Inhibition training: Automatic inhibition or response conflict

Student: Gerben Kraaij Student number: 5749433 Supervisor: Thomas Gladwin Co-assessor: Denise van Deursen

Inhibition training: Automatic inhibition or response conflict

Abstract

In previous studies it has been shown that drinking behavior in alcoholics and hazardous drinkers can be adapted using cognitive bias modification strategies. A number of these strategies involve consequent inhibition of a response to alcohol-related stimuli. In the present study we sought to disentangle the contributions of two not mutually exclusive hypotheses (automatic inhibition and response conflict) that have been proposed to provide a working mechanism for modification of implicit cognition in these cases. A joystick task which entails a combination between an approach-avoidance and a go/no-go task was used as an experimental intervention. Eighty hazardous drinkers, selected from a student population, were assigned randomly to one of four conditions. In two experimental conditions participants never responded to alcohol stimuli and either pushed away (push condition) or pulled (pull condition) soft drink stimuli. We measured weekly alcohol intake, implicit attitude towards alcohol, approach-avoidance tendencies and attentional bias both before and after the intervention. Before the intervention AUDIT score and response inhibition were also measured. Before the intervention there were condition-related differences in implicit biases, which made comparisons between conditions more difficult. After the intervention only the experimental pull condition had developed a negative bias for alcohol, which provides support for the response conflict hypothesis. In the experimental push condition reaction times on the joystick task for pull responses to alcohol stimuli were slowed down relative to soft drink stimuli, which might reflect the development of automatic stimulus-stop associations. Neither of these changes was reflected in actual drinking behavior, which was most likely due to top-down inhibitory control causing explicit rejection of influences from implicit cognition. Future studies might benefit from using a clinical population in order to achieve stronger consistency between implicit and explicit attitude changes. The current study also indicates that in future studies the time of the day at which participants are tested must be considered an influential factor for the effect of interventions. Keywords: alcohol, implicit, bias, inhibition, conflict, approach

Introduction

Addiction to alcohol is one of the most common addictions worldwide (WHO, 2011; Grant et al., 2006). Next to the profound societal costs, which were in 2002 estimated to be around 184.6 billion dollars in the United States alone (ONDCP, 2002), the addiction also has profound influences on an individual’s health (Rehm et al., 2009) and ability to maintain a normal social life. Prolonged alcohol abuse causes profound changes in cognitive and neurological functioning (Brown et al., 2000; Goldstein et al., 2001; Oscar-Berman & Marinkovic, 2003; Ratti et al., 2002;) leading to severe impairments in everyday functioning.

To counter this problem many different treatments have been developed over the years (Jaffe, 1991), often involving cognitive behavioral training in order to change drinking behavior (Longabaugh & Richardson, 1999; Ouimette et al., 1997). But despite the diverse spectrum in

therapies and even medicines (Krystal et al., 2001), effective and reliable treatments are still lacking. A national survey in the United States found that one year after receiving treatment about one third (35.9%) of individuals were either abstinent or low risk drinkers, indicating that relapse rates are still relatively high (Dawson et al., 2005). Moreover relapse rates appear to get higher when the number of years of follow-up increases, meaning that people might still encounter total relapse even after years of abstinence (Jaffe, 1991). Although not favorable, from a theoretical perspective these relapse rates are interesting since they suggest that many conventional treatments might keep the actual core of alcohol dependency relatively intact. Answering the question of what causes these relapse rates is crucial for developing treatments that are effective on the long term.

Implicit and explicit cognition

Most addicts acknowledge the self-destructive effects of their behavior, but still fail to stay abstinent. This dissociation between explicit convictions and behavior illustrates the well known fact that explicit convictions and intentions are not always enough to guide behavior. In psychology a major distinction regarding cognitive processes is that between reflective and automated, or

impulsive, processes (Strack & Deutsch, 2004). It is these processes together that guide behavior. The reflective system is involved in relatively slow evaluations, which are propositional in nature,

operating largely within conscious awareness (Gawronski & Bodenhausen, 2006). In this paper these processes will be referred to as explicit cognition. The automatic system, which is associative in nature, works on a shorter timescale and operates largely outside of conscious awareness. It is primarily concerned with the motivational and emotional value of thoughts and actions supported by the fact that implicit evaluations are often accompanied by activations in areas like the amygdala and striatum (Cunningham et al., 2004; Cunningham et al., 2007). Processes belonging to the automatic system will be referred to as implicit cognition.

Since most addicts know perfectly well what they should be doing on an explicit level, it might be implicit cognition that is most interesting in explaining the paradoxical behavior (Stacy, 2010). Implicit cognitions can be spontaneously initiated upon perception of a stimulus which activates pre-existing associative structures (Gawronski & Bodenhausen, 2006). Without intervention from the reflective system these cognitions will give rise to affective actions, whereas the reflective system gives rise to evaluative actions. In reality an individual’s actions, attitudes or thoughts are almost never process pure in the sense that they are neither the result from solely implicit nor explicit cognition. Also these two do not operate independently, for example, Gawronski and

Bodenhausen present several possibilities how bidirectional interaction between implicit and explicit cognition can explain attitude formation and change (Gawronksi & Bodenhausen, 2006).

Implicit cognition and addiction

In implicit cognition broadly three sets of processes are distinguished which are all adjusted over the course of addiction. Implicit cognition encompasses; i) attentional bias, which refers to automatic reallocation of attentional resources concerning perception and processing of addiction related information (Field & Cox, 2008), ii) action tendencies, which are automatically prepared motor responses to approach addiction related stimuli upon perceiving or thinking about the substance of abuse (Palfai & Ostafin, 2003; Wagner et al., 2011), and iii) memory associations, in existing associative structures a range of addiction related concepts can be spontaneously activated upon perception of an addiction related stimulus through spreading of activation. Together

automated processes might be able to initiate, plan (Veling et al., 2011; Wagner et al., 2011) and guide behavior outside awareness (Aarts et al., 2007; Childress et al., 2008). Central might be a strong incentive motivational state that drives these automated processes (Palfai & Ostafin, 2003).

According to Robbinson and Berridge (2003) this motivational state is the result of a sensitization of the brains reward system to addiction related stimuli that develops over the course of addiction (Everitt et al., 2008). It is in this light that in the literature implicit cognition has been ascribed an important role in establishing and maintaining an addiction (Wiers & Stacy, 2005). Bringing together these characteristics it appears reasonable that implicit cognition and automated processes could also play a valuable role in explaining paradoxical behavior and relapse rates seen in addicts (Stacy, 2010). This is underscored by two studies which found that the extent to which attentional bias for

addiction related stimuli had formed, contained significant predictive value for treatment outcome and relapse rates (Cox et al., 2002; Waters et al., 2003), indicating that changes in implicit cognition and relapse rates might in some way be connected.

Measuring implicit cognition

Important for measuring implicit cognition is the consideration that the contents upon which the automated processes operate are only indirectly accessible, both to ourselves and others (Stacy, 2010). Over the years several psychological tests have been developed to gain access to them. To assess attentional bias for example an emotional version of the Stroop test can be used (Cox et al., 2006) in which problem drinkers show slowed response times in color naming of alcohol related words (Sharma et al., 2001). For measuring both action tendencies and memory associations several modified versions of the implicit association test (IAT) have been developed. The IAT is a member of a bigger category of response interferences tasks (Gawronksi et al., book chapter), designed to reveal implicit associations between a target concept and an attribute by alternately mapping two

discrimination tasks onto a single pair of responses (Greenwald, 1998). Response time slowing could reflect the effort of a subject in trying to overcome pre-existing implicit associations that conflict with the response required by the task. Response time acceleration could reflect facilitation of response preparation and execution by existing associations. For this cause the IAT has proven to be both a reliable and flexible measurement tool, sometimes even accounting for unique variance in drinking behavior not accounted for by explicit report (Stacy, 2010; Thush & Wiers, 2007). By varying the target concept it can be used to measure both memory associations (Greenwald, 1998) or approach-avoidance tendencies (Palfai & Ostafin, 2003). Although for the latter more direct and exact

measurement tools are often preferred (Field et al., 2011; Krieglmeyer & Deutsch, 2010). Using the IAT it was also found that the strength of implicit associations and drinking behavior are highly correlated throughout adolescence (Stacy, 2010), suggesting that these two reciprocally feed on each other.

The rider and the horse

Although the exact relation between automated and reflective processes in addiction is of course complicated, their relation has been usefully illustrated by Friese et al. (2011) using the image of a rider and his horse. The rider might have the intention to go to a certain destination, but when he is for some reason not in control of the horse and the horse takes off because it might smell food nearby, the rider might end up somewhere he did not intend to go. In this illustration the rider and the horse respectively represent the reflective and the automated system. From this perspective it might be more easy to see why many conventional treatments fall short. The rider might have his own convictions and goals clear, but the aims of the horse can be dissociated from, and even opposed to, the goals of the rider. It is expounded by Strack and Deutsch (2006) that this is indeed what is happening in the development and persistence of addiction. The impulsive system will strive for the emotional salience of addictive behavior while the long term consequences of this behavior, which are not considered by the impulsive system, are opposed to the addicts desire to stay abstinent. Like stated above automated and reflective system do not operate independently, under normal

circumstances the rider has strength to control and direct the movements of the horse. This relates to general executive functioning which normally controls and directs behavioral influences of automatic processes. An interesting example in this respect was provided by Houben and Wiers (2009). They demonstrated that it is the reflective systems ability to exert response inhibition that mediates the relation between implicit alcohol-related associations and drinking behavior.

Differences in drinking behavior they observed arose as a result of differential ability of the reflective system to override automatic evaluation through response inhibition. To strengthen this point, the opposite was also found, in individuals with high executive functioning drinking behavior was best predicted by explicit attitude (Thush et al., 2008). One of the problems in addiction is that executive functioning is impaired, as suggested by both behavioral (Lawrence et al., 2009) and neurological evidence (Goldstein et al., 2011; Ratti et al., 2002). And when executive functioning is impaired it is more easy for implicit cognition to direct behavior (Jentsch & Taylor, 1999). Here we can see that although conventional treatments might give the rider a good set of instructions, if he does

not have the strength to control his horse he can not prevent it from running the wrong way. Therefore executive functioning and response inhibition have also been target of treatment strategies.

Taming the horse

Outside the field of addiction specific training of self-control and executive functioning has already proven to be effective (Cohen et al., in press). In the same line of reasoning attempts have been made to strengthen executive functioning and response inhibition in order to control drinking behavior in addicts (Brady et al., 2011; Houben et al., 2011a; 2011b). For example, Houben et al., (2011b) focused on increasing working memory capacity and found that successful training caused a reduction in drinking behavior in a population of heavy drinkers. Although fruitful, what this approach achieves is that it strengthens the rider and teaches him how to deal with a wild horse, but it does not tame it. This leaves open the possibility that addicts relapse into their old drinking habit that will been preserved in the automated system and shows up again if their executive functioning fails at some point (Heatherton, 2011).

When it comes to taming the horse itself, a prominent possibility is to retrain automated processes and direct them away from addictive behavior. One such attempt was performed by Wiers and colleagues (2010) by retraining action tendencies. In addiction strong positive valuation of alcohol stimuli is accompanied by automatic approach tendencies. Wiers and colleagues let subjects respond to alcohol or non-alcohol pictures by either pulling or pushing a joystick. They found that subjects who consequently pushed the joystick in response to alcohol pictures showed reduced approach behavior both on an explicit and an implicit level. Avoiding a positive stimulus would require the selection of a response that is incongruent with the automatically activated response, this is believed to lead to reduced availability or modification of these automated processes. Most promising might be the fact that a simple intervention like this was also shown to improve treatment outcome in a clinical population (Wiers et al., 2011).

Another interesting attempt was performed by Houben and colleagues (Houben et al., 2011). Using a classical go/no-go task they sought to improve inhibitory control over impulses to alcohol cues in heavy drinkers. They did so by consequently pairing alcohol related cues with a no/go signal in one of the two experimental condition. In the other condition alcohol cues always required a go response. Behavioral results of this intervention were measured in three ways; i) valence version of the IAT (Greenwald, 1998), ii) bogus taste test, iii) self-reported drinking behavior using TimeLine Follow Back questionnaire (Sobell & Sobell, 1990). It was found that the inhibition training resulted in a more negative implicit attitude towards alcohol in the alcohol no-go condition. Importantly this result was accompanied by a decrease in self-reported drinking behavior proving its effectiveness outside the lab environment. A trend towards reduced drinking in the taste test did not reach

statistical significance. The most important point here is that the implicit attitude towards alcohol had been changed in such a way that implicit and explicit attitude were now more aligned and more consistent with actual behavior. From a clinical perspective these results look promising, but for this study the most important question to answer will be what causes the effectiveness of this

intervention?

Automatic inhibition

When searching for explanations it is good to have in mind what is required to change an implicit attitude. According to Gawronksi and Bodenhausen this either involves an incremental change in associative structure, or a change in pattern activation of already existing structures (Gawronski & Bodenhausen, 2006). While the second notion is important in assuring whether the effects of implicit retraining will generalize from the lab to real life situations, our primary concern will be with the first notion. One example of how this can be achieved is given in a study by

Verbruggen and Logan who have researched how effective automated stimulus-stop associations can be developed (Verburggen & Logan, 2008). Using both classical go/no-go and stop-signal tasks they found that such associations only developed when stimuli are consequently paired with a stop/no-go signal. After consistent pairing of stimulus-response pairs in the training condition, they reversed mappings in the testing phase. They inferred the existence of automatic stimulus-stop associations

from the slowing of reaction time to stimuli that had been consistently paired with stop/no-go signals during training. Their results provide support for the automatic-inhibition hypothesis. According to this hypothesis impulsive responses to stimuli can be inhibited by creating automatic response inhibition associations through explicit training. The training used by Houben et al, shows a similar design and therefore renders the automatic-inhibition hypothesis a possible explanation. As an additional benefit, it was found that automatic inhibition effects are not influenced by alcohol consumption while intentional inhibition is impaired (Abroms et al., 2006). In that sense it might prevent treated individuals from experiencing a total relapse after a single violation of abstinence, as is often seen (Heatherton et al., 2011).

One concern is that the studies by Verbruggen and Logan did not involve addiction related stimuli but rather neutral stimuli which do not normally induce strong action tendencies. It does explain how automated responses can be directed or redirected, but it does not explain how to overcome the strong affect that normally drives existing action tendencies in addiction. As stated before, the automated system is heavily involved in the processing of emotion and value, enabling it to induce strong intrinsic motivation for responses favored by the automated system. Studies demonstrate that the brains reward circuitry already shows a response upon perception of addiction related stimuli, regardless of whether they were perceived consciously (Heatherton et al., 2011) or unconsciously (Childress et al., 2008), while this is not the case in non-addicted individuals (Robinson & Berridge, 2003). Additionally the emotional salience of a stimulus can cause redirection of attention as a bottom up process (Ohman et al., 2001). It is the affective valence that is associated with

addiction related stimuli that drives the preparation and execution of action tendencies and the redirection of attentional resources (Veling et al., 2011), and it might be that the key in changing the automated processes lies in changing these affective associations.

Response conflict and stimulus devaluation

From this line of thought Aarts and colleagues commenced by arguing that stimulus perception in addicts can unconsciously lead to motivated goal pursuit (Aarts et al., 2007). In their experiments a pre-existing goal is subliminally primed, accompanied by negatively valenced

information. They find that priming a goal in temporal proximity of negatively valenced information reduces motivation to pursue that goal, even without conscious awareness. According to Verbruggen and Logan, if the goal would consequently be paired with negative affective information, they would become associated in the automated system which could possibly reduce positive affect in alcoholics. Consistently, Houben et al. (2010) performed a likewise experiment with alcohol related stimuli in regular drinkers. They found that consequently presenting alcohol related cues in a negative context resulted in a more negative attitude towards alcohol demonstrating that affective implicit

associations had indeed been changed. This change was further accompanied by reduced drinking behavior, again arguing for an important role of implicit cognition in driving behavior.

According to Veling and colleagues, additional presentation of negatively valenced

information is not necessary because under the right circumstances the brain is capable of generating negative affect itself (Veling et al., 2008). They outline that when people refrain from executing their automatic response to a positive stimulus because of situational demands, this will automatically lead to devaluation of that stimulus. The devaluation is a consequence of the conflict that arises in

response selection between the response triggered in the automated system and the response required by situational demands. The brain is thought to solve this conflict by adding negative affect to the automatic approach response. In the light of the results obtained by Houben et al. (2011a), stimulus devaluation through solving of a response conflict might be a second possibility to explain the intervention effect. This second possibility is referred to as the response-conflict hypothesis. Although behavioral data render this explanation possible, Veling et al., only mention indirect evidence for how the brain could generate negative affect.

However, a number of recent studies provide a more valid basis for their argument. In a new study be Houben et al. (2012), it was found that inhibition training results in a reduction of weekly alcohol use. This reduction was accompanied by a more negative implicit attitude towards alcohol while action tendencies remained unaltered. This can be seen as a strong indication that conscious

inhibition of a response to alcohol-related stimuli causes a decrease of affective associations. This leaves open the question how the brain could ‘spontaneously’ generate negative affect. A plausible answer to this question can be found in the work by Kiss et al. (2008), who provide both behavioral and electrophysiological evidence that response inhibition and emotional evaluation are linked in the brain. They found that in a go/no-go task, faces that were presented with a no-go signal were

successively evaluated as less trustworthy than faces that required a go-response. And interestingly, within the no-go trials it was the amplitude of the N2-peak, which is used as a measure for the effectiveness of response inhibition, that correlated with stimulus devaluation. This indicates that more successful response inhibition in the motor domain leads to greater devaluation in the emotional domain. That is, intentional inhibition in one domain can ‘spill over’ and cause

unintentional inhibition in another domain. Recent fMRI studies indicate that orbitofrontal cortex (OFC) and right inferior frontal cortex (rIFC) are both possible loci of a common inhibitory mechanism across domains (Berkman et al., 2009; Doallo et al., 2011). These studies showed that activation of OFC (Doallo et al., 2011) or rIFC (Berkman et al., 2009) in no-go trials was accompanied by reduced activation in limbic affect regulating areas like the amygdala (Berkman et al., 2009; Doallo et al., 2011), and ventral striatum (Berkman et al., 2009). Combined, these results provide a solid

neurological basis for affect regulation of stimulus processing through inhibition of motor responses, as proposed by response-conflict hypothesis. However it does not directly demonstrate the necessity of response conflict in explaining inhibition training effects.

Current study and hypotheses

In the current study firstly we would like to replicate the inhibition training effect found by Houben et al. (2011). Given that we are able to do so, we want to examine the contribution of the processes proposed by the automatic-inhibition hypothesis and response-conflict hypothesis in establishing this effect. It is worth to explicitly mention that these hypotheses are not mutually exclusive. That is, stimulus devaluation might occur as a result of response conflict solving while at the same time automatic response inhibition associations might develop because of the consistent pairing between stimuli and no-go indicators. To tell their contributions apart a modified joystick go/no-go task will be used with two experimental conditions. In the pull condition subjects will always pull the joystick towards themselves in response to non-alcohol pictures, while they refrain from responding to alcohol pictures. In the push condition subjects will always push the joystick away from themselves in response to non-alcohol pictures and again refrain from responding to alcohol pictures. In the latter condition the response the subject has to inhibit is not congruent with the automatic response tendency to approach alcohol stimuli and therefore no response conflict will be present. In the first condition the automatic approach tendency is congruent with the to be inhibited response thus causing conflict during inhibition. By comparing the results from assessments sessions before and after this intervention in which subjects respond to both alcohol and non-alcohol pictures, the effect per condition can be quantified. The results from both conditions will be compared to two control conditions where subjects receive placebo training.

To measure the effect of the intervention we will use measures for; i) implicit cognition, ii) explicit attitude, and iii) drinking behavior. To assess influence on implicit cognitive processes both the valence IAT and approach/avoidance IAT, will be performed by participants both before and after intervention. Additionally implicit attitude towards alcohol will be measured by a bogus taste test in which the amounts people consume are measured under the cover story of a taste test (Field and Eastwood, 2005). Drinking behavior will be monitored for two weeks before and one week after training using the Time Line Follow Back (TLFB) questionnaire (Sobell & Sobell, 1990).

It is expected that if stimuli are indeed devaluated through response conflict, that a more negative implicit attitude will develop specifically in the experimental pull condition. Differences between experimental conditions are also expected for the approach/avoidance IAT. Wiers et al. (2010) found that consistently pushing a stimulus away induced avoidance associations while consistently pulling a stimulus towards yourself left action tendencies unaltered. In our experimental conditions, soft drink stimuli are either pushed or pulled. And since the IAT measures action

push condition might become more approaching or will remain unaltered. Action tendencies in the experimental pull condition are expected to become more avoiding because participants consistently inhibit a positively valued pull response to alcohol. For the joystick task specifically we expect

reaction times for alcohol pull go trials to become slower relative to soda pull go trials. We expect to find this in both experimental conditions although the effect will be strongest in the experimental pull condition.

In line with evidence that affective alcohol associations and drinking behavior reciprocally influence each other (Houben et al., 2010; Rooke et al., 2008; Stacy, 2010) we expect that a more negative implicit attitude towards alcohol will be accompanied by a reduction in self-reported drinking behavior. Although the focus of this study is mainly theoretical, from a clinical perspective it would be interesting to find such a reduction, which would be supportive for previous work by Wiers et al. (2011). Additionally a classical Stroop test will be used to obtain a measure of response

inhibition. Houben et al., found that the ability to inhibit a response mediates between the existence of implicit associations and drinking behavior (2009). Based on these findings we would like to know whether changes in implicit cognition will be only accompanied by changes in drinking behavior for participants whose response inhibition ability is low. Lastly, a version of the addiction-Stroop test can help to determine whether any found effects are likely to be attributable to bottom-up or top-down processes (Cox et al., 2006).

Methods

Participants were 95 young adults, mainly students from the University of Amsterdam (35 male, 60 female; mean age=21,7, SD=5,59). Participants were recruited via an online registration board for psychological tests and through distribution of promotional material such as flyers and posters inside the university buildings. As a screening procedure participants filled out the alcohol use disorder identification test (AUDIT) (Daeppen et al., 1998; Saunders et al., 1993) which they accessed online. A minimal score of 8 on this test was set as an inclusion criterion for participation. In our communication it was stated that we were looking for regular drinkers between 18 and 30 years of age, but it was never explicitly mentioned that there would be a cut-off score for participation. Average score of included individuals was 13.2 (SD 4,78). Average score in general was 10,97 (SD 5,35). For a number of analyses our pool of participants was split into a high (ME = 19,6, SE = 3,35, N = 29) and a low (ME = 9,1, SE = 0,88, N = 20) AUDIT group. This split was made mainly for explorative and supportive analyses for cases in which quantitative differences in effect size were expected based on AUDIT score. Participants that were included in the low- and the high AUDIT group were

participants who scored respectively half a standard deviation below or above average AUDIT score. Another split that was made was based on whether subjects were tested in the morning or in the afternoon. This split was thought to be of importance since implicit attitudes might not be stable during the day but might differ between different times of the day. This is because in the morning the sight of beer might activate different (probably more negative) associations than the sight of the same beer might induce during the afternoon or evening. This issue will be discussed in more detail later on. An overview of the number of participants throughout the various stages of the study, including reasons for exclusion or dropout, is depicted in the supplementary materials (suppl. fig. 1). Subjects were rewarded either 3 student credits or 21 euro for participation.

Alcohol use

For participants who made it through the screening, alcohol use was measured before and after intervention using a modified version (Wiers et al., 1997) of the TLFB questionnaire (Sobell & Sobell, 1990), which has proven to be both a reliable and valid measure for alcohol use (Carey, 1997). Participants were asked to indicate how many alcoholic beverages they consumed during each day of

the past 2 weeks. Before the intervention, on average, participants consumed 28 (SD=18,06; range=82) Dutch standard drinking units of 10g of alcohol per week. For analysis of the TLFB questionnaire we used the average consumed amount of standard drinking units per day. IAT

The IAT measures the strength of associations between concepts. In this study two versions of the IAT were used. The valence IAT was used to measure positive and negative associations with alcohol stimuli, the approach-avoidance IAT was used to reveal approach and avoidance associations with alcohol stimuli. In both versions stimuli were categorized in two target categories (alcohol – soda for both versions) and two attribute categories (valence; positive - negative, action; approach - avoidance). Both versions followed a typical seven block design (Greenwald et al., 1998).

Classification was performed by pressing the corresponding response key to the presented stimulus (either ‘f’ or ‘j’, instructions were given at the start of each block). In the first block stimuli were classified based on the target categories only. In the second block classification was based on

attribute categories only. In both blocks all stimuli appeared 2 times with 32 trials in total. In the third and fourth block a target category and an attribute category were combined under one response key, the third block was used for practicing (32 trials), the fourth block for testing (64 trials). The extent to which a participant’s reaction times for the first combined sorting condition are faster than for the second was used to estimate the relative strength of associations between the involved target category and the two attribute categories . The fifth, sixth and seventh block were a repetition of respectively the first, third and fourth block. The difference was that both response keys and positioning of the category labels on the screen were reversed. For example, whereas in the first block categorizing a stimulus as alcohol required pressing the ‘f’ button with the alcohol label being presented on the left of the screen, in the fifth block this required pressing the ‘j’ button with the alcohol label being presented on the right of the screen.

The assignment of the attribute categories to the left and right response keys was

counterbalanced across participants. We also aimed to counterbalance the order of the combined sorting conditions (concerning blocks 3, 4, 6 and 7), so that half of the participants categorized the first target category with the first attribute category first, and the other half categorized the first target category with the second attribute category first. However, due to a mistake in the code counterbalancing was not perfect (36 participants performed order 1; 44 participants performed order 2). The imbalance in performed block-order can proof problematic since not all conditions were affected equally by it. Additional measurements were performed in a later stage to resolve this issue, unfortunately the additional data is not included here yet. Stimuli were always presented in the middle of the screen, whereas category labels were presented just above the stimulus to the left and the right side (see Supplementary Materials figure 3). Stimuli remained on-screen until a response was given. If a stimulus was categorized incorrectly the stimulus was replaced by the word ‘incorrect’, after which the stimulus reappeared and had to be recategorized. The intertrial interval was set to 250ms.

For both the valence and the approach-avoidance IAT the D600 scoring algorithm was used to derive a reaction measure. This algorithm deviates from conventional algorithms on a number of points. Importantly it takes into account all trials from all four combined classification blocks. It deletes trials above 10,000ms and eliminates participants with more than 10% or response times under 300ms, but has no further extreme value treatment. This algorithm has been adopted in similar research by Wiers et al. (2011). A detailed description and in-depth discussion is provided by Greenwald et al. (2003).

Approach-avoidance go/no-go task

The approach avoidance go/no-go (AAGNG) task used in this study is a modified version of the one used by Houben et al. (2011). It comprises a combination between a classical go/no-go task and an approach-avoidance task as used by Wiers et al. (2011). Stimuli were a set of alcohol and non-alcohol pictures (further described below). Stimuli were always presented in the middle of the screen

together with a go or a no-go cue for 1500ms. Participants had to push or pull a joystick (Pro flight 2 – Logic3) when a go cue was shown, and refrain from responding when a no-go cue was shown. As go/no-go cues the letters ‘p’ and ‘q’ were used, for which counterbalanced response instructions were given at the start of each block of trials. If subjects did not respond within 1500ms, a message was shown in red to tell participants they had not reacted in time. If subjects did not respond

correctly, the required responses to the letters ‘p’ and ‘q’ for that block were shown again for 1500ms before the next stimulus was presented. Between trials a blank screen was presented for 750ms. Participants performed the following sequence of sessions over two testing days; i) assessment session, ii) training session, iii) training session, iv) assessment session, v) booster session.

Both assessment sessions consisted of 12 blocks of 24 trials of which the first 4 blocks of the first assessment session were used for participants to get familiar with the task, these blocks were not considered for analysis. Both training sessions consisted of 8 blocks of 48 trials. The final booster session consisted of 4 blocks of 24 trials and had identical instructions to the training sessions. The number of alcohol and non-alcohol trials in a block was always balanced for every 4 trials.

Participants could either be placed in the; i) control pull condition, ii) control push condition, iii) experimental pull condition, or iv) experimental push condition. During the assessment session participants from both the control pull and experimental pull condition were instructed to pull when a go-cue was presented and refrain from responding when a no-go cue was presented. Participants from both the control push condition and experimental push condition were instructed to push when the go-cue was presented and refrain from responding when the no-go cue was presented. Between blocks it was counterbalanced whether participants would respond to alcohol or non-alcohol pictures. In this way, over the whole session participants made an equal number of pull or push responses to alcohol and non-alcohol pictures. In both control conditions the same pattern of responses was required again during the training sessions and the booster session, this was the placebo training. In the experimental pull condition, non-alcohol stimuli were consistently presented with the letter indicating a pull response, while in the experimental push condition non-alcohol stimuli were consistently presented with the letter indicating a push response. In both conditions a no-go cue was presented together with alcohol stimuli.

For all conditions, per session 4 incongruent trials were included. This was done because it reduces the chance that contingency awareness arises in participants. This is important because contingency awareness can cause participants to mistrust their associative evaluations so that actual changes in associative evaluations might be rejected by propositional reasoning and are therefore not reflected in evaluative judgments and behavior (Gawronski & Bodenhausen, 2006).

In the instructions of the task, pushing was phrased as ‘push the joystick away…’, while pulling was phrased as ‘pull the joystick towards…’. This was done intentionally since it has been shown that phrasing of instruction is important to ensure successive negative and positive framing of the mental representation of these responses (Eder & Rothermund, 2008; Krieglmeyer & Deutsch, 2010). This was further reinforced by a feedback mechanism which made pictures gradually become smaller during pushing and gradually become larger during pulling, thus creating the illusion of avoiding or approaching the stimulus (see Supplementary Materials figure 4). This again is important since it has been shown that visual feedback can cover any biases that might arise because of ambiguously verbalized instructions (Leotti & Wager, 2010), it also prevents subjects from easily recategorizing their responses.

The used picture set consisted of 48 alcohol and 48 non-alcohol pictures. Both sets contained an equal number of pictures showing; i) a drink and a filled glass, ii) a drink being opened.

The picture sets were adapted for commonality of locally consumed drinks. This is important to make the stimuli as representative as possible for the target category they are to represent.

In the AAGNG task the trajectory of the joystick movement was sampled every 40 ms with a resolution of 2000 sample points between the extreme push and pull positions. This allowed us to derive reaction measures for the AAGNG task. This was done by fitting the response trajectories to a logistic curve using Matlab (Cavallini, 1993). The maximum and minimum of the second derivative of this curve were used to define the times of the initial joystick movement and the end of the

movement (at the extreme position). We analyzed these Initial and End reaction times, as well as the movement duration, i.e., the End RT – Initial RT. Besides, the time point for which joystick

acceleration was maximal was also used as an alternative reaction measure. Stroop test

Both the original Stroop test and the addiction-Stroop test (Cox et al., 2006) were performed by participants. In both tests participants were handed plasticized cards (A4 format) on which colored words or signs were printed on a grey background. Cards were presented upside down and were turned around by participants on instruction of the experimenter. Participants were to determine the color of the printed words and verbally report them as quickly as possible. The time they took was measured using a stopwatch and was rounded off at one decimal, errors made in color naming were also noted. The addiction-Stroop test involved one card with neutral words and one card with alcohol related words. Both cards contained 56 words in total. For the neutral card 14 different words were each presented 4 times throughout the card. For the addiction-related card by mistake only 13 different words were used. Each word was presented 4 times with exception of the word ‘rum’ which occurred 8 times. (see Supplementary Materials figure 2). Whether participants performed the alcohol or neutral card first was counterbalanced across participants. Words used in the addiction-Stroop test were adapted for commonality of locally consumed drinks and matched for frequency of occurrence (Keuleers et al., 2010) and number of syllables. The original Stroop test involved a congruent, a neutral and an incongruent card respectively. These cards contained 24 words each.

To obtain a measure for response inhibition from the classical Stroop task, the time needed to read the incongruent card was divided by the time needed to read the neutral card (Lansbergen et al., 2007). For the addiction-Stroop test we subtracted time needed to read the neutral card from the time needed to read the alcohol card.

Bogus taste test

Participants rated their thirst for an alcoholic and a non-alcoholic consumption on a ten point scale, and reported their favorite alcoholic and non-alcoholic drink. They were presented with 150gr of 4 different drinks; i) diet coke (Coca-Cola), ii) malt beer (Grolsch 0.5% alcohol), iii) white grape juice (albert heijn), iv) white wine (Garret – vinho verde). Wine was included as a drink since it is often the preferred beverage amongst female drinkers instead of beer. Also it was shown that there are relatively strong gender differences in positive implicit associations towards beer, which are more prevalent in men (Houben et al., 2011). Before the taste test participants were instructed to consume as much or as little as they wished in order to judge the drinks on different aspects (unpleasant– pleasant; tasteless–strong tasting; bitter–sweet; flat–gassy) which they did by means of a digital questionnaire. They were given 10 minutes to complete the test. The questionnaire was not analyzed for this study since the consumed amount was our primary variable of interest. After the participant had left the test room the experimenter weighed how much had been consumed of each drink.

For analysis of the taste test it was necessary to take into account the fact that people may in general tend to drink more from one type of beverage compared to others (i.e. beer compared to wine). This would complicate the combination of consumed amounts of beer and wine into a single score for consumed alcoholic beverages. The same problem applies to combining the cola and grape juice amounts into a non-alcoholic amount. Therefore, consumed amounts for each of the four drinks were transformed into z-scores. The amount of the soft drink of which participants had consumed the most was then subtracted from the amount of the alcoholic beverage from which the subject had consumed the most. Alternatively the average of soft drinks was subtracted from the amount of the beverage they had indicated to be their favorite on one of the questionnaires. Also, for the taste test a number of analyses were performed separately for all drink types. This was done because the used drink types were a deviation from the standard protocol and we did not want to miss any unexpected differences this might cause.

Awareness check

Using open questions we determined whether participants were aware of the critical go/no-go manipulation. Also we asked them to report what they considered to be the go/no-goal of the study. Procedure:

Participants who made it through the screening filled out a series of questionnaires which they accessed online via a link that was send to them. This series was part of the larger test battery used in the study but will not be

reported about here. This study was conducted in a double-blind, parallel-group, placebo-controlled fashion with imbalanced randomization for conditions [2:2:1:1] and was carried out at the Roeterseiland building of the University of Amsterdam.

For each participant two sessions in the lab were scheduled. Lab sessions were performed between 9 am and 6 pm. Sessions were planned preferably on 2 consecutive days and as much as possible at the same time of the day. Duration of both sessions was about 60 minutes. The sequence and arrangement of the various

components of these sessions is displayed in figure 1. After the first subject the length of the booster session was increased from 24 to 48 trials in

order to increase training effects on the taste test. After 7 subjects the length of the assessment sessions was increased from 96 to 144 trials in order to gain more statistical power. One week after the second lab session participants filled out the TLFB questionnaire again and were probed regarding the goal of the experiment by open questions.

Statistical analysis

Four main outcome measures were obtained in this study: i) changes in action tendencies (as measured with the AAGNG task), ii) implicit affective and approach associations (as measured by the IAT), iii) consumed amounts on a bogus taste test, iv) self-reported drinking behavior ( as measured by the TLFB questionnaire), and v) subjective craving (measured on a visual analog scale (VAS)).

The analytical sample that was used for all tasks was 80: 9 participants (4 from exp. push cond., 2 from control push cond., 3 from control pull cond.) were excluded from analysis because of missing data. The main portion of this data loss was caused by short-term blackouts of the power system in our testing room that were caused by technical maintenance elsewhere in the building. No participants had to be excluded for excessive error rates on any task.

The effect of condition on AAGNG reaction times in go trials was tested with a 2 (time: pre-, post-intervention) x 2 (movement: push, pull) x 2 (drink: alcohol, soft drink) x 4 (condition:

experimental, control; push, pull) repeated measures ANOVA. The effect of condition on response accuracy on the AAGNG task was examined using a 2 (time: pre-intervention, post-intervention) x 2 (movement: push, pull) x 2 (drink: alcohol, soft drink) x 2 (go/no-go; go, no-go) x 4 (condition: experimental, control; push, pull) repeated measures ANOVA. No-go trials were also included in this analysis. One-way ANOVA’s were used to test whether the difference between the D600 score on the approach/avoidance and valence IAT’s before and after the intervention differed between conditions. A one-way ANOVA was also used to determine the effect of condition on taste test score. An ANCOVA Figure 1.Overview of procedure during lab sessions

Session 1 * Questionnaires : - TLFB (Carey, 1997) * Stroop task: - classical Stroop - emotional Stroop

* IAT (Greenwald et al., 1998): - positive/negative - approach/avoidance * AAGNG: - assessment session - training session Session 2 * AAGNG: - training session - assessment session * IAT: - positive/negative - approach/avoidance * Inhibition training:

- booster session (training) * Stroop task:

- emotional Stroop * Bogus taste test

was performed to test for conditional differences on TLFB score while controlling for the effect of gender.

Results

For both versions of the IAT task we observed no condition-related differences. However, both before and after the intervention for both IAT’s the size of the bias was different between conditions. All groups which already had a bias at the measurement before the intervention (figure 2) retained their bias after the intervention. Only the experimental pull condition changed from no bias at the first measurement (ME=-0.18, SE=0.07) to a negative bias for alcohol (ME=-0.18, SE=0.06) on the second measurement, t(22) = 2.82, p = 0.01, r = 0.52 (figure 3).

In explaining the pre-existing condition-related differences in bias for both IAT’s, we analyzed both pre- and post-intervention IAT scores separately for subjects that were tested during the morning and for subjects that were tested in the afternoon. For subjects tested in the morning there was no bias on any of the two IAT’s neither pre- nor post-intervention. For subjects tested in the afternoon there was a significant avoidance bias for alcohol both pre- and post-intervention, and a significant negative valuation bias for alcohol both pre- and post-intervention (figure 4).

At pretest only afternoon participants demonstrated both negative valuation and avoidance biases for alcohol, while these biases were absent in morning subjects. These biases were stable between the first and second measurement. This indicates that the time of the day could really influence implicit bias. However, the fact that biases were negative and avoiding in the afternoon was unexpected.

Figure 3. Difference in bias score (D Scores) between pre- and post-intervention

measurements, for both versions of the IAT. Figure 2. Bias scores (D Scores), for the

valence- and approach/avoidance IAT for each condition before intervention. Illustrating pre-existing differences in bias score between

conditions. Figure 5. Reduction of average response time

between first and second assessment session for pull go trials.

Figure 4. Difference in Bias Score on both the

approach/avoidance- and the valence IAT between morning and afternoon participants

At pretest significant positive correlations were found between the extent to which people were craving for alcohol and both the score on the valence (r = 0.29, p < 0.009), and the

approach-avoidance IAT (r = 0.23, p < 0.003). Positive correlations with score on the valence IAT were also found for AUDIT score (r = 0.36, p = 0.001), and score on the TLFB at pretest (r = 0.38, p = 0.001).

AAGNG

On AAGNG reaction times main effects were found of time, F(1, 78) = 119.5 p < 0.001, ηᵨ²=0.61, and movement, F(1,78) = 43.3, p < 0.001, ηᵨ²=0.36. There was an interaction effect for Time x Movement x Condition, F(1,78) = 3.18, p = 0.026, ηᵨ²= 0.11.

Follow-up analyses were then performed for both movement directions separately. The design of the mixed ANOVA was kept the same but Movement was removed as a factor. For pull trials a main effect was found for time, F(1,78) = 76.3, p < 0.001, ηᵨ²= 0.56. An interaction effect was found for Time x Drink F(1,78) = 5, p = 0.049, ηᵨ²= 0.06. For Push also a main effect was found for time, F(1,76) = 105.1, p < 0.001, ηᵨ²= 0.58, but the interaction effect between Time and Drink was absent.

Zooming in on the interaction effect found for pull trials, paired t-tests indicate that

regardless of condition reaction times for pulling in response to soft drink images were significantly more reduced (ME=52.5, SE=5.5), than reaction times for pulling in response to alcohol images (ME=43.7, SE=5.13), t(79) = -2.06, p = 0.043, r = 0.23. Performing these tests separately for each of the conditions it appears that this reduction was only significant in the experimental push condition (soft drink: ME=61.3, SE=10.3 ; alcohol: ME=43.0, SE=10.4), t(17) = -2.51, p = 0.022, r = 0.52 (figure 5).

Looking at the reaction times for the first and second assessment sessions separately it appears that in the first assessment session in all conditions, except for the control pull condition, subjects were faster to make a pull response than to make a push response, irrespective of drink type (figure 6). During the second assessment session this discrepancy between pull and push responses was absent in both the control pull condition and the experimental pull condition. It was maintained in both push conditions (figure 7). In the experimental pull condition participants were now faster to make a push response to soft drinks (ME=656.1, SE=10.78) than to alcohol (ME=668.7,

SE=11.25), t(22) = -2.759, p = 0.011, r = 0.51. And in the control pull condition participants were faster to make a pull response to soft drinks (ME=665.7, SE=17.22) than to alcohol (ME=680.9,

Figure 7. Difference in reaction time between push and pull trials during the second

assessment session Figure 6. Difference in reaction time between

push and pull trials during the first assessment session

SE=14.30, t(17) = -2.587, p =0.019, r = 0.53.

In detailed comparisons between the first and second assessment sessions it is found that all effects are superseded by a general practice effect. Reaction times during the second assessment were found to be faster in general (ME=669.4 , SE=6.07) than in the first assessment session (ME=722.0 , SE=6.23), t(78) = 11.825, p < 0.001. Next difference scores were calculated to represent the difference in reaction time between pull trials and push trials. These difference scores were calculated for both drink types separately and for both assessment sessions separately. They were compared using paired t-tests. We found that the difference in reaction time between push and pull responses decreased significantly from the first to the second assessment session in the experimental pull condition specifically. This was true for both drink types (alcohol: t(22) = 2.177, p = 0.04, r = 0.42 ; soft drink: t(22) = 2.177, p = 0.025, r = 0.42). A decrease for the difference score between pushing and pulling alcohol was also found for the control pull condition (t(17) = 2.558, p = 0.02, r = 0.53).

The above described comparisons were also performed using the alternative reaction measures; i) movement duration, and ii) point of maximal acceleration. No noteworthy deviations from the results described above were found for these alternative measures.

After looking at the reaction measures for the go trials, the accuracy of responding on the AAGNG task was also analyzed. Main effects were found for Time, F(1,78) = 72.2, p < 0.001, ηᵨ²= 0.48, Movement, F(1.78) = 12.3, p = 0.001, ηᵨ²= 0.14, Drink, F(1,78) = 4.98, p < 0.05, ηᵨ²= 0.06 and Go/no-go, F(1,78) = 104.9, p < 0.001, ηᵨ²= 0.57. All interaction effects were superseded by the Go/no-go factor. Therefore follow-up analyses were performed for go and no-go trials separately.

For go trials there was a main effect for movement direction for alcohol trials (F(1,78) = 12.8, p = 0.001, ηᵨ²= 0.14), but not for soft drink trials. Paired t-tests reveal that this accuracy difference for trials requiring either a pull or a push response arose because accuracy for pushing alcohol stimuli was significantly lower (ME=0.95, SE=0.01), than for pulling alcohol stimuli (ME=0.98, SE=0.003), t(80) = -4.87, p < 0.001, r = 0.48. This was only so for the go trials.

Using a paired t-test it was found that subjects are less accurate in withholding a pull

response to alcohol stimuli (ME=0.99, SE=0.003) than withholding a push response to alcohol stimuli (ME=0.995, SE=0.002), this difference was significant, t(80) = 3.58, p = 0.001, r = 0.37. The same did not hold true for soft drink stimuli.

Overall subjects responded faster and more accurately post- than pre-intervention, but there were no differences between conditions. More specifically, subjects from the experimental push condition were the only ones to significantly improve more on pulling soft drink stimuli relative to pulling alcohol stimuli. Accuracy testing indicated that subjects are more accurate in pulling alcohol stimuli than pushing alcohol stimuli, regardless of condition. Also regardless of condition subjects are more accurate in withholding a push response to alcohol stimuli than a pull response.

Taste test

Since for this study a number of deviations from the standard protocol had been made, we performed a number of other tests to validate our protocol. Importantly a strong positive correlation was found between AUDIT score and taste test score (r = 0.33, p = 0.003) (figure 8), and between response inhibition, as determined using the classical Stroop test, and taste test score (r = 0.29, p = 0.009). Also in general the quantity of consumed beer correlated positively with craving (r =0.33, p = 0.002), and response inhibition (r = 0.25, p = 0.026), while the quantity of consumed wine did not correlate with any of these variables. Comparable results were found when analyses were split for preferred type of alcoholic drink, indicating that also specifically in participants who prefer to drink wine, there was no correlation between consumed amounts of wine and any of the three above named variables. Because of the expected effect of gender we split the analysis on these comparisons for gender. This revealed that in women craving score was a good predictor of consumed amounts of beer (r = 0.33, p < 0.017), but not for consumed amounts of wine. In men craving score did not

correlate with consumption of either of the two types of alcoholic beverages. For men there was a strong correlation between response inhibition and the taste test summary score (r = 0.50, p = 0.006), while this correlation was absent in women.

Concerning the summary taste test score, no significant differences were found between the experimental and the control groups, nor between any of the separate conditions. This also held true for the alternative calculation of the taste test score we performed based on preferred drink type. Next we sought to examine whether the time of the day on which subjects were tested could have

influenced the results. This could be because it was found that subjective craving was significantly higher in the afternoon than in the morning (F(1,78) = 13.52, p < 0.001, ω = 0.20). Therefore the same one-way ANOVA was also performed separately for subjects who performed the taste test in the morning and subjects who performed the test in the afternoon. However, in the morning as well as in the afternoon significant differences between conditions were absent. Splitting our analysis for gender did also not return any significant results. Performing the above mentioned analyses separately for all drink types did also not yield any meaningful results. It is however noteworthy that for morning participants there were positive correlations between taste test score and: i) craving, ii) response inhibition, and iii) AUDIT score. The same three positive correlations were found for consumed quantity of beer. For afternoon participants all these correlations were absent. Also for morning participants there was a trend for a positive correlation between score on the valence IAT and consumed amount of beer, but this did not reach statistical significance (r = 0.33, p = 0.07).

Stroop test

For the classical Stroop task the ratio between the time needed to read the neutral and the incongruent card was used as a measure for inhibitory control. This means that the more time people needed to read the incongruent card relative to the neutral card, the larger the measure for response inhibition becomes. There was a strong correlation between this ratio and AUDIT score (r = 0.39, p < 0.001), indicating that people who scored higher on the AUDIT took relatively longer to read the incongruent relative to the neutral card than people with a lower AUDIT score.

For the emotional version of the Stroop task it was found that in the first session there was a significant correlation between AUDIT score and time needed to read the card with alcohol words (r = 0.28, p < 0.05), as well as time needed to read the neutral card (r = 0.23, p < 0.05). It was also found that in the second session subjects were significantly faster to read both the alcohol and the neutral card, regardless of condition. However, there were no significant differences in improvement on either of the cards between conditions. The same holds true when looking at the time difference between the alcohol and the neutral card (neutral –

alcohol). Besides AUDIT score, the measure for inhibitory ability that was derived from the classical Stroop task was also found to correlate strongly with time needed to read the alcohol (r = 0.25, p < 0.05), as well as the neutral card (r = 0.32, p < 0.01), but did not correlate with the time difference between the alcohol and the neutral card.

TLFB

Figure 8. Correlation between AUDIT Score and Taste Test Score, r = 0,33, p = 0,003.

Regarding alcohol use, there was an effect of gender on alcohol consumption in the week before the experimental intervention (F(1,78) = , p < 0.049, ηᵨ²= 0.22). Therefore in the analysis of the effect of the intervention on alcohol consumption, gender was included as a covariate. No significant differences were found for any of the conditions in average alcohol consumption before and after the training, nor for the difference in consumption between the two measurements. Also, regardless of condition alcohol consumption slightly increased from the first (ME=2.03, SE=0.15) to the second (ME=2.38 , SE=0.2) measurement, t(79) = -2.33, p < 0.023, r = 0.25 (figure 9). There was a strong significant correlation between AUDIT score and average consumption before and after the

experimental manipulation. The same holds true for the inhibition measure derived from the classical Stroop task. Partial correlations were therefore performed to isolate their unique contributions. When controlling for AUDIT scores, the correlations between alcohol consumption and response inhibition were not significant anymore. Controlling for inhibitory ability AUDIT score still was a significant predictor of alcohol consumption, both before (r = 0.57, p < 0.001), and after the intervention (r = 0.57, p < 0.001).

Awareness check

Analysis of the awareness questions indicated that two subjects showed considerable awareness of study objectives and the goal of our manipulation. They were therefore excluded from analysis. Apart from these individuals, participants did not show awareness of either the manipulation or goal of the study.

Discussion

IAT:

During our research we encountered two possible limitations to the interpretability and validity of the results from the IAT. First, as described in the method section we aimed to

counterbalance the order in which participants received the combined classification blocks. This was done because previous research indicated that the size of the bias might be dependent on the presentation order of the blocks (Greenwad et al., 2003). Therefore we tested whether IAT scores differed between the two presentation orders with a one-way ANOVA. For the valence and the approach-avoidance IAT respectively p-values were 0,89 and 0,97, with corresponding F-values F(1,78) = 0,020 and F(1,78) = 0,001 Therefore we do not expect that the incorrect counterbalancing will cause problems for the interpretability of the data. A second issue concerns the used set of words for the valence IAT. Erroneously ‘koffie’ (coffee) and ‘thee’ (tea) were part of the category of words that had to be labeled as ‘soda’, while this is theoretically incorrect. When given the choice between ‘soda’ and ‘alcohol’ people will tend to classify coffee and tea as ‘soda’ rather than ‘alcohol’, but still the fact that it is not theoretically correct might have caused general slowing of reaction times for these items. This slowing might have blurred results from the valence IAT. In a later stage IAT results could be reanalyzed after removing ‘koffie’ and ’thee’ trials.

Pre-test bias scores for the approach-avoidance IAT indicate that both the experimental pull group and the control push group exhibited an avoidance bias for alcohol while the control pull group showed an approach bias. For the valence IAT only the control push group had a negative bias for alcohol at pre-test. These condition-related differences in implicit biases at pre-test are a

complicating factor for interpretation of the IAT data, and also for making comparisons between conditions. This will be discussed in more detail in the general discussion.

Comparing pre- and post-measurements of both versions of the IAT, no main effects were found for condition. Looking at the conditions separately only the experimental pull condition

Figure 9. Average daily consumed number of standard drinking units of alcohol, both before and after the intervention.

changed from no bias on the valence IAT at pre-test, to a negative bias for alcohol post-test. This result is in line with our expectations and therefore could be interpreted as being the result of the repeated resolving of response conflicts during the training sessions. Supporting this claims is the fact that this change in implicit attitude was not observed in the experimental push condition where repeated inhibition of a response to alcohol stimuli did not lead to a response conflict. For the approach avoidance IAT no changes in implicit attitude were observed for any of the conditions. Stil this result fits reasonably with our expectations. First, because we did not expect approach bias to change in the experimental push condition. And second, because the experimental pull group, in which we expected implicit bias to become more avoiding, already exhibited a strong avoidance bias at the measurement. We did not expect this avoidance bias to be present at the

pre-measurement, but given that it is there we would expect it to be maintained at the post-measurement.

Searching for an explanation for the condition-related differences in pre-existing biases we found that regardless of condition, biases only became apparent in subjects that were tested in the afternoon. But if this effect found for time of day is to explain differences between groups, then by chance in certain groups more participants should have been tested during either the morning or the afternoon. But between conditions the ratio of participants tested during the morning and during the afternoon was roughly equal. In itself the discrepancy between IAT scores in the morning and the afternoon is an interesting finding that could be seen as an indication that implicit bias is not always stable and that contextual factors can influence implicit activation. Further implications of this idea for the study will be discussed in the general discussion.

Taken together, bias scores for both versions of the IAT seemed to depend upon the time of the day at which participants were tested. Only the experimental pull condition developed a negative bias for alcohol between pre- and post-test measurements, which is in line with response conflict hypothesis.

AAGNG:

The AAGNG task entails a combination between the concepts of approach-avoidance and go/no-go tasks. Since it is a newly developed task it is important to determine whether it functions properly. Its foremost function is to provide differential possibilities for experimental intervention to modify or generate implicit associations. Besides, reaction times and response accuracy both pre- and post-test can function as an additional measure to validate the effects of the experimental

intervention. Therefore we also looked carefully at the consistency of the measures derived from the AAGNG task with the other measures we obtained.

Concerning the reaction times for the first assessment session it was expected that subjects would be faster to make a pull response to alcohol stimuli than a pull response to soft drink stimuli, but this was not observed in any of the conditions. This was expected since the AUDIT score that was used as an inclusion criteria for this study is usually accompanied by an approach bias for alcohol. This implicit tendency to approach alcohol relative to soft drink stimuli is then expected to be reflected in reaction times on the AAGNG task. Splitting this analysis for the high and low AUDIT group did also not reveal relatively faster response times for alcohol pull trials in either of these groups.

For the first assessment session, the only trial types for which reaction times were

significantly different were the trials for which opposing joystick movements were required (i.e. push vs. pull). We interpret this as a reflection of the fact that both alcohol and soft drink stimuli are positively valuated and therefore it is easier for participants to make a pull response than a push response. Alternatively it could be argued that the fact that reaction times are slower for push responses is because it is in general easier to make a pull response than a push response. However, this explanation does not suffice since previous research has shown that what determines reaction time is not the movement per se, but the evaluation of the movement as being positive or negative. According to the evaluative response-coding framework, which states that approach and avoidance are evaluated at a mental representational level, absolute movement direction does not matter (Eder