Increased conflict-induced slowing, but no differences in conflict-induced positive or negative prediction error learning in patients with schizophrenia

Abstract

People with schizophrenia (PSZ) often fail to pursue rewarding activities despite largely intact in-the-moment hedonic experiences. Deficits in effort-based decision making in PSZ may be related to enhanced effects of cost or reduced reward, i.e., through the amplification of negative prediction errors or by dampened positive prediction errors (here, positive and negative prediction errors refer to outcomes that are better or worse than expected respectively). We administered a modified Simon task to people with schizophrenia (PSZ; N = 46) and healthy controls (N = 32). The modification included a reinforcement learning component, where positive and negative prediction errors are dampened or boosted through the use of cognitively-effortful response conflict. EEG was recorded concurrently to investigate potential differences in conflict enhanced mid-frontal theta power between PSZ and controls. We found an enhanced effect of response conflict on response time in people with schizophrenia, but no discernible difference in conflict processing as reflected by the lack of a difference in theta-power enhancement to conflict in mid-frontal regions. Using the reinforcement learning transfer phase of the modified Simon task, PSZ also showed clear deficits in selecting the most rewarding stimulus during the ‘easy’ (most discriminable in terms of value) stimulus contrasts. However, we failed to find a difference between patients and controls in their gain or avoidance learning bias, nor did these biases correlate with negative symptoms. Previous studies had failed to find significant conflict effects on the Simon task likely due to its modest effect size. Our results show that PSZ do indeed possess subtle impairments in response-conflict, suggesting an increase in cognitive effort required for appropriate responding. In addition, while the lack of an overt positive or negative prediction error bias (i.e., a bias towards punishment or reward learning) was unexpected, it is consistent with recent work showing intact estimation of value when the reinforcement learning system is isolated from other contributors to value learning.

1. Introduction

People with schizophrenia (PSZ) often fail to pursue rewarding activities despite largely intact in-the-moment hedonic experiences (Cohen and Minor, 2010; Strauss and Gold, 2012). Interestingly, we (Gold et al., 2013) and others (Fervaha et al., 2013b; Barch et al., 2014; Treadway et al., 2015) have shown that PSZ are less willing to exert effort in pursuit of rewards compared with controls. Altered effort-cost behaviour has been linked with negative symptom expression, highlighting an important role of effort-cost estimation in clinical phenomenology (Gold et al., 2013; Barch et al., 2014). It was argued that the pattern of behaviour seen in patients was more consistent with increasing effort acting on the positive outcome to devalue the reward (Gold et al., 2013). A reward devaluation hypothesis is consistent with a similar body of work where it has been shown that patients with schizophrenia are especially poor at learning from positive relative to negative outcomes. Moreover, this learning profile may be particularly prominent in patients with a high burden of negative symptoms (Gold et al., 2012). However, findings in this regard have been inconsistent (Fervaha et al., 2013a; Barch et al., 2017), with some studies even reporting an opposite association with negative symptoms and reward vs loss sensitivity (Albrecht et al., 2016b). Thus, it is unclear the extent that these acquisition findings are generalisable despite showing some early consistency across tasks and clinical samples (Strauss et al., 2011; Gold et al., 2012). Therefore, there is a need to further identify the parameters under which selective impairments of reward learning or maintenance of learning from punishments or negative prediction errors (PEs; the deviation between expected and obtained outcome). occur and their relationships with negative symptoms (note, both positive and negative PEs can occur in both reward and punishment contexts e.g., if the expectation is a punishment, but no punishment is received, then a positive PE would occur).

Recently, Cavanagh et al. (2014) examined whether cognitive conflict – thought to be represented as a cost that enhances perceived effort – can modify the learned positive or negative reward values of stimuli in a reinforcement learning task. To do so, they modified the basic Simon task (Simon and Rudell, 1967), a classical task manipulating response conflict, to include probabilistic reinforcers and punishers. The Simon task requires participants to respond either congruently or incongruently with stimulus location depending on stimulus characteristics. Response conflict occurs, eliciting increased reaction times, when the task demands an incongruent response, e.g., when a right response is required to a stimulus in the left hemifield. This type of conflict has been proposed to act as a cost signalling increased cognitive effort, leading to avoidance of stimuli and tasks that give rise to it (Botvinick, 2007). Indeed, Cavanagh et al. (2014) found that conflict in the Simon task acted as an implicit cost during learning, reducing experienced reward values and enhancing experienced losses, as assessed by subject preferences in a subsequent transfer phase. Moreover, they found that mid-frontal EEG activity in the theta band, a marker of conflict originating from midcingulate cortex (MCC; Cavanagh and Frank, 2014), was predictive of the degree to which conflict altered experienced reinforcement values. Furthermore, these cost of conflict effects were altered as a function of striatal dopaminergic function (as assessed via pharmacological manipulations and individual differences related to genetic variants; see also Cavanagh et al., 2017 for cost of conflict effects in patients with Parkinson’s). These findings are consistent with evidence that 1) striatal dopamine reflects reward prediction errors, and acts to modify learning about costs and benefits of choices (Montague et al., 1996; Collins and Frank, 2014), and 2) that negative PEs in particular are reflected by activity in MCC and theta signalling therein (Cavanagh et al., 2010, 2012). Thus cognitive conflict may give rise to aversive learning in part by signalling negative PEs. These findings are also consistent with pharmacological studies across species showing that striatal dopamine manipulations bidirectionally affect the willingness to expend effort for reward (Salamone et al., 2007; Wardle et al., 2011; Treadway et al., 2012).

Interestingly, performance in PSZ on the Simon task is assumed to be relatively spared. Several studies have failed to find a statistically significant difference between performance measures in patients and controls, suggesting intact response conflict control (Gastaldo et al., 2002; Behrwind et al., 2011; Sevos et al., 2013; Cieslik et al., 2015; Smid et al., 2016). However, a failure to find statistically significant differences in these previous studies may be due to small sample sizes used to detect a relatively small effect. We recently demonstrated that PSZ patients actually showed reductions in the impact of conflict between Pavlovian and instrumental reward values during learning (Albrecht et al., 2016b), though it is unlikely that this reduction was mediated by enhanced executive control per se. Overall, this literature suggests that schizophrenia is not substantially associated with deficits in response conflict as presented using the Simon task and that this component of executive function may be relatively spared (as opposed to other components of executive function, e.g., Dickinson et al., 2007; Westerhausen et al., 2011). However, it remains unclear whether such conflict could differentially impact reward or cost learning in PSZ; resolving this issue could potentially inform whether deficits in effort-based decision making are related to enhanced cost or reduced reward values.

We administered a modified Simon task to people with schizophrenia and healthy controls. The modification included a reinforcement learning component to investigate individual differences in positive versus negative PE sensitivity by using response conflict that acts as a cognitive cost during learning. We anticipated no difference between patients and controls on the effect of response conflict on reaction times during the task. By contrast, we hypothesised that negative symptoms would correlate during the transfer phase with avoidance of the stimulus associated with conflict (and thus higher cost) due to enhanced negative PE signalling and poor learning from positive PEs. We also anticipated that blink rate (an indirect measure of dopamine function) would correlate with transfer preference of conflict stimuli as we had shown previously (Cavanagh et al., 2014), and that performance on the easy transfer condition would be associated with cognitive ability (due to a heavy reliance on learning and memory). Given previous findings that mid-frontal theta EEG signals related to the impact of conflict on cost and benefit learning, we also recorded EEG concurrently with the task and investigated these effects in PSZ.

2. Methods

2.1. Participants

Forty-eight participants with a diagnosis of schizophrenia or schizoaffective disorder (according to DSM-IV diagnostic criteria; final n = 46 [see below]) and 32 controls were recruited for the experiment. Patients were clinically and pharmacologically stable (no change in drug or dose for at least 4 weeks) outpatients from the Maryland Psychiatric Research Center (MPRC) or other nearby clinics. Most PSZ were being treated with second-generation antipsychotic medications. Our PSZ group has a higher than usual rate of clozapine administration compared to other research groups because of its capacity to closely monitor the adverse effects of clozapine adverse effects. This potentially limits the generalisability of our findings to high clozapine cohorts and partly motivates the supplementary analysis of clozapine. Controls were free from a lifetime diagnosis of a psychotic disorder, current Axis I disorder (including drug dependence), neurological disorder, or cognitively-impairing medical disorder, and had no family history of psychosis in first-degree relatives. Controls were screened with the Structured Clinical Interview for DSM-IV (First et al., 1997).

2.2. Symptom and neuropsychological assessment

Participants were administered the Wechsler Abbreviated Scale of Intelligence (WASI; Wechsler and Hsiao-pin, 2011), the Wechsler Test of Adult Reading (WTAR; Wechsler, 2001) and the MATRICS Consensus Cognitive Battery (MCCB; Nuechterlein et al., 2008). Participants with schizophrenia were further administered the Scale for the Assessment of Negative Symptoms (SANS; Andreasen, 1984), and the Brief Psychiatric Rating Scale (BPRS; Overall and Gorham, 1962).

2.3. Experimental task

The task (see Fig. 1) was derived from Cavanagh et al. (2014), using a modified Simon task (Simon and Rudell, 1967). On each trial, participants were presented with one of four unique simple shape stimuli and instructed to respond by pressing a button (“Left”) or (“Right”) to gain a reward (+1 point). Incorrect responses resulted in no reward (0), and button-presses outside of the response window (850 ms) resulted in punishments (−3). Two stimuli were rewarded deterministically: one stimulus (A) was rewarded at 100% for correct responding regardless of the side it was presented; a second stimulus (D) was never rewarded. The critical stimuli ‘B’ and ‘C’ were both reinforced at 50% rates on average, but their reinforcements were dependent on the experience of conflict: stimulus B was reinforced 100% on congruent trials and 0% on conflict trials, whereas stimulus C was reinforced only on conflict trials. Participants completed four blocks of trials, with each block continuing until the participant gave 20 correct responses for each stimulus (minimum total trials per block = 80; minimum per participant = 320; mean number of trials for PSZ = 368, range = 325, 551; mean number of trials for controls = 347, range = 326, 414) so that the total experienced reward was equivalent across participants. Following each block, a transfer phase ensued where participants were presented with the stimuli from that block in pairs, i.e., ‘AB’, ‘AC’, ‘AD’, ‘BC’, ‘BD’, and ‘CD’. Each pair was presented four times in forward (‘AB’) and reverse order (‘BA’) per block for a total of 32 presentations of each stimulus pair over the course of the experiment. Participants were asked to select the most rewarding stimulus for each pair.

Fig. 1.

Cost of conflict task (adapted from Cavanagh et al. 2014). a) During the training phase, participants were presented with 4 stimuli in each block and possessed different reward probabilities as a function of response requirement. Notably, the conflict stimuli ‘B’ and ‘C’ were reinforced only on congruent and conflict stimuli respectively. b) Illustrates the flow of the presentation to participants during the training phase. c) During the testing phase, the stimuli from the preceding training block were presented to the participant in pairs and d) they were asked to select the most rewarding stimulus. e) The role of conflict on positive prediction error devaluation versus negative prediction error enhancement. A bias for B > C should occur in participants who learn better from reward because conflict on the reward obtained from C diminishes its relative positive PE (usually generated by rewards) value. By contrast, a bias for C > B should occur in participants who learn better from negative PEs (usually generated by punishments) because conflict on the negative PE obtained from B enhances its negative PE value. PE = prediction error +ve = positive −ve = negative.

According to our framework, conflict reduces the value of rewards and enhances the value of losses. Thus rewards experienced during conflict trials (i.e., for stimulus C) will be reduced compared to those experienced for congruent trials (i.e., for stimulus B). Based on rewards, then, subjects should prefer B over C in a direct choice. If instead conflict primarily amplifies the aversive impact of a loss, then losses experienced during conflict trials (i.e., for stimulus B) will be more aversive than those experienced during congruent trials (i.e., for stimulus C). Thus, subjects with a disposition to make choices based on prospective reward values should prefer B over C, but those disposed to avoid losses should instead prefer C. For a detailed illustration of this effect, see Fig. 1 and Cavanagh et al. (2014).

At the completion of the task, points were converted into money, with each accumulated point worth 5 cents. Total winnings for each subject ranged between $2.00 and $8.00. Participants were informed about the monetary reward before the study. The stimulus presentation sequence and timings were as follows: the pre-stimulus period consisted of a cross hair presented for 1000 ms, the stimulus was then presented until the participant made a response (maximum response time allowed = 850 ms), a cross hair was again presented for 170 ms (with a ± 10 ms jitter), finally feedback was presented for 1000 ms (only for the training phase, no feedback was presented during the transfer phase).

2.4. EEG recording and data processing

EEG was recorded from a 64 channel BrainVision system. Data were recorded unreferenced with the ground at AFz using a sampling rate of 1000 Hz with 500 Hz hardware filters. Data were imported into EEGLAB (Delorme and Makeig, 2004), and pre-processed using the prepPipeline package (Bigdely-Shamlo et al., 2015; high pass filtered @ 1 Hz, 60 Hz line noise removal, referenced to robust average reference, and electrode interpolation). A low pass filter was then applied at 50 Hz. Data were epoched from −1500 to 3000 ms around stimulus event codes ensuring adequate capture of feedback events and sufficient resolution of theta (4500 ms total analysed epoch allowing for a minimum of 18 cycles at 4 Hz). Epochs with large potential fluctuations were removed using EEGLAB’s pop_autorej procedure (starting probability was set at 5 SD and the maximum % of epochs to reject per iteration was set at 5). Data were then down-sampled to 500 Hz. The first pass cleaned EEG data underwent independent components analysis (ICA) using the AMICA algorithm (Palmer et al., 2017) before further artifact rejection was applied based on detection of significant linear trends over the epoch in component space or abnormal component signal strength in the 0–3 Hz range and the 20–50 Hz range (Delorme et al., 2007). A second ICA was applied to the second pass cleaned data, which was used to subtract activity associated with eye blinks and eye movements. Epochs with voltages greater than ± 90 μV were removed. The mean number of epochs per group and condition were as follows: PSZ congruent response trials = 121; PSZ conflict response trials = 120; PSZ positive PE (reward) feedback trials = 121; PSZ negative PE feedback trials = 120; controls congruent response trials = 109; controls conflict response trials = 110; controls positive PE feedback trials = 111; controls negative PE feedback trials = 108.

Current source density (CSD) was calculated using the CSD toolbox (Kayser and Tenke, 2006) with spline flexibility equal to 4. ERPs were baseline corrected to a 100 ms baseline (see Supplementary material). Time-frequency analysis used the functions from EEGLAB applied at logarithmically spaced frequencies. Time-frequency power was baseline corrected using the average of the power response from −300 to −200 ms pre-stimulus onset. Stimulus-locked theta power was averaged from −200 to 50 ms surrounding the response. Feedback-locked theta power was averaged from 200 to 500 ms post-feedback onset.

Data for this project is available from: http://predict.cs.unm.edu/downloads.php and http://dx.doi.org/10.5281/zenodo.889204.

2.5. Statistical analysis

One control and four patients obtained overall accuracies less than 0.6 on the easy conflict cost transfer condition. Due to the small proportion of participants performing poorer than the cut-off, the use of robust statistical methods, and the behaviour likely being representative of the population, these participants were kept for all analyses. Two patients did not generate usable EEG data due to issues during the recording session.

Bayesian repeated measures ANOVA-style models and Bayesian t-tests were used to analyse the behavioural and ERP data (Kruschke, 2011, 2013) using the ‘BEST’ package (Kruschke and Meredith, 2014) and models derived from Kruschke (Kruschke, 2011) in R. Details on the models used are presented in (Albrecht et al., 2016a, 2016b). The advantages of these models include: incorporation of t-distribution to render the analysis robust to outliers and some distortions of the normal distribution; model unequal variances; and shrinkage to improve estimation and control for multiple comparisons. Given the highly skewed nature of performance during the ‘easy’ transfer condition, analysis of this data was conducted by fitting beta distributions to the data and the expected value of the distribution (alpha/(alpha + beta)) was compared between the two groups to assess differences in accuracy. This was additionally supplemented with a Bayesian t-test and a non-parametric Wilcoxon test to ensure consistency in results.

Between group differences were obtained by contrasting the posterior for PSZ and controls. Credible intervals were obtained using the highest density interval (HDI) that covered 95% of the posterior distribution.

Correlation analysis used standard non-parametric Spearman rank correlations. Given our hypotheses, relationships between negative symptoms and conflict cost were planned and therefore not corrected for multiple comparisons. Similarly, no corrections for multiple comparisons were applied to correlations examining associations with cognitive variables, as there is significant covariance among neuropsychological measures (making it difficult to identify an independent comparison). It is also to be expected that some performance directly relevant to simple learned stimulus values should be associated with memory ability (and thus cognitive ability more generally), e.g., simple transfer accuracy that asks participants to recollect which stimulus possessed greater value between stimuli with unambiguous differences in value. We report the full spectrum of correlations for completeness.

3. Results

3.1. Demographics and baseline theta

Demographic characteristics of the sample are presented in Table 1. Participants were well matched across age, sex, race, and parental education. Patients were found to have lower education and cognitive ability compared with controls. Blink rates were assessed using the electro-oculogram (EOG) channels from the EEG during two minutes of resting state preceding the experiment because of their association with dopamine tone. Patients were found to have a higher blink rate than controls. There was no difference in blink rate by clozapine status.

Table 1.

Demographics, symptoms, and neuropsychological performance for the participants with schizophrenia (PSZ) and controls.

		Controls		PSZ
		Mean	SD	Mean	SD	t/X stat	p
Biological	Age	37.1	10.4	38.0	8.8	0.4	0.70
	Sex (F | M)	11 | 21		16 | 30		0.0	1.00
	Race (Black | White | Other)	(12 | 17 | 3)		(18 | 26 | 2)		0.8	0.67
Education	Level of Education Achieved (yrs)	15.1	2.2	13.2	2.5	3.5	0.001
	Mother’s Education Level	13.8	2.2	14.1	3.0	0.6	0.52
	Father’s Education Level	14.2	3.7	13.9	3.7	0.4	0.72
Neuropsych	Estimated IQ (WASI)	110.8	13.4	96.3	15.8	4.3	< 0.001
	Processing Speed	52.8	11.9	41.0	12.8	4.2	< 0.001
	Attention/Vigilance	52.3	11.4	41.8	12.4	3.8	< 0.001
	Verbal Working Memory	51.5	11.5	41.6	10.9	3.8	< 0.001
	Verbal Learning	50.3	8.8	38.0	8.0	6.3	< 0.001
	Visual Learning	45.7	10.8	36.9	11.0	3.5	0.001
	Reasoning	49.3	9.8	44.2	11.0	2.2	0.033
	Social Cognition	54.1	7.6	43.8	11.5	4.8	< 0.001
	MCT Overall	51.0	11.0	35.2	13.1	5.7	< 0.001
BPRS	Affective			5.2	2.9
	Negative			6.4	2.4
	Reality Distortion			7.6	4.6
	Disorganisation			3.0	0.2
	Total			32.4	8.7
SANS	Anhedonia			8.8	4.1
	Role Functioning			8.7	4.6
	Affective Blunting			10.4	5.5
	Alogia			1.1	1.7
	Total			29.0	11.7
Medication	Haloperidol Equivalents			8.6	5.9
	Clozapine (No | Yes)			(22 | 24)
Blink	Blink Rate (Total over 2 min)	32.7	21.3	45.2	27.8	2.3	0.027

We had previously shown increased baseline theta power, and subsequently reduced event-related theta power, in patients on clozapine. In line with this, patients on clozapine (n = 24) had the greatest baseline theta power compared to patients on other antipsychotics (n = 22) (clozapine – other contrast = 0.42, 95% HDI = 0.21, 0.65) and controls (clozapine – control contrast = 0.48, 95% HDI = 0.27, 0.69). Controls and patients not on clozapine did not show baseline theta differences (other – controls contrast = 0.06, 95% HDI = −0.12, 0.24). Despite these differences in baseline power, clozapine did not substantially influence the findings reported below (for EEG or behaviour) and are not reported on further.

3.2. Behavioural performance

3.2.1. Training performance and the effect of conflict on response time

The PSZ group made numerically more errors during the training phase, which appeared to be driven by three particularly poor performers (Fig. 2 left; effect size contrast = 0.54, 95% HDI = − 0.06, 1.15). Fig. 2 (middle) illustrates the response time (RT) profile to congruent and conflict stimuli across participant groups. Both groups showed a credible slowing to conflict stimuli relative to congruent stimuli, demonstrating the classic Simon effect. The PSZ group were slower overall and were more affected by conflict compared to controls, as indicated by an interaction between group and conflict, showing a 10.8 ms increase in conflict-induced RT compared to controls (Fig. 2 right; interaction effect = 10.8 ms, 95% HDI = 0.76, 21.4; effect size = − 0.52, 95% HDI = − 1.02, − 0.04). We supplemented the between group contrast in conflict induced slowing analysis by evaluating proportional difference scores (i.e., [RT_conflict – RT_congruent]/RT_congruent). This analysis indicated a slightly weaker effect size compared to the direct difference score analysis, but the overall effect was in the same direction (effect size = −0.43, 95% HDI = −0.91, 0.06). Both groups exhibited significant post-error slowing (mean RT following an error – mean RT for HC = 51.5 ms, 95% HDI = 31.0, 73.5; PSZ = 38.9 ms, 95% HDI = 23.0, 55.1); however, there was no difference between the groups on the degree of post-error slowing (contrast = 12.6 ms, 95% HDI = − 13.8, 39.4).

Fig. 2.

Training phase behavioural summary and conflict effect (Blue – controls; Red – Participants with schizophrenia). a) Number of errors made during the training phase summed over all blocks. b) Training phase reaction times across congruent and conflict conditions with between subject 95% HDIs. There were main effects of conflict (note HDIs are between subject HDIs) and group, with response conflict delaying RTs in both groups and patients making slower responses overall. Patients were also more influenced by response conflict, with an interaction between group and conflict indicating slower responses to conflict relative to congruent stimuli in patients.

3.2.2. Transfer performance and the effect of conflict on value

Fig. 3 (left) illustrates performance accuracy for the ‘easy’ transfer condition that contrasts participant preferences for stimuli that were always rewarded vs. those that were rarely rewarded. While both groups achieved mean accuracies > 80%, patients were less able to select the stimulus with higher value (difference in accuracy: 0.078, 95% HDI = 0.014, 0.14).

Fig. 3.

Transfer behaviour for the easy comparison of reward values is presented on the left and the effect of conflict on reward is presented on the right. Left) Patient and control preferences during the transfer phase for the highly rewarded stimulus ‘A’ compared to the rarely rewarded stimulus ‘D’. Patients were less able to select the stimulus with a higher expected value. Right) Participant preferences for stimuli that have only been rewarded during congruent ‘B’ or only rewarded during conflict ‘C’. As a group, neither controls or patients showed a preference for B > C or C > B. (Note: the 95% HDIs in left figure are from a binomial model due to the significant ceiling performance while the right 95% HDIs are from a robust Bayesian t-test).

Fig. 3 (right) illustrates participants’ stimulus preference for the ‘conflict cost’ condition that compares the influence of conflict on reward omission or reward. As noted above and in Cavanagh et al. (2014), preference for B over C or C over B – stimuli that had both been reinforced at 50% but differentially for congruent or conflict trials – is indicative of whether conflict primarily interacts with positive or negative prediction errors. Contrary to expectations, there was no difference between patients and controls in their preference for C > B or B > C (preference contrast = 0.01, 95% HDI = − 0.08, 0.11; effect size = 0.071, 95% HDI = − 0.4, 0.55).

3.3. EEG

3.3.1. Response-locked activity

Previously, we found that response conflict resulted in enhanced theta (4–7 Hz) power around the time of the response. Fig. 4 shows that this effect was replicated here in patients and controls who both showed increased response-locked theta power immediately preceding the motor response to conflict stimuli (conflict – congruent contrast = 0.67, 95% HDI = 0.39, 0.95), albeit with patients showing less response-locked theta power overall (during both conflict and congruent conditions; patient – control contrast = −1.77, 95% HDI = −2.48, −1.11). There was no group by response conflict interaction (interaction contrast = − 0.33, 95% HDI = − 0.92, 0.26).

Fig. 4.

Left) Response-locked time-frequency decomposition at electrode FCz (CSD transformed) for response congruent and response conflict conditions. Right) Time-course of the response-locked theta power (theta - average power between 4 and 7 Hz) obtained from electrode location FCz (CSD transformed). Conflict clearly enhanced theta power around the time of response in both groups and patients had lower response-locked theta compared to controls. However, there was no interaction between group and response conflict that would suggest differential conflict processing as reflected by changes in mid-frontal theta power.

3.3.2. Feedback-locked activity

Fig. 5 illustrates the time-frequency power response to rewards (gains of 1 point) and reward omissions (no gain). Consistent with our previous reports (Cavanagh et al., 2010, 2012) there was an enhancement of theta power to reward omission stimuli relative to reward stimuli across the groups (reward omission vs reward contrast = 0.17, 95% HDI = 0.01, 0.32). Again, patients had lower overall theta compared to controls (patient vs control contrast = − 0.52, 95% HDI = − 0.91, − 0.10). There was no group by feedback type interaction (interaction contrast = − 0.10, 95% HDI = − 0.38, 0.14).

Fig. 5.

Left) Feedback-locked time-frequency decomposition at electrode FCz (CSD transformed) for positive PE (reward) and negative PE (punishment) stimuli. Right) Time-course of the feedback-locked CSD theta power (theta - average power between 4 and 7 Hz) at FCz. In both groups, theta power was increased 200–500 ms post-feedback during negative PE trials and again patients showed reduced theta power over both conditions. There was no credible interaction between group and feedback.

Next we examined the effect of conflict on feedback-evoked theta power. Previously, Cavanagh et al. (2014) reported a significant modulatory influence of conflict on the relationship between feedback-locked theta power and future bias in learning from feedback (see Cavanagh et al., 2014, Fig. 4). Principally, we evaluated the difference in the association between feedback and theta power at Cz as a function of conflict condition, and similarly the difference in association between feedback and theta power at FCz as a function of learning bias. These electrodes were chosen as they were shown to possess the greatest association in Cavanagh et al. (2014). However, we failed to find the expected associations in the current sample; the absolute value of all ρs was < 0.2, p > 0.25. We also looked more posteriorly at CPz and Pz, but also failed to find an association.

3.4. Correlation analyses

Tables 2 and 3 present the Spearman correlation coefficients relating behavioural measures, EEG-theta power, symptom variables, and neuropsychological performance for patients and controls respectively (expanded tables with the full set of neuropsychological assessments are presented in the Supplementary material: Supp Tables 1 and 2). A cluster of correlations were found between accuracy on the easy transfer condition and general cognitive ability in both patients and controls.

Table 2.

Spearman correlation coefficients for the schizophrenia sample.

		Training RT Conf-Cong	Transfer Easy	Conflict Cost B>C	Resp Theta Conf-Cong	Fdbk Theta Pun – Rew
BPRS	Anxiety/Depression	−0.04	0.05	−0.11	−0.05	0.14
	Negative	0.09	0.02	0.08	−0.03	0.17
	Reality Distortion	0.04	−0.15	−0.03	−0.45	−0.14
	Disorganisation	−0.33	0.09	0.12	0.12	0.08
	Total	0.08	−0.03	−0.16	−0.24	0.08
SANS	Anhedonia	0.08	0.06	−0.07	0.01	0.15
	Role Functioning	−0.16	−0.06	−0.01	0.10	0.17
	Affective Blunting	−0.02	−0.01	0.14	−0.10	0.26
	Alogia	−0.27	−0.02	0.12	−0.05	0.18
	Total	−0.11	−0.01	0.05	0.02	0.29
Medication	Haloperidol Equivalents	−0.01	0.10	0.04	−0.27	0.06
Neuropsych	Estimated IQ	−0.10	0.46	0.09	0.32	−0.05
Blink	Blink Rate	−0.02	−0.01	0.21	−0.22	−0.06
Response Theta	Conflict – Congruent	−0.07	0.31	0.05	-	0.05
Fdbk Theta	Punishment – Reward	0.10	0.11	0.05	-	-

Red = p < 0.05*.

Bold Red = p < 0.01*.

Table 3.

Spearman correlation coefficients for the control sample.

		Training RT Conf-Cong	Transfer Easy	Preference Direct CC	Resp Theta Conf-Cong	Fdbk Theta Pun – Rew
Neuropsych	Estimated IQ	−0.25	0.36	0.27	0.03	−0.33
Blink	Rate	0.19	−0.11	−0.28	0.32	−0.05
Response Theta	Conflict – Congruent	0.14	−0.11	−0.12	-	−0.18
Fdbk Theta	Punishment – Reward	0.46	−0.03	−0.20	-	-

Red = p < 0.05.

Bold Red = p < 0.01.

4. Discussion

We used a modified Simon task that included rewards and reward omission balanced over cognitively-effortful (conflict) and cognitively-easy (congruent) responding to assess whether patients with schizophrenia (PSZ), particularly those with relatively-severe negative symptoms, show a bias towards avoidance learning. Interestingly, PSZ and controls displayed a very similar effect of conflict on stimulus preference during transfer, suggesting no differences in reward processing due to conflict in PSZ. This is consistent with recent work showing that model-free/basal ganglia-driven reinforcement learning (more habitual learning system based primarily on past outcomes) is, by and large, intact in PSZ, especially when assessed through a transfer phase (Collins et al., 2014, 2017). By contrast, impairments in prefrontal cortex-dependent reinforcement learning processes, reliant on resource-limited working memory systems, appear to be severe in PSZ. Surprisingly, patients demonstrated modestly greater RT slowing to conflict stimuli compared to controls, indicating that spatial response conflict is affected in PSZ. This was unexpected given previous literature that had repeatedly shown lack of a significant effect in PSZ. However, despite an increase in behavioural response conflict and reduced overall theta power in PSZ, theta enhancement to conflict was of a similar magnitude as controls.

4.1. Behavioural effects of conflict

Contrary to expectations, PSZ showed greater conflict-evoked RT slowing than controls. Patient RTs were slowed by ca. 10 ms with a corresponding standardised effect size of ca. 0.5. Previous studies using the Simon effect in studies with schizophrenia patients have failed to detect this effect (Gastaldo et al., 2002; Behrwind et al., 2011; Sevos et al., 2013; Cieslik et al., 2015; Smid et al., 2016). The RT slowing between patients and controls as a function of conflict in these previous studies ranged between 0 and 21 ms, with a relatively even spread of effects in between. Our find of a 10 ms exaggerated delay in patients falls exactly in the middle of the distribution of conflict delay times. The combined evidence indicates that there is, in fact, a response conflict effect in patients with schizophrenia that has not been reliably detected – likely because of its modest effect size. Given that there does seem to be a modest conflict effect in patients, this argues against suggestions of spared spatial response conflict control in schizophrenia. However, this effect size is in the bottom range of cognitive effect sizes seen in the literature (e.g., compared to digit symbol coding at ca. 1.6, episodic memory/learning at ca. 1.25, and category fluency at ca. 1.4), suggesting that a deficit in spatial response conflict control is not likely a core neuropsychological feature of schizophrenia. Nevertheless, subtle deficits in response conflict control may still play a role in schizophrenia, with the possibility of a subgroup of patients who may show more severe deficits and for whom this might be a significant problem. However, we did not find evidence in the current study of a strong association between symptoms or neuropsychological performance and response conflict effects (see Table 2 and Supp Tables 1 and 2). The current task is conceptually similar in many ways to the classic Stroop task. In particular, both the Stroop and Simon task require responding to one stimulus dimension, whilst ignoring another dimension that biases a prepotent response. In the Simon task, responses contingent to the same spatial location can be considered a learned ‘rule’, where a considerable history of spatially-dependent action prioritises same side responding. When the current rule demands top-down over-riding of the prepotent response, the resolution of this competition is responsible for the slowing of responses (when spatial location and colour are incompatible). A review of attention deficits in schizophrenia finds consistent evidence for impairments in rule selection and control of selection in PSZ, whereas input selection and the implementation of selection are relatively unimpaired (Luck and Gold, 2008). Increased slowing during the Simon task is consistent with the notion of impaired top-down rule selection primarily associated with prefrontal cortical function (Posner and Petersen, 1990). Indeed, the strong mid-frontal theta response to conflict in the present study and in our previous study (Cavanagh et al., 2014) supports a role of prefrontal involvement. The effect sizes seen in the Stroop task are generally higher than those seen in the present task (ca. 0.99; Dickinson et al., 2007). However, Stroop effect sizes are considerably reduced when going from traditional pen and paper tasks to a task structure that optimised for event-related brain activation (Westerhausen et al., 2011), perhaps also explaining the divergence of effects between the Simon task and the Stroop in schizophrenia studies.

4.2. Transfer performance – conflict-cost effect

We observed no between-group difference in the effect of conflict condition on transfer behaviour and no association with negative symptoms. We had hypothesised that greater negative symptoms would show a preference for C > B responding relative to controls, indicating relatively enhanced aversive value of conflict as opposed to diminished reward. We and others had previously found greater reluctance to engage in high effort choices to obtain reward, and greater deficits in positive relative to negative prediction error learning, in PSZ, particularly in subgroups manifesting more severe negative symptoms. There are several possible explanations for the failure of the present study to find such a difference. First, the presence of a negative PE-driven learning bias is present in only a subgroup of PSZ making it less generalisable than initially considered. Furthermore, the sampling of the schizophrenia population may have yielded a subset of patients in the current study (who are amenable to EEG studies) who do not show the anticipated negative PE biases, suggesting alternative reasons for the expression of negative symptoms. Second, dopamine tone has been shown to be an important modulator of learning biases in this task. Previously, we had shown that transfer performance bias is sensitive to dopamine tone in a way that suggests greater preference for B > C (higher scores in Fig. 3b) is associated with higher dopamine tone and reward sensitivity. Conversely, C > B preference was associated with lower dopamine tone and higher negative prediction error sensitivity. It would be anticipated that controls administered antipsychotics might be pushed towards a C > B (negative prediction error) preference. By contrast, a disorder associated with high dopamine tone, in the absence of antipsychotic treatment, may be pushed into a B > C (reward) preference. It is unclear how combining dopamine hyperactivity with medication, as in the current study, would affect the relative preference for positive versus negative prediction errors. More heavily medicated patients may end up with a C > B negative prediction error profile; however, they may be more heavily medicated because of their innate high pre-synaptic dopamine capacity. Moreover, we found increased blink rates in patients, consistent with a number of previous studies (Karson et al., 1981; Karson, 1983; Mackert et al., 1991), suggesting higher overall dopamine tone despite stable D₂ blockade (n.b. blink rate did not correlate with antipsychotic equivalents). Despite all of this, there was no difference between patients and controls in C > B | B > C preference and there was no correlation with blink rate on bias in either group. Finally, the effect may depend strongly on task parameters. For example, the probabilistic stimulus selection task that found relative reward/punishment differences associated with negative symptoms used incremental updates to expected value to guide behaviour. This can be compared with the current experiment that used an explicit unchanging rule, making learning stimulus-response requirements independent of a history of positive or negative prediction errors.

Finally, if we are consistently unable to find a negative PE (or punishment) learning bias across several experimental designs, we will have to re-consider the role of positive vs negative PE learning in the formation of negative symptoms. We have shown in two different experiments a failure to conceptually replicate the initial positive vs negative PE bias findings (Albrecht et al., 2016b; Barch et al., 2017), in addition to the current experiment. This then would prompt the search for other mechanisms that more strongly predict negative symptom formation. For example, in our initial modelling experiment (Gold et al., 2012), negative symptoms were also strongly associated with a modelled reduction in relative contributions from prefrontally oriented Q-learning system (reflecting a more versatile, direct action-value system) relative to the more basal-ganglia oriented actor-critic system (where value is updated separately to action-selection). This suggests that higher-order more adaptive systems of stored value are responsible for negative symptoms, potentially little impact of the sign of the PE.

4.3. EEG theta power

We found an overall reduction in response-evoked theta in the patient sample consistent with previous schizophrenia studies. This is also consistent with findings indicating that PSZ patients show reduced BOLD activation in brain regions associated with cognitive control during reversal learning (Culbreth et al., 2016). However, despite lower theta power overall, patients were able to appropriately modulate theta power in response to conflict and there was no interaction indicating differential conflict processing via theta modulation in patients compared to controls. The lack of a between-group difference in the effect of conflict on theta power is interesting in the context of greater RT slowing to conflict in patients. This highlights the possibility of a downstream mechanism responsible for translating ACC signalling of conflict into a rectification of prepotent action being delayed in PSZ. In partial support, RT inversely correlated with response-locked theta power for both conditions in the patient sample, indicating a link between theta power and RT; however, there was no correlation between change in RT to conflict and change in theta power to conflict that might more directly assesses the relationship between conflict theta and RT slowing. We scanned other frequency bands around the time of the response that might give more insight into RT slowing in response to conflict but did not find any significant association. In contrast, it is commonly shown in Stroop fMRI tasks no difference in interference, but yet present significant differences in ACC activation in response to conflict (e.g., Kerns et al., 2005).

Consistent with Cavanagh et al. (2014), we found an enhancement of theta power to reward omission in both controls and patients when processing feedback. However, despite theta power reductions in patients, we failed to find a significant group by outcome valence interaction. We and others have shown intact early brain responses in patients to worse-than-expected outcomes (as measured by EEG; Morris et al., 2008, 2011; Horan et al., 2012); however, analysis of later processing stages in our recent study indicated ERP differences to explicit punishers (i.e., − 1, not 0) and rewards were blunted in patients (Albrecht et al., 2016b). This was corroborated by model based EEG analysis, where the correlation between PE and EEG voltage was consistent between patients and controls at early processing stages, before deviating around the time of the P3a. Differences between the present study (where a ‘0’ was used) and our previous study (where a ‘ −1’ was used) may be due to the nature of the negative prediction error form used in each study. The more explicit loss likely enhances the brain’s negative prediction error response that would be necessary to show relative deficits between patients and controls during dynamic allocation of resources to punishing stimuli.

4.4. Conclusion

We found an enhanced effect of response conflict on RT in people with schizophrenia, which may be interpreted as an effect of an increase in required cognitive effort. This finding highlights an impairment in spatial response conflict that was previously elusive due to the modest effect size. Despite the increased behavioural conflict effect, there was no discernible difference in conflict processing as assessed by the EEG-theta response to conflict in the mid-frontal regions. There was also no difference between patients and controls in gain or avoidance learning biases assessed during transfer. While this was unexpected from the perspective of a framework accounting for impairments in reward-learning, in the presence of intact avoidance learning previously seen in schizophrenia, it is consistent with our previous work showing intact estimation of value when the RL-system is isolated from other contributors to value learning.