Network Analysis

History

Network analysis was developed to account for individuals' relatively complex and sophisticated explanations of human behavior. It is underpinned by the notion of a perceived causal structure, which Harold Kelly described as being implicit in the cognitive representation of attributions. The perceived causal structure constitutes a temporally ordered network of interconnected causes and effects. Properties of the structure include the following: direction (past–future), extent (proximal–distal), patterning (simple–complex), components of varying stability–instability, and features ranging from actual to potential. The structure produced might be sparse or dense in nature, depending on the number of causal factors identified. Network analysis comprises a group of techniques developed in sociology and social anthropology, and it provides a method for generating and analyzing perceived causal networks, their structural properties, and the complex chains of relationships between causes and effects.

Network Analysis Techniques

Network analysis can be conducted using semi-structured interviews, diagram methods, and matrix methods. Although interviews provide detailed individual networks, difficulties arise in that individual structures cannot be combined, and causal structures of different groups cannot be compared. The diagram method involves either the spatial arrangement of cards containing putative causes or the participant directly drawing the structure. Participants can both choose from a given set of potential causal factors and incorporate other personally relevant factors into their network. In addition, the strength of causal paths can be rated. Although these methods have the virtue of [Page 896]ensuring only the most important causal links are elicited, they might potentially oversimplify respondents' belief structures, often revealing only sparse networks.

The matrix technique employs an adjacency grid with the causes of a focal event presented vertically and horizontally along its top and side. Participants rate the causal relationship for every pairwise combination. Early studies used a binary scale to indicate the presence/absence of causal links; however, this method does not reveal the strength of the causal links. Consequently, recent studies have used Likert scales whereby participants rate the strength of each causal relationship. A criterion is applied to these ratings to establish which of the resulting causal links should be regarded as consensually endorsed and, therefore, contributing to the network.

Early studies adopted a minimum systems criterion (MSC), the value at which all causes are included in the system, to determine the network nodes. Accordingly, causal links are added hierarchically to the network, in the order of mean strength, until the MSC is reached. It is generally accompanied by the cause-to-link ratio, which is the ratio of the number of extra links required to include a new cause in the network. Network construction stops if this requirement is too high, reducing overall endorsement of the network. An alternative criterion is inductive eliminative analysis (IEA), wherein every network produced when working toward the MSC is checked for endorsement. Originally developed to deal with binary adjacency matrices, networks were deemed consensual if endorsed by at least 50% of participants. However, the introduction of Likert scales necessitated a modified form of IEA, whereby an item average criterion (IAC) was adopted. The mean strength of a participant's endorsement of all items on a network must be above the IAC, which is usually set at 3 or 4 on a 5-point scale, depending on the overall link strength. In early research, the diagrammatic networks produced using these methods were topological, not spatial. However, recent studies have subjected the matrices of causal ratings to multidimensional scaling analysis to determine the spatial structure of networks. Thus, proximal and distal effects can be easily represented. The matrix method has the advantage of ensuring the exhaustive investigation of all possible links, and as it does not rely on participants' recall, it would be expec-ted to produce more reliable results.

...