The Cultural Meaning of Suicide What Does That Mean? Review

• Loading metrics

Open Admission

Peer-reviewed

Research Article

The Cultural Dynamics of Copycat Suicide

The Cultural Dynamics of Copycat Suicide

• Alex Mesoudi

x

• Published: September 30, 2009
• https://doi.org/10.1371/journal.pone.0007252

Figures

Abstract

The ascertainment that suicides sometimes cluster in space and/or time has led to suggestions that these clusters are caused by the social learning of suicide-related behaviours, or "copycat suicides". Signal clusters are clusters of suicides localised in both time and infinite, and have been attributed to directly social learning from nearby individuals. Mass clusters are clusters of suicides localised in time merely non space, and have been attributed to the dissemination of information concerning glory suicides via the mass media. Here, agent-based simulations, in combination with scan statistic methods for detecting clusters of rare events, were used to clarify the social learning processes underlying point and mass clusters. It was constitute that social learning between neighbouring agents did generate signal clusters equally predicted, although this effect was partially mimicked by homophily (individuals preferentially assorting with similar others). The 1-to-many transmission dynamics characterised by the mass media were shown to generate mass clusters, merely simply where social learning was weak, perhaps due to prestige bias (only copying prestigious celebrities) and similarity bias (just copying similar models) interim to reduce the subset of available models. These findings can help to clarify and formalise existing hypotheses and to guide time to come empirical work relating to real-life copycat suicides.

Commendation: Mesoudi A (2009) The Cultural Dynamics of Copycat Suicide. PLoS ONE 4(9): e7252. https://doi.org/10.1371/journal.pone.0007252

Editor: James Holland Jones, Stanford University, Usa of America

Received: March xi, 2009; Accustomed: August 26, 2009; Published: September 30, 2009

Copyright: © 2009 Mesoudi et al. This is an open-access commodity distributed under the terms of the Artistic Commons Attribution License, which permits unrestricted utilise, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: The author has no support or funding to report.

Competing interests: The author has declared that no competing interests be.

Introduction

While suicide is undoubtedly a complex phenomenon with multiple and diverse causes [i], [2], show accumulated over recent years suggests that one of these causes may be social learning. These "copycat" suicides are proposed to be caused at to the lowest degree in part by exposure to another private's suicide, for case through the faux of suicidal behaviour. Ii general patterns of suicide clusters take been documented and taken every bit evidence for copycat suicides [3]: point clusters, which are localised in both time and space, and mass clusters, which are localised in time simply.

A point cluster is defined equally a temporary increase in the frequency of suicides within a small customs or establishment, relative to both the baseline suicide charge per unit before and subsequently the signal cluster and the suicide rate in neighbouring areas [four], [5]. For example, Haw [half dozen] documented 14 suicides within a psychiatric hospital during a one-year period, while Brent et al. [7] documented ii suicides and seven suicide attempts during a fourteen-day flow in a single school. Beyond anecdotal case studies, Gould et al. [5] used statistical analyses designed to find the clustering of disease infections to decide whether suicides occur in spatiotemporal clusters. On average around ii% of suicides amongst 15–xix year olds in the U.S. were found to cluster spatially and temporally beyond that expected by hazard, although this figure was equally high equally 13% in some states. Given this spatiotemporal clustering, point clusters are frequently explained in terms of copycat suicides, with suicidal behaviour spreading through a local network via social learning [four], [5].

A mass cluster is divers as a temporary increase in the full frequency of suicides inside an entire population relative to the period immediately before and subsequently the cluster, with no spatial clustering. Mass clusters are typically associated with high-profile celebrity suicides that are publicised and disseminated in the mass media. Analyses take shown that national suicide rates rise immediately later the suicides of amusement celebrities, and to a lesser extent political figures, have been highly publicised in the mass media [8]–[ten]. The implication here is that this ascent is caused past social learning: people across the state imitate the suicide behaviour of the celebrity. Consequent with a social learning effect, this increase is found to be proportional to the amount of media coverage, e.one thousand. the number of column inches devoted to the suicide [8] or the number of television networks covering the suicide [10]. Moreover, suicide rates practise non show a corresponding drop some time afterwards the publicised suicide, suggesting that the immediate increment is not caused by already-vulnerable people committing suicide earlier than they otherwise would have [8]. The consequence appears to be restricted to the suicides of famous people who are afforded some degree of prestige in their order (e.chiliad. entertainment celebrities); in contrast, not-celebrities and famous figures who have negative reputations (e.grand. common cold war spies), both of whom lack prestige, accept smaller or non-meaning effects on national suicide rates [9], [x]. At that place is also evidence that people are more likely to imitate the suicides of celebrities who friction match them in gender and nationality [9], although this effect is less robust than the celebrity effect [11]. Similar increases in suicide rates in response to media-publicised suicides have been observed in Germany [12], Japan [13], Taiwan [fourteen] and Austria [15].

The overall aim of this study is to use agent-based simulations to formally explore how unlike social learning dynamics might generate these different spatial and temporal clusters of suicides. Agent-based models are typically used in cases such as these to determine the population-level patterns generated by underlying interactions betwixt individuals [16], [17], making them a useful tool in the case of copycat suicides. Specifically, the amanuensis-based model addresses possible explanations that accept been posited for each of the two kinds of suicide clusters - indicate and mass - as discussed beneath.

Point clusters: Social learning or homophily?

Joiner [iii] has challenged the assumption made by researchers such as Gould et al. [5] that spatiotemporal indicate clusters are necessarily caused by social learning. Joiner [3] hypothesised that point clusters may instead be a past-product of homophily, the trend for like individuals to preferentially associate with one another [eighteen]. If people preferentially associate on the basis of factors that increment the gamble of (non-copycat) suicide, then spatial clusters of loftier-take a chance people will sally. These high-risk clusters may class suicide clusters due to each member'due south independently high gamble of suicide, without any social learning occurring within the cluster. Joiner [3] suggests that many spatiotemporal suicide clusters observed in hospitals and schools may be cases of independent suicides within homophilous groups of high-risk individuals. All the same, while in that location is extensive testify for the general phenomenon of homophily [18], no direct empirical test of Joiner'southward hypothesis has yet been conducted in relation to suicide, and without such tests it is difficult to determine which of these explanations - copycat suicide via social learning or independent suicide within a homophilous network - is responsible for point suicide clusters.

The first aim of the present study is to make up one's mind whether homophily tin can mimic social learning in generating indicate clusters, and if and then to guide future empirical research by exploring the conditions under which this is almost likely to occur.

Mass clusters: Prestige bias, similarity bias, and/or the mass media?

Explanations for mass suicide clusters have centred around three characteristics of such clusters: (i) that they are associated with prestigious celebrities only, (2) that the effect is greater when the celebrities are similar to the target individual, and (iii) that the mass media is involved in the dissemination of suicide information. Regarding the first two of these, Henrich and McElreath [19] suggest that mass suicide clusters upshot from two social learning biases: prestige bias, where individuals preferentially copy the behaviour of prestigious or high-status models [twenty], and similarity bias, where individuals preferentially copy the behaviour of models who are like to them in ethnic markers such as dialect, language or dress [21]. Evolutionary models suggest that both prestige bias and similarity bias are adaptive means of acquiring accurate information compared to both costly trial-and-error private learning and the unbiased copying of other randomly-called people [22]. Prestigious individuals have usually acquired high prestige considering their behaviour is adaptive, so copying prestigious individuals will, on boilerplate, lead to the conquering of that adaptive behaviour [20]. Copying like individuals is probable to lead to the acquisition of adaptive behaviour because similar individuals face similar adaptive challenges and so should have appropriate solutions to such challenges [21]. Crucially, withal, both prestige and similarity bias are vulnerable to the occasional acquisition of maladaptive behaviour when such behaviour is exhibited past prestigious or similar individuals. Thus copycat suicide can be seen as a maladaptive by-product of these by and large adaptive social learning rules [19].

Other researchers often cite the mass media as a driver of mass suicide clusters [10], [11], [23], with suicide-related behaviour assumed to exist disseminated via newspapers, magazines, goggle box and radio. Indeed, this assumption has led to the institution of guidelines and safeguards concerning the reporting of suicides in the media [xi], [23]. Formally, mass media dissemination resembles "1-to-many" cultural transmission [24], where a unmarried individual can influence a large number of other individuals simultaneously. Cultural evolution models suggest that the farthermost one-to-many transmission that is permitted by the mass media can greatly increment the rate at which behavioural traits spread [24], thus potentially generating temporal clusters. Note that prestige/similarity bias and 1-to-many transmission are not mutually sectional hypotheses: the "one" individual from whom the "many" acquire may be more prestigious than, or similar to, the "many".

The second aim of the present written report is to explore which of the aforementioned social learning biases - prestige bias, similarity bias and/or 1-to-many manual - are necessary and sufficient to generate mass suicide clusters.

Hypotheses

Based on the literature reviewed above, the following predictions are made:

1. 1a. Social learning generates spatiotemporal bespeak suicide clusters.
2. 1b. Homophily can generate spatiotemporal point clusters in the absence of social learning.
3. 2a. Prestige bias generates temporal (but not spatial) mass suicide clusters.
4. 2b. Similarity bias generates temporal (but not spatial) mass suicide clusters.
5. 2c. One-to-many cultural transmission, i.e. the cultural dynamics characterised by the mass media, generates temporal (only not spatial) mass suicide clusters.

Methods

The amanuensis-based model was programmed in Borland C++ Architect. An executable (.exe) version is available for download as Supplementary File S1 and may exist used to recreate the results presented below. Source code is available upon asking from the author.

The freely-available program SaTScan™ [25] was used to observe spatial, temporal, and spatiotemporal clusters in the suicide frequency data generated by the agent-based model. This program is ordinarily used to detect clustering of diseases in space and time, such every bit leprosy [26], West Nile virus [27] and gonorrhoea [28]. Previous simulation studies have found that SaTScan™ is constructive in detecting clusters of rare events [29] making it especially applicable to suicides. SaTScan™ uses the scan statistic [30] to identify statistically significant clusters, i.e. clusters that deviate from frequencies expected under a random distribution. A window of varying size is gradually moved across time and/or space and the number of observed events (here, suicides) is compared with the number expected under a random, no-clustering distribution. This window is either an interval in fourth dimension (for temporal scanning), a circle (for spatial scanning) or a cylinder with a circular spatial base and a fourth dimension interval as its length (for spatio-temporal scanning). The maximum cluster size was set here at 50% of the full area for spatial clusters and 50% of generations for temporal clusters to avoid biasing the detection with a priori target cluster sizes. For each location and/or size of the window, the expected frequency under the null hypothesis of no clustering is calculated assuming a Bernoulli distribution, and the window with the maximum likelihood is identified. The statistical significance of this window is calculated using a Monte Carlo simulation method. The scanning procedure and maximum likelihood examination is repeated for 999 randomly generated replications of the information generated under the null hypothesis. The statistical significance (p value) is given by the rank of the maximum likelihood calculated from the real data compared with all ranked maximum likelihoods from the fake data sets; if the real maximum likelihood falls inside the top α proportion of ranked faux maximum likelihoods, and so the null is rejected (e.grand. if α = 0.05, the real maximum likelihood must be inside the largest 5% of fake maximum likelihoods to be assigned statistical significance).

To further increase the robustness of the analysis in the nowadays study, data from ten independent runs of the agent-based model for each set of parameter values were analysed using SaTScan™. The results below are given as the proportion (X) of these 10 runs that yielded pregnant clusters at the p<0.005 level (given ten tests, α is Bonferroni corrected to 0.05/10 = 0.005), either in space (X_s), fourth dimension (X_t), or both fourth dimension and space (X_st). Thus where X_s = 0, X_t = 0 or Ten_st = 0 and then there is no spatial/temporal/spatiotemporal clustering beyond that expected due to chance, and where X_s = 1, X_t = i or X_st = i and then it is statistically most likely that at least one spatial/temporal/spatiotemporal cluster is present in the data generated by the model.

Results

Basic model assumptions

The model assumes N = m agents inhabiting a two-dimensional ten×ten grid, with 100 groups each located at a different Cartesian coordinate and x agents in each group. This organisation was intended to simulate the kind of social structure often examined in suicide cluster studies (eastward.g. [5]) and every bit such may be bathetic to different levels of social organisation, e.1000. a collection of schools/hospitals within a town, towns within a state, or states within a state. The population then undergoes T = 100 generations. During each generation, every agent is cycled through in a random society and commits suicide with a probability that is adamant by various parameters described in the post-obit sections and summarised in Table ane. If an agent commits suicide, it is replaced with a new agent and, in the social learning weather condition, affects the surrounding agents' probabilities of suicide in the following generation. Each suicide is recorded as a case and the entire 100-generation dataset is analysed for clusters using SaTScan™.

During the assay information technology became credible that analysing 100 generations from an initial no-suicide land generated artifactual temporal clusters as suicides emerged during the beginning few generations due to social learning, homophily or other processes. Given that real-life suicide cluster data does not offset arbitrarily at zilch suicides, in the agent-based model 110 generations were run in full, the start 10 generations were ignored and generations 11–110 analysed for clusters.

Each agent is initially given the aforementioned stock-still probability of committing suicide, p₀ . This baseline probability is then modified co-ordinate to a prepare of risk factors, intended to capture individual differences in suicide rates. For example, data from the U.Due south. [31] advise gamble factors of gender (men are iii.ix times more likely to commit suicide than women), ethnicity (white people are 2.ii times more likely than non-white people) and historic period (over 65s are ane.five times more likely than xv–24 twelvemonth olds). These gamble factors announced to combine additively, east.one thousand. white men aged over 65 have the highest compounded risk of suicide. Risk factors are represented in the model as a set up of six binary bits, k_i , where i indexes the six risk factors (i = {1, 2…6}) and k_i ∈ {one−q, 1+q}. Each bit therefore indicates whether an agent is at college (ane+q) or lower (i−q) chance of suicide (e.g. male person vs. female), and are randomly generated for each agent (except in the example of homophily, encounter below). The probability of suicide later on modification by the risk factors, p₁ , is then the product of these take a chance factors (Equation 1). (ane)

The magnitude of q thus determines the individual variation in p_ane within the population. Except where indicated otherwise, in the simulations below q = 0.2; 6 chance factors with q = 0.ii gave a suitable range of individual variation across the population, from p₀ (ane−q)⁶ = 0.26p₀ to p₀ (1+q)⁶ = 2.99p₀ . Evidently risk factors in the real world are much more than circuitous than this (e.g. age is continuous not dichotomous and there may exist more or less than six factors that may interact non-independently). However, the above implementation captures the essential phenomena of private differences in hazard factors in an abstract, simplified way that is easily implemented in silico. An example fourth dimension series with a small baseline risk of suicide of p₀  = 0.005 and private differences of q = 0.two is provided in Figure 1A, which shows rare suicide events distributed randomly in fourth dimension and space.

Figure ane. Three time serial indicating (A) baseline suicide occurrences with no clustering, (B) a spatiotemporal cluster resulting from social learning, and (C) a spatial cluster resulting from homophily.

Each square inside the 10×10 grid indicates one 10-agent sub-group, with the colour of the foursquare indicating the frequency of suicide from green (0%) to cherry (100%). In A, randomly distributed suicide events tin exist observed due to the non-copycat probability of suicide (p₀  = 0.005). No clustering is detected nether these conditions. In B, a spatiotemporal point cluster generated by social learning (s = 5) is marked with a red circumvolve, and can be seen persisting over a flow of 3 generations from t = 73 to t = 75 inclusive, thus showing localisation in both time and space. In C there is no social learning (s = 0), but homophily (h = 1) and large inter-group differences (q = 0.iv) causes i sub-group, marked with a ruddy circle, to exist composed entirely of high suicide gamble agents. This group repeatedly features suicides throughout the simulation run, forming a spatial (but not temporal) cluster despite the lack of social learning.

https://doi.org/10.1371/journal.pone.0007252.g001

Social learning (south)

Whenever an amanuensis commits suicide, it increases the probability that every other agent in its 10-agent group will commit suicide during the following generation according to the parameter due south (s≥0) as in Equation 2, where p_ii is the modified probability of suicide following social learning and x_n is the number of agents in the same group in the previous generation who committed suicide. (2)

Thus social learning is assumed to be additive, with each suicide "observed" by an agent increasing their suicide take a chance p_ii in the adjacent generation by an equal amount. Note that this does non apply to the new agent that replaced the suicide agent, or whatever other new agents in that generation who did non "observe" the suicide in the previous generation. Also annotation that the (i+ 10_n southward) term in Equation 2 constitutes a measure out of relative risk, RR (where RR  = p₂ /p₁ ): when there is no social learning (s =  0) so there is a relative run a risk of 1, and exposed individuals have the same probability of suicide as unexposed individuals; the relative risk then increases every bit s increases.

Social learning within local groups is predicted to result in the reliable spatiotemporal clustering of suicides as agents acquire suicide behaviour from members of their local grouping. Table ii shows the incidence of spatial, temporal and spatiotemporal suicide clusters nether different values of both p₀ and south. First, note that clusters never occur when s = 0, i.eastward. in the absence of social learning (illustrated in Figure 1A). As s increases in magnitude, the probability of observing clusters increases, but only for sufficiently big values of p₀ (p₀  = 0.005 or p₀  = 0.01). For these values, spatiotemporal clusters are most likely to sally, followed by purely spatial clusters, and then purely temporal clusters (eastward.g. for p₀  = 0.005 and s = 5: X_s = 0.2; 10_t = 0; 10_st = 0.8). Figure 1B shows a time series of an instance spatiotemporal signal cluster, in which a single group temporarily exhibits unduly more suicides than surrounding groups. These results therefore back up Hypothesis 1a that social learning generates suicide clusters, and specifically spatiotemporal betoken clusters.

Homophily (h)

To simulate homophily, new agents created at the first of the simulation re-create the yard_i bits of a previously created agent in the aforementioned group with a probability h (0≤h≤one). The first agent in the group takes random thousand_i values as described to a higher place. Thus where h = 0 in that location is no homophily and agents never share bits across that expected past chance. Where h = i at that place is strict homophily: every agent in the same grouping shares identical g_i bits and different groups vary in their $.25 (i.e. no inside-group variation and high between-group variation). As some of these groups will by chance have uniformly high gamble factors due to the variation acquired by q, these are the groups we would look to form suicide clusters even with no social learning. New agents introduced to supercede agents that have committed suicide accept the same 1000_i bits of a randomly selected agent in their group in club to maintain the same level of homophily throughout the simulation run. The use of binary bits to simulate homophily is based in part on previous amanuensis-based simulations [32], [33], although in the present model homophily is assumed to accept occurred before the simulations begin, rather than emerging during the simulations.

Tabular array iii shows the probability of observing clusters in response to dissimilar levels of h and the parameter q (the extent of individual differences in baseline suicide risk), which was institute to strongly moderate the upshot of h. When in that location is naught private variation in suicide adventure (q = 0) so no clusters are observed even nether maximum homophily (h = 1). As q increases, clustering becomes more than frequent under high levels of homophily. Here, purely spatial clusters are more common than spatiotemporal clusters, while purely temporal clusters are never observed (due east.1000. for q = 0.2 and h = 0.75: X_south = 0.7; X_t = 0; Ten_st = 0.2). An example of a homophily-generated spatial cluster is illustrated in the time series in Figure 1C, in which a unmarried loftier-risk group repeatedly experiences a disproportionately high frequency of suicides throughout the entire simulation run. The model therefore lends only fractional support to Hypothesis 1b, that homophily on suicide hazard factors tin can mimic the spatiotemporal clustering shown above to result from social learning, with the two qualifications that (i) private differences in suicide risk factors must be sufficiently large and (two) while spatiotemporal clusters are observed, purely spatial clusters are more than likely to be observed, the reverse of that documented for social learning, in which spatiotemporal clusters are more likely than purely spatial clusters.

Prestige bias (c)

Two parameters were used to simulate a minority of prestigious "celebrities" whose suicides have an increased social influence on other agents' suicide risks. These parameters are c_p (0≤c_p ≤1), which specifies the probability that a new amanuensis is assigned celebrity condition, and c_s (c_southward ≥0), which specifies the increase in p_i of another agent in the aforementioned group as a upshot of observing a celebrity agent committing suicide in the previous generation. Thus if x_n is the number of non-glory agents in a detail group who in the previous generation committed suicide, and x_southward is the number of celebrity agents in the same group who committed suicide in the previous generation, then the suicide risk of surviving agents in that group is now given past Equation 3. (3)

Thus c_s replaces southward for celebrity agents, and prestige bias is operating when c_s > s such that celebrities have a greater social influence than non-celebrities.

Table four shows the upshot of prestige bias on the probability of clustering, bold values of p₀ and southward that would commonly not generate clustering (p₀  = 0.01, southward = 1; come across Table 2). Increasing the strength of prestige bias c_s increases the probability of observing spatiotemporal clusters, and to a lesser extent purely spatial and purely temporal clusters (e.k. for c_s  = 20 and c_{due south}  = 0.1: X_southward = 0.3; X_t = 0.2; X_st = 1). Nonetheless, the strength of prestige bias (c_southward ) must be essentially larger than the non-prestige social learning force (south), with a 20-fold increment in suicide risk in response to celebrity suicides needed to reliably generate spatiotemporal point clusters. Moreover, Tabular array iv also shows that the strength of prestige bias must be larger equally the proportion of agents who are prestigious celebrities gets smaller (i.e. c_p decreases). Overall, so, prestige bias can mimic non-prestige biased social learning in generating spatiotemporal clusters when prestige bias is sufficiently stiff to counteract the lower frequency of prestige-based suicides. Hypothesis 2a, notwithstanding, states that prestige bias lonely should generate mass (temporal) clusters rather than spatiotemporal clusters, and thus was non supported by the model.

Similarity bias (m)

Here it is assumed that agents only influence each others' probability of suicide if they share at least 1000 (0≤m≤six) of the six chiliad_i bits that draw private differences in hazard factors. When yard = 0, none of the one thousand_i bits need to be shared, and similarity bias is not operating. When k = 6, learners and models must share all vi k_i $.25 in order for the learner'south p₂ to be affected past s. Thus the college the value of chiliad, the stronger is the similarity bias (i.e. the more than similar the model must be to the learner in society for the learner to exist influenced by their behaviour).

Table 5 shows that increasing thousand from 0 (no similarity bias) to 6 (agents must be identical to appoint in social learning) reduces the frequency of all types of clusters, with no clusters occurring in the extreme case where chiliad = half dozen (10_s = X_t = X_st = 0). This might be expected, given that similarity bias reduces the set of models from whom suicide behaviour can be learned. Given that social learning generates clusters (Table 2), in blocking social learning similarity bias also eliminates clusters. Hypothesis 2b, that similarity bias generates temporal (mass) suicide clusters, is therefore not supported. However, Tabular array 5 also shows that homophily (h = one) removes the inhibitory upshot of similarity bias, with clusters virtually universally observed for big values of yard (e.g. for h = 1 and m = 6: 10_southward = one; 10_t = 0.viii; Ten_st = 1). This is to be expected: when h = i, all agents within a sub-group are identical, and so even when similarity bias is at its strongest (thousand = vi) social learning still occurs. However, given that these clusters are spatial besides every bit temporal, and that homophily acts to partially mask the clustering issue of social learning, this farther undermines Hypothesis 2b that similarity bias generates mass (temporal only) clusters as a result of social learning.

One-to-many manual (r)

The one-to-many transmission consequences of the mass media is faux by manipulating the radius of a "zone of social influence" across which social learning of suicide behaviour occurs. Thus when an agent commits suicide, in the post-obit generation every amanuensis in every grouping that is inside r (0≤r≤ix) sectors from the suicide agent's grouping has their suicide probability p₁ updated according to Equation three. Where r = 0, only the suicide agent'southward group is afflicted, as assumed in all of the simulations discussed previously. Where r = i, every amanuensis in the eight groups immediately surrounding the suicide agent's group is afflicted (or fewer groups if the focal grouping is on the border of the grid). In the extreme case where r = nine, the zone of social influence encompasses the entire grid and all m agents in the population are afflicted by every suicide.

Table half dozen shows that, for values of p₀ and s that would not normally produce clusters (p₀  = 0.005, southward = ane), a small increase in r increases the probability of detecting clusters, predominantly spatial and spatiotemporal clusters (e.g. for r = 3: X_south = 1; Ten_t = 0.6; X_st = 1). This is because at that place are now more than agents who are affected by s, thus increasing the probability of a cluster occurring. Still, large values of r neglect to generate clusters of whatsoever kind (e.g. for r = 9: 10_s = 10_t = 10_st = 0). The reason clusters were not observed at large values of r was that the widespread social learning causes a suicide pandemic such that most the entire population constantly committed suicide during every generation. Such a pandemic is illustrated in Figure 2A. As suicide rates are at a constantly high rate, there are no clusters in either fourth dimension or infinite. Obviously, such a pattern of constant mass suicide is highly unrealistic. Overall, Hypothesis 2c, that one-to-many manual generates mass clusters, was therefore not supported nether whatsoever of these values of r.

Figure two. Two time series illustrating the effects of strong one-to-many transmission (r = 9).

In A, when the baseline suicide rate and the strength of social learning are relatively loftier (p₀  = 0.005, s = 1), a pandemic causes the unabridged population to commit suicide at extremely high rates throughout the simulation run. Neither spatial nor temporal clusters are observed nether these conditions, which are obviously highly unrealistic. In B, when the frequency of social learning is reduced by introducing prestige bias (p₀  = 0.005, s = 0, c_p  = 0.01, c_{due south}  = v) such that but a pocket-size minority of agents have social influence, mass (temporal simply non spatial) clusters sally. Here, one of the iv suicides that occur in generation t = 84 was a prestigious "celebrity", resulting in a mass cluster in the following iii generations. Suicide rates and so drop dorsum to baseline pre-cluster levels at generation t = 88.

https://doi.org/x.1371/journal.pone.0007252.g002

However, Table six also shows three cases where mass clusters were observed. In these cases the outcome or frequency of copycat suicide is reduced such that suicide pandemics fail to take off, yet copycat suicides are non so weak or infrequent that clusters do non occur. The offset is when the strength of social learning (southward) is direct reduced (east.g. for r = 9 and south = 0.i: X_s = 0; X_t = 0.7; X_st = 0.iii). The 2nd is where similarity bias operates to reduce the frequency of social learning events (e.thou. for r = 9, m = v: X_s = 0; Ten_t = 0.8; X_st = 0.iii). The third is where prestige bias reduces the subset of agents who have social influence (due east.g. for r = 9, c_p  = 0.01, c_s  = five: 10_s = 0; X_t = 1; 10_st = 0.9). In each of these cases the probability of pandemics such as those observed in Figure 2A is reduced either past reducing the forcefulness of social learning (southward = 0.1) or reducing the frequency of social learning events (c_p  = 0.01 or m = 5). Instead, temporary clusters occur that are localised in time earlier returning to baseline suicide rates. When r is large, these clusters bear on all agents in the population equally and so are not spatially localised. Such a mass cluster is illustrated in Figure 2B. Hypothesis 2c is therefore supported only under the weather condition where ane-to-many transmission is strong enough to eliminate spatial clustering and where social influence is strong enough to generate statistically pregnant clusters yet non so strong as to cause population-wide suicide pandemics.

Discussion

Evidence accumulated during contempo years suggests that suicide may be subject to social learning, potentially resulting in distinct clustering of suicides in time and/or space. Point clusters are clusters of suicides in both time and infinite, and accept been attributed to social learning within local groups [5]. Mass clusters are clusters of suicides in time merely non space, and have been attributed to prestige and similarity bias (preferentially copying prestigious or like models: [19]) and the mass media [11], [23]. The present written report used agent-based modelling techniques, in combination with rigorous statistical cluster-detection analyses, to assess the validity of these proposals. Naturally, abstruse simulation models cannot give definitive answers to questions concerning copycat suicides that are ultimately empirical. Still, they tin can assistance to clarify definitions of different processes with greater precision than informal verbal explanations, they can lend plausibility to hypotheses by demonstrating that assumed consequences logically follow from premises, and they can guide future empirical work by identifying the kinds of variables that might exist important and that future empirical work should focus on.

The prediction that social learning within groups of agents generates spatiotemporal signal clusters was supported. An additional hypothesis, that homophily generates spatiotemporal clusters in the absence of social learning considering individuals who are independently at high chance of suicide congregate in space and form not-social suicide clusters [3], was only partially supported. Homophily only generated clusters when at that place was relatively high individual variation in agents' (non-copycat) suicide adventure, such that high-risk clusters occur. Furthermore, these homophily clusters were most probable to be spatial, to a lesser extent spatiotemporal, and never purely temporal. This makes sense given that groups maintained their relative levels of gamble throughout the simulation, and there is no reason why the agents would cluster their suicides in time without social learning. These findings might be used to guide futurity empirical tests of Joiner's [3] homophily hypothesis, by specifically taking into account the degree of individual variability in known suicide risk factors (e.g. historic period, sexual activity, ethnicity) in a region, and by distinguishing betwixt the spatial-but-not-temporal clusters generated by homophily and the spatiotemporal clusters generated past social learning.

A second set of simulations found that neither prestige bias (preferentially copying prestigious celebrities) nor similarity bias (preferentially copying others who are similar to oneself) generate mass (temporal-just-not-spatial) clusters alone. Both prestige and similarity bias act to reduce the subset of potential models from whom suicide-related behaviour tin be learned. For prestige bias, this is because merely a minority of the population tin 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 1 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 atmospheric condition, such equally increasing the forcefulness of prestige bias and introducing homophily, which made neighbouring agents similar to one some other and therefore more likely to copy each other even at high levels of similarity bias. Yet even under these conditions (strong prestige bias, homophily) mass clusters were no more likely to sally than purely spatial clusters or spatiotemporal clusters.

Notwithstanding, the mass media, represented here by one-to-many transmission, did generate mass clusters, but only under certain conditions. When social influence was also strong, all-encompassing 1-to-many transmission gave ascent to suicide pandemics in which all agents committed suicide with an extremely high probability. These pandemics neither contained any clusters nor were very realistic. Mass clusters did emerge, still, when social influence was weak, either direct via a reduced forcefulness of social learning, or indirectly via prestige bias or similarity bias, which both reduced the subset of models that agents could exist influenced by. In summary, prestige and similarity bias were neither necessary nor sufficient for mass clusters, while one-to-many manual was necessary only not sufficient. The three processes in combination generated mass clusters, which is consequent with sociological evidence for each in bodily cases of mass suicide clusters. Still, the model highlights the very different roles that each plays: one-to-many transmission acts to spread suicide behaviour beyond the entire population thus eliminating spatial clustering, while prestige and similarity bias somewhat counter-intuitively (and in dissimilarity to previous suggestions: [19]) prevent copycat suicides from persisting and becoming pandemic.

Obviously several assumptions of this model are farthermost simplifications of a complex real-life miracle. For example, the implementation of prestige and similarity bias in the present model but incorporated certain, simplified aspects of these processes, ignoring for example potential runaway prestige furnishings [22], prestige hierarchies [xx] and the consequences of similarity bias on private variation [32], [33]. There is also no consideration of the mechanism by which 'social influence' occurs: social influence via the transmission of practical cognition regarding suicide methods might have quite different consequences to social influence via the emotional result of a close friend's suicide. A further source of potential inaccuracy is the mismatch betwixt parameter values in the model and equivalent real-life estimates. The baseline suicide rate that is required in the model (0.001≤p₀ ≤0.01) to detect statistically pregnant clusters is higher than actual national suicide rates (e.grand. xi in 100,000, or 0.00011 in the USA in 2005: [31]), although this is peradventure considering of the much smaller population size in the model compared to bodily national populations. The assumed strength of social influence south might as well be considered large (e.k. s = 5, or for prestige bias c_s  = 20) compared to estimates that publicised suicide stories increase the national suicide rate by just two.v% [ten] or that simply two–4% of suicides testify any spatiotemporal clustering [5]. However, information technology should be noted that under some conditions of the model much smaller values of s reliably generated clustering (eastward.chiliad. when r = 9, clusters occurred when due south = 0.one), and more detailed individual-level studies have institute relatively large estimates of social influence. For example, 1 study found that teenagers who knew some other person who had committed suicide were three times more than likely to commit suicide than teenagers who did not know anyone who had committed suicide [34]. Still, even with simplified assumptions and exaggerated parameter values, the findings reported higher up can exist useful in showing qualitatively how a change in 1 variable (east.1000. the magnitude of individual differences) interacts with another (e.yard. homophily) to cause some effect (e.g. increased spatial clustering). These relationships tin can and then be tested in actual datasets.

In supporting the assumptions made by sociologists that point and mass clusters can be taken equally show that suicide may spread via social learning, the model reinforces the need for efforts to counter the social transmission of suicide-related information. The findings related to signal clusters suggests that social learning and homophily generate distinct types of clusters (predominantly spatiotemporal versus predominantly spatial); by using this cognition to distinguish between copycat signal clusters and homophilous betoken clusters, efforts to reduce social transmission might be more than effectively targeted at the former. The findings related to mass clusters in detail highlight the need for media guidelines that restrict the dissemination and glorification of suicides, as already introduced in many countries [eleven], [23]. More specifically, the model suggests that increasing the range of one-to-many transmission (r), increasing the social influence of prestigious celebrities (c_s ) and increasing the proportion of the population who are assigned celebrity status (c_p ) can all increment the probability of widespread suicide pandemics. Anecdotally, all 3 of these trends appear to exist occurring in many countries in contempo years: satellite television receiver and the internet take increased the global range of the mass media; celebrities such equally film actors and popular singers are existence assigned increasing importance relative to politicians and intellectuals (whose suicides do non elicit copycat suicide attempts); and reality idiot box programmes are increasing the number of celebrities within society. This highlights how media guidelines on suicide reporting will become all the more important in the futurity.

Supporting Information

Author Contributions

Analyzed the data: AM. Wrote the paper: AM. Designed the model: AM. Programmed the model: AM.

References

1. 1. Stack Southward (2000) Suicide: A 15-year review of the sociological literature. Role I: Cultural and economical factors. Suicide and Life Threatening Behavior thirty: 145–162.
   • View Article
   • Google Scholar
2. two. Stack S (2000) Suicide: A fifteen-twelvemonth review of the sociological literature. Part II: Modernization and social integration perspectives. Suicide and Life-Threatening Behavior 30: 163–176.
   • View Article
   • Google Scholar
3. three. Joiner JTE (1999) The clustering and contagion of suicide. Electric current Directions in Psychological Science 8: 89–92.
   • View Article
   • Google Scholar
4. 4. Gould MS, Wallenstein Due south, Davidson L (1989) Suicide clusters: A critical review. Suicide and Life-Threatening Beliefs 19: 17–29.
   • View Article
   • Google Scholar
5. 5. Gould MS, Wallenstein S, Kleinman M (1990) Fourth dimension-space clustering of teenage suicide. American Periodical of Epidemiology 131: 71–78.
   • View Article
   • Google Scholar
6. half-dozen. Haw CM (1994) A cluster of suicides at a London psychiatric unit. Suicide and Life-Threatening Behavior 24: 256–266.
   • View Article
   • Google Scholar
7. 7. Brent DA, Kerr MM, Goldstein C, Bozigar J, Wartella M, et al. (1989) An outbreak of suicide and suicidal behavior in a high school. Journal of the American Academy of Child & Adolescent Psychiatry 28: 918.
   • View Article
   • Google Scholar
8. viii. Phillips DP (1974) The influence of suggestion on suicide: Substantive and theoretical implications of the Werther consequence. American Sociological Review 39: 340–354.
   • View Article
   • Google Scholar
9. 9. Stack Southward (1987) Celebrities and suicide: A taxonomy and analysis, 1948–1983. American Sociological Review 52: 401–412.
   • View Article
   • Google Scholar
10. ten. Stack S (2000) Media impacts on suicide: A quantitative review of 293 findings. Social Science Quarterly 81: 957–971.
    • View Article
    • Google Scholar
11. 11. Stack South (2003) Media coverage as a risk factor in suicide. Journal of Epidemiology and Community Health 57: 238–240.
    • View Article
    • Google Scholar
12. 12. Jonas G (1992) Modelling and suicide: A test of the Werther effect. British Journal of Social Psychology 31: 295–306.
    • View Commodity
    • Google Scholar
13. 13. Ishii K (1991) Measuring common causation: Effects of suicide news on suicides in Nippon. Social Scientific discipline Enquiry 20: 188–195.
    • View Commodity
    • Google Scholar
14. fourteen. Cheng ATA, Hawton K, Chen THH, Yen AMF, Chang JC, et al. (2007) The influence of media reporting of a glory suicide on suicidal behavior in patients with a history of depressive disorder. Periodical of Melancholia Disorders 103: 69–75.
    • View Article
    • Google Scholar
15. 15. Sonneck One thousand, Etzersdorfer Due east, Nagel-Kuess S (1994) Imitative suicide on the Viennese subway. Social Scientific discipline and Medicine 38: 453–457.
    • View Article
    • Google Scholar
16. 16. Gilbert North (2007) Agent-based models: Quantitative applications in the social sciences. London: SAGE Publications.
17. 17. Epstein JM (2007) Generative social science: Studies in agent-based computational modeling. Princeton, NJ: Princeton University Printing.
18. eighteen. McPherson M, Smith-Lovin L, Cook JM (2001) Birds of a feather: Homophily in social networks. Annual Review of Folklore 27: 415–444.
    • View Article
    • Google Scholar
19. 19. Henrich J, McElreath R (2007) Dual inheritance theory: The evolution of man cultural capacities and cultural evolution. Oxford Handbook of Evolutionary Psychology Oxford University Press, Oxford.
20. 20. Henrich J, Gil White FJ (2001) The evolution of prestige. Evolution and Man Behavior 22: 165–196.
    • View Article
    • Google Scholar
21. 21. McElreath R, Boyd R, Richerson PJ (2003) Shared norms and the evolution of indigenous markers. Current Anthropology 44: 122–130.
    • View Article
    • Google Scholar
22. 22. Boyd R, Richerson PJ (1985) Civilisation and the evolutionary process. Chicago: Univ. Chicago Press.
23. 23. Gould MS, Jamieson P, Romer D (2003) Media contagion and suicide amongst the young. American Behavioral Scientist 46: 1269.
    • View Commodity
    • Google Scholar
24. 24. Cavalli-Sforza LL, Feldman MW (1981) Cultural manual and evolution. Princeton: Princeton.
25. 25. Kulldorff Thousand (2006) Satscan: Software for the spatial and space-time scan statistics. 7.0 ed: Information Direction Services Inc.
    • View Article
    • Google Scholar
26. 26. Bakker MI, Hatta M, Kwenang A, Faber WR, van Beers SM, et al. (2004) Population survey to determine run a risk factors for Mycobacterium leprae transmission and infection. International Journal of Epidemiology 33: 1329–1336.
    • View Article
    • Google Scholar
27. 27. Mostashari F, Kulldorff 1000, Hartman JJ, Miller JR, Kulasekera V (2003) Dead bird clusters every bit an early warning organisation for West Nile virus activity. Emerging Infectious Diseases 9: 641–646.
    • View Article
    • Google Scholar
28. 28. Jennings JM, Curriero FC, Celentano D, Ellen JM (2005) Geographic identification of high gonorrhea transmission areas in Baltimore, Maryland. American Journal of Epidemiology 161: 73–80.
    • View Commodity
    • Google Scholar
29. 29. Aamodt G, Samuelsen SO, Skrondal A (2006) A simulation study of three methods for detecting affliction clusters. International Journal of Wellness Geographics 5: xv.
    • View Article
    • Google Scholar
30. 30. Kulldorff Chiliad (1997) A spatial browse statistic. Communications in Statistics-Theory and Methods 26: 1481–1496.
    • View Commodity
    • Google Scholar
31. 31. Kung HC, Hoyert DL, Xu J, Murphy SL (2008) Deaths: Final data for 2005. National Vital Statistics Reports 56: 9.
    • View Article
    • Google Scholar
32. 32. Axelrod R (1997) The dissemination of civilization: A model with local convergence and global polarization. Journal of Conflict Resolution 41: 203–226.
    • View Article
    • Google Scholar
33. 33. Durrett R, Levin SA (2005) Can stable social groups be maintained by homophilous faux lone? Journal of Economic Behavior and System 57: 267–286.
    • View Article
    • Google Scholar
34. 34. Cutler DM, Glaeser EL, Norberg M (2001) Explaining the ascent in youth suicide. In: Gruber J, editor. Risky behavior among youths: An economic analysis. Chicago: University of Chicago Printing. pp. 219–269.

taitrourts.blogspot.com

Source: https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0007252

Share this post