Name: Evelien Mols Student number: s1229028 Date: 29 August 2019
Supervisor: Dr. Roy de Kleijn Second reader: Dr. Jop Groeneweg Word count: 11.463
Cognitive Psychology
Thesis Msci Applied Cognitive Psychology
DECISION-MAKING DURING
POLICE–CITIZEN ENCOUNTERS
A research on the kinematic differences between
Abstract
Analyses of the kinematics of pointing and shooting movements of suspects may improve accuracy of decision-making during the police-citizen encounter. In this research, we investigated whether there is a predictable difference between movements associated with shooting a weapon and showing an identification card, from either the front- or back pocket and which body segment is most predictive for distinguishing between those movements. We also explored whether there are differences between those movements across different levels of experience in using firearms. We hypothesised that segments from the dominant hand will be most predictive and that experienced shooters will show less within- and between-subject variability during the shooting movement. After performing a Stacked Penalized Logistic Regression over the two different movements of 10 (front condition) and 11 (back condition) right handed subjects, we conclude that the right elbow is the most predictive body segment for frontal movement with an accuracy of 81%. For back pocket movement the right wrist is most predictive with an accuracy of 92%. Visualization of the movement trajectories showed that experienced shooters seem to make faster shooting movements, but we were not able to conduct statistical analyses to prove these differences. In this research we have shown that kinematic movement analyses might be valuable for the development of decision-making models for police officers. However, more research that includes a larger sample size and also takes compound variables instead of solely local kinematic variables into account is needed to further generalize the claims made.
Keywords: pointing, shooting, police, police–citizen encounter, gun use, shooting expertise, police training, decision-making, identification
Introduction
Police and shooting behaviour
Responding adequately to the intentions and behaviour of others can be crucial. Police officers encounter situations where responding competently involves using a firearm in order to maintain security for themselves and their environment. However, encountering a life-threatening situation is very stressful (Anderson, Litzenberger & Plecas, 2002). Under these stressful circumstances, one has to decide whether shooting is necessary and legitimized. During high threat situations, police officers experience high levels of anxiety and execute more false positive firearm incentives compared to low-threat situations (Nieuwenhuys, Savelsbergh & Oudejans, 2014). A false positive decision can result in unnecessarily injured suspect, whereas false negative errors can lead to serious risk of injury for bystanders and officers. This puts a lot of weight on adequately judging specific situations.
The prevalence of violent crimes has increased by 22% within Europe for the period 1997-2001 (Barclay et al., 2003). The prevalence of offences resulting in bodily injury, including through use of weapons, has remained stable across 22 European countries between 2003 and 2007: the European average per 100.000 was 35 offences in both 2003 and 2007 (WODC, 2010). Germany was the only country within the research that showed an increase, with a rise in offences of 17% between 2003 and 2007 (M = 188 in 2007). Within the Netherlands the amount of crimes involving weapons decreased by 2% between 2007 and 2014 (Kalidien & Heer-de Lange, 2014).
Because of ongoing European integration, Timmer & Pronk (2009) argue for the importance of a Europe-wide standard for police duties and norms. In their exploratory research Timmer and Pronk (2009) found a large number of differences within the degree of authority, training and gun use allowed for police officers across Europe. As can be seen in Table 1, the difference in deaths resulting from firearms use by police officers across four major western countries is enormous.
Table 1. Amount of people killed annually by police bullets in different countries, as mentioned by Timmer and Pronk (2009)
Country Amount of people killed in 2009 by police bullets
United States 1:1.000.000
Canada 1:3.000.000
The Netherlands 1:5.000.000
The relatively low number in Germany might be explained by the fact that German police officers have more restrictions for using their firearm than those of neighbouring countries do. German police is only authorized to use lethal force under the following conditions: (i) they need to arrest someone that is accused of a crime of violence; (ii) said subject does not cooperate with their arrest; (iii) the circumstances of arrest can be reasonably judged as life threatening (NOS op 3, 2016). In general, there is an international jurisdiction that violence use must always be proportionate to the severity of the crime (Timmer, 2005).
Although there is a lack of international comparisons in the literature surrounding crime rates and gun use by both police and suspects, we can conclude that violent crimes occur regularly, during which police officers need to make accurate life or death decisions. Making a wrong decision has far-reaching consequences. In the U.S. 153 of such fatalities were reported in 2015 (The Force Science Institute, 2016). A common cause of such fatalities are for instance that a suspect takes an harmless object, such as a wallet, phone or lighter out of his pocket and it is mistakenly seen for a weapon (Binder & Scharf, 1980) (Fritsch, 2000).
In general, the behaviour of police officers will rest on their rapid believes and judgements whether there is a threat or not. This can lead to wrong decisions, but it must be kept in mind that equally rapid responses by officers have also saved the lives of officers and bystanders (Bayley, 1986). On the other hand, there are also cases known where police officers failed to shoot, when they should have (BBC, 2018). These types of errors were studied by Cox, Devine, Ashby Plant and Schwartz (2014). They tested shooting reactions of police officers to video footage of suspects reaching either for a gun or neutral object (cell phone, soda can or wallet). The officers made almost no errors of shooting unarmed suspects (true negative) but they commonly failed to shoot armed suspects (false negative).
A more recent study about the situational characteristics of a shooting incident, was conducted by Petersson et al. (2017). They conclude that the main cause for threat
misidentification is a lack of previous in-field experience, where a sudden emerging threat might be overwhelming and cause psychological distress. In addition, they state that apart from increasing the time required for perceiving and responding to a scenario, the
psychological distress can also affect the execution of their otherwise well-trained skills. In accidental shooting events (false positive), wo circumstances are common. First, holstered pistols are not allowed to be loaded, resulting in a two-step procedure to be performed during a possible dangerous, attention-demanding encounter. During this loading procedure, setting the trigger force is commonly forgotten which results that the firearm can be triggered with a lower force than officers are used to (Petersson et al., 2017). Another factor contributing to
accidental gunfire Petersson et al. (2017) mention is the physical element: involuntary clenching of the hand due to reflex responses, for instance due to loss of balance, being startled or due to mirroring of movements.
Finding ways to improve accuracy during those kinds of high danger situations can potentially save lives. To our knowledge, the specific kinematics of shooting and pointing movements of suspects have not been examined before, and analyses thereof might be a starting tool to adequately train police officer responses.
Decision-making
Bayley (1986) investigated how police officers choose their tactics in problematic, stressful situations and found that decisions made at earlier stages effect decisions made later in the encounter. There is a continuous sequence of decisions, behaviour and action-reaction which all contribute to the probability of an eventual use of deadly force (Bayley, 1986). Binder and Scharf (1980) characterized five important decision phases in police-citizen encounters: (i) anticipation, (ii) entry and initial confrontation, (iii) dialogue and information exchange, (iv) final decision and (v) aftermath. Important situational factors that contribute to the actual use of physical force of police officers are the issue of actual and perceived danger, which can be triggered by citizen age, gender, race and perceived disrespect (Bayley, 1986). The information exchange phase can be limited, and only a matter of one extreme short visual or auditory contextual cue, during which a police officer has to make a decision whether or not to use force. The phase of final decision is a mixture of intellectual and emotional factors (Bayley, 1986). Intellectual factors can for instance be the information an officer heard over the radio about the specific suspect, the visual appearance of an armed burglar or gunshots heard. Examples of emotional factors are fear or surprise.
White (2002) also examined the situational predictors of deadly police force. He found that the characteristics of the environment, such as the violent crime arrest rate, also affected behaviour and perception during the police – citizen encounter. Evidence also suggests a relationship between the amount of violence a police officer uses, the amount of exposure to violence one has encountered and the level of community violence (Fyfe, 1980). There are also direct external efforts to control deadly force, for instance by court ruling and laws, although this is country dependent. In addition, the characteristics of a police force’s organization such as their administrative policy turned out to influence behaviour during violent citizen encounters. These external efforts are however more important during low danger situations where suspects are unarmed (White, 2002). Situational factors play a more
determining role during situations where there are armed suspects (White, 1999) (White, 2001). For example, White (2002) found that victims of police shooting are most likely to be males with a median age of 24.5. Another conclusion derived from this research is that when the police encounters a suspect at a critical stage, i.e. when the suspect is still focused on the crime instead of escaping, police officers are more likely to u se force. Cox et al. (2014) mention the importance of real-world neighbourhoods for shooting decision, while neighbourhoods have reputations of either relative safety or danger. They address the
complexity that comes with shooting decisions: it is more than identifying a static object. If a suspect pulls an object from within a jacket or pocket, an officer needs to determine very quickly whether he is faced with a gun or not. Cox et al. (2014) also mention that he broader social context such as the nesting within specific neighbourhoods and cities, the suspect’s race and the prevalence of violent events contribute to gunfire usage decision-making.
A schematic overview of the decision phases and contributing factors as described before is found in Figure 1. We are interested in finding correct visual cues to identify threat during the information exchange phase in order to improve the ‘intellectual’ factor in the final decision phase. Ideally, this will result in less false positive perceptions of danger, improving the objective decision-making of a police officer in the use of lethal force.
Experience Various research has been conducted on specific traits that can contribute to an increased shooting performance. Vickers and Lewinski (2011) examined the differences in decision errors and accuracy between rookie and elite officers: rookies shoot significantly more often (61.5% of the times) in situations where an assailant drew a phone (a false positive situation), than elite officers (18.5%).
Landman, Nieuwenhuys and Oudejans (2015) researched whether certain personality traits could predict shot accuracy. They found that thrill-and adventure seeking was positively related to shot accuracy. In a follow-up study, they also found a correlation between self-control strength and shot accuracy under high pressure (Landman et al., 2016). As expected, shooting experience turned out to be a far stronger predictor: officers in the study that were part of a special arrest unit, perceived less anxiety and showed a much higher shot accuracy (Landman et al., 2016). Questionnaires and a continuous heart rate monitoring were used to measure their state of anxiety. The mean heart rate of special arrest units officers was lower during both low (M = 100, SD = 8.3) and high pressure conditions (M = 99, SD = 7.7), compared to a group of officers wanting to join a specialised arrest unit during low pressure (M = 107, SD = 19.3) and high pressure (M = 110, SD = 19.3) tasks. Regular police officers showed the highest heart rate during low (M = 110, SD = 16.9) and high pressure (M = 111, SD = 15.7) conditions (Landman et al., 2016). This can be attributed to the fact that officers that are part of special arrest units are typically more accustomed to handling dangerous and violent situations and often operate in situations where the use of force is very likely to be required (Williams & Westwall, 2003). It was also noted that regular officers spend
significantly more time on shot performance while focussing on the opponent’s gun than their more experienced colleagues while special arrest officers spend more time aiming, having a seemingly higher target-related focus (Landman et al., 2016). The duration of the aiming and firing phases between the rookies and elite did not differ significantly in the research of Vickers and Lewinski (2011). The elite officers focussed more on the weapon or phone of the assailant before firing, whereas ‘rookies’ focussed on their own weapon.
The aforementioned experiments focussed on the possible improvement of accuracy and identification of underlying variables that influence decision-making and attention. Applying this knowledge can result in minimizing the earlier mentioned emotional factors during the decision phase in a highly threatening situation. However, few studies have investigated possible ways to improve the intellectual aspect of the decision-making process that is involved during those dangerous encounters.
Sensorimotor skills and degrees of freedom
Observing and analysing kinematic data is a way to learn more about our movement behaviour and sensorimotor skills. According to Bernstein (1967) coordination of movement entails both a neuromotor process and an observable structural form. This brings some challenges within this domain. Bernstein, was the first to mention the degrees of freedom (DOF) problem: human movement happens accurately yet fast, leaving a lot of questions about how the motor cortex control system exactly operates during complex movements. This entails an enormous amount of degrees of freedom (Bongaardt & Meijer, 2000). Some
neurons share a largely common set of target muscles, called ‘synergies’ (Omrani et al., 2017). A muscle can be part of multiple synergies and one synergy can activate more than one muscle. DOF can therefore sometimes be compounded, but this depends on the specific task and experience of the person who conveys the movement. In that sense, coordination of movement is of a much more flexible nature than the traditional reflex (Bongaardt & Meijer, 2000). Apart from the DOF problem, there is another issue. Exact repetitions of movements do not occur, because you and the world around you constantly change and goals can be reached in multiple ways (Smyth, Morris, Levy & Ellis, 1987). Repetitions are never
identical and every movement needs fine-tuning, which creates challenges for the analysis of movements (Schuijers, 1997). Walmsley and Williams (1994) emphasize that our body maximizes functionality and minimizes effort in a way that we master redundant DOF and can control our biological system within the environment. For instance, humans tend to perform all their actions in as many straight lines as possible.
Kinematics in gun shooting
Speed and accuracy are two important variables during rapid-fire pistol shooting (Walmsley & Williams, 1994). According to the information-processing approach, there are two popular theories towards speed and accuracy. Fitts’ law states that speed and accuracy are a product of a logarithmic relationship between movement distance, movement time and target size. The theory of Schmidt and Meyer incorporates the same variables, albeit in a linear model. It turns out that the second model is more apt when time is limited, as is the case during the need to adequately react to an armed suspect (Walsmsley & Williams, 1994).
Scholz, Schöner and Latash (2000) address two non-independent problems that multi-joint movements entail in tasks such as reaching, pointing or shooting from a classic control-theory perspective: (i) the planning phase during which the shape of the movement gets specified and (ii) the control phase, during which the time it takes for all the DOF involved in
the planned path to be identified. Scholz et al. (2000) investigated the variability between trials during shooting movement, in order to analyse partition of joint variance. According to Suzuki et al. (1997) the centre-of-mass position is important for the control of skilled
extreme tasks, especially reaching for a firearm. Stability is the capacity of a system to restore its state after a phasic disruption and identifying differences in stability is a tool to distinguish task specific variables controlled by the nervous system. In the research by Scholz et al. (2000) subjects shot from the same position each time, in order to limit the trial-to-trial variations and to decrease inconsistency in performance. For the analyses, they performed time normalization of the movement paths. They found that the final joint configuration was dependent on the start position of the arm and that the success of the pistol shooting was dependent on the alignment of the barrel to the target (Scholz et al., 2000).
In an experiment about assessing differences of gun motion during high and low anxiety conditions, it showed that gun movement efficiency as well as the success rate decreased during a high anxiety condition (Causer et al., 2011). According to the processing efficiency theory, as mentioned by Eysenck and Calvo (1992) this occurrence might be explained by the idea that anxious thoughts consume resources of working memory, and leave less memory available for the processing of the addressed task.
Sport sciences
In 1973 a method was pioneered to isolate motion of a human body and it was found that point-light stimuli created impression of human motion according to Johansson (1973). Ever since, biological motion and anticipation of human action have been studied thoroughly, especially in sport sciences. Diaz et al., (2012) used kinematic data to investigate whether they could anticipate movement based on previous movements during penalty kicking in soccer. They found that the angle of the non-kicking foot, was predictive in the shooting direction. In addition, more complex movements, with a higher DOF and formation of synergies, turned out to be predictive. They proposed to use this knowledge for training. However, one should take into account that these advantages might diminish during high-pressure situations.
Lopes, Jacobs, Travieso and Araujo (2014) performed a similar research, although they used a regular size goal and a goalkeeper, as opposed to Diaz et al. (2012). They also found that the non-kicking foot plays a predictive role and their regression analysis showed that compound, higher order variables were more useful than single kinematic variables in predicting shot direction. A difference in outcomes between the researches was that Lopes et
al. (2014) found that the correlation kept increasing until the ball contact, while Diaz et al. (2012) found a correlation peak around 250 ms before ball contact.
The present research
In this research we will investigate the action movements of potential suspects in order to identify possible predictors in the movement of reaching out for a firearm versus a neutral object. More specifically, we address the movement of both (i) reaching for and firing a firearm and (ii) fetching and handing over an identification card to an imaginary police officer in front of the participant. Both conditions, shooting and identifying, get examined from two different angles: as if they were retrieved from the participants front pocket or from their back pocket. Our ‘suspects’ have different degrees of shooting skills: novice (about to start police academy), intermediate (completed police academy) or expert (part of special arrest units). It is important to take these differences into account, as literature mentioned multiple differences between shooting behaviour depending on level of expertise.
Furthermore, real suspects will also have differences in shooting experience, thus it is important to investigate whether this would make difference for the anticipation of movement. The movement of these ‘suspects’ will be recorded by 3D kinematic analysis techniques. We do not test the accuracy of the shots fired. The goal of the current research is to find out whether it is possible to predict how suspects will act based on the early onset of a movement. If this is found to be the case, we are furthermore interested in (i) identifying the variables that predict whether a subject will pose a threat or not and (ii) whether there are differences among levels of expertise. If we succeed to predict the outcome of the
movements, we can improve police training by focussing on specific body parts of the
suspect in an early stage of an encounter. This would contribute to the process of determining whether someone is dangerous or not and what would be the appropriate (re)action. This knowledge can contribute to a more accurate decision-making process by increasing the perceptual and action control skills of police officers. Ultimately, implementation of this knowledge can lead to fewer fatal false positive and false negative shooting decisions of the police force and thus improvement in safety for both police officer and citizen.
We try to gain more insights in the movements of suspects during both shooting and pointing actions, in order to learn whether specific suspect behaviour can be predicted in an early phase of a police – citizen encounter.
Research questions and hypotheses
There are multiple research questions that we want to examine, although dependent on each other:
1. Is there a difference between the movement of shooting a weapon and showing an identification card, from either the front- or back pocket and is it possible to predict the movement based on kinematic data?
2. If so, which body segment is most predictive for the movement?
3. Are the movements - and potential differences - the same across different levels of shooting experience?
We expect to distinguish different movement patterns between the shooting and card pointing condition. This can for instance be caused by the weight difference between a card and a gun. However, there might be huge differences within observations, as literature showed that exact repetitions of movements do not occur (Smyth et al., 1987). Nevertheless, we hypothesise that the general movement patterns between the 2 x 2 x 3 conditions (front/back x
shoot/pointing x novice/intermediate/ expert) are structurally different enough to be statistically observable and thus, that we are able to predict in an early phase what kind of movement will occur.
It is expected that synergies of the dominant hand will be predictive in the movement patterns. Both shooting and pointing is a movement you mostly carry out with one arm, and we do not expect that the significant differences will be found at the non-dominant hand like is the case with penalty kicking in soccer. We expect many differences related to experience level of a subject. Especially subjects who have never shot before, are expected to show much more variability between movement patterns across trials compared to more experiences subjects. For experts, it is expected that their shooting movements are more consistent within the group while extensive training is expected to have improved their sensorimotor skills. We do not expect much differences within and between the movement trajectories during the pointing condition, as this is not a specifically acquired skill.
Method Design
All participants had to perform 4 sets of 20 trials each. These trials consisted of four different tasks: (i) take an ID card from their front pocket and point it towards a fictitious police officer; (ii) shoot at fictitious police officer with weapon from front pocket; (iii) take
ID from back pocket and point it towards the fictitious police officer and (iv) shoot at fictitious police officer with weapon from back pocket. The trials were randomized by using a randomizer calculator (https://www.randomizer.org/). There were five different
permutations of the 4x20 trials and every combination was used a maximum amount of three times. The conditions are shooting/pointing from either the front/back, with a
novice/intermediate/expert skill level. This means we have a 2 x 2 x 3 design for the study. It is a complex design, with the 2 x 2 condition as within-subject variables and the x 3 condition as between-subject variable.
Participants
We cooperated with the Hessische Hochschule für Polizei und Verwaltung (police academy) from Kassel, Germany. The police academy gathered 15 participants between 18 and 40 years of age (M = 25). Of these participants, five were experts from a Special Arrest Unit and five finished police academy. Another five participants were students who were about to start with police academy and did not have any prior shooting experience. None of the participants was colour blind and all had normal or corrected-to-normal vision. During the testing of one expert subject, there was an emergency that resulted in being able to complete only 66 out of 80 trials. Because of the same reason, another expert was not able to
participate in the study.
Table 2. Descriptive statistics of the participants between the 3 weapon experience conditions (table continues at the next page).
N Mean Std.
Deviation
Std.
Error Min Max
Height (cm) Novice 5 172.0 6.82 3.05 163 180 Intermediate 5 174.4 9.56 4.27 165 187 Expert 4 178.8 8.46 4.23 168 186 Total 14 174.8 8.17 2.18 163 187 Weight (kg) Novice 5 69.0 11.71 5.24 52 80 Intermediate 5 76.6 10.74 4.80 65 90 Expert 4 87.0 16.31 8.16 64 100 Total 14 76.9 13.94 3.73 52 100 Age Novice 5 18.6 .89 .40 18 20 Intermediate 5 22.0 1.87 .84 20 25 Expert 4 36.8 5.19 2.59 29 40 Total 14 25.0 8.32 2.22 18 40
Novice Intermediate Expert Total
Dominant hand Left 2 0 1 3
Right 3 5 3 11
Sex Male 3 3 3 9
Female 2 2 1 5
In Table 2 descriptive statistics of the subjects are listed across levels of expertise. Preparation, pre-testing and the experiment itself occurred in a laboratory at The Institute for Sports and Sports Science of the University of Kassel over a period of one week. The
participants however had to imagine that they were at the Wolfhager Strasse, a more
notorious part of Kassel, as the environment plays an important role in dangerous encounters according to Cox et al. (2014). To ensure safety, pre-testing was conducted under supervision of a police chief inspector of the Hessian Police Academy. All the participants gave their written informed consent and ethics approval, as provided by the University of Kassel, prior start of the testing.
Setup
To monitor the movements of the subjects, 28 reflective markers were placed
bilaterally at prominent bony landmark of their bodies. A six-camera motion capture system (Oqus 3+, Qualisys AB, Gothenborg, Sweden) was used to collect the three-dimensional kinematic data. Three cameras were mounted at the left side of the participant and three cameras were positioned at the right wall, of which one was standing on a tripod. The three-dimensional coordinates were recorded with 100Hz using Qualisys Track Manager (version: 2018.1). A webcam recorded all the participants during their tasks, in order to make sure we could re-watch possible malfunctions or unexpected behaviour if needed. There was a
demarcated zone, including footprint marks, in which the participants had to stand. In front of them was a shooting frame, simulating their target. The distance between the participant and the frame was determined by the police chief inspector and was a realistic distance in real life situations. A little bit above the shooting frame, a slide show was projected that gave the different task-cues. In Figure 2, a photo of the setup can be seen. The guns that were used were the same across trials: a P30 Heckler, Koch V2. This is the same weapon officers train with at the police academy, meaning intermediate and expert subjects were familiar with this gun. The bullets used were Force on ForceTM coloured cartridges.
The projected slides, that were the cue to either shoot or show identification, consisted of colour, text and a photo of a police officer. The slides with ‘identification cues’ were green and showed a police officer with an ‘open armed’ positioning. The slides with the ‘shooting cues’ had a red background and a photo of the same policemen, with his weapon in a ‘low ready’ state. These differences were created for participants to recognize the tasks faster, without having to read the text on the
slide over and over again, which wouldmight have influenced reaction times and movements of the
participants.
Procedure
Upon entering the lab, participants had to fill in a questionnaire to answer standard demographic questions, such as age, education, weight, colour blindness and weapon expertise. The
participants were given a heartrate measurement device, which they had
to wear below their shirt. Next, they received a holster with two weapons: one in the back pocket and one in the front pocket. During preparations we found out that carrying two weapons would be the best solution to enable randomization of the front and back trials. These were placed at the most common places to put a weapon in real life situations, as stated by the attending police chief inspector. In both pockets behind the weapon, identification cards were stored. To simulate that the front weapon was actually hidden in an inner pocket, we covered the front jacket with a piece of fabric. This way, the participant could not directly see the weapon when he was given the cue to use it. In order to guarantee relative similarity of holster positioning among different sized participants, pictures were taken for the positions of each weapon and compared with previous subjects. Two researchers of the Institute of Sports and Sports Science of the University of Kassel who were specialized in motion capture, placed infrared markers across joint centres of the body of the participants (see appendix, A1). Figure 3 shows the positioning of the weapon and identification card at the back is visible.
Instructions
A slide deck was projected to the wall the participants were facing. First, they were presented with a short story to picture themselves in the following situation: “You have just robbed someone at the Wolfhager Strasse in Kassel. While trying to leave the crime site, you got into a police control but do not know whether the police is already looking for you or whether it is a routine control.” The next slides would cue to either shoot the officer as soon as possible with either their front or
back gun, or make a pointing movement with their identification card from either their front or back pocket. The participants were given the instructions that they had to shoot one-handed, stay in the marked area without moving their feet and fire only one shot per trial. During the experiment all people in the room, including the subject observed, wore hearing protection.
Next, the participants performed four test trials (one of each condition) to get
comfortable with the task and to check whether all instructions were clear. Subsequently the experimental phase began, consisting of 4 x 20 trials. In between each set, the background story was projected again to remind the subject the situation he or she was in. All the slides with shoot/point cues were projected for 20 seconds. Slides with text were shown until the participant gave us the signal that reading was completed. All the communication during the experiment, including the instructions on the slides, was in German in order to guarantee full understanding for the test subjects.
Observations
Testing continued until a weapon was completely empty, after which munition was reloaded. This only took a few seconds, and we do not expect much discrepancies within the movement of the participants by doing so. The expert officers reloaded their own gun while novice and intermediate gun-users had their weapon reloaded by the police chief that was present. Sometimes there were small intermezzos between movements, for instance, when a candidate had trouble reholstering his/her weapon or when the Qualisys software had to
Figure 3. Positioning of the back holster and ID card for a right handed participant.
restart. Because we carefully listed and corrected for these moments, we do not expect that this affected our eventual dataset. The guiding police chief inspector noticed that the movements of some novice gun users was ‘better’ than the movement of some people with intermediate shooting skills. One expert (subject 2) was not able to finish the trial, because he was called away for duty at cue 66/80.
Sometimes participants used two hands or moved their feet despite the clear
instructions that this was not allowed. One participant did not notice the appearance of a new slide, that was the same as the previous, and therefore missed one movement. It was
noticeable that some participants did not watch their hands when reaching for their gun, but did look at their hands when reaching for their ID card. Some participants pointed the identification card towards the photo of the policemen on the slide deck, instead of towards the shooting frame which might have resulted in movement discrepancies. It seemed as if the subjects became faster and more accurate in their movements as the experiment continued. Participant 13 made a movement with his hand towards his nose, which was recorded by Qualisys as a movement but we were able to remove this movement from our data.
Data preparation
The kinematic data was labeled using Qualisys Track Manager and further processed using Visual3D (V3D; c-motion, Rockville, MD, USA; Version: v6.01.29). After the marker data was inspected and corrected, a body model was applied by a researcher from Institute of Sports and Sports Science of the University of Kassel to calculate 3D body segment
coordinates for all subjects and trials. The Kassel researcher determined the segments and joint centers as outlined by de Leva (1996) and the joint angles of the movements were calculated via the inverse kinematic model in Visual 3D.
The data was scaled to the movement by individual thresholds to define the start and end of either a shooting or pointing movement. For the start of the movement, the velocity of the elbow was used as a reference. Movement start was defined at the point of reaching 10% of the maximum velocity in the Y direction (front pocket) or in the Z direction (back pocket). The end of the movement was defined as the elbow joint angle reaching 90% of maximum extension, after full extension.
We exported this data to RStudio (version 1.1.463), where we built a longlist of the values of the coordinates of specific body segments of our subjects during the four different movements, including demographic data. This resulted in a file with X,Y and Z coordinates of 14 proximal- and 14 distal locations of the body with a measurement frequency of 100Hz
from the start- to end of a movement. Joint centres of shoulder, hip, wrist, elbow, ankle and knee were defined by the proximal and distal end of the segments articulating the joint, which meant that the data for some of these locations were identical (e.g. the distal marker of the left shank and the proximal marker of the left foot contained the same values). We removed duplicates, which resulted in movement data for 19 unique locations of the body (see
appendix A2). Furthermore we decided to leave out the data about the centre-of-mass, because the dataset became too complex. We also decided to not include the heart rate measures, as we were unable to match the heart rates to the four specific movements. Two separate subsets were created for the data: one for movements from the front and one for movements from the back.
Data normalization
The duration of the four different movements showed a high degree of variance (between 80 and 400 ms) across trials and subjects. In general, weapons were drawn much faster than ID cards. In order to analyze the data, we normalized movement frames to the mean length of the movement, as has also been done in other kinematic studies (Scholz et al.,2000; Diaz et al., 2012). To do so we used linear interpolation, which resulted in trail lengths of 170 ms for front pocket movements and 175 ms for back pocket movements. Using the same approach, we also normalized the Z-coordinates to the mean subject length of 176 cm.
Analyse method
For our data analysis we cooperated with students from the Methodology and Statistics Master program of Leiden University. To answer the research questions related to movement differences between a shooting- and pointing movement, a stacked penalized logistic regression (StaPLR) analysis was used. The predictors within this logistic regression were the values of the coordinates (X, Y, Z). The outcome variables was whether the subject pointed with an ID ( = 0) or fired a weapon ( = 1 ). The within-subject design of the two different movements was accounted for by not splitting the data across trials, but across participants.
Next to the StaPLR analyses, we also fitted an Extreme Gradient Boosting Machine (XGBoost). For the third question related to the difference in expertise levels, we compared the specific X, Y and Z values over time for the body segment found to be the strongest predictor for shooting versus pointing across novices, intermediate and expert gun users. All analysis were conducted separately for the front- and back pocket movements.
Results Data exclusion
Because of the enormous differences in movement patterns between left- and right-handed subjects, we decided to exclude the data from left-right-handed subjects. This meant we had to continue our analyses without data of two novices and one expert. An additional subject was excluded from the ‘front pocket’ condition because of unsolvable errors during the data pre-processing steps. This resulted in data for 10 subjects (n pointing = 196, n shooting = 189 ) for analysis of front pocket movement and data for 11 subjects for analysis of back pocket movement (n pointing = 235, n shooting = 192). For the different levels of experience, we differentiated between beginner (n = 3), intermediate (n = 5) and expert (n = 3) gun users.
Performance measures
A stacked penalized logistic regression analysis was performed on both the front pocket and back pocket condition. In our study a pool of base-learners was trained on the 19 separate body parts, with values of each coordinate (X, Y and Z) as predictors. Next, the probability vectors were used as input for the meta-learner. An elastic net penalty with a value of alpha = .5 was selected to keep grouped features together (Beck & Makonnen, 2019). In this step, coefficients with low probability vectors shrunk to zero. An overview of the applied StaPLR model that Beck and Makonnen (2019) based on literature from Loon, Fokkema, Szabo and de Rooij (2019), is shown in the appendix (A3).
Through leave-one-out cross validation (LOOCV) we were able to assign one subject to the test set and the other subjects to the training set. Multiple performance measures were computed and averaged: sensitivity, specificity, accuracy and the area under the curve (AUC) (Beck & Makonnen, 2019). We also fitted XGBoost to achieve the highest possible
prediction accuracy, which might not be captured by a StaPLR approach (Beck & Makonnen, 2019). The results of the performance measures for the StaPLR and XGBoost analysis are presented in Table 3. As can be seen, XGB showed a higher accuracy than StaPLR at the front pocket condition but mean sensitivity and specificity are almost the same across the front (M=0.851) and back (M =0.881 ) subgroup. Overall, the XGBoost classifier (M =0.870 ) performed slightly better than the StaPLR procedure (M = 0.862). Interestingly, StaPLR outperformed XGBoost at the back condition. A remark on this outcome is that the accuracy of StaPLR fluctuated considerably depending for which person the prediction was made for (Beck & Makonnen, 2019).
Table 3. Overview of performance measures for StaPLR and XGBoost classifier
Condition Method Accuracy AUC 95% confidence
interval AUC Sensitivity Specificity
Front StaPLR 0.808 0.993 0.966 - 1.020 0.920 0.717 XGB 0.893 0.978 0.931 - 1.025 0.881 0.910 Back StaPLR 0.915 0.969 0.889 - 1.023 0.956 0.834 XGB 0.846 0.947 0.874 - 1.021 0.879 0.797
Note. All values are averaged across LOOCV runs; nfront= 10 and nback= 11. Research questions
1. Is there a difference between the movement of shooting a weapon and showing an identification card, from either the front- or back pocket and is it possible to predict the movement based on kinematic data?
Analysis of the kinematic data showed marked differences between pointing- and shooting movements, both from a front- and back pocket starting position. With StaPLR, 81.8% of front pocket movements and 91.5% of back pocket movements were classified correctly (Beck & Makonnen, 2019). Using the XGBoost prodecure, 89.3% of front pocket movements, and 84.6% of back pocket movements could be correctly predicted. Although the XGBoost technique shows high accuracy for the front pocket
movement, the algorithm is not useful for more in depth interpretation of the data and does not show which specific markers are most important for the movement prediction (Beck & Makonnen, 2019). Therefore, we will focus on the results of our StaPLR analyses to answer the following research question.
2. Which body segment is the best predictor for the movement?
Our main interest was to investigate whether we are able to generalize the outcome across the subjects, rather than finding specific coefficients per person. The body parts were ranked on the relative size of their corresponding mean odds ratio. An averaged Spearman’s rank correlation coefficient was computed to test whether this ranking was stable across runs (Beck & Makonnen, 2019).
Front pocket. Table 4 shows the data of the mean odds ratio of the coefficients
retrieved with the StaPLR (meta-learner) procedure, for the movements from the front pocket. Note that left shoulder, right hand and left foot are not listed, because those body parts shrunk to zero by the elastic net regularization. This means that they do not influence predictions of the movement.
The relative size of the coefficients is an indication of how important the specific body part is in the prediction. This means that the right wrist seems to be the most important factor in predicting front pocket movements (Odds Ratio = 200.85, SD = 124.53), followed by between hips (OR= 110.41, SD = 48.70), right elbow (OR = 53.03, SD = 21.46) and right foot (OR = 19.60, SD = 10.73). Noticeably, the right hand is not listed while the left hand received the seventh rank, an unexpected result given that we hypothesized the non-dominant hand would not feature in shooting versus pointing predictions. Furthermore the standard deviations for all coefficients are large.
Table 4. StaPLR meta learner coefficients for most predictive body parts: Front pocket movements
Body part Occurrence OR* SD**
Right wrist 10 200.85 124.53 Between hips 10 110.41 48.70 Right elbow 10 53.03 21.46 Right foot 10 19.60 10.73 Right shoulder 10 15.32 27.90 Right hip 10 14.90 12.02 Left hand 10 9.82 10.17 Right ankle 10 6.67 3.35 Below neck 10 5.96 4.89 Left hip 3 5.21 6.72 Left wrist 10 4.94 3.59 Right knee 10 4.25 2.47 Head 2 3.21 3.01 Left elbow 2 1.85 0.66 Left ankle 1 1.83 - Left knee 5 1.30 0.27
Note. * Mean Odds Ratio across all LOOCV Runs, ** SD of the size of the estimate across all LOOCV Runs. Body
parts are ranked by importance (size of OR) in descending order. ‘Occurrence’ describes the number of runs during which the respective body part was determined to be influential.
This indicates that the coefficients varied majorly across cross-validation runs and therefore the data is not very generalizable (Beck & Makonnen, 2019). Spearman’s averaged rank correlation indicated a somewhat unstable ordering for the coefficients of the front condition, with a value of .72. As can be derived from Table 4, the values of the markers at the right wrist, between hips and right elbow are the most predictive to distinguish either a
pointing or shooting movement from the front pocket according to the stacked penalized logistic regression. A visualization of these differences between the two different movements and the corresponding values can be seen in Figure 4. The thin lines are all the individual normalized trajectories of our subjects over time, and the thick lines the means of these movement trajectories. If we look at our best predictor, the right wrist, it is noticeable that the two different movements start almost at the same position in space at both the X and Y coordinate, after which the weapon movement reaches slightly higher values. In addition, a high level of variance is noticeable for the weapon condition from the X coordinate. If we look at the Z coordinate from the same movement, we do see completely different movement patterns for the two different conditions.
For our second most predictive body location, between hips, we see completely different trajectories. There is a lot of variation across all the trajectories, but the means of shooting and pointing movements show high degrees of overlap over time in the X and Z coordinates. Only the Y coordinate portrays a slightly higher mean difference for shooting versus pointing. The lack of difference in mean trajectories between the shooting and pointing movements for the between hips position is unexpected, given that our model identified movements at this body part as the second best predictor for shooting versus pointing behaviour. However, if we look at the trajectories we see much outliers at the weapon condition compared to the ID condition. It might be possible that the relatively high level of predictability is due this variance and less related to the differences in mean
trajectories.
If we look at the trajectories of the right elbow, we see the highest amount of difference between movements types (shooting versus pointing) across the Z coordinate. Shooting a weapon results in a much faster upward movement of the right elbow, than pointing with an ID does. Based on the data from Table 4 in combination with the
visualization of Figure 4, the differences of the Z values of the right wrist and right elbow seem most predictive for either a pointing or shooting movement from the front. It is also striking that the weapon trajectories show much more variance and a higher mean value at the Y coordinate of the between hips body segment.
Back pocket. An overview of the results of the same analysis for the back pocket
movement are presented in Table 4. Fewer body parts seem to influence the prediction of this movement compared to the front movement. For the back condition the coordinates of the right elbow (OR = 125.11, SD = 138.91) were found to be the strongest predictors for
shooting versus pointing movements, followed by the location of between hips (OR = 26.67, SD = 24.50) and the right wrist (OR = 18.62, SD = 25.30).
Figure 4. (Mean) trajectories of the front movement over time of the X,Y,Z coordinates at the three most predictive body segments according to StaPLR.
According to the StaPLR meta learner analysis the body parts below neck, right hand, left hip, right knee, left knee, right ankle, left ankle, right foot and left foot do not play a role in predicting whether a subject is going to shoot or pointing their identification card.
Spearman’s averaged rank correlation of the back movement indicated a more stable coefficients than the front subset, with a value of .94. In summary less body segments are important for the movement prediction during a back pocket movement in comparison to a movement from the front pocket but the ones that do are more stable and carry more weight for the prediction.
In Table 5 we see that for the back pocket movement the strongest predictors are first the right elbow, then between hips and thirdly the right wrist in destinguishing between the pointing and shooting movement, according to our StaPLR analyses.
Table 5. StaPLR meta learner coefficients for most predictive body parts: Back pocket movements
Body Part Occurrence OR* SD**
Right elbow 11 125.11 138.91 Between hips 11 26.67 24.50 Right wrist 11 18.62 25.30 Right hip 11 8.26 5.64 Right shoulder 11 4.83 2.35 Head 11 2.21 0.63 Left elbow 7 1.34 0.15 Left shoulder 10 1.30 0.24 Left wrist 2 1.21 0.03 Left hand 3 1.07 0.01
Note. * Mean Odds Ratio across all LOOCV Runs, ** SD of the size of the estimate across all LOOCV Runs. Body parts are
ranked by importance (size of OR) in descending order. ‘Occurrence’ describes the number of runs during which the respective body part was determined to be influential.
A visualization of these differences between the two different movements and the
corresponding values can be seen in Figure 5. The thin lines are all the individual normalized trajectories of our subjects over time, and the thick lines the means of these movement trajectories.
Like the predictors for front movements, the Z coordinates seem to be most important during the back pocket movements, as the mean trajectories differ the most. At the X
trajectory of the right elbow there is much variance between subjects, in particular for the shooting movement. Again, the trajectories of the between hips body segment seem very similar, the means of the X coordinates are exactly the same. In addition, in the back
condition, individuals showed higher values at the between hips Y segment during shooting than during pointing. The trajectories of the right wrist are similar to those of the right elbow, due to the interdependence of these two body parts. The X coordinates has consequently higher values during shooting, and at the Z coordinate we again see that the trajectories of the two different movements intercept.
Based on the data from Table 4 in combination with the visualizations of Figure 5, the differences of the Z values of the right elbow and right wrist seem most strongly predicting for a pointing or shooting movement during the back pocket condition. It is also striking that the weapon condition shows much more variation, and often higher outliers, at the Y
coordinate of the between hips body segment. These outcomes are highly similar to the forward movement, although the importance of the right elbow and right wrist is reversed.
Figure 5. (mean) Trajectories of the back movement over time of the X,Y,Z coordinate at the three most predictive body segments as found by StaPLR.
3. Are the movements – and potential differences – the same across different level of shooting expertise?
In our previous analyses, we combined the data of all our subjects. However, it seems plausible that people with a different level of weapon expertise show completely different trajectories over time. Due to the small sample size, we are not able to conduct a statistical analysis on this data. However, to explore whether there are specific trends between the novice, intermediate and experts, we visualized the trajectories of the most predictive body segment for both the front pocket condition (right wrist) and back pocket condition (right elbow).
Front pocket. In Figure 6 an overview of the right wrist trajectories for the front
condition of the three different expertise levels beginner (n = 3), intermediate (n = 4) and expert (n = 3) can be seen. If we look at the pointing condition, we do not see much variance between the three different kind of expertise levels. However, if we look at the shooting condition, there are observable trends in the mean trajectories. Looking closer at the X and Y coordinates, it is remarkable that the experienced officers show much more variance than their less experienced peers. At the Z coordinate the movement seems not to differ that much
between expertise levels, although the values for the novice subgroup are lower across most of the movement.
The combined trajectories present a somewhat unreliable image, as we have compared the movement of only a few subjects per subgroup. To take a closer look at the differences and see whether there is much variation between subjects, the individual trajectories of the
Figure 6. (Mean) trajectories of front movement for pointing and shooting over time for X, Y and Z coordinates at the right wrist, across three different colour-coded experience levels. Individual trajectories per subject of the red marked coordinates can be found in de Appendix (A4 to A6).
three coordinates of the shooting condition can be found in the Appendix (A4 to A6). In these figures, it is noticeable that for both the X and Y coordinate, there seem to be a trend that experts reach higher values faster, than intermediate and beginners.
Overall, the individual trajectories show that the within-subject variance for experts is not that big as it seems in Figure 5, and that the main source of variance is between subjects. Furthermore, the trajectories almost follow the same path across levels of expertise, but experts are faster in this movement than intermediate and novice gun users.
Back pocket. In Figure 7 we can see the trajectories of the back pocket movement of
the right elbow for both the pointing and shooting movement. For these analyses we had sufficient data to compare X, Y and Z coordinates for beginners (n = 3), intermediate gun users (n = 5) and experts (n = 3). Similar to the front pocket movement, the ID condition
seems to not fluctuate much across expertise level. If we look at the shooting condition however, we see pronounced differences. Especially at the X coordinate there seems to be a big fluctuation across novice, intermediates and expert level of gun experience.
Figure 7. (Mean) trajectories of back movement for pointing and shooting over time for X, Y and Z coordinates at the right elbow, across three different colour-coded experience levels. Individual trajectories per subject of the red marked coordinates can be found in de Appendix (A7 to A9).
To further examine the results in depth, we also looked at the individual trajectories of the weapon condition which can be found in the Appendix (A7 to A9). Overall, it is
noticeable that there is more within-subject variance for the back pocket motion than
observed at the front condition. When comparing the mean trajectory for experts in Figure 7 with the individual expert movement paths, we see that the mean trajectory is highly
dependent on individual differences. Furthermore, intermediate gun users seem to be slightly faster in their movements than the beginners. There is also a large amount of variance within subjects for the Z coordinate during the back pocket movement. Taken together, we see some experts making a relatively fast shooting movement during the back pocket condition but this does not hold for all experts. Some intermediate shooters are also very consistent in their movements. Overall, the back pocket shooting movements do not seem to be very stable and there seems to be both between- and within-subject variability.
Summary of results
According to our analyses, we are able to predict whether a movement entails drawing a weapon or pointing an ID from a front- or back pocket, through both a stacked penalized logistic regression analysis (front: 81% , back: 92%) as with as through an Extreme Gradient Boosting Machine (front: 89%, back: 85%). The StaPLR procedure found out that the right wrist, between hips and right elbow body segments are most predictive for the front
condition. For the back pocket condition the right elbow, between hips and right wrist are most predictive. The pointing movement seems similar across different expertise levels. Experts seem to perform shooting movements faster than intermediate and novice gun users, although there is a lot of between-subject variability. The shooting movement seems more unstable between individuals.
Discussion Movement type prediction
The aim of the current study was to investigate whether it is possible to predict whether a potential suspect is going to draw a weapon or show his or her identification card and if so, which body parts are most predictive. Additionally, we wanted to investigate whether this is dependent on one’s shooting experience. Looking at the results we can confirm that it is possible to distinguish the ID-pointing movement from the shooting
movement with fairly high accuracy. This is in accordance with our hypotheses. If we look at the trajectories (Figure 4 and 5) it seems that the Z coordinate of the right wrist and right elbow are most important in this distinction, but we also see that there is a high degree of variance across movements. This can be explained by the fact that exact repetitions of movements never occur, as theorized by Smyth et al.(1987) and by Schuijers (1997). The specific body parts that we expected to be important are the body parts of the dominant hand, such as right wrist and right elbow. We did not expect that the between hips segment would play such an important role. Therefore, our hypothesis that the synergies of the dominant hand will be most predictive, is partly accepted. If we look at the trajectories of the between hips segment at Figure 4 and Figure 5, we see that especially for the Y coordinate in the shooting condition there are many high outliers. This is an interesting result, indicating that subjects often tend to lean more to the left while shooting, which might be explained because the subjects need to adjust the position of their centre-of mass during this movement. In
contrast, the mean trajectories of this coordinate stay close to the mean trajectory of the pointing condition.
Variability for expertise
We have found multiple interesting trends in our exploratory analyses of movement differences between different levels of expertise. We observed that experts tend to make the same movement faster than less experienced shooters. There is however a large amount of between-subject variability. We hypothesized that the less experienced a subject is, the more within-subject variability they would show in their trajectories, and that experts would be more consistent within-subject because they are more trained and experienced regarding a shooting movement. If we compare the individual trajectories (Figure A4 to A9) however, this is not necessarily the case.
This is consistent with the research of Shim et al. (2005), who found that expert tennis players reacted significantly faster at returning balls than novice players. However, they still found a large amount of between-subject variability within the different level of experience subgroups, as well as among shooters of different levels of expertise. Some intermediate shooters were quite inconsistent with their movements, while others showed trajectories similar those of experts. As expected, there was not much variability between individuals or levels of expertise for the pointing condition. Almost all subgroups showed the same mean movement trajectories in this condition. Because of the low number of subjects per group it is not possible to form conclusive statements on this matter and we cannot concisely accept or reject our hypothesis regarding the difference in expertise levels. Another observation regarding the variability of the movements, is that the trajectories of the back pocket
movements are less consistent. This implies that this movement is harder for the subjects and thus less stable. This confirms our observations during the experiment, where subjects often could not immediately find their back pocket, meaning this movement occurred less smoothly than the front pocket condition.
A possible reason for the variability across movements could be due to gender differences. We did correct for height, but did not distinguish between female (n = 3) and male (n = 8) subjects. There are however structural differences between female and male bodies, such as the position of the hips and the position of the elbows when walking (Troje, 2002). Initially we gathered data of five female subjects, but two were left-handed. As such, we only used data of female subjects #5, #6 and #9. If we look at the individual trajectories in the Appendix, (A4 to A9) we cannot distinguish observable trends from the graphs between
male and female subjects. If they would have been visible, it is hard to tell whether they are caused by differences in experience level. We did not have enough subjects to perform a 2 x 2 x 3 x 2 analysis.
Implications
The practical outcomes for this study might be limited at this moment. However, to our knowledge a kinematic study performing a comparison between shooting and pointing movements has never been conducted before. The promising results of this exploratory research show that it could be possible to predict whether a suspect is drawing a weapon or drawing an ID card based on body part movement and urges for further research.
Another possible use of the data concerning shooting movement trajectories, could for instance be to test performance of police officers before or after schooling and improve shooting training. It was interesting to see that some intermediate subjects showed similar movement patterns as novice shooters while other intermediate subjects showed trajectories similar to experts. This data could be useful for recruiting and training purposes to identify good gunmen. This is relevant, as police departments have long struggled with how to effectively measure, and potentially promote, high-quality performance (White, 2008). However, it would also be necessary to also measure accuracy of the shots fired.
Limitations
Although we could reject and accept some of our hypotheses, there are limitations to our research. For instance our subjects received static pictures as their stimuli, which do not reflect the dynamic context in which shooting decisions are made (Cox et al., 2014). The measuring of the movement trajectories took place in a lab at the Institute for Sports and Sports Science of the University of Kassel, which is not a realistic location for an actual police- citizen encounter. Shooting decisions are not made within a controlled environment, meaning the real-world context is very important. Shooter bias patterns change between experimental settings of different studies, and therefore it is important to be aware that the psychological processes that involve critical decision-making are hard to simulate in laboratory settings (Cox et al., 2014).
Police officers (and offenders) are likely to experience stress and anxiety during an encounter involving gunfire, and as Eysenck and Calvo (1992) explained with their
processing efficiency theory, performance is negatively affected by anxiety. According to Petersson et al. (2017) extreme stress has a motor effect on shooting, as gunmen tend to squeeze the sidearm grip harder of a gun than in absence of stress factors. As such, it is
uncertain to which extent we can generalize the outcomes from this study. Furthermore, we asked real life officers to perform the movements and pretend they just robbed someone. A possible limitation in this setting is that real life officers, by virtue of their role, may experience a degree of cognitive dissonance when asked to perform an action during an imaginary situation that is counterintuitive to their expected intentions. Shooting is a skill that entails constraints for success which depend very strongly on external information, such as the location of the target (Scholz et al., 2000). However, in our study the target was a wooden frame and we did not measure the accuracy and thus what were actual successful shots. Also, it was not always completely clear for the subjects if they had to point their ID to the slides presented or towards the wooden frame.
The limitations of our dataset and methods of analysis are as follows. With XGBoost it was not possible to adequately interpret the results of the 89% (front) and 85% (back) (Beck & Makonnen, 2019). The results of the StaPLR analyses showed high standard deviations (Table 4 and Table 5), which means there is a lot of variability for the
predictability of specific body parts for the pointing and shooting movement. Our data sample was fairly small, resulting in us only being able to qualitatively observe trends concerning the movement trajectories of different shooting experience levels, but not being able to
statistically quantify the suspected differences.
For our analyses we did normalize for time duration of the movement and height of the participant but we did not correct for weight, which could also be a contributing factor to difference in movements. Especially when a subject has more belly fat, this could impact the X coordinate during a movement.
Recommendations
As mentioned in the METHODS section, we also measured heart rate and centre-of-mass during the movement of our subjects. However, because of time restrictions and a lack of meaningful applications for this data in our analyses, we have excluded this data for our analyses. As heart rate is a correlate for perceived amount of stress, it could be interesting in future research to include this to compare heart rate fluctuations over the different conditions. The centre-of-mass is important to adjust for stability during movements (Suzuki et al., 1997). Our research showed that the between hips body part is quite predictive for shooting versus pointing, which could be related to this adjusting for one’s centre-of-mass. Future research is required to test this. Another interesting element in future research would be to measure eyesight orientation during shooting, as Vickers and Lewenski (2011) found that
elite officers focus more on the weapon or phone of the assailant before firing, while
‘rookies’ focus more on their own weapon. It was also notable during our data gathering, that some participants did not watch their hands when reaching for their gun but did look at their hands when reaching for their ID card. Therefore, eye gaze measurement would an
interesting addition to the measurement of pointing and shooting performance. We have conducted our research in a laboratory, with police officers instead of offenders, which affects the transferability to a real life situation. However, it should be obvious that performing similar tests in a real-life environment brings with it substantial ethical and practical objections, especially since a motion capture set-up is required to record the three-dimensional coordinates.
Lopez et al. (2014) mention the importance of compound variables compared to local kinematic variables when analysing movement data. For instance, considering that the coordinates of the right wrist are an indicator for a specific movement result, this will automatically also influence the position of the right hand and right elbow. In this research however, we did not examine body parts as conjunction with motion of other body parts, contrary to recommendations by Diaz et al. (2012). This could be an interesting direction for future research. We also tried to incorporate shoulder flexibility of our subjects, but did not succeed due to measurement errors. It would be compelling to account for one’s flexibility as this might explain some of the between-subject variance.
The most important recommendation would be to test the difference in movements between pointing and shooting movements over a larger sample size, to improve
generalizability of claims regarding the predictability of movement outcome (shooting versus pointing) during movement onset. If we can replicate the outcomes of this research in a future experiment with larger sample size, it is possible to train officers to pay more attention to specific body parts of suspects. This knowledge will improve the intellectual decision-making process of officers, by increasing their perceptual and action control skills during dangerous encounters. This will ultimately lead to fewer fatal false positive and false negative shooting decisions during the police-citizen encounter.
Conclusion
We found that specific body parts accurately predict whether someone is going to draw his identification card or shoot with a weapon. The prediction between these two movements from a front pocket, is most accurate from deviations in the X, Y and Z
coordinates of the right wrist, whereas for movement from the back pocket the right wrist, between hips and right elbow seem to be the most predictive. This accounts for a
heterogenous group of 10 (front condition) and 11 (back condition) right-handed subjects across 812 measurements. There is not much variability in the movement of pointing an ID across different experience levels. However, regarding the movement trajectories for shooting a weapon there seem to be observable trends differentiating novice, intermediate and
experienced gun users. More research is needed to generalize the claims made regarding the influence of gun use experience on movement trajectory.
Acknowledgements
We would not have been able to conduct this research without the help from Kassel University’s Institute for Sports and Sports Science and the Kassel Hochschule fur Polizei und Verwaltung (police academy). We would like to thank in particular Armin Kibele, Sebastian Fischer and Lisa Claussen from Kassel University. From the Kassel police department our thanks go to Martin Silbersack. The analyses for the most predictive body parts were not possible without the consulting of Clara Beck and Sarah Makonnen from the Leiden University Methodology and Statistics Master program and therefore we would also like to thank them for their help.
