Psychometric Properties of the Gaze Anxiety Rating Scale: Convergent, Discriminant, and Factorial Validity

Abstract

Fear and avoidance of gaze are two features thought to be associated with problematic social anxiety. Avoidance of eye contact has been linked with such undesirable traits as deceptiveness, insincerity, and lower self-esteem. The Gaze Anxiety Rating Scale (GARS) is a self-report measure designed to assess gaze anxiety and avoidance, but its psychometric properties have only been assessed in one preliminary study. We further investigated psychometric properties of the GARS by assessing convergent and factorial validity. We obtained a two-factor solution: gaze anxiety and avoidance across situations (1) in general (GARS-General) and (2) related to dominance communication (GARS-Dominance). The GARS-General factor related more strongly to social anxiety than the GARS-Dominance, and convergent validity of the factors was supported by expected relationships with personality and social anxiety variables. Our results indicate that the GARS subscales are psychometrically valid measures of gaze aversion, supporting their use in future study of the relationship between social anxiety and eye contact behavior.

Introduction

The tendency to avoid eye contact has been linked to problems in social functioning and has been documented in individuals with higher social anxiety (Hietanen, Leppänen Peltola, Linna-Aho, & Ruuhiala, 2008; Horley, Williams, Gonsalvez, & Gordon, 2004). Despite the relevance of assessing gaze behavior for understanding and treating social anxiety disorder (SAD), eye contact remains a difficult behavior to measure through observation. Thus, a psychometrically valid self-report instrument that captures gaze anxiety and avoidance would aide researchers and clinicians in addressing this behavior. The Gaze Anxiety Rating Scale (GARS; Schneier, Rodebaugh, Blanco, Lewin, & Liebowitz, 2011) assesses gaze aversion via self-report and has shown initial psychometric support. A crucial next step is thus to further test its psychometric properties.

The link between forms of psychopathology such as social anxiety and gaze avoidance is an important one to consider because a tendency to avoid eye contact is associated with making unfavorable impressions. For example, those who engage in gaze avoidance have been rated as more anxious, less sincere (Iizuka, 1992), and more deceptive (Bond, Omar, Mahmoud, & Bonser, 1990).¹ Therefore, it is likely that gaze avoidance further exacerbates social difficulties and contributes to less social acceptance. Indeed, in a study on safety behaviors, it was found that dropping safety behaviors such as gaze avoidance was associated with a more favorable impression on an interaction partner (McManus, Sacadura, & Clark, 2008). Gaze avoidance may be a particularly important behavior to target in treatment and the GARS may, therefore, be useful for assessing whether gaze avoidance is a problem and for tracking improvement for both clinical interventions and treatment research.

Schneier et al. conducted the first investigation of the psychometric properties of the measure and found that patients with SAD of the generalized subtype reported higher scores on the GARS than healthy controls. After completion of treatment for SAD with paroxetine, patients’ scores on the GARS decreased. The authors also reported positive correlations with a measure of social anxiety and a measure of submissive behavior, and a lack of relationship with a measure of depression.

The initial findings in support of the psychometric properties of the GARS nevertheless left several issues untested; perhaps most strikingly, the possibility of gender differences. Previous research indicates that the personality characteristics and judgments associated with gaze avoidance tend to differ by gender (Larsen & Shackelford, 1996). Larsen and Shackelford found that, for women, gaze avoidance was associated with higher psychopathology, hysteria, and traditional femininity, but less anxiety and depression than for men. For men, gaze avoidance was associated with higher emotional inhibition, lower self-esteem, higher depression, and higher rates of psychosomatic and physical symptoms. Engaging in gaze avoidance was also associated with being judged negatively for women, whereas, for men, there was no association between gaze avoidance and being judged negatively. Given that women may be more negatively judged for engaging in gaze avoidance, it will be important for clinicians and researchers to consider gender when assessing this behavior.

To expand upon the findings of Schneier et al. (2011), we tested the factorial and convergent validity of the GARS. Based on the evidence outlined earlier that gaze behavior and associated judgments may differ based on gender, we initially explored the factor structure in one gender at a time. We also tested whether the factor structure and convergent relationships varied by gender. We did not have any a priori predictions about the exact factor structure of the GARS, but we expected that the relationship between pairs of items (i.e., fear versus avoidance for each situation) might influence the factor structure; we also expected that the different types of situations would potentially group together by situation similarity. For example, performance situations might load on one factor and interaction situations on a second. In the absence of prior evidence to support particular factors, we first explored the factor structure and then confirmed it in random halves of the sample.

Once a factor structure was confirmed, we tested convergent validity with measures of social anxiety, the big five factors of personality, and the dimensions of dominance and warmth that comprise the interpersonal circumplex (Markey & Markey, 2009). We expected to replicate previously established positive relationships between scores on the GARS and measures related to social anxiety (Schneier et al., 2011). Based on previous research suggesting that individuals with SAD tend to be higher in neuroticism and lower in extraversion, openness, and agreeableness (Bienvenu et al., 2004; Trull & Sher, 1994), we further expected scores on the GARS to show similar relationships with these personality factors. Specifically, we hypothesized that GARS scores would be negatively related to extraversion, openness, and agreeableness, and positively related to neuroticism.

We also tested the relationship between scores on the GARS and the International Personality Item Pool-Interpersonal Circumplex (IPIP-IPC; Markey & Markey, 2009). The interpersonal circumplex measures dominance and warmth, which are proposed to be the primary dimensions related to social behavior (e.g., Leary, 1957). Because of the previously documented relationship between eye contact and dominance communication (Coss, Marks, & Ramakrishnan, 2002; Strongman & Champness, 1968), we predicted that GARS scores would correlate inversely with dominance. In addition, given that people tend to show a preference for direct eye contact in non-threatening encounters (Hietanen et al., 2008; Larsen & Shackelford, 1996), we predicted that GARS scores would also relate negatively to warmth. Finally, we predicted that the relationships between scores on the GARS and the convergent measures would show at least some differences depending on gender.

Methods

Participants

A sample of undergraduates (N = 698) completed a series of self-report measures at Ohio University. Participants received course credit through introductory psychology courses for their participation. They completed the measures on an online survey system that could be accessed at their leisure. Participants were predominately female (n = 449, 64.3%) and white (n = 614, 88%), with a mean age of 19.03 (SD = 1.58). Other ethnicities reported were Black/African-American (n = 28, 4%), Latino (n = 7, 1%), American-Indian or Alaskan Native (n = 4, 0.6%), Asian (n = 19, 2.7%), Native Hawaiian or Pacific Islander (n = 2, 0.3%), and 8 (1.1%) participants indicated their ethnicity was not captured by the above list.

Measures

The GARS (Schneier et al., 2011) is a measure of the amount of anxiety and avoidance one experiences in response to making eye contact across a variety of social situations. There are 17 straightforwardly worded items describing a variety of situations (e.g., giving a speech, speaking to someone you find attractive, and receiving a compliment). Respondents are asked to rate fear and avoidance of making eye contact in each situation over the past week; thus, similar to the Liebowitz Social Anxiety Scale there are two items for each situation. Scores range from 0 (no anxiety) to 3 (a lot of anxiety) for anxiety making eye contact, and from 0 (no avoidance) to 3 (avoid a lot) for avoidance (i.e., not holding eye contact). Total scores on the measure are formed by adding all items. In this study, the avoidance and anxiety subscale scores were found to be highly correlated (r = .90, p < .001). Internal consistency for this study was calculated after the factor analyses were conducted, and is reported in the “Descriptive statistics” section below.

The Social Interaction Anxiety Scale (SIAS; Mattick & Clarke, 1998) is a 20-item measure employing a 0–4 Likert-type scale. The SIAS has shown good psychometric properties, including strong reliability and discriminant and construct validity (Brown, Turovsky, Juster, Brown, & Barlow, 1997; Heimberg, Mueller, Holt, & Hope, 1992). The three reverse-scored items were omitted in the present analyses because they fail to load on the same factor as the other items (Rodebaugh, Woods, Heimberg, Liebowitz, & Schneier, 2006) and may be less related to social anxiety and more related to extraversion, and generally limit the validity of the scale (Rodebaugh, Woods, & Heimberg, 2007). In this study, the straightforward items of the SIAS showed good internal consistency (α = .85).

The Social Phobia Scale (SPS; Mattick & Clarke, 1998) is a 20-item self-report measure designed to measure social performance anxiety. Items are rated on a Likert-type scale ranging from 0 to 4. Researchers have found support for the reliability and validity of the SPS (Gore, Carter, & Parker, 2002; Mattick & Clarke, 1998; Peters, 2000). In this study, the SPS showed excellent internal consistency (α = .93).

The Brief Fear of Negative Evaluation Scale (BFNES; Rodebaugh et al., 2004; Weeks et al., 2005) is a 12-item self-report measure based on the original Fear of Negative Evaluation Scale (Watson & Friend, 1969). The BFNES measures fear and distress related to negative evaluation and has shown very similar psychometric properties to the original scale (Leary, 1983). Weeks et al. (2005) and Rodebaugh et al. (2004) found the straightforwardly worded items of the scale to be more valid than the reverse-scored items in both clinical and undergraduate samples, respectively. Carleton, Collimore, McCabe, and Antony (2011) compared three variants of the BFNES and found support for the use of the 8-item variant with the original straightforwardly worded items. Therefore, only the eight original straightforwardly worded items were used in the current analyses. In this study, the BFNES showed good internal consistency (α = .89).

The Fear of Positive Evaluation Scale (FPES; Weeks, Heimberg, & Rodebaugh, 2008) is a 10-item (although only 8 items are scored; see Weeks, Heimberg, & Rodebaugh, 2008) self-report measure of fear of positive evaluation in a variety of situations. Items are rated on a 0–9 Likert-type scale. Weeks, Heimberg, and Rodebaugh (2008) reported good internal consistency (α = .80), 5-week test–retest reliability (ICC = .70), and convergent and discriminant validity. Weeks, Heimberg, and Rodebaugh (2008) and Fergus et al. (2009) also reported that the FPES is factorially distinct from fear of negative evaluation (a related construct). Fergus et al. (2009) and Weeks, Heimberg, Rodebaugh, Goldin, and Gross (2012) have reported good convergent and discriminant validity of the FPES in independent clinical samples, and Weeks, Heimberg, Rodebaugh, and Norton (2008) found that the FPES predicted anxiety-related responses to receiving positive feedback. In this study, the FPES showed good internal consistency (α = .84).

The Mini-International Personality Item Pool Inventory (Mini-IPIP; Donnellan, Oswald, Baird, & Lucas, 2006) is a 20-item short form measure of extraversion, neuroticism, agreeableness, conscientiousness, and openness. It is based on the International Personality Item Pool (Goldberg, 1999), and has been shown to have consistent and acceptable internal consistencies, similar coverage of personality facets as other Big Five measures, and good convergent, discriminant, and criterion-related validity with other Big Five measures (Donnellan et al., 2006). Baldasaro, Shanahan, and Bauer (2013) report acceptable reliability, good convergent validity, and some support for a five-factor structure for the scales. One of the items from the neuroticism subscale (item 19) was dropped because internal consistency analyses suggested that α would increase from .59 to .63 if this item were excluded. In this study, the five factors showed acceptable to good internal consistency (α’s ranged from .63 to .73).

The IPIP-IPC (Markey & Markey, 2009) is a 32-item measure of the interpersonal circumplex. The interpersonal circumplex is a circular array of personality characteristics that consists of four quadrants formed by the dimensions of dominance and warmth: two traits that are thought to be important for understanding interpersonal interactions (Gurtman, 1993; Leary, 1957). Warmth is characterized by nurturance, communion, love, and affiliation, and dominance is characterized by status and agency (Gurtman, 1993). Markey and Markey (2009) confirmed the circular structure of the IPIP-IPC across three samples. The authors reported that the IPIP-IPC demonstrated convergent validity in its relationships with the five-factor model of personality and that it maintained the psychometric properties of longer interpersonal circumplex measures. In this study, internal consistencies for the dominance and warmth dimensions were good (reliability coefficients > .83; see Nunnally & Bernstein, 1994, p. 269).

Data analytic procedure

Missing data occurred in no more than 3.2% of the cases for any given analysis. For exploratory factor analyses (EFA), we used geomin (oblique) rotation. As per the recommendations of Rhemtulla, Brosseau-Liard, and Savalei (2012), we considered the 0–3 scale of the GARS to be categorical. We therefore used the weighted least squares with robust standard errors and mean- and variance-adjusted chi-square (WLSMV) estimator, implemented in the Mplus program (version 4, Muthén & Muthén, 1998–2009), which is appropriate for categorical data. For all other measures, we analyzed scale totals; they were therefore treated as continuous data. The WLSMV estimator estimates missing data, allowing the use of participants with partially missing data in factor analyses. In determining factor structure, global model fit was evaluated using the following: Tucker–Lewis index (TLI; Tucker & Lewis, 1973), comparative fit index (CFI; Bentler, 1990), and the root mean square error of approximation (RMSEA; Steiger & Lind, 1980). The following values indicate a good fit of the model to the data: TLI and CFI ranging from .95 to 1.0 and RMSEA below .06 (cf. Hu & Bentler, 1999). We began our factor analyses by conducting an EFA in the women’s data. We did so because of previous evidence indicating gender differences in gaze behavior (Exline, Gray, & Schuette, 1965; Larsen & Shackelford, 1996).

Results

Factor analyses

Exploratory and confirmatory factor analyses by gender.

We conducted an EFA in a randomly selected half of the data set (n = 343), further divided into one data set for women (n = 217) and one data set for men (n = 126). Given that more participants were women, we first determined the best factor structure for women and then tested whether this structure could hold for men. We conducted a parallel analysis (Hayton, Allen, & Scarpello, 2004) on the women’s data set using 50 randomly generated data sets that indicated a 95% confidence interval (CI) of 1.97–2.24 (M = 2.09) for the eigenvalue of the first random factor, a CI of 1.85–2.03 (M = 1.94) for the eigenvalue of the second random factor, and a CI of 1.72–1.94 (M = 1.82) for the eigenvalue of the third random factor. We planned to only retain factors that had eigenvalues greater than the upper limit of these CIs.

For the women’s data EFA (n = 217) a one- or two-factor solution was indicated. We conducted a confirmatory factor analysis (CFA) in the same data set for the one- and two-factor models. The one-factor (CFI = .90, TLI = .90, RMSEA = .11 [90% CI = .10–.11]) model had borderline fit, and the two-factor model (CFI = .94, TLI = .93, RMSEA = .09 [90% CI = .08–.09]) had adequate fit.

In an attempt to improve on the borderline to adequate fit, we reasoned that the above-tested models may not have fit well because they failed to account for the anxiety and avoidance items for each situation depending on the same item stems describing the relevant social situation. Thus, we next tested a model in which all items loaded on two factors and the error variances of each pair of anxiety and avoidance items was correlated. This model showed good fit (CFI = .97, TLI = .97, RMSEA = .06 [90% CI = .05–.06]) and had significantly better fit (χ²[9, n = 217] = 195.79, p < .001) than a one-factor model that also included error covariances between anxiety/avoidance item pairs. Having arrived at a model that fit well for women, we tested this model in the first random male sample. The two-factor model with error covariances also fit well for men (CFI = .98, TLI = .98, RMSEA = .05 [90% CI = .04–.06]).

We next tested whether loadings and thresholds could be constrained across gender in a multiple-group CFA utilizing the two-factor structure with error covariances. The constrained model fit well (CFI = .98, TLI = .98, RMSEA = .05 [90% CI = .04–.06]). The unconstrained model fit about as well (CFI = .98, TLI = .98, RMSEA = .05 [90% CI = .05–.06]), but was significantly better fitting than the constrained model in terms of absolute fit (χ²[103, n = 344] = 128.72, p = .044). Based on previous research suggesting that a change in CFIs smaller than .01 between models indicates that measurement is effectively invariant (Cheung & Rensvold, 2002), we determined that the unconstrained model was not meaningfully better fitting than the constrained model. Therefore, we concluded that factor loadings and item thresholds could be constrained across gender. However, because we conducted all of the above tests in the same half of the data set, cross-validation in the second random half of data was essential.

The CFA in the other half of the data set produced good fit (CFI = .97, TLI = .96, RMSEA = .06 [90% CI = .05–.06]). We also conducted the test of whether loadings and thresholds could be constrained by gender in this half of the data set. Results were equivalent. We then tested our model in the full data set to obtain the most precise estimates; this model also fit well (CFI = .97, TLI = .97, RMSEA = .06 [90% CI = .06–.06]). The two factors were significantly and positively correlated (r = .78, p < .001). As seen in Table 1, the factors appear to capture anxiety and avoidance of eye contact across: (a) situations in general (GARS-General) and (b) performance-based situations such as giving a speech or discussing work with a boss (GARS-Dominance). Notably, these performance-based situations also involved more potential conflict involving dominance than other situations listed on the measure. Supporting this interpretation are the inverse loadings for having a routine talk with a close family member, a situation not expected to be associated with dominance issues. Therefore, we hypothesized that the second factor (GARS-Dominance) would show stronger positive relationships with the SPS (which better captures social performance anxiety than our other measures) and the dominance dimension of the IPIP-IPC.

Table 1.

Factor loadings for the Gaze Anxiety Rating Scale General factor and Dominance factor

Item	General factor	Dominance factor
Giving a speech
1. Anxiety		0.65
2. Avoidance		0.66
Speaking to a group of people at a party
3. Anxiety	0.77
4. Avoidance	0.81
Speaking up at a meeting
5. Anxiety		0.76
6. Avoidance		0.89
Speaking in a discussion with a few people
7. Anxiety	0.79
8. Avoidance	0.83
Dealing with a cashier when buying something
9. Anxiety	0.83
10. Avoidance	0.81
Being introduced
11. Anxiety	0.76
12. Avoidance	0.80
Greeting an acquaintance passing by on the street
13. Anxiety	0.79
14. Avoidance	0.73
Speaking with someone you do not know well
15. Anxiety	0.49	0.36
16. Avoidance	0.59	0.21
Speaking to someone you find attractive
17. Anxiety	0.15	0.63
18. Avoidance	0.50	0.31
Inviting someone you do not know well on a date
19. Anxiety		0.72
20. Avoidance	0.44	0.34
Feeling close to someone you love
21. Anxiety	0.74
22. Avoidance	0.78
Discussing the quality of your work with a boss or a teacher
23. Anxiety	0.34	0.42
24. Avoidance	0.75
Having a routine talk with a close family member
25. Anxiety	1.22	−0.53
26. Avoidance	1.24	−0.51
Listening while a person speaks to you, in general
27. Anxiety	0.82
28. Avoidance	0.84
Speaking while a person listens to you, in general
29. Anxiety	0.84
30. Avoidance	0.80
Expressing a disagreement
31. Anxiety	0.69
32. Avoidance	0.75
Receiving a compliment
33. Anxiety	0.71
34. Avoidance	0.75

Note: Standardized factor loading estimates for the factors in CFA.

Factor relationships

To test how the factors related to the convergent measures, we included these measures in our factor analysis model. To determine the relative strength of correlations, we compared models in which the compared relationships were constrained to a model in which they were unconstrained. If the model that is constrained is significantly worse fitting, this means that the relationships are significantly different. Specifically, we included correlations between both GARS factors and the SIAS, the SPS, the FPES, the BFNES, the big five personality factors, and the dominance and warmth dimensions, as well as intercorrelations among the convergent measures. Full results are given in Table 2. Against hypothesis, the SPS showed a significantly stronger relationship with the GARS-General than with the GARS-Dominance factor, and constraining these correlations to be equal resulted in significantly worse fit (χ²[1, n = 696] = 46.62, p < .001). In contrast, as predicted, the GARS-Dominance factor showed a stronger relationship with the IPIP-IPC dominance dimension than did the GARS-General factor, and constraining these correlations to be equal resulted in significantly worse fit (χ²[1, n = 696] = 25.98, p < .001). In addition, the GARS-General factor related significantly to agreeableness and conscientiousness, whereas the GARS-Dominance factor did not.

Table 2.

Factor relationships with all study measures (n = 696)

	GARS-General r	GARS-Dominance r
SIAS	0.68**	0.62**
SPS	0.64**	0.51**
FPES	0.50**	0.43**
BFNES	0.42**	0.47**
NEUR	0.29**	0.24**
EXTRA	−0.40**	−0.39**
CONSC	−0.18**	−0.08
OPEN	−0.25**	−0.10*
AGREE	−0.33**	−0.04
DOM	−0.15**	−0.36**
WARMTH	−0.48**	−0.19**

Notes: SIAS, social interaction anxiety scale; SPS, social phobia scale; FPES, fear of positive evaluation scale; BFNES, brief fear of negative evaluation scale; NEUR, neuroticism; EXTRA, extraversion; CONSC, conscientiousness; OPEN, openness; AGREE, agreeableness; DOM, dominance dimension; WARMTH, warmth dimension.

^*p < 0.01 and

^**p < 0.001.

Factor relationships by gender

We tested whether the relationships between the factor scores and convergent measures were equal across gender. Due to limitations in the statistical program we were unable to constrain correlations between the factors and the convergent measures across gender when using the WLSMV estimator. We therefore computed factor scores and constrained the correlations between the factor scores and the convergent measures across gender. We ran separate models to obtain factor scores for each factor. For the GARS-General factor score, the model with constrained convergent relationships had only reasonably good relative fit (CFI = 1.00, TLI = .94, RMSEA = .07 [90% CI = .04–.10]) and fit less well than an unconstrained model (χ²[11, n = 684] = 28.79, p = .002). For the GARS-Dominance factor score, the constrained model fit well (CFI = 1.00, TLI = .96, RMSEA = .06 [90% CI = .02–.09]), but had significantly worse fit than the unconstrained model (χ²[11, n = 684] = 23.07, p = .017). Therefore, overall correlations (between factor scores and convergent measures) were significantly different across gender for each factor score.

We examined the correlations in the unconstrained models to see whether specific correlations showed large differences by gender. We tested whether unconstraining the correlations that showed the greatest differences in magnitude by gender would be equivalent to unconstraining all correlations. For the GARS-General factor score, the model with the openness, neuroticism, and conscientiousness correlations unconstrained fit well (CFI = 1.00, TLI = .98, RMSEA = .04 [90% CI = .00–.08]) and was not significantly different from the fully unconstrained model (χ²[8, n = 684] = 12.18, p = .14). For the GARS-Dominance factor score, a model with only the conscientiousness correlation unconstrained fit well (CFI = 1.00, TLI = .97, RMSEA = .05 [90% CI = .00–.08]) and was not significantly different from the fully unconstrained model (χ²[10, n = 684] = 16.86, p = .078). Therefore, we concluded that most of the relationships were invariant across gender, with the exception of openness, conscientiousness, and neuroticism. As seen in Table 3, these variables were more strongly related to the respective factors for men than for women.

Table 3.

Factor score relationships that showed the greatest differences by gender (n = 684)

	Men r (n = 240)	Women r (n = 444)
GARS-General with
NEUR	0.39**	0.21**
CONSC	−0.25**	−0.08
OPEN	−0.32**	−0.12*
GARS-Dominance with
CONSC	−0.22**	−0.01

Notes: NEUR, neuroticism; CONSC, conscientiousness; OPEN, openness.

^*p < 0.01 and

^**p < 0.001.

Descriptive statistics

We calculated subscale totals based on the factor loadings. Items that showed negative loadings on the factors were reverse scored. Notably, these items are not reverse worded; all items of the GARS are straightforwardly worded. Rather, these were straightforwardly worded items that showed negative loadings. When subscales were calculated based on adding items with reverse scoring as needed based on factor loadings, the internal consistencies for the subscales were good (both α’s > .90). Scores on each of the GARS subscales (formed from the two-factor structure) did not differ significantly by gender (p’s > .078).²

Discussion

Our results provide replication and extension of Schneier et al. (2011). Our results indicate that the GARS is a psychometrically valid measure of anxiety and avoidance of eye contact with two underlying factors. The two factors appear to capture gaze anxiety and avoidance (a) across a variety of social situations (GARS-General) and (b) in situations related to dominance (GARS-Dominance). We found that the factor loadings and thresholds could be constrained across gender. Convergent validity was supported by expected relationships with measures related to social anxiety, though there were some differences in how the factors related to some of the measures. Finally, there was some evidence for gender differences in the strength of the relationships between the factors and the convergent measures.

Although the two factors showed similar relationships to the convergent measures, a one-factor solution did not show good fit in factor analyses, indicating that the scale cannot be considered to be unifactorial. These findings suggest that the GARS subscales have validity superior to that of the total score. In support of this conceptualization, the GARS-General factor tended to show stronger relationships with the convergent (i.e., social anxiety-related) measures, suggesting that it may be more strongly related to social anxiety. The GARS-Dominance factor showed similar, but less strong, relationships, as well as some differential relationships. Specifically, the GARS-Dominance factor showed a stronger relationship with a measure of dominant interpersonal behavior than did the GARS-General factor. However, against hypothesis, the GARS-Dominance factor did not show a stronger relationship with the SPS than did the GARS-General factor. It should be noted that the SPS does not measure dominance behavior per se, perhaps explaining this finding.

As predicted, and supported in previous research (Exline et al., 1965), there was some indication that gender differences should be considered in relation to gaze aversion. We found that the correlations between the GARS-General and GARS-Dominance factor scores and conscientiousness as well as the relationships between the GARS-General factor score and openness and neuroticism could not be constrained across gender. In line with previous research showing that gaze avoidance is more strongly tied to self-report of emotional difficulties for men (Larsen & Shackelford, 1996), our results suggest that self-report of gaze anxiety and avoidance is more strongly related to self-reported undesirable personality traits such as lower conscientiousness and higher neuroticism for men than women. We propose that gaze aversion may be more socially normative for women than men as suggested by the relationship between higher femininity and gaze avoidance in Larsen and Shackelford. It would therefore follow that gaze avoidance would be more strongly tied to less desirable personality characteristics for men than women. If this finding is replicated, targeting gaze avoidance in treatment may be particularly important for male clients.

The results of this study should be considered in light of its limitations. First, all of our measures were self-report; the convergent and discriminant validity of the GARS should be investigated utilizing other sources of information, including behavioral observations and eye tracking. Second, the participants in our sample were primarily Caucasian women and were all undergraduates, which limits the generalizability of our findings. Tests of replication in samples that are diverse with respect to ethnicity, age, socioeconomic status, and form and severity of psychopathology will be crucial for establishing recommendations for the use of the GARS. Third, we conducted many correlational tests; the relationships that were hypothesized a priori deserve greater emphasis. Finally, the GARS contains fewer response options than would be ideal; future researchers might improve the scale by offering a wider range of response options (e.g., 5 or 7 instead of 4).

We recommend that researchers utilizing undergraduate samples consider using GARS subscales based on our factor structure. However, replicating our findings in different samples, particularly non-undergraduate and clinical samples, is necessary before making a broad recommendation for these subscales. Based on previous research indicating cultural differences in eye contact conventions (McCarthy, Lee, Itakura, & Muir, 2006), future researchers should also include tests of how culture influences eye contact behavior as well as responses on the GARS. Given the high degree of overlap between social anxiety symptoms and gaze anxiety and avoidance and the documented negative effects of this behavior (Bond et al., 1990; Iizuka, 1992), gaze behavior is an important behavior to consider for clinical intervention.

The GARS factors and subscales demonstrated convergent and discriminant validity through relationships with social anxiety-related measures, personality variables, and interpersonal circumplex dimensions. Overall, our results indicate that the GARS is a psychometrically valid measure that captures two constructs of gaze aversion tendencies: (a) a construct related to gaze aversion across a variety of situations and (b) a construct related to gaze aversion in situations that tend to elicit dominance behaviors.