Mel Slater's Presence Blog

Thoughts about research and radical new applications of virtual reality - a place to write freely without the constraints of academic publishing,and have some fun.

31 January, 2013

In the Presence of Violence


In the Presence of Violence

The Scene

You’re in a rather empty bar minding your own business, enjoying a quiet drink, and eventually a stranger walks in and starts to talk to you. He is pleasant and you find that you have some things in common, in particular you both support the same football team. The friendly chat lasts a while, and suddenly someone you hadn’t noticed, who’d been sitting by the bar, a third person, strides over and starts to accuse your new found acquaintance of “staring” at him. Of course he denies this, saying that he’d just been talking to you. This third person becomes increasingly aggressive in tone and uses threatening body language towards your acquaintance. Whatever the latter says is immediately turned around to make it sound like an escalation of the argument, and soon it becomes clear to you that the only goal of the aggressor is to cause a fight. Occasionally the victim, the one you’d been having the friendly chat with, looks towards you, but he doesn’t actually ask you for help. Obviously this is going to end in some pretty bad violence, in spite of the fact that your acquaintance is trying everything he can to de-escalate.

The question of interest here is – what do you do? Do you intervene? Do you quietly leave the bar? Do you stand there trying to think what to do but not actually do anything? Do you freeze? What should you do? What is the right thing to do? Would it make things worse or better if you intervened?

Now replay the above scene. The man walks into the bar, and he starts chatting with you. You find that you and he really have nothing much in common. Although he seems to be interested in football, it is apparent that he does not support the same team as you. Otherwise, everything is the same as in the first version. The aggressor clearly wants to start a fight and this is going to end in violence.

What do you do in these slightly different circumstances?

Bystander Intervention

This is an example of the so-called ‘bystander’ problem – how people respond when they come face to face with violent emergencies. I wrote about this in an earlier blog entry ‘The Illusion of Violence’. As pointed out earlier it is very difficult to study this type of situation experimentally – to investigate the factors that might lead to someone intervening or not. There have been many studies of this type of bystander situation, but experiments do not actually include violence, or even if they do it is not a face-to-face type of violent scene, but may be something that is shown on a video.

Our interest was to find out what people actually do when they are face-to-face with this type of violent confrontation. Yet we cannot carry out experiments in ‘real life’. However, research over the past 25 years has shown that people immersed in a virtual reality tend to behave realistically – carry out actions, have emotional responses, even have thoughts that would be appropriate for a situation occurring in reality. By being ‘immersed’ I mean that they are in a computer generated surrounding environment, that they see in 3D stereo, where everything is fully life-size, where they can perceive using their body in a natural way – for example, turning their head to look to the side, bending down to see underneath something, and so on. This is especially important interacting with virtual humans – these are life-sized, they talk with you, they look you in the eye, it would seem as if they could touch you – they seem to respond to you, they have a life-like presence in the same space as you. Now under these circumstances you’re talking with a (virtual) acquaintance, and the third (virtual) man suddenly appears and violently threatens your acquaintance, how do you respond?

The Bystander in Virtual Reality

We recently had a paper published that addresses this issue. The particular factor of interest that we looked at was the extent to which your responses are modulated by the sense of group identity between yourself and the victim. In order to be able to manipulate group identity in a natural way we used football (soccer) club affiliation. We recruited 40 male supporters of the Arsenal football team. They went through the experience described in the opening paragraphs. However, for 20 of them the eventual victim was clearly himself an Arsenal supporter, and for the other 20 he clearly was not an Arsenal supporter. So 20 of them, with respect to this situation, were ‘in-group’ with respect to the victim and the other 20 ‘out-group’ (or at least definitely not ‘in-group’). The aggressor was clearly ‘out-group’ since he made it very clear many times that he ‘hated’ Arsenal and thought that they were an extremely useless soccer team (I’ve put it more politely here than the way that he expressed it).

We wanted to see whether group affiliation could predict helping behaviour. We measured the latter by the number of physical or verbal interventions that the participants made once the argument had started. A physical intervention might be something like trying to step between the two characters, or gesticulating towards them.  Also for half of the participants we programmed the scenario so that the victim would sometimes look toward the participant, and for the other half not. So finally this experiment had 4 conditions: in-group, looking; in-group, not looking; out-group looking; out group, not looking – with 10 participants arbitrarily assigned to each group (in fact data for 2 of the participants were not usable, so we ended up with 38 not 40).

What we found was interesting – that the ‘in-group’ people intervened on the average more than the ‘out-group’ people. In a questionnaire after the experience we had asked how much the participants thought that the victim had been looking towards them for help. For those in the ‘in-group’ the stronger their belief that the victim was looking towards them for help the greater the number of interventions. However, for those in the ‘out-group’ condition there was no such relationship.

Now the finding is on the face of it fairly obvious. If a fight breaks out between two people in your presence you might be inclined to be more likely to help the victim if he had some affiliation with you (in this case supporting the same football team), other things being equal. However, the interesting aspect is that this appears to occur also in virtual reality, even though everyone knows that nothing ‘real’ is actually happening.

Another aspect of the results that is hard to convey in a journal paper is the actual reactions of the participants – that they were disturbed by the situation, reporting things like a racing heart, and also irrational worries like if they had intervened then the aggressor might have turned on them, and so on. As I’ve said before, some part of the brain does not understand about virtual reality – and simply takes what is happening at face value – and responds. Of course you ‘know’ that nothing real is happening, and therefore a slower cognitive response might then act to dampen down your responses compared to those that might occur if the events were actually taking place in reality. In virtual reality studies in fact typically we are trying to capture that first automatic response, the one that happens before you have ‘time to think’. This is the genuine response, and the one most likely to be similar in virtual and physical reality.

Another interesting result of this experiment is that it illustrates that the quality of the computer graphics is not vital. If you look at the video there will be something striking – when the virtual characters talk their mouths do not move! There is no lip sync. After the experiment we asked participants what things they thought took them out of the experience – how could the scenario have been improved? Very few people actually mentioned the lack of lip sync. I think that they became so involved in the situation that they somehow didn’t notice it.

One thing that they did tell us though was that ‘a fight like that would never happen in a bar like this’ – in other words the décor of the bar was wrong, it was not a bar that would be frequented by this type of football supporter. This aspect of plausibility is extremely important, and requires research on the domain to be simulated for any kind of experiment that is supposed to be depicting events that could happen in reality. (See ‘Illusion is Part of the Definition’).

Statistical Diatribe and Symbolic Regression

One other aspect of this paper is quite new. Research in this field follows the conventions of psychology (in this case social psychology) in terms of statistical analysis and reporting. In psychology there is convention of the 5% significance level, enforcement of the frequentist interpretation of statistics rather than Bayesian, and the tyranny of linearity. When you carry out a standard analysis such as regression or analysis of variance, even if using a generalised linear model,  somewhere in this is a very strong assumption of a linear (in fact affine) relationship between the response variable and the independent and explanatory variables (even if the variables themselves might be transformed e.g. to a log scale). But why should everything be linear? In fact it is safe to say that conventional statistics makes the linearity assumption mainly because the mathematics and computation is easier – and certainly the latter is a factor that should not make us stop and think too much today, compared to the nearly a century ago when many of these techniques were invented.

In this paper we had both the observed intervention data (number of physical and verbal interventions) and questionnaire data. I wanted to see what the relationship was between the number of interventions and the responses to the questionnaires. I used a method called ‘symbolic regression’ (which is a specific aspect of Genetic Programming). In particular I used a system called Formulize (or Eureqa) in order to analyse the relationship between the numbers of interventions and the subjective questionnaire responses. See the paper by Michael Schmidt and Hod Lipson in Science.  Supporting Text S3 of our paper briefly explains how this works. The important thing is that this discovered something that I don’t think would be possible with conventional statistics. For the number of physical interventions (N), the resulting equation was of the following form:

N = group*exp(LookAt + VictimLooked) + f(… other questionnaire variables…)

The f() represents some function which isn’t important in this particular discussion. The variable group = 0 means ‘out-group’ and group = 1 means ‘in-group’. LookAt = 1 for the case when the victim occasionally looked towards the participant during the argument and LookAt = 0 for the group where this did not happen. VictimLooked is the response to the statement: “After the argument started, the victim looked at me wanting help.” This (as all questions) was scored on a 1-7 scale where 1 meant least and 7 most agreement with the statement.

Now if we look at this we see that the whole first term on the right hand side of the equation vanishes for the ‘out-group’. Hence only for the ‘in-group’ were the ‘look at’ factor and the strength of the belief that the victim was looking for help important. For the ‘out-group’ these factors seemed to have no effect.

This equation captures 85% of the variation in the original data. The point is that it is much easier to look at this equation and try to understand what it signifies than looking at the original data (that it very well) represents. Reviewers in the psychological and social sciences have to accept that the world has changed since the 1920s when methods such as ANOVA were invented, and that this type of data exploration is a valid way to understand data. There is a world beyond formal 'hypothesis testing'. For example, now that we have this type of equation, what would be wrong with an experimental replication that ran the symbolic regression on the new data and then compared the form of the equation with the original data? Or, having found this equation it tells us that the more we foster the idea that the victim is looking to the participant to help the greater the number of interventions should be. We could set up an experiment to test that specific hypothesis. This is not a challenge to conventional methods, but a statement that there is more - and different techniques should not lead to suspicion. As another example, when we first analysed the 'number of interventions' data we did not use standard ANOVA, which is based on the assumption of a continuous response variable, and a normally distributed error structure. This is because 'count data' (the number of times something happens) is better modelled with a generalised linear model with a Poisson error structure (log-linear regression). This was treated as something 'suspect' by a first round of reviews, even though generalised linear models with Poisson error have been around for at least 60 years!

So What?

What are the conclusions from this type of experiment? First is that using VR in this way allows us to carry out lab based experiments where participants are confronted with a situation that has high ‘ecological validity’ – it is almost like real life. Such experiments don’t come out of nowhere, but they are guided by theory. Here it was the idea suggested by Mark Levine amongst others that group affiliation plays a major role in bystander behaviour – the apparent relationship between the bystander, the victim and the perpetrator. Having carried out the experiment we obtain data so that we can now look again at theory with this additional information, and perhaps formulate a revised theory, leading to another experiment. On the latter we have since carried out an experiment where we change the number of bystanders (not just one – the participant) and have examined the effect of that. The results of the new study are in preparation.

On the practical side we could suggest, for example, that if you are a victim, then yes do explicitly ask people around you for help. This might not be effective if those around do not share some group affiliation with you, but it should help if they do. Even if they do not share affiliation with you, this is something that itself is open to reframing. For example, fans of two rival football clubs might be bitter enemies, but if the situation is redefined so that they are both ‘football enthusiasts’ (compared to say rugby enthusiasts) then at that level they do have a joint affiliation. This was explored in http://psp.sagepub.com/content/31/4/443.short where bystanders to a (non-violent) emergency behaved differently depending on whether they had been primed to think of themselves as fans of a specific football club, or general football supporters. So for a victim it might always be possible to appeal to higher level affiliations if it is possible to seek help from bystanders.

For the bystander him- or herself this experiment cannot say what they ‘should’ do. This depends entirely on circumstances and on the moral choice made by the bystander taking into account many factors – including most importantly their own safety and of others around. But for authorities this type of experiment might be very useful for the formation of policy. Even from this simple experiment, authorities could give advice that victims should explicitly ask for help if there are any potential bystanders around, even perhaps with advice about how to ‘reframe’ group affiliation (e.g., “Think about how my kids will feel”, might appeal to bystanders around who happen to have children!). Or another example, it is widely believed in the social psychology community that there is a ‘bystander effect’ such that the greater the number of bystanders the less the chance that anyone will intervene – because there is a diffusion of responsibility (“Why should I be the one to stop this?”). If that were the situation what should the victim do to break this? Should for, example, he or she choose someone at random in the crowd and appeal specifically to that person? Or should the victim somehow try to raise group consciousness towards prosocial behaviour (“You are all party to this attack by not helping me!”). We don’t know the answer, but with an experimental study we could gain some insight into this.

(WARNING -  this video includes bad language and
depicts a violent confrontation).

Bystander Responses to a Violent Incident in an Immersive Virtual Environment

  • Mel Slater, 
  • Aitor Rovira,
  •  
  • Richard Southern,
  •  
  • David Swapp,
  •  
  • Jian J. Zhang,
  •  
  • Claire Campbell,
  •  
  • Mark Levine