Susceptibility to interference between Pavlovian and instrumental control is associated with early hazardous alcohol use

Pavlovian‐to‐instrumental transfer (PIT) tasks examine the influence of Pavlovian stimuli on ongoing instrumental behaviour. Previous studies reported associations between a strong PIT effect, high‐risk drinking and alcohol use disorder. This study investigated whether susceptibility to interference between Pavlovian and instrumental control is linked to risky alcohol use in a community sample of 18‐year‐old male adults. Participants (N = 191) were instructed to ‘collect good shells’ and ‘leave bad shells’ during the presentation of appetitive (monetary reward), aversive (monetary loss) or neutral Pavlovian stimuli. We compared instrumental error rates (ER) and functional magnetic resonance imaging (fMRI) brain responses between the congruent and incongruent conditions, as well as among high‐risk and low‐risk drinking groups. On average, individuals showed a substantial PIT effect, that is, increased ER when Pavlovian cues and instrumental stimuli were in conflict compared with congruent trials. Neural PIT correlates were found in the ventral striatum and the dorsomedial and lateral prefrontal cortices (lPFC). Importantly, high‐risk drinking was associated with a stronger behavioural PIT effect, a decreased lPFC response and an increased neural response in the ventral striatum on the trend level. Moreover, high‐risk drinkers showed weaker connectivity from the ventral striatum to the lPFC during incongruent trials. Our study links interference during PIT to drinking behaviour in healthy, young adults. High‐risk drinkers showed higher susceptibility to Pavlovian cues, especially when they conflicted with instrumental behaviour, indicating lower interference control abilities. Increased activity in the ventral striatum (bottom‐up), decreased lPFC response (top‐down), and their altered interplay may contribute to poor interference control in the high‐risk drinkers.


| INTRODUCTION
To behave efficiently in one's daily life and to adapt one's actions to a dynamic environment, a response selection system is frequently engaged. Critical control components involved when making such choices include Pavlovian and instrumental control. Through Pavlovian conditioning, inborn and hard-wired responses (e.g., approach or avoidance) to biologically potent (unconditioned) stimuli can be associated with neutral stimuli. Thereafter, such conditioned responses to Pavlovian cues are independent of their outcomes. Conversely, instrumental behaviour, more specifically, goal-directed instrumental behaviour, is controlled by the contingencies between actions and outcomes. Pavlovian cues can influence ongoing instrumental behaviour, even though the responses to the Pavlovian cues were acquired separately from the instrumental responses-this process is called Pavlovianto-instrumental transfer (PIT). To elaborate, a food's enticing scent (Pavlovian) may encourage people to partake in eating behaviour (Instrumental), whereas an unpleasant scent may hinder eating behaviour. In a typical human PIT task, 1,2 participants need to perform learned instrumental responses (press a button for approach or avoidance) in the presence of previously and independently trained Pavlovian cues (appetitive or aversive).
Most previous human PIT studies investigated how Pavlovian cues influence instrumental approach behaviour. Accordingly, appetitive Pavlovian cues were found to promote instrumental approach responses compared with the neutral cues, [3][4][5][6][7][8][9] whereas aversive Pavlovian cues were found to reduce instrumental approach behaviour. 10,11 Additionally, some studies have examined PIT effects in the avoidance context by rewarding successful instrumental avoidance behaviour, in which aversive Pavlovian cues were shown to promote instrumental avoidance behaviours. [12][13][14] Moreover, in an orthogonal experimental design with the appetitive-aversive Pavlovian axis and the approach-avoidance instrumental axis, instrumental behaviour was impaired by incongruent Pavlovian cues (instrumental approach behaviour by aversive Pavlovian cues or instrumental avoidance behaviour by appetitive Pavlovian cues) but was promoted by congruent Pavlovian cues. 10,15 Freeman, et al. 16 used a go-no-go/PIT task, which resembles a classical go-no-go task. In this task, participants learned to respond to one stimulus in the go trials while withholding their responses to another stimulus in no-go trials. The authors modified the proportion of no-go trials where appetitive Pavlovian cues were presented. It was then found that when the proportion of incongruent no-go trials out of all no-go trials was higher, the provocation of the appetitive cues on instrumental approach behaviour (go trials) in the subsequent trials was reduced. Additionally, in one EEG study, Cavanagh et al. 17 used another variant of a go-no-go task to investigate how Pavlovian biases influence instrumental learning during the conflict between both systems. It was found that midfrontal theta power, sensitive to conflict and the following adaptive control, was associated with the ability to overcome Pavlovian biases when they interfered with the instrumental behaviour. Taken together, these four studies imply that cognitive control is to be allocated to overcome the conflict between Pavlovian and instrumental control.
Linked to alcohol drinking behaviour, previous studies from our group have found associations between the stronger motivational effect of Pavlovian cues on instrumental behaviour and alcohol dependence, [18][19][20] as well as high-risk drinking during young adulthood. 21 In addition to the enhanced behavioural effect, the neural correlates of the motivational PIT effect in the nucleus accumbens 19,20 and the amygdala 21 were also associated with alcohol dependence and high-risk drinking during young adulthood, respectively. Notably, when whether the Pavlovian cue interferes with the instrumental behaviour was taken into account, alcoholdependent patients committed more errors compared with healthy controls when Pavlovian stimuli and instrumental responses were in conflict, especially when participants needed to inhibit instrumental approach responses during the presence of appetitive Pavlovian cues 15 ; this behavioural impairment was also stronger for future relapsers. 22 As of yet, whether this interference effect along with its neural correlates was associated with high-risk drinking during young adulthood is not clear.
We thus investigated interference control during a PIT task in a group of healthy, young men at age 18, who were drinking occasionally but did not fulfil the criteria for Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) alcohol dependence. The rationale behind this is that social drinking behaviour is influenced by numerous environmental cues during social occasions, which reflects the PIT task in the experimental settings to some extent. A reduction in the ability to allocate cognitive resources in order to control the response to cues that look tempting but violate the long-term goals may contribute to hazardous drinking development. From this perspective, we assumed that the ability to allocate interference control when the Pavlovian cues conflict with the instrumental behaviour, along with its associated neural responses, could be potential (bio)markers of hazardous drinking behaviour during early adulthood. More specifically, on the behavioural level, it was hypothesised that error rates (ERs) would increase in the incongruent condition, that is, when Pavlovian cues and instrumental stimuli are incongruent, as compared with the congruent condition. Importantly, individuals with higher levels of risk in drinking should show more susceptibility to this effect, that is, show lower interference control.
On the neural level, previous literature has found neural correlates of motivational effects of Pavlovian cues in the amygdala, 11,[23][24][25] the ventral striatum (VS), 11,23,25 and the dorsal striatum. 12,26 Accordingly, the VS and amygdala were expected to show responses during the PIT task. Importantly, referring to the meta-analysis of tasks that require different dimensions of inhibitory or interference control, 27 we also hypothesised that conflict between Pavlovian cues and required instrumental behaviour would elicit responses in cognitive control areas-the lateral prefrontal cortex (lPFC) and the dorsomedial prefrontal cortex (dmPFC). Further, low-risk drinkers were hypothesised to allocate more top-down interference control as compared with high-risk drinkers. If this were to be the case, we would expect the effective connectivity between the aforementioned brain regions to be altered in the high-risk drinkers, which we would explore with dynamic causal models.

| Participants and general procedure
Invitation letters were first sent to 1937 males at age 18 who were randomly sampled from the local registration offices in Dresden and Berlin, Germany. At the baseline of the longitudinal study, only males were recruited because of the higher prevalence of risky drinking behaviour. After screening 445 respondents, those with the inclusion criteria of right-handedness, no history of major mental disorders including substance dependence (except for nicotine dependence), eligibility for magnetic resonance imaging (MRI) and having had at least two drinking occasions in the past 3 months were further invited. Of those who met the inclusion criteria, 201 participants completed the behavioural and MRI assessment. After excluding participants with incomplete behavioural data because of technical issues, 191 participants were included for the final analysis.
Participants went through the experimental procedure with two appointments. During the first appointment, participants finished the Munich Composite International Diagnostic Interview (M-CIDI 28,29 ) according to the DSM-IV 30 , along with cognitive ability assessment (details in Supporting Information S2). The risk status of our subjects was defined according to their binge drinking behaviour based on World Health Organization (WHO) standards 31 : as recommended, an average intake of more than 60 g of ethanol on a given drinking occasion was used as a cut-off for high-risk and low-risk drinkers.
According to the self-reported alcohol intake per occasion during the last year reported in the M-CIDI, 97 participants were classified as low-risk drinkers and the other 94 as high-risk drinkers (drinking behaviours of the two groups shown in Table 1).
During the second appointment, approximately 9 days (SD = 16 days) later, participants performed the PIT task consisting of four phases. The Pavlovian phase and the PIT phase were done within the MRI scanner, whereas the instrumental phase and the forcedchoice phase were conducted outside the scanner. As briefly mentioned above, participants were presented with images of various shells whose quality (good or bad) was randomly assigned. During the instrumental training, participants were asked to learn the quality of each shell through trial-and-error instrumental responses. When collecting or leaving the shells, the participants received probabilistic feedback that dictated whether their action resulted in a monetary gain or loss. To collect a shell, the participants were required to press the left mouse button five or more times. Each button press resulted in a visual cue (a small dot) moving closer to the image of the shell. To Information S1). In the last phase, participants were presented with two CSs within 2 s and were required to choose one. A more detailed PIT task description is provided in Figure 1 T A B L E 1 Drinking behaviour of the sample (also see Garbusow et al. 19 ). Participants also rated their subjective experience with the five Pavlovian fractals after the experiment. The analyses for the subjective ratings and forced-choice query trials are presented in Supporting Information S6.

| Behavioural analysis
It is important to note that the same dataset was used in a previous study from within our group 21  The behavioural data were analysed with R 3.4.0 (R Core Team, Vienna, Austria). ER was used as a primary measurement of F I G U R E 1 Pavlovian-to-instrumental transfer (PIT) experiment procedure (also see Garbusow et al. 18,19 ). (A) Instrumental phase: participants learned to collect the good shells (press the button more than five times to move the dot towards the shell) and leave the bad shells (no action was required) according to the probabilistic feedback. After 60 trials, instrumental training ended once participants reached the learning criterion (80% correct choices over the last 16 consecutive trials) or at a maximum of 120 trials. (B) Pavlovian phase: participants passively learned the association between five types of compound conditional stimuli (CSs, consisting of fractal-like images and pure tones) and positive (€1, €2), negative (€−1, €−2) or neutral (€0) unconditioned stimuli (USs). There were 80 trials in total with 16 trials of each type. (C) PIT phase: Participants performed the instrumental task again with the tiled fractal images of the CSs in the background. Each trial lasted 3 s, with the fractal images shown 0.6 s before the instrumental shells. Therefore, participants had a response window of 2.4 s. There were 90 trials in total. This phase was done under nominal extinction to avoid further learning. Additionally, there were 72 trials with alcohol/water pictures presented in the background in combination with the two instrumental stimuli (details about the alcohol/water PIT trials shown in Supporting Information S6).
(D) Query trials: in order to verify the acquisition of Pavlovian expectations, participants needed to make forced choices between two CSs within 2 s. Each possible pair of the CSs was presented three times in a randomised order task performance in the PIT phase. Correct responses were defined as at least five button presses in collect trials, or less than five button presses in the leave condition, regardless of the background stimuli.
To check whether our approach for PIT data analysis is suitable, we first compared the ER across the 14 conditions (7 Pavlovian cues × 2 instrumental behaviours), which confirmed that the main difference in ER arises from the incongruent versus congruent contrast ( Figure S1). Within the incongruent condition, the ER showed a symmetric pattern: collecting a good shell with a negative Pavlovian background did not differ from leaving a bad shell with a positive background. This symmetric pattern held true when assessing the association between the ER and the drinking behaviour; a detailed description and the exploratory analyses of alcoholic/water beverage background trials are displayed in Supporting Information S1.
The interference PIT effect score was calculated by subtracting the ER in the congruent condition from the ER in the incongruent condition for each individual. To test whether the participants make more errors in the incongruent condition compared with the congruent condition, a one-tailed, one-sample t test was conducted on the interference PIT effect score. The one-tailed test was used on the basis of our a priori hypothesis that the ER is higher in the incongruent compared with the congruent condition.
The association between performance during the PIT task and the alcohol drinking behaviour was then tested, particularly binge drinking behaviour. Again, on the basis of our hypothesis, a one-tailed two-sample t test was performed accordingly to test whether the interference PIT effect was higher in the high-risk compared with the low-risk drinking group. analysed. The dmPFC, lPFC and VS masks were defined on the basis of the 12 mm spheres around the peaks from previous review papers. 27,33 The amygdala mask was defined anatomically (details in Figure 2). The mean individual parameter estimates were then extracted within the four ROIs from the first-level incongruent versus congruent contrast. To examine the neural incongruency effect on the group level, the mean parameter estimates from the four ROIs were tested in 4 one-sample t tests. Following this, the association between the brain response to interference and the behavioural interference PIT effect (ΔER) was tested with Pearson correlation tests for the four ROIs separately. These results were corrected for four comparisons with Bonferroni correction, with p corr. = 0.05 (p uncorr. = 0.0125) as the threshold.
These ROI analyses were followed by an exploratory whole-brain analysis of the incongruent versus congruent contrast and its association with the behavioural interference PIT effect (i.e., covariate effect on the second level) at an uncorrected threshold of p < 0.001, cluster size k ≥ 50. Whether or not the association between behavioural and neural incongruency effect differs from risk status was also explored.

The detailed description for this analysis is shown in Supporting
Information S5.
To further explore how the effective connectivity modulated by the incongruent condition differs between the two groups, especially regarding the interplay between the VS and the dmPFC and lPFC areas, dynamic causal modelling (DCM) analyses were applied to the data. 34 The time series were extracted from the peak voxels within the VS, lPFC and dmPFC that showed more activation during the conflict (i.e., incongruent-congruent contrast) because no regions were less activated during the conflict. Accordingly, for each individual, the time series of the three regions were extracted from 8 mm spheres centred on the individual local maxima, which were allowed to vary within 5 mm spheres around the group peak voxels during the conflict (incongruent-congruent contrast). The amygdala was excluded for this exploratory analysis, as there was no neural response in the amygdala within our contrast of interest; detailed information about this can be found in  (C) ventral striatum (VS) mask: defined on the basis of the peak of a previous meta-analysis on functional magnetic resonance imaging (fMRI) reward-related tasks. 33 The conjunction of the two 12 mm spheres around the peak Montreal Neurological Institute (MNI) coordinates: −12/10/−6 and 12/10/−6 were defined as the VS mask. (D) Amygdala mask: the bilateral amygdala mask was defined anatomically on the basis of the AAL atlas in the WFU PickAtlas toolbox 53

| Association between risk status and PIT effect
To further examine whether the PIT effects were associated with risk status, logistic regression was employed with risk status as the dependent variable. Possible predictors included the behavioural interference PIT effect and parameter estimates from the neural activated clusters in the incongruent condition (after adjusting for the behavioural interference PIT effect to avoid collinearity in predicting). In a stepwise backward selection process, the best combination of predictors was examined. Data-driven clusters were again used for this analysis because it was expected that these regions would reflect the neural responses within the PIT task more precisely compared with the ROIs. The driving input consisted of all Pavlovianto-instrumental transfer (PIT) trials that entered the ventral striatum. The red arrows specify the intrinsic connectivity: all three regions were assumed to be intrinsically connected to each other and to themselves. The incongruent condition was assumed to modulate the connectivity between each pair of regions in three ways: forward, backward or bidirectional, which resulted in 27 modulatory structures in total 3 | RESULTS

| Behavioural results
The ER was found to be, on average, approximately twice as high in the incongruent condition (30.8%) as compared with the congruent condition (15.6%, Figure 5A). This increase of ER was highly signifi-

| Neural incongruency effect-ROI analysis
In the ROI analyses, the four one-sample t tests of the parameter estimates within the four ROIs did not survive the correction for multiple comparisons, thus indicating no significant difference in the congruent condition compared with the incongruent condition on the group level.

| Neural incongruency effect-Whole-brain analysis
With respect to the explorative whole-brain analysis, the second-level In contrast, no brain region showed higher activity during the congruent compared with the incongruent PIT trials at the same statistical threshold (whole-brain p < 0.001, cluster size k ≥ 50).

| Neural correlates of the behavioural interference PIT effect-Whole-brain analysis
In the next step of the whole-brain analyses, whether or not the neural response to interference was associated with the behavioural interference PIT effect was investigated by conducting a one-sample t test on the behavioural interference PIT effect covariate. Neural correlates of the behavioural interference PIT effect were seen in the VS  Table 2). To illustrate the brain correlates of the behavioural interference PIT effect (ΔER), the neural activation within the three activated clusters was plotted in response to incongruent over congruent trials (neural incongruency effect) against the behavioural interference PIT effect ( Figure 7). As can be seen, the neural response to incongruency in the VS, lPFC and dmPFC was higher in subjects with a stronger behavioural interference PIT effect. However, not all the individuals showed responses to incongruency-this effect was driven by around half of the individuals who committed more errors in the incongruent condition as compared with the congruent condition. The association between the behavioural interference PIT effect and the neural incongruency effect was stronger for low-risk drinkers compared with high-risk drinkers in the VS and the lPFC, but the difference was marginal in the dmPFC (detailed result in Figures S3 and S4).

| Effective connectivity difference between high-and low-risk drinkers
The model selection was first performed in order to select an optimal family of models among the six families in Figure 4. The selection was performed separately for the high-and low-risk drinking groups to test whether the winning family of models was different for the two groups. The selection was based on the exceedance probability: a higher exceedance probability suggests one family of models has more evidence compared with other specified families of models.
According to the family exceedance probability, the winning family for   Generally speaking, with around twice the exceedance probability of the winning family compared with the second-best family, it was concluded that there was only weak support for the two different winning families for the two groups (plotted in Figure 8).
Because of the different winning families, the strength of the connectivity was further obtained through BMA across the entire model space for both groups; this ensured the parameter estimates were comparable. The BMA does not make inferences about the model structure, but it rather computes a weighted average of the effective connectivity parameters from all the specified models.
The weights are given by the posterior probabilities of different models. 36 On the basis of the BMA results, one can directly compare whether the effective connectivity parameters between certain brain regions are different for the two groups. According to the criteria that the posterior mean is larger than zero at a probability threshold of 95%, the incongruent condition significantly modulated the connection from the VS to the lPFC and the bidirectional connection between the lPFC and the dmPFC for the low-risk but not the high-risk drinkers (Table 3). By comparing the modulatory parameters between the two groups, significantly higher effective connectivity was found from the VS to the lPFC modulated by the incongruent condition in the low-risk compared with the high-risk drinking group (p = 0.004 after Bonferroni correction for six comparisons) ( Table 3).

| Association between risk status and PIT effects
In the backward stepwise logistic regression with risk status as the dependent variable, the best model (χ

| DISCUSSION
In this study, we investigated whether interference between Pavlovian and instrumental control, assessed with a PIT task, is associated with risky alcohol use in a cohort of healthy males aged 18 years. In contrast, when the interindividual differences in interference were considered, it was found that the VS, lPFC and dmPFC activation correlated positively with the behavioural interference PIT effect.
Previous literature repeatedly reported the VS to reflect the influence of the Pavlovian cue on instrumental behaviour. 11,23,25 The VS cluster that was found also extended to the dorsal striatum; this has also been shown by two previous studies. 12,26 In contrast to previous studies, we did not find amygdala activation. 11,21,[23][24][25] As suggested by these studies, the amygdala may compute the affective valence of Pavlovian cues in the PIT task. Notably, one difference between the previously mentioned studies and the current study involves the valence signal. In the aforementioned PIT studies, when comparing the positive/negative Pavlovian cue condition with the neutral condition, the finding reflected a mixture of salience and valence signal.
Conversely, in the current analysis, the valence signal was averaged out when pooling the different combinations of Pavlovian cues and instrumental stimuli into incongruent and congruent conditions. This may begin to explain why activation in the amygdala was not found.
Taken together, the signal seen in the VS may reflect a salience signal indicating that the Pavlovian cue is at odds with the required instrumental behaviour.
The response elicited by incongruent trials was also found in the dmPFC. This region has been extensively linked to conflict-related performance monitoring, in which it plays an important role in deciding the subsequent adjustments in performance. 39,40 Additionally, incongruent trials also evoked a response of the lPFC, which is a critical structure that gathers task-related information and exhibits topdown cognitive control 41,42 in relation to conflict monitoring, error monitoring and response selection. 43 To summarise, the activation It is worth noting that a previous paper from our group found that the association between the valence of the Pavlovian cues and response rates (indicating response vigour) was stronger for high-risk than low-risk drinkers. 21 However, in this study, the main focus was to investigate the motivational effect of Pavlovian cues on the ongoing instrumental behaviours, regardless of whether they promote (congruent condition) or hinder (incongruent condition) the required instrumental response. Despite using the same dataset, the main focus of the current study was to examine the interference effect of Pavlovian cues when they are in conflict with the necessary instrumental behaviour. By doing this, the motivational and cognitive control perspectives were able to be examined simultaneously, as both perspectives were present during trials with interference from Pavlovian cues. Therefore, these results connect previous research in the fields of cognitive control and motivated behaviour. Even though the interplay of cognitive control and motivated behaviour is essential to understand addictive behaviour, most experimental approaches either focus on one or the other. An exception would be the go-no-go/PIT task, 16,46 which assesses the influence of non-drug Pavlovian cues on response inhibition. So far, go-no-go/PIT tasks have not been used to study substance use or dependence. These results, therefore, complement previous studies that reported an association between binge drinking and impaired interference control in young adults. 47 Importantly, the conflict between Pavlovian and instrumental control substantially differs from conflict seen in traditional interference tasks such as the classical colour-word Stroop task (conflict at stimulus level) 48,49 or the Simon task (conflict at response level). 50,51 In these 'cold' interference tasks, responses are instructed and are not the result of learning based on rewards or punishments. Interference in these tasks mainly results from automated response tendencies (i.e., neither the colour representation in the Stroop task nor the location cue representation in the Simon task triggers motivational responses). In contrast, in our 'hot' interference task, Pavlovian cues trigger a motivational response, that is, approach or avoidance behaviour and interfere with motivated instrumental behaviour. On the basis of the hypothesis about the difference between the 'cold' and 'hot' interference task, future studies could investigate whether the PIT effect we found could (to some extent) be explained by these 'cold' interference tasks or it involves fundamentally different mechanisms.
To conclude, the results of the current study show that the susceptibility to Pavlovian interference during a PIT task is linked to hazardous drinking behaviours at age 18. Although the imbalance between the top-down and bottom-up systems has been suggested to be associated with addictive behaviour, previous studies have tended to consider either the perspective of cognitive control or motivated behaviour but not both at the same time. Using a PIT task, we assessed the top-down control and its interaction with bottom-up Pavlovian and instrumental processes. Our experimental data indicate that a poor interplay between top-down and bottom-up processes may contribute to early hazardous alcohol use.

| LIMITATIONS
We investigated a sample of 18-year-old social drinkers. In this sample, some participants did not commit any errors during the PIT task.
It is thus unclear whether these participants experienced no interference at all or they had better interference control. Another explanation could be that the PIT task was not sensitive enough to capture the very subtle effects that may have been present in these participants. Therefore, a possible solution to this issue could be found in further refinement of the PIT task to increase the sensitivity to more subtle effects. Additionally, the classification of high-and low-risk drinkers based on the self-reported alcohol consumption data during the past year may not be entirely accurate because of the possible memory bias; future studies may improve this by using more frequently assessed electronic diary data. Another limitation of the current study is that these results cannot be generalised to nonmale populations.