Revisiting the demeanour effect: a video-observational analysis of encounters between law enforcement officers and citizens in Amsterdam

ABSTRACT We investigate the ‘demeanour hypothesis’, stating that police officers are more likely to arrest and use force against citizens who display a ‘bad attitude’. We observed 78 encounters captured on surveillance cameras in the city of Amsterdam. Video material allowed us to code specific behaviours (‘citizen pointed at officer’) instead of the more ambiguous interpretation of behaviour (‘citizen was disrespectful’) used in prior studies. We employ two regression analyses to estimate the extent to which different types of citizens’ behaviour – ‘bad attitude’, non-compliance, and aggression and crime – relate to physical coercive behaviour by law enforcement agents. After controlling for non-compliant, aggressive and criminal behaviours, as well as situational and individual features, citizens’ ‘bad attitude’ behaviours remain associated with physical coercion. However, our data also shows that the effects of aggressive and criminal behaviours are far stronger than that of ‘bad attitude’ behaviours. Yet, there is an observable ‘demeanour effect’ in our sample. Conceptually, we provide a more thorough behavioural description of what a ‘bad attitude’ looks like. Practically, our findings can be used in training, such as scenario or VR training, in order to raise officers’ awareness of citizens’ behaviours, and may assist them to prevent escalation in their encounters with the public.


Introduction
Every day, law enforcement officers around the world interact with citizens on the streets, and most encounters are characterised by mutual respect and cooperation. However, in a small portion of these encounters, officers use threats or physical forcewith estimates suggesting around 0.1% in the Netherlands (Politie 2022) to 2% in the US (Harrell and Davis 2020). The literature indicates that citizen demeanour can be an important predictor of use of force and arrest. More specifically, the 'demeanour hypothesis' states that officers are more likely to initiate an arrest and/or use force when they are confronted with citizens who display a 'bad attitude', that is, behaviours perceived as disrespectful or hostile, yet not unlawful. In this paper, we explore this hypothesis using video observations of encounters on the streets of Amsterdam, the Netherlands. seriousness of the offense, the number of officers on the scene, the presence of weapons, whether the encounter was initiated by officers, and whether officers intervened in a dispute , Terrill 2003, Alpert 2015, Bolger 2015. Klahm and Tillyer's (2010) review revealed that an increased number of officers would increase force (see also Paoline and Terrill 2007), whereas others reported the opposite (e.g. Lawton 2007), or no relationship (e.g. McCluskey et al. 2005. Mastrofski et al. (2002, p. 529) also note that officers discredit citizens who display 'irrational' behaviour, which they may perceive as resistance. Finally, most studies conclude there is no relationship between the number of bystanders and force outcomes (Klahm and Tillyer 2010, but see Crawford andBurns 1998, Engel et al. 2000). We will now consider the demeanour hypothesis to further specify the role of citizen behaviours in law enforcement agent's use of force at the situational level.

The demeanour hypothesis and Klinger's critique
Specifically, we are interested in investigating the claim made by van Maanen (1978) that officers are more likely to use force against 'assholes'. When asked directly, Weisburd et al.'s (2000) respondents reported that they were more likely to arrest citizens displaying 'bad attitudes'. Others (Lundman 1994, Engel et al. 2000 found 'hostile' and 'disrespectful' behaviours to be the strongest predictors of force. Such studies claimed to find support for the hypothesised relationship between 'demeanour', and arrest and force. However, Klinger (1994) questioned whether it was 'demeanour or crime' that explained why 'hostile' or 'disrespectful' citizens were more likely to be arrested or subjected to force. When reviewing 17 studies between the 60s and 80s, he found that none of the studies had controlled for crime against or in the presence of police, and seven studies did not control for prior crimes. Klinger (1994) noted that those who did, used imprecise indicators such as a felony-misdemeanour dichotomy (e.g. Smith and Visher 1981). Klinger (1994) criticised the operationalisation of the demeanour concept. The prevailing notion of 'legally permissible behaviour of citizens during interactions with police officers that indicates the degree of deference or respect they extend to the involved officers' (p. 477), does not say much about actual behaviours, and neither did the research. In fact, only three studies included attempts of clarifying this, and these included fighting (Lundman 1974), violence (Visher 1983) and other attacks (Black 1980) in their hostility measures, conflating the effects of demeanour and crime. Klinger's (1994) reanalysis showed that 'hostile' and 'disrespectful' citizens were more likely to be arrested because they ' […] are more likely to commit crimes against and in the presence of the police, not because their demeanour connotes a lack of respect for police authority ' (1994, p. 489). Conclusively, he noted that research claiming to identify a demeanour effect, instead measured the effect of crime. Later, he vouched for bringing crime back into the equation to 'provide a clearer picture of the relative effects of crime and extra-legal variables […] [to] increase substantially our understanding of police arrest decisions' (Klinger 1996, p. 336). Interestingly, Klinger (2010) did find a significant demeanour effect on non-lethal force when experimentally studying the effect of roleplay-based training. Here, demeanour was measured as friendly, moderately hostile and highly hostile. The effect was, however, smaller than the effect of attacks against the police, in line with his previous argument.
Beyond Klinger's critique, these studies tend to suffer from methodological limitations. Relying on systematic social observation in which observers make notes on the spot, they suffer from validity issues due to the a-priori instructions 'given to observers and the judgements they make' (Worden et al. 1996, p. 329). Another limitation is the reactivity issue, as the presence of observers may influence the behaviours of citizens and officers they observe (Spano 2005). SSO-based research typically also struggles with establishing a clear temporal order of the sequences in the encounters (Klinger 1994), which makes it hard to determine if a citizen's behaviour triggers use of force, or if force occurs as a prelude to citizens turning 'hostile' and 'disrespectful' (Dunham and Alpert 2008).
We attempt to address these concerns by providing empirical descriptions of the behaviours we identify as bad demeanour, by observing and controlling for the effects of crime in each model, and by exploiting the non-intrusive, precise eye of the CCTV cameras.

Material and methods
We rely on a systematic video observational method, utilised previously in studies of conflicts in public (e.g. Philpot et al. 2019, Friis et al. 2020, Liebst et al. 2021. The videos were recorded by surveillance cameras in Amsterdam, from the period of March to June 2020. We gained permission to use the video for scientific purposes by the Ministry of Justice, and since it is practically impossible to ask for informed consent, the Ministry provided consent on behalf of the individuals recorded. The city of Amsterdam owns the cameras, and the mayor approves their placement based on disturbances of public ordertypically drug dealing, pickpocketing, robberies or other crime-related issues. This provides a somewhat biased perspective on activity in public space. However, the cameras are placed in a variety of residential and shopping areas, some involving night-life activities, and cover a vast area of the city, providing a large variety of encounters (for a discussion on camera location bias, see Lindegaard and Bernasco 2018, pp. 170-172).
The cameras are operated continually by municipality employees who can adjust the angle or zoom in if they observe suspicious behaviour. When this occurs, they notify a supervisor and determine if relevant authorities should be alerted. Upon registration, they note the camera number, time of the event, and a small description of what happened. The footage is stored for 28 days. We received a list of cases every month and selected all cases involving close contacts between citizens and/or law enforcement. 2 This resulted in around 1000 h of video material, from which we selected our sample.

Selection of encounters
We sampled relevant videos based on the following criteria (informed by Philpot et al. 2019). First, a video had to show an encounter between law enforcement and citizen lasting at least 30 s. Second, the officer had to be in the encounter performing social control, and not as a helper (e.g. just tending to injured people). Third, the footage had to have the technical quality to allow us to distinguish facial expressions. 3 Fourth, the recording should have no, or negligible, breaks in vision (a bus passing by would not exclude the video, but a tram parking in front of the camera would).
We defined a 'law enforcement officer' as either a police officer or a public enforcement officer (handhaver). The police are employed by the Dutch state and their role is well-known: preventing, investigating and persecuting crime as well as maintaining social order in public areas. Handhavers are employed by the municipality and their role is to supervise neighbourhoods or town centres. They deal with 'minor nuisance and antisocial behaviour' (van Steden 2017, p. 41), have limited police authority, and can, for example, issue fines and detain people until the police arrive, but they cannot use force or arrest people other than in self-defence (for in-depth analyses of handhavers, see van Steden 2017, Eikenaar 2019). Both types of officers were present in a substantial part of our dataset. Since our focus is on law enforcement behavioural responses, regardless of the type of officer, we report our models with the two groups lumped together. Separate models, in which we excluded the handhaver-only encounters can be found in the supplementary material.

Inductive coding
Based on the inclusion criteria, around 450 videos were included in the preliminary dataset. We then created an ethogram based on an inductive analysis of ten encounters from the preliminary dataset (5 including force, 5 without), that were randomly selected from each subset. This is a behavioural inventory of 'carefully defined behavioural patterns that can be recognized and reliably documented' (Jones et al. 2016, p. 489). It describes the observed behaviours of officers and citizens as detailed as possible, with concrete operationalizations of the codes. The next sections briefly describe the contents of the ethogram, and how observed behaviours relate to variables.

Citizen behavioural variables
Inspired by Nix et al. (2019), we identified four behavioural modes that make up our independent variables. First, we coded compliant gestures, such as nodding or smiling. This code constitutes the dependent variable compliance, implying an absence of bad attituded, non-compliant and/or aggressive/criminal behavioural modes, which we used as the reference category.
Secondly, we coded non-verbal bodily actions that are '[…] compliant […] but disrespectful and hostile towards the officer' (Nix et al. 2019, p. 621). The codes argumentative gesturing, pointing directly at an officer, videotaping or photographing an officer, yelling at an officer, and non-friendly touching, make up this behavioural mode. Such behaviours are not unlawful, yet they indicate a disagreement with, or symbolic resistance towards the officer. These codes constitute the variable bad attitude.
Third, we coded behaviours in which citizens failed to comply, consisting of refusing to follow orders ('inactivity' where officers would expect activity), resisting arrest (but not fighting back), and trying to escape. These codes make up the variable non-compliance. Here, we deviate slightly from Nix et al.' (2019) definition of non-compliance. They defined it as behaviours displaying both disrespect and refusal to follow orders, but we limit it to behaviours in which citizens refused to complycitizens may 'just' refuse to provide their ID without being disrespectful, and disrespect should not be a necessary condition.
Lastly, we coded aggressive and criminal behaviours. Whereas Nix et al. (2019) considered aggressive and criminal behaviours as part of their non-compliant variable, we observed them independently to account for Klinger's (1994) critique. We observed invasion of personal space (i.e. moving very close) and threatening gestures as aggression, and trespassing, stealing, public drunkenness, possession of a weapon, and violence as crime. Due to the rarity of these events, we lumped them together in one variableaggression and/or crime. 4 Non-compliant behaviours may constitute criminal offenses, yet we argue they differ behaviourally and therefore, conceptually. We define non-compliant behaviours as acts of passive or defensive nature intended to disapprove, oppose or avoid the police, and aggressive and criminal behaviours as active, physical behaviours directed offensively towards the police or other actors in the encounter (Terrill 2003, p. 63).
All behaviours were initially coded as counts, capturing each observation in the encounters. This provides a multitude of analytical options. For our analyses, we computed dichotomous predictor variables, indicating if we had observed any of the behavioural modes in the encounter. The behavioural modes are not mutually exclusive, meaning that a citizen could show bad attitude behaviours such as yelling, non-compliance, such as resisting arrest, and aggression or crime, such as violence against the police, in the same encounter. The only mutually exclusive category is the compliant mode (the complete absence of conflict behaviour), which excludes the occurrence of the other behavioural modes.

Officer behavioural variables
Officer behavioural codes are founded in a force continuum (e.g. Henriksen and Kruke 2020). Our baseline behaviour for police was presence and communication with citizens. The physical force categories start with what we call spatial and bodily control as the lower threshold. This code indicates that police delimited citizens' space, without being impactful by, for example, blocking a potential escape route, grabbing, pushing or holding a citizen. To our knowledge, such coercive use of the body to gain control over citizens is not explicitly identified in prior work, even though some research consider 'firm grips' as force (e.g. Klinger 2010). We also coded frisking as spatial and bodily control.
Further, we distinguished two types of arrests, without and with the use of pain stimuli, based on the officers' usage of their bodies. The prior involved 'traditional' arrest behaviour, handcuffing a citizen who put their hands on their back. In arrests with pain stimuli, we observed more forceful techniques such as restraining grips and wrestling the person to the ground. We also observed instances that involved threats with the baton. We coded this as threats with baton. Finally, we observed striking techniques such as punching, and the use of batons.
These behaviours were initially coded as counts, and later computed to three outcome variables. First, we constructed a dichotomous outcome variable indicating whether officers used force or not. Secondly, we constructed an overall force count variable, summing each occurrence of physical force together to measure cumulative usage of physical force. We call this force frequency. This variable, however, does not capture the severity of the force. Therefore, we also constructed an outcome variable indicating force severity. Here, we followed the force continuum logic, and coded 0 for no physical force, 1 for spatial and bodily control, 2 for arrests without pain stimuli, 3 for threats with baton, 4 for arrests with pain stimuli, and 5 for striking techniques 5 (baton, hits, and kicks).

Control variables
We controlled for the following variables. First, we considered citizen gender and race. These were first coded on the individual level. For gender, we determined if a citizen was male or female based on visual cues, and for race, we coded white or non-white based on skin tone. Later, we aggregated the variables to ratio variables indicating the proportions of males and non-whites. Also, at the situational level, we considered the duration of the encounter, the number of citizens interacting with officers, the number of law enforcement officers, and the number of bystanders present.

Coding procedure
After the inductive phase, we trained a colleague with extensive experience in video analysis to apply the ethogram. Together, we re-watched the ten videos used in the inductive phase and discussed the codes. When we disagreed, we adjusted the codes based on our discussion, to ensure reliability between coders. Thereafter, the first author applied the ethogram. Behavioural coding of CCTV videos is time-consuming, and coding the full sample was not feasible, and similar research typically conducted analyses on sub-samples of videos (Liebst et al. 2022, Lindegaard et al. 2022. To estimate the number of required observations, we ran an a-priori power analysis using the G*Power tool (Faul et al. 2009). We specified that we were interested in observing medium effects ( f 2 = 0.15), with an error term of 0.05, and the conventional power of 0.8. The analysis showed that 77 observations was sufficient. Due to force being rare, we decided to oversample forceful encounters. Therefore, we randomly selected videos from the dataset, with the only condition being that 50% should involve some kind of physical force, given our aim to study the relationship between different citizen behavioural modes and force outcomes. The analytical sample consists of 78 videos, where half involved force. Such sample size is comparable to research using body-cam footage (Willits andMakin 2018, Friis et al. 2020) and CCTV footage (Piza and Sytsma 2016).
When applying the ethogram, we viewed each video once to identify key situational characteristics: point of contact, most severe form of physical force, and end of the encounter, as well as identifying how many actors were present. Officers were identified by their uniforms, indicating POLITIE and HANDHAVING respectively. We considered a citizen as an active part if they engaged directly in interaction with an officer, or when the officer(s) directed their primary actions towards them. We identified bystanders through their awareness to the encounter (e.g. people who stopped to watch or tried to meddle). Thereafter, we visually assessed the gender and race of each officer and citizen.
In the next round of coding, we followed each actor through the encounter, and coded their behaviours in individual viewings. In cases where several different behaviours were performed in close succession, the video was slowed down to capture each behaviour correctly. In cases of doubt, the encounter was replayed several times. To ensure the correct temporal order of demeanour and force, we stopped the behavioural coding at the most severe point of force. After coding each video, we replayed it one last time to double-check the codes and scan for errors. As such, each video was viewed at least the same number of times as there were active actors, plus two, and a video with two officers and two citizens would be viewed at least six times for coding.

Analytical approach
We address the research questions statistically using SPSS. First, we investigated the association between our predictor variables and the likelihood of force being employed, through a simple crosstabulation. We also use bivariate logistic regression to report odds ratios (OR) for each predictor.
Secondly, we investigated the overall force count as an outcome to address research question two. Count variables have Poisson distributions, and like expected, the variance (12.23) was greater than the mean (2.32), meaning that the data is overdispersed 6 (Piza 2012). Therefore, we employed negative binomial regression models specifically fit for overdispersion (Berk and MacDonald 2008). We regressed the independent variables on the force frequency variable. We report incidence rate ratios (IRR, expressed as exp(β) in SPSS), showing the predicted change in the outcome, given a one-unit increase in the predictors (Piza 2012).
Lastly, we ran two multiple linear regression models to estimate the effect of each predictor on our force severity outcome. We report the regression coefficients for each predictor variable. Since our predictors are dichotomous, the coefficient is the average difference in force severity between the two groups. Regression diagnostics revealed that linear regression assumptions were met, and that there were no correlations between variables that exceeded 0.25, and no VIF value over 2.5.

Results
First, we provide descriptive statistics. Secondly, we present a crosstabulation of the predictor variables and the force employment outcome with bivariate odds ratios. Finally, we turn to the findings of our regression models: negative binomial models estimating the force frequency, and linear regression models estimating the force severity. Table 1 provides descriptive statistics for the variables in our models, as well as the behavioural codes within each category. Spatial and bodily control occurred in 43% of the cases, being by far the most prevalent form of force. On average, we observed 1.3 instances of such force, typically through blocking or grabbing. Arrests without any pain stimuli occurred in 23%, and the more severe forms of force observed (arrests with pain and striking techniques) both occurred in 3% of the cases. We observed bad attitude behaviours in 62% of the encounters, averaging 6.64 behaviours per encounter, with argumentative gesturing as the most prevalent. Non-compliance, and aggression and/or crime, were both observed in 21.8% of the encounters, averaging 0.4 observations of non-compliance (e.g. trying to escape), 0.28 observations of aggression (e.g. invasion of personal space) and 0.28 observations of crimes (e.g. violence) per encounter. The average number of citizens, officers, and bystanders in the encounters were 2.15, 2.49, and 3.19 respectively. The proportion of male citizens was close to 90%, and the proportion of non-white citizens was 47%. Finally, the encounters lasted 8 and a half minutes on average. Table 2 shows a crosstabulation of the force employment outcome and the predictor variables, revealing a relationship in our data. The occurrence of bad attitude behaviours is positively associated with the likelihood of physical force being employed. However, we also see that non-compliance, and aggression and/crime both have a stronger tendency to occur in forceful encounters. A clear majority of encounters where those behaviours occurred ended with physical force. Similarly, only one encounter with completely compliant citizens ended in some form of physical force being used. To address these relationships, and our first research question, we estimated the odds ratios for the three predictors variables and the compliance only variable. These odds ratios compare the odds of force being employed in the group of cases in which the predictor occured (column 'yes') to the odds of force for all other cases in which the predictor did not occur (column 'no'), for each of the four predictors. The results indicate that there is an observable demeanour effect in our sample, and that bad attitude behaviours increase the likelihood of force being employed, but not as much as non-compliance or aggressive and criminal behaviour do. The effect sizes are large or very large (Rosenthal, 1996). More specifically, the occurrence of a bad attitude is associated with an approximately four times higher odds of force (OR = 4.3), while the non-compliance and aggression and crime variables were clearly more influential (OR = 11.6 and OR = 26.4 respectively), which was in line with our expectations. To address our second research question, we turn to a negative binomial regression model to estimate the relationship between the behavioural modes and the force frequency outcome (Table 3). Again, we see a demeanour effect on force. The models show how each behavioural mode is associated with the overall cumulative count of force behaviours. The bad attitude variable produces an incidence rate ratio of 3.5. This means that when we observed an encounter with citizens displaying bad attitude behaviours, the force frequency was predicted to increase with a factor of 2.5, as compared to cases in which we only observed compliant behaviours. Furthermore, we find a positive, and stronger, association between non-compliance and the force frequency (IRR = 3.8). The same was true for the IRR for our aggression and crime variable (IRR = 4.1), indicating that when citizens display any such behaviours, the force was predicted to increase with a factor of 3.1 as compared to cases in which we did not observe these behaviours.
Related to our second research question, we see that encounters involving citizens who display bad attitude behaviours are predicted to involve more force behaviours, even when controlling for non-compliant, aggressive and criminal behaviours. However, like we expected, the effects of the non-compliance, and aggression and crime variables are once again stronger. This means that as the behaviour of the citizen becomes physical, the force intensifies. Interestingly, the effect of the duration of the encounter is not significant, and practically meaningless. From this, we can infer that when police use more severe force, they also perform a broader set of force behaviours. We interpret this finding as an indication that police will try lower forms of force before escalating the force continuum. Finally, to address our third research question, we turn to our multiple linear regression model (Table 4). Here, we estimate the effects of the behavioural modes on the force severity. Once more, our models show a demeanour effect on the force outcome. When estimating the effects on force severity, the bad attitude variable produces an unstandardised regression coefficient of 0.5. Non-compliance produces a non-significant result, and we cannot conclude that this effect is significantly different from the effect of bad attitude. The aggression and crime variable is once again the strongest predictor (β = 1.3). Practically speaking, the effect of bad attitude is just marginally significant (p = 0.046) and modest in size, predicting an increase of half a level of force. It should thus be interpreted with some caution. It is noteworthy that it remains significant when controlling for aggression and crime. This result indicates that officers in some situations use physical forcemost typically spatial and bodily controlwhen citizens display bad attitudes. When aggression or crime occurs, the models suggest that the force severity is increased into a more severe form of force.
Based on the regression analyses presented, we propose that demeanour mattersbut not as much as non-compliance, aggression and crimewhen it comes to the likelihood of physical force, the frequency of force, and severity of force in encounters between officers and citizens. Thus, our data provides evidence of a demeanour effect, independently of other citizen behaviours that are correlated with physical force.

Discussion
The aim of the study was to explore the relationship between different citizen behavioural modes and the use of force by law enforcement. More specifically whether bad attitude behaviours are associated with the use of physical force, and in what way it matters for the probability of force, frequency of force behaviours, and severity of force. We see that, independently of non-compliant, aggressive and criminal behaviours, citizens' bad attitudes increased the risk of being subjected to force, the frequency of force, and the severity of force. Further, our analysis suggests that noncompliant, aggressive and criminal behaviours have stronger associations with each of the force outcomes. 7 Our findings emphasise that demeanour matters, even when controlling for crime. Even though the force in general is mild, citizens who argue with, yell at, videotape, or point at the police are indeed more likely to be subjected to force, and subjected to more frequent, and more severe force compared to compliant and respectful citizens, despite their behaviour being lawful.
Our findings are largely in accordance with similar research. For example, results reported by Engel et al. (2000) indicated that 'disrespectful citizens are 2.2 times more likely to be arrested' (p. 246) and that 'non-compliant citizens or verbally resisting citizens 8 are 5.8 times more likely […] to be subjected to force' (p. 249). Klinger's (2010) analysis revealed that demeanour predicted an increase of around 0.6 'steps' on the force ladder when going from 'civil' to 'highly hostile'. He also found that attacking an officer had a stronger association with non-lethal force than demeanour. He did not, however, consider non-compliance. Nix et al. (2019) suggest that behaviours indicating a bad attitude increase police officers' antagonistic emotions in public order violations 9 , traffic stops and dispatch-initiated encounters, as compared to compliant citizens. This tension is likely to increase the chances of an encounter ending forcefully. Nix et al. (2019) argue that such forceful responses are related to officers' suspicion of these behaviours (p. 618).
Following this reasoning, officers may not use force because the citizen is 'an asshole', but rather because these behaviours might indicate that the citizen is going to breaking the law or pose a threat. As such, their use of force might be in anticipation of further escalation into aggression and crime. In an interview study of robbers, respondents described similar anticipatory use of force in the beginning of their robberies when their victims displayed non-compliance and, in their view, 'asshole attitudes' in order to avoid further escalation (Lindegaard et al. 2015). Relatedly, officers might infer a risk of criminal behaviour when they are confronted with 'bad attituded' citizens, prompting them to act forcefully to prevent escalation.
Furthermore, our findings suggest that officers, on the interpersonal level, did not use force differentially when they were confronted with men or women, nor when they were engaging with what we evaluated as 'non-white' or 'white' citizens, as can be seen by the close-to-null, non-significant effects of these control variables in all models. On an aggregate level, however, some groups may be in contact with police more frequently than others, due to factors such as where officers patrol, or who officers decide to engage with. The cameras may also capture these groups disproportionately often due to their placement. The proportion of non-white citizens (48%) in our sample is higher than the official figures (35.8%) for people with a non-Western background (Gemeente Amsterdam 2019), which may indicate a bias in 'target groups' (Çankaya 2020). On the other hand, Crul (2016) estimated that around 35% of the younger population (under 15) was of Dutch, European or US descent, as opposed to 65% of overall population, suggesting that age may play a role in this association. The police in Amsterdam explicitly focus on equality across ethnic groups, which might explain the lack of effect on the force outcomes on the interactionlevel (see also Lindegaard et al. 2023, for a research report on profiling in Amsterdam suggesting no effect of race, but an effect of age). This finding adds to the existing body of literature, which emphasises the importance of interactional dynamics of conflict encounters, as opposed to individual dispositions (e.g. Engel et al. 2000, Collins 2008, Klinger 2010. Our main focus is on forceful encounters. However, some patterns can also be seen in encounters that do not end with force. Here, we observed fewer bad attitude behaviours, no (or very low counts of) non-compliance, or aggressive and criminal behaviours. Based on this, the officers in our data seem unlikely to enter an encounter with forceinstead they seem to await citizens' behaviour before they respond with the most fitting behavioural response (which most often is no force at all). This suggest that officers are reactive when they encounter citizens, and that they rarely escalate the chains of actions without prior citizen behaviour to react upon. Instead, they tend to mirror the resistance . This impression is strengthened by our force frequency modelshere we find that officers use more force behaviours as the citizen behaviour turns aggressive or criminal. We interpreted this as a willingness to attempt lowers forms of force before escalating to the higher forms. This finding is in line with the research of Terrill (2001Terrill ( , 2003Terrill ( , 2005 who found that officers generally would not leap the force continuum unprovoked (see also Terrill 2010, pp. 64-65).
In some encounters in our data, officers could have been more forceful than they were (see also Terrill 2005, p. 134). Eski (2018) argued that security guards' perception of force was characterised by 'heightened awareness, if not fear, of accountability' (p. 995) and that this caused them to be hesitant to employ force. While we do not believe officers we observed were afraid of accountability, we should not disregard the possibility that their heightened awareness can play a role in preventing or reducing the use of force, not just for the observed handhavers, but also for the police. Attention to procedural justice and accountability increased in the public and research the last decades (such as the anthology edited by McCoy 2010). Future research should scrutinise how changes in views on legitimacy, accountability as well as relevant societal events may affect the officers' perceptions and use of force and other social control techniques in action (see, e.g. Boehme and Kaminski 2023, for an analysis on the effect of the George Floyd murder).
Moreover, our innovative methodology allows us to study what people do, rather than what they say they do (Bennell et al. 2021), as well as what they do not do. An interesting observation was that officers relatively rarely engaged in calming gestures. In 52 encounter we recorded no observations of such behaviour, and in 21 videos showing any form of physical force, no calming gestures were used. In just six encounters we observed more than two calming gestures. This is remarkable, as studies of conflicts show that non-verbal, calming gesturing is one of the most important techniques used to de-escalate conflicts. Simply 'shouting' orders or commands without any attention to body posturing and the use of non-verbal communication, is less effective than if it was combined with a clear, calming body language (e.g. Philpot et al. 2020, Ejbye-Ernst 2022, Weenink et al. 2022).
An increased focus on calming gestures (and other de-escalatory bodily techniques) in intervention training and practice could make officers more confident and capable to calm down citizens rather than using force to install compliance. Our ethogram could be the empirical foundation for what a 'bad attitude' constitutes, and if applied in scenario-based trainingspotentially even through the use of virtual reality environments (Bosse et al. 2016)officers can experience these behaviours up-close. Klinger (2010), for example, showed that officers attending such training used significantly less force than those who did not. Such training, we recommend, should prompt officers to focus on their use of body language, and in particularly calming gesturing. Future studies should take this behaviour into account, and tests its effect on the use of force and officer victimisation.
Lastly, whereas prior research typically separates arrests from physical force, we chose to include both in our force outcomes. Furthermore, we included the use of spatial and bodily control as physical force. This decision was based on our assessment of the behaviour as inverted from the video observations. Ideally, we would have liked to confirm this assessment with experiences of officers employing these forms of coercion, and people subjected to them, to strengthen our confidence in our assessment of it as physical force. The use of physical force should not be limited to restraining techniques, striking techniques or the use of weapons, as such operationalisation will lead to a narrow understanding of the wide array of bodily techniques of social control officers can impose on citizens. Therefore, we argue that both handcuffing (regardless of how painful it may be) and any coercive use of the body must be included in observations of use of force in relation to citizens' demeanour.

Limitations
A limitation of our study is the lack of sound, and we cannot evaluate the impact of verbal communication on the interaction. However, previous studies showed that verbal and non-verbal communication tends to be used in combination (Friis et al. 2020), and that non-verbal communication is the most influential (Nassauer and Legewie 2021). Future studies should take this into consideration, and drawing on body-warn camera videos includes the possibility to study both verbal and non-verbal communication, especially in scenarios where BWC videos could be triangulated with CCTV videos.
It would be preferable to combine the sound and closeness of the BWC, with the birds' eye overview of the CCTV.
Furthermore, another limitation is the selection of study objects. We chose a broad definition of law enforcement officers for our analyses. The Dutch police and the Dutch handhavers were present in a substantial part of the videos we had access to. Police officers were present in 62 videos, and handhavers were present in 22. There were 56 police-only observations, and 16 handhavers-only observations. To control for potential biases between the groups, we ran our regression models with police-only encounters. Generally speaking, this did not affect the interpretation of the results: the direction of the effects remained the same in all models, despite most of them turning non-significant, due to the sample size falling under the threshold of 77. Relatedly, the standard errors of the estimates increased substantially. As such, it made the estimates more uncertain. Therefore, we encourage future research to address the differences between police and other law enforcers and front-line workers, such as the public enforcement officer included in this study, correctional officers, ticket inspectors and security guards, to improve the understanding of how various officers' use of force is affected by citizens' behaviours.
Lastly, our sample size is relatively small, which leads to our estimates being uncertain. It is also not a representative sample of encounters between law enforcers and members of the public. Therefore, we recommend future research to reproduce our analyses on larger, more representative samples, in order to confirm our findings for broader generalisation.

Conclusion
In this paper, we bring an older concept into a novel context. We have used a systematic video observational method to scrutinise the demeanour hypothesis, claiming that law enforcement are more likely to arrest and use force against 'assholes'. In order to empirically describe their behaviour and its role in use of force, we observed real-life encounters and developed an ethogram, presenting a detailed understanding of what constitutes a 'bad demeanour'. We move beyond general conceptualizations like 'impolite', 'disrespectful' or 'hostile' behaviour, to more specifically formulated codes of observable behaviours. Using videos and ethograms, researchers can move towards more objective measures of the behaviours we intend to study. Such descriptions can also be useful for officers, in order to know what behaviours they should try to deter early in an encounter, to prevent it from escalating.
Furthermore, we statistically examined the associations between these behaviours, and different use of force outcomes. Our study revealed a demeanour effect across the different analysesand behaviours such as argumentative gesturing, yelling, pointing and videotaping were positively associated with a greater likelihood, frequency and severity of force. We addressed Klinger's (1994) points of critique, and controlled for the effects of crime, which did not undermine the demeanour effect in the sample. Use of force, especially in the form of officers' mild coercive techniques to delimit citizens' use of their bodies and manoeuvring room, is, in other words, not just related to citizens breaking the law, but also their demeanour.