Incorporating Neuropsychological knowledge in intervention techniques : the effectiveness of emotion perception training on the reduction of aggression of offenders

U

A

I

N

K

I

T

;

E

E

P

T

ON THE REDUCTION OF

A

O

.

___________________________________________________________________________

M

Clinical Forensic Pyschology and Clinical Neuropsychology

AUTHOR

Name Anna van der Schoot Student number 10172815

SUPERVISORS

Within program group Lieke Nentjes External supervisor Niki Kuin

RESEARCH FACILITY Penitentiare Inrichting, Vught; Penitentiary Institution, Vught DATE 13-08-2016

A

The aim of the present research was to see whether emotion perception training could reduce aggression in violent offenders. Limited research has focused on investigating the effectiveness of the training that encourages the perception of happiness over anger in

ambiguous expressions in the target group of adult offenders. The present research was able to demonstrate that aggressive offenders exhibit the tendency to attribute hostile intent to ambiguous emotional expressions of others. Subsequently, it was found that this Hostile Attribution Bias (HAB) could be related to aggression in offenders directly. Results showed that with the implicit method of targeting the bias in emotion perception, HAB could be altered with subsequent positive effects for aggressive behaviour. An important find was that brain damage seemed to moderate the training effects. The findings underline that to reduce aggression and violent recidivism, it is essential to incorporate fundamental neuropsychological knowledge in intervention techniques.

Introduction

The release of offenders back into society is a sensitive topic. A particular pressing issue is the high rate of recidivism among offenders. In exact figures a staggering 47.7% of

ex-offenders re-offend within two years of their release. The biggest and most concerning issue is that 14.6% of these crimes committed by re-offenders is accompanied with violence (Ministerie van Veiligheid en Justitie, WODC, 2012). Following from this, penitentiary institutions try to reduce the violent recidivism numbers with use of standardized treatments. Some prove to be effective, such as Enhanced Skill Thinking (EST; Dutch: Cognitieve Vaardigheden Training; COVA) a form of cognitive behavioral therapy (CBT) (Buysse & Loef, 2012). However, there still is a large proportion of offenders that are not able to change as can be concluded from the remaining recidivism rate.

The current judicial interventions are mainly based on assumptions from social science and law. Kogel (2008) addresses that the problem with current behavioral interventions for antisocial behavior, hence violence and aggression, is that they are not focused enough on individual differences and needs. Most importantly, Kogel’s point is that in current treatment protocols there is a lack of incorporation of knowledge about neuropsychological processes that contribute to violent and aggressive behavior in offenders.

In the present research a new method for changing aggressive behavior was studied. In support of Kogel’s claim, the current study focused on a training that is based on assumptions from a neuropsychological perspective on aggression in offenders. In particular, it was tested whether this method could effectively be used to reduce aggressive behavior in a prison population in the Netherlands.

Aggression may be thought of as any behavior that is executed with the proximate goal of causing physical or psychological harm to another individual (Tapscott, Hancock, & Hoaken, 2012). Furthermore, a violent offense is defined as an offense that is accompanied by a high level of aggression (Tapscott et al., 2012). Aggression can therefore be seen as an evident behavioral component of violent offending. A current explanation for aggressive behavior is that aggressive individuals perceive others as more hostile and therefore respond with aggression. The intent to perceive others as more hostile is called “hostile attribution bias” (HAB) and is especially apparent when situations are highly uncertain and ambiguous (Dodge 2006; cited by Schönenberg & Justyte 2014a).

Social cognition, the processing of relevant social information, is key in guiding social behavior. It encompasses how we perceive, interpret and react to others (Adolphs, 2009). The social information-processing model of aggressive behavior, describes how the attribution of hostile intent to peers leads to maladjusted social behavior in reactive aggressive children (Crick and Dodge, 1994). Based on studies of the role of social information processing in aggression Dodge and Schwartz (1997; cited by Schwartz et al., 1998) showed that misinterpretation of

ambiguous emotional cues results in inappropriate social responses, such as acting aggressively of violently. Following from this, aggressive individuals might experience more ambiguity in social interactions due to difficulties with perception of emotions and therefore interpret others’ intent as hostile, adversely leading to a more aggressive approach to others. The misperception of emotions might play a key role in aggression. In support of this claim, Marsh and Blair (2008) found that antisocial individuals have difficulty identifying several basic emotions such as sadness and fear, but not anger. Moreover, aggressive individuals are more sensitive to facial expressions of anger than less aggressive individuals (Wilkowski & Robinson, 2012). Subsequently, in a study by Schönenberg and Justyte (2014a) it was found that antisocial violent offenders were not only more prone to detect aggressive facial expressions in an ambiguous morphed continuum than healthy controls, they also rated facial expressions as more hostile. These results suggest that more aggressive individuals not only perceive others as more hostile, they preferentially infer hostile intent from others. The tendency to misperceive emotional facial cues in social

interactions and infer hostile intent from others (HAB) possibly underlies aggressive behavior in violent offenders.

The assumption that aggressive individuals, especially offenders, respond more

frequently -and are more sensitive to angry emotions is also supported by neurological evidence. In a study by Pawliczeka et al. (2013), brain activity was measured with FMRI during a facial expression version of the stop-signal task. The stop-signal task is a classic approach to measure response inhibition: a previously learned response has to be inhibited when a probe is presented. It was found that when participants had to respond to pictures of facial expressions, relatively aggressive individuals had a faster response to angry faces in comparison to low aggressive individuals. Furthermore, highly aggressive individuals had difficulty to withhold responses for angry faces when the stop signal was presented. This response behavior was related to higher activity in the right pre-supplementary motor area (SMA), the right middle frontal (MF), and the motor cortex (MC). Therefore, brain activity in these regions seems to be related to response behavior in aggressive individual. Subsequently, response inhibition improved during anger trials in both groups. This gives insights in the possibilities for incorporating neuropsychological knowledge in behavioral interventions for aggression, since it suggests that the related brain regions might change through repetitive training.

Activation of the neural pathway (SMA-MF-MC) during the recognition of facial expressions is related to a faster -and more intense signal to the muscles to prepare for action. Impulsivity and inhibitory problems in this brain pathway can therefore be related to aggression. Also, the improvement of response inhibition after repetition suggests aggressive behavior can be changed with repetition training aiming on the reaction to angry faces. This indicates that training targeting the perception of emotions might be a good method to facilitate change on underlying structures linked to aggression. Therefore, it might be a more effective method for tackling

problem behavior than the current non-neurological based behavioral treatment methods, as Kogel (2008) predicts.

In support of this idea, Schönenberg and Justyte (2014b) were able to demonstrate that a brief computerized training protocol aimed on recognition of the six basic emotions (anger, disgust, fear, happiness, sadness and surprise) was sufficient to significantly improve recognition of facial affect in violent offenders directly after the training. Furthermore, Penton-Voak et al. (2013) showed that promoting recognition of happiness over anger in ambiguous facial

expressions reduced anger and aggressive behavior in a youth offender sample. Not only did this training improve the perception of faces as happy (non-hostile), it also had a direct effect on self-rated aggression and aggressive behavior self-rated by the staff that lasted for at least two weeks. Subsequently, in a very recent study, Stoddard et al. (2016) replicated these results in a group of youth with Disruptive Mood Dysregulation Disorder (DMDD). DMDD is often related to high irritability, accompanied with acting out in the form of aggression. Using the same happy-angry training paradigm as Penton-Voak et al. (2013), this research showed that DMDD youth, compared to healthy controls, rated ambiguous faces as more angry rather than happy. Active training was associated with a shift towards more “happy” judgments for both groups and lower levels of irritability accompanied this reduction of HAB in de DMDD group. More importantly, this reduction was associated with changes in neural activity in the orbitofrontal cortex (OFC) and amygdala.

These results strengthen the evidence of a causal role between the perception of emotion and aggressive behavior. More importantly, it shows that simple emotion training techniques based on neurological evidence could facilitate behavioral changes. This sort of training can be administered with speed and ease at a low level of instruction or action of the trainer and might be a cost-effective means to reduce aggressive behavior in target populations.

The strength of the emotion perception training also lies in its implicit way that it targets the underlying cognitive processes. Research on implicit measures to evaluate

aggression/aggressive tendencies such as the Implicit Association Test (AIT, Greenwald, McGhee, & Schwartz, 1998) proved that this is of additional value to explicit measures such as questionnaires (Uhlmann & Swanson, 2004), since these measures are able to assess the

underlying implicit cognitive constructs that contribute to aggression. Although implicit measures have proven their value, there never has been an implicit intervention that targets the underlying cognitive constructs. The emotion perception training is unique in the way that is the first and only intervention technique that targets aggression on a more implicit level. Therefore, it was of importance to test this intervention for clinical application in the target population of violent offenders.

The previous studies suggest that aggressive behavior might be causally related to biased emotion recognition and, subsequently, emotion training can positively modify this hostile

attribution bias (HAB) on both behavioral and neuronal levels. To find better means of

interventions that aim to reduce violent offending, it is important to see if this sort of training is applicable for clinical administration in a forensic setting. Therefore the main purpose of this research was to investigate the effects of emotion perception training on aggression in offenders. The above research setup by Penton-Voak et al. (2013) was replicated in a Dutch adult prison institution. The study aims were to determine whether (1) Hostile Attribution Bias (HAB) is related to aggression in the present sample of incarcerated offenders and (2) if an emotion recognition training procedure designed to promote the perception of happiness over anger in ambiguous emotional expressions results in a change in self perceived and observer rated anger and aggressive behavior (3) if specific offender characteristics account for differences in training effectiveness.

A pattern that is consistent with the hostile attribution bias is that aggressive individuals interpret neutral and ambiguous facial expressions more negatively (Schönenberg and Justyte, 2014a; Penton-Voak et al., 2013; Stoddard et al., 2016). Therefore, it was first expected that aggression as behavioral component is positively related to the tendency to rate ambiguous facial expressions as “angry” as opposed to “happy” in a forced choice paradigm.

Second, it was expected that an emotion recognition training procedure designed to promote the perception of happiness over anger in ambiguous emotional expressions will result in a change in self perceived and observer rated anger and aggressive behavior (Penton-Voak et al., 2013). In this study emotion recognition was altered by giving participants biased feedback. This feedback is based on a “balance point”, which is a calculation of the number of “happy” responses in a categorical choice paradigm (happy vs. angry) as a proportion of the total number of trials. This gives the balance point at which each participant is equally likely to perceive happiness or anger. The expectation was that, after receiving biased feedback (intervention condition), participants would rate ambiguous faces as happy instead of angry more readily. Furthermore, it was expected that the participants who receive biased feedback, showed a decrease in pre-training vs. post-training aggression on both self-report and staff-rated measures. Participants in the control condition would show no change on pre-training vs. post-training measures of aggression. In line with the assumption that the emotion perception training also promotes changes on a neural level, it was expected that these effects would last over an extended period of time.

Lastly, it was hypothesized that certain population characteristics can cause different training outcomes. In line with the reasoning that follows from the study by Pawliczeka et al. (2013), high aggressive individuals show different neural activity in reaction to angry faces as opposed to low aggressive individuals. If emotion perception training facilitates neural changes and subsequent behavior, it can be reasoned that intended change can be difficult if brain structures are damaged. It was therefore expected that differences in task performance could be

related to brain damage. Brain damage was therefore included as a moderating factor in the relationship between treatment (control vs. intervention) and aggression. Furthermore, for some people it is hard to interpret social cues because of mild intellectual disability (ID). Little is known about emotion recognition and ID (Wishart, Cebula, Willis, & Pitcairn, 2007), but since social impairment is characteristic for ID, it was expected that ID would have an influence on training effect. A small number of studies have suggested that syndromic differences in emotion recognition skills may indeed be present in children with ID (Kasari & Sigman 1996; Turk & Cornish 1998; Wishart & Pitcairn 2000; Kasari et al. 2001; Gagliardi et al. 2003; Williams et al. 2005; cited by Wishart et al. 2007), although the evidence at older ages is currently much more limited (see e.g. Mazzocco et al. 1994; Simon & Finucane 1996; Plesa-Skwerer et al. 2006; cited by Wishart et al, 2007). Therefore, ID was another moderating factor that was considered when looking at the effect of emotion training on aggressive behavior.

Method

PARTICIPANTS

A-priori analysis using the program G*power, showed that when computing the required sample size given alpha (α=.05), power (1 – β=.80) and effect size (d=.15) an n of approximately

68 would be needed to obtain statistical power at the recommended .80 level.

One hundred and sixteen male adult incarcerated offenders serving their sentence in the Dutch Penitentiary Institution (PI) Vught, participated in this study. They all were convicted for violent crimes at that moment of testing or in the past. Participants were not included if they suffered from a major and currently apparent psychiatric disorder (e.g. major manic, depressive, or

psychotic episode or an autism spectrum disorder). The intervention took place in groups varying in size from 10 to 15. During the intervention, participants could not take part in any other treatment for aggression, with exception for medicated treatment (which will be monitored). Participants did not receive compensation for their participation, other than a break from daily routine.

I

Emotion perception training instrument

The emotion perception task: The emotion perception task that was used in this study was developed

by Penton-Voak et al. (2013). Prototypical happy and angry composite images were derived from 20 individual male faces showing a happy facial expression and the same 20 individuals showing an angry expression. The original images came from the Karolinska Directed Emotional Faces (Lundqvist, Flykt & Öhman, 1998; cited by Penton-Voak et al., 2013). These prototypical images were used as endpoints to generate a linear morph sequence of 15 stimuli that consist of images

changing incrementally from unambiguously happy to unambiguously angry facial expressions, with emotionally ambiguous expressions in the middle.

The task has two goals: (1) estimate the balance point at which participants are equally likely to perceive happiness or anger on a morphed continuum (a balance point more towards angry represents the possibility of HAB) and (2) to alter this balance point towards the happy spectrum of the morphed continuum. The task is designed in a way that in the intervention condition, the feedback received by the participant is biased, while in the control condition it is not. In the control condition feedback is directly based on the participant’s baseline balance point. In this case, responses were classified as “correct” when the participant identified images below the original balance-point as happy, and faces above that point as angry. Otherwise a response was classified as “incorrect”. In the intervention condition, feedback was also based on the participant’s baseline balance point, but the feedback that was received was biased: the “correct” feedback is shifted two morphed faces towards the happy end of the spectrum; that is, the two images nearest the balance point that the participant would have classified as angry at baseline, were classified as happy in the feedback phase. A participant in the intervention

condition would in this case receive the feedback “incorrect” if he classified a face as angry when it fell within a two-morphed step reach of the original balance point.

In the Penton-Voak et al. (2013) study, training effects were related to outcomes on aggression measures (self-report and staff rated); this task was successfully used to determine what the balance point is and it also identified that altering this point has positive effects on aggressive behaviour. These findings support that the reliability, validity, and therefore the applicability of this instrument for the purposes of this study.

Instruments for assessment of aggressive behavior

Several measurements of aggression were administered in this study to determine training effects over time. The instruments used are:

1. Observation Scale for Aggressive Behavior (OSAB): The OSAB is a tool to measure the observed

behavior of the inmate on the ward (by ward staff). Though standardized self-report

questionnaires have high internal consistency and are stable over time, social desirability is found to have a moderately high negative relationship with the aggression scales suggesting that social desirability may influence responses provided by aggression questionnaires (Harris, 1997). Therefore, it is important that the participant’s behaviour is assessed from other’s perspectives. The OSAB is a 40 item observational list that consists of the following subscales:

irritation/anger, fear/depression, aggression, prosocial behavior, triggers and sanctions. Behaviour is rated on a 4 point likert scale: 1 = ‘never’, 2 = ‘seldom’, 3 = ‘sometimes’ and 4 = ‘often’. This scale has good internal consistency and inter-rater reliability, as well as convergent validity (Hornsveld, Nijman, Hollin, & Kraaimaat, 2007). The internal consistency of the

subscales (Cronbach's α) were, respectively, .82, .79 and .93, the inter-rater reliabilities are .79, .81 and .70 and test-retest reliabilities are .59, .57 and .76 in a group of 220 inpatients (Hornsveld et al., 2007). Convergent validity was supported by significant correlations with corresponding subscales of the Forensic Inpatient Observation Scale (FIOS; Carpenter, Vastenburg, & Emmelkamp, 2001; cited by Hornsveld et al., 2007) and self-report questionnaires measuring hostility, anger and aggression (Hornsveld et al., 2007), which are considered to measure the same concepts. For the purpose of this study, only the scales irritation/anger and aggression were analysed. The internal consistency of the questionnaire in this study could not be calculated since only total scores were registered for data analysis, not the scores on separate items, which are needed for the calculation of reliability.

2. Social Dysfunction and Aggression Scale, self-report (SDAS-11-sr): The SDAS-11 was developed by

Wisted et al. (1990) and translated to Dutch by van der Werf (1997). It is a self-report

questionnaire, consisting of 11 items scored with a 5-point likert scale (in order from 1 to 5): ‘not present’, ‘sometimes present’, ‘regularly present’, ‘often present’, ‘almost always present’. The questionnaire takes 5 minutes to complete. The SDAS-11 was originally an observational instrument. For the present study it was rephrased so it could be used as a self-report questionnaire.

The SDAS-9 consists of 9 items covering outward aggression and 2 items (SDAS-2) covering inward aggression. In a study by Wisted et al. (1990), it was found that convergent validity was good: in a pilot study on 82 inpatients from different centres in Denmark and Sweden, the SDAS was compared to 3-item scales for outward and inward aggression and to a global scale for outward aggression. The results showed that the SDAS-9 correlated positively with the other outward observer-scales, and the SDAS-2 correlated positively with the other inward scale. A divergent validity was seen between the outward and inward scales, indicating that it is necessary to measure both dimensions (Wisted et al., 1990). The inter-observer reliability was also found to be adequate (ERAG, 1992). For the purpose of this study, the SDAS was administered every week of the intervention period (11 weeks). The SDAS was selected for its brevity (which makes it easy to administer) and its applicability for repeated measures within a short interval. The internal consistency of the questionnaire in this study could not be calculated since only total scores were registered for data analysis, not the scores on separate items, which are needed for the calculation of reliability.

3. Novaco Anger Scale – Provocation Inventory (NAS-PI): Because the SDAS measures the behaviour

of the past week, a measurement of self-perceived aggression in general is also used. The NAS-PI (Novaco, 1994) is a self-report questionnaire consisting of 73 items that are divided into two components: the NAS scale (48 items focus on self-perceived anger in three dimensions:

cognition, arousal, and overt behaviour). Answers are given on a 3-point likert scale: ‘never true’, ‘sometimes true’ and ‘always true’. The second scale (PI) is consists of 25 items assessing

sensitivity for provocation. Participants consider what degree of anger they would experience in certain proposed situation, options are: “not angry at all”, “a little bit angry”, “angry” and “very angry”. Both scales have good internal consistency and test-retest reliability (Hornsveld, Muris, & Kraaimaat, 2011). The internal consistency of the questionnaire in this study could not be calculated since only total scores were registered for data analysis, not the scores on separate items, which are needed for the calculation of reliability.

4. Aggression Questionnaire -short version-: (Agressie Vragenlijst, Aangepaste Versie: AVL-AV): Lastly, to

give a better view of different sorts of aggression and also to add more credibility to the overall measure of aggression, this extra questionnaire will be used. The AVL-AV was developed by Buss and Perry (1992) as the Aggression Questionnaire and translated for the Dutch population by Meesters et al. (1996). The AVL is a self-report questionnaire, consisting of 29 items with a 5-point likert scale. Scores are distributed into 4 factors: physical aggression, verbal aggression, rage and hostility. In a Dutch population of aggressive offenders, the psychometric characteristics appeared worse for the AVL than for the shorter 12-item form, the AVL-AV (Hornsveld et al., 2009). Therefore, in this research the 12-item AVL-AV was used, which has the same subscales as the longer AVL. The internal consistency of the questionnaire in this study could not be calculated since only total scores were registered for data analysis, not the scores on separate items, which are needed for the calculation of reliability.

Tools to assess sample characteristics

1. Semi structured interview: This interview was developed for a previous study conducted by

external supervisor and head researcher N. Kuin. It is a semi-structured interview administered by the researcher. It takes 10 minutes to complete. The participants are asked to answer questions regarding their age, highest completed level of education, current medication, drug abuse (defined as weekly use for at least 3 months), contact with a psychologist and current treatments. Prior brain damage is assessed by asking questions regarding head trauma (“did you ever get hit in the head?” “did you ever blackout”? and “have you been in a coma?”), further questions on hospitalization for head trauma, any diagnosis related to head trauma and long term drug abuse. Group differences for possibility of brain damage vs. no brain damage, was assessed to see whether training effects were dependent on brain damage.

2. Justicial Records: These records are used to obtain information about the participant’s criminal

history (based on prior convictions), the index offence and current regime. Regimes can vary depending on the sort of conviction or the personal offenders needs. In this study this can vary

between a “regular regime”, a regime for “systematic offenders”, “high care” regime or a

psychiatric ward (for offenders with psychological problems or offenders who are too vulnerable to participate in regular regimes). Self-report on this topic is often difficult for participants, especially when asked for within the judicial context. Therefore, checking the criminal record provides the most factual information about criminal behaviour and whether it is aggressive in nature.

3. Screening Checklist Intelligence (SCIL): To detect intellectual disability a screener for intelligence

and mild intellectual disability is used, in Dutch called: ‘screener voor intelligentie en licht verstandelijke beperking’ (SCIL; Kaal, Nijman & Moonen, 2013). The outcome of the test is a statement regarding the possibility that the examinee has a mild intellectual disability. It differentiates between below or above 85 IQ-points. The cut-off score for ID or no ID is 19 (below represents ID). Kaal (2013) tested its applicability for detecting ID in a forensic

population. It was determined that to screen for the possibility of the presence of ID, the use of 14 items gives the most differentiating information. The items include questions about education, special mental care, contact with family and reading behavior. In addition, the respondent is asked to perform tasks in the areas of math, reading, writing, spelling and clock drawing. The reliability of the 14-item scale is 0.82 in terms of Cronbach's alpha (n = 318 forensic population).

Nunnaly (1978) has indicated 0.7 to be an acceptable reliability coefficient, a Cronbach's alpha of 0.80 and higher is the generally considered to be close to optimal. . The 'predictive' value of the instrument to accurately estimate either the presence or absence of ID was examined with ROC analysis. The AUC values are fairly high (respectively 0.92 and 0.93). The sensitivity is 88%; which means almost 9 out of 10 people with ID actually are classified as ID. The specificity is 83%, which means that more than 8 out of 10 people without ID are also classified as people without ID. In Kaal’s study (2013), the test-retest reliability is 0.92 (p = 0,05, n = 33). Kaal

underlines that the instrument is a screener. To be certain of the presence of ID, additional diagnostic examination is needed. In the present study no further diagnostic examination will be done, therefore the term ID in this study not absolute, and will thus have to be read as the “possibility that ID is present with a certainty of 90% (9 out of 10)”.

P

Before any participants were recruited, the psychologist was consulted to have an overview of the psychological condition and information about the treatments that were taken part in. A list was made of who could be approached for participation. The researchers (also referred to as trainers) individually approached the remaining potential candidates. The

candidates were informed about the study in writing and person. After a consideration time of at least 24 hours informed consent forms were signed.

The total duration of the time frame in which participants completed the whole test program was 11 weeks; this included four weeks of pre measure administration, one week in which the emotion perception training took place, and 6 weeks of follow-up post-measure administration. After the participant gave notice that he was willing to participate, the semi-structured interview and the SCIL were administered during an intake. For the following four weeks the participants and the staff filled in the SDAS and the OSABs. Then, in the fifth week, the intervention (training) took place. Table 1. gives an overview of all the measures that were used for the total time that each participant took part in the study.

Table 1. Administration of each questionnaire during the study time frame

During a period of five days (Monday to Friday) the training sessions took place on a daily basis. The training was conducted in groups varying from 10 to 15 participants. Each training session took about 30 to 45 minutes to complete. Participants individually sat behind a Time-path Week number 4 weeks prior to intervention 1-4

During intervention Follow-up Final follow-up

testing Day 1 Day 1, 2, 3, 4, 5 5 Day 5 + 5 weeks 6-10 + 1 week 11 Instrument Emotion recognition task Happy - Angry 3-phased training procedure (daily) 3-phased training procedure (daily) 3-phased training procedure (daily) Baseline measurement (from training procedure) OSAB*

(staff rated) Once a week inventory Once a week inventory Once a week inventory Once a week inventory

SDAS-11 –sr2

(self report diary) Once a week inventory Once a week inventory Once a week inventory Once a week inventory Official records of

rule breaking and aggressive behaviour (‘reports’)

Once a week

inventory Once a week inventory Once a week inventory Once a week inventory

AVL-AV3 x NAS-PI4 x x x SCIL5 Semi-structured interview x x

*_{OSAB=Observation scale of aggressive behaviour,} 2 _{SDAS=Social dysfunction and}

aggression scale, 3_{AVL-AV=Agressie Vragenlijst, Aangepaste Versie,} 4_{NAS-PI=Novaco}

laptop that was started by the trainer. The training program was activated after filling in the participant number, age, session number, gender and condition code. This condition code started either the intervention or the control condition. The procedure was conducted with double-blind, randomized controlled trials of active (intervention condition) versus sham balance-point training (control condition). An independent researcher, not directly involved in the study,

provided this computer-generated code (number: 1 or 2) to the trainer. It was not possible to differentiate which condition was activated at first sight by either the trainer or the participant.

On the first training day, before the computer task was started, participants filled in the NAS-PI and AVL-AV questionnaires. After this the computer task was activated. Each computer training session (one slot of 30-45 min) consisted of three phases: baseline, training -and test phase. This was similar for both the intervention and control condition. The representations of the baseline and test stimuli were identical (15 images) and were presented in 45 trials; each stimulus from the morph sequence was presented 3 times. Images were presented in random order, for 150ms, preceded by a fixation cross (visual noise). Then a prompt was presented asking the participant to respond by tapping either a green dotted key: happy (keyboard letter C), or a red dotted key: angry (keyboard letter M).

In the baseline phase each participant’s balance point between these categorical

judgements (happy vs. angry) was estimated. This estimation was done by calculating the point at which each participant was equally likely to perceive happiness or anger in the presented faces by calculating the number of “happy” responses as a proportion of the total number of trials. Following this baseline phase, a “training phase” was implemented. The presentation of the pictures and requested response were the same. However, the element of feedback was added: the response (happy or angry) was followed by the message “correct!” or “incorrect!”. Lastly, following the training-phase, a test-phase was implemented. Here the procedure to assess the baseline balance point was used to assess the post training balance point. The aim of this procedure is represented in the illustration below.

Figure 1. Illustration of the stimuli and design of the intervention. Each participant’s balance point between

the categorical judgments (happy vs. angry) was estimated (top row). Participants in the intervention condition receive biased feedback intended to shift this balance point (participants in the control condition received feedback that was directly based on their baseline categorization). The modification training was successful in shifting the balance point (bottom panel) (Penton-Voak et al, 2013).

The self-report questionnaires NAS-PI and SDAS were administered again, after all the trials were finished on the fifth day.

After the intervention week, six weeks of post measures followed. Every week the self-report questionnaire (SDAS) and a staff rated aggression questionnaires (OSAB) were filled in again. In the last week, week 11 of the full testing period, the emotion recognition task from the intervention week was again administered. The same measurement for the baseline procedure was used. This meant that participants rated the morphed faces as “happy” or “ angry” again and balance point was estimated. Also, in the 11th _{week a NAS-PI was administered again.}

D

,

To assess the first hypothesis that hostile attribution bias was related to aggression, the relationship between the baseline balance point and baseline aggression scores on all pre-training administered questionnaires was assessed. For this first hypothesis, correlations were examined. The expected direction was that, when HAB is related to aggression as the paradigm predicts, aggression scores must be negatively correlated with threshold for detection of an anger face. This can be explained as that the higher the scores on aggression questionnaires, the lower the threshold to detect an angry face (more choice of anger as opposed to happy). Data was derived from the pre-training aggression scores on the questionnaire SDAS, OSAB, AVL-AV and NAS-PI. The total scores of AVL-AV and NAS-PI were used, but not for the SDAS and OSAB. These questionnaires were administered four times; mean scores were calculated to have a clear estimate of the general average aggressive behaviour of the participant (less dependent on cofounding influences at time of assessment) and to control for missing data. Eventually, the four values of pre-training mean SDAS, mean OSAB, total AVL-AV and total NAS-PI were compared to the pre-training balance point with correlation analysis.

Second, to test whether the manipulation was successful repeated measures ANOVA was conducted over the baseline balance point difference for each training day. The expectation was that there would be an effect of time intervention condition; balance point would shift each day. For the control condition, no effect was expected: balance point would not shift.

For support for hypothesis two, the effect of training was evaluated with use of repeated measures ANOVA since there were several moments in time at which the questionnaires were administered. This led to three separate repeated measures ANOVA’s for each different tool:

SDAS, OSAB and NAS-PI. It was expected that there would be a main effect of time since aggression scores should differ over time as a result of the training. Especially the effect of interest is the interaction time*group since the effect should only appear dependent on the condition (intervention or control). If this effect appeared, repeated measures were conducted again after splitting the file based on condition. Then, according to the expectations, the effect of time should only appear for the intervention group, since the aggression score must differ over time as result of training. To assess training effects, aggression after the training was compared to baseline aggression. Baseline aggression was therefore added as a covariate in the analysis.

Data was prepared so that there were 3 values for the SDAS and OSAB tool after the training to compare: aggression was measured 6 times (each week, over six weeks) after training, two averages were created to control for missing values. The first one consists of week 7, 8 and 9 and the second one of week 10 and 11. This resulted in three moments in time for analysis: directly after (which is the direct score in week 6, the first week after training) 3 weeks after and 6 weeks after training. These were the moments in time that were compared in the repeated measures ANOVA analysis. As a last part of the hypothesis that training will lead to a decrease in aggression, correlations were again examined between aggression scores and shift of HAB. Correlations were calculated between the weighted mean difference of HAB (post-training

balance point minus baseline balance point) and the weighted mean difference of aggression

(post-training aggression minus baseline aggression). This was done for all the repeatedly

administered aggression questionnaires: NAS-PI, SDAS and OSAB. The expectation for the relationship was that aggression difference scores would be positively correlated with balance point difference scores: the bigger the change in balance point, the bigger the change in aggression scores.

Lastly, to investigate whether training effects were depending on the sample characteristic brain damage (BD) and intellectual disability (ID), the repeated measures ANOVA’s with values explained under hypothesis two were conducted again, now with the addition of BD and ID as separate covariates. Again, the effects of interest were time*group: aggression scores should change over time according to the group a person was in (intervention and control). In addition the interaction effect time*group*BD or ID was of importance, since this will reveal whether the effect of time dependent on the group a person is in is dependent of this person suffering from BD or ID. From this follows the expectation that this effect will be revealed when BD and ID are incorporated in the analysis as covariates.

The values for BD and ID are dichotomous and coded as: yes (it is present) and no (it is not present). For BD this is based on results from the semi structured interview. When a person answered yes to the question: “have there been any incidents in the past in which you lost consciousness as a result off a blow to the head?” this was valued as BD possibly being present (dichotomous code “yes” for BD). In all other cases, the person would be assigned to the no BD

group. For ID the dichotomous code “yes” was assigned when the score on the Screener Checklist Intellegence (SCIL) was 19* _{or below. Participants with scores of 20 and above are} assigned to the no ID group (dichotomous code “no”).

Results

Hypothesis one: Hostile Attribution Bias is Related to Aggression

Of the 116 participants, 98 participants completed day 1 of the training. Eighteen participants dropped out before the training started for three main reasons: nine of the

participants did not like the research, eight of the participants got either transported or released and one was diagnosed with autism and could therefore not be included. The drop-out group was compared against the non-drop out participants with independent sample t-test. This revealed that, on average, dropped-out participants were rated higher by staff on the observer rated aggression questionnaire OSAB (M = 29.48, SE = 2.78), than the non dropped-out

participants (M = 20.76, SE = 0.55). This difference was significant t (10.81)= -3.08, p < .05.

Furthermore, other variables that were analysed for differences were: age, education, possibility of brain damage, possibility of intellectual disability, prison regime, index crime, total of aggressive crimes in the past and pre-training aggression on self rated SDAS questionnaire†_{. No} significant differences were revealed.

Because these 98 participants completed day 1 of the training procedure, the data of these 98 participants was used to assess hypothesis 1: Hostile Attribution Bias (HAB) is related to aggression in violent offenders. Forty-eight participants were included in the intervention group and 50 in the control group. To see whether the randomization was successful, again

independent sample t-tests were conducted. This time over the variables: age, education, possibility of brain damage, possibility of intellectual disability, prison regime, index crime, total of aggressive crimes in the past, pre-training aggression on the OSAB, SDAS, AVL-AV and NAS-PI questionnaires and pre-training baseline balance point on the emotion perception task. This analysis revealed no significant differences between the intervention and the control group. The means and standard deviations of the sample characteristics are reported in the appendix (table 1.). To reveal the variability in the variables of interest, the means, standard deviations and ranges of each variable used to assess hypothesis 1, are displayed in table 2.

Table 2. Descriptive Statistics of the Variables for Assessment of The Relationship Between Hostile Attribution Bias and Aggression

* _{This is the cut-off value Kaal et al. (2013) determined.}

†_{The NAS-PI and AVL-AV were not included since these were administered at the first day of training; at} this point participants already dropped out.

Mean, standard deviations and ranges of scores on aggression questionnaires used for assessment of baseline aggression and pre-training baseline balance point.

Note: Baseline values (before the training was implemented), (n = 98). AVL-AV = Agressie Vragenlijst,

Aangepaste Versie the total score at baseline, NAS-PI=Novaco Anger Scale – Provocation Inventory total score at baseline, Baseline Balance Point = the point at which a participant is equally likely to perceive a face as “happy” as “angry”, Pre-training SDAS = the value is the average of four weeks of baseline assessments on Social Dysfunction and Agression Scale and the Pre-training OSAB= the value of this variable is the average of four weeks of baseline assessments on Observation scale of aggressive behaviour.

Skewness and Kurtosis were checked to see if the assumption of normality was met. The values showed that the data significantly deviated from normality. Subsequently histograms visually revealed that the directions of the distributions were all skewed to the left, meaning that the pre-training mean aggression scores on all questionnaires were on the lower end of the range. Transforming the data still lead to deviation from normality. Therefore, pre-training aggression was compared to baseline balance point using the non-parametric test Spearman’s correlation (𝑟𝑟𝑠𝑠).

Rs was calculated for the four different tools SDAS, OSAB, AVL-AV and NAS-PI. The

outcomes are displayed in table 3 below.

Table 3. The Relationship Between Aggression and Pre-Training Balance Point

Spearman’s correlations (𝑟𝑟𝑠𝑠) between baseline balance point and aggression measures

Baseline Balance Point

SDAS Pre-Training Mean -.07

OSAB Pre-Training Mean .09

AVL-AV Total Score -.06

Variable M SD Minimum – Maximum

Total score AVL-AV 25.84 8.98 12-51

Total score NAS-PI 129.64 29.07 72-223

Baseline balance point 7.83 1.76 2-12

Pre-training SDAS 6.49 5.34 0-27

NAS-PI Total Score -.27*

Note: *Correlation is significant at the p<.01 level (1-tailed). n= 98 for all scales, except AVL-AV: n= 97.

AVL-AV = Agressie Vragenlijst, Aangepaste Versie total score at baseline, NAS-PI=Novaco Anger Scale – Provocation Inventory total score at baseline, Baseline Balance Point = the point at which a participant is equally likely to perceive a face as “happy” as “angry”, Pre-training SDAS = the value is the average of four weeks of baseline assessments on Social Dysfunction and Agression Scale and the Pre-training OSAB= the value of this variable is the average of four weeks of baseline assessments on Observation scale of

aggressive behaviour.

There was a significant relationship between the NAS-PI total score and baseline balance point, 𝑟𝑟𝑠𝑠 = -.27, p (one-tailed) < .01, n= 98. There was no significant relationship between the other measures of aggression and baseline balance point.

Results of the Spearmans 𝑟𝑟𝑠𝑠 indicated that the higher the score on the NAS-PI scale, the lower the threshold for detecting a face as “angry”. This is in line with the expectations. The lack of a significant relationship between the other measures of aggression and baseline balance point is not in line with the expectations.

Hypothesis two: an emotion recognition training procedure designed to promote the perception of happiness over anger in ambiguous emotional expressions results in a change in self perceived and observer rated anger and aggressive behavior

Not all 98 participants that started the training completed it. Eventually 89 participants completed the full 5-day training. Nine of the 98 participants did not complete the full week because: the training did not live up to the expectations, they were absent two or more days because they were sick, participants left the institution (immediate release) or the training schedule was to much to keep up with. The data of the remaining 89 participants was used for analysis to find support for hypothesis 2. The intervention condition consisted of 45 participants

and the control condition consisted of 44 participants. Analysis with independent sample t-test of the 9 dropped out participants revealed no significant differences with the remaining 89 non-drop out participants.

Manipulation check

As a manipulation check, a repeated measures ANOVA with within subject variable: training outcome per day (with 5 levels: balance point difference score day 1, balance point difference score day 2, balance point difference score day 3, balance point difference score day 4, balance point difference score day 5) and between factor: group (intervention or control), was conducted to reveal the effect of training on balance point. Descriptive statistics of the baseline balance differences per group are displayed in table 2. in the appendix.

Mauchly’s test indicated that the assumption of sphericity had been violated 𝑋𝑋2₍₉₎₌ 35.3, p <.05, therefore multivariate tests are reported (ε= .83). The results revealed a main effect

for training outcome V= .15, F (4,77)= 3.42, p <.05, 𝜂𝜂2_{= .15, indicating that the balance point} difference changed over time. Pairwise comparisons showed that there was a significant

difference (p < .05) of the change in balance point between day 1 and all the other days, but not

between days 2 to 5, suggesting a decrease/stabilization of the effect over time. Most importantly the test of between subject effects showed that there was a significant difference between groups:

p= .01, see figure 1.

Figure 1. The Effect of Training on Balance Point Shift compared between the Intervention and the Control Condition (n= 82)

Figure 1. Plot of difference scores: pre-training balance point minus post-training balance point for each day

(training day 1 to 5) in the intervention and the control condition.

Test of within subject effects further revealed that there was a significant interaction effect between training effect and group, V= .27, F (4,77)= 7.13, p < .05. This indicated that the

training had a different effect on the balance point depending on the group a participant was in. To see whether the effects were significant for the intervention condition and not for the control condition, the repeated measures ANOVA was conducted again after the data was split based on groups. The results revealed a significant main effect of training outcome per day in the

intervention group V= .49, F (4, 37)= 8.96, p < .01, 𝜂𝜂2_{= .49. This main effect did not appear for} the control group: V=.10 F(4,37)= 1.04, p> .05, 𝜂𝜂2_{= .10.}

This indicates that the balance point difference changed over time, for the intervention condition only, which is in line with the expectations. This also meant that the manipulation was successful. Furthermore, the effect size for change over time could be considered close to large in the intervention condition (𝜂𝜂2_{= .49), but not for the control condition (𝜂𝜂}2_{= .10). Results}

showed that the difference in balance point between pre-training and post-training was strongest on day one.

Effect of training on aggression

From 89 of the participants that remained for the analysis under hypothesis two, 7 more participants dropped out during the follow up period (6 weeks). There was a group of 12 participants that did not complete the full 11-week participation period before the data in this research was analysed. Eventually, 70 participants completed the full 11-week program. Data of these participants was used for the analysis of long-term effects.

The effect of training on aggression was assessed by conducting separate repeated measures ANOVA for each tool: SDAS, OSAB and NAS-PI. The AVL-AV questionnaire was not included since it was only administrated before the training to assess (pre-training) baseline aggression. Descriptive statistics (M and SD) for each tool and the values at each moment in time

are displayed in table 3 (SDAS),4 (OSAB) and 5(NAS-PI) in the appendix. Repeated measures ANOVA for SDAS with between factor “group”

(intervention/control), within factor “time of assessment” (3 levels: direct after, 3 weeks after and 6 weeks after training) and covariate baseline mean SDAS score, revealed that there was no main effect for the time of assessment, F (1.69, 116.82) = .16, p> .05, n=72 ‡_{. This indicates that} the overall mean scores on SDAS were not significantly different over time over the entire sample. However, there appeared to be a significant interaction effect between the time of assessment and group, F (1.69, 116.82) = 3.82, p < .05, 𝜂𝜂2_{= .052. This indicates that the mean} scores on the SDAS at the different times of assessment differed according to the group the participants were in. To break down this interaction, in other words, to see at which time of assessment the mean scores did differ; separate ANOVA’s (with pre-training SDAS mean as covariate) were performed, comparing the means on each moment of assessment across the intervention and the control group. This revealed a significant effect of group for mean

aggression score on SDAS after 3 weeks, §_{F (1, 69) = 4.00, p <.05, 𝜂𝜂}2_{= .055. This effect did not} appear for the other moments of assessment: directly after training, F (1, 69) = 2.00, p >.05 and 6

‡_{Mauchly’s test indicated that the assumption of sphericity had been violated 𝑋𝑋}2_{(2)= 18.70, p <.05, also the}

Greenhouse-Geisser appeared to be greater than .75. To avoid rejecting to many false null hypotheses, degrees of freedom were corrected using Huynh-Feldt estimates of sphericity ε= .87.

§ _{The ANOVA’s for each moment of assessment dependent on group, were conducted across the same valid}

observations for the variable moment of assessment (same n) as had been included in the repeated measures ANOVA.

Eventually, this led to the analysis of n= 37 valid cases in the intervention condition, and n= 35 valid cases in the

control condition. When including all cases in the ANOVA analysis not controlling for missing cases that were not taken into account in the repeated measures ANOVA, the significant effect of SDAS score after 3 weeks for group did not appear. Although this distinction seemed crucial, independent samples t-test revealed that there were no significant differences between the groups (inclusion/exclusion from repeated measures) based on sample characteristics (age, education, presence of brain damage, presence of intellectual disability, prison regime, index crime (aggressive or not), total amount of violent crimes in the past, and pre-training aggression on AVL-AV, OSAB and NAS-PI questionnaires) that can explain this effect.

weeks after training, F (1, 69) = .29, p >.05.

Figure 2. Aggression Scores on Social Dysfunction and Agression Scale Directly after, Three weeks after and Six weeks after training (n=72).

Figure 2. Plot of self report measure Social Dysfunction and Aggression Scale (SDAS) score at each

moment of time of assessment, displayed for both the intervention and the control condition.

The significant effect of SDAS after 3 weeks indicates that participants in the intervention condition had significantly lower aggression scores (M=5.3, SD= 6.0) than participants in the

control group (M= 6.9, SD= 8.5) at that moment in time, see figure 2 and appendix table 3 for

total overview. The size of this effect can be considered small.

The second repeated measures ANOVA with between factor “group”

(intervention/control), within factor “time of assessment” (3 levels: direct after, 3 weeks after and 6 weeks after training) and covariate baseline mean score for the OSAB tool, revealed a significant main effect of time of assessment. Mauchly’s test indicated that the assumption of sphericity had been violated 𝑋𝑋2_{(2)= 20.81, p <.05, therefore multivariate tests are reported (ε=} .76). The results revealed a main effect for time of assessment, V= .11, F (2,56)= 3.58, p <.05,

𝜂𝜂2_{= .11, n=60. The effect size of this result can be considered medium to large. Pairwise} comparisons revealed, however, that there was no significant difference between any of the times of assessments. Next to that, no interaction effect was apparent for time*group V= .04, F (2, 56)

= 0.19, p >.05, indicating that the scores on the OSAB did not differ across the time period

dependent on group. The test of between subject effects revealed that there was no significant difference between groups, F (1, 57) = .08, p >.05 over time. This indicates that both the

intervention and the control group showed a difference in OSAB scores over time at any given point in time. The trend of this difference is revealed by the graph, see figure 3.

Figure 3. Aggression scores on Observational Scale for Aggressive Behavior after training

Figure 3. Plot displaying Observational scale for Aggressive Behavior score at each moment of time of

assessment, displayed for both the intervention and the control condition.

The third repeated measures ANOVA for the effect on the NAS-PI assessment tool, with between factor “group” (intervention/control), within factor “time of assessment” (2 levels: direct after, 6 weeks after training) and covariate baseline NAS-PI score, revealed no main effects for either time of assessment, F (1, 84) = .08, p > .05 or group, F (1, 84) = .85, p > .05. There

was also no significant interaction effect of time*group, F (1, 84) = .18, p > .05. Meaning that the

scores on NAS-PI were not different over time dependent on group. Figure 4 reveals the trend of the scores over time, indicating that both groups had lower NAS-PI scores directly after the training in comparison to 6 weeks after the training. Although not significant, the trend of the figure shows that lines are parallel: both groups showed a similar change of aggression scores over time with an increase at 6 weeks in comparison to 3 weeks.

Figure 4. Plot displaying NAS-PI score directly after and 6 weeks after training, displayed for both the

intervention and the control condition.

Although the manipulation seemed to be successful in reducing HAB, the results seemed to partly support our second hypothesis. In the intervention condition training did not result in overall decrease of aggression over time when compared to the control condition. Although, there seemed to be an effect on the self-report questionnaire after three weeks, this effect did not last to six weeks. Our hypotheses were based on the assumption that a reduction in HAB would be related to a reduction in aggression. To analyze the possible direct relationship between the reduction of HAB and the (non significant) reduction of aggression, correlations were calculated between difference scores (pre-training minus post-training) of HAB and aggression for each condition. The assumption of normality was not met for some variables (also not after transformation), therefore, difference scores were compared using the non-parametric test Spearman’s correlation (𝑟𝑟𝑠𝑠). Descriptive statistics are displayed in table 5 of the appendix, outcomes 𝑟𝑟𝑠𝑠 are displayed in table 2.

Table 2. The Relationship Between Change in Aggression Scores and Change in Balance

Point.

Correlations 𝑟𝑟𝑠𝑠between difference scores of HAB and aggression measures for SDAS, OSAB and NAS-PI

compared based on group; the control –and intervention group.

Control condition Intervention condition

Difference score HAB directly after

training

Difference score HAB long- term 6 weeks after training

Difference score HAB directly after

training

Difference score HAB long- term 6 weeks after training Direct SDAS difference score N .00 (44) .14 (33) .26* (45) .29* (35) OSAB difference score N -.23 (41) -.26 (32) .05 (37) .02 (31) NAS-PI difference score N .18 (45) (34) .19 -.09 _{(45) (35)} .01 3 weeks

SDAS difference score N .13 (39) .22 (34) .29* (40) .28* (35) OSAB difference score N -.12 (37) -.20 (33) .01 (38) .09 (33) 6 weeks SDAS difference score N .06 (36) .01 (34) .18 (37) .05 (35) OSAB difference score, N -.42** (31) -.03 (30) -.11 (33) -.16 (32) NAS-PI difference score N .11 (39) .31* (34) .02 (40) .10 (35)

Note: To see the possible relationship between differences in HAB and aggression dependent on time, the

values of change in aggression at three points in time (directly, after 3 and after 6 weeks) were compared against the balance point difference. The difference scores for aggression were created through subtracting the pre-training aggression score from the score of that moment in time (directly after training, three weeks and six weeks after). Subtracting the pre-training balance point from the balance point directly after training created the balance point differences. Difference score HAB long term (6 weeks) represents the effect of training after 6 weeks and was created by subtracting the pre-training baseline balance point from the balance point at the end of the 11 weeks study (6 weeks after training).

**Correlation is significant at the p <.01 level (1-tailed) * Correlation is significant at the p <.05 level (1-tailed)

There seemed to be a strong significant relationship between the OSAB difference score after 6 weeks and the HAB difference score measured directly after training 𝑟𝑟𝑠𝑠= -.48, p (one-tailed) < .01 in the control condition. This negative relationship indicated that participant’s baseline balance point change was lower when aggression difference scores got higher. This is not in line with the expectations because there should be no change in aggression when balance point doesn’t change. Looking at the Mean of the difference score of aggression 6 weeks after the training for OSAB (table 5, appendix) shows a positive value: aggression increased on OSAB compared to pre-training scores when there was small change in balance point. Also, NAS-PI long term difference score (after 6 weeks), seemed to be positively related to the HAB difference score 6 weeks after training 𝑟𝑟𝑠𝑠=.31, p <.05. Indicating that the higher the difference of pre and post-training aggression, the higher the change in baseline balance point. The meaning of this

relationship can be explained as: more aggression is higher threshold for detecting an angry face. This is not in line with the expectation for the control condition and this effect would be more in place if it occurred in the intervention condition. In the intervention condition, a positive

relationship was revealed between SDAS long term difference scores and baseline balance point difference directly after and 6 weeks after training 𝑟𝑟𝑠𝑠=.29 and 𝑟𝑟𝑠𝑠=.28, p <.05. The meaning of this relationship is that the bigger the change in aggression, the higher the change in balance point (higher threshold for detecting angry faces). This is in line with the expectation for the intervention condition. This effect supports the previous finding that was detected with the repeated measures ANOVA for SDAS, as this also showed that there was an effect of training after three weeks. However, there seemed to be an elevation of aggression as could be seen in figure 2 in the control condition. This might partly be responsible for the significant difference in SDAS score after three weeks. This point will be evaluated further in the discussion.

Hypothesis three: training effects depend on sample characteristics brain damage and intellectual disability.

To see if the effect of training on aggression was dependent on brain damage (BD**_{) or} intellectual disability (ID), the analyses above were conducted again with these variables as covariates. Crosstabs revealed that both ID and BD occurred in sufficient numbers for revealing group effects in the intervention and the control group; sample distribution is displayed in table 3.

Table 3. Size of Groups Brain Damage and Intellectual Disability across Conditions

Distribution of participants (n) for brain damage and intellectual disability across intervention and control group

Intervention Control

Brain damage 29 25

No brain damage 19 24

Intellectual disability 21 18

No intellectual disability 27 31

The repeated measures analysis above with between factor “group”

(intervention/control), within factor “time of assessment” (3 levels: direct after, 3 weeks after and 6 weeks after training) and covariates baseline mean aggression score for the tools SDAS, OSAB and NAS-PI were conducted again. The variables “brain damage” and “intellectual disability” were separately included as covariates. Because the expectation is that BD and ID

**_{ID and BD among respondents is not a certainty, since screeners and clinical impression were used to} define the groups. Always read the presence of ID and BD as a mere possibility that these participants really have ID or BD.

would affect the benefit a participant has from training, it is important to check if the effects for both groups are different dependent on if someone has BD or ID.

Tables 4, 5 and 6 display the results for each aggression tool of repeated measures analysis with BD and ID added as covariates: F values, degrees of freedom, error and p values are

reported. Significant results (p values < .05) are displayed in bold font. Social Dysfunction and Aggression Scale

Table 4 (a and b). Results of Analysis with Covariates for Social Dysfunction and Agression Scale

a.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Brain Damage (BD) as covariate (n = 72).

Effect df ε F p

Time 1.72 116.84 .36 > .05

Time x group 1.72 116.84 4.15 < .05

Time x group x BD 3.39 117 1.15, >.05

Group x BD 2 69 2.83 > .05

b.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Intellectual Disability (ID) as covariate (n = 72).

Effect df ε F p

Time 1.65 113.67 1.81 > .05

Time x group 1.68 107.31 4.12 < .05

Time x group x BD 3.30 113.67 2.88 >.05

Group x BD 2 69 .37 > .05

When adding BD and ID to the repeated measures ANOVA as covariate for SDAS, the effects were similar to the analysis under hypothesis two (without BD and ID as covariates). There was an interaction effect of time and group, 𝜂𝜂2_{= .058}††_{. However, the interaction effects} group*BD/ID and group*BD/ID*time were not significant. This indicates that BD and ID had no influence on the effect of training dependent on the group a participant was in (intervention

††_{Mauchly’s test indicated that the assumption of sphericity had been violated 𝑋𝑋}2_{(2)= 18.70, p <.05, also the}

Greenhouse-Geisser appeared to be greater than .75. To avoid rejecting to many false null hypotheses, degrees of freedom were corrected using Huynh-Feldt estimates of sphericity ε= .86.

of control) and subsequently this also did not change over time. This is not in line with the expectations since, in the intervention condition; one might expect that BD and ID influence the training effects.

Observational Scale for Aggressive Behavoir

Table 5 (a and b). Results of Analysis with Covariates for Observation Scale for Aggressive Behavior

a.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Brain Damage (BD) as covariate (n = 72).

Effect df ε F p

Time 1.75 97.75 .16 > .05

Time x group 1.75 97.75 3.79 < .05

Time x group x BD 3.49 97.75 2.78 < .05

Group x BD 2 56 .39 > .05

b.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Intellectual Disability (ID) as covariate (n = 72).

Effect df ε F p

Time 1.71 95.77 .05 > .05

Time x group 1.71 95.77 .59 > .05

Time x group x BD 3.42 95.77 .38 > .05

Group x BD 2 56 .47 > .05

Results from repeated measures with covariate BD for OSAB revealed an interesting result. First the interaction effect time*group was significant, meaning that there now was an effect of time dependent on the group the participant was in. Most importantly the significant interaction effect between time*group*BD reveals that this effect of time is dependent on BD. To break down this effect repeated measures ANOVA for OSAB were again conducted, but now after the file was split based on groups: brain damage and no brain damage. This revealed that the effect of time*group(intervention of control) only appeared in the no brain damage group, F (1.81, 46.96)= 4.53, p <.05 and not in the brain damage group, F (1.64, 45.87)= 1.93, p

damage only. This is in line with the expectations.

Novaco Anger Scale – Provocation Inventory

Table 5 (a and b). Results of Analysis with Covariates for Observation Scale for Aggressive Behavior

a.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Brain Damage (BD) as covariate (n = 69).

Effect df ε F p

Time 1 82 4.52 <.05

Time x group 1 82 .01 > .05

Time x group x BD 2 82 .72 > .05

Group x BD 2 82 .73 > .05

b.) Effects displayed with degrees of freedom, error, F and p value for the results of repeated measures ANOVA with Intellectual Disability (ID) as covariate (n = 69).

Effect df ε F p

Time 1 82 1.05 > .05

Time x group 1 82 .12 > .05

Time x group x BD 1 82 .13 > .05

Group x BD 1 82 .24 > .05

Adding BD and ID to the repeated measures ANOVA as covariate for NAS-PI, the effects were similar to the analysis under hypothesis two (without BD and ID as covariates). However, there was one notable result for the analysis with BD as covariate. There appeared to be an effect of time. As was revealed in figure 4. (analysis without BD, page 22.) both the intervention and control group showed an effect over time, though this was not significant. Figure 4 reveals the trend of this effect, indicating that both groups had significantly lower NAS-PI scores directly after the training in comparison to 6 weeks after the training. Now, when adding BD as covariate, this effect was revealed to be significant. This is an unexpected result and there will be elaborated on this further. ID had no influence on the effect of training dependent on the group a participant was in (intervention of control) and subsequently this also did not change over time.