P300 component updating? Behavioral and Brain Sciences, 11, 357-374.
a manifestation
of context
W. Sommer and S. Schweinberger / Conditioning P300
49
Miltner, W., Larbig, W., & Braun, C. (1986). Biofeedback of visual evoked potentials. International Journal of Neuroscience, 29, 291-303. Miltner, W., Braun, C., & Larbig, W. (1987). Biofeedback of event related potentials evoked by nociceptive stimulation. Journal of Psychophysiology, 1, 283 (Abstract). Rockstroh, B., Elbert, T., Canavan, A., Lutzenberger. W., & Birbaumer, N. (1989). Slow cortical potentials and behauior. Baltimore: Urban & Schwarzenberg. Roger, M., & Galand, G. (1981). Operant conditioning of visual evoked potentials in man. Psychophysiology, 18, 477-482. Rosenfeld, J.P., Stamm, J., Elbert, Th., Rockstroh, B., Birbaumer, N., & Roger, M. (1984). Biofeedback of event-related potentials. In R. Karrer, J. Cohen & P. Tueting (Eds.), Brain and information: Event-related potentials. Annals of the New York Academy of Sciences, 425, 653-666. Ruchkin, D., & Glaser, E.M. (1978). Simple digital filters for examining CNV and P300 on single trial basis. In D. Otto (Ed.), Multidisciplinary perspectives in euent-related brain potential research (pp. 579-581). Washington DC: US Environmental Protection Agency, EPA-600/977-043. Ruchkin, D.S., & Sutton, S. (1983). Positive slow wave and P300: association and disassociation. In A.W.K. Gaillard & W. Ritter (Eds.), Tutorials in ERP research: Endogenous components (pp. 233-2.50). Amsterdam: North-Holland. Shabsin, H.S. (1982). Feedback controlled increases in P300 visual evoked potential amplitudes and neurological functioning and academic performance in learning disabled children. (Doctoral Dissertation, University of Tennessee, 1982). Dissertation Abstracts International, 43, 3067-B (University Microfilms No. DA8303716). Sommer, W. (1987). Operant conditioning of P300. Psychophysiology, 24, 612 (Abstract). Verleger, R. (1988). Event-related potentials and cognition: A critique of the context updating hypothesis and an alternative interpretation of P3. Behaoioral and Brain Sciences, II, 343-356.
Operant conditioning Werner
Sommer
and Stefan
37 0301-0511/92/$05.00
of P300 * Schweinberger
Department of Psychology, University of Konstanz, Germany
This study demonstrates that operant conditioning may increase the P300 component of the event-related potential above a level obtained without contingent training. An experimental group of subjects was rewarded for producing large P300 amplitudes and was compared with a yoked control group which was rewarded on a random basis. During training the experimental subjects increased both the amplitude of the P300 and of a subsequent frontal negative slow wave relative to the control group. These training effects were independent of prestimulus potential shifts and occurred likewise for target and nontarget stimuli. Keywords: ERPs,
P300, operant
conditioning,
biofeedback
1. Introduction Conditioning techniques are an alternative tool to the more common task-manipulative approach in studying the functional significance of eventrelated brain potentials (ERPs) (Rosenfeld, Stamm, Elbert, Rockstroh, Birbaumer, & Roger, 1984). One candidate for this technique is the P300 component which is elicited after task-relevant stimuli and whose theoretical interpretation is still controversial (e.g. Donchin & Coles, 1988; Verleger, 1988). Roger and Galand (1981) were the first to operantly condition a positive ERP component in the latency range of the P300. They reported increases and decreases in the amplitude of the visual ERP to checkerboard stimuli in the time segments between 280 to 300 ms. However, only one type of stimulus was used, which is unfavourable for eliciting the P300, and only occipital activity was recorded where P300 amplitude is small. Therefore, it is impossible to decide whether the reported component can be regarded as a P300. Shabsin (1982) rewarded five children with learning problems for large Correspondence to: Dr. W. Sommer, Fachgruppe Psychologie, Universitat Konstanz, Postfach 5560, 7750 Konstanz, Germany. * This research was supported by the Deutsche Forschungsgemeinschaft. We thank Ursula Lommen for assisting in data acquisition and R.B. Freeman, Jr., A.W.K. Gaillard, and two anonymous reviewers for helpful comments on earlier versions of this paper.
3x
W Sommer and S.’ Schweinberger
/Conditioning
P300
P300 amplitudes elicited by visual stimuli. However, no clear increases in amplitude as a function of training were found. Miltner, Larbig, and Braun (1986) showed differential effects of training for increases and decreases (up- vs. downtraining) the amplitude of ERPs between 200 and 600 ms after the onset of visual target stimuli. Recently, these authors have extended their findings to the somatosensory P260 component, elicited by pain stimuli (Miltner, Braun, & Larbig, 1987). However, as no data were avaiIable from a baseline or a random feedback condition, the differential effects of up- and downtraining might have been due to P300 enhancement in uptraining, P300 reduction in downtraining, or a combination of these. Using an intrasubject design, Sommer (1987) confirmed the reports of Miltner and coworkers of differential effects of up- and downtraining P300 for auditory stimuli, but P300 amplitude during training was not larger than in an initial condition with random feedback. The training effects seemed to be superimposed on a general amplitude decline. If downtraining the P300 is in fact easier than uptraining, as suggested by these data, differential effects of up- and downtraining may be predominantly related to the contributions of downtraining. The present experiment was designed to investigate whether P300 amplitude can be increased above the course it would take without training, regardless of whether this course is declining, stable, or increasing. Therefore, an experimental group was rewarded for large P300 amplitudes and compared with a control group receiving random reinforcement that was not contingent on P300 amplitude. This design allows for a direct comparison of the course of P300 with and without uptraining, instead of relying on the comparison of two opposite training conditions or on comparing the level at the start with that at the end of the uptraining, which may be confounded by time-dependent changes. In addition, a long interstimulus interval (ISI) of 3 s was chosen to allow the recording of slow prestimulus shifts. Also prestimulus shifts may become subject to voluntary control (Rockstroh, Elbert, Canavan, Lutzenberger, & Birbaumer, 1989) and may contribute to P300 amphtudes. 2. Method 2. I. Subjects Thirty-four subjects who reported normal or corrected-to-normal visual acuity, normal colour vision, and hearing, were recruited for this study. Most of them were psychology students and had previously participated in other experiments, but not in ERP studies. The data of six subjects were excluded because of technical problems or large amounts of ocular artifacts. The 28
W. Sommer and S. Schweinberger / Conditioning P300
remaining subjects were divided into an experimental and a control balanced for age (M = 23.2 and 22.4 years) and gender (women/men and 8/6).
39
group, = 7/7
2.2. Stimuli Auditory stimuli were 65 dB (A) sinusoidal tones of either 1000 or 1500 Hz and 65 ms duration, including 15 ms rise/fall times. The tones were presented at equal probabilities in random order through a loudspeaker 1m in front of and slightly below the eye level of the subjects. Ventilation provided a masking noise of 42 dB (A). Visual stimuli were dim (1.5 cd rn-‘) green, red, and white lights presented for 65 ms behind an opaque disk of 1.5” visual angle at eye level; ambient light was 0.9 cd m-*. 2.3. Procedure A total of eight blocks of 125 stimuli each were presented, consisting of a random feedback condition, succeeded by seven blocks of training. Blocks were separated by short breaks. In all conditions the high-pitched tones were assigned as targets. The interval between stimuli was 3 s. Subjects were given feedback about the amplitude of the P300 to each target tone, by presenting one of the light stimuli instead of the next tone. A white light indicated that an ocular or muscular artifact had occurred. Otherwise, the target was followed by a green or a red light, depending on the experimental group, condition, and the magnitude of P300. During the random feedback block, the green and red lights were presented randomly at a ratio of 2: 1, to obtain for each subject a feedback criterion for the first training block. During training the experimental group received a green light after a target if the corresponding ERP was artifact free and if the P300 amplitude exceeded the criterion. An amplitude below this criterion was followed by a red light. The criterion was defined as the average P300 amplitude during the preceding block minus 0.4 standard deviations. With constant response characteristics this algorithm provides a reward-punishment ratio of about 2: 1. As changes in response magnitude or distribution due to learning or other processes affect this proportion, each control subject was yoked blockwise to the reinforcement ratio of green to red lights of a particular experimental subject. Thus, experimental and control subjects should be comparable with respect to feedback quantities, differing only in feedback contingency. 2.4. Task Prior to the experiment subjects received and procedures of the study. They were
written information on the goal told to find out how to deal
40
W Sommer and S. Schweinberger / Conditioning P300
“mentally” with the target tones in order to produce a particular kind of electrical brain response. Whether or not the desired response was achieved would be indicated by the presentation of the green or the red light. In addition to the basic bonus of DM 15.00, subjects with above- and belowcriterion responses were respectively rewarded with or fined 0.10 DM. Information about the current balance was given after each block. Subjects were also advised to avoid muscular activity and eye movements or blinks, in which case the white light would be presented without monetary consequences. 2.5. Questionnaire At the end of the experimental session, a questionnaire was presented concerning the strategies used and considered as suited for achieving rewards: attending, valuing, expecting, counting the targets, comparing them with other events or memories, or any other. Subjects also indicated on visual analogue scales from 0 to 100 the experienced strength of association between their “behaviour” and the feedback, of the training success, and how much interest they had taken in the study. The latter questions were aimed at possible group differences in motivation and at whether the control subjects had noticed the noncontingency of the feedback stimuli. 2.6. Recording and averaging The electroencephalogram (EEG) was recorded from Fz, Cz, and Pz with linked earlobes as common reference. In addition, recordings were made of the vertical electrooculogram (EOG) from above and below the right eye and of the electromyogram (EMG) from two submandibular points. Ag/AgCl electrodes (Grass ESSH) were used, affixed with adhesive collars and filled with Beckman Electrode Electrolyte paste. The time constants for all signals were 5.5 s and low pass filters were set to 40 Hz (-3 dB attenuation, 12 dB roll-off/octave). The sampling rate for all signals was 100 Hz. As the EMG was only screened for excessive excursions, it does not matter that this sampling rate cannot faithfully represent its high frequency activity. All signals were digitized for 2 s, starting 1 s prior to stimulus onset, and stored on magnetic tape for offline processing. For the online measurement of P300 amplitudes in single-trial ERPs, the EEG from Pz was digitally low pass filtered at 3.4 Hz (- 3 dB attenuation; Ruchkin & Glaser, 1978). P300 amplitude was quantified as the most positive value within 250 and 390 ms after stimulus onset (criterion segment) referred to the average in the 100 ms period prior to stimulus onset. For feedback purposes a response was considered artifact free if a criterion of 100 PV was not exceeded by the EOG or EMG activity within a 700 ms period starting 100 ms before stimulus onset.
W. Sommer
and S. Schweinberger
/ Conditioning
41
P300
When the ERPs were averaged, these artifact criteria were applied to the whole sampling epoch of 2 s. Averaging was performed separately for the random feedback, each of the seven blocks of training, and separately for target and nontarget tones. Prior to analysis the averaged ERPs were digitally low pass filtered at 8.8 Hz (- 3 dB).
3. Results Table 1 shows the percentage of positive feedback trials in each block for the two groups. The results show that the matching procedure had worked well for the training blocks; the difference in rewards between the experimental and control group did not exceed 6.5%. Figure 1 depicts the grand average ERPs for the random feedback and the seven training blocks. Preceding the stimulus there is a gradually rising negativity. After the stimulus, a vertex potential is followed by an N200, a parietally distributed P300 around 330 ms, and finally a broad fronto-central, negative component, henceforth termed negative slow wave (NSW). 3.1. Amplitude measures The amplitude of the P300 was quantified as the average voltage between 260 and 400 ms referred either to the 100 ms before stimulus onset (prestimulus baseline) or to the time point at the beginning of the recording epoch, 990 ms before stimulus onset (early baseline). The two baseline measures were used in order to assess contributions of prestimulus activity to the P300 amplitude. In addition, we measured the NSW amplitude between 360 and 740 ms referred to the early baseline. Each amplitude measure was submitted to an analysis of variance (ANOVA) with a group factor and repeated measures for electrode site, stimulus (target vs. nontarget) and the seven blocks of training. Degrees of freedom were adjusted for violations of the sphericity assumption with the Huynh and Feldt correction method. Both measures of P300 amplitude confirmed the parietal scalp distributions, for the prestimulus baseline CM (Fz to Pz) = - 1.2, 0.6, and 3.3 pV; F
Table 1 The percentage
of positive
feedback
trials
Group
Random Feedback
Block of Training 1
2
3
4
5
6
I
Experimental Control
75 87
57 64
65 61
64 65
68 65
61 58
68 66
67 67
Pz
Fz Random
Feedback
Block
6
7
Fig
I. ERPs
averaged
for the random
feedback
over target and nontarget
and control
group (broken
condition
and seven blocks of training.
tones and superimposed
lines). Tick marks
denote
intewals
Waveshapes
for the experimental
of 500 ms; stimulus
are
(solid lines)
onset is at
I s.
W. Sommer
and S. Schweinberger
/ Conditioning
P300
43
(2/52) = 60.3; p < 0.001) and the early baseline (M (Fz to Pz) = -0.3, 2.0, and 5.6 pV; F (2/52) = 100.2; p < O.OOl), and were strongly affected by the
stimulus as a main effect (F (l/26) = 14.5 and 22.8, p < 0.001) and in interaction with the electrode site (F (2/52) = 50.6 and 45.0, p < 0.001). Both effects reflect differences in the P300 between target and nontarget stimuli; while differences were small at Pz, the amplitude reduction over Cz to Fz was more pronounced for nontargets than for targets. The interaction between the groups and training blocks was significant for both amplitude measures, but only as an interaction with electrode site (F (12/312) = 2.38 and 3.32, p < 0.05 and 0.01, prestimulus and early baseline, respectively). Figure 1 shows that during the initial blocks of training both groups displayed similar P300 amplitudes, but during the final blocks the central and parietal P300 was markedly larger for the experimental than for the control group; at Fz, P300 amplitude was negligible throughout. An additional ANOVA of the prestimulus baseline amplitude, referred to the early baseline, confirmed that there were no training effects on the prestimulus activity (Fs < 1.3; ps > 0.20). As the training effects in P300 amplitude were present with both baseline measures and absent for the prestimulus baseline alone, we may safely exclude significant prestimulus slow potential contributions. There were, however, training effects in interaction with group and electrode site, F (12/312) = 2.83, p < 0.01, in the NSW following P300. This component was of fronto-central scalp topography (M (Fz to Pz) = -3.6, -3.1, and -0.5 pV>, F (2/52) = 39.0, p < 0.001, and considerably larger for targets (M = - 3.4 PV) than nontargets (M = - 1.1 pV1, F (l/26) = 24.6, p < 0.001. The training effects for the NSW were complex with more negative values for experimental than control subjects during the final blocks at Fz (e.g. Training Block 7: -4.6 vs. -3.0 PV) but more positive values at Pz (1.5 vs. - 1.4 pV>. As the NSW was generally small at Pz, this training effect at Pz may stem from overlap with the P300 component. In order to obtain measures free of component overlap we performed a principal components analysis (PCA) with subsequent varimax rotation. 3.2. Principal components analysis The PCA was based on the covariance matrix of the ERPs, referred to the first data point and with every other sampling point omitted. Figure 2 shows the loadings of the seven-factor solution, accounting for 85.4% of the data variance. The scores of each PCA-derived factor were submitted to the same type of ANOVA as the amplitude measures. The only two factors showing significant interactions involving groups and training blocks were Factor 1 (40.1% of data variance) and Factor 5 (5.6%), which represented most of the ERP
44
W. Sommer
and 5’. Schweinberger / Conditioning
P300
45
W. Sommcr and S. Schweinberger / Conditioning P300
NSW
P300 1 0.5-
Fz
o
:
-0,5-1
1
Pz -0,5-
-0,5I
-I R’Fil
1 d
I 5
,
I 7
Blocks Fig. 3. P300 and NSW factor scores for the random blocks of training, averaged over target and nontarget mental (solid lines) and control group (broken lines).
-1
I RF1
1
I
I
,
3
I 5
I
1 7
Blocks feedback condition (RF) and the seven tones, and superimposed for the experi-
activity during the segment that was used as a criterion for feedback (see Fig. 3). The loadings of Factor 1 increased at about 100 ms after stimulus onset towards a plateau, lasting from around 250 ms until the end of the recording epoch. This factor obviously represents the fronto-central NSW (M (Fz to Pz) = - 0.32, - 0.23, 0.51, F (2/52) = 44.47, p < 0.001). It was more pro-
46
W. Sommer and S. Schweinberger
/ Conditioning
P300
nounced after targets (M = -0.25) than after nontargets (M = 0.21, F (l/26) = 19.3, p < O.OOl>, and showed a three-way interaction between the groups, training blocks and electrode site (F (12/312) = 2.88, p < 0.01) which was even clearer for the linear trend over blocks (F (linear: l/26) = 7.23, p < 0.05). Post-hoc comparisons indicated that the differential linear trends between the groups were restricted to the Fz electrode (F (l/26) = 5.91, p < O.lO>, and absent at Cz and Pz (F < 1). The absence of training effects on the parietal scores indicates that the PCA had been successful in disentangling the P300 and the NSW components. At Fz the NSW of the controls progressively declined over blocks while it was constant or even slightly increased in the experimental subjects (Fig. 3). During the initial random feedback condition there were no group differences either in the P300 or in the NSW (ps > 0.10). ERP activity almost exclusively restricted to the time segment of the P300 component is represented by Factor 5 peaking around 310 ms. For the training blocks the scores showed a parietal scalp distribution (M (Fz to Pz> = -0.05, -0.14, 0.241, F (2/52) = 6.76, p < 0.01, which was more pronounced for targets (M (Fz to Pz) = -0.11, -0.14, and 0.42) than for nontargets (M (Fz to Pz) = 0.01, -0.13, and 0.061, F (2/52) = 19.46, p < 0.001). The interaction between the groups and the training blocks approached significance in the P310 factor scores, (F (6/156) = 2.36, p = 0.06); moreover, the groups differed in the linear trends of the P300 over blocks, (F (linear: l/26) = 4.77, p < 0.05). As in the amplitude measures, the training effects in P300 factor scores were independent of the target/nontarget factor (p > 0.27). Factor 2, which covers the prestimulus time segment, did not show any group-related effects (p > 0.35). This converges with the independence of the P300 amplitude measures from the choice of baseline. To provide an idea of the intersubject variability of the training effects, we calculated for each subject the linear regression slopes of the P300 factor scores for targets over the seven blocks of training. Table 2 shows the slopes of these regression lines, indicating that 9 out of 14 experimentals, but only 3 out of 14 controls, had been able to increase their P310, x2 = 5.25, df = 1, p < 0.05. 3.3. Questionnaire data Six subjects of the experimental group and five of the controls named “waiting for the target, preparing for its arrival” as the strategy primarily used for achieving rewards. This strategy was also experienced as the most successful one by five experimental and three control subjects. The strategy used next most frequently (three subjects in each group) was to “indifferently accept the stimulus”. ‘Both groups had taken much interest in the experi-
W. Sommer Table 2 Slopes of the linear individual subjects SubjectPair
Mean SD
regression
and S. Schweinberger
of target
P300 factor
/ Conditioning
scores
47
P300
over the 7 blocks
of training
for
Group Experimental
Control
-0.13 + 0.03 + 0.01 +0.15 + 0.02 - 0.08 + 0.01 - 0.02 +0.16 + 0.06 + 0.02 - 0.09 + 0.06 -0.18
- 0.30 -0.13 - 0.32 -0.16 + 0.02 - 0.04 0.00 - 0.28 +0.12 - 0.06 -0.12 +0.16 - 0.08 - 0.03
0.00 0.10
- 0.09 0.15
merit, the controls even more so than the experimentals (M = 67.5 and 84.2 scale units, respectively; Kruskal-Wallis test statistic = 5.39, p < 0.05). The association between “behaviour” and feedback had been considered as fairly strong by experimental (M = 65.7, SD = 21.5) as well as control subjects (A4 = 64.6, SD = 14.0) and both groups had been about equally satisfied with their training success (it4 = 68.2, SD = 14.5, and A4 = 58.0, SD = 23.2, respectively).
4. Discussion In line with the hypothesis, a differential trend of the P300 amplitude was observed for the two groups across the seven blocks of training, with a larger amplitude increase in the experimental group than in the control group. This finding demonstrates genuine effects of operant conditioning on P300 amplitude, which could not be found by prior investigators who had not separated the effects of up- and downtraining. One might argue that this result merely represents the restoration of the ERP differences in the random feedback block to be seen in Figs. 1 and 3, which might have vanished somehow during the first few blocks of training. However, these group differences were not significant. Even if one takes
48
W. Sommer and S. Schweinberger / Conditioning
P300
these subtle differences seriously, they do not invalidate the conditioning results. Specifically in the random feedback block the control group had received more rewards (87%) than the experimental group (75%) - because of a programming error, as noticed later. This may have temporarily affected factors such as task involvement, leading to group differences in P300 amplitude. For the training conditions reinforcement ratios had been kept almost identical for the two groups and, initially, also the P300 amplitudes were very similar; it was only during the course of training that group differences in P300 amplitude emerged. Are the training effects specific for P300 or are they related to other ERP components? Firstly, no changes in prestimulus negativity were found which might explain the effects of training on P300. Of the other components, a group-specific trend over blocks was only to be seen in the NSW component overlapping P300 and covering the last part of the recorded epoch. The frontal negativity of the NSW in the control group decreased during the training. However, it is hard to conceive how the reduction of a frontal negativity might diminish a preceding parietal positivity (P300) - in fact, the opposite would have to be expected. While both the NSW and P300 changes appear to be related to the training, their mutual connection, if there is any, seems to be more indirect. It is possible that the decline of the P300 and NSW in the controls are both related to a gradual reduction in attention or motivation (Ruchkin & Sutton, 1983). The lack of stimulus specificity of the training effects for both P300 and NSW is in line with this view. While both components were more pronounced for targets than for nontargets, the training effect was equally present for both stimuli. This may indicate that uptraining did not so much affect stimulus-specific processes but rather helped to maintain a general responsiveness in the experimental group. On the other hand, the questionnaire reports do not suggest a decline in motivation or attention in the control group. The controls had not experienced the rewards as less related to their “behaviour” than the experimentals, nor had they been less satisfied with their training success. Moreover, both groups, and particularly the controls, had claimed much interest in the study. The other evidence available was similarly inconclusive on the issue of how control over P300 might have been achieved. The most often cited strategies were equally present in both groups. In conclusion, the present study shows that operant conditioning is able to increase the P300 component above the level observed in a randomly rewarded control group and therefore demonstrates a genuine uptraining effect. References Donchin, E., & Coles, M.G.H. (1988). Is the P300 component updating? Behavioral and Brain Sciences, 11, 357-374.
a manifestation
of context
W. Sommer and S. Schweinberger / Conditioning P300
49
Miltner, W., Larbig, W., & Braun, C. (1986). Biofeedback of visual evoked potentials. International Journal of Neuroscience, 29, 291-303. Miltner, W., Braun, C., & Larbig, W. (1987). Biofeedback of event related potentials evoked by nociceptive stimulation. Journal of Psychophysiology, 1, 283 (Abstract). Rockstroh, B., Elbert, T., Canavan, A., Lutzenberger. W., & Birbaumer, N. (1989). Slow cortical potentials and behauior. Baltimore: Urban & Schwarzenberg. Roger, M., & Galand, G. (1981). Operant conditioning of visual evoked potentials in man. Psychophysiology, 18, 477-482. Rosenfeld, J.P., Stamm, J., Elbert, Th., Rockstroh, B., Birbaumer, N., & Roger, M. (1984). Biofeedback of event-related potentials. In R. Karrer, J. Cohen & P. Tueting (Eds.), Brain and information: Event-related potentials. Annals of the New York Academy of Sciences, 425, 653-666. Ruchkin, D., & Glaser, E.M. (1978). Simple digital filters for examining CNV and P300 on single trial basis. In D. Otto (Ed.), Multidisciplinary perspectives in euent-related brain potential research (pp. 579-581). Washington DC: US Environmental Protection Agency, EPA-600/977-043. Ruchkin, D.S., & Sutton, S. (1983). Positive slow wave and P300: association and disassociation. In A.W.K. Gaillard & W. Ritter (Eds.), Tutorials in ERP research: Endogenous components (pp. 233-2.50). Amsterdam: North-Holland. Shabsin, H.S. (1982). Feedback controlled increases in P300 visual evoked potential amplitudes and neurological functioning and academic performance in learning disabled children. (Doctoral Dissertation, University of Tennessee, 1982). Dissertation Abstracts International, 43, 3067-B (University Microfilms No. DA8303716). Sommer, W. (1987). Operant conditioning of P300. Psychophysiology, 24, 612 (Abstract). Verleger, R. (1988). Event-related potentials and cognition: A critique of the context updating hypothesis and an alternative interpretation of P3. Behaoioral and Brain Sciences, II, 343-356.
