To our knowledge, this is the first social network analysis of a group with both high rates of Internet use and at high-risk for suicide ideation and attempts. It is also among the first to demonstrate the potential of online social networks as a prevention platform. This may be particularly relevant to future research and interventions targeting LGB youth, who have been historically quite difficult to reach.
Representing a much larger scale network with higher numbers of personal social connection than typically described, our global structural analysis is consistent with a power-law distribution of nodal degrees. This pattern is consistent with a scale-free network structure (M. E. J. Newman et al., 2006
). Scale-free networks are frequently found in interactive networks, and have a range of interesting properties that affect diffusion processes within them. Perhaps the most relevant is the dual potential to foster the diffusion of both preventive and risk-enhancing influences for suicide in this population more rapidly than other types of network structures (Caldarelli, 2007
The results of our Monte Carlo simulations are interesting and demand further exploration using other empirically-observed social networks. The most counterintuitive of our findings is that, under most of the conditions studied, the final number of people reached is not substantially increased by starting with double or even triple the number of randomly-selected individuals. Since each starting individual represents a new chain of potential recruitment, the effect of including more chains would be expected to be additive. However, given that the peer-recruitment process is limited within a bounded sample of 100,014 individuals, and the interconnectedness within this sample is high, these recruitment chains will likely begin to merge quickly and limit the sample size that can be reached. This could also be true in certain sub-types of scale-free networks, namely, those with higher degrees of modularity (M. E. Newman, 2006
). Respondent-driven methods would be expected to exhaust local modules, which are clusters of connections between individuals, before spreading to other modules in subsequent recruitment rounds. Thus, both the global and local structural features of networks can affect diffusion processes within it.
Our approach has several important limitations. The selection of an arbitrarily chosen individual from which to start the process of constructing the network dataset is a reasonable first step. If the overall network size is expected to be relatively small, the breadth-first searching approach will achieve a complete map of the network. If the network is large enough, however, breadth-first searching may capture a less-representative slice of the overall network if data collection is terminated too soon. Although biases that may result are diminished as the sample size increases, we cannot confidently predict the overall size of the network. We sampled more than 100,000 individuals in order to help reduce any bias, further expansion of the network dataset will be needed to more definitively assess the accuracy of our findings.
A related limitation is in the choice to compile the empiric network database using a single seed node. Respondent-Driven Sampling can be used to statistically characterize a network using a subset of members within it starting from a convenience sample of initial seed nodes, but is thought to require using at least nine such seed nodes and completing at least five waves of subsequent recruitment from each seed in order to be able to overcome any initial sampling bias (Heckathorn, 2007
). Since we do not expect that we have fully explored the network of Young LGB within MySpace, using a single seed need implies that our findings may not accurately reflect the characteristics of the overall population.
In addition, this sample is limited to those individuals who participate in an online social network, and who self-identify their sexual orientation. Sexual orientation is best measured using a multiaxial items that address sexual orientation identity, sexual behavior, and sexual attraction (Saewyc, Bauer, Skay, Bearinger, Resnick, Reis et al., 2004
). Although not strictly analysis to survey methods, it could be argued that these adolescents and young adults represent a better-supported, more affluent segment of the population. However, online participation among this population appears to be becoming nearly universal, at least in the developed world (Hillier et al., 2001
). The second limitation related to participation in online networks is that we have been very conservative in requiring LGB self-identification for inclusion, thus potentially significantly underestimating the network size and characteristics. Although RDS methods have been successfully used to address this very concern in the area of HIV/AIDS (Magnani, Sabin, Saidel, & Heckathorn, 2005
). There is no a priori reason to suspect that LGB-identified individuals will differ significantly in terms of the numbers of sexual minority youth who are not LGB-identified (and therefore could potentially recruit them based on that knowledge), although this could not be simulated in this exploratory work. Future research will be needed to address whether systematic biases exist in this population so that they may be directly accounted for.
The Internet and social media technologies represent an exciting set of future possibilities for not only suicide prevention research, but also in addressing a wide array of public health concerns. However, it is important to note the ethical concerns raised by such methods (Moreno, Fost, & Christakis, 2008
). Although the data available online has been made public by the individuals studied, questions remain about whether these individuals fully appreciate the ease with which others can obtain such details, and thus have given fully-informed consent in making their information accessible to others (Allen, Burk, & Ess, 2008
). Moreover, the field is constantly in flux as technologies advance and new technologies are created. While this study used passive methods only to gather data, and the hypothetical models posed minimal risks to subjects, most authors would agree that active application of peer-driven methods to directly recruit study participants in real world settings or that would pose more than minimal risks would require parental or subject consent (Moreno et al., 2008
). There is a fleetingly small distance between our approach and activities such as the “warrantless surveillance” reputedly practiced by Federal agencies such as the NSA (Anonymous, 2009
). The ethical principles pertaining to the protection of human subjects do not change, but it is hoped that the scientific community would learn from this example and consider the ethical questions posed by these emerging technologies with more care.
This preliminary exploration demonstrates the feasibility of our approach, and its potential use in suicide prevention research with much larger samples of LGB adolescents and young adults than previously possible. Although LGB youth suicide prevention is the focus of our work, our methods can be analogously applied to other populations and prevention research questions. This would be especially true of other populations that have pose challenges to reach or “hidden” using conventional approaches (such as drug users or sex workers), or who commonly use Internet or related social media (including other adolescents and young adults) to communicate with individuals and resources outside of their immediate surroundings. Determination of network characteristics has important implications for the development of epidemiological models of both suicide-related contagion (Dodds & Watts, 2005
; Watts, Muhamad, Medina, & Dodds, 2005
) and of innovation diffusion such as social marketing or other interventions targeting these youth (Mahajan & Peterson, 1985
; Rogers, 2003
). We are only beginning to explore the importance of local network neighborhood structural factors on dissemination processes. Using empirically observed networks such as ours to model and study these structural effects may hold great importance across a wide range of public health concerns far beyond the narrow confines of suicide prevention research.