A sample size of 300 was targeted for both EFA and CFA, which prior simulations with more complex five-factor structures have shown to be sufficient, and was therefore deemed more than adequate for the present three-factor (or simpler) model (Mundfrom et al., 2005; Myers et al., 2011; Reio and Shuck, 2014). Participants were recruited by a group of trained postgraduate students and were instructed to fill out online questionnaires by computer or mobile phone. No additional inclusion/exclusion criteria were applied beyond being a Chinese college student and providing informed consent. Participants received 10 Chinese Yuan (CNY) as compensation for their time. Participation was voluntary and they could withdraw at any time without negative consequences. Group 1 collected in Wave 1 comprised 350 Chinese college students (35.71% males, M_age = 20.13, SD = 1.34) and was used for the EFA. For the subsequent CFA, 309 Chinese college students (Group 2, 44% males, M_age = 20.87, SD = 1.78) were recruited. Study 1 and 2 were approved by the Ethics Committee of the College of Psychology, University (approval number: PN-202200049).

Contrast Avoidance Questionnaires (CAQs). The original 25 item CAQ-GE, and 30 item CAQ-W developed by Llera and Newman (2017) were translated into Chinese by a researcher and masters-degree student in Psychology with fluent English skills. All items were back-translated into English by a bilingual speaker who was blinded to the original version. Inconsistencies between the back-translated items and the original ones were fully discussed by the translators and the developers of the original CAQs until a consensus was reached. Prior to the large-scale administration, a pilot test was conducted to examine participants’ comprehension of the translated items, and minor modifications were made to improve clarity. As with the original CAQs, the Chinese version entailed two subscales: CAQ-Worry (CAQ-W; 30-items), and CAQ-General Emotion (CAQ-GE; 25-items). These were rated on a 5-point Likert scale ranging from 1 (not at all true) to 5 (absolutely true). Sum scores were created for the respective subscales, with higher scores indicating higher levels of contrast avoidance.

All analyses and data processing in this study were performed using R (version 4.4.1; R Core Team, 2024), with analyses implemented through relevant R packages or other statistical software as needed. The EFA was run separately on each CAQ measure by Psych package (Revelle, 2025). We applied oblique rotation (Oblimin) based on the assumption of factor intercorrelations, used a polychoric correlation matrix to treat the 5-point Likert items as ordinal variables, and employed the minimum residual (MINRES) extraction because it does not require the assumption of multivariate normality. A parallel analysis (Hayton et al., 2004; Horn, 1965), Velicer’s MAP (Velicer, 1976), the Hull method (Lorenzo-Seva et al., 2011) scree test, and eigenvalues were used to determine the accurate number of factors to be extracted. After factors were identified, we retained items that had loadings above 0.40 onto one component while with a loading difference of at least 0.20 on all other components (Matsunaga, 2010). Finally, to verify the factor structure, we conducted CFA in Mplus 8.3 via MplusAutomation (Hallquist and Wiley, 2018), using the robust diagonally weighted least squares estimator (WLSMV), which treats items as categorical. We further ran a CFA using the maximum likelihood estimator with the robust standard errors estimator (MLR) as a sensitivity analysis, in which items were treated as continuous. Post hoc modification indices (MIs) were not applied to refine the CFA model.

For both measures, most Kaiser-Meyer-Olkin (KMO) (Shrestha, 2021) statistics of sample adequacy were above 0.9, except for items 28 and 29 from the W subscale, and items 1, 14, 19 and 22 from the GE subscale, which ranged from 0.82 to 0.88; Bartlett’s test of sphericity was significant (CAQ-W: χ² [435] = 8842.59, p < 0.001; CAQ-GE: χ² [300] = 6996.89, p < 0.001), indicating that the sample was suitable for factor analysis.

CAQ-W. The initial EFA revealed three factors with eigenvalues > 1.0; results of parallel analyses, Velicer’s MAP and Hull method (based on CAF, CFI and RMSEA), and scree plot (see Supplementary Figure A1) consistently supported a 3-factor model, which was consistent with the factor structure in the original study by Llera and Newman (2017). Hence, we re-ran the EFA constrained to three factors, and dropped items that did not meet screening rules. Specifically, Item 26 (“When I worry I can hold my emotions in a steady state”) had a low factor loading (< 0.40), and Items 22 (“I would rather worry and be surprised if something good happens, than feel good and be distressed if something bad happens”) and 11 (“If I worry, I feel so much better when things turn out okay than if I had not been worried”) cross-loaded on F1 and F3. Thus, these three items were deleted. Item 25 (“I find worrying most rewarding when something good happens in the end”) was removed in the second EFA, due to cross-loadings with F1 and F3. Additionally, Item 30 (“It is better to have worried first and then get the best outcome than to expect the best all along”), which was originally developed for F3, consistently loaded more strongly on F1 than on F3 across all EFAs, including the EFA conducted after the aforementioned items were removed. Given that F1 already contained 13 items, Item 30 was removed. All remaining items met screening rules in the final EFA. The final CAQ-W retained 25 items loading onto three factors. Factor loadings and communalities of each item are presented in Table 1. F1 (Worry to Avoid Negative Emotional Shifts; 13 items) showed a moderately positive correlation with F2 (Worry Creates and Sustains Negative Emotion; 9 items; r = 0.27) and F3 (Worry to Create Positive Contrast; 3 items; r = 0.36). F2 was moderately positively correlated with F3 (r = 0.40). No correlations exceeded r = 0.70, suggesting no problems with multicollinearity (Kline, 2005).

CAQ-GE. The initial EFA demonstrated three factors with eigenvalues > 1.0, Velicer’s MAP also suggested three factors, whereas parallel analysis suggested four factors, and results from both the Hull method (based on CAF, CFI and RMSEA) and scree plot (see Supplementary Figure A2) indicated two factors. We therefore re-ran the EFA constrained to two, three, and four factors. The additional factors beyond the two-factor solution were not supported, as they did not yield sufficient numbers of items meeting the loading criteria. Specifically, in the three-factor solution, Items 2, 4, and 13 exhibited cross-loadings on Factors 1 and 3, leaving only one item (Item 1) with a salient loading on the third factor. In the four-factor solution, Items 4, 8, and 13 showed cross-loadings on Factors 1 and 3, leaving only two items (Items 1 and 2) on the third factor, and no items demonstrated adequate loadings on the fourth factor. Accordingly, the two-factor solution was retained, as all items had distinctively high loadings on their respective factors, with one exception. Item 13 (“When I have already been in a bad mood, it has been easier to endure bad news”) was deleted for low factor loading (< 0.40). The second EFA demonstrated that all remaining items met screening rules, which indicated a stable factor structure. The final CAQ-GE retained 24 items loading onto two factors. Factor loadings and the communalities of each item are presented in Table 2. F1 (Creating and Sustaining Negative Emotion to Avoid Negative Contrasts; 17 items) was moderately positively correlated with F2 (Discomfort with Emotional Shifts; 7 items; r = 0.543).

Details including the KMOs and 95% bootstrap confidence intervals of the factor loadings for each item under the aforementioned factor solutions are presented in Supplementary Tables A1–A8.

CAQ-W. To verify the factor structure for the CAQ-W, we compared the 3-factor 25 item model, as determined by EFA, with rival 1- and 2-factor models (i.e., combining all factors for the 1-factor model, and combining F1 and F3 to create a single factor reflecting worry to manage contrast for the 2-factor model). In addition, the original 3-factor model per Llera and Newman (2017), in which none of the original items were deleted, was also included for comparison. The CFA generally supported a 3-factor model with 25 items that fit better than a 1-factor model and slightly better than 2-factor model, and was comparable with the original 3-factor model (see Table 3). Overall, most indices for this model indicated an acceptable model fit (χ² = 863.32, df = 272, χ²/df = 3.17, CFI = 0.95, TLI = 0.95, SRMR = 0.07, RMSEA = 0.08 [90% CI = 0.08–0.09]). For this model, standardized factor loadings ranged from 0.69 to 0.86 (all ps < 0.001). Thus, the 3-factor 25 item model was retained for subsequent analyses in Study 2.

CAQ-GE. The CFA indicated the 2-factor model with 24 items fit better than a 1-factor model, and comparable with the 2-factor without deleting any items (Table 3). Key fit indices for this model were acceptable (χ² = 729.23, df = 251, χ²/df = 2.91, CFI = 0.97, TLI = 0.97, SRMR = 0.05, RMSEA = 0.08 [90% CI = 0.07–0.09]). Standardized factor loadings ranged from 0.54 to 0.87 (all ps < 0.001) for the 2-factor model with 24 items. Thus, this model was retained for subsequent analyses in Study 2.

Sensitivity analysis using an MLR estimator consistently suggested a 3-factor model with 25 items for the W scale, and 2-factor model with 24 items for the GE scale (see Supplementary Table C1), which demonstrated acceptable model fits and sufficient factor loadings. The parameters for CFA are presented in the Supplementary Material B (for WLSMV) and Supplementary Material C (for MLR). Details about the final version of the two scales (instructions, response format, and scoring) are presented in Supplementary Material D.

For Study 2, 388 Chinese college students (Group 3, 40.21% males, M_age = 20.11, SD = 1.11) were recruited in Wave 3, which was used to examine reliability and construct validity. A subset of participants (N = 134, 38.1% male, M_age = 20.104, SD = 1.020) in Group 3 completed the CAQ questionnaires a week later, to estimate retest reliability. The recruitment procedures and ethical safeguards were identical to those used in Study 1.

The modified Chinese-language CAQ measures (i.e., 25-item CAQ-W; 24-item CAQ-GE), as developed in Study 1, were used in this study.

Generalized Anxiety Disorder (GAD-7). The GAD-7 (Spitzer et al., 2006) with 7 items assessed participants’ anxiety symptoms over the past 2 weeks. He et al. (2010) translated the GAD-7 into Chinese, and found that a cut-off score of 10 could identify Chinese adults with GAD with good sensitivity (86.2%) and specificity (95.5%). Prior literature has indicated the Chinese version had fair to good reliability and validity (e.g., Zhang et al., 2021). This scale is scored on a 4-point Likert scale ranging from 0 (never) to 3 (almost every day). A total score was created by summing all items. Internal consistency for the total score in the current sample was excellent (Cronbach’s α = 0.911).

Penn State Worry Questionnaire (PSWQ). Trait worry was assessed with the Chinese version (Sha et al., 2006) of the PSWQ (Meyer et al., 1990), which had fair to good reliability and validity in the Chinese samples (Zhong et al., 2009). This scale consists of 16 items classified by two subscales: Engagement of worry and absence of worry, scored on a 5-point Likert scale ranges from 1 (not at all typical of me) to 5 (very typical of me). Items were averaged to create a total score, with higher scores indicating higher levels of trait worry. In the current sample, Cronbach’s α was 0.885 for the total scale (0.64 for worry absence, 0.93 for worry engagement, respectively).

Perceptions of Threats Questionnaire (PTEQ). The PTEQ is a 72 items scale developed by McCubbin and Sampson (2006) and was included to assess participants’ beliefs about the threat posed by each of seven basic emotions (i.e., sadness, happiness, anger, fear, disgust, guilt, and lust) and strong emotions in general. The original scale was translated into Chinese by the authors. The current study focused on six of the basic emotions (excluding ‘lust’) and strong emotion, which were rated on a 5-point scale ranging from 1 (not at all) to 5 (definitely). Cronbach’s α was 0.816 for sadness, 0.68 for happiness, 0.82 for anger, fear, and disgust, 0.81 for guilt, and 0.84 for strong emotion. A CFA also supported that each subscale fit well in terms of CFI (0.89–0.91) and TLI (0.82 ~ 0.88), except for the Happiness subscale (CFI = 0.80 and TLI = 0.73).

Sensation Seeking Scale (SSS). The trait of seeking intensity and novelty in sensory experiences was assessed with the SSS for Chinese undergraduates (Zhao, 2004). The SSS includes 36 items with 18 items for each subscale (i.e., Thrill and Adventure Seeking, and Disinhibition). It is scored on a Likert scale ranging from 1 (Do not want to do) to 3 (Want to do, do it if you have the chance). The Chinese version of the SSS had adequate reliability and validity (Zhao, 2004). An average score was derived for sensation seeking, with higher values indicating higher levels of the trait of seeking intensity and novelty. For the current study, the measure demonstrated excellent internal consistency reliability (Cronbach’s α = 0.93, 0.93, and 0.88 for the total scale, Thrill and Adventure Seeking, and Disinhibition subscales, respectively).

Firstly, Cronbach α was estimated to evaluate the internal consistency reliability for total scales and each subscale of the two CAQ measures with a threshold of 0.70, 0.80, and 0.90 indicating adequate, good, and excellent relatability, respectively (Youngstrom et al., 2017). Using the semTools package (Jorgensen et al., 2025), Omega total (ω) was estimated for total and subscale scores, and hierarchical omega (ω_h-cat) was estimated for the total scales based on a bifactor model (Flora, 2020) if the model resulted in adequate fit (e.g., CFI > 0.95, RMSEA < 0.08). The average inter-item correlation (AIIC) was evaluated, and a range of 0.15–0.50 was recommended as it indicates a balance between desirable commonality and the avoidance of redundancies between items; item-total correlations were also examined for each item, with values below 0.30 considered unacceptable (Stefana et al., 2024). Then, bivariable correlations and intraclass correlation coefficients (ICC [2,1]) were used to evaluate retest reliability, with values 0.40–0.59, 0.60–0.74, and ≥0.75, respectively, considered fair, good and excellent for correlation (Cicchetti, 1994), and values < 0.50, 0.50–0.75, 0.75–0.90, and > 0.90, respectively, indicated poor, moderate, good, and excellent for ICC (Koo and Li, 2016). Secondly, to estimate construct validity we ran correlation analyses between CAQ scales and all convergent and discriminant measures, and used Steiger’s Z tests (Hoerger, 2013; Steiger, 1980) to determine distinctions between convergent and discriminant correlations. Thirdly, we created High and Low-symptom subgroups from within our sample (see below) and used independent-sample t-tests to estimate whether the CAQ measure total score and individual subscale scores in the High-symptom group were significantly higher than the Low-symptom group. Receiver operating characteristic (ROC) analyses were then conducted with the pROC package (Robin et al., 2011) to evaluate the ability of each CAQ measure to predict GAD symptomatology accurately. The area under the ROC curve (AUC) was calculated as an index of overall discrimination ability, with confidence intervals obtained using DeLong et al.’s (1988) nonparametric method. To explore the optimal cut point, we used the Youden index (J), which identifies the threshold that maximizes the combined sensitivity and specificity (Youden, 1950).

We created High- and Low-symptom subgroups according to recommendations by He et al. (2010) using the GAD-7, as well as scores on the PSWQ. Individuals with GAD-7 total scores ≥ 10 and PSWQ total scores greater than one standard deviation above the overall sample mean (M = 50.38, SD = 10.90) were included in the High-symptom group (N = 37; GAD-7: M = 14.649, SD = 3.327; PSWQ: M = 69.324, SD = 4.773). The PSWQ total score for this group was comparable to those with clinically diagnosed GAD found in other samples (M = 67.16, SD = 9.16; Z = 2.474, p < 0.01; Startup and Erickson, 2006). Participants were included in the Low-symptom group (N = 61) if they scored below 10 on the GAD-7 (M = 2.328, SD = 2.343), and one standard deviation below the overall sample mean on the PSWQ (Low-symptom group: M = 34.639, SD = 3.578). The PSWQ total score in the Low-symptom group was significantly lower than the score for those without clinically diagnosed GAD (M = 37.5, SD = 9.16, Z = −6.239, p < 0.001; Startup and Erickson, 2006). To balance the sample size between the two groups for t test, 37 Low-symptom participants were selected randomly among the Low-symptom group (N = 61) as a comparison sample. The demographic variables (i.e., sex, age, education, major, and ethnicity) did not differ significantly between the two groups (t (72) = 0.93, p = 0.35 for age; χ²(1) = 2.38, p = 0.123 for education; χ²(2) = 1.61, p = 0.45 for major; χ²(1) = 0.14, p = 0.71 for ethnicity and χ²(1) = 2.69, p = 0.10) for sex, which suggests lack of significant difference between the two groups.

The CAQ-W demonstrated excellent internal consistency reliability for the total scale and F1 and F2 subscales (Cronbach’s α = 0.93, 0.93 and 0.94, Omega total ωs = 0.96, 0.93 and 0.94, respectively), and good for F3 (α = 0.81, ω = 0.82). As the bifactor model with three factors did not converge, ω_h-cat was unavailable for the W subscale. The AIICs were 0.59, 0.68 and 0.66, respectively for F1, F2, and F3. Item-total correlations ranged from 0.54 to 0.82 across the three subscales. For the CAQ-GE, the internal consistency reliability was also excellent for the total scale and F1 (αs = 0.93 for both, ωs = 0.95 and 0.94, respectively), and good for F2 (α = 0.88, ω = 0.88). ω_h-cat were 0.14, 0.78 and 0.87, respectively, for the general factor, F1 and F2, which indicated the variance was largely attributable to the specific factors rather than the general factor. The AIICs were 0.55 and 0.56, respectively for F1 and 2. Item-total correlations ranged from 0.44 to 0.80 across the two subscales (see Supplementary Tables E1, E2 for more details).

Correlation coefficients indicated generally good to excellent retest reliability on the CAQ-W for the total score, as well as F1 and F2 subscores (rs = 0.76, 0.77, 0.74, respectively, ps < 0.001); while ICCs indicated moderate to good retest reliability (ICCs = 0.76, 95% CI [0.68, 0.82], 0.77 [0.69, 0.83], 0.74 [0.65, 0.81], respectively). However, the CAQ-W F3 subscale showed fair (r = 0.54, p < 0.001) to moderate (ICC = 0.51 [0.35, 0.64]) retest reliability. The retest reliability of the CAQ-GE total score and its subscales was excellent or good (rs = 0.85, 0.86, and 0.76, ps < 0.001; ICCs = 0.85 [0.79, 0.89], 0.86 [0.80, 0.90] and 0.76 [0.68, 0.83]). Overall, the results suggested both CAQ measures and their subscales were repeatable within a sample over time, except for weaker stability for CAQ-W F3.

Table 4 shows that, within both CAQ-W and CAQ-GE scales, all the subscales were moderately correlated with one another (range: 0.28 to 0.48, ps < 0.001), and strongly associated with the total scale (range: 0.70 to 0.95, ps < 0.001). The CAQ scales and subscales mostly showed moderate correlations with convergent construct measures (GAD-7, PSWQ, and the PTEQ negative and strong emotion subscales), with some weaker (e.g., r = 0.26 between W-F1 and PTEQ-fear) or stronger (e.g., r = 0.61 between W-F2 and PSWQ) associations observed. Additionally, these scales overall demonstrated lower correlations with divergent construct measures (SSS and PTEQ-Happy).

Table 5 presents Steiger’s Z comparisons between correlations with convergent and discriminant measures. In these tests, both CAQ measures and their subscales were overall significantly more positively correlated with convergent measures than with measures of sensation seeking and perceived threat of happiness (Z = 2.17~8.97). However, there were some exceptions. For instance, the CAQ-W F1 and CAQ-GE F1 subscales did not demonstrate significantly higher correlations with most convergent measures as compared to the PTEQ-Happy subscale (Z = −2.1~1.01). They also did not demonstrate higher correlations with PTEQ-Fear versus sensation seeking (Z = 1.74 and 1.27, respectively). Finally, the CAQ-GE Total did not correlate more highly with PTEQ-Fear versus PTEQ-Happy (Z = 1.08).

The High-symptom group scored significantly higher than the Low-symptom group on both CAQ measures and their subscales, with moderate to strong effect sizes for all comparisons (Cohen’s d = 1.04–2.07; see Supplementary Table E3). ROC curve analyses indicated that CAQ measures were strong predictors of GAD symptomatology (see Supplementary Figures E1, E2), with AUC of 0.91 (95% CI [0.86, 0.97]) for CAQ-W and 0.88 (95% CI [0.81, 0.95]) for CAQ-GE. This suggested the probability of a person with high GAD symptoms scoring higher than a person with low GAD symptoms on the CAQ-W and CAQ-GE was 91 and 88%, respectively. For CAQ-W, Youden index (J) indicated the cut-off point ≥2.50 gave the best combination of sensitivity (92%) and specificity (82%) for predicting GAD symptomatology, with 86% of participants correctly classified. For CAQ-GE, the cut score 2.31 provided the best combined results of sensitivity and specificity (78 and 84%) for accurately predicting GAD symptomatology, with 82% of participants correctly classified (see Supplementary Table E4 for details).