The Media Noose: Copycat Suicides and Social Learning

This post was chosen as an Editor's Selection for ResearchBlogging.orgResearchBlogging.orgI always remember 2008 as the year when the entire UK media descended upon the former mining town of Bridgend. The reason: over the course of two years, 24 young people, most of whom were between the ages of 13 and 17, decided to commit suicide. At the time I was working in Bridgend, so I’m able to appreciate the claims of local MP, Madeleine Moon, that media influence had become part of problem. After all, most editors will tell you: the aim is to sell newspapers. And when this rule is rigorously applied, it should not come as a surprise at the depths some journalists will sink to recycle a news story. Even at a local-level, where you’d think some civic responsibility might exist, journalists clambered over themselves to find a new angle, generating ridiculous claims such as: electromagnetic waves from mobile phones caused the suicides.

In fact, the role media plays in suicide dates back to the publication of Johann Gothe’s 1774 novel, The Sorrows of Young Werther, where the protagonist shoots himself after being involved in a love triangle. It was subsequently banned, after several men throughout Europe were believed to have taken their own lives imitating Werther. Myths aside, and there’s plenty of research showing suicides increase following intense media coverage (Phillips, 1974; Stack, 2003; Ishii, 2004). And while the causes of suicide are immensely complicated, their localisation in time and/or space may partially be the result of social learning: exposure to another individual’s suicide may lead to the imitation of suicidal behaviour. In a recent paper, Alex Mesoudi (2009) looks at several hypotheses relating to two general patterns of suicide clusters: point clusters and mass clusters.

The first of these, point clusters, are confined temporally and spatially: that is, there’s a temporary increase in the frequency of suicides within a small community or institution. This is an example of the above copycat suicides. Here, suicidal behaviour spreads through a localised network via social learning mechanisms, such as imitation and emulation. Meanwhile, mass clusters are differentiated by the lack of spatial clustering, with a temporary increase in the total frequency of suicides across an entire population. An example of this is when a celebrity suicide will garner a ridiculous amount of media attention, which, due to the wide-reach of the mass media, leads to people across the country imitating the suicide behaviour. As Mesoudi notes:

Consistent with a social learning effect, this increase is found to be proportional to the amount of media coverage, e.g. the number of column inches devoted to the suicide [8] or the number of television networks covering the suicide [10]. Moreover, suicide rates do not show a corresponding drop some time after the publicised suicide, suggesting that the immediate increase is not caused by already-vulnerable people committing suicide earlier than they otherwise would have [8] […] There is also evidence that people are more likely to imitate the suicides of celebrities who match them in gender and nationality [9], although this effect is less robust than the celebrity effect [11].

To further investigate these two types of clustering, Mesoudi uses agent-based simulations, combined with statistical cluster-detection analyses, to determine the population-level patterns of behaviour generated by interactions between individuals. One of these questions relates to whether or not spatiotemporal point clusters are caused by social learning, or some other mechanism, such as homophily: where a tendency exists for preferential association between similar individuals. These factors will increase the risk of suicide, and show spatial clustering, due to each member being independently at a high risk of suicide — that’s to say it’s not transmitted via social learning (non-copycat).

For mass suicide clusters, the tendency for researchers is to focus on three explanations: (i) that only the prestige bias of the celebrity is responsible; (ii) the effect is enhanced because celebrities share common characteristics with the target individual (similarity bias); and, (iii) the dissemination of suicide information by the mass media is responsible for suicide behaviour. For the first two hypotheses, evolutionary models suggest both prestige bias and similarity bias are an adaptive method for acquiring accurate information, which, for the most part, is far more efficient than trail-and-error individual learning and unbiased copying. Although these behaviours are generally considered adaptive social learning rules, there is also room for maladaptive behaviours, such as suicide, to spread throughout the population when exhibited by prestigious and/or similar individuals.

The last hypothesis places the mass media as a central component of mass suicide clusters. It’s important note that, rather than being mutually exclusive, the mass media is actually an amplification of our social learning rules:

Formally, mass media dissemination resembles “one-to-many” cultural transmission [24], where a single individual can influence a large number of other individuals simultaneously. Cultural evolution models suggest that the extreme one-to-many transmission that is permitted by the mass media can greatly increase the rate at which behavioural traits spread [24], thus potentially generating temporal clusters.

The general results of the simulations provide support that social learning generates spatiotemporal point clusters. As for the other hypothesis, of homophily generating spatiotemporal clusters, Mesoudi found it was partially supported (see figure 1 below). Homophily-based clusters only occurred when there was large individual variation in agents’ suicide risk, with these clusters mainly being spatial but never solely temporal. This is fairly obvious when you consider these are high risk agents, who, without social learning, have no reason to cluster their suicides across some point in time. On the basis of these findings, Mesoudi suggests future empirical tests should taking into account the “degree of individual variability in known suicide risk factors (e.g. age, sex, ethnicity) in a region, and by distinguishing between the spatial-but-not-temporal clusters generated by homophily and the spatiotemporal clusters generated by social learning.”

For the second set of simulations, neither prestige bias nor similarity bias were capable of generating mass clusters alone:

Both prestige and similarity bias act to reduce the subset of potential models from whom suicide-related behaviour can be learned. For prestige bias, this is because only a minority of the population can be, by definition, prestigious. For similarity bias, requiring that models must be similar to oneself in some respect reduces the number of potential models from whom one can learn. Both biases therefore reduce the frequency of social learning events and reduce the probability of clustering. This reduction in the probability of clustering was counteracted under certain conditions […] Yet even under these conditions (strong prestige bias, homophily) mass clusters were no more likely to emerge than purely spatial clusters or spatiotemporal clusters.

But what about the one-to-many transmission produced by the mass media? Well, this did generate mass clusters, but only when social influence was relatively weak — either directly via a reduced strength in social learning, or indirectly via prestige bias or similarity bias. When it was unrealistically strong the social influence resulted in suicide pandemics where all agents eventually committed suicide (see figure 2 below). Of course, this isn’t really reflective of real world situations. As Mesoudi notes in his discussion:

In summary, prestige and similarity bias were neither necessary nor sufficient for mass clusters, while one-to-many transmission was necessary but not sufficient. The three processes in combination generated mass clusters, which is consistent with sociological evidence for each in actual cases of mass suicide clusters. However, the model highlights the very different roles that each plays: one-to-many transmission acts to spread suicide behaviour across the entire population thus eliminating spatial clustering, while prestige and similarity bias somewhat counter-intuitively […] prevent copycat suicides from persisting and becoming pandemic.

Obviously there are many assumptions made in this model and, after all, this is an extremely complex phenomenon. On that note, however, the current model does highlight the need to restrict the dissemination and glorification of suicides. This is backed up by real world evidence. From 1983 to 1986 a large number of people were throwing themselves in front of trains on a recently introduced Vienna subway system. The reporting during this period provided detailed expositions of the victims and their lives. Finally, in an effort to combat the media influence, strict guidelines were introduced to stop reporting on suicides. From the first to the second half of 1987, subway-suicides and -attempts dropped by more than 80% — and are still at a low level ever since (Etzersdorfer & Sonneck, 1998).

Despite the Viennese experience, media coverage, in the UK at least, is still very dubious. We can clearly see this in media attempts to find any excuse to recycle, and continually maintain, a particular news story. But this isn’t just about suicides either. Recently, the British media is providing non-stop coverage of serial murderer Raoul Moat, not too long after they lavishly presented Derek Bird‘s rampage across Cumbria. I’m not saying there’s necessarily a causal relation between the two, even though it makes me wonder whether we’ll once again find ourselves in a similar situation.

Main Reference

Mesoudi A (2009). The cultural dynamics of copycat suicide. PloS one, 4 (9) PMID: 19789643

3 thoughts on “The Media Noose: Copycat Suicides and Social Learning”

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.