Psychometric Evaluation and Item Optimization of the Resilience Scale CD-RISC-10 with Chinese Airline Pilots

Abstract

Brief resilience measures may facilitate group-level aviation research and training feedback, but evidence on the CD-RISC-10 and its shortening in airline pilots remains limited. We evaluated the scale and its shortening potential in a development sample of 192 Chinese male airline pilots; 106 had paired occupational burnout data. Classical test theory, confirmatory factor analysis (CFA), generalizability theory, item response theory, and qualitative item-content review were integrated. The CD-RISC-10 showed high internal consistency (α = 0.895) and expected negative associations with occupational burnout. CFA and graded response model results were consistent with an approximate one-factor representation, despite global-fit uncertainty and localized model strain. This integrated process yielded a candidate six-item form, the CD-RISC-6. In the development sample, it reduced item count by 40%, retained 70.38% of the original test information, and showed high post-selection internal consistency (α = 0.866). In a separate, non-overlapping same-airline sample (N = 58), the fixed form showed high internal consistency (α = 0.891) and expected negative associations with occupational burnout. CFA results were also consistent with the prespecified one-factor structure. These findings provide preliminary within-context evidence for further evaluation of the CD-RISC-6 as a low-burden measure in comparable recurrent-training settings.

1. Introduction

Pilot psychological resilience refers to pilots’ capacity to regulate cognition, emotion and behaviour, and to maintain or restore adaptive functioning when facing challenges, stressors or unexpected events in flight operations and occupational contexts [1,2,3,4]. Increasing automation has improved the precision and efficiency of routine flight tasks and reduced repetitive workload. At the same time, it has shifted pilots’ work toward monitoring automated systems, understanding system modes, maintaining situation awareness, detecting abnormal states and intervening or taking over when necessary [5,6,7]. These demands are particularly salient during mode confusion, automation surprises, system degradation and takeover situations. In such contexts, pilots must judge system status with incomplete information, interpret the source of anomalies and adjust action strategies under uncertainty, which may increase cognitive stress and operational risk [8,9,10]. Adaptive performance in complex automated operations therefore depends not only on technical skill, but also on the ability to maintain cognitive control under stress, regulate emotional responses and translate situational assessment into appropriate action. This set of capacities aligns with the theoretical content of psychological resilience, which emphasizes stress adaptation and the maintenance or recovery of functioning. Psychological resilience may therefore help explain how pilots maintain stable psychological functioning and reliable task performance in complex operational contexts [11,12]. Previous studies have also shown that pilot resilience is closely associated with psychological status, fatigue, work performance and operational performance in abnormal flight situations [1,3,13]. Accordingly, valid and reliable assessment of pilot psychological resilience may help identify pilots’ stress-adaptation characteristics and potential support needs during recurrent training, psychological status surveys and training feedback, and may provide a measurement basis for research on fatigue, occupational burnout and non-technical skills training [7,14,15,16,17].

Existing measures of pilot resilience have relied largely on the 25-item Connor–Davidson Resilience Scale (CD-RISC-25) or its adapted versions [18]. The CD-RISC-25 is a widely used measure of resilience with broad content coverage. It captures multiple aspects of adaptive responses to stress, adversity, and adverse events [4,19]. It has therefore been applied in studies of resilience across different pilot populations [1,13]. However, evidence from pilot samples suggests that its item applicability and latent structure may not be consistent across contexts. In a study of female pilots in the United States, several items were removed during scale adaptation, and a two-factor structure was identified [2]. Among Chinese civil aviation pilots, Zhao [16] found that item 3, “sometimes fate or God can help”, and item 20, “having to act on a hunch”, were poorly aligned with the cultural background and occupational cognitive characteristics of this population. Their final model included two factors: decisiveness and adaptability. These findings indicate that, although the CD-RISC-25 has provided an important basis for pilot resilience research, its factor structure and decisions about item retention may depend on cultural context, occupational characteristics, and sample composition. For applications focused on overall resilience assessment, group comparison, and subsequent intervention research among civil aviation pilots, it remains unclear whether a brief measure with a more focused structure, clearer interpretation, and lower respondent burden is available.

The CD-RISC-10 provides a potential methodological solution to this issue. When re-examining the CD-RISC-25, Campbell-Sills and Stein [19] found that the factor structure of the original scale was unstable across samples and that some factor interpretations were limited. They therefore proposed a more concise 10-item unidimensional version for measuring overall resilience. Subsequent studies have reported favourable evidence for the internal consistency, structural validity, or measurement invariance of the CD-RISC-10 in Chinese college students and patients with depression [20], competitive athletes [21], and male military personnel with and without PTSD [22]. Direct comparisons of different CD-RISC versions have also suggested that the CD-RISC-10 provides a favourable balance between reliability, validity, and practical convenience [21,23,24]. Notably, the CD-RISC-10 does not retain items 3 or 20, which previous research identified as poorly aligned with the cultural or occupational context of Chinese civil aviation pilots [16]. It may therefore avoid some item-applicability problems previously reported for the CD-RISC-25 in pilot samples. Given its unidimensional structure and lower respondent burden, the CD-RISC-10 represents a promising candidate instrument for assessing overall resilience among pilots.

However, before the CD-RISC-10 can be used with confidence to assess overall resilience among pilots, its item-level measurement performance requires further examination. In a highly selected and extensively trained occupational group, some items may show weaker discrimination because their content is broad, response distributions are concentrated at higher score levels, or their incremental information about the latent resilience trait is limited. Retaining such low-contribution items may add little to measurement quality while increasing unnecessary respondent burden [25,26]. Previous studies have noted that repetitive or perceived redundant items can reduce respondents’ favourable evaluations of questionnaire design [27,28]. Greater respondent burden may also reduce willingness to participate, affect completion quality, and increase the risk of careless responding and compromised data quality [29]. Therefore, after establishing the structural fit of the CD-RISC-10, it is both methodologically and practically important to examine item quality, information contribution, and shortening potential [30,31,32].

Existing psychometric evaluations of the CD-RISC-10 commonly draw on classical test theory (CTT) and item response theory (IRT) [20,21,33,34,35,36]. CTT evaluates score reliability and associations with relevant external variables at the scale level, thereby indicating the basic measurement quality of total scores in the target population. Confirmatory factor analysis (CFA) provides complementary evidence of structural validity by examining whether the observed relationships among items are consistent with a hypothesized measurement structure [33,37,38]. Establishing a sufficiently tenable one-factor representation is particularly important before fitting and interpreting a unidimensional IRT model. However, CTT and CFA provide limited direct information about item discrimination, threshold distributions, and information contributions across different levels of the latent trait [39,40,41]. IRT complements these scale- and structure-level analyses by estimating item discrimination, threshold or location parameters, and item information. It also uses information functions to examine measurement precision across levels of latent resilience. IRT can therefore provide evidence for identifying items with comparatively lower measurement contributions or weaker item-level fit [42,43]. Item optimization, however, also requires determining whether score dependability remains adequate after item reduction. For this reason, the present study further incorporated generalizability theory (GT) [44]. Through a G-study, GT decomposes sources of score variance, including pilots, items, and their interaction. Through a D-study, it estimates the generalizability coefficient and phi coefficient under different item-number conditions, providing quantitative evidence for defining a provisional statistical reference length for scale shortening [16,45,46]. Accordingly, the present study integrated CTT and CFA for scale- and structure-level evaluation, GT for establishing a provisional reference length, and IRT for evaluating item-level measurement contributions, supplemented by qualitative item-content review.

Within this integrated framework, the present study used two non-overlapping samples from the same Chinese airline to address two primary questions. The first concerned whether the CD-RISC-10 showed adequate structural and score-level psychometric properties in Chinese male airline pilots, encompassing score reliability, the tenability of an approximate one-factor representation, and theoretically expected concurrent associations with occupational burnout. The second concerned whether the complementary psychometric evidence could support a defensible shortening of the CD-RISC-10, encompassing item optimization in the development sample and subsequent evaluation of the fixed candidate short form in a separate, non-overlapping sample. Before the second-stage evaluation, the candidate item set was fixed, and its score reliability, consistency with the prespecified one-factor structure, and concurrent associations with occupational burnout were then examined. This two-stage design distinguished evidence generated during item selection from evidence obtained after the candidate item set had been fixed. The study was intended to provide a psychometric basis for further evaluating low-burden resilience assessment for group-level research and training feedback in comparable recurrent-training settings and to inform future cross-airline and cross-population validation research.

2. Materials and Methods

2.1. Participants and Recruitment

2.1.1. Development Sample

For the development sample, flight instructors at an airline in China (hereafter, the participating airline) were asked to distribute a standardized study invitation and a link to an online questionnaire to male airline pilots attending recurrent training. The invitation explained the purpose and procedures of the study, the questionnaire content, the voluntary nature of participation, and the confidentiality of the responses. All participants provided electronic informed consent before completing the questionnaire.

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Aeronautical Science and Engineering (approval no. ASE-2025-040). Data were collected between 16 September and 16 October 2025. A total of 267 questionnaires were returned. Questionnaires were excluded if the respondent failed the attention-check item or selected the same response option for all CD-RISC-10 items. After data-quality screening, 192 questionnaires were retained for analysis, representing 71.9% of the returned questionnaires.

All 192 retained participants were male airline pilots. Their demographic and occupational characteristics, including age, flight role, cumulative flight time, and relationship/family status, are summarized in

Table 1. Demographic and occupational characteristics of the development and separate validation samples.

Characteristic	Category	Development Sample (N = 192)	Separate Validation Sample (N = 58)
Sex	Male	192 (100.0%)	58 (100.0%)
Age	21–30	89 (46.4%)	35 (60.3%)
	31–40	70 (36.5%)	18 (31.0%)
	>40	33 (17.2%)	5 (8.6%)
Flight role	First officer	102 (53.1%)	38 (65.5%)
	Captain	37 (19.3%)	8 (13.8%)
	Flight instructor	53 (27.6%)	12 (20.7%)
Cumulative flight time (h)	≤3000	75 (39.1%)	28 (48.3%)
	3001–7000	48 (25.0%)	14 (24.1%)
	>7000	69 (35.9%)	16 (27.6%)
Relationship/family status	Single	24 (12.5%)	8 (13.8%)
	In a relationship	33 (17.2%)	12 (20.7%)
	Married without children	35 (18.2%)	10 (17.2%)
	Married with children	100 (52.1%)	28 (48.3%)

. These 192 participants were included in the psychometric evaluation of the CD-RISC-10 and in the derivation and within-sample post-selection evaluation of a candidate short form. Among them, 137 also completed the occupational burnout questionnaire, and 106 provided valid paired resilience and occupational burnout data. Analyses involving occupational burnout in the development stage were based on these 106 paired observations.

2.1.2. Separate Validation Sample

To obtain preliminary validation evidence from participants who were not involved in item selection, a separate, non-overlapping sample of male airline pilots attending recurrent training at the same participating airline was recruited between 10 and 22 August 2026. Flight instructors distributed a standardized study invitation and a link to the online validation questionnaire to pilots attending the training. Participation was voluntary, and all participants provided electronic informed consent before completing the questionnaire.

The validation study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Aeronautical Science and Engineering (approval no. ASE-2026-061). The validation sample did not overlap with the development sample. A total of 60 pilots completed the online validation survey, which included the fixed candidate short form and the occupational burnout questionnaire. Data from two respondents were excluded because they failed the attention-check item. The remaining 58 participants had valid data for both measures, yielding a retention rate of 96.7%. All 58 retained participants were male airline pilots, and their demographic and occupational characteristics are presented in Table 1. Data from all 58 participants were used to evaluate the fixed candidate short form, and all 58 paired observations were included in the criterion-related validity analyses involving occupational burnout.

The item composition, scoring rule, and one-factor measurement model of the candidate short form were fixed before the validation data were examined. The validation sample was not used for further item selection, deletion, scoring modification, or post hoc model modification. Because the development and validation samples were recruited from the same airline and recurrent-training context, this sample was treated as a preliminary within-context validation sample rather than as evidence of external validation across airlines or pilot populations.

2.2. Materials

Connor–Davidson Resilience Scale. The Connor–Davidson Resilience Scale was originally developed by Connor and Davidson as a 25-item instrument for assessing psychological adaptation to stress, adversity, and adverse events [18]. Campbell-Sills and Stein later developed and validated a 10-item version, the CD-RISC-10, based on the CD-RISC-25 [19]. Their findings showed that the CD-RISC-10 measures a unidimensional resilience construct and has good internal consistency (Cronbach’s α = 0.85). This study used the Chinese version introduced by Yu [47]. The scale uses a 5-point Likert response format, with responses ranging from 0 (“not true at all”) to 4 (“true nearly all the time”). Total scores range from 0 to 40, with higher scores indicating greater psychological resilience. The 10 retained items correspond to items 1, 4, 6, 7, 8, 11, 14, 16, 17, and 19 of the original 25-item scale. Their item contents are “able to adapt to change”, “able to deal with whatever comes”, “try to see the humorous side of problems”, “coping with stress can strengthen me”, “tend to bounce back after illness or hardship”, “able to achieve goals despite obstacles”, “able to stay focused under pressure”, “not easily discouraged by failure”, “think of myself as a strong person”, and “able to handle unpleasant feelings”. The original CD-RISC-25 item numbers were retained to maintain comparability with previous studies.

In the development sample, participants completed the full CD-RISC-10. After item selection, scores for the candidate short form were calculated from the retained responses within the original CD-RISC-10 administration and therefore did not represent an independent administration of the short form. The fixed candidate short form was administered directly in the separate sample. Its final item composition, name, and scoring range are reported in Section 3.

Occupational burnout questionnaire. Because psychological resilience is significantly associated with occupational burnout [48,49], the occupational burnout questionnaire was used as an external criterion for examining the criterion-related validity of the resilience measure. The Maslach Burnout Inventory-General Survey (MBI-GS) is a widely used instrument in burnout research [50]. It contains 16 items and assesses three dimensions: emotional exhaustion, depersonalization, and personal accomplishment [51]. Each item is rated on a 7-point scale. According to the scale scoring rules, items related to personal accomplishment were reverse-scored so that the total burnout score and all dimension scores had the same direction, with higher scores indicating greater occupational burnout. Subsequent analyses used the total occupational burnout score and the scores for emotional exhaustion, depersonalization, and reduced personal accomplishment after reverse scoring.

2.3. Statistical Analysis

Analyses proceeded in two stages. In the development sample, the CD-RISC-10 was evaluated using descriptive statistics, score-reliability indices, associations with occupational burnout, and one-factor confirmatory factor analysis (CFA). Generalizability theory (GT), item response theory (IRT), and qualitative item-content review then informed the derivation of a candidate short form, which underwent post-selection evaluation using retained responses from the same participants. The fixed candidate short form was subsequently evaluated in the separate, non-overlapping sample.

SPSSAU (version 26.0) was used for descriptive statistics, reliability analyses, correlations with occupational burnout, CFA, and exploratory demographic comparisons. Development-sample adequacy for CFA was assessed using Monte Carlo simulation in Mplus 7.0 [52]. Generalizability theory analyses were conducted using the PC implementation of GENOVA. [53,54]. IRT analyses used the mirt package (version 1.46.1) in R version 4.6.1 [55], whereas robust CFA and local model diagnostics used the lavaan package (version 0.7-2). All tests were two-sided, with statistical significance set at p < 0.05.

2.3.1. Score Reliability and Associations with Occupational Burnout

Score reliability was estimated for the CD-RISC-10, the candidate short form, and the occupational burnout questionnaire using Cronbach’s α, Spearman–Brown split-half reliability, McDonald’s ω, and the theta coefficient. Cronbach’s α values greater than 0.80 were interpreted as indicating good internal consistency. In the development sample, correlations of resilience scores with total occupational burnout and its three dimensions were examined. Their direction and magnitude were interpreted against the theoretical expectation that greater resilience would be associated with lower occupational burnout [49].

2.3.2. Confirmatory Factor Analysis

In the development sample, a one-factor CFA was first fitted to the CD-RISC-10 using maximum-likelihood (ML) estimation in SPSSAU. After item selection, the same model was fitted to the retained item responses from the same 192 participants to characterize post-selection structural performance within the development sample. The prespecified CFA of the directly administered candidate short form in the separate sample is described in Section 2.3.6. Model fit was evaluated using the chi-square statistic, comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA) with its 90% confidence interval, and standardized root mean square residual (SRMR) [33,37,38].

To assess the sensitivity of the one-factor findings to potential non-normality in the five-category item responses, robust maximum-likelihood estimation (MLR) was additionally applied to the CD-RISC-10. The primary maximum-likelihood results were retained, and the robust results were used to assess whether the interpretation of the one-factor structure was sensitive to potential non-normality. Standardized factor loadings and standardized residual correlations were also examined to identify localized areas of model strain.

Average variance extracted (AVE) and composite reliability (CR) were calculated as supplementary indices of convergent validity and construct reliability [56]. Monte Carlo simulation assessed development-sample adequacy using parameter bias, standard-error bias, 95% confidence-interval coverage, and statistical power [57,58].

Modification indices (MIs) were calculated for all 45 pairwise residual covariances fixed to zero in the one-factor CD-RISC-10 model to examine potential local model misspecification. The same maximum-likelihood model was re-estimated in lavaan, and each MI was interpreted together with its unstandardized expected parameter change (EPC) and lavaan fully standardized expected parameter change (sepc.all). An MI of 3.84 was used only as a descriptive reference corresponding to a one-degree-of-freedom chi-square value at p = 0.05; it was not treated as a multiple-testing-adjusted significance threshold or an item-deletion criterion. No residual covariances were freely estimated on the basis of these diagnostics.

2.3.3. Generalizability Theory Analysis

A one-facet crossed p × i design was used, with pilots as the objects of measurement and items as the measurement facet [59]. The G-study estimated variance components for pilots, items, and the pilot-by-item interaction, whereas the D-study estimated the G coefficient for relative decisions and the phi coefficient for absolute decisions across alternative item-number conditions [37]. The smallest item number for which both point estimates reached 0.80 was carried forward as a provisional statistical reference for scale length. Because these projections treat items as exchangeable, the resulting reference length neither identified which items should be retained nor guaranteed equivalent performance for every item subset of that length. After the candidate item set had been selected, the coefficients were re-estimated for the retained items, with pilot-level 95% confidence intervals obtained from 5000 bootstrap resamples.

2.3.4. Item Response Theory Analysis

To examine the measurement characteristics of each CD-RISC-10 item in the development sample of Chinese male airline pilots, this study used the graded response model (GRM) within IRT [60,61]. The GRM is appropriate for items with multiple ordered response categories [32,62] and was therefore suitable because the CD-RISC-10 uses a five-point Likert response format. For the CD-RISC-10, item discrimination parameters, threshold parameters, item fit, and item information were estimated for each item. Category response curves and the test information function were then used to evaluate response-category functioning and item-level measurement contribution. After the candidate item set had been selected, GRM parameters were re-estimated for the candidate short form to examine the measurement performance of the retained items.

Global fit was assessed using the limited-information C₂ statistic, CFI, TLI, RMSEA with its 90% confidence interval, and SRMSR. C₂ was used as the primary statistic because M₂* can provide insufficient degrees of freedom in short polytomous models. AIC and BIC were reported descriptively within each fitted item set and were not used for direct comparisons between the original and candidate item sets. Item fit was examined using S−X² with Holm adjustment. Standard errors and 95% confidence intervals were reported for discrimination and estimable threshold parameters. Adjusted Q3 residual correlations with absolute values greater than 0.20 were treated as diagnostic flags rather than automatic item-deletion criteria.

Item reduction followed a sequential rather than numerically weighted decision process. The procedure first required a sufficiently tenable one-factor representation in the development sample. If this prerequisite was met, GT provided a provisional reference length. Within this reference length, item information was the primary ranking criterion, whereas discrimination, threshold location, response-category functioning, item fit, and local-dependence diagnostics provided supporting evidence. Lower-contributing items were then reviewed qualitatively to safeguard coverage of adaptive coping and recovery, goal-directed persistence, and attentional stability under pressure. This review was not numerically weighted and did not constitute a formal expert-rated content-validity analysis. The resulting candidate item set was subjected to post-selection analyses of score reliability, CFA, GT, IRT, and associations with occupational burnout in the development sample.

2.3.5. Exploratory Demographic Comparisons

Secondary exploratory comparisons of latent trait estimates from the candidate short form were conducted across age, flight role, cumulative flight time, and relationship/family status. Normality was assessed before group comparisons, and Kruskal–Wallis H tests were used when parametric assumptions were not met. Results were summarized using means and standard deviations and medians with interquartile ranges. Because these analyses were exploratory, unadjusted p values were interpreted descriptively. Sex differences were not examined because all participants were male.

2.3.6. Preliminary Validation Analyses in the Separate Sample

The fixed candidate short form was administered directly in the separate, non-overlapping sample after its item composition, scoring rule, and one-factor model had been specified. The validation data were not used for further item selection, scoring modification, residual-covariance specification, or other data-driven model revision. Item distributions and corrected item–total correlations were examined. Score reliability was estimated using Cronbach’s α with a bootstrap 95% confidence interval, Spearman–Brown split-half reliability, McDonald’s ω, and the theta coefficient. Score dependability was evaluated using a one-facet p × i GT analysis.

A prespecified one-factor CFA model was fitted using ML estimation. Given the modest sample size, the CFA was treated as a preliminary assessment of compatibility with the fixed measurement model. The fit indices described in Section 2.3.2 were reported.

3. Results

3.1. Descriptive Statistics

Descriptive statistics for each CD-RISC-10 item are shown in

Table 2. Descriptive statistics for CD-RISC-10 items.

Item	Proportion of Each Option					Mean	SD
	0	1	2	3	4
1	0	1.04	9.38	71.88	17.71	3.063	0.557
4	0.52	5.21	37.5	50	6.77	2.573	0.72
6	0.52	2.6	8.85	70.83	17.19	3.016	0.643
7	1.56	10.94	28.13	52.6	6.77	2.521	0.837
8	2.6	5.73	20.83	59.38	11.46	2.714	0.841
11	0.52	1.04	22.4	60.94	15.1	2.891	0.674
14	0.52	2.6	17.71	70.83	8.33	2.839	0.622
16	1.04	3.13	15.1	64.06	16.67	2.922	0.73
17	0.52	2.6	20.83	63.02	13.02	2.854	0.686
19	0	3.65	15.63	69.27	11.46	2.885	0.637
CD-RISC-10 total score	-	-	-	-	-	28.28	5.02

Note: Values under response options 0–4 are percentages. 0 = not true at all; 1 = rarely true; 2 = sometimes true; 3 = often true; 4 = true nearly all the time. SD = standard deviation. The CD-RISC-10 total score was calculated by summing the 10 items and ranged from 0 to 40. Response-option percentages are not applicable to the total-score row.

. The mean scores of the 10 items ranged from 2.521 to 3.063, falling between “sometimes true” and “often true” and closer overall to “often true”. Standard deviations ranged from 0.557 to 0.841, indicating moderate score dispersion across items. Item 1, “able to adapt to change”, had the highest mean score (M = 3.063, SD = 0.557), whereas item 7, “coping with stress can strengthen me”, had the lowest mean score (M = 2.521, SD = 0.837).

For response-option distributions, the proportion of responses scored as 0 (“not true at all”) ranged from 0% to 2.60%, 1 (“rarely true”) from 1.04% to 10.94%, 2 (“sometimes true”) from 8.85% to 37.50%, 3 (“often true”) from 50.00% to 71.88%, and 4 (“true nearly all the time”) from 6.77% to 17.71%. The mean CD-RISC-10 total score was 28.28 (SD = 5.02). Across items, responses were concentrated in category 3 (“often true”), whereas the lower response categories were infrequently endorsed. Because the study did not include a matched reference group or prespecify a norm-referenced classification, these values were not interpreted as indicating low, average, or high resilience. The response distributions were examined further using GRM category-response and item-information analyses.

3.2. Score Reliability and Concurrent Associations with Occupational Burnout

Based on CD-RISC-10 data from 192 pilots in the development sample, the scale showed high score reliability. Cronbach’s α was 0.895, Spearman–Brown split-half reliability was 0.868, McDonald’s ω was 0.914, and the theta coefficient was 0.896. These results showed that the CD-RISC-10 had good internal consistency and score reliability in the present sample. In the 106 cases with valid data for both psychological resilience and occupational burnout, the occupational burnout questionnaire and its dimensions also showed good reliability. For emotional exhaustion, depersonalization, and reduced personal accomplishment, Cronbach’s α values were 0.948, 0.836, and 0.803, respectively; Spearman–Brown split-half reliability values were 0.932, 0.810, and 0.861, respectively; McDonald’s ω values were 0.960, 0.895, and 0.863, respectively; and theta coefficients were 0.948, 0.860, and 0.812, respectively. For the total occupational burnout score, Cronbach’s α was 0.903, McDonald’s ω was 0.921, and the theta coefficient was 0.919. These results indicated that the occupational burnout measure used in this study also had high reliability and was suitable as an external variable for testing the criterion-related validity of the resilience measure.

As shown in

Table 3. Correlations between the CD-RISC-10 and occupational burnout.

Variable	CD-RISC-10
Total occupational burnout score	−0.482 **
Emotional exhaustion	−0.402 **
Depersonalization	−0.440 **
Reduced personal accomplishment (after reverse scoring)	−0.315 **

Note: N = 106. ** p < 0.01.

, the CD-RISC-10 total score was significantly negatively correlated with the total occupational burnout score (r = −0.482, p < 0.01), emotional exhaustion (r = −0.402, p < 0.01), depersonalization (r = −0.440, p < 0.01), and reduced personal accomplishment after reverse scoring (r = −0.315, p < 0.01). These findings were consistent with the theoretical expectation that higher psychological resilience is associated with lower occupational burnout, providing concurrent criterion-related evidence for the CD-RISC-10 in the development sample of Chinese male airline pilots.

3.3. Suitability of the CD-RISC-10 Based on Confirmatory Factor Analysis

To examine the structural fit of the CD-RISC-10 in the development sample of Chinese male airline pilots, confirmatory factor analysis (CFA) was used to test its one-factor model. Because CFA parameter estimation is affected by sample size, Monte Carlo simulation was further used to evaluate whether the sample size supported model estimation. The relative bias of parameter estimates ranged from 0% to 1.92%, all below the recommended threshold of 10%. The relative bias of standard errors ranged from −4.08% to 2.12%, within the acceptable range of ±5%. The 95% confidence interval coverage ranged from 0.91 to 0.98, and the statistical power for all target parameters was 1.00. These results indicated that the sample size provided acceptable support for parameter estimation in the one-factor CFA model.

The one-factor CFA model converged successfully. All standardized factor loadings were greater than 0.40, indicating that all items reflected the common latent resilience factor to some extent. The average variance extracted (AVE) was 0.466, slightly below the commonly recommended threshold of 0.50. Composite reliability (CR) was 0.896, indicating high overall consistency in the extent to which the items represented the common latent construct. Model fit results are shown in

Table 4. Model fit indices for the one-factor CD-RISC-10 model.

Estimator	χ2	df	p	CFI	TLI	RMSEA	RMSEA 90% CI	p-Close	SRMR
ML, primary	81.21	35	<0.001	0.944	0.928	0.083	0.059–0.107	0.013	0.050
MLR, sensitivity	63.62	35	0.002	0.954	0.941	0.074	0.044–0.102	0.089	0.050

Note: The MLR row reports the scaled chi-square statistic and robust CFI, TLI, and RMSEA estimates. The p-close test evaluates H0: RMSEA ≤ 0.05. ML = maximum likelihood; MLR = robust maximum likelihood; CFI = comparative fit index; TLI = Tucker–Lewis index; RMSEA = root mean square error of approximation; SRMR = standardized root mean square residual.

.

The primary ML model yielded χ²(35) = 81.21, p < 0.001, CFI = 0.944, TLI = 0.928, RMSEA = 0.083 (90% CI [0.059, 0.107]), p-close = 0.013, and SRMR = 0.050. The RMSEA results therefore did not support a close fit, whereas the CFI, TLI, and SRMR supported an adequate approximate one-factor representation. Standardized factor loadings ranged from 0.554 to 0.772, and all loading 95% confidence intervals remained above 0.40, with the lowest lower confidence limit being 0.447. The maximum absolute standardized residual correlation was 0.132. In the MLR sensitivity analysis, the scaled chi-square statistic was χ²(35) = 63.62, p = 0.002, with robust CFI = 0.954, robust TLI = 0.941, robust RMSEA = 0.074 (90% CI [0.044, 0.102]), p-close = 0.089, and SRMR = 0.050. Taken together, the primary and sensitivity analyses supported the one-factor model as an adequate working representation for subsequent item-level analyses, but they did not provide unequivocal evidence of close or exact fit.

Among the 45 pairwise residual covariances fixed to zero, nine had MI values of at least 3.84. Seven of these nine pairs involved at least one item later removed during the shortening procedure, and all four later-removed items appeared in at least one of these pairs. The three largest MIs were observed for items 16 and 17 (MI = 13.17; sepc.all = 0.290), items 17 and 19 (MI = 10.79; sepc.all = 0.259), and items 16 and 19 (MI = 10.12; sepc.all = 0.251). The largest MIs involving items 1 and 6 were 7.79 for items 1 and 17 (sepc.all = −0.220) and 9.86 for items 6 and 7 (sepc.all = 0.255), respectively. Across the nine pairs, the absolute sepc.all values ranged from 0.191 to 0.290. Notably, retained item 16 was included in two of the three pairs with the largest MIs, showing that comparatively large MIs were not confined to subsequently removed items. The MIs were therefore interpreted as pair-specific diagnostics of localized residual dependence rather than as item-level deletion statistics. No residual covariance was freed, and the item-selection procedure remained unchanged.

3.4. Measurement Tool Optimization Analysis

3.4.1. Item-Number Analysis Based on Generalizability Theory

Given that the one-factor structure of the CD-RISC-10 was supported as an adequate working representation, generalizability theory (GT) was further used to examine sources of score reliability and to evaluate whether the scale could maintain acceptable measurement reliability after item reduction.

The variance component estimates from the G-study are shown in

Table 5. Model Variance Components from G-Study.

Effect	df	Variance Component Estimate	Standard Error
p (pilots)	191	0.2252533	0.0256431
i (items)	9	0.0296799	0.0132440
p × i interaction	1719	0.2649381	0.0090317

. The variance component for the pilot main effect was 0.2252533, accounting for approximately 43.33% of the total variance. This indicated clear individual differences in psychological resilience scores among pilots. The variance component for the item main effect was 0.0296799, accounting for approximately 5.71% of the total variance. This indicated small differences in average item scores, suggesting limited differences among items in overall difficulty or average response level. The variance component for the pilot-by-item interaction was 0.2649381, accounting for approximately 50.96% of the total variance, and was the largest source of variance in the model. This result showed that pilots differed in their response patterns across items. It also indicated that total-score reliability indices alone were insufficient for judging the retention value of specific items.

The D-study results are shown in

Table 6. G and phi coefficients by item number in the D-study.

Pilots’ Sample Size	Items’ Sample Size	G Coefficient	Phi Coefficient
192	4	0.77277	0.75359
	5	0.80956	0.79265
	6	0.83610	0.82102
	7	0.85615	0.84257
	8	0.87182	0.85948
	9	0.88442	0.87311
	10	0.89476	0.88433

. Under the full 10-item condition, the CD-RISC-10 had a G coefficient of 0.89476 and a phi coefficient of 0.88433, indicating high reliability for both relative and absolute decisions in the present sample. As the number of items decreased, both the G coefficient and phi coefficient declined gradually. This indicated that item reduction reduced measurement stability, although the overall decrease was relatively modest. When the number of items was reduced to five, the G coefficient remained above 0.80 (G = 0.80956), indicating that five items could maintain acceptable reliability if the primary purpose was to compare relative differences in resilience among pilots. However, the phi coefficient under the five-item condition was 0.79265, slightly below 0.80, and therefore did not meet the prespecified point-estimate criterion for absolute decisions. When six items were retained, both the G coefficient (0.83610) and phi coefficient (0.82102) exceeded 0.80. Thus, under the point-estimate criterion and when both relative and absolute decisions were considered, six items provided the more conservative reference length.

Pilot-level bootstrap analyses indicated uncertainty around this boundary. For the five-item condition, the 95% bootstrap intervals were [0.761, 0.844] for G and [0.740, 0.829] for phi. For the six-item condition, the corresponding intervals were [0.792, 0.867] and [0.773, 0.854]. Because these intervals overlapped 0.80, the six-item result was interpreted as a development-sample point-estimate reference rather than as a universally guaranteed threshold.

Overall, the GT results identified six items as the point-estimate-based reference length for subsequent item-selection analysis, while the bootstrap intervals indicated uncertainty around this boundary.

3.4.2. Item Information Analysis Based on Item Response Theory

After the unidimensional structure of the CD-RISC-10 was shown to be generally acceptable, the graded response model (GRM) within item response theory (IRT) was further used to analyse the measurement characteristics of each item. The model information criteria were AIC = 3095.96 and BIC = 3252.32. Item parameter estimates are shown in

Table 7. Item parameter estimates from the graded response model for the CD-RISC-10.

Item	a	b1	b2	b3	b4	Item Information	S−X2	p (S−X2)
1	1.956	—	−3.106	−1.648	1.226	5.020	13.26	0.21
4	2.159	−3.401	−1.957	−0.236	2.001	7.255	8.28	0.51
6	1.636	−3.866	−2.697	−1.671	1.367	4.191	14.93	0.38
7	2.647	−2.583	−1.327	−0.314	1.844	9.222	14.39	0.35
8	2.874	−2.266	−1.569	−0.638	1.427	9.423	13.21	0.43
11	2.610	−3.084	−2.522	−0.833	1.235	8.477	14.15	0.17
14	2.845	−3.070	−2.128	−0.924	1.685	9.895	15.26	0.08
16	2.246	−2.919	−2.097	−1.040	1.205	6.932	14.88	0.32
17	2.135	−3.373	−2.302	−0.876	1.452	6.753	15.78	0.40
19	2.134	—	−2.206	−1.072	1.575	5.585	15.31	0.43

Note: a = discrimination parameter; b1–b4 = threshold parameters between adjacent response categories; — indicates that the threshold parameter could not be stably estimated because the lowest response category was not selected.

.

Limited-information fit assessment for the CD-RISC-10 yielded C₂(35) = 79.78, p < 0.001, CFI = 0.976, TLI = 0.969, RMSEA = 0.082 (90% CI 0.058–0.105), and SRMSR = 0.062. The CFI, TLI, and SRMSR therefore indicated favourable incremental and residual fit, whereas the significant C₂ test and RMSEA confidence interval indicated some remaining global misfit. The GRM was consequently interpreted as providing adequate but not unequivocal fit for item-level evaluation.

As a sensitivity diagnostic, M₂*(7) = 28.33, p < 0.001, CFI = 0.688, TLI = 0.243, and RMSEA = 0.126 (90% CI [0.080, 0.176]). Because M₂* had only seven degrees of freedom, it was not treated as the primary global-fit diagnostic. Nevertheless, its less favourable result reinforced the need for a cautious interpretation of the GRM fit.

The discrimination parameters (a) of the 10 items ranged from 1.636 to 2.874, with standard errors ranging from 0.255 to 0.438. Item 6 had the lowest discrimination estimate (a = 1.636, SE = 0.255, 95% CI [1.138, 2.135]). Except for item 6, which had a discrimination parameter of 1.636 and was slightly below the reference criterion of 1.70 for “very high discrimination”, the remaining nine items all had discrimination parameters greater than 1.70. This indicated that most items distinguished relatively well among pilots with different levels of psychological resilience. In the relative ranking, items 6, 1, 19, and 17 showed lower discrimination than the other items, with item 6 showing the lowest discrimination.

The threshold parameters generally increased as response categories increased, and no threshold disordering was observed. This indicated that the response categories of the CD-RISC-10 were ordered overall. For items 1 and 19, the lowest threshold parameter, b1, could not be stably estimated because no participant selected the 0-point response option (“not true at all”). For items 4, 6, 11, and 17, the absolute values of b1 were greater than 3, indicating that the transition point between the lowest response category and higher categories was located at the very low end of the latent trait distribution. This finding reflected the low use of the 0-point response option in the present sample. For item 6, b1 was −3.866, and the subsequent threshold parameters were also located toward the lower end of the latent trait distribution, indicating an uneven response-category distribution and relatively limited discrimination among pilots with lower resilience levels. Across the estimable threshold parameters, standard errors ranged from 0.105 to 0.774. The largest standard error was observed for the lowest threshold of item 6 (b1 = −3.866, SE = 0.774, 95% CI [−5.383, −2.349]), consistent with the sparse use of the lowest response categories. In terms of item fit, the unadjusted S−X² p values ranged from 0.084 to 0.506, and the corresponding Holm-adjusted p values ranged from 0.839 to 1.000. Thus, no CD-RISC-10 item showed statistically significant item-level misfit after adjustment for multiple testing, and the item-fit results did not identify an item that required deletion solely on the basis of S−X².

Four adjusted Q3 residual correlations exceeded 0.20 in absolute magnitude: items 6 and 7 (0.271), items 16 and 17 (0.270), items 17 and 19 (0.250), and items 16 and 19 (0.248). All four pairs involved at least one item later removed during the shortening procedure. These values were treated as post hoc local-dependence diagnostics and were not used as item-deletion criteria.

In terms of item information, item information values ranged from 4.191 to 9.895. The total information across the 10 items was 72.753, and the mean item information value was 7.275. Item 14 had the highest information value (9.895), followed by item 8 (9.423), item 7 (9.222), and item 11 (8.477), indicating that these items contributed more information for measuring the latent resilience trait. In contrast, item 6 had the lowest information value (4.191), followed by item 1 (5.020), item 19 (5.585), and item 17 (6.753). Item 16 also had an information value below the mean (6.932), but its value remained higher than those of the four lower-information items. These results indicated uneven measurement contributions among CD-RISC-10 items.

The category response curves are shown in Figure 1. Because no participant selected the 0-point response option for items 1 and 19, their category response curves did not fully represent all five response categories. Overall, lower response categories, especially scores of 0 and 1, had low probabilities of being selected across several items, indicating sparse use of the lowest response options in the development sample. For items 6 and 11, the response curve for the 1-point category was relatively flat, indicating limited discrimination across different levels of latent resilience for this category. Some category response curves for item 6 also had low peaks, further indicating insufficient use of response categories and potentially reduced measurement efficiency for this item. Although item 11 also showed limited use of lower response categories, it had high item information and a good discrimination parameter. Therefore, it was not prioritized for deletion solely on the basis of its category response curves.

The test information function of the CD-RISC-10 was obtained by summing the information functions of its 10 items (Figure 2a). Test information was relatively high in the lower region of the latent resilience continuum, approximately from θ = −2.5 to −1.0. Information was also relatively high in the upper-middle region, approximately from θ = 1.0 to 2.0, indicating greater measurement precision in these regions. In contrast, test information was comparatively lower near the centr of the latent trait scale. Consistent with the relationship $SE(\theta )=1/\sqrt{I(\theta )}),$ regions with higher information showed lower conditional standard errors, whereas the central region showed a comparatively higher standard error (Figure 2b).

Overall, the CD-RISC-10 showed generally strong item discrimination and no statistically significant item-level misfit after Holm adjustment, although the global fit evidence was mixed and several local-dependence flags were observed. The measurement contributions of individual items were nevertheless uneven. Compared with the other items, items 6, 1, 19, and 17 showed relatively weaker performance in discrimination, threshold distribution, response-category use, or item information. These items were therefore prioritized for further consideration in the subsequent shortening analysis.

3.4.3. Scale Shortening Based on Integrated Evidence from CTT, GT and IRT

The evidence from classical test theory, CFA, GT, and IRT was combined sequentially rather than through a numerical weighting scheme. CFA provided the prerequisite for considering item reduction within a common latent construct, whereas GT provided a point-estimate-based reference length of six items. Within this reference length, IRT discrimination, threshold and response-category functioning, and item information were used to compare item-level measurement contributions. Reliability results and qualitative item-content review served as safeguards, whereas the MI and adjusted Q3 diagnostics were not used as item-deletion criteria. Items 1, 6, 17, and 19 showed relatively lower measurement contributions. Item 6 had the lowest discrimination (a = 1.636), the lowest item information (4.191), and relatively extreme lower thresholds, indicating insufficient response-category use and relatively limited measurement efficiency in the present sample. Items 1 and 19 had no stable estimates for the lowest threshold parameter because no participant selected the 0-point response option, and their item information values, 5.020 and 5.585, respectively, were both below the mean item information. Although the discrimination and item information of item 17 remained within an acceptable range, both were relatively low among the 10 items, and its lower threshold was also located toward the lower end of the latent trait distribution. Based on this integrated evidence, items 1, 6, 17, and 19 were removed, and items 4, 7, 8, 11, 14, and 16 were retained to form a candidate six-item short form, the CD-RISC-6. Each retained item used the original 0–4 response scale, yielding total scores ranging from 0 to 24. The qualitative item-content review served as a coverage safeguard and was not a formal expert-rated content-validity analysis.

The shortened CD-RISC-6 still showed good internal consistency. Cronbach’s α was 0.866, Spearman–Brown split-half reliability was 0.874, McDonald’s ω was 0.901, and the theta coefficient was 0.869, all exceeding 0.80. These results indicated that the shortened scale retained high score reliability after four items were removed. Compared with the original CD-RISC-10, the CD-RISC-6 reduced the number of items by 40% while maintaining acceptable to high internal consistency and split-half reliability in the present sample. This provided preliminary support for its feasibility as a candidate brief measure. When GT was re-estimated for the six items actually retained, the G coefficient was 0.866 (95% bootstrap CI [0.816, 0.899]) and the phi coefficient was 0.854 (95% bootstrap CI [0.800, 0.889]). These findings provided additional evidence of acceptable score dependability for the selected six-item set in the development sample.

Further CFA results showed that the one-factor model of the candidate CD-RISC-6 yielded χ²(9) = 13.22, p = 0.153, CFI = 0.991, TLI = 0.985, RMSEA = 0.049 (90% CI [0.000, 0.102]), SRMR = 0.027, AVE = 0.527, and CR = 0.869. In the development sample, the candidate CD-RISC-6 yielded CFI, TLI, RMSEA, and SRMR values that were more favourable than those of the CD-RISC-10, and its AVE exceeded 0.50. Because the same sample was used for item selection and post-selection evaluation, this pattern may partly reflect sample-dependent optimization. It should therefore be interpreted as post-selection structural performance within the selection sample, not as independent cross-validation or evidence of superiority over the CD-RISC-10.

From the perspective of item information, the six retained items had a total information value of 51.204 based on the original IRT parameters of the CD-RISC-10, accounting for 70.38% of the total information of the original 10-item scale (72.753). Thus, after a 40% reduction in item number, the six retained items still preserved more than 70% of the measurement information of the original scale. After GRM parameters were re-estimated for the CD-RISC-6, the total information value of the six items was 53.32, with model information criteria of AIC = 2004.57 and BIC = 2102.29. These values were reported descriptively and were not compared directly with those of the CD-RISC-10 because the models were fitted to different item sets. The item parameters are shown in

Table 8. Item parameter estimates from the graded response model for the CD-RISC-6.

Item	a	b1	b2	b3	b4	Item Information	S−X2	p (S−X2)
4	2.46	−3.39	−1.89	−0.22	1.89	9.10	10.16	0.12
7	2.33	−2.80	−1.38	−0.31	1.90	8.07	8.51	0.39
8	3.05	−2.29	−1.54	−0.61	1.39	10.28	12.66	0.049
11	2.69	−3.23	−2.58	−0.82	1.22	9.20	7.21	0.21
14	2.81	−3.28	−2.18	−0.92	1.66	10.19	3.69	0.45
16	2.05	−3.12	−2.19	−1.07	1.25	6.48	12.51	0.13

Note: a = discrimination parameter; b1–b4 = threshold parameters between adjacent response categories; S−X² = item-fit statistic. For item 8, the unadjusted S−X² p value was 0.049, and the Holm-adjusted p value was 0.293. No retained item showed statistically significant item-level misfit after Holm adjustment.

. The re-estimated CD-RISC-6 GRM yielded C₂(9) = 12.55, p = 0.184, CFI = 0.995, TLI = 0.992, RMSEA = 0.045 (90% CI [0.000, 0.099]), and SRMSR = 0.038. Discrimination standard errors ranged from 0.297 to 0.459, and threshold standard errors ranged from 0.108 to 0.548. M₂* could not be calculated because the six-item model provided insufficient degrees of freedom. No adjusted Q3 residual correlation exceeded 0.20, with a maximum absolute value of 0.184. Item 8 had an unadjusted S−X² p value of 0.049 but was not significant after Holm adjustment (adjusted p = 0.293); no retained item showed statistically significant item-level misfit after adjustment. The IRT-based latent trait estimates (θ) from the CD-RISC-6 were highly positively correlated with the original CD-RISC-10 total scores (r = 0.969, p < 0.001), indicating high consistency in individual ranking between the candidate short form and the original version. Because the CD-RISC-6 items were derived from the CD-RISC-10, this correlation mainly reflects consistency between the two scoring results and should not be interpreted as independent validity evidence.

The concurrent associations with occupational burnout are shown in

Table 9. Correlations between the CD-RISC-6 and occupational burnout.

Variable	CD-RISC-6
Total occupational burnout score	−0.494 **
Emotional exhaustion	−0.418 **
Depersonalization	−0.457 **
Reduced personal accomplishment (after reverse scoring)	−0.311 **

Note: ** p < 0.01.

. The CD-RISC-6 remained significantly negatively correlated with the total occupational burnout score and all burnout dimensions. Specifically, the correlation between the CD-RISC-6 and total occupational burnout was −0.494. The correlations with emotional exhaustion, depersonalization, and reduced personal accomplishment after reverse scoring were −0.418, −0.457, and −0.311, respectively, all reaching p < 0.01. These correlation coefficients were close to those observed for the CD-RISC-10, indicating that the shortened version showed directionally consistent associations of similar magnitude with the external criterion after item reduction. This result provided preliminary support for the feasibility of the CD-RISC-6 as a candidate brief measure of psychological resilience among Chinese male airline pilots.

Figure 3 shows the distribution of IRT-based latent trait estimates (θ) among the 192 pilots. The estimates were concentrated mainly between approximately θ = −1.0 and 0.7, with observations extending across a wider range. This distribution is described relative to the modelled latent-trait scale and was not interpreted as a norm-referenced classification of low, average, or high resilience.

Taken together, the integrated CTT, CFA, GT, and IRT results provided preliminary support for shortening the CD-RISC-10 to a candidate six-item version. After item reduction, the CD-RISC-6 maintained high internal consistency, a favourable post-selection one-factor fit pattern, substantial item information, and directionally consistent associations with the original CD-RISC-10 and occupational burnout in the development sample. Therefore, the CD-RISC-6 can be considered a candidate short form, and its psychometric significance and applied value are further examined in the Discussion.

3.5. Demographic Differences in Latent Trait Estimates

As a secondary exploratory analysis, CD-RISC-6 latent trait estimates from the development sample were compared across demographic and occupational groups, as shown in

Table 10. Exploratory demographic and occupational comparisons of CD-RISC-6 latent trait estimates in the development sample.

Characteristic	Group	n	Mean ± SD	Median (P25, P75)	Kruskal–Wallis H	p
Age	21–30	89	−0.14 ± 0.93	−0.234 (−0.7, 0.4)	5.765	0.056
	31–40	70	0.01 ± 0.88	0.041 (−0.7, 0.4)
	>40	33	0.36 ± 1.01	0.388 (−0.3, 1.1)
Flight role	First officer	102	−0.17 ± 0.93	−0.275 (−0.7, 0.4)	8.674	0.013
	Captain	37	0.31 ± 0.80	0.388 (−0.3, 0.7)
	Flight instructor	53	0.11 ± 0.98	0.041 (−0.7, 0.6)
Cumulative flight time (h)	≤3000	75	−0.19 ± 0.96	−0.271 (−0.7, 0.4)	5.935	0.051
	3001–7000	48	−0.04 ± 0.85	0.055 (−0.6, 0.4)
	>7000	69	0.23 ± 0.93	0.388 (−0.5, 0.8)
Relationship/family status	Single	24	−0.22 ± 0.96	−0.275 (−0.5, 0.4)	8.505	0.037
	In a relationship	33	−0.30 ± 0.95	−0.291 (−0.8, 0.4)
	Married without children	35	−0.02 ± 0.85	−0.291 (−0.6, 0.4)
	Married with children	100	0.16 ± 0.94	0.284 (−0.5, 0.7)

Note: Values are presented as mean ± SD and median (P25, P75). The p values are unadjusted omnibus Kruskal–Wallis test results; no post hoc pairwise comparisons were conducted.

. The Kruskal–Wallis tests yielded unadjusted p values of 0.013 for flight role and 0.037 for relationship/family status, whereas the corresponding p values for age and cumulative flight time were 0.056 and 0.051, respectively. The omnibus tests did not identify which specific groups differed, and no post hoc pairwise, interaction, or multivariable analyses were conducted. Given the exploratory nature of these analyses, the absence of multiplicity adjustment, the limited subgroup sizes, and the interrelationships among age, flight role, cumulative flight time, and relationship/family status, these findings were interpreted as within-sample descriptive patterns rather than evidence of independent demographic or occupational effects.

3.6. Preliminary Validation of the Candidate CD-RISC-6 in a Separate Sample

A separate, non-overlapping validation sample of 58 male airline pilots recruited from the same participating airline was used to evaluate the candidate CD-RISC-6, which was administered directly as a six-item questionnaire. The demographic and occupational characteristics of the validation sample are presented in Table 1. The six-item composition, scoring rule, and one-factor measurement model were fixed before the validation data were examined. No further item selection, item deletion, or post hoc model modification was performed using the validation sample.

Item-level results are presented in

Table 11. Item characteristics and standardized factor loadings of the CD-RISC-6 in the separate validation sample.

Item	Mean	SD	Corrected Item–Total Correlation	Standardized Factor Loading	n (Score 0)	n (Score 1)
4	2.707	0.749	0.775	0.836	0	4
7	2.500	0.800	0.685	0.723	0	9
8	2.793	0.744	0.729	0.792	0	3
11	3.000	0.649	0.644	0.676	0	0
14	2.810	0.783	0.756	0.819	1	3
16	2.983	0.688	0.674	0.708	0	3

Note: N = 58. The original CD-RISC-25 item numbering is retained. Each item was scored from 0 (“not true at all”) to 4 (“true nearly all the time”). SD = standard deviation. The final two columns report the numbers of participants selecting response categories 0 and 1, respectively; the reported values are frequencies rather than item scores.

. Mean item scores ranged from 2.50 to 3.00, and standard deviations ranged from 0.65 to 0.80. Corrected item–total correlations ranged from 0.644 to 0.775, indicating that all six items contributed positively to the total score. The mean CD-RISC-6 total score was 16.79 (SD = 3.56; median = 17.50). Responses were concentrated in the 3-point category (“often true”). Six of the 30 item-by-category cells were empty: the 0-point category was not selected for items 4, 7, 8, 11, and 16, and neither the 0-point nor the 1-point category was selected for item 11. These sparse lower-response categories were considered when interpreting the model-based results, and a new graded response model calibration was therefore not undertaken in the validation sample.

The CD-RISC-6 showed high internal consistency. Cronbach’s α was 0.891, with a bootstrap 95% confidence interval of 0.824–0.925. Spearman–Brown split-half reliability was 0.859, McDonald’s ω was 0.892, and the theta reliability coefficient was 0.892. In the one-facet pilot-by-item generalizability analysis, the estimated variance components for pilots, items, and the pilot-by-item interaction were 0.313, 0.030, and 0.230, respectively. The G coefficient for relative decisions was 0.891, and the phi coefficient for absolute decisions was 0.878. Thus, both coefficients exceeded 0.80, indicating acceptable score dependability for the fixed six-item form in this validation sample.

The prespecified one-factor CFA model converged without adding residual covariances. All standardized factor loadings were positive and ranged from 0.676 to 0.836. The model yielded χ²(9) = 8.509, p = 0.484, χ²/df = 0.945, CFI = 1.000, TLI = 1.000 (unbounded estimate = 1.005), RMSEA = 0.000, 90% CI [0.000, 0.143], and SRMR = 0.036. The average variance extracted was 0.580, and composite reliability was 0.892. Although the point estimates were compatible with the prespecified one-factor model, their precision was limited by the modest sample size, as reflected in the RMSEA 90% confidence interval extending to 0.143. Accordingly, the CFA findings were interpreted as preliminary evidence of compatibility with the prespecified structure rather than confirmation of model stability.

The occupational burnout measure and its dimensions also showed acceptable to high internal consistency. Cronbach’s α was 0.956 for emotional exhaustion, 0.890 for depersonalization, 0.802 for reduced personal accomplishment, and 0.927 for the total occupational burnout score. As shown in

Table 12. Correlations between the CD-RISC-6 and occupational burnout in the separate validation sample.

Variable	r	95% CI	p
Total occupational burnout score	−0.540	[−0.701, −0.328]	<0.001
Emotional exhaustion	−0.502	[−0.673, −0.280]	<0.001
Depersonalization	−0.503	[−0.674, −0.282]	<0.001
Reduced personal accomplishment (after reverse scoring)	−0.347	[−0.555, −0.097]	0.008

Note: N = 58. Reduced personal accomplishment was reverse-scored so that higher values indicated greater occupational burnout. CI = confidence interval.

, the CD-RISC-6 score was negatively correlated with total occupational burnout (r = −0.540, 95% CI [−0.701, −0.328], p < 0.001), emotional exhaustion (r = −0.502, 95% CI [−0.673, −0.280], p < 0.001), depersonalization (r = −0.503, 95% CI [−0.674, −0.282], p < 0.001), and reduced personal accomplishment after reverse scoring (r = −0.347, 95% CI [−0.555, −0.097], p = 0.008). The directions and approximate magnitudes of these associations were consistent with those observed in the development sample.

Overall, the separate sample showed high score reliability and expected concurrent associations with occupational burnout, whereas the precision of the CFA estimates remained limited by the modest sample size.

4. Discussion

4.1. Psychometric Performance of the CD-RISC-10 Among Chinese Male Airline Pilots

This study first examined the structural fit and basic measurement quality of the CD-RISC-10 among Chinese male airline pilots. The results showed that the CD-RISC-10 had high score reliability and significant negative associations with the total occupational burnout score and each burnout dimension, whereas the CFA and GRM were consistent with an adequate approximate one-factor representation but also indicated uncertainty in global fit and several localized areas of model strain. Taken together, these results provide preliminary evidence that the CD-RISC-10 may index overall self-reported psychological resilience within this airline and recurrent-training context.

This finding is broadly consistent with the original unidimensional purpose of the CD-RISC-10, which was developed to assess overall resilience [19]. Compared with the CD-RISC-25, the CD-RISC-10 emphasizes brief assessment of overall resilience rather than differentiation among multiple latent dimensions. In applied contexts such as recurrent pilot training [16], psychological status surveys [20], and human factors research, a brief, structurally focused, and easy-to-administer instrument has practical advantages.

At the same time, the findings also suggest that the CD-RISC-10 leaves room for further optimization in this sample. Although the CFA incremental and residual fit indices supported an adequate one-factor approximation, the RMSEA confidence interval and p-close test did not provide unequivocal evidence of close fit. The GRM similarly showed mixed global-fit evidence and several adjusted Q3 local-dependence flags, while the AVE was slightly below the commonly recommended threshold. Descriptive statistics and IRT results further indicated limited use of some lower response categories and uneven item information across items.

Thus, within the present airline and recurrent-training context, the CD-RISC-10 may serve as a basic measure of overall self-reported psychological resilience, but not all items appear to operate with equal measurement efficiency in this highly selected and extensively trained occupational group. On this basis, examining item-level measurement performance and the potential for scale shortening is methodologically meaningful and practically relevant.

From the perspective of the aviation occupational context, this pattern is plausible. Airline pilots typically undergo rigorous selection and long-term standardized training, and they routinely operate in task environments characterized by strong procedural norms, high responsibility, and stringent safety requirements [25,26]. In this context, responses to general resilience items concerning adaptation to change, stress coping, and emotional control may be more likely to cluster in the upper response categories. Items that discriminate well in general populations may therefore provide less measurement information in a pilot sample. This response pattern helps explain the lower information of several items and provides a rationale for examining item optimization in this sample.

4.2. Derivation, Theoretical Interpretation, and Preliminary Validation of the CD-RISC-6

The rationale for the candidate CD-RISC-6 was based on converging evidence concerning scale length, item functioning, item content, and score performance rather than on a single statistical deletion rule. These evidence sources played sequential and non-interchangeable roles rather than being combined through numerical weights: CFA provided the structural prerequisite, GT provided a statistical reference for the target scale length, IRT ranked item-level measurement contributions, and qualitative item-content review served as a coverage safeguard. The MI and adjusted Q3 analyses were post hoc diagnostics and were not used as item-deletion criteria. The GT D-study identified six items as a point-estimate-based reference length, although bootstrap intervals indicated uncertainty around this boundary. Because the D-study assumes exchangeable items, whereas IRT item information ranged from 4.191 to 9.895, this reference did not guarantee equivalent performance for every six-item subset. IRT evidence and qualitative item-content review were therefore used to select the specific retained items. The resulting CD-RISC-6 retained 70.38% of the original test information, alongside high internal consistency and a favourable post-selection one-factor fit pattern. Taken together, these within-sample findings were consistent with the selected set being adequate for further evaluation as a candidate short form, but they do not imply that the removed item content is theoretically irrelevant [17,18,63,64].

The lower measurement contributions of the four removed items can be interpreted at both measurement and content levels. Items 1 (“able to adapt to change”) and 19 (“able to handle unpleasant feelings”) received no responses in the lowest category, indicating restricted response variation at the lower end of the latent trait. Item 6 (“try to see the humorous side of problems”) showed the lowest discrimination and item information, together with limited use of its lower response categories. Humour may function as one possible coping or affect-regulation strategy, but the appropriateness and effectiveness of any particular regulation strategy depend on the characteristics of the stressor and the broader context [65]. Item 17 (“think of myself as a strong person”) represents a relatively global self-evaluation and may be more sensitive to occupational identity or self-presentation than items referring to specific adaptive responses. In a highly selected and extensively trained pilot sample, these broad or globally desirable statements may elicit relatively homogeneous upper-category responses, thereby reducing their ability to differentiate individuals. These explanations remain theoretically plausible interpretations of the observed response patterns rather than mechanisms directly tested in the present study.

The content of the six retained items can be understood as representing two closely connected themes within an overall resilience score, rather than as empirically established subdimensions. Items 4, 7, 8, and 16 concern adaptive engagement with stress, recovery after difficulty, and the maintenance of functioning following setbacks. Items 11 and 14 concern goal-directed persistence and the ability to maintain attention under pressure. This qualitative review was intended to examine item-content coverage and did not constitute a formal expert-rated content-validity analysis. These themes are consistent with aviation human-factors accounts of performance under unexpected conditions. Unexpected technical failures and automation surprises may require pilots to interrupt an inappropriate automatic response, reinterpret the situation, regulate stress reactions, maintain attention, and select an appropriate course of action [8]. Individual reactions to stress have also been associated with attentional control and performance during demanding flight assessment [66]. Moreover, pilots’ responses to abnormal events become less consistent when familiar events are presented unexpectedly [67], whereas unpredictable and variable simulator training can improve the transfer of learned skills to novel situations [68]. The retained items therefore align with individual psychological resources that may support adaptive functioning during stress and unexpected events. However, the CD-RISC-6 does not directly measure situation awareness, abnormal-event management, or actual decision-making performance.

This distinction is important because the one-factor structure of the CD-RISC-6 is a property of its measurement model and should not be interpreted as evidence that resilience itself is inherently unidimensional. Contemporary frameworks distinguish resilience-related attributes or resources, adaptive processes, and outcomes [69], and also emphasize that successful adaptation depends on interactions among individual, interpersonal, organizational, and environmental systems [70]. The CD-RISC-6 primarily captures self-perceived individual capacities related to coping, persistence, attentional stability, and recovery. It does not assess social support, crew coordination, organizational resources, operational competence, or the broader environmental conditions that may contribute to resilient functioning. It should therefore be understood as a brief indicator of overall individual psychological resilience for group-level research and supportive assessment, rather than as a comprehensive representation of the resilience construct or a complete replacement for the broader content of the CD-RISC-10 or CD-RISC-25.

The prespecified evaluation of the fixed CD-RISC-6 in the separate, non-overlapping sample provided a more stringent test than the post-selection analyses in the development sample. In this sample, the fixed six-item questionnaire was administered directly, and its item composition, scoring rule, and one-factor model were specified before the data were examined. The scale showed high internal consistency (Cronbach’s α = 0.891), and both the G coefficient (0.891) and phi coefficient (0.878) supported score dependability under the fixed six-item design. All standardized factor loadings were positive and ranged from 0.676 to 0.836, and the negative associations with occupational burnout were consistent with those observed in the development sample. Nevertheless, the validation sample included only 58 participants, and the wide RMSEA confidence interval indicated limited precision of the CFA estimates. Sparse lower response categories also precluded a new graded response model calibration. Because both samples consisted of male airline pilots recruited from the same airline and recurrent-training context, the findings constitute preliminary within-context replication but do not establish structural stability, measurement invariance, or external applicability across sexes, airlines, aviation sectors, or operational settings.

4.3. Construct Interpretation and Implications for Aviation Human Factors Research

Within the present occupational setting, the CD-RISC-6 provides a low-burden summary of self-reported psychological resilience. The six retained items primarily represent perceived capacities for coping with stress, recovering from difficulty, maintaining goal-directed effort, sustaining concentration under pressure, and recovering from failure. In the separate, non-overlapping same-airline sample, the fixed short form showed high score reliability, CFA point estimates compatible with the prespecified one-factor structure, and negative concurrent associations with occupational burnout. Because both samples came from the same airline and recurrent-training context, this constitutes preliminary within-context evidence rather than validation for unrestricted operational application.

Importantly, brevity should not be equated with comprehensive coverage of the resilience construct. Resilience has been conceptualized as a set of personal resources or capacities, an adaptive process, and an outcome following exposure to adversity. It may also emerge through interactions among individual, interpersonal, organizational, and broader contextual systems [65,69,70]. Accordingly, the unidimensional CD-RISC-6 score should be interpreted narrowly as a concise indicator of pilots’ self-perceived individual capacity to cope with and recover from stress. It does not encompass the full cognitive, affective, behavioural, interpersonal, team, organizational, and context-dependent dimensions of resilience. Nor does it directly assess non-technical skills, crew resilience, abnormal-event management, decision-making stability, or actual operational performance. The lower measurement contribution of the four removed items in the development sample should therefore not be interpreted as indicating that adaptation to change, humorous coping, management of unpleasant emotions, or global self-perceived strength is theoretically irrelevant. Rather, it indicates that these items contributed less incremental measurement information under the specific response distribution observed in this sample. When comprehensive construct coverage is required, the CD-RISC-10 or a multimethod assessment approach should be preferred [63,64].

Subject to further validation, the principal value of the CD-RISC-6 may lie in group-level aviation human factors research, the examination of relationships between resilience and occupational variables, and supportive training feedback. It should not be used as a standalone instrument for clinical diagnosis, fitness-for-duty decisions, personnel selection, classification of operational risk, or punitive management. Individual scores should be interpreted, where appropriate, alongside interviews, validated mental health measures, simulator or training performance, instructor observations, and information about organizational and operational demands. This supportive and non-diagnostic interpretation is particularly important in aviation settings, where concerns about occupational consequences may affect pilots’ willingness to disclose psychological difficulties or seek professional support [71,72,73].

The demographic comparisons conducted in this study were secondary and exploratory. Although differences were observed across groups defined by flight role and relationship/family status, age, flight role, cumulative flight time, and relationship/family status were interrelated in the present sample, and the cross-sectional analyses could not separate their independent contributions. Moreover, sex and operational context did not vary within the samples, and measurement invariance or differential item functioning across demographic subgroups was not examined. The observed group differences should therefore not be interpreted as evidence of causal effects of flight role or relationship/family status on resilience, nor do they justify role-specific intervention recommendations. Instead, they provide preliminary hypotheses for future studies using larger stratified samples, multivariable analyses, and formal tests of measurement invariance.

4.4. Limitations and Future Research

Several limitations define the scope of inference of the present findings. First, the development and validation samples consisted exclusively of male airline pilots recruited from a single airline in China while attending recurrent training. The current evidence should therefore be interpreted as applying to this specific population and measurement context. It does not establish the applicability of the CD-RISC-6 to all Chinese civil aviation pilots, female pilots, student pilots, general aviation or military pilots, helicopter pilots, pilots from other airlines or cultural settings, or pilots assessed under different operational conditions. The all-male composition precluded examination of sex-related differences [74], whereas recruitment from a single airline and training context precluded examination of variation attributable to organizational safety climate, route structure, aircraft type, operating environment, or training system. Although age, flight role, and cumulative flight time varied within the samples, the subgroup sizes were limited and no measurement-invariance or differential-item-functioning analyses were performed. Consequently, the demographic score differences reported in Section 3.5 may reflect true latent differences, subgroup-specific item functioning, sampling variation, or a combination of these influences.

Second, the cross-sectional and self-report design limits causal and predictive interpretation. The observed associations with occupational burnout provide concurrent evidence only and do not establish temporal direction or predictive validity. Responses may also have been affected by common method variance and self-presentation. In aviation settings, perceived occupational or aeromedical consequences may further influence pilots’ willingness to disclose psychological difficulties [71].

Third, the GRM calibration and item-selection results were derived from the development sample of 192 pilots. Responses were concentrated in the higher categories, and several lower categories were sparsely endorsed or unused. Methodological studies indicate that GRM parameter recovery can be affected by sample size and the distribution of the latent trait [75,76]. Changes in category use in a new sample could therefore alter the estimated thresholds, discrimination parameters, and relative item-information values. In addition, the more favourable CFA and GRM fit indices obtained for the CD-RISC-6 in the development sample may partly reflect sample-dependent optimization because the same data were used for item selection and post-selection evaluation. A new GRM calibration was not performed in the validation sample because its modest size and sparse lower-response categories would not have supported sufficiently stable estimation. The CFA in this sample was likewise based on only 58 participants; its estimates were imprecise and should not be regarded as definitive confirmation of structural stability. Consequently, the separate-sample findings did not establish the stability of either the factor structure or the IRT calibration and item-information ranking across samples.

Future research should first replicate the fixed CD-RISC-6 in a larger, separate sample drawn from the same target population. The six-item form should be administered directly and evaluated using a prespecified analysis plan without further data-driven item modification. Such replication should re-estimate the GRM parameters and examine the stability of category thresholds, discrimination parameters, item-information functions, and measurement precision across latent-trait levels. Subsequent studies should include pilots differing in sex, age, flight role, cumulative flight time, airline, cultural background, and operational context. Multi-group CFA and differential-item-functioning analyses should then determine whether the six items operate comparably across these groups.

Construct validity should also be evaluated more directly by comparing the CD-RISC-6 with established resilience measures based on different conceptualizations of resilience. In addition, item-content coverage was evaluated qualitatively rather than through structured expert ratings or cognitive interviewing; the formal content validity of the candidate CD-RISC-6 therefore remains to be established. Future validation should include independent expert assessment of item relevance and representativeness, together with cognitive interviews examining how pilots interpret the retained items and response options. Because the six candidate items are nested within the CD-RISC-10, the correlation between the CD-RISC-6 and CD-RISC-10 alone is insufficient to establish convergent validity. Future studies should therefore administer non-overlapping external instruments, such as the Brief Resilience Scale, which focuses on the perceived ability to recover from stress, and the Resilience Scale for Adults, which assesses multiple personal and interpersonal protective resources [77,78]. A priori hypotheses should be formulated regarding the direction and relative magnitude of correlations with these resilience measures and with theoretically distinct constructs. Comparisons with clinically characterized or independently defined high- and low-stress groups could provide additional known-groups validity evidence.

Finally, longitudinal and performance-based studies are required to determine whether CD-RISC-6 scores predict outcomes beyond concurrent self-report measures. Future research could examine prospective associations with occupational burnout, fatigue, psychological well-being, recurrent-training outcomes, simulator performance, decision stability, and the management of unexpected or abnormal events [3,8]. Test–retest reliability, sensitivity to change, and incremental validity beyond existing psychological and operational indicators should also be evaluated before the scale is considered for routine research or supportive assessment.

5. Conclusions

This study evaluated the psychometric performance and item-reduction potential of the CD-RISC-10 in a development sample of Chinese male airline pilots. The CD-RISC-10 showed high score reliability and expected negative concurrent associations with occupational burnout, while the CFA and GRM were consistent with an adequate approximate one-factor representation, with some uncertainty in global fit and localized model strain. A sequential decision process, in which CFA assessed the structural prerequisite, GT provided a statistical reference for scale length, IRT ranked item-level measurement contributions, and qualitative item-content review safeguarded coverage, yielded a candidate six-item form. In the development sample, the candidate CD-RISC-6 retained high score reliability, substantial item information, and a favourable post-selection one-factor fit pattern. A separate, non-overlapping same-airline sample provided preliminary within-context evidence regarding the score reliability and prespecified one-factor structure of the fixed six-item form, together with its expected negative concurrent associations with occupational burnout. These findings position the CD-RISC-6 as a candidate low-burden measure for further psychometric evaluation in comparable recurrent-training settings. However, larger and more diverse samples are required to establish external validity, measurement invariance, formal content validity, and the stability of the IRT calibration before broader operational use.