The main focus of this work is to develop a BN model for analyzing SAD, based on some cognitive-behavioral predictors, observed symptoms and a priori known causal relationships, with the precise objective of helping physicians and mental health professionals in their decisional process. Whereas Bayesian network has two parts, the process of BN construction is to construct structure and conditional probability tables.
The Bayesian network construction
There are three methods to construct Bayesian networks: manual construction, learning and a combination of them (Yu et al. 2009). In this work, we construct the structure of our BN model manually using domain knowledge and interviews with experts. In this process, variables and relationships between them should be determined. The first stage in manual construction is the identification of the important variables generally based on interviews with experts and descriptions of the domain. It is important to limit variables and choose important variables which are target variables and observation (evidence) variables. Target variables are outputs of network and what we want to know and observation variables are inputs of network (Yu et al. 2009). After that, the dependence and independence relationships among the variables have to be analyzed and expressed in the graphical structure (Lucas et al. 2004).
After the modeling stage, the Bayesian inference is used to update the network statistical knowledge based on current observations and the Bayes theorem (Curiac et al. 2009). Inference in a BN means computing the conditional probabilities for some variables, given evidence concerning other variables. Evidence is produced by responses to clinical questions (tests, signs or symptoms) (Olmus 2004).
The BN graphical model is presented in Fig. 1. Analyzing the structure presented in Fig. 1, we can observe a number of nodes having a significant influence inside the network. We describe them below.
There has been a continuing controversy between categorical and dimensional approach to social anxiety disorder. Researchers with the dimensional approach consider social anxiety as a continuum; in line with this view, Rapee and Spence’s (2004) model is one of the models that are based on the dimensional approach to social anxiety. Rapee and Spence (2004) introduced some general factors in their model that contribute to the development of social anxiety disorder. According to this model and literature, we chose some factors.
Shyness (X4 node) is one of the important genetic factors in the field of social anxiety disorder. The definitions of shyness and social anxiety disorder in somatic, cognitive and behavioral symptoms are alike (Heiser et al. 2009). In this research, Stanford Shyness Survey (Zimbardo 1977) was used for screening shyness and measuring severity of it.
Another related temperament to SAD is behavioral inhibition (BI), which refers to a child response to novel stimulus by behavioral withdrawal, timidity, increased vigilance and excessive arousal (Kagan et al. 1988). Some evidence indicate that BI contributes to development of anxiety disorders (Perez-Edgar and Fox 2005) especially social anxiety disorder (e.g., Coplan et al. 2006; van Brakel et al. 2006). For measuring the two variables of childhood behavioral inhibition (X1 node) and adulthood behavioral inhibition (X2 node) respectively, Retrospective Measure of Behavioral Inhibition and Adult Measure of Behavioral Inhibition (Gladstone and Parker 2005) were used.
To consider the interpersonal deficits in SAD, attachment theory can be useful (Eng et al. 2001). This theory suggests that the experiences with earlier significant others can be generalized to future interpersonal functioning. In this regard, Michelson et al. (1997) found that social anxiety disorder was positively related to avoidant and anxious styles. Furthermore, for measuring the two variables of anxious attachment style (X5 node) and avoidant attachment style (X6 node), the Adult Attachment scale questionnaire (AAS; Collins and Read 1990) was used.
Cognitive-behavioral models (e.g., Rapee and Heimberg 1997; Clark and Wells 1995) emphasized the dysfunctional cognitive processes in the maintenance of SAD (Hofmann 2007). Social anxious people showed different kinds of bias in information processing such as interpretation bias. Individuals with social anxiety disorder misinterpret the ambiguous social situations (Heinrichs and Hofmann 2001). Interpretation bias has various components like negative self-evaluation (X8 node) and perceived negative evaluation by others (X9 node) (Heimberg and Becker 2002). For measuring these two variables, the Consequences of Negative Social Events Questionnaire (CONSE-Q) (Wilson and Rapee 2005) was used.
Self-efficacy refers to “the conviction that one can successfully execute the behavior required to produce the outcomes” (Bandura 1977). High socially anxious individuals are likely to devalue their social performance; they succeed objectively though (e.g., Clark and Wells 1995). In this regard, Gaudiano and Herbert (2006) noted that self-efficacy might be especially important for understanding the SAD and it has a moderately inverse correlation with social anxiety. For the measurement of the variable of social self-efficacy (X13 node), the social situation scale questionnaire (SESS; Gaudiano and Herbert 2003) was used.
Thus, in the current research, we assumed shyness and behavioral inhibition (childhood and adulthood) as genetic factors and attachment (secure, avoidant and anxious) as parent influence. Social self-efficacy stemmed from poor social skills construct and interpretation bias (negative self-evaluation and perceived negative evaluation by others) arose from interrupted social performance construct.
In general, the symptoms of SAD reflect a fear of being embarrassed or humiliated in front of others. According to Rapee and Heimberg’ (1997) model, anxiety can be seen in behavioral, cognitive and physical symptoms in socially anxious people. Therefore, the physical symptoms reported by many persons with social phobia include sweating, trembling, blushing, palpitations, nausea, twitching, shaky voice and dry mouth (X11 node) (Brunello et al. 2000). In this research, for collection data, Behavioral symptoms of Anxiety (X12 node) were also rated by asking a number of questions. To consider mentioned symptoms, social anxiety disorder is associated with some impairment in social function (Eng et al. 2005) and significant interference to different domains of life especially in the career, academic and interpersonal functioning (X10 node) (Hofmann and Barlow 2002).
For the three aforementioned variables, a number of questions were designed in the form of a questionnaire and they were rated so that their value is determined using the sum of the scores.
In sum, in each questionnaire, the range of variation of the total scores was different and we classified this range for each variable into three groups: mild, moderate, and severe. In fact, all continuous variables except ‘social anxiety’ were discretized into three relatively equal-width intervals.
In Fig. 1, we have the qualitative representation of the Bayesian network. We need to specify the quantitative representation of our Bayesian network that is the set of conditional probability tables of the nodes.
There are several commercial and research tools designed for BN model authoring and testing. Among the most popular of these tools are Hugin, Netica, and GeNIe. We used Netica software from Norsys (Netica, www.norsys.com) for the construction of the Bayesian network, because of its simplicity and high performance. Netica allows network construction and parameter learning from data. Parameter learning determines the conditional probability table at each node. According to data, we can achieve prior and conditional probabilities. In order to learn the CPTs, data were gathered from a number of university students.
Data collection for learning and testing the model
The statistical population consisted of the students of a university, from which a sample from the five educational groups of human sciences, technical sciences, medical sciences, basic sciences and arts was selected randomly and in multiple stages. Three faculties from each group and a number of students from each faculty were randomly selected. The questionnaires were given to the volunteers in each faculty. In sum, 438 students of a university (218 male and 220 female) participated in this study (first phase). All participants completed Social Phobia Inventory, Stanford Shyness Survey, Adult Measure of Behavioral Inhibition, Retrospective Measure of Behavioral Inhibition, Consequences of Negative Social Events Questionnaire, Self-efficacy for Social Situation Scale and Adult Attachment Scale. After data collection, the average age of the participants was 21.37 ± 2.43.
In the next phase, whereas the total score higher than 19 in Social Phobia Inventory indicates on likelihood of social anxiety disorder, some students with the total score of over 19 were invited to participate in the diagnostic interview. Consistent with the nature of SAD that discussed or some other reasons that were not identified in the current research, some of these students refused to participate in the interview. After interviewing with Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) (First et al. 1994), 22 of the participants were certainly diagnosed with SAD. Among these individuals, the lowest total score in the SPIN questionnaire was 25. Therefore, in this article, the cutoff value for social anxiety variable for distinguishing people without social phobia from others, score 24 was used (instead of score 19 which is used in Western countries). (It is noteworthy that there is a need to establish the Iranian version of the SPIN and find the cutoff for the Iranian population, due to the different culture and social conditions in Eastern countries like Iran.)
This data was split into 2/3 of cases for a training set and 1/3 for a test set. The BN was trained with the training set using Netica (Fig. 2).
In the following section the BN model described above is evaluated using a case file of test data. Then the results obtained using the BN model in the diagnosis of social anxiety are presented.