Abstract

Introduction

Materials & Methods

Results

Discussion & Conclusion

References

Discussion
Board

Participants. The RCBSHY¹² was administered to 315 consecutive admissions to a university affiliated anxiety disorders treatment clinic in Hamilton, Ontario, Canada. In all, 281 returned a completed RCBSHY. A separate sample (N=162) of patients completed a self-report version of the LSAS²⁵. The anxiety disorders treatment clinic serves people with a variety of primary DSM-IV¹ anxiety disorders including panic disorder (5.4%), panic disorder with agoraphobia (39.7%), social phobia (20.3%), obsessive compulsive disorder (21.3%), specific phobia (1.0%), posttraumatic stress disorder (1.3%), generalized anxiety disorder (0.3%) and a variety of other disorders (10.7%). Diagnoses were based on the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID)¹⁶ administered by nurses trained in the use of the structured interview.

Measures. The Revised Cheek and Buss Shyness Scale (RCBSHY)¹² is a 13-item revision of the Cheek and Buss Shyness Scale¹¹. Items on the measure (e.g. "I am often uncomfortable at parties and other social functions") assess the degree to which a person feels awkward or tense when with unfamiliar others, and explicitly do not assess a preference for being with others (sociability).

The Liebowitz Social Anxiety Scale (LSAS)^34,53 assesses a range of social interaction and performance situations (24 items), that cover the broad domain of social anxiety producing situations reported by people with social phobia. Each item is rated twice; the first rating is of fear, and the second is about avoidance.

Taxometric procedure. The MAXCOV-HITMAX procedure^37,39,41 involves analyzing the internal relationships among an array of indicators that are presumed to detect a latent typology, and are scored positively in the direction of the type. The method has been described in detail by Meehl and his colleagues^37,41, and concise descriptions of the method can be found in abundance^43,58,60,63. In brief, the method involves plotting the covariance of a pair of indicators (the y axis) across a scale created by summing the other indicators (the x-axis). These plots are examined qualitatively, where curved (inverted-U) graphs indicate a taxonic solution, and flat or U-shaped graphs are consistent with the existence of a latent dimension in the scores. The question of whether underlying types consistently produce inverted U-shaped graphs and whether underlying dimensions consistently produce flat graphs has been examined in extensive Monte-Carlo investigations, the results of which are encouraging^{19,37,39,41,58}.

Data simulation procedure. Despite the apparent robustness of MAXCOV-HITMAX as evidenced by Monte-Carlo simulations, there is evidence^23,43 that suggests that taxometric methods may be seriously prone to false positive results. In an effort to correct for this possibility, we used a data simulation strategy to generate a sampling distribution of null results with which to compare the real results. The idea is to simulate the data under investigation, reproducing important properties of the data set as a whole, but generating data that is known to be dimensional with which to compare the real data under investigation. This idea has been suggested by Tellegen³⁹ and has been used before^19,58. The procedure replicates the scoring pattern for each item while preserving the factor structure of the measure. A concise explanation of this procedure is available⁴³.

Data resampling procedure. The data simulation strategy described above provides a means of addressing the false positive rate of the MAXCOV methodology that is evident from previous research. While the data simulation procedure addresses this problem, it is also known to be overly conservative, and MAXCOV results of known typologies often look no different than simulated controls, except when very large sample sizes are employed. In an effort to address the false-negative rate of the control strategy, we used a bootstrapping procedure¹⁵ that involves resampling the original data (with replacement) to estimate expected results with a larger sample size. The original samples of data were resampled to produce sample sizes of 1000. As resampling involves randomly selecting new cases from the original sample, resampled data sets will vary with respect to case composition, and results of analyses of these data sets are also expected to vary somewhat. We chose to create 100 sets of bootstrapped data with a sample size of 1000. To guard against any unanticipated effect of resampling the real data, we used the same resampling procedure to generate the larger samples of simulated dimensional data. Each set of simulated data generated as controls for the real data was resampled to produce the larger sample sizes to serve as controls for the bootstrapped real data.

Back to the top.